goldeneye-pc-port/docs/dev/findings.md

# PC port — finding log

The raw, chronological engineering record for the port: every runtime bug,
its root cause, and the fix. Kept unedited for transparency. Recurring
patterns are distilled in [../porting-notes.md](../porting-notes.md).

---
## 11. Scaffolding review: findings & resolutions

The initial scaffolding was reviewed against the sources. Each finding below
was verified and resolved in the same change. Code comments in `CMakeLists.txt`
and `port/src/` reference these labels (A1–A4, B1–B4).

### A. Build-blocking issues

* **A1 — Region/version macros.** The N64 `Makefile` defines a full per-region
  macro set (`VERSION_*`, `LANG_*`, `REFRESH_*`, `LEFTOVERDEBUG`,
  `LEFTOVERSPECTRUM`, `BUGFIX_R*`, `BYTEMATCH`) that the game code `#ifdef`s
  heavily (e.g. `VERSION_EU` in 60+ files). **Resolved:** `CMakeLists.txt`
  now emits the matching `REGION_DEFS` per `ROMID`, mirroring the Makefile
  exactly.
* **A2 — Missing source files.** The initial `SRC_ENGINE` list + "exclude all
  `io/*.c`" strategy missed files the included code depends on. **Resolved:**
  added `src/cfb.c` (defines `cfb_16`, the framebuffer),
  `src/libultrare/audio/*.c` (`drvrNew`/`env`/`reverb` — the audio "New"
  driver `synthesizer.c` calls), and `src/libultrare/io/vitbl.c` (defines
  `osViModeTable`, referenced ~15× by `fr.c`). `motor.c` / `pfsinit.c` /
  `pfsisplug.c` are **not** compiled — `joy.c` calls `osMotor*`/`osPfs*` only
  for accessory detection, so they are shimmed as no-ops (see B2).
* **A3 — `mainproc()` / `init.c` + the excluded engine files' symbols.**
  `mainproc()` is defined only in `src/init.c`, which the initial build
  excluded, contradicting `main.c` calling it. **Resolved:** `init.c` is now
  **compiled** (it provides `mainproc()` + the thread-setup helpers). Its
  N64-only `init()` is compiled but not called.

  Compiling `init.c` (and the other included engine files) pulls in a set of
  symbols that live in the EXCLUDED files (`sched.c`, `rmon.c`, `vi.c`,
  `stacks.c`). Each is now provided:
  * **`src/stacks.c`** (added to `SRC_ENGINE`) — thread-stack arrays
    `sp_boot`/`sp_rmon`/`sp_idle`/`sp_shed`/`sp_main`/`sp_audi` (+ `sp_debug`
    under `LEFTOVERDEBUG`). Pure data; the `SP_*_SZ` sizes in
    `bondconstants.h` match exactly, so `sizeof()` is correct.
  * **`port/src/gesched.c`** (replaces `sched.c`) — scheduler globals
    `os_scheduler`, `gfxClient[3]`, `g_schedViCurrentFrameBuffer`,
    `g_ViChangeVideoModes`, `g_ViModes`, `g_ViModePtrs`, plus
    `get_counters()` and `permit_stderr()`.
  * **`port/src/n64stubs.c`** — `init()`'s N64 symbols (segment starts,
    segment boundary pointers, `jump_decompressfile`, TLB, FPU CSR) and the
    rmon host-I/O functions (`rmonMain`, `osReadHost`, `osWriteHost`,
    `rmonGetToken`, `rmonStatus`, `osSyncPrintf`).
  * **`port/src/libultra.c`** — the libultra OS calls `init.c` makes, plus
    `osTvType` (region-based), `viInit`, `vi_c_debug_MQ` (from the excluded
    `vi.c`), and `osPiReadIo` (cartridge token read, called by `token.c`).

  Every external symbol referenced by the compiled set was cross-checked
  against its definition; no duplicates (all other `sp_*`/`rmon*` references
  are `extern` decls).
* **A4 — `libultra.c` missing the high-level message API + others.** The
  initial stub had low-level `osEnqueueMesg`/`osDequeueMesg` but not
  `osSendMesg`/`osRecvMesg` (used in 13/26 files) or several others. **
  Resolved:** `port/src/libultra.c` now provides `osSendMesg`/`osRecvMesg`/
  `osSetEventMesg`, the controller query/read API, `osPfs*`/`osMotor*`,
  `osSetTimer`/`osStopTimer`/`osSetIntMask`, `osViBlack`/`osViSetYScale`/
  `osViSetXScale`/`osViSetEvent`, `osUnmapTLB`, `__osGetFpcCsr`/
  `__osSetFpcCsr`, `piCreateManager`, and `viDebugRemoved`. The scheduler API
  (`osCreateScheduler`/`osScAddClient`/`osScGetCmdQ`) lives in
  `port/src/gesched.c` (it's game scheduler API, not libultra).

### B. Research-doc corrections

* **B1 — `G_SETTEX` appears unused.** `gsSPUseTexture` (the only emitter) is
  never called in the game code or `gmain.s`; the game uses standard texture
  commands the PD `fast3d` already handles. **Resolved:** §5, §8 (Phase 2),
  and §9 updated — Phase 2 shrinks to verifying the custom CC/RM modes; a
  `G_SETTEX` decode path is kept only as a safety net.
* **B2 — Saves are EEPROM-based, not Memory Pak/PFS.** The save I/O
  (`src/game/file2.c`) uses `joyGamePak*` -> `osEeprom*`; the PFS/motor code
  in `joy.c` is accessory detection only. **Resolved:** §7 and Phase 4
  updated — Phase 4 is a file-backed EEPROM + no-op PFS/motor shims.
* **B3 — `src/libultrare/` was unmentioned.** **Resolved:** new §3.2 documents
  the two libultra trees and exactly which files the PC build compiles vs.
  shims, citing `Makefile.libultrare` as ground truth.
* **B4 — `src/spectrum.c` does not exist.** **Resolved:** removed from the
  `CMakeLists.txt` excluded-files comment.

### C. Assembly (`.s`) file symbols

The A/B sweeps covered excluded **`.c`** files. A second symbol source is the
MIPS assembly (`src/*.s`), also not compiled for the PC. Sweeping every global
label in the `.s` files against the compiled C set found **10 more** undefined
symbols, in two distinct classes:

* **C1 — PRNG (`src/random.s`) — ported to C, not stubbed.** `randomGetNext`,
  `randomGetNextFrom`, `randomSetSeed`, and `g_randomSeed` are used for real
  gameplay logic: `randomGetNextFrom()` feeds the CRC (`src/game/crc.c`),
  `g_randomSeed` is persisted in replay state (`src/game/ramromreplay.c`), and
  `randomGetNext()` drives `RANDOMFRAC()`/`RANDOMGETNEXT_F32()`. A no-op would
  silently corrupt CRCs and replays. **Resolved:** `port/src/random.c` ports
  the three functions **verbatim** (each MIPS instruction mirrored with explicit
  `dsll32`/`dsrl32` masks for bit-exactness) and defines
  `u64 g_randomSeed = 0xAB8D9F7781280783ULL` (the two `.words` in `random.s`).
  Signatures match `src/random.h`. Verified bit-exact against a Python
  simulation of the assembly (the PRNG is effectively 32-bit stored in a u64;
  the initial seed's high 32 bits are exercised only on the first call).
* **C2 — RSP/ASP/GSP microcode segment markers — dummy definitions.**
  `rspbootTextStart`/`rspbootTextEnd` (`rspboot.s`), `gsp3DTextStart`/
  `gsp3DDataStart` (`gspboot.s`), `aspMainTextStart`/`aspMainDataStart`
  (`aspboot.s`) are ROM addresses of the RSP/ASP microcode, referenced by
  `src/audi.c` and `src/game/rsp.c` (declared `extern long long int <name>[]`).
  **Resolved:** `port/src/ucode.c` defines all six as `long long int <name>[1] = {0}`.
  The graphics path is safe permanently (fast3d interprets the GBI list directly
  and never runs the gsp3D microcode, so the size differences are never read).
  **Carried into Phase 3:** the audio path is the one place it could matter — if
  the port emulates the ASP by *running* the `aspMain` microcode, those two
  markers must point at the real microcode bytes in the ROM (PD `pd.ld`
  `RSP_TEXT_SEGMENT` model); if audio is CPU-only, the dummies are fine.
  A `TODO(Phase 3)` on the `aspMain*` dummies records this decision.

### D. Host-compiler portability (GCC 16 vs IDO)

The N64 build targets IDO (MIPS). The PC build targets GCC 16 (x86-64). IDO
tolerates a number of non-standard constructs that the decompilation relies on;
GCC rejects them as hard errors. Each finding below was verified and resolved
in the same change. Where a fix could not be made purely in `port/`, the minimal
`src/` change is `-DPORT`-gated (N64 build untouched) and flagged with a
`PC port:` comment. These are **portability** fixes, not game-logic changes.

* **D1 — `inherits` = struct inlining + duplicate member name.**
  `src/bondtypes.h:44` does `#define inherits struct`, so `inherits X;` inside a
  struct becomes `struct X;`. IDO resolves this by **inlining** X's members into
  the enclosing struct (C++-style base), which the decomp depends on (offset
  comments like `CCTVRecord.unk84` at `0x84`, and positional initializers like
  `New_CCTVRecord(pad)` = `{New_PropDefHeaderRecord(6), 0, pad+0}` only line up
  with the inlined layout). GCC 16 also inlines `struct X;` (correct layout) but
  **hard-errors** on duplicate member names; Clang treats `struct X;` as a no-op
  nested tag (wrong layout). No flag/pragma downgrades the duplicate-name error.
  A sweep of `bondtypes.h` found exactly **one** parent/child member-name
  collision: `CCTVRecord` redeclares `pad` (s32 @0x80) which `ObjectRecord` also
  has (s16 @0x08). The game code's `->pad` (e.g. `setupCctv`, prop.c) is used as
  the **pad index**, which `New_CCTVRecord` stores at 0x08 (the *inherited* pad),
  so IDO resolves the ambiguous `->pad` to the inherited one; CCTVRecord's own
  `pad` (0x80, "lookpad") is never accessed by name. **Resolved:** under `-DPORT`
  only, CCTVRecord's own `pad` is renamed to `lookpad` (layout byte-identical);
  the N64 build keeps `pad` (IDO tolerates it).
* **D2 — `port/shim/PR/gbi.h` needed an include guard.** The shim redefines the
  `Gfx` union members as little-endian `G*_le` typedefs. With no guard, a TU that
  includes it twice (directly + via another header) re-creates the anonymous
  `Gdma_le`/`Gtri_le`/… struct types → "conflicting types for 'Gdma_le'". The real
  header's guard is `_GBI_H_`; the shim now uses a distinct `_PORT_SHIM_GBI_H_`.
* **D3 — `New_Vector()`/`New_Coord3d()` called with zero args.** The decomp's
  macros are declared with exactly 3 params (`x,y,z`) and use the
  `IF_ELSE(IS_EMPTY(..))` trick to default each to 0, but the game code calls
  them with **zero** args (`New_Vector()` at chrai.c:1358/1378, `New_Coord3d()`
  at chrai.c:4491), relying on IDO's leniency with empty macro args. GCC rejects
  `New_Vector()` against a 3-param macro. A sweep found these are the **only**
  call sites (all zero-arg). **Resolved in `port/`:** `port/shim/bondtypes.h`
  redefines both as `#define New_Vector(...) {0,0,0}` / `New_Coord3d(...) {0,0,0}`
  after including the real header (the shim is found first via the include path).
* **D4 — local `#define osSyncPrintf()`/`(x)` arg-count mismatch.** Five files
  locally disable `osSyncPrintf` with a fixed-arity macro (`#define
  osSyncPrintf()` or `(x)`) but then call it with more args (2–4), relying on
  IDO's leniency. GCC errors ("passed N arguments, but takes just M").
  **Resolved:** the five local defines (bg.c, debugmenu_handler.c,
  debug_camera.c, deb_loadallmodels.c, initexplosioncasing.c) are made variadic
  (`#define osSyncPrintf(...)`); they still expand to nothing, so behavior is
  unchanged.
* **D5 — array-initialized-from-array.** `chraction.c:2485` had `s16 mrs[3] =
  metal_ricochet_SFX;` (a local array "initialized" by a global array). C requires
  a constant-expression initializer; IDO treated it as a runtime copy. **
  Resolved:** replaced with an explicit 3-element copy (no new includes needed).
* **D6 — flexible array member in a nested context.** `bg.c` declared `s_special
  portal specialportalarray[]` where `s_specialportal` has a flexible array member
  (`u8 portallist[]`). C forbids initializing a FAM in a nested context (an array
  element); GCC enforces it, IDO tolerated it. The code treats the data as a flat
  byte array (cast to `u8*` in `sub_GAME_7F0B37EC`, walked byte-by-byte). **
  Resolved:** defined `specialportalarray` as a flat `u8[]` with the identical
  byte sequence.
* **D7 — AI X-macro system (`chraidata.c`) — RESOLVED.** The AI command system
  (`bondaicommands.h` + `aicommands.def` + `CPPLib.h`) uses deep preprocessor
  metaprogramming (`SWITCH` with 49 fixed params + `IF_VA`/`IS_EMPTY`,
  `DEFINED(SETUPSUBROUTINES(ID))` token-pasting, `_AI_CMD_POLYMORPH` redefined
  per-include) that relies on IDO-specific `##`/expansion behavior. GCC rejects
  several of these (e.g. `pasting ')' and '_'` in `DEFINED(SETUPSUBROUTINES(ID))`,
  `SWITCH requires 49 arguments, but only 20 given`). Note `src/aicommands2.h`
  is a **pre-generated** header (from `tools/cmdbuilder.c`) already included by
  `bondaicommands.h:864`; the failing path is the *raw* `aicommands.def` include
  (bondconstants.h:731 for the `AI_CMD` enum, chrai.c:172/920 for the command
  table).

  A general GCC-clean reimplementation of `SWITCH()` was investigated and ruled
  out: its content arguments are single preprocessor arguments that expand to
  top-level comma lists, and the C preprocessor cannot detect where one content
  ends and the next `CASE`/`VAL` begins (arity is not recoverable after
  expansion). The three active `SWITCH()` call sites in chraidata.c (m_IdleAnimations,
  m_BashKeyboard, m_RunToBondPersistent) each have a **hand-written equivalent
  already present in the file behind `#if 0`** — the author's own reference form.
  Each was verified byte-identical to the IDO expansion of the adjacent `SWITCH`
  call (including the `IFNewRandomGreaterThan(N, lbl)` == `SetNewRandom()` +
  `IFRandomGreaterThan(N, lbl)` identity, confirmed against the runtime check in
  chrai.c:1605). **Resolved:** under `-DPORT` the three `#if 0` blocks are
  activated (the `SWITCH` calls become dead `#else` branches); `port/shim/bondaicommands.h`
  keeps the original 49-param `SWITCH` defined but replaces it with a marker
  that fails loudly if any *new* game code uses `SWITCH()` on the port. See also
  D8–D11 for the sibling paste/comma issues in the same macro system.
* **D8 — `MODELSKELETON`/`New_ModelSkeleton` paste failure.** The model-record
  macros in bondconstants.h write `SKELETON(##NAME##)`-style pastes that IDO
  tolerates but GCC hard-errors on ("pasting does not give a valid
  preprocessing token"). **Resolved:** `port/shim/bondconstants.h` re-emits the
  affected macros with equivalent byte-identical expansions.
* **D9 — `CPPLib.h` helpers.** The CPPLib metaprogramming header's
  `IS_EMPTY`/`IF_VA`/`DEFINED` family uses paste tricks that break under GCC.
  **Resolved:** `port/shim/CPPLib.h` provides a paste-free reimplementation with
  identical results for every usage in the tree (intercepted via the include
  path; inert on N64).
* **D10 — file-record macros paste `&` onto NAME.** `CHRFILERECORD`/
  `GUNFILERECORD`/`SUIT_LFRECORD` and `GUNSTATS` write `{& ## NAME ## _header, …}`.
  IDO tolerated the failed `&##NAME` paste; GCC hard-errors. **Resolved:**
  `port/shim/bondconstants.h` rewrites them with the `&` kept out of the paste;
  expansion is byte-identical (`&NAME_header`, `&NAME_stats`).
* **D11 — generated `CALL()` double trailing comma.** The pre-generated
  `CALL()` (aicommands2.h) concatenates `SetReturnAiList()` and
  `SetChrAiList()`, each ending in its own trailing comma, then appends its own
  separator → `…, ,` inside the array initializer. IDO accepted it; GCC
  hard-errors. The artifact byte is never executed: `AI_SetChrAiList(CHR_SELF)`
  switches to the called list at offset 0 and `AI_Return` resumes the return
  list at offset 0 (chrai.c), so anything after the `SetChrAiList` record in a
  `CALL` is dead. **Resolved:** `port/shim/bondaicommands.h` re-emits `CALL`
  without the artifact byte.
* **D12 — 64-bit pointer→integer in static initializers.** On a 32-bit target
  (MIPS) storing an array address in a 32-bit field is fine; on x86-64 GCC 16
  makes it a hard error ("initializer element is not computable at load time")
  that no warning flag suppresses. Two sites, two fixes:
  * `process_monitor_animation_microcode` jump targets: the monAnim script tables
    (chrai.h) stored raw `monAnim*` array addresses in the tvcmd word's 32-bit
    field. **Resolved:** under `-DPORT` the initializers store an *index* into
    `_PORT_monAnimPtrs[]` (defined in propobj.c after all 35 monAnim arrays, via
    a shared `_PORT_MONANIM_LIST` x-macro in chrai.h) and the interpreter
    resolves index→pointer at `TVCMD_SETCMDLIST`/`TVCMD_RANDSETCMDLIST`. On-script
    layout unchanged (12-byte tvcmd words, same opcode bytes).
  * `assets/obseg/setup/{u,e,j}/UsetuplenZ.c`: the `intro[]` table stored
    `&credits_data_0` in an `s32` slot. **Resolved:** PORT-gated replacement with
    `0` (the N64 initializer is kept in the `#else` branch). The value is only
    consumed by romCopy-style size arithmetic, which is inert until Phase 2.
    (These files are currently excluded from the PC build per D16; the patch
    keeps them compilable should any setup data be pulled into the host link.)
* **D13 — assorted IDO leniency hard errors.** Small strictness failures with no
  semantic content:
  * `front.c:2405`: bare `return;` in an `s32` function (GCC error). The sole
    caller ignores the value → PORT-gated `return 0;`.
  * `audi.c` `audioInit`: C++-style array initializer `s32 sp48[…] =
    CUSTOM_FX_PARAMS_N;` (IDO accepted, GCC rejects) → PORT-gated explicit
    `memcpy` of the same bytes.
* **D14 — `_Printf` prototype clash (xstdio.h vs xprintf.c).** The IDO printf
  engine is compiled for the PC (`src/libultrare/libc/xprintf.c` + helpers
  `src/libultra/libc/xlitob.c`/`xldtob.c`, added to CMakeLists) because
  `src/sprintf.c`'s `sprintf()` calls `_Printf` directly. xstdio.h declares it
  with `u8 *` params; xprintf.c defines it with `char *` — IDO-compatible,
  GCC-fatal. **Resolved:** `port/shim/libc/xstdio.h` renames the declaration
  (`_Printf_u8decl`) before pulling in the real header via `#include_next`, so
  xprintf.c's definition is the sole prototype. The only caller (sprintf.c)
  doesn't include the header and passes its char*-based outfun, matching the
  definition exactly.
* **D15 — host libc lacks IDO/K&R symbols.** MinGW provides none of: `bcopy`/
  `bzero` (declared in PR/os.h with `int` sizes), `__libm_qnan_f` (quiet-NaN
  helper used by gu/cosf.c and game/zlib.c), the `tlbmanage*` API (boss.c calls
  two of them; tlb_manage.c is excluded — N64 TLB management), and
  `g_ViXScales`/`g_ViYScales` (defined in vi.c on N64, written at runtime by
  fr.c). Also `chrObjRandomGetNext/SetSeed` + `g_chrObjRandomSeed` live in
  `src/game/chrObjRandom.s` (not built for PC). **Resolved:** all provided in
  `port/src/n64stubs.c` / `port/src/libultra.c`; the chrObj PRNG is ported
  verbatim from chrObjRandom.s into `port/src/random.c` (same xorshift as
  `randomGetNext`, separate state).
* **D16 — asset data strategy for the single host link.** The N64 build links
  each model/level file into its own ELF and `.incbin`s compressed ROM blobs;
  none of that transfers. Three buckets:
  * **Compiled for real** (self-contained, unique symbol names):
    `assets/animationtable_data.c`, `animationtable_entries.c`,
    `oddtextures.c`, `font_chardatae.c`, `font_chardataj.c`, `font_dl.c`,
    `rarewarelogo.c` → added to CMakeLists (`SRC_ASSETS`). Real font/image/
    animation-table data is therefore live in the binary.
  * **Excluded — symbol collisions:** every `assets/obseg/{bg,brief,setup,stan}/*.c`
    defines generic file-local globals (`header`, `room_data_table`, `intro`,
    `padlist`, …) that collide in a single link. The N64 build isolates them by
    per-file ELF linking; the PC build cannot. Their top-level symbols are
    stubbed instead.
  * **Stubbed — ROM-derived / absent:** `port/src/assetstubs.c` defines the 758
    model/level/text symbols referenced by ob.c's `file_resource_table`
    (one zero word each), the 14 `ramrom_*` replay pointers (NULL), and the
    `unknown2`/`unknown2_end` pair (same address → zero-length romCopy in
    title.c). All `*Segment*` linker-script markers (normally from ge007.ld)
    are defined as NULL `u32 *` in n64stubs.c, so address arithmetic computes
    zero lengths and any Phase-1 romCopy is a safe no-op. Real data arrives
    with ROM loading in Phase 2 (`port/src/romdata.c`), which will also replace
    the table's address arithmetic.
* **D17 — fast3d placeholder.** `port/fast3d/` contains only `gfx_api.h`; the
  real software RSP lands in Phase 2. `port/src/gfxstub.c` provides no-op
  implementations of every entry point (plus `gfx_current_dimensions`) so
  video.c links. **Delete it when the real fast3d sources are added** (same
  symbols → duplicate-definition error if both remain).
* **D18 — GE ROM header offsets.** The N64 header is GE-specific
  (`src/rom_header.s`): magic `0x80371240` @0x00, ROM name "GOLDENEYE" @0x20,
  cartridge ID "GE" **@0x3C**, country byte **@0x3E** ('E' US / 'P' EU /
  'J' JP), version @0x3F. The standard-N64 offsets (ID @0x38, country @0x3A)
  are wrong for GE — verified against `baserom.u.z64` (hexdump: `... 47 45
  45 00` at 0x3C–0x3F). romdata.c validates all four fields.
* **D19 — music table is big-endian in the ROM.** The `.music` section of
  `assets/music/music.s` (the `RareALSeqBankFile` header + 63 × 8-byte
  `music_table_entry` records) is stored in the ROM in **big-endian** word
  order: at 0x419790 the bytes are `00 3f 00 00 | 00 00 01 fc` = BE u16
  seqCount=63 (= NUM_MUSIC_TRACKS), BE u16 unk=0, then BE u32 offset 508 for
  the first track. The N64 build evidently assembles that section with
  big-endian output; the PC port never reassembles it — the CSV offsets are
  unaffected and `musicSeqPlayerInit()`'s romCopy + `musicSeqFileNew()`
  patching work on the raw bytes unchanged.
* **D20 — US music tracks are RLE-compressed in ROM.** The `music_file`
  macro incbins `build/u/assets/music/<name>.rz` under `.ifdef VERSION_US`, so
  every M* track lives in the ROM 1172/RLE-compressed; the filelist sizes are
  the compressed lengths and per-track `end_` symbols = start + csv_size are
  correct boundaries. Decompression happens at playback (Phase 3 audio), not
  at load — no Phase-1/2 action needed.
* **D21 — scanner-prefixed manifest names.** `filelist.<r>.csv` contains
  entries whose file column is prefixed with the scan origin, e.g.
  `assets/ge007.u.117880.jfont_dl.bin` (real name `jfont_dl.bin`).
  gen_romassets.py aliases such names to their suffix after the
  `ge007.<r>.<hexoffset>.` prefix, so font/image/oddtexture segments resolve.
* **D22 — CMake silently skips `.s` files without ASM enabled.** The first
  assembly source in the PC build (`romassets_u.s`) was added to the target
  but never compiled: `project(ge007 C CXX)` had no ASM language, and CMake
  does not error on unknown extensions — it just drops them, surfacing only as
  a wall of undefined asset symbols at link. Fixed with
  `project(ge007 C CXX ASM)`. Lesson: when adding a new source *extension*,
  verify the object file actually appears in the build graph.
* **D23 — absolute cart-address asset symbols (Phase 1 approach).** Instead of
  zeroed stubs (D16), every ROM asset symbol is now an ABSOLUTE cart address:
  `scripts/gen_romassets.py` emits `port/src/romassets_<r>.s` defining all
  obseg labels (in `file_resource_table.inc.c` order — ob.c computes file
  sizes as `table[i+1].hw_address - table[i].hw_address`, so order matters),
  ramrom files, music tracks/markers, and the ge007.ld segment markers as
  `0x10000000 + rom_offset` from `filelist.<r>.csv`. romdata.c maps the .z64
  at exactly 0x10000000 (VirtualAlloc preferred address), so `&symbol` is a
  live host pointer and the PI shims (`osPiStartDma`/`osPiRawStartDma` in
  libultra.c) service reads as bounds-checked memcpys. This makes both DMA'd
  assets (models, banks) and direct-read assets (fonts, image display lists)
  work with unmodified game code. `assetstubs.c` is deleted; n64stubs.c keeps
  only the pure-RAM segment symbols (`_bssSegmentEnd`, `_csegment*`,
  `_inflate/_gameSegmentVaddr*`). US-only for now: EU/JPN manifests have
  naming inconsistencies (trailing-Z add/drop, region subdirs) that need the
  same treatment before those regions build.

### E. Compile + link milestone (status)

All ~235 translation units compile and the target **links**: clean build from
scratch is `236/236` steps, zero errors (`ninja ge007 -k 0`), producing
`ge007.x86_64.exe`. Remaining warnings (~4.5k) are expected IDO-leniency noise
(int-conversion, implicit declarations, etc.), demoted via the CMake warning
flags.

**Phase 1 (boot to window) is done:** `romdata.c` loads and validates the ROM
(`baserom.u.z64` at the repo root, or a final .z64 in `data/` / next to the
exe), maps it at cart base 0x10000000 (D23); `video.c` opens an SDL2 640×480
window with a GL context and clears/presents every frame; `main.c` runs a
demo loop (ESC quits) in place of `mainproc()`, which is deferred until the
software RSP + scheduler can service real frames. Verified: ROM header passes
validation, 12 MB mapped at 0x10000000, window renders for the full test
duration. Asset symbol spot-checks against the ROM confirm correct offsets
(e.g. music bank BE seqCount=63 @0x419790, first track offset 508 → Mno_music
@0x41998C; sfx/instrument banks start with GE's `B1` header).

### F. Phase 2 runtime findings (threads, DRAM, addresses)

**Dxx index — jump to a finding, do not linear-read this section.** §F
covers D24–D69; the log continues in §H (D32 procedure, D70–D121).

| Range | Topic | Status |
|---|---|---|
| D24–D30 | green threads / kernel / crash handler | resolved |
| D31–D42 | boot ABI: ANIM_DATA, music seq-table, libaudio banks, Globalimagetable rebase, GL context, rsp toggle | resolved |
| D43–D50 | model-file loading → offline sidecar "Plan B" (`d43_emit.py`); D50.6 texCopyGdls full-slot copy | resolved |
| D51–D58 | font pixeldata, osGetCount tick rate, model RW pools, cseq BE header, synth slots, RLE menu bg, watch raw offsets, gun-barrel DL | resolved |
| D59–D68 | intro render: blood-RLE clobber, DMA validate, OSMesgQueue, HEADS/BODIES sentinels, romCopy width, image_entry, Globalimagetable BE→LE | resolved |
| D69 · D78–D84 | stage load (`load_bg_file`): bg/stan offline sidecar (`d69_emit.py`) + StandTile/bg_room_data ABI | resolved |
| D70–D74 | intro-logo pixels: C-array bswap, UV path, sinf/cosf `DVAL()`, texture-import truncation | resolved |
| D75 | front-end 3D model transforms (D77 audio split out below — resolved, unrelated cause) | OPEN (parked). **M-32 triage (see §F "D75 ADDENDUM"): (a) D73-scope-gap RULED OUT (gu tree fully endian-clean). Splits in two: Bug 1 = logo/photo transform = the parked D114/D116 fast3d viewport mirror (not game code). Bug 2 = absent animated models = category (b). M-32b runtime probe (`GE_D75=1`): `render_pos`/`dynAllocate`-arena hypothesis RULED OUT (render_pos valid + fresh each frame), model instances valid (nMtx 21/1), zero fast3d DL warnings — failure is downstream in `drawjointlist`/`dotube` vtx/node-DL resolution or an off-screen `basemtx`. Needs a drawjointlist-level probe. M-33 (upright captures, D168 fixed): "Bug 1 = D114/D116 mirror" retracted; **gun-barrel Bond RENDERS fine** (walk + fire, upright) — the "absent" reports were the flipped capture; **Nintendo logo genuinely broken** (renders as 2 white blobs, shifted left — a real `logoinst` transform/geometry bug, not a flip); cast roll not re-captured. See §F "D75 Bug 2 — M-33 UPDATE". **v0.2.0-pre playtest (user):** Nintendo logo still broken; Rareware logo nearly right but its texture filtering looks off; MISSION COMPLETE / mode-select screens have no 3D models at all, so the earlier "absent mode-select model" framing was inaccurate.** |
| D76 · D164 | disclaimer/legal screen only draws line 1 — **root-caused (M-31)**: `constructor_menu00_legalscreen` text loop bounds `legal_text_end` on `&legalscreen_MRD`, a linker-adjacency assumption that mingw breaks (`legalscreen_MRD` links 0x60 *before* `legalpage_text_array`) → `do{}while(ptr<end)` runs once. NOT an image-table bug (screen references zero `sImageTableEntry`). | fix proposed (not applied — `front.c` owned by another agent) |
| D159 | front-end wallet-Bond photo "interlaced"/combed (RC1 / D149) — `texSwapAltRowBytes` odd-row 8-byte pre-swap (N64 RDP odd-line TMEM XOR compensation) not reversed by fast3d | FIXED (`#ifdef PORT` no-op the swap in `image.c`) |
| D161 | Depot (`-level_30`) ceiling = bright-blue speckle + radial rays (B2). A CI8 tile drawn with `gsDPSetTextureLUT(G_TT_NONE)` was decoded against the stale `rdp.palette` → garbage. Fix: `#ifdef`-free narrow route CI→I when `palette_fmt == G_TT_NONE` in `gfx_pc.cpp import_texture()`. | FIXED (`port/fast3d/gfx_pc.cpp`) |
| D165 · D166 | input-layer polish (M-31, port-only, `port/src/input.c`): D165 front-end cursor is now a true ~1:1 pointer (P-controller estimating the game cursor via front.c's own integrator) instead of velocity²; D166 hipfire mouse pitch emits speed-proportional C-button pulses instead of a fixed digital threshold. | FIXED |
| D180 · D181 | native-PC input pass (QoL run, `qol/native-pc-input-menu`): D180 = `Input.MouseCaptureMode` click-to-lock + absolute-cursor menu tracking + aim-speed default retune (`port/src/input.c`, `video.c`, port-only); D181 = `Game.ScreenShakeIntensity` route-(b) hook in `src/fr.c viShake` (`#ifdef PORT`, default 1.0 = no-op). | LANDED, config-gated, feel-checks owed |
| D186 | **`Video.FpsCap` throttles the sim, not just presentation (M-39 diag, M-42 clamp).** The frame-pacing `sysSleep`+busy-wait in `sync_framerate_with_timer` (`port/fast3d/gfx_sdl2.cpp`) runs inline on the scheduler thread (`osSpTaskStartGo`→`gfx_run`→`swap_buffers_begin`), so a low cap blocks VI-retrace delivery and drags every game thread down to the cap rate (no N64 analogue — RSP/RDP are separate silicon). **Landed:** `0 < FpsCap < 30` → treated as `0` (uncapped) + warning, at both `videoInit` (normalises a bad `ge007.ini` for the next `configSave`) and `gfx_sdl_set_target_fps` (covers the F10-overlay live path); caps ≥ 30 unchanged. Verified: default `FpsCap=0` framediff 3/3 golden-identical; `FpsCap=10` now runs at the uncapped rate; `FpsCap=60` still paces. **Owed:** move pacing off the scheduler thread so any cap only drops presented frames. See §F "D186" + porting-notes.md §E. | CLAMP LANDED (`port/fast3d/gfx_sdl2.cpp`, `port/src/video.c`); real fix owed |
| D187 | **HUD ammo-type/weapon icon vanished at `-O2` (M-46 root-cause, M-47 fix + verify).** — full `## D187` entry at file tail | FIXED (`src/game/gunfire.c`, `#ifdef AVOID_UB`); PR #19 |
| D188 | **Linux `-O2` boot crash: `_Printf` passes `&args` of a by-value `va_list` param (M-48).** — full `## D188` entry at file tail | FIXED (`src/libultrare/libc/xprintf.c`, `src/sprintf.c`, `#ifdef PORT`); PR #22 |
| D189 | **Linux `-fstack-protector` aborts on the decomp's latent stack-buffer overruns (M-48) — FIXED.** — full `## D189` entry at file tail | FIXED (`CMakeLists.txt`, `src/game/stan.c` `#ifdef PORT`); PR #23 |
| D190 | **Linux `-level_45` SIGSEGV at load in `add_ptr_to_objective` (`objective.c:64`) (M-48, open).** — full `## D190` entry at file tail | **CLOSED — not a port bug; level 45 is MP-only and has no solo setup file in the ROM (root-caused M-105, reproduced on Windows).** Full chain… |
| D191 | **Windows `-O2` release-bundle crash: SIGSEGV in `modelGetNodeRwData` (`src/game/model.c:478`) — the `root->Parent` node-tree walk (M-49, user playtests of the…** — full `## D191` entry at file tail | FIXED + PLAYTEST-VERIFIED (M-79).… |
| D193 | **NPC AI locomotion is too slow — characters don't move to their destinations at the original game's rate (M-49, user playtest; clarified after the FULL 21-level…** — full `## D193` entry at file tail | **FIXED + user-playtest-verified (M-81) —… |
| D194 | **Mouse input needs another tuning pass — RMB aim over-sensitive + sensitivity couples to frame/sim rate (M-49, user playtest).** — full `## D194` entry at file tail | **PARTIAL — M-120 human playtest… |
| D195 | **Transparency / alpha surfaces broken on Control (M-49, user playtest)** — full `## D195` entry at file tail | **FIXED (`port/fast3d/gfx_pc.cpp`, `gfx_sp_vertex`) —… |
| D197 | **Character face/head textures wrap around the head, seen on Silo (M-49, user playtest).** — full `## D197` entry at file tail | CLOSED — user playtest v0.2.0-pre |
| D198 | **`alAdpcmPull`/`_decodeChunk` used `K0_TO_PHYS` (an unconditional `& 0x1FFFFFFF` mask) on 64-bit heap pointers (`src/libultra/audio/load.c`).** — full `## D198` entry at file tail | FIXED (`src/libultra/audio/load.c`) — part of the Phase-3 mixer landing, not yet build/runtime-verified in isolation |
| D199 | **Phase-3 software audio mixer: macro-swap `aXxx` opcode execution + reverse-engineered `aSetBuffer` persistent-context semantics (`port/src/mixer.c`…** — full `## D199` entry at file tail | LANDED (`port/src/mixer.c`, `port/include/mixer.h`, `include/PR/abi.h`) — functional on `-level_09`, context-inference risk noted, needs a wider level sweep |
| D200 | **`-level_09` segfault: reverb's `ALDelay.input/output` are u32 — `-d->output` zero-extends to a +4GB pointer offset on 64-bit (D3x class, FIXED).** `src/libultrare/audio/reverb.c` back-references the delay ring as `&r->input[-d->output]`; on N64 (s32 `ptrdiff_t`) the u32 negation wraps to a small negative index, on PC x86-64 it zero-extends to +4294967136 samples → `_loadBuffer`/`_saveBuffer` write 13–52KB wild, trampling the delay array / `r->base` / adjacent heap in a self-propagating loop until unmapped memory. Fixed with `(s32)` negation casts at 5 sites (PD ground truth `n_reverb.c` does exactly this); ramalign line gets `(s64)`. Verified: `-level_09` runs 120 s+ crash-free, DRAM guard window over the FX region sees zero OOB writes. | FIXED — 5-line cast in reverb.c (this session) |
| D196 | **OS mouse cursor stays visible after closing the F10 options overlay in click-to-lock mode (M-49, user playtest).** — full `## D196` entry at file tail | **FIXED (`6e758f1d`, merged to `main`) —… |
| D192 | **Front-end mission/level-select grid pointer still can't reach the outer cells — but only under `Input.MouseCaptureMode=0` (legacy always-grab) (M-49, user…** — full `## D192` entry at file tail | CLOSED — the buggy mode 0 was removed entirely in the v0.2.0 prep pass (see D239); click-to-lock is now the only capture mode, so the unreachable-cells path no longer exists |
| D221 | **SELECT FILE screen: the mouse hit-box for a file slot is vertically offset from its visible box — you have to click *below* a file to select it, hovering directly over it doesn't (M-87, user QA, default `MouseCaptureMode=1`).** Different symptom from D192 (D192 is an unreachable-outer-cells *clamp range* bug under legacy mode 0; this is a per-row hit-test/highlight *offset* bug under the shipped default mode 1). | **FIXED (M-91) — root cause was NOT a coordinate/origin mismatch: it's a compiler stack-layout dependency in `interface_menu05_fileselect` (`front.c`).** The hit-test projects each wallet's model extents via `projectRectCornersTo2D`, which takes `{min,max}` `coord2d` *pairs* — but the file-select caller passed `&xmin, &ymin` (four separate f32 locals), relying on the N64 compiler laying out the max adjacent to each min. On PC (declaration-order locals) `arg1->f[1]` reads **ymax** as an X coordinate and `arg2->f[1]` reads an unrelated stack float (observed: `2.25734`, later `nan`) — so the vertical hit band collapsed to a ~14 px strip at the bottom edge of each wallet, sometimes NaN (never selectable that frame). Exactly "click below the file." The other caller (`objective_status.c:525`) passes real `coord2d` structs and was always fine. Fix: pass explicit `{xmin,xmax}`/`{ymin,ymax}` pairs under `#ifdef PORT` (identical values to what the N64 binary used; N64 path untouched). Verified headless: pre-fix A-press at cursor (154.7,127.8) over folder 0 never selected; post-fix same press selects folder 0 and advances. **Human mouse-playtest (this session): confirmed clicking directly over each visible file wallet selects it immediately, no aiming low required; consistent across all slots and the Copy/Erase icons; no crash.** Full details in §H "D221 — M-91". |
| D222 | **High `Video.FovScale` breaks culling — NPCs and some world objects stop rendering near the screen edges (M-87, user QA).** Split out of D211's noted residual: `currentPlayerSetCameraScale()` / the game's frustum-cull and screen↔world math still run at the **nominal, unscaled** fovy while the actual render projection is widened by `portFovScale` (`fr.c:737`) — so objects that are on-screen in the widened view but outside the nominal-FOV frustum get culled before they reach the renderer. Confirmed **not cosmetic** — a real, noticeable functional regression at high FovScale, not just "culls a hair early." | **FIXED, awaiting visual verification (M-121).** Root cause traced one step further than the original write-up: `currentPlayerSetCameraScale()` (`game/bondview.c`) derives both the frustum-cull plane normals (`c_cameratopnorm`/`c_cameraleftnorm`) *and* the fog/LOD distance-visibility-fade scale (`c_scalelod`/`c_lodscalez`, consumed by `chrobjFogVisRangeRelated()` et al. in `propobj.c`) from `c_perspfovy`, which every `viSetFovY`/`viSetAspect`/`viSetFov`/`viSetZRange` call in `fr.c` was setting from the **nominal** `g_ViBackData->fovy` — never the `portFovScale`-widened value D211 already computes for the actual render projection two lines away in `viSetupCurrentPlayerView`. Fix: factored the existing widen-FOV math out of `fr.c` into `f32 portScaleFovY(f32 fovy, s32 isTitleScreen)` (`port/src/video.c`, D222 — identical logic, same `LEVELID_TITLE` guard, no game-code duplication) and fed it into all 4 `currentPlayerSetPerspective()` call sites via a small `#ifdef PORT` local helper in `fr.c` (`frCullFovY`), same as the existing D211 projection-matrix chokepoint now also calls the shared helper instead of its own inline copy. Identity (bit-for-bit no game-code-visible change) at `FovScale=100` by construction — `portScaleFovY` returns `fovy` unchanged whenever `portFovScale == 1.0f`. Audited per the original risk note: `c_cameratopnorm`/`c_cameraleftnorm`/`c_scalelod`/`c_lodscalez` are consumed only in `propobj.c`'s fog-visibility-fade math and `bondview.c`'s own screen-space transform — no AI-visibility/awareness call sites found reading these fields, so `portFovScale` still only touches rendering-adjacent visibility, not NPC behavior. Rebuilt clean, crash-free at `FovScale=150` on `-level_33` (Dam). `bunker1` golden framediff was inconclusive in this dev box — the *unmodified* baseline build also fails it by a similar or larger margin (confirmed via an A/B stash/rebuild), matching the documented pre-existing frame nondeterminism (`porting-notes.md` §E), not a regression from this change. **Needs a human's eyes**: does high-`FovScale` screen-edge NPC/object culling actually look fixed now. |
| D223 | **QoL ask: mouse-wheel weapon cycling should be directional like a normal PC shooter — wheel up = next weapon, wheel down = previous — instead of both directions cycling forward (M-87, user QA).** **Root-caused, port-only fix, no game-logic change needed** — GE's engine already has separate forward/backward weapon-cycle signals: `bondview2.c` computes `moveData.weaponForwardOffset = edge(invButtons) && !(shootButtons held)` and `moveData.weaponBackOffset = (invButtons held) && edge(shootButtons)` (`bondview2.c:5338-5347`, the classic N64 "hold A, tap Z to cycle backward" trick). Default control scheme (the `else` branch used unless a Domino/Goodhead/Galore/Plenty preset is active, `bondview2.c:5177-5179`): `invButtons = A_BUTTON` (= port `GE_CONT_A`, `input.c:86`), `shootButtons = Z_TRIG` (= port `GE_CONT_G`, `input.c:88`, the fire button). `moveData.triggerOn` is explicitly suppressed while `invButtons` is held, so holding A never causes an accidental shot when the Z edge lands. The port's `inputPostWheel()` (`input.c:977`) currently does `if (notches < 0) notches = -notches` — **discards wheel direction** — and only ever synthesizes a bare `GE_CONT_A` pulse (`wheelPulse`, `input.c:631-634`), i.e. always "forward." | **FIXED (M-105, `port/src/input.c`) — directional wheel cycling implemented per the plan below.** Wheel-up keeps the existing 2-poll `A` pulse (`weaponForwardOffset`: fresh A edge, Z released). Wheel-down runs a 2-poll backward sequence: poll 1 presents `A` only, poll 2 presents `A`+`Z_TRIG` — the fresh-Z-edge-while-A-held combo `weaponBackOffset` requires (`bondview2.c:5337-5340`); ordering matters because the edge is computed against the game's own previous poll. `moveData.triggerOn` (`bondview2.c:5419`) requires A to be released, so the synthesized Z edge can never fire a shot. Roles are stable: the port forces `CONTROLLER_CONFIG_HONEY` (`bondview.c:1484`), whose scheme gives `invButtons=A_BUTTON`/`shootButtons=Z_TRIG` — the same roles the pre-fix forward pulse already relied on. Known limitation (pre-existing, not a regression): while RMB aim is held the game flips `controldef` to KISSY (`bondview2.c:6106-6112`), under which *neither* direction cycles — exactly as the old forward-only pulse behaved in that state. Owes a human play-test (wheel up = next weapon, wheel down = previous, no accidental shots). Plan: keep `wheelPulse` (or a new `wheelPulseFwd`) for wheel-up as today; add a `wheelPulseBack` path for wheel-down that instead holds `GE_CONT_A` across ≥2 polls while presenting a fresh `GE_CONT_G` edge partway through (mirroring the invButtons-held + shootButtons-edge combo) — sequence matters (A must already read "held" `oldbuttons`-wise before G's edge, per `weaponBackOffset`'s check order). Must confirm which N64 control preset (`controldef`) the port's save data actually carries — `invButtons`/`shootButtons` differ under KISSY/GOODNIGHT/SOLITARE/etc; if the port doesn't force one, the synthesized combo needs to match whatever's active, or the port should just force the default (HONEY-equivalent) preset since its own `Input.Bind` system already replaces N64-style scheme selection. Port-only, `#ifdef PORT` in `input.c`, no `src/` game-logic touch. **M-87: PD-PC-port precedent checked (user asked what PD did about N64 control presets).** PD does **not** route mouse/gamepad through the N64 preset system at all — it defines its own logical control-key layer (`CK_*` in `port/src/input.c`/`input.h`, e.g. `CK_DPAD_L`/`CK_Y`) fully decoupled from the N64 controller-config enum, and even treats mouse-wheel-up/-down as **two independent bindable virtual keys** (`VK_MOUSE_WHEEL_UP`/`_DN`, defaulted to distinct `CK_*` actions) rather than one "wheel = cycle" event — exactly the shape of the fix this row wants. PD's `controldef` field is hardcoded (`player.c`/`playermgr.c` `controldef = 2`), not left to the original in-game control-style picker. **This GE port already has the equivalent layer for this purpose: D214's `Input.Bind` system.** Recommendation for D223 (and the user's broader ask — N64 presets aren't meaningful for mouse/keyboard, and modern-gamepad (Xbox/PS) support should be a first-class port concept, not routed through N64 presets either): add two new `Input.Bind`-able logical actions (e.g. `IA_WEAPON_NEXT` / `IA_WEAPON_PREV`) that synthesize the invButtons/shootButtons combos under the hood, default-bound to wheel-up/-down (keyboard/mouse) and left open for a future bumper/shoulder-button default under real gamepad support; and have the port force a single fixed `controldef` (matching PD's approach) so the synthesis target never depends on the original N64 control-style menu. That menu itself becomes a candidate for hiding/no-op in a future QoL pass once `Input.Bind` fully supersedes it — not proposed as a change here, just noted as the natural conclusion. **M-87 — D194 gives a concrete answer for *which* preset: `CONTROLLER_CONFIG_SOLITARE` (1.2).** It's the scheme that gives free continuous full-stick mouselook (see D194) — and conveniently, `SOLITARE` is **not** in the `KISSY`/`GOODNIGHT` button-role group (`bondview2.c:5168-5179`), so it keeps the **same** `shootButtons=Z_TRIG`/`invButtons=A_BUTTON` roles this row's plan already assumed — no change needed to the wheel-cycle synthesis. (`GOODNIGHT` (1.4) also gets natural pitch/turn but groups with `KISSY` for button roles — `shootButtons=A_BUTTON`, `invButtons=L_TRIG|R_TRIG` — so it would need this row's combo re-derived; prefer SOLITARE unless something else needs GOODNIGHT specifically.) **M-87 survey note, filed for later (candidate 4, no action needed now):** GE's port has no per-player control-*mode* concept — keyboard/mouse is always controller 0 (`inputOpenPads()`, `connectedMask=0x1`), covering the common "P1 mouse+kb, P2+ pads" local-MP case fine. PD adds a 5th per-player control-mode entry ("PC controls") with a save-compatibility shim (written as an existing N64 mode so the save format never changes, overridden back to PC mode on read by config) to let *any* player be on mouse+keyboard, not just P1. Only relevant if "P1 on pad, P2 on keyboard" (or more than one keyboard player) is ever a real ask — bigger feature than anything currently planned; not worth a Dxx yet. |
| D224 | **QoL gap (M-87, from a PD-PC-port survey): the N64 Rumble Pak is fully dropped instead of modernized.** GE genuinely has a rumble subsystem (`src/motor.c`, `src/joy.c`) but the port's `osMotorInit/Start/Stop` (`port/src/libultra.c:1220-1223`) are all hard stubs — "no accessories on the PC," `osMotorStart`/`osMotorStop` return `-1` unconditionally, `osPfsFindFile`-family returns `PFS_ERR_NOPACK`. So every in-game rumble event (weapon recoil, explosions, etc.) computes and calls into a dead path. **PD's PC port wires the equivalent straight through to real gamepad rumble:** `inputRumbleSupported(idx)` checks `SDL_GameControllerHasRumble`/`SDL_JoystickIsHaptic` (with a Windows fallback: some pads report no haptics but rumble anyway), `inputRumble(idx, strength, time)` calls `SDL_GameControllerRumble()` scaled by a per-pad `RumbleScale` config value (options-menu slider, `Input.PadN.RumbleScale`), and `osPfsFindFile` reports a "Rumble Pak" present exactly when a real pad supports rumble (`libultra.c` — same "pretend the N64 accessory is there" pattern GE's port already uses for `osMemSize`/Expansion Pak, just applied to a real capability instead of faked). Cross-ref `QOL-INVENTORY.md` (folded into the existing "per-pad tuning" row). | **DEPRIORITIZED (M-87, user call).** Gamepad-only — irrelevant to mouse/keyboard, which is the primary target for this next release; no keyboard/mouse play is affected by NOT doing this. Root cause + fix still valid whenever gamepad support becomes a priority (no `src/` game-logic touch: GE's `motor.c`/`joy.c` call sites already exist and already call the `osMotor*` shims; the fix is entirely in `port/src/libultra.c` + `input.c` — add `inputRumbleSupported`/`inputRumble` per the PD shape, wire `osMotorStart`/`osMotorStop` to them, add `Input.PadN.RumbleScale` config). **M-87 PD-legacy-survey addendum (`docs/dev/notes/PD-LEGACY-SURVEY.md` candidate 1, high confidence) — the D224 fix as originally stated is NOT sufficient by itself.** Two more upstream blockers sit before the stubbed `osMotorInit/Start/Stop` ever gets reached: (1) `port/src/libultra.c:1096-1098` sets `g_contStatus[i].status = 0` for every pad — `CONT_CARD_ON` is never set, and `src/joy.c:184` gates the *entire* rumble-pak init path on that bit, so `joyRumblePakInit()` never even calls `osPfsInit`. (2) `src/joy.c:186-190` only attempts `osMotorInit` if `osPfsInit`'s return is `PFS_ERR_ID_FATAL` or `PFS_ERR_DEVICE` — but the port's stub (`libultra.c:1214`) returns `PFS_ERR_NOPACK`, so motor init is never attempted even once (1) is fixed. PD's exact trick (`libultra.c:213-227,349-351`): set `CONT_CARD_ON` for connected pads, and make `osPfsInitPak()` return `PFS_ERR_DEVICE` specifically when the pad supports rumble. **Whenever D224 gets implemented, both of these must land alongside the `SDL_GameControllerRumble` wiring** or the game's state machine parks at `RUMBLEPAKINITSTATE_NOT_READY` forever and nothing audibly changes despite looking wired up correctly. |
| D225 | **Dev-tooling QoL (M-87, user ask): a debug-build/env toggle that unlocks all missions + all cheats, for testing.** Root-caused a clean single choke point — **both** gates route through the same function: `bool fileGetIsCheatUnlocked(save_data *save, s32 cheat)` (`src/game/file2.c:391`) reads the save's `unlocked_cheats_1/2/3` bitfields and returns whether bit `cheat` is set. Mission-select lock (`src/game/file.c:59`, gates whether a briefing page's mission is selectable) and the cheat-menu availability check (`frontCheckIfCheatIsUnlocked`, `front.c:1009`, itself calling `fileGetIsCheatUnlocked` for the per-level `SP_LEVEL_*` cheat IDs) **both** call it — it's a pure, side-effect-free bitfield read (no save-data mutation), so short-circuiting it to always-true under a debug env var unlocks mission-select *and* the cheat menu in one change, matching exactly what was asked ("those two go a long way"). **Bonus, free once cheats unlock:** the ROM already contains real Rare-era developer cheats that are otherwise unreachable in normal play — `CHEAT_LINEMODE` (wireframe render — would make D222-style culling bugs trivial to see), `CHEAT_BONDPHASE`, `CHEAT_DEBUG_POS`, `CHEAT_DEBUG_UNK5` (`cheat.c` ~L687-707, `CHEAT_MASK_GLOBAL`) — worth trying once this lands, may be useful dev tools in their own right. | **OPEN — root-caused, not implemented.** Plan: `#ifdef PORT`, env-gated (e.g. `GE_DEBUG_UNLOCKALL=1`, catalogue in `GE-ENV-PROBES.md` once added) short-circuit at the top of `fileGetIsCheatUnlocked` — `if (getenv-cached-flag) return TRUE;` before the bitfield read. Default off ⇒ zero behavior change for normal play; when set, purely a query-time override (save file itself is never touched, so it can't corrupt a save or leak into the shipped default). No `src/` logic change beyond the guarded early-return (same class of edit as other `#ifdef PORT` dev-convenience hooks, e.g. `GE_STARTMENU`). |
| D226 | **QoL ask: player-adjustable HUD scale for the ammo counter, bottom-left pickup/item status text, and top-of-screen dialogue/subtitle text.** All three route through the same primitive, `textRenderOutlined()` (`src/game/textrelated.c:688`, doc comment: *"Used for ammo counter, bottom left HUD messages, countdown timers"*), called from two specific in-game (not menu) sites in `bondview2.c`: **ammo + bottom-left pickup/status** at `bondview2.c:10005` (`stringbuffer_lowerleft[...]`, anchored at `view_left`/`view_vert`), and **top-of-screen dialogue** at `bondview2.c:10163`/`10165` (`stringbuffer_top[...]`, anchored at `msg.x`/`msg.y`). `textRenderOutlined` → `textDrawGlyphQuad` (`textrelated.c:615`) emits each glyph as a plain RDP `gSPTextureRectangle` at the glyph's native font-metric pixel size, `dsdx`/`dsdy` = `0x400` (1:1 texel:pixel, no scale factor anywhere in the call chain). Same function is also used by menus/options/multiplayer UI (`front.c`, `options.c`, `mpmenu.c`) — **do not scale generically**, those have their own layout at fixed screen fractions and D211 already established HUD anchors must stay put; scale only the 3 in-game HUD call sites named above, about their existing anchor (so bottom-left stays bottom-left, top stays top — same anchor-preserving principle as D211's HUD-static check). | **OPEN — root cause / call sites identified, not designed or implemented.** Two viable directions: (a) a `#ifdef PORT` wrapper at just these 2-3 `bondview2.c` call sites that scales the emitted `gSPTextureRectangle` geometry (`xl/yl/xh/yh`) by a new `f32 portHudScale` (default 1.0 ⇒ byte-identical) while inversely scaling `dsdx/dsdy` to keep the same source glyph texels mapped onto the larger/smaller rect, re-anchored so the opposite edge (bottom for the lower-left block, top for dialogue) doesn't move; or (b) thread a scale parameter into `textRenderOutlined`/`textDrawGlyphQuad` itself, defaulted to 1.0 at every other call site. (a) is more surgical (matches the "don't touch shared menu text" constraint) but duplicates a little geometry math; (b) is more central but touches a much more widely shared function. No PD precedent found (PD's port has no equivalent HUD-scale option). F10 overlay slider candidate once implemented, alongside `Video.FovScale`. |
| D227 | **Sky renders as two independently-moving halves — reads as "two sky processes" (M-93, user QA on PR #43's debug build, `-level_22` Statue and Cradle).** Each half of the screen shows a different sky pattern, and the two halves scroll/move at different rates relative to each other. Surfaced *while* verifying D176(a)'s perspective-correct-interpolation fix — D176(a)'s fix measurably improved the scroll-speed symptom on the visible pattern but did not touch this; this is a distinct defect, not a regression from that PR (unclear yet whether it pre-dates PR #43 or was merely easier to notice once the scroll speed calmed down enough to see the seam). User flagged one confound worth ruling out first: might be a widescreen/aspect-ratio artifact of rendering at modern resolutions rather than a logic bug — original N64 output is 4:3/{something narrower}, and the sky-quad tiling math (`skyPortRenderPoly`, `sub_GAME_7F097388` upstream) may assume the original screen split. | **FIXED (M-100b, PR #43) — root cause was a unit error, and it also fixed the "sky scrolls too fast" complaint (M-96) and the over-tiling.** `unk20`/`unk24` are already in the RDP's native S10.5 1/32-texel units; M-82 read them as texel counts and baked `tc = (S - fold) * 32`, making every sky texcoord 32x too large and overflowing the `s16` `tc`. Human play-test confirmed by the user on the previously-failing levels. Full write-up: §H "D227 — M-100b". Historical investigation trail below. **OPEN — visual evidence captured + a concrete mechanistic hypothesis found (M-94), not yet fixed.** Headless `-level_22` capture confirms a real, visible crease/discontinuity in the cloud texture — not a rendering artifact of viewing it live, and present in a single static frame (rules out the "just camera pan" explanation for at least this instance). Screenshot: `docs/img/bugs/d227-sky-seam-statue.png`. **Leading hypothesis:** `skyPortRenderPoly` (D176(a)'s fix) computes `wScale` per call — i.e. per individual triangle in the sky fan (`skyRenderTri`/`skyRenderFull` issue up to 5 separate `skyPortRenderPoly` calls per frame from a 5-vertex fan, `sky.c:900-920`/`1372-1393`) — from that triangle's own 3 vertices' `unk0c`. The position math is exact for any `wScale` (verified algebraically: `clip.x = (ndcX·w/wScale)·wScale = ndcX·w` regardless of `wScale`), so this is not a logic bug in the fix itself, but the `s16`-quantized `ob[]` storage means each triangle's shared-edge vertices get **independently re-quantized at a different granularity** depending on that triangle's own `wScale` — a vertex sitting on the seam between two triangles can round to a measurably different `ob[]` value in each triangle's independent call, producing a small position/continuity crack exactly at triangle boundaries. This would read as "two adjacent regions with different patterns" without requiring two draw calls or two cloud layers — one `skyPortRenderPoly` call producing a slightly different quantization per neighboring primitive is sufficient. **Not yet confirmed** — did not isolate camera motion from sky-scroll motion cleanly enough to independently verify the "each half moves at a different rate" part of the user's report, only the static crease. **M-95: FIXED (the static-crease part) + headless-verified.** Implemented the fix direction above: `skyPortBeginFan()`/`skyPortEndFan()` (new, `#ifdef PORT`, `src/game/sky.c`) compute one shared `wScale` across every vertex in the whole fan (`sp274[]`/`sp94[]`, up to 5 verts), called once right after each fan's vertex array is populated and before its 1-3 `skyRenderTri`/`skyRenderFull` calls; `skyPortRenderPoly` now uses that shared value when set, falling back to its old per-call-local computation otherwise (defensive — no other caller exists today, checked: `grep` finds `skyRenderTri`/`skyRenderFull` used nowhere outside this file). **Verified:** same-frame before/after screenshots at `-level_22` frame 345 — `docs/img/bugs/d227-sky-seam-statue.png` (before, visible kink ~x=1000/1280) vs `docs/img/bugs/d227-sky-seam-statue-fixed.png` (after, same frame, continuous streaks, kink gone). Full 21-level `verify.sh sweep`: 21/21 PASS both before and after this change (no regressions). `bunker1`'s `worst_cell` framediff number is unaffected by this change — reproduced identically on the pre-fix build, it's the already-known stale-golden-baseline issue (see M-84/M-83 notes; owed a re-baseline, unrelated to sky.c). **Cradle checked too:** the "black rectangle" artifact visible there is confirmed pre-existing (identical on a pre-fix build capture) — that's the separate, already-documented FOV-coupled black-sky-past-frustum-edge symptom (D218/D222 family), not this defect; didn't chase it here. **Not fully closed:** did not re-verify the "moves at a different rate" half of the original report post-fix (the M-94 addendum's static-camera diff measurement wasn't repeated on the fix); the aspect-ratio-artifact alternate theory was never separately confirmed or ruled out, though the fix visibly closing the crease at a fixed resolution is itself decent evidence the crease specifically was a code-side seam, not a display-scaling artifact. Human eyeball still owed before merge, same as every other rendering fix landed without display access this session. **M-96: human eyeball done — seam/crack STILL PRESENT on Statue with the M-95 fix in.** The static-crease theory (verified by the frame-345 screenshot diff) is therefore not the whole story, or not the dominant contributor to what's actually visible in live play; the never-repeated "moves at a different rate" temporal theory from the original report is the next thing to check — a single-frame diff can't see it. **D227 NOT resolved; PR #43 stays in draft.** **Also M-96: user reports the N64 sky should be *almost static* and this build's scroll is *extremely fast* — stronger than M-93's "still reads too fast," bears on D176(a)'s fix magnitude itself, not just this seam; possibly the same wrong-magnitude tc-delta drives both symptoms, not yet investigated.** Full write-up: §H "D227 — M-96". **M-98: seam residual FIXED (same bug class as M-95, one field over) + scroll-rate quantitatively measured.** Re-derivation confirmed the tick-rate side is clean: `skyTick()` (unmodified N64 logic) adds `g_ClockTimer` (verified via `frametiming.c` to run ~1/tick under normal load, D193/D204's territory) to `g_SkyCloudOffset` once per logic tick — not a per-render-frame or double-counted path, ruling out a global-timing explanation for "extremely fast" (that would show up as everything sped up, not just the sky). The seam, though: M-95 shared `wScale` across a whole sky fan but **left `skyPortRenderPoly`'s S/T fold (`foldS`/`foldT`) computed per-call, from only that triangle's own 2-4 verts** — the exact same class of bug M-95 fixed for position, just one field over, and never touched. Two triangles in one fan can independently floor a shared vertex's S/T to a different multiple of 64 texels; nominally invisible (GL_REPEAT period 64) but each triangle bakes `(S - fold) * 32` into its own `s16` Vtx buffer, and two folds 64 texels apart produce tc values 2048 apart — large enough to crack at the shared edge. Fix: extended `skyPortBeginFan()`/`skyPortEndFan()` to also compute one shared fold across the whole fan (`s_skyFanFoldS/T`), threaded into `skyPortRenderPoly` the same way as `wScale`. Verified: `-level_22` frame 345 (same frame M-95's before/after used) now shows continuous streaks with no visible crease at the ~x=1000/1280 location, vs. the still-visible crease in `docs/img/bugs/d227-sky-seam-statue.png`; `verify.sh bunker1` PASS (worst_cell unchanged, the known stale-golden-baseline number, not a regression). **Scroll-rate data point (the M-96/M-97-requested measurement):** cross-correlating consecutive-frame captures (`GE_PCDUMP=345-355:1`) at a static camera gives a consistent **~7 px/tick horizontal shift at 640px capture width** — a full 640px screen-width sweep roughly every 90 ticks (~1.5s at 60Hz), which is hard quantitative confirmation of "extremely fast," not just a subjective read. **Root cause of that magnitude is NOT yet found** — `g_SkyCloudOffset`'s accumulation (1 texel/tick) and the vertex S/T geometry it feeds (`sub_GAME_7F097388`) are unmodified N64 logic, so if the rate is genuinely wrong relative to N64, the bug must be in how many on-screen pixels one texel-tick maps to (camera-to-sky-plane geometry, FOV, or a scale baked into the D176(a) Path B substitution) rather than in the accumulator itself — not investigated this pass. Human eyeball still owed on both the seam-close and the (unfixed) scroll-speed complaint. **M-99: human playtest — seam STILL PRESENT after M-98, AND confirmed view-angle-dependent (moves as the mouse/view moves); also reported still bad on Cradle and present on Surface.** Ruled out the cross-quad water/cloud scale-mismatch theory on the two `IsWater==0` levels tested (Statue, Cradle) — untested on any `IsWater==1` level. Headless camera-pan repro via `GE_INPUTSCRIPT` confirmed NOT to work (doesn't reach real mouselook) — this needs either a human playtest or a new debug hook to chase further. Full write-up + ranked candidate list for next session: §H "D227 — M-99". **M-100: ROOT CAUSE FOUND AND PROVEN — `s16` texture-coordinate overflow.** A user-supplied emulator-vs-ours screenshot pair on Surface (same framing) showed the real symptom is a *radial starburst*, not a crack — which redirected the investigation off the quantization-mismatch track M-94→M-98 had been on. New `GE_D227V` per-vertex probe on the Cradle frame-350 camera: the sky is one quad spanning **29,830 texels in S and 29,929 in T**, but `Vtx.tc` is S10.5 in an `s16` = **±1024 texels**, so **7 of its 8 texture coordinates overflow and wrap to arbitrary values**; the fold cannot help against a span 29x the representable range. The quad is split `gSP2Triangles(0,1,3, 0,3,2)` on the 0–3 diagonal and vertices 0 and 3 are exactly the two that keep one coordinate in range — hence "two halves with different patterns, split by a line", the original D227 report verbatim. All four w positive, so no projective pole. Apparent view-dependence needs no view-dependent mechanism: S/T depend on view, so the wrap point moves. **Confirmed by diagnostic build:** baking tc at 1-texel instead of 1/32-texel granularity (32x range, deliberately wrong scale) makes the starburst and the split *completely disappear* (`docs/img/bugs/d227-cradle-tcscale-diagnostic.png`) — no second defect hiding behind it. **D227 == D176(a) "Defect 2" == M-98's "Cradle tower starburst"; M-98 misfiled that capture as unrelated, which is what made D227 look headlessly-unreproducible — it reproduces statically.** M-95/M-98 stay (per-fan consistency is what makes subdivision seamless). **Fix not implemented:** subdivide until span < ~960 texels/primitive; subdivision math and cut placement are worked out (S is linear in w, so cut uniformly in w — uniform-in-screen is badly wrong), but the vertex budget is an open decision (unchecked 50 KB/frame pool; pure tessellation wants ~40% of it). See §H "D227 — M-100". **M-100b: FIXED — and the real cause is one step further back: a UNIT ERROR.** M-82 ("Defect 1") read `unk20`/`unk24` as texel counts and baked `tc` as `(S - fold) * 32` to convert to S10.5. They are not texels: `skyRender` sets them as `worldpos * 0.1f` and the N64 path hands them to the RDP coefficient block unscaled (`skyRenderTri`: `S * unk34 * sp368`), so they are **already in the RDP's native S10.5 1/32-texel units**. That stray `* 32` made every sky texcoord 32x too large and caused **three symptoms tracked as separate bugs**: (1) D227 (spans of ~30,000 *texels* vs an `s16` holding ±1024 → overflow → starburst); (2) over-tiling (~466 repeats of the 64-texel cloud per quad instead of ~15); and (3) **M-96's "the N64 sky is nearly static, ours races"**, measured at ~7 px/tick in M-98 — `g_SkyCloudOffset` advances 1 unit/tick, which is 1/32 texel, not 1. **The scroll complaint and D227 were the same bug**, which is why M-98's tick-path re-derivation kept coming back clean: the accumulator was never wrong, its unit was. **Fix:** `tc = (S - fold)` 1:1, fold rebased to a whole tile period (2048 units); the span then fits an `s16` unaided — no tessellation, no tile shift, which is also why the original N64 code never needed either. **Verified:** Cradle frame 350 (`docs/img/bugs/d227-cradle-fixed-m100.png`) soft cloud bands at correct scale, no starburst/split/moiré; `GE_D227V` reports `kS=kT=0` and max tc 31,513 of 32,767; Statue frame 345 clean; `verify.sh bunker1` PASS. M-95/M-98 kept (per-fan consistency); the tile-shift machinery kept as a graceful-degradation safety valve (only ~4% headroom left on the observed span). **Human play-test still owed** — both prior "fixes" passed static checks and failed live. See §H "D227 — M-100b". |
| D229 | **Water on `IsWater` levels renders green and pulses blue<->green on an interval (M-101, user playtest, Frigate `-level_26`; Surface 2 also an IsWater level).** Reported alongside green weapon textures on Frigate/Surface 2 — **that half was NOT a bug**: the branch under test was 2 commits behind main and those 2 commits were exactly the D217 fixes (#47 texpool alignment, #49 CI cache palette-content key); after merging main the user confirms the gun is good. | **FIXED (port-layer) — root cause: water image is ZLIB+CI8; pool holds raw 8-bit indices; N64's custom ucode expanded them to RGBA16 in TMEM at load, fast3d read index byte-pairs as 16-bit texels (g=2·idx mod 32 dominates → green). Fix: track `src_fmt` per TMEM load; RGBA/16b tile over a CI-sourced slot now imports via the CI8 palette path (`gfx_pc.cpp` import_texture). Frigate verified live (user: "looks better"); sweep + Surface 2 by-eye still owed.** Pre-existing (A/B proven), mechanism identified as above. Built plain `main` with none of the D227 work: `-level_26` frame 340 shows the **same green** (only the tiling density differs), so D229 is independent of D227. The quad is `sky.c`'s `IsWater` block1 (Frigate confirmed `IsWater=1` via `GE_D176` — which also settles M-99's ranked item 1: that path is textured and does go through `skyPortRenderPoly`). Draw state from `sub_GAME_7F09343C` (`unk_092E50.c:278`): tile 0 and tile 1 both RGBA/16b at **the same TMEM 0**, `masks=maskt=5`, tile 1 offset by `gDPSetTileSize(1, uls=90, ult=150, lrs=0, lrt=0)`, 2-cycle `LERP(TEXEL1, TEXEL0, PRIM_LOD_FRAC)` with cycle 2 `= COMBINED * SHADE` and the LOD fraction driven by **`sinf()`** — that sine IS the reported pulsing, so **one texunit resolves blue and the other green** and the sine fades between them. **Ruled out: the shade path** — `skyChooseWaterVtxColour` with measured `env RGB=16,48,96` / `WaterRGB=255,255,150` spans dark blue to pale yellow and can never be green; measured pixels ~`(5,70..175,20..37)` (R and B crushed) put the green on the texel side. **Not established:** why one texunit decodes green when both tiles share TMEM and format. One unproven oddity: `uls/lrs` are `uint16_t` and `width = ((lrs-uls)>>2)+1` gives tile 1 `-22` -> **65514**, though the matching negative `tex_width2` makes the sub-tile-window branch skip and fall through to the correct mask-based 32-texel wrap, so it may be harmless. **Next:** widen the `GE_D172` multitex probe (it filters on `tile1.tmem != tile0.tmem`, equal here, so it never fires) to dump both texunits' decoded state; capture a *sequence* across the sine period, since a single frame is useless (frame 340 reads green on all builds); and check `skywaterimages[2]`'s real format vs the RGBA/16b `sub_GAME_7F09343C` re-declares. **D227 is deliberately decoupled:** M-100b's safety valve takes `allowShift` and the water quad passes FALSE, leaving that path byte-identical to pre-D227 — shifting both tiles was observed live to flatten the cross-fade into constant green. Full write-up: §H "D229 — M-101". |
| D230 | **In-level music quality: a bass instrument sounds corrupted on PC, correct on N64 reference (M-105, user reports; repro refined M-106).** Filed to give the post-D77 residual its own label — D77 (in-level music *absence*) is RESOLVED (PR #42, missing `return` in `sub_GAME_7F0C0BF0` zeroed the seqp volume on the only in-level trigger path), so "music doesn't play" claims in tracked docs are stale; what actually remains is this quality gap. **M-106 repro, deterministic (not intermittent):** Control's elevator track and its regular in-level track both carry the bad instrument throughout; Cavern's elevator track has it too; Runway's track is otherwise fine but the same instrument enters ~5–7s after the track starts and sounds bad from then on. User confirmed against original N64 reference audio that a bass instrument belongs there on Runway — this is **not** a wrong-instrument/wrong-bank selection (rules out the D77 trigger-path family and the seqp-volume fix's neighborhood); the instrument choice is correct, only its *sound* is wrong on PC. Determinism (same track, same spot, every time) points at a static decode/synthesis bug for this instrument's wavetable/ADPCM data — most likely low-frequency-specific (pitch-shift/downsample artifacting, a wrap/loop-point error, or a clipping/gain issue that only bites bass-range samples) — rather than a live mixing/voice-stealing issue (D147/D152/D202 family), which would be intermittent. The mixer layer and the trigger layer are both proven working by every track that sounds right, so the suspect neighborhood is the seq/instrument-bank decode path (`csplayer.c`/`synthesizer.c`, bank loading), not `port/` audio plumbing. | **OPEN — static-analysis pass done (M-108), not runtime-confirmed.** Repro: Control (elevator + regular track, throughout), Cavern (elevator, throughout), Runway (~5–7s in). Traced the suspect neighborhood to the software ADPCM decoder in `port/src/mixer.c` — the PC replacement for the RSP's `aLoadADPCM`/`aADPCMdec` ucode ops (`csplayer.c`/`synthesizer.c` only emit ucode command lists; they don't decode samples themselves). Found two hardcoded assumptions that don't hold for the general ADPCM-book format: (1) `sAdpcmTable` (`mixer.c:174`) is declared `[8][2][8]`, hardcoding predictor `order=2`, but `order`/`npredictors` are per-instrument authored fields (`include/PR/libaudio.h` `ALADPCMBook.order`; `src/libultra/audio/load.c:428-429` computes `bookSize` from them) that the bank-relayout code (`port/src/romdata.c:718-736`) already handles generically — if this bass instrument's authored book uses `order != 2`, the 2-tap recursion at `mixer.c:223-240` would decode it wrong; (2) even at `order=2`, `aLoadADPCMImpl` (`mixer.c:177-192`) silently truncates any book over 256 bytes (`mixer.c:183`, `npredictors > 8`), while the decode loop's 4-bit predictor-select nibble (`mixer.c:219`) can legitimately address up to 16 rows — an instrument with `npredictors > 8` would read truncated/garbage rows for some frames. **Caveat weakening confidence:** N64 SDK ADPCM in practice is conventionally always `order=2` (the RSP ucode's own decode algorithm is essentially fixed to a 2-tap recursion industry-wide), so candidate (1) may not apply to a real N64 title's data; candidate (2) (`npredictors > 8`) is the more plausible of the two absent contrary evidence. Not confirmed against GE's actual ROM-authored book data (no live capture taken this pass). Next: extend the existing `GE_MIXERTRACE=1` trace's `[LOADADPCM]` line (`mixer.c:187`, already logs `count` — `count/16 = order*npredictors`) to flag any book where `count > 256` (proves truncation) during a Control run, and correlate the anomalous book to its channel/instrument via `seqplayer.c:1120`'s program-change. If no anomaly shows, this whole neighborhood is likely a dead end and the next suspect is the resample (`mixer.c:295-341`) or envelope/gain (`mixer.c:379+`) paths for bass-range-specific behavior. **M-120 (local-Qwen triage dispatch): a third hypothesis proposed and CHECKED, RULED OUT.** The dispatched read-only pass proposed that `afFixupBook`'s `port/src/romdata.c:733-737` conversion of the ADPCM book coefficients (BE ROM → LE host, via `afRd16`/`afWr16`) is backwards and should stay big-endian. Verified against the actual code and disproven: (1) that conversion is the identical BE→LE pattern used for every other field in the same function (`order`/`npredictors` two lines above) and in the sibling `afFixupAdpcmLoop`'s ADPCM state array a few lines down — not something special-cased for this one book; (2) `port/src/mixer.c`'s `aLoadADPCMImpl` does a raw `memcpy` of the already-converted book straight into a host-native `s16[8][2][8]` table with no further byte-swap, consistent with the coefficients being intentionally pre-converted to host order; (3) decisively, `afFixupBook` is one shared code path applied uniformly to every ADPCM book in the game — if this conversion were actually backwards, every ADPCM-encoded sound would decode wrong, not one bass instrument on three tracks, and the vast majority of the game's audio is confirmed correct. Do not re-propose this specific byte-order theory without new evidence; the standing M-108 leads above (predictor-count truncation / `order` mismatch) remain the live ones. **FIXED (M-121, v0.2.0 finalization) — root cause was a plain byte-order bug in a corner neither M-108 nor M-120 looked at: RAW16 wavetables.** Offline decode of the local NTSC ROM's instrument bank found exactly two `AL_RAW16_WAVE` (type 1) wavetables — instruments 2 and 51, both sharing base 0x2118 / len 790 in the `_instrumentstbl` segment — which decode as a clean ~110 Hz soft-attack bass synth sample when read as **big-endian** s16. That is the missing bass instrument on Control/Cavern/Runway. On N64 the CPU reads those bytes BE, so it plays fine there; on PC the mixer's resampler (`port/src/mixer.c` `alRaw16Pull` path) consumes host-native (little-endian) s16, so every sample arrives byte-swapped — a deterministic per-instrument corruption, exactly matching the M-106 repro (and explaining why every ADPCM-based instrument in the game sounds right: nibble data is byte-order agnostic). Both M-108 candidates are moot — every ADPCM book in both bank trees (scanned offline from the ROM) is `order=2, npredictors∈{1,4}`, so neither the `[8][2][8]` shape nor the 256-byte truncation can ever bite GE's data. **Fix (port-only, D55 in-place-fixup idiom):** `romdataFinishCartMap()` (`port/src/romdata.c`) now walks both bank trees (SFX + instruments; same ALBankFile layout as `romdataFixupAudioBank`) and bswap16s each RAW16 wavetable's `[tbl+base, +len)` range in place in the cart map — where the sample data lives for the mixer (`alBnkfNew()` points at the tbl segments directly; only the bank structs are re-laid out, D37). Shared ranges are deduped (swapping twice would undo the swap); every read is bounds-checked against the ROM size. Startup log: `romdataInit: D230 byte-swapped 1 RAW16 wavetable range(s) in place` (one range: the two instruments share the sample data). Byte-level verification done (BE decode of the ROM range = clean tone; post-swap LE decode identical by construction); **owes a by-ear pass** on Control (elevator + level track), Cavern elevator, and Runway ~5–7 s in. |
| D233 | **All levels (intermittent): indoor level geometry invisible near doors — you can see through walls/floors into rooms you can't see into, while world objects (terminals, furniture) in that space still draw (M-106, user QA report; screenshots `docs/img/bugs/20260912-cullingbug-bugged.png` vs `-notbugged.png`).** Root-caused M-107 (session-016TaKy6, port/fast3d only): **fast3d's CPU-side "trivial reject" AND-test in `gfx_sp_tri1` (`port/fast3d/gfx_pc.cpp:1455-1460`) has a latent w-sign bug independent of D222/D106.** The per-vertex outcode bits computed in `gfx_sp_vertex` (`x<-w`/`x>w`/`y<-w`/`y>w`/`z>w`, `gfx_pc.cpp:1372-1386`) are only valid half-space tests when `w>0`; for a vertex that has crossed behind the camera plane (`w<0` — common when the camera is close to large room-structure polygons, e.g. walls/ceiling trusses near a doorway) the comparison sense is flipped, so its outcode can come out wrong. If that spurious bit happens to match the other two (genuinely correct) vertices' bits, `v1->clip_rej & v2->clip_rej & v3->clip_rej` fires and the whole triangle is dropped before it ever reaches OpenGL — even though GL's own homogeneous clipper (which *does* handle `w<0` correctly) would have rendered the visible part of it. Confirmed against the two screenshots: same room/camera/ammo-HUD, one frame is missing the room's ceiling trusses/walls (replaced by flat sky) while the console/terminal props in the same room still draw fully — exactly the "big background polys vanish, small prop models don't" split you'd expect, since large near-camera polygons are far more likely to have a vertex cross behind the eye than small self-contained prop meshes. The backface-cull code 6 lines below (`gfx_pc.cpp:1462-1495`) already special-cases this exact `w<0` hazard for its own cross-product sign (`(v1->w<0)^(v2->w<0)^(v3->w<0)`) — the trivial-reject block above it never got the same treatment. **Independent of D222** (D222 is a game-logic frustum-cull bug in `src/game/fr.c`/`bondview2.c` that only manifests at non-default `Video.FovScale`; this is a GPU-pipeline clip-rejection bug in `port/fast3d/` that fires at any FovScale, including the default 100, whenever the camera gets close to a large polygon — matching the report's "all levels, near doors" framing with no FovScale mentioned). Also independent of D106 (D106 fixed the *game-side* portal-visibility bbox computation for a different z==0 edge case in `bg.c`; this is the *renderer-side* per-triangle clip decision, a different stage of the pipeline entirely) — same neighbourhood/symptom vocabulary ("things vanish near a portal-adjacent boundary"), different bug. | **FIXED (port/fast3d/gfx_pc.cpp, `gfx_sp_tri1`) — high confidence on mechanism, UNVERIFIED visually.** Fix: skip the trivial-reject AND-test entirely whenever any of the triangle's 3 vertices has `w<0`, deferring to GL's clipper instead (mirrors the cull code's existing w-sign handling one block below). Change only ever *adds* triangles that were previously dropped — cannot make anything newly invisible, so it's a safe, narrowly-targeted fix with no game-logic touch (fast3d is PC-only code, no N64 build parity to preserve, no `#ifdef PORT` needed). **Build verified green** (`./build-pc.sh` equivalent — MSYS2 mingw64 cmake+ninja, full `ge007.x86_64.exe` link succeeds, only one pre-existing unrelated `-Wformat` warning in the same file). **VISUALLY VERIFIED (M-108):** user played `-level_26` (Frigate — corrected level id, the old `-level_02` note above predates the `bondconstants.h` numbering check) and confirmed the doorway/ship-corridor geometry no longer flickers or vanishes near room boundaries. Closed. |
| D234 | **Dam: guard towers along the dam structure invisible except the first one (M-106, user QA report).** Investigated M-107 (session-016TaKy6, port/fast3d scope): **NOT root-caused — no fast3d-side mechanism found that plausibly explains "only the first of N repeated instances renders."** Ruled out: fast3d has no matrix/model instance cache that could explain a first-instance-only bug (the D217-family "keyed by address, not content" cache-bug class doesn't recur here — the *only* content-addressed cache in fast3d is the texture cache, already excluded from this investigation per the M-106 triage note); the modelview matrix stack (`gfx_pc.cpp:105`, depth 11, push properly gated `< 11` at `gfx_pc.cpp:1220`) is correctly bounded and would affect every level's nested-DL geometry, not just Dam's towers specifically; the D233 w<0 clip-rejection bug (above) doesn't fit either — towers are viewed from a distance along the dam, not close enough for a camera-behind-vertex condition, and D233's mechanism is per-triangle/per-frame flicker, not a stable "always the same instances missing" pattern. **Leading hypothesis, NOT verified:** GE's road/structure-instancing for repeated architecture typically goes through the room-graph's bytecode "vis program" (`VISOP_*` opcodes interpreted in `bg.c` ~L4450-4620, e.g. `VISOP_PUSH_IF_ROOM_IN_RANGE`, `VISOP_ADD_VISIBLE_ROOM`) or through per-instance room/portal definitions in the level's compiled BG data — both are level *data*, not code reachable via static analysis without the ROM (this worktree has none extracted). Could equally be a per-instance object/prop transform or cull-volume bug in `src/game/propobj.c`/`chrprop.c` (out of this task's `port/fast3d/` scope, and out of the `src/game/*` model/anim exclusion in the brief regardless). **No fix attempted — do not have enough evidence to touch code.** | **VISUALLY VERIFIED FIXED (M-108) — mechanism still unexplained.** User confirmed on a live Dam playthrough (post-D233 fix, correct `-level_26`/Frigate session's build) that all guard towers along the dam now render, not just the first. No code change was made specifically for D234, and the M-107 investigation's own reasoning argued D233's mechanism doesn't fit this symptom (towers viewed from a distance, not a close-camera w<0 case) — so either that reasoning was wrong, D233's fix had a broader effect than analyzed, or the original symptom was intermittent/state-dependent rather than deterministic. Not closed with high confidence in *why* it's fixed, just closed on live confirmation that it *is*. If it resurfaces, re-open with a fresh `GE_PCDUMP` capture rather than assuming this write-up's ruled-out list still holds. |
| D244 | **Self-hosted CI runner (`corpyvt6agent01`) link failure: `undefined reference to 'objectiveGetStatus_WEAK'` (2026-09-12, first-ever build attempt on the reinstalled box, run `34707798707`).** `src/game/objective_status.c:162` uses `#pragma weak objectiveGetStatus_WEAK = get_status_of_objective` (a decomp ground-truth alias, referenced 2-arg from `chrai.c:2538` — comment calls it "horrible hack to get ai matching"). The box's `g++`/`ld` reported as `x86_64-w64-mingw32/16.1.0`; a minimal local repro of the identical `#pragma weak` cross-TU function alias **links cleanly** on this session's own machine (`mingw-w64-x86_64-gcc 16.2.0-3`, `binutils 2.47-3`). Not a game-logic bug (the alias is unmodified decomp code, ground truth) and not a CMakeLists/link-flag issue (no `--gc-sections`/LTO in play) — the leading, unverified explanation is the self-hosted box's MSYS2 packages predate a PE/COFF weak-external fix or behavior change versus current MSYS2, consistent with rollout step 2's earlier finding that the box's mingw64 package set was stale/incomplete at initial setup. | **OPEN, not yet fixed.** Next: on the box, `pacman -Syu` (or at minimum `pacman -S mingw-w64-x86_64-gcc mingw-w64-x86_64-binutils`) to bring the toolchain current, then re-dispatch `selfhosted.yml`. If updating doesn't clear it, the fallback is a narrow `#ifdef PORT` replacement of the alias with an explicit thin wrapper function in `objective_status.c` (ABI-preserving, no logic change) — but try the toolchain update first since the local repro suggests this is a version/environment gap, not a fundamental PE/COFF limitation. |
| D228 | **AK47 (and other) weapon models still show WHITE missing-texture patches after D217's texpool-alignment fix — a separate failure mode from the green wrong-palette defect (M-97, user playtest of PR #47, `-level_34` FACILITY).** Split out of D217 defect ①, exactly as the M-90 write-up anticipated ("do not close D217 defect ① on this merge — reopen/retitle as the AK47 + NPC-world-model white-texture residual"). **Playtest result that motivates the split:** with PR #47's `texInitPool` 8-byte arena alignment in, the *green* symptom is GONE — PP7/PPK grip and hammer render correct, Bond's watch face correct, no green cast on hands or the 3rd-person world model. But **NPC-held AK47s still render with white patches**, i.e. tiles sampling a missing/undecoded texture rather than a wrong-but-present palette. White ≠ wrong colour: a wrong palette binds a real TLUT with wrong contents (D217's mechanism, now fixed), whereas white is the signature of a texture that never got decoded/uploaded at all, or a tile descriptor pointing at an empty/cleared texpool slot. The M-87 matrix described these as distinct from the start ("AK47 viewmodel has white (missing-texture) patches on the magazine + barrel" vs. the green/blue colour corruption) — this playtest confirms they have distinct root causes, since fixing one left the other untouched. | **FIXED (`port/fast3d/gfx_pc.cpp`, `palette_to_rgba32`) — a real byte-order bug, not a texpool/decode-never-ran issue as originally guessed.** `GE_TEXDUMP` on Facility (`-level_34`) found ~14% of LOD-textured uploads (fmt=CI, `palette_fmt=G_TT_IA16`) rendering solid near-white; a `GE_D228` correlation probe ruled out the texpool/RC2-clip hypothesis (the decoded size exactly matched the tile's base image every time — the data reaches `import_texture` fine, sized correctly). The actual bug: `palette_to_rgba32`'s `G_TT_IA16` branch read `intensity = palentry & 0xff` (low byte) and `alpha = palentry >> 8` (high byte) — backwards relative to the already-correct direct-IA16 decode two functions above it in the same file (`import_texture_ia16`: `intensity = addr[2*i]` (first/high byte after `gfx_dp_load_tlut`'s `PD_BE16` swap), `alpha = addr[2*i+1]`). A dim, fully-opaque palette entry (e.g. intensity=0x58, alpha=0xff) decoded backwards as bright and mostly-transparent (intensity=0xff, alpha=0x58) — exactly the white-patch symptom, on any CI-format material whose TLUT happens to be IA16 rather than RGBA16 (weapon metal/grip materials commonly are, since IA is a natural fit for grayscale metal with a separate opacity channel). Fix: swap the two lines to match the working convention. **User-verified live:** AK47 and other metallic guns render correctly (no more white patches). Verified crash-free with a `-level_34` headless run before the live check; a full 21-level sweep is still owed (not run this session — the user was using their own game instance concurrently, `verify.sh`/`taskkill` would have killed it). D217 · D161 lineage — same general "CI/LUT format-assumption" bug class, distinct specific bug. |
| D204 | **PC audio ran ~2 % below real time permanently: GE's AI feedback loop (`src/audi.c:531`) has a ~3 ms setpoint that PC scheduling jitter clears, so the SDL queue starves and the device pads playback with silence (FIXED, measured).** `amMain` wakes at 30 Hz (`sched.c:334` forwards every 2nd retrace to the audio client) and asks for `g_MinFrameSize`=720 samples per block = 21600/s against a 22050 Hz device — structurally 2 % short, relying on 784-sample top-up blocks that `audi.c` only requests once the reported AI length falls under ~69 frames (3 ms). Fine on N64 (double-buffered AI, exact VI interrupt); on PC the queue empties before the loop reacts, and the padded silence is unrecoverable time. **Fix (F5, port-only):** `audioGetAiLengthBytes()` subtracts a 1024-frame (~46 ms) cushion before reporting, so the loop tops up while slack remains; `audi.c`'s control law untouched (rule #2 clean). Measured A/B in one binary via `GE_D204_OLD`: **rt 0.980 → 1.000, q min 0 → 368, drop 0.** Also: F1 correct AI_LEN_REG single-buffer semantics (robustness, measures as a no-op), F3 `Audio.QueueLimit` 8192→2880 + non-silent drops (NB: an existing `ge007.ini` pins it), F4 oversize invariant guard, and `GE_D204=1` cheap audio-health monitor (`rt`/`q`/`drop`/`max`, safe to leave on for a full playtest). **Falsified in-session, do not re-open:** the u32 wrap at `audi.c:531` is real (high-water 2064 frames > the 789 threshold) but harmless — `frameSamples` is `s16` (`audi.c:145`), so it lands negative and audi.c's own lower clamp catches it; max block stays 3136 B vs the 3156 B allocation, no heap overrun. **Corrects two M-63 claims:** `soundIndex=109` is guards' return fire (count was double-logged; the "32.6 ms cadence" is just the 2880-byte block quantum every `dumppos` is rounded to) — closed negative; and M-63's "13 % of real time" was `GE_MIXERTRACE`'s own 25 MB unbuffered log starving the audio thread (same repro without it: 58.0 s / 60 s). Does NOT explain D202. | FIXED (`port/src/audio.c`, `port/src/libultra.c`, `port/include/audio.h`) — measured; a by-ear pass on a real playthrough still owed |
| D203 | **Steam Deck (SteamOS, x86_64 Linux) v0.1.0 release bundle: Facility (`-level_34`) crashes "after loading in as James Bond" (user bug report).** — full `## D203` entry at file tail | CLOSED — root cause + fix in D253. |
| D201 | **`src/libultra/audio/bnkf.c` bank/instrument/sound/wavetable relocation offsets were `s32`, truncating the real 64-bit base pointer.** — full `## D201` entry at file tail | FIXED (`src/libultra/audio/bnkf.c`) — part of the Phase-3 mixer landing, not yet build/runtime-verified in isolation |
| D202 | **Silenced PPK plays a "slap" instead of a gunshot, PLUS user reports broader real-time audio corruption (M-52 → M-59 → REOPENED M-60 → M-61/M-62 closed the data chain, recommended close → REOPENED M-63 on a fuller symptom report).** M-56–M-62: exhaustively verified the index/bank/pointer/decode/ROM-offset/source-data chain — soundIndex 46 is correctly requested and the data is ground truth (byte-match checksum corroboration, M-62). **M-63: user's actual complaint is broader — wrong sound always plays, correct one SOMETIMES plays too (both heard together), and sounds "pile up and spirally glitch out" over a session until only a stuck looping sound (e.g. Bunker's door) remains.** Got a live build+run repro this session (`-level_09`, 60s, `GE_MIXERTRACE`+`GE_AUDIOTRACE`+`GE_AUDIODUMP`, 58 scripted PPK fires). **Ruled out voice-pool exhaustion/leak** — only 16 distinct voice slots used, properly recycled (`sndDeactivate` count tracks `sndPlaySfx` count 1:1). **Found an unexplained anomaly**: `soundIndex=109` (AK47 bolt-action SFX) fires 94 times in 60s despite no AK47 in the repro, some retriggering the same voice slot at a suspiciously uniform ~32.6ms cadence (4-in-a-row) — far faster than plausible weapon fire; looks like something re-issuing `sndPlaySfx` every audio tick instead of once per game event. Not yet confirmed as the cause of the user's symptom, but the top concrete lead. Traces preserved: `scratchpad/d202-m63/`. (Side note, not a real bug: chased a `load_bg_file` crash this session that turned out to be an incomplete test-data setup on the investigator's part — see M-63 write-up.) **M-65 PARTIALLY resolved:** the "eventual silence" cascade is root-caused and mitigated (ownerless `SOUND_FLAG_LOOPED` SFX — sound 203, `METAL_SLIDE_CLOSE_SFX`, `decayTime=-1`, played by `doorPlayCloseSound0/1` with `NULL` owner — leaks 7 of 8 voices; a 4-line guard in the preemption scan reclaims only provably-ownerless loops, measured 171→629 acquisitions), and a second independent cause is fixed (`sndCreatePostEvent` was stubbed out per D138, removing ALL distance attenuation). The audible stuck door loop is still not reproduced — every candidate mechanism ruled out by measurement (see M-65); needs a user capture at the failing door with `GE_AUDIOTRACE=1 GE_AUDIODUMP=1`. | ROOT CAUSE ESTABLISHED (M-66); DISPOSITION C IMPLEMENTED + MEASURED (M-66b) — the audible stuck loop is sound 203 (`METAL_SLIDE_CLOSE_SFX`): ROM infinite ADPCM loop (`count=-1`) + `decayTime=-1`, played fire-and-forget by `doorPlayCloseSound0/1` with NULL owner, never stored in any slot, exempt from preemption (flag 0x12) and from decay-stop scheduling, priority 0x41 makes it unstealable — **faithful N64 behaviour / original quirk, not a port bug**. 327 ms period = -550-cent pitch (keyBase 54 + detune 50 − shift 6000 → ratio 0.7278; 5248 samples @ 22050×0.7278 = 327.007 ms). M-65's guard fires only under pool pressure (twice in the user capture) and cannot silence the idle loop. **Disposition C implemented + measured (M-66b):** PC-only `AL_SNDP_PORT_EXPIRE_EVT` fades provably-ownerless infinite-loop SFX out ~2.5 s after start (2 s delay + 0.5 s fade); per-second spectral check: pre-fix drone flat at max t≈39→361 s, post-fix transient bursts with full silence between; run ends allocated=0/8. **M-67: static analysis exhausted — full-bank data scan clean (all 261), DSP audit complete vs PD, A_LOOP branch dead code in GE, reference clip B00I00S2D.wav REFUTED as an in-game capture (best waveform CC 0.19 ≈ noise; M-61's +0.999 was envelope-only correlation); new `GE_VOICEDUMP` per-voice probe + timestamped WIRE/AUDIOTRACE for the decisive runtime A/B; self-test found the offline resample direction had been inverted all along (y[k]=x[ratio·k]) and that GE "music" is an SFX-sequence system with its own ~8-sample set sharing the 8-voice pool; parked anomaly: sound 232 requested at level start produces no audible audio. **M-68 (user captures ×3): all three complaint classes verified per-voice at runtime — doors CC 0.91–0.99, PPK 46 CC 0.90–0.97, armor 81 full 833 ms vs ROM-predicted 837 ms @ ratio 0.5612; the earlier "instant death" metric was a line-count artifact (µs-timestamped `[EVT]` probe shows every STOP lands on schedule). No port bug: the heard "wrong" sounds are the original design (overlapping SFX, short decays, quick-stop truncation).** **M-69 (full-corpus exact match, 320 scripted-run requests): 286/320 CC ≥ 0.85 — every allocated voice plays the exact ROM sample at the exact ROM pitch; the initial 44 "BAD" were tooling artifacts (voicedump stamps are block-END times; coarse-slide missed sharp onsets; refs not clipped at voice death). Remaining failures = dropped allocations (no `[VOICE+]` — suspected 8-voice pool exhaustion, unverified) + shared-wavetable wire misattribution. No corruption.** **M-70 (decisive clean re-capture, 209 requests, matcher v3): every voiced request plays the exact ROM sample — 200/209 auto CC ≥ 0.85; the 4 MID are idx=109 verified at CC 0.946 by manual ratio sweep (residual = wall-clock-bursty dump stamps + short live windows); the 5 NO-WIRE are 1 matcher window artifact + 4 never-voiced requests, NOT pool exhaustion (alloc ≤ 5/8 at each drop). Zero wrong-sample playback — the D202 wrong-sample/mixer-corruption hypothesis is refuted.** **M-71: user by-ear pass on M-66b = PASS ("the metal door sound does not loop forever now") — Disposition C validated; D202 ready for probe removal + close. The user's remaining audio complaints (slap/glass over gunfire, explosion→scream, armor pickup wrong sample — N64 A/B confirms non-fidelity) are a DIFFERENT bug: PC requests different/extra sound indices than N64; bank converter + playback chain exonerated → split out to D205.** **M-76/M-77c: D202 CLOSED.** The silenced-PPK "slap" was D206 (bank index +1) — `ALInstrumentAlt_s.soundArray` sits at struct offset 12 on N64 but 16 on PC (pointer width), so `sndPlaySfx` resolved `GUN_SILPPK_A`=46 to on-disk slot 46 (`PUNCH1`, a slap) instead of slot 45; M-61's rejected "index 45, +0.999" was right. Fix `snd.c:1001` `#ifdef PORT` `-1`, committed `378386d1`; **user by-ear A/B vs N64 = PASS (M-77c)**. Mixer/pool/decode were correctly exonerated by M-56–M-70. The M-65/M-66b ownerless-infinite-loop mitigations stay (independent, validated M-71). |
| D206 | **PC played every SFX one bank slot too high — a pointer-width layout shift in `struct ALInstrumentAlt_s`, not a game-logic bug (M-76 FIXED).** Root cause of D205 (3) armour + (4) melee→Klobb and of D202's silenced-PPK "slap". `sndPlaySfx` resolves a sound via `soundBank->instArray[0]->soundArray[soundIndex]` (`snd.c:1001`), casting the on-disk standard `ALInstrument` (12-byte u8 header, `s16 bendRange`@12, `s16 soundCount`@14, `ALSound* soundArray[]`@16) through GE's `struct ALInstrumentAlt_s { s32 unk0, unk4, unk8; ALSound *soundArray[1]; }` (`src/snd.h:165`). On N64 (4-byte pointers) that struct's `soundArray` sits at **offset 12** — it deliberately aliases the on-disk `bendRange`/`soundCount` words, so `soundArray[N]` == on-disk table entry **[N-1]**, i.e. GE's `SFX_ID` values are **1-based** into the sound table (`SFX_ID 0 = NOTHING_SFX`, short-circuited at `snd.c:992`, never dereferenced). On PC the 8-byte pointer + 8-byte alignment pushes `soundArray` to **offset 16**, and the converted bank (`port/src/romdata.c` `afFixupInst`, which faithfully reproduces the standard layout) is packed to match — so `soundArray[N]` landed on entry **[N]**, one slot high, on every SFX. Verified by static struct/converter arithmetic **and** at runtime: a BUNKER1 `GE_AUDIOTRACE` capture post-fix shows all ~90 distinct requested indices resolve to `rom_sfx_decode.py` slot **N-1** (0/90 at N). e.g. `ARMOUR_COLLECT_SFX`=81 → slot 80 (= rip `S50`, the user's "correct" armour); `skorpion_stats.Sound`=0x6A=106 → slot 105 (= rip `S69`, "the actual Klobb sound"); `wppksil_stats.Sound`=0x2E=46 → slot 45 (= rip `S2D`, M-61's rejected "+0.999" clip). The earlier "262 enum vs 261 bank ⇒ needs a -1" arithmetic was right about the *symptom*; the *mechanism* is the alias, and `BIG_CLANK_SFX`=261 → slot 260 is in-bounds (never an OOB). Upstream `n64decomp/007` `snd.c` byte-matches with no `-1` and no `NON_MATCHING` guard — correct, because the N64 struct offset supplies the -1. **NOT D205 (2)** explosion→scream (169-183 → slots 168-182, all still explosions) — stays with D205. | **FIXED (M-76, `src/snd.c:1001`, `#ifdef PORT` → `soundArray[soundIndex - 1]`, N64 line verbatim under `#else`).** ABI/layout-only, D3x class (pointer-width struct-layout reconciliation); no game-logic change. Build clean; BUNKER1 60s crash-free; `-level_09` golden framediff 3/3. **User by-ear A/B vs N64 = PASS (M-77c)** — armour / melee / silenced PP7 all correct now → D202 closed, D205 down to symptom (2). Committed `378386d1`. Scratchpad `rom_sfx_decode.py` / `exact_match.py` / `diff_bank.py` still walk slots 0-based — correct for "what is at bank slot i", but a check against a `SFX_ID` must compare to slot `id-1`. | 
| D207 | **Alarm klaxon starves combat SFX and (per user) doesn't recover after it stops — surfaced by D206 (M-77, ROOT-CAUSED, fix designed not applied).** Before D206, `ALARM3_SFX` (id 163) resolved to the wrong bank slot 163 — a **one-shot** sample: played once, freed its voice. After D206 it correctly resolves to slot 162 = the **real infinite-loop klaxon** (`decayTime=-1` -> `SOUND_FLAG_LOOPED`; ADPCM `loop.count=-1`). Played by `handle_alarm_gas_timer_calldamage()` (`src/game/propobj.c:14457`) with owner `&ptr_alarm_sfx` while `alarmIsActive()`. It now **permanently holds 1 of only 8 SFX voices** (`MUSIC_SFX_SEQ_MAYBE_MAX_SOUNDS=8`, `src/music.c`), at priority `0x41` (looped sounds get `decayTimeFlag + 0x40`, `sndSetupSound` `snd.c:820`) — and is **exempt from the port's preemption scan** (`snd.c` ~369: `!(unk3e & 0x12)` refuses to steal any LOOPED/RETRIGGER voice; the M-65 carve-out only covers *ownerless* loops, and the alarm is owned). **M-77b:** the voice-starvation hypothesis did NOT hold up — a BUNKER1 `GE_FORCEALARM` + scripted-fire soak showed **0 true drops** (`sndDisposeSound` no-voice sites instrumented), the designed last-resort preemption pass **never fired**, and `g_sndAllocatedVoicesCount` recovered to 0 after every alarm-off (no leak). The earlier "~20% drop" was a log-latency artifact. The alarm klaxon (`ALARM3` slot 162) is a *continuous, gapless* infinite loop — a real GE alarm pulses — so the leading remaining explanation is **perceptual** (the drone masking gunfire, and/or slot-162 mixed too loud vs N64), or genuine 8/8 starvation only on a heavy level not yet tested. Needs a user repro. | MONITORING (M-77c: user reports the alarm is acceptable in play after D206; no fix shipped) — **the voice-starvation hypothesis did NOT hold up (M-77b).** The designed last-resort LOOPED-preemption pass was built, and instrumented drop counters (`[D207-DROP]` at both `sndDisposeSound` no-voice sites) + a `[D207-YIELD]` trace were added. Result on BUNKER1 (`GE_FORCEALARM` + scripted sustained fire, ~85 s): **zero true drops**, the last-resort pass **never fired** (pass-1 always found a non-looped victim), and `g_sndAllocatedVoicesCount` sat at 1 during quiet alarm and recovered to **0** after every alarm-off — no leak, no permanent starvation. The earlier "~20% dropped" figure was a measurement artifact (log-line latency, not real drops; every requested sound eventually got a voice via the `unk38` retries). So the fix was **reverted** (kept: nothing in `snd.c`; D206 `-1` only). BUNKER1 combat is too light to peg the 8-voice pool. **Next session needs a real repro from the user:** which level + weapon, and a `GE_AUDIOTRACE=1` capture at the failing alarm — is it (a) genuine 8/8 starvation on a heavy level (Facility/Silo/Statue firefight), (b) a leak that only triggers on a specific alarm-stop path, or (c) perceptual — the continuous klaxon drone masking gunfire / the alarm mixed too loud (check `ALARM3` slot 162 volume vs N64). Temp `GE_FORCEALARM` probe kept in `src/game/propobj.c` for that work. **M-78: re-added just the `[D207-DROP]` counter (NOT the reverted last-resort pass) at both `sndDisposeSound` no-voice sites in `snd.c` (gated on `GE_AUDIOTRACE`), logging `state`/`prio`/`flags`/`count=N/max`, so the user's real-level capture is decisive between (a)/(b)/(c). Static re-check of `ALARM3` slot 162 this session: `sampleVolume=80` (vs gunfire 90-110, punches 100-110 — the klaxon is quieter per-sample, not louder, weakening the "mixed too loud" half of (c)); `decayTime=-1` + `keyMax=8` (`&0xf0=0` → NOT `SOUND_FLAG_RETRIGGER`) → it is a continuous looped sound by design on N64 too, the pulse is baked into the 225→7512-sample ADPCM loop body. So (c) narrows to "the continuous drone masks gunfire" or a loop-point playback defect; (a)/(b) still need the heavy-level capture. Launcher `build-pc/d207_run.ps1` added (forced-alarm auto-repro on level_09).** |
| D208 | **Player weapon-fire sound goes permanently silent mid-firefight (all other SFX unaffected) — `struct` field `bondview.h:200 s32 field_A48` is used as a second…** — full `## D208` entry at file tail | FIXED + user playtest-verified (M-78d) — Bunker firefight: player gunshots persist through the whole engagement and survive the klaxon stopping. Committed 7601b9db (PR #29, merged). |
| D209 | **Every AI character in the game is locked to a *walk* animation regardless of the speed the ailist commands — `act_ubytes.padding[45]`, a raw-byte alias into…** — full `## D209` entry at file tail | **FIXED + user-playtest-verified (M-81).** `chraction.c:3665`, `#ifdef PORT` → `get_sound_at_range(self, self->act_gopos.unk59, ...)`; the N64 line kept verbatim under `#else`.… |
| D210 | **`chrToPatrol()` leaves `act_patrol.lastvisible60` uninitialised on PC — a raw-byte union alias (`act_init.padding[0x13]`) that shifts when `act_patrol` widens…** — full `## D210` entry at file tail | **FIXED (`chraction.c:3872`, `#ifdef PORT` → `self->act_patrol.lastvisible60 = -1;`, N64 line under `#else`).… |
| D205 | **PC plays wrong/extra SFX for gameplay events the user can A/B against N64: general gunfire has a slap/glass layer audible on PC but silent on N64; an explosion plays a soldier-scream sample; armor pickup plays a wrong sample; unarmed melee (slap) plays Klobb's shooting sound idx 106 instead of the slap effect (M-71/M-72, user by-ear + N64 A/B).** **M-74 WITHDRAWS M-73's root cause.** M-73 read the user's combat capture as guards stuck re-triggering fire sound idx 109 ("256x/29.7 s, no bullets") and pinned it on `stanTestLineUnobstructed` LOS over converted collision geometry. Re-counting the same file: idx 109 is requested **128x** not 256x (the double-log artifact D204 already corrected for M-63 on this very index), = 4.31/s across ~11 guards = ~0.4/s each, **~6x UNDER** the AK47's own `SoundTriggerRate` (`RATE_AK47`=4 ticks => 15/s ceiling) — so the `field_178` gate never binds; and impacts are plentiful in the same window (~30 ricochet/wall-hit 19-41, 69 x2 flesh, 12 body-falls 123-132, 14 yelps 134-147). The eye-catching monotone sweeps 134->147 / 123->132 are the ground-truth round-robin `male_guard_yelp_counter` (`chraction.c:2454-2470`), faithful, not a broken selector. **The 109 traffic is normal in-spec guard combat**; do NOT open the pccg-stan/LOS geometry investigation on this evidence. Everything upstream of presentation is proven: bank 261/261 identical (M-71), sample+pitch exact (M-68/M-70), and every index in the capture correct (182 explosion x1, 81 armor x1, 46/47/48/49/105). **Remaining untested layer = spatial presentation (per-voice volume, pan, concurrency)** — it explains all four complaints at once (a wrongly-placed sound is by ear a sound at your own position) and is the youngest code in the stack (`sndCreatePostEvent` was fully stubbed by D138 until M-65). **M-78: CLOSED.** All four symptoms were D206 (the +1 bank-slot alias). (3) armour + (4) melee->Klobb + (1) slap/glass over gunfire: user by-ear A/B vs N64 = PASS (M-77c). (2) explosion->scream was thought NOT explained by D206 (169-183 -> 168-182, all still explosions) but the user re-tested on Bunker post-merge and the explosion SFX is now correct by ear — the earlier "scream" was a coincident guard death-yelp (GET_HIT_MALE 134-147 seen +/-1s of the blast in M-73's capture), faithful N64 behaviour, not a wrong sample. No presentation-layer bug: pan audit (M-74) already showed GE never spatially pans SFX (252/261 bank sounds `samplePan==64`, no `AL_SNDP_PAN_EVT` poster). | CLOSED (M-78) — root cause was D206; all four symptoms resolved + user-confirmed by ear vs N64. No separate presentation-layer fix needed. | 
| D152+ | **D152 addendum (M-31):** — full `## D152+` entry at file tail | FIXED (`src/snd.c`, `port/src/libultra.c`, `#ifdef PORT`) — fade-out repro playtest-gated, not headless-verified |
| RC3 · D167 | **Non-power-of-two texture wrap period (`docs/dev/TEXTURE-GLITCH-ANALYSIS.md` §6 RC3 — "textures repeat oddly", residual Depot-ceiling noise after D161).** The N64 RDP masks the texel coordinate of a wrapping render tile at `1<<mask`, and GE sets `mask = texDimensionToMask(dim) = ceil(log2(dim))` (`src/game/tex.c:361`), so a non-PoT tile (Depot's 65×65 / 96×48 / 56×56 room surfaces) repeats at the **next power of two**, not at its image size the way GL `GL_REPEAT` does → the pattern is squashed/stretched and the seam lands in the wrong place. fast3d never stored `masks`/`maskt` at all (`gfx_dp_set_tile` dropped them) and wrapped purely at the uploaded image dimension. **Fix (`port/fast3d/gfx_pc.cpp`, behind the existing `Video.WrapFix` knob, default OFF):** store `masks`/`maskt` on the tile; in the hoisted per-texunit pre-wrap block in `gfx_sp_tri1` (D74 block — already lifted out of the vertex loop, indexed by texunit `t` not vertex `i`), when the tile is WRAP (no CLAMP/MIRROR bit) and `1<<mask != tex_width`, fold the UV at the N64 period `1<<mask` and clamp the `[dim, 1<<mask)` overflow band (which is a TMEM smear on console, no real texels) to the last texel so it reads as an edge streak instead of a bogus early image restart. `GE_WRAPFIX=0/1` env override added (env wins over the ini, matching `Debug.FrameDump`). **Per-level captures (`-level_09`/`-30`/`-34`/`-20`, WrapFix OFF vs ON, `GE_PCDUMP` 6-frame windows):** no crashes, 6/6 frames each; on settled/comparable frames Silo is ~pixel-identical (phash 0–11), Facility 180–260 pixel-identical, Depot shows small localized texel changes on ceiling/wall cells (dmean 5–9, no structural break); the large per-run deltas are all the D117 intro-camera-pan nondeterminism, not the fix. Default kept **OFF** — no regression, but a headless structural diff can't confirm the Depot ceiling actually looks *better*; needs a human eyeball with `Video.WrapFix=1`. Default-off is byte-identical to golden (all new behaviour is inside `if (g_wrap_fix)`). D74's dead in-vertex-loop wrap block was already reworked/hoisted at M-30; this only adds the mask-period trigger. Confidence: **medium** (mechanism correct; overflow-band handling is an approximation, not exact TMEM-smear emulation; visual win unconfirmed). porting-notes.md §D. | KNOB ADDED, default OFF (`port/fast3d/gfx_pc.cpp`). Needs user visual check on Depot. |
| D168 | **`GE_PCDUMP` / F12 PPM captures were vertically flipped — the entire source of the bogus D114/D116 "HUD/text X-mirror" (M-33, developer-confirmed).** — full `## D168` entry at file tail | FIXED (`port/fast3d/gfx_opengl.cpp`); D114 / D116 reclassified as capture-orientation artifacts (below). |
| D169 | **Front-end mouse pointer can't reach the outer cells of the mission/level-select grid — only an inner ~3×3 selectable (M-33, developer bug report).** — full `## D169` entry at file tail | **FIXED (M-33, `port/src/input.c`, port-only, no `#ifdef PORT`).** The `menuMode` pointer branch now derives its clamp bounds from the live virtual screen —… |
| D178 | **Pre-mission briefing screen: objectives blank / missing, briefing pages blank (also the D143 "briefing text blank" side effect) — FIXED (M-36).** — full `## D178` entry at file tail | **FIXED (M-36)** — high confidence |
| D175 | **In-game stutter / brief hang during normal play** — full `## D175` entry at file tail | OPEN — observed, not investigated |
| D176 | **Surface exterior renders wrong (`-level_36`), two independent defects.** **(a)** sky solid black — env data is correct (`Clouds=1`, warm `CloudRGB`), the cloud-sky path runs, but `skyRenderTri`/`skyRenderFull` emit only `G_RDPHALF_*` immediates which `gfx_pc.cpp:2901` deliberately no-ops → **root-caused M-37** (see "D176(a) — M-37 UPDATE"); fix = decode the RDPHALF sky-tri stream in fast3d. **(b)** cliff/rock walls = grey diagonal static — NOT a texture decode bug (D183 disproved the shear hypothesis, 0/166 loads strided); re-scope from ROM ground truth of the wall texnum (M-37: inconclusive, leaning tiling-density). See `GRAPHICS-BACKLOG.md` D176. | (a) **"scrolls too fast/tiles too visibly" ROOT-CAUSED + FIXED** (see "D176(a) — fidelity pass" below): `skyPortRenderPoly` interpolated tc *linearly* under a `w=1` ortho projection when the source S/T data assumes RDP-style perspective-correct interpolation (per-vertex camera-space `w` differed up to 38x within one primitive); fixed by giving each vertex its real `w` via a custom projection matrix, `#ifdef PORT` in `src/game/sky.c` only. Verified: `bunker1` regression-clean, `-level_22`/`-level_41` (Cradle) before/after captures show the expected uniform-aliasing → perspective-converging-fan change; Cradle sky at default FOV is no longer black. Confidence HIGH on root cause/fix direction, MODERATE on full N64 parity (no N64 reference capture available to diff against). (b) open. Defect 2 (horizon-band coverage/adaptive tessellation) and the separate FOV-coupled black-sky-past-frustum-edge symptom (D218/D222-adjacent) are **unchanged, still open** — not attempted this pass. **M-93 human playtest of PR #43 (debug build):** confirms Cradle default-FOV black sky is FIXED, and Statue's scaling is measurably better but still reads too fast. Confirms the already-flagged FOV-coupled black-sky symptom too — Cradle's black skybox reappears at max FOV (D218/D222 family, unchanged). **New, previously undocumented residual found this pass, split out as D227:** the sky visibly renders as two independently-moving halves ("two sky processes") on both Statue and Cradle — see D227. **M-98: Defect 2 (horizon-band tessellation) visually confirmed on Cradle with a concrete screenshot** — checking Cradle for D227 regressions (it wasn't) turned up a much more visible, distinct artifact: looking up the suspension-tower catwalk near the level's start, roughly the left half of the sky renders as a sharp "starburst" of radiating light-ray-like streaks converging toward a point near the top of the screen, stitched right up against normal-looking horizontal cloud streaks on the right half — `docs/img/bugs/d176a-cradle-tower-starburst.png` (`-level_41` frame 350). **Confirmed pre-existing, not a regression from M-98's fold fix or M-95's wScale fix** — identical pixel-for-pixel with `src/game/sky.c` reverted to its pre-M-98 state, same camera/frame. **Confirmed static, not the D227 "moves at a different rate" symptom** — pixel-identical across frames 350 and 380 (30 ticks apart) at this static camera. This matches the already-known, already-open Defect 2 exactly: the code comment at `skyPortRenderPoly`'s fold logic already predicts it — "Triangles whose S/T span exceeds ~1000 texels (only those straddling the horizon) still need Defect-2's finer tessellation." Looking nearly straight up the tower is exactly the camera angle that produces one sky triangle spanning from near-camera to horizon-grazing texel coordinates, i.e. the >1000-texel-span case the fold logic already flags as unhandled. Not fixed this pass (tessellation work, separate from the D227 fold/wScale-sharing fix) — logged here so whoever picks up Defect 2 has a concrete repro + screenshot instead of the prior "inconclusive, leaning tiling-density" note. |
| D177 | **Ladders non-functional — progression blocker (user QA report) — FIXED (M-36).** Not the input path and not the ladder state machine: `MoveBond` gates the ladder-collision path on `stanGetLocusCount(&curLocus)`, which was pinned at 0 on PC. `stanCheckLinkedSpecialTile` writes the LADDER signal via raw `outFlags[1] = 1` into a `struct StandTileLocusCallbackRecord` whose first member `s32 *rooms` is pointer-widened on PC → `[1]` is the high half of `rooms`, and `count` (moved +4→+8) is never written. Compounded by `curLocus` being declared as the 8-byte placeholder `move_bond_temp_struct` (too small for the widened record) and a `(s32)coords` pointer-truncation AV waiting in `stanGetMoveBondCollisionTiles`. **Fix:** `#ifdef PORT` — write record fields by name, declare `curLocus` as the real struct, `PORT_PTRADD` for the truncating cast (`src/game/stan.c`, `src/game/bondview2.c`; no game logic). Full detail: §F "D177". | **FIXED (M-36)** — high confidence; interactive climb test owed |
| D172 | **Bullet-impact / blood / spark particles render magenta or cyan instead of dark red (M-82, ROOT CAUSE FOUND + FIXED).** — full `## D172` entry at file tail | **FIXED (`port/fast3d/gfx_pc.cpp` `gfx_lod_tile_offset`)… |
| D174 | **"No blood effect" (user QA report).** — full `## D174` entry at file tail | OPEN — observed; first step = BUNKER1 firefight with regenerated sidecars |
| D183 | **M-36 Family A ("texture line/pitch shear") is DISPROVEN for the Surface repro (`-level_36`).** The importers' pitch assumption is real but never fires there; the Surface cliff "grey static" is not a shear. Full evidence + what the walls actually are: §F "D183" below. A defensive, provably-no-op de-stride landed in `import_texture()` anyway (covers the `gfx_dp_load_tile` case the `SUPPORT_CHECK`s assert against), plus `GE_DTEX` STRIDED marker + `GE_TEXRAW` raw-source dump. | Hypothesis DISPROVEN + diagnostics shipped; D176(b)/D182(2) still OPEN |
| D173 | **Third-person Bond model spawns too high** — full `## D173` entry at file tail | OPEN — observed, not investigated.… |
| D179 | **Packaged build crashes after the logos — the D43/D69 model + bg sidecars are ROM-derived and absent from any build that has no ROM (M-34, alpha-release QA).** — full `## D179` entry at file tail | Diagnosed, not fixed — release packaging gap + latent `romdata.c` fallback |
| D180 | **Native-PC input pass (M-XX QoL run, `qol/native-pc-input-menu`, port-only).** — full `## D180` entry at file tail | LANDED, port-only, no `#ifdef PORT`.… |
| D181 | **`Game.ScreenShakeIntensity` — first route-(b) `src/` gameplay-cosmetic hook (M-XX QoL run).** — full `## D181` entry at file tail | **Menu row pulled (v0.2.1, user testing).** The hook is correct on its own narrow terms —… |
| D184 | **F10 port-layer options overlay — the approach-(C) surface from `OPTIONS-MENU-PLAN.md` (M-37).** — full `## D184` entry at file tail | **LANDED, port-only, no `#ifdef PORT` in `src/`.** Build links 241/241.… |
| D211 | **PC FOV slider — `Video.FovScale` (M-83 QoL wave, port-only, `WIDESCREEN-FOV-PLAN.md` Phase 4).** fast3d applies a multiplicative half-angle scale to every *perspective* projection matrix as it is loaded (`gfx_apply_fov_scale()` in `gfx_pc.cpp gfx_sp_matrix`, on the `G_MTX_PROJECTION | G_MTX_LOAD` path only — GE always reloads its view projection via `guPerspectiveF` at `fr.c:725`, so a relative `G_MTX_MUL` onto an already-scaled matrix cannot compound). Columns 0 and 1 are scaled by the same factor `tan(half_orig)/tan(half_adj)` → H and V FOV widen uniformly, aspect ratio exactly preserved (no stretch), and because the game rewrites fovy every frame for aim-zoom the scale composes with zoom for free. Ortho matrices (`P[3][3]==1`) are left untouched so HUD/menu 2D is unaffected. `Video.FovScale` = percent of the original vertical FOV, 50–150, **default 100 ⇒ `g_fov_scale == 1.0f` ⇒ guarded early-return, zero FP ops on the matrix ⇒ golden dumps byte-identical**. F10 overlay row "FOV scale %". Math checked offline (`s=1.25 → fovy 60°→75°`, aspect held at 1.3333). | **PARTIAL — M-84 refix: game-side FOV scale (`src/fr.c`, `#ifdef PORT`), front-end gated, user-verified.** Original M-83 note follows. LANDED, port-only; build clean; `bunker1` framediff worst_cell 12.219 vs 12.221 baseline (= render noise, no-op confirmed). **M-84 refix — the fast3d hook only scaled *pure* perspective loads; GE draws the world through `g_CurrentPlayer->field_10E0`, a projection × view matrix pre-combined on the CPU (`bondview2.c:8224`) that failed the `gfx_apply_fov_scale` guard, so only the pause/watch model + sky FOV changed (user report).** Moved the scale game-side to the `guPerspectiveF` chokepoint (`fr.c:718`, `#ifdef PORT`, `frFovY *= portFovScale`, clamp 160°), which lands *before* the CPU combine; fast3d `gfx_apply_fov_scale`/`gfx_set_fov_scale` removed; `port/src/video.c` sets `f32 portFovScale`. Also gated to `lvlGetCurrentStageToLoad() != LEVELID_TITLE` — the front end (main menu / file+mission select / briefing) renders through the same path (`lv.c lvlRender`) and its 3D is authored at a fixed FOV. **User-verified: front end unaffected + in-level view widens (Dam tunnel at high FovScale).** M-85 headless nit check (`-level_09`, FovScale 100 vs 120): **HUD-static confirmed** — the ammo counter ("7 | 93") sits at the identical bottom-right screen position in both, and the door-arrow stays square (no 4:3 stretch); viewmodel scales down + stays bottom-right anchored. **aim-zoom-compose — M-86 static analysis: correct by construction.** `bondview2.c:3123` lerps `zoominfovy` then `viSetFovY()` stores it in `g_ViBackData->fovy`; `fr.c:737` applies `frFovY = fovy * portFovScale` **once**, at the per-frame chokepoint, *after* the zoom lerp. Hipfire and zoomed states scale by the same factor so the zoom magnification ratio is preserved exactly; no double-application; the smooth zoom interpolation is unaffected (scale is post-lerp). Residual human eyeball (low risk, cosmetic) at FovScale ~120–150 + full aim-zoom: the viewmodel (`c_perspfovy`, `bondview.c:644`) and frustum-cull (`currentPlayerSetCameraScale`) use the *unscaled* fovy, so a subtle viewmodel-scale or edge-cull-early mismatch is possible; the scope reticle is 2D screen-space, unaffected. Caveat (same root): edge geometry culls a hair early at large scale (cosmetic). **User-verified good at big FovScale.** New adjacent report → **D220**: 1P muzzle flash position shifts upward at higher FovScale (Jungle, M16) — likely the same unscaled-viewmodel-fovy root as the caveat above. Wide FOV also exposes the per-level draw-distance boundary → **D218** (logged, parked). **M-87: the frustum-cull residual is a real bug, not cosmetic** → split to **D222**. |
| D212 | **Anisotropic-filtering config wiring — `Video.Anisotropy` (M-83 QoL wave, port-only).** — full `## D212` entry at file tail | **LANDED, port-only.** Build clean. **Owed:** eyeball at 16× on a grazing surface (Depot roof / Dam walkway) vs 1×. |
| D213 | **On-screen FPS readout — `Video.DisplayFPS` (M-83 QoL wave, port-only, PD parity).** — full `## D213` entry at file tail | **LANDED, port-only.** Build clean; `bunker1` no-op confirmed. **Owed:** eyeball that the counter reads a sane ~30 and doesn't overlap the HUD. |
| D214 | **Keyboard rebinding — `[Input.Bind]` config section (M-83, port-only, PD parity, Phase 4 debt).** — full `## D214` entry at file tail | **PARTIAL — M-84: code PASS (alt/ctrl); one doc-only fix.** LANDED, port-only, no `#ifdef PORT`; build clean.… |
| D215 | **1P weapon viewmodel never drew in-level — `gunRenderFirstPersonGunModels` builds its `ModelRenderData` from a reinterpret across separately-declared adjacent…** — full `## D215` entry at file tail | **FIXED (`src/game/gunfire.c`, `#ifdef PORT`).** Build clean.… |
| D216 | **`Game.SkipIntro` — boot straight to the file-select menu (M-83, port + one `#ifdef PORT` game-code line, PD parity, Phase 4 quick win).** — full `## D216` entry at file tail | **LANDED + VERIFIED this session.** Build clean.… |
| D217 | **1P weapon viewmodel + third-person Bond model textures render garbled — wrong/stale palette on model-DL CI textures (M-84, user QA; surfaced by D215 now that the viewmodel draws).** BUNKER1/Facility: the PP7 viewmodel has yellow+blue scrambled texels and a solid **green** patch by the hammer; Bond's in-game **watch face** shows the same green mismatch; user reports the **third-person Bond world model** does this "sometimes" too. Signature matches the **D161 family** — a CI-format tile drawn without the expected TLUT state, so fast3d's `import_texture` indexes a **stale `rdp.palette`**. D161 fixed the explicit `G_TT_NONE` case; this looks like model DLs not emitting `gDPLoadTLUT` through the port path, or the palette segment not resolving for weapon/character geometry. **Not a D215 regression** — D215 only un-gated the DL. | **FIXED (M-90/M-92/M-97; PR #47 texpool alignment + PR #49 CI-cache palette key) — the green-texture defect is fixed and merged, confirmed by two independent human playtests. The white missing-texture residual (AK47 / NPC-held world-model weapons) is split out as D228.** Status cell had gone stale at OPEN; corrected M-105. History: root cause characterised M-86; headless repro: `-level_34` FACILITY ~frame 300–400, PP7 **grip = flat green block** + green ring at the barrel joint + green-cast hand; barrel/slide upper correct black. Also green on `-level_30` DEPOT. **Clean on `-level_09` BUNKER1** (A/B control — same PP7, identical DL). **Two separate defects found:** ① **Wrong palette bound (the visible green — upstream of fast3d).** The grip is a CI4 solid-fill of palette index 0. Green `(24,65,0)` == RGBA5551 `0x1A00` == `rdp.palette[0]` at grip-import time. The grip's `count=1` model `G_LOADTLUT` is fed **different source bytes per level** for the same tile: BUNKER `00 01`→`PD_BE16`→`0x0001` (black ✓), FACILITY `1a 00`→`0x1A00` (green ✗). fast3d byteswaps + stores faithfully (proven correct by BUNKER) — the wrong bytes *arrive* from the model-DL / segment-address path. Strong tell: every *larger* model TLUT has `raw[0]==raw[1]==0x0000` with real colour from index 2 ⇒ GE's model-TLUT payload likely carries a **4-byte (2-entry) leading zero pad**, OR the port's model-TLUT source pointer is **4 bytes / 2 entries too low** (then the `count=1` grip load reads the tail of the preceding CI blob — BUNKER lucky, FACILITY not; fits "widespread across levels + weapons"). Ruled out this session: not tail-overflow (index 0 is in-range; matches M-85 `GE_D217Z`), not the fast3d byteswap. **Next:** dump the grip TLUT `G_SETTIMG` target vs the ROM/`pcmodels` sidecar PP7-grip material bytes; confirm the 4-byte offset vs N64; fix in model-DL segment resolution or `tools_pc/d43_emit.py` blob placement (sidecar regen). ② **CI texture-cache collision (proven, real, separate; fix landed).** `gfx_texture_cache_lookup` (`gfx_pc.cpp:522`) keyed CI textures on `{addr, palette_addrs[0/1], palidx, size_bytes}` with **no palette-content check**; GE reissues `gDPLoadTLUT` from repeated scratch-arena addresses with different content, so a later material could take a stale HIT and sample another material's decoded palette. **Fix on branch `fix/d217-model-texture-palette` (commit `f6fc0399`):** FNV-1a hash of the 512-byte `rdp.palette` computed only in `gfx_dp_load_tlut()`, added to the CI cache key (`port/fast3d/gfx_pc.{cpp,h}`, fast3d-internal, no `#ifdef`). Cheap, cannot worsen a decode; **owes a full 21-level `verify.sh sweep`** before merge (touches every palettized world texture's import path — cache-churn risk only). Does **not** fix defect ①. Screenshots: session scratchpad `d217/`, `fac_gun_crop.png`. Cross-ref `docs/dev/GRAPHICS-BACKLOG.md`. **User playtest matrix (M-87, raw notes — not yet triaged):** DAM/PPK: watch still green; AK47 viewmodel has white (missing-texture) patches on the magazine + barrel. FACILITY: PPK garbled **green AND blue**; 3rd-person Bond world model shows greenness on level load-in; AK47 has "weird blue / glitched" textures. JUNGLE: M16 viewmodel is orange in spots. Confirms multi-weapon (PP7/AK47/M16), multi-color (green/blue/white/orange) — consistent with the upstream model-TLUT-source theory (different trailing garbage per weapon/material) rather than one single fixed offset. **Scope clarified (M-87):** not just Bond's 1P viewmodel / 3P world model — **NPC-held weapon models (other characters' guns in third person) show the same glitched-color/white-missing texture symptom.** So the defect is in the shared weapon-model-DL/TLUT path used by *any* character holding that weapon, not something specific to Bond's model or the 1P viewmodel rig — widens the likely fix surface (weapon model loading in general) beyond "Bond's gun." **M-87 clarification:** not uniform across levels — severity and glitch *type* both vary per level (some levels show it lightly, some not at all in a given session, and the specific wrong-color/missing pattern differs), consistent with the per-material trailing-garbage theory rather than one fixed global offset. **M-88 update: defect ① is NOT a `tools_pc/d43_*.py` sidecar bug — the theory in this row is refuted; the model converter never touches texture/palette bytes at all.** Root cause narrowed to the runtime `texWriteLoadToTmemAddr`/`gfx_dp_load_tlut` TLUT-address path (`src/game/tex.c` + `port/fast3d/gfx_pc.cpp`, both out of this session's file scope). See full write-up: §H "D217 — M-88 defect ① root-cause narrowing". **M-89 update: the `gfx_dp_load_tlut` pitch/width=1 hypothesis is REFUTED** (hand-derived: it already computes `pitch=1` for every model-DL TLUT call via the GBI `width-1` packing — a proposed `pitch=1` patch would be a no-op, do not apply it) — **the addressing math is correct by inspection**. New leads, neither yet probed: (a) does `g_Textures[texturenum]`/`_imagesSegmentRomStart` actually resolve to identical ROM bytes for this weapon's grip texture on BUNKER1 vs FACILITY (an image-table/asset-resolution path distinct from the already-ruled-out per-model `TextureID`/`d43_emit.py`/`Globalimagetable`/D68 path), or (b) does the shared `texpool` arena (`image.c` `texLoad`, two-ended `leftpos`/`rightpos` allocator) serve a stale/aliased `tex->data` for what should be a fresh per-level decode. See full write-up: §H "D217 — M-89: the `gfx_dp_load_tlut` width=1/pitch hypothesis is REFUTED". **M-90: defect ① ROOT CAUSE FOUND + FIXED** — neither lead (a)/(b) was it: the texpool arena base for the weapon pools lands **%8==4 on PC** (memp bump drift from pointer-bloat allocation sizes), and `texAlignIndices` pads CI rows to *absolute* 8-byte boundaries, so every palette in a misaligned pool sits **4 bytes early** while the model-DL `G_LOADTLUT` still reads the N64 offset — the grip's count=1 load lands on the next tenant's header slot (`texnum=26` → LE s16 → `0x1A00` = the green), and larger loads read two zero-pad entries then the real palette (this is M-86's unexplained "leading zero pad" tell). **Fix:** `texInitPool` rounds the arena base up to 8 under `#ifdef PORT` (no-op on N64; ≤7-byte pool shrink) — ABI/layout-class edit, no logic touched. Verified: FACILITY 188/188 decodes aligned + draw reads `pal0=0001` (was `1a00`) + 0 green pixels in frames 250–450 (was ~1300); BUNKER1 + DEPOT still clean. **Pending:** full 21-level sweep result + human playtest eyeball of the M-87 weapon matrix. See §H "D217 — M-90". **M-92: sweep 21/21 PASS + human eyeball done — PP7/PPK/watch CONFIRMED FIXED by this alignment fix. AK47 (viewmodel + world model) and NPC-held world-model weapons (white/missing-texture patches, e.g. magazine/receiver) are a DISTINCT, still-OPEN residual — not fixed by this PR, not yet root-caused. PR #47 should still merge (real, confirmed partial fix); reopen defect ① scoped to the AK47/NPC-world-model white-texture failure mode. See §H "D217 — M-92". **M-97: independently re-confirmed by a second human playtest** (this branch merged up to date with `main`, fresh build, `-level_34` FACILITY): PP7/PPK grip+hammer clean, watch face clean, no green anywhere; **NPC-held AK47 still shows white missing-texture patches.** Two independent playtests, same split — the green defect is fixed, the white defect is not. The white residual is now split out under its own label, **D228** (rather than leaving defect ① ambiguously half-open), so this row can close on the green mechanism it actually describes. |
| D218 | **Wide FOV exposes the per-level draw-distance / fog boundary — "blue artifacting" down long peripheral sightlines (M-84, user QA on the D211 FOV slider).** BUNKER→**Dam main tunnel** at `FovScale` ~150: geometry / backdrop past the level's fog-end distance shows at the widened frustum edges (blue = Dam's fog tint). Not a render bug — `fogLoadCurrentEnvironment` (`bgfog.c:305`) does `viSetZRange(Visibility.BlendMultiplier, Visibility.FarFog)`, so **`Visibility.FarFog` is both the far clip plane and the fog-saturation distance**; levels tune it for the stock ~60° FOV. | **Logged (M-84), fix proposed + parked.** Plan: `Video.DrawDistance` (percent, 100–400, default 100 ⇒ no-op) `#ifdef PORT` multiplier on `FarFog` at `bgfog.c:305` (+ the direct `arg0->Visibility.FarFog` reads just below, so fog stays consistent) plus `Video.DrawDistanceAutoFov` (0/1, default 1 — couples draw distance to `FovScale` unless `DrawDistance` is set explicitly). Clamp effective multiplier ≲2.5× (small znear ⇒ far-field z-fighting risk). **M-87: user confirms still occurring** at high FOV, and — reversing the earlier "not implemented by user request" — **now wants `Video.DrawDistance` implemented as a real QoL feature**, plus scope it to cover **LOD/detail-distance in general**, not just the fog/far-clip boundary this row root-caused. Open question for that broader ask: does GE have a *separate* geometry/object LOD-swap or draw-culling distance mechanism (distinct from `Visibility.FarFog`) worth exposing too, or is fog-distance the only meaningful "how far can you see" knob in this engine? Check for a per-object visibility-radius / simple-model-swap system before assuming `FarFog` is the whole story. **FIXED as designed, awaiting visual verification (M-121).** Implemented exactly the parked plan: `f32 portDrawDistanceMultiplier(void)` (`port/src/video.c`) — an explicit `Video.DrawDistance != 100` wins outright, otherwise `Video.DrawDistanceAutoFov` (default on) couples the multiplier 1:1 to `Video.FovScale`, clamped to `[1.0, 2.5]` (never shrinks draw distance; the 2.5x ceiling is the finding's own far-field z-fighting caution). Applied in `fogLoadCurrentEnvironment` (`bgfog.c`) to a new local (`scaledFarFog`) used at *both* the `viSetZRange` far-clip call and the `g_CurFogDetails.scaled_far_fog_dist` fog-blend read just below it, so the clip plane and the fog math stay consistent per the original plan's own note — the level's `EnvironmentRecord` itself is never mutated (it's shared/reloaded, not a per-frame copy). Both new config keys added to the F10 overlay (`Video.DrawDistance` slider, `Video.DrawDistanceAutoFov` toggle) alongside the existing `FovScale` row. Identity at both defaults (`DrawDistance=100`, `FovScale=100` ⇒ `portFovScale==1.0f`) by construction. **Broader LOD-swap question from M-87 ANSWERED — yes, and now exposed too (same session, M-121 continued).** `src/game/model.c` has a real, distinct geometry-LOD system: `MODELNODE_OPCODE_LOD` nodes whose children are shown/hidden per-frame by `modelUpdateDistanceRelations()` testing a view-space `distance` (scaled by `getPlayer_c_lodscalez()`, itself now FOV-consistent per this session's D222 fix) against each node's `MinDistance`/`MaxDistance` — a genuine detail-swap mechanism, separate from the fog-visibility-fade this row already fixed. It even has an existing ground-truth multiplier hook, `g_ModelDistanceScale`, already used transiently by the game itself (`chr.c`'s 0.3125f close-up override, `front.c`/`title.c`'s menu-model `modelSetDistanceDisabled`). Rather than write to that shared global (would fight those existing overrides), added a second, composed multiplier at the same call site: `f32 portLodDistanceMultiplier(void)` (`port/src/video.c`), driven by new `Video.LodDistance` (percent, 25-400, default 100) and `Video.LodDistanceAutoFov` (**default OFF**, unlike `DrawDistanceAutoFov` — this is meant as a standalone performance lever the user can pull *down*, not something that silently adds cost as FovScale widens) — `distance *= portLodDistanceMultiplier()`, inverse-scaled so `LodDistance=200` keeps full detail twice as far (more cost) and `LodDistance=50` drops to the lower LOD twice as soon (less cost). Both new keys in the F10 overlay alongside `DrawDistance`. Identity at `LodDistance=100`. Rebuilt clean, crash-free at `LodDistance=40` (aggressive) on Dam and at defaults on bunker1. **M-121 live playtest, iterated to a working state.** Two real issues found and fixed while testing live on Dam with the user: (1) **regression** — the D222 fix (above) correctly widened `c_lodscalez` for the screen-edge cull planes, but that same value is *also* read directly by `propobj.c`'s `chrobjFogVisRangeRelated()`/`sub_GAME_7F054C58()` (the character/prop fog-visibility-fade cutoff) — uncompensated, this made guards fade/disappear *sooner* at wide FOV than before D222, the opposite of the intent. Fixed by dividing that same `c_lodscalez` reading by `portDrawDistanceMultiplier()` at both call sites (`#ifdef PORT`), so the "how far can you see" question these two share is answered consistently. (2) **tuning** — a straight 1:1 `Video.FovScale` auto-coupling on `portDrawDistanceMultiplier()` was not aggressive enough; live-tested up twice (1x → 2x → 4x `FovScale`, clamp ceiling raised 2.5x → 4.0x → 6.0x each time) before guards held up and the Dam tunnel's blue-glow artifact receded to the user's satisfaction. Re-verified end-to-end that the fog *ramp* itself (not just the far-clip cutoff) scales with the multiplier — `g_ScaledFarFogIntensity`/`scaled_far_fog_dist` both derive from the same scaled far value — so if a similar artifact resurfaces on another level, the multiplier magnitude or a level-specific tunnel-length edge case are the first things to check, not this math. **User confirms this is good for now.** All three (D222 cull fix, D218 draw-distance + fog-visibility-fade fix, D249 LOD-distance) verified crash-free and landing in the same PR. |
| D219 | **PPK explosions render purple on Dam (M-87, user QA, "not 3D but noting").** Effect-color defect, not a texture-decode one — likely the D172 particle-color family (`gfx_lod_tile_offset` / `G_CC_INTERFERENCE` channel-order class) recurring on a different particle type/weapon than the D172 fix covered, or a fresh instance of the same class. | **RESOLVED (M-115, 2026-09-12)** — byte-order fix in `src/game/image.c` scoped by pixel format (16-bit family bswap16-at-store, 32-bit identity); user playtest-confirmed. Superseded history below, which began: **OPEN — observed, not investigated.** Next: `GE_PCDUMP`/`GE_TEXDUMP` a Dam PPK explosion; check whether it's the same `gfx_lod_tile_offset` tile-sampling defect D172 fixed for blood, on an explosion DL D172's fix didn't touch. **Status: re-checked at runtime M-90 — mechanism covered by the D172 fix on current `main`; close pending a human by-eye pass (folds into PR #47 playtest).** **M-89 triage (local-Qwen, static analysis only, not yet runtime-confirmed):** the explosion DLs (`assets/oddtextures.c` `globalDL_0x078/0x120/0x1c8/0x270…`) use the same `G_CYC_2CYCLE`+`G_CC_INTERFERENCE,G_CC_MODULATEIA2`+`G_TL_TILE`, tile0=IA smoke/tile1=RGBA fire mechanism as the already-fixed D172 blood/spark records — no separate/unfixed code path found in `explosion.c` or the asset DLs. If still purple after D172's fix (which the user's Dam report postdates — needs re-confirming on current `main`), the next suspect is the GLSL combiner implementation of `G_CC_INTERFERENCE`+`G_CC_MODULATEIA2` for cycle 2 specifically (`gfx_generate_cc`, `port/fast3d/gfx_pc.cpp:312-494`), not the tile-offset fix. **Next, concrete:** run Dam (`-level_33`) with `GE_D172=1` while triggering a PPK/grenade explosion and check for "BUG" lines in the log (silence = D172's fix already covers this DL and the row can likely close on a re-test; needs a scripted-fire repro, not yet built). **M-90 runtime re-check DONE (scripted-fire repro built with `GE_INPUTSCRIPT`): the D172 fix demonstrably covers Dam's particle DLs — row can close on a human by-eye pass.** Ran `-level_33` headless with `GE_D172=1`, 64 scripted Z-pulses (`GE_INPUTSCRIPT="200:Z;215:Z;..."`, ~15 controller-reads apart) and a `GE_PCDUMP` window over the same span. Result: **40/40 particle multitex tris log `SAMPLED1=tile1 (tmem=392)`** — the correct RGBA fire tile, i.e. exactly the post-D172-fix behaviour (pre-fix these read `SAMPLED1 = tile0 (tmem=0 fmt=IA)`, the smoke texture); **zero** "BUG"/unknown-GBI/D146-abort lines; **0** purple/magenta pixels (`r>120, b>120, g<90, r-g>50, b-g>50`) across all 28 captured frames. Caveat: the user's M-87 "purple" report may predate the D172 fix on their binary (playtests run on whatever build was current); this re-check proves the sampling path is correct on current `main`, which is as far as a headless pass can take it — a final by-eye look at a live Dam PPK firefight is the honest closer (cheap for the user, folds into the D217 PR #47 playtest). **M-108 — REOPENED, the by-eye closer FAILED.** User played a live Dam (`-level_33`) PPK firefight on the current `main` build (post-D233/D234 fixes, same session as D243) and confirmed explosions are **still blue/purple**, directly contradicting M-90's clean scripted-repro result (40/40 correct-tile samples, 0 purple pixels, 28 frames). This means the M-90 `GE_INPUTSCRIPT` scripted-fire repro was very likely **not exercising the same explosion instance/DL the live firefight hits** — plausible gaps: a different explosion variant (grenade vs. bullet-impact spark vs. enemy-death vs. PPK-specific muzzle/hit effect) using a DL the scripted test's Z-pulses didn't trigger, or a state/lighting/prim-color condition (M-89's alternate suspect: the `G_CC_INTERFERENCE`+`G_CC_MODULATEIA2` cycle-2 GLSL combiner implementation, `gfx_generate_cc`, `port/fast3d/gfx_pc.cpp:312-494`) that only manifests under conditions the synthetic repro didn't reproduce. **Do not trust the M-90 all-clear for this finding going forward** — it proved one particle path clean, not that Dam's live-fire purple is fixed. Next: a live `GE_PCDUMP`/`GE_D172=1` capture during an actual purple explosion in real play (not a scripted approximation) is now the only way forward — need to pin down exactly what triggers it (PPK gunshots hitting a surface/enemy? a thrown/tossed explosive? which). **User has now confirmed (2026-09-12) this is NOT Dam-exclusive — game-wide, not just suspected.** Treat as confirmed scope, not pending a second-level check. **User wants this actively fixed, not just triaged** — next session should prioritize a live `GE_PCDUMP`/`GE_D172=1` capture during a real explosion (any level) over further static analysis, since the M-89/M-90 static passes are exhausted on this finding without producing a fix. **New HIGH-interest lead (M-110): `explosion.c` never clears `G_LIGHTING` before drawing any of the 15 particle DLs** (`globalDL_0x078` through `0x9a8` in `assets/oddtextures.c`, all structurally identical, all `G_CC_INTERFERENCE`+`G_CC_MODULATEIA2`) — grepped every particle-draw call site in `explosion.c` (`gSPClearGeometryMode` calls at lines 883/1495/1718/2227 only ever clear `G_CULL_BOTH`/`G_FOG`/`G_CULL_FRONT`, never `G_LIGHTING`) and confirmed `G_CC_INTERFERENCE,G_CC_MODULATEIA2` is used nowhere else in the codebase, so it's safe to key a probe on it. The port's vertex path (`port/fast3d/gfx_pc.cpp:1301-1346`) branches hard on `rsp.geometry_mode & G_LIGHTING`: if lighting is on when these particles draw, the authored RGBA tint in `v->v.cn[]` gets reinterpreted as a normal vector, and the resulting SHADE color (fed into `G_CC_MODULATEIA2`'s `COMBINED*SHADE`) becomes a scene-light color instead of the DL's actual red/orange fire tint — a mechanism that would plausibly read as "wrong color, state-dependent, game-wide" exactly as reported, and would explain why a synthetic scripted-fire repro (M-90, no nearby lit models) came back clean while live firefights (guards/lit models on screen) don't. **Not yet confirmed** — this depends on whatever last touched `geometry_mode` before the particle draw in a given frame, which the game's own N64-accurate command order controls; it is not yet established whether this is a genuine port bug (fast3d misordering/dropping a `G_GEOMETRYMODE` command relative to real hardware) or something that would reproduce identically on N64 (in which case it isn't actually the D219 mechanism, since N64 doesn't show purple explosions). **Instrumented, not fixed:** added a `#ifdef PORT`/`GE_D172=1`-gated log line to the existing D172 multitex-tri probe (`port/fast3d/gfx_pc.cpp` ~line 1740, `gfx_sp_tri1`) that now also prints `geometry_mode` and flags `LIGHTING=ON <-- D219 SUSPECT` whenever one of these particle triangles draws with lighting on. Sanity-checked against the known-clean M-90 repro (headless `-level_33`, 40 scripted Z-pulses): probe fires exactly 40/40 as before, all show `LIGHTING=off` — consistent with the already-known-clean case, confirming the probe works and doesn't false-positive. **Next, concrete:** run `build-pc/d172_silo.bat` (or equivalent, `GE_D172=1`) and trigger a real explosion (any level, any explosive source) near visible character models; grep the log for `D219 SUSPECT`. A hit confirms the mechanism and turns this into a real fix (clear `G_LIGHTING` in the port's particle-draw path or in fast3d's geometry-mode tracking, whichever turns out to be the actual divergence from N64); a miss across a real purple-explosion capture rules this whole lead out. **M-112 (this session): decode-chain code review + two new inert diagnostic probes; runtime capture NOT obtained (environment limitation, see below). Confidence: LOW-MEDIUM on a narrowed next step, no fix applied.** Read the full N64 texture decompression ground truth for these assets end to end: `texLoadFromDisplayList`→`texLoad`→`texInflateNonZlib`/`texInflateZlib`→(`texReadUncompressed` | `texInflateHuffman`/`texInflateRle`+`texChannelsToPixels` | `texBuildLookup`+`texInflateLookup`/`texInflateLookupFromBuffer`) (`src/game/image.c`), plus `gimgFixupGlobalimagetable`/`gimgSyncCompiledGlobalDLs` (`port/src/gimgfixup.c`) and `import_texture_rgba16` (`port/fast3d/gfx_pc.cpp:653-681`). Findings: (1) `IMAGE_FIRE_0..14` are **not** `sImageTableEntry`-table images (unlike `IMAGE_IMPACTMULTI`/`IMAGE_IMPACTREDBRICK*`/flare/ammo images, which live in the 32 `sImageTableEntry` arrays at segment offset 0xAC8+ and get their `G_SETTIMG` marker synthesized at runtime by `texSelect`, `src/game/othermodemicrocode.c:318`) — FIRE_N's `gsDPSetTextureImage(...IMAGESEG(IMAGE_FIRE_N))` is baked directly into the compiled `globalDL_0x078..0x9a8` display lists (segment offset 0x000-0xA50) alongside the interleaved, correctly-rendering `IMAGE_SMOKE_0..14` (IA16). Both FIRE's and SMOKE's markers go through the *same* `gimgFixupGlobalimagetable` part-1 DL-bswap step and the *same* `texLoadFromDisplayList`→`texLoad` resolution — no separate/divergent table-offset math applies to FIRE_N specifically (this narrows away part of the open brief's option (a): there is no `sImageTableEntry`/`s_tbl_offs[]` bounds bug possible here, because FIRE_N never goes through that array class at all). (2) `texChannelsToPixels`'s `TEXFORMAT_RGBA16` case and `texReadUncompressed`'s `TEXFORMAT_RGBA16` case are decomp ground truth (byte-exact, `#ifdef PORT` free) and are format-generic, not texture-specific — a bug there would require a same-format sibling to also be wrong, and IMPACT/FLARE RGBA16 sprites (same `G_IM_FMT_RGBA,G_IM_SIZ_16b`, same underlying `texLoad`) are the only same-format comparison point on hand, **but this session did not runtime-confirm whether those actually render correct colors in game** — flagging as an unverified assumption inherited from earlier sessions' "shared decoder is proven correct" framing, not re-proven here. (3) `texSwapAltRowBytes` is already an unconditional no-op on `PORT` (D159, `src/game/image.c:2219`, `return;` before any format switch) so it cannot be format- or texture-selective on the port — ruled out as a contributor for FIRE specifically (it's equally a no-op for every other texture). **New, added, NOT yet exercised:** two matching `GE_D219RAW`-gated diagnostic probes (inert unless the env var is set, `#ifdef PORT`, no behavior change): (a) in `texInflateNonZlib` (`src/game/image.c`, right after the per-image header read, and again right after decode) — logs `texnum`, `format`, `width`, `height`, `compmethod` from the bitstream header, then the first row of decoded RGBA5551 pixel words, for every image shaped 16x14 (the FIRE tile shape); (b) in `import_texture_rgba16` (`port/fast3d/gfx_pc.cpp`) — logs the same pixel words as read back at upload time, so a session can diff "as decoded" vs "as consumed" to rule pool corruption/mis-addressing in or out. **Environment blocker hit this session:** could not get the built `ge007.x86_64.exe` to produce *any* stderr/log output when launched from this delegated agent's shell, across four different invocation strategies (direct foreground exec, backgrounded `&`/`nohup`, `cmd /c start foo.bat` detached, `cmd /c foo.bat` synchronous with a hard timeout) — the process either exits or is killed before a single `osSyncPrintf`/`sysLogPrintf` line reaches the redirected log file, even though every print call in this codebase does an explicit `fflush(stderr)`. A plain GUI sanity check (launching `notepad.exe` the same way) worked fine, ruling out a blanket "no GUI" sandbox restriction, so the failure is specific to this binary/environment combination and its cause was not identified. Per `env-no-runtime-verify` (session memory): runtime verification is session-dependent; treat this session's empty captures as inconclusive, not as evidence of anything. **Next, concrete (for a session that CAN capture runtime output, interactive or otherwise):** (i) run any level with `GE_D219RAW=1` during an explosion/PPK firefight (or simply on level load, since these are likely eagerly cached) and read the logged `compmethod` for `IMAGE_FIRE_N` — cross-reference against a `GE_D219RAW`-instrumented capture of a *confirmed-correct-on-screen* RGBA16 sprite from the sibling `sImageTableEntry` families (e.g. `IMAGE_IMPACTMULTI`) to see whether FIRE uses a different `TEXCOMPMETHOD_*` (the compressed-bitstream compression method, not the pixel format) than a texture confirmed to render correctly — `TEXCOMPMETHOD_LOOKUP`/`_HUFFMANLOOKUP`/`_RLELOOKUP` (`texBuildLookup`+`texInflateLookup*`) build/consume a packed-color lookup table with a completely different bit-layout code path than `_HUFFMAN`/`_HUFFMANPERHCHANNEL`/`_RLE` (`texChannelsToPixels`, per-plane R/G/B/A order); if FIRE's method differs from the working comparison texture's, that pinpoints exactly which decode function to scrutinize next, rather than continuing to treat "the RGBA16 decoder" as one monolithic suspect. (ii) if the compmethod matches, the bug is genuinely upstream of decode (ROM/asset-extraction data itself, not the decompression code) — at that point compare the decompressed byte stream against a reference N64 emulator's texture-dump of the same asset (this repo has no such reference on hand) to establish definitively whether the ROM bytes for FIRE really do decode to blue/purple on real hardware too (in which case the bug is actually downstream in the color-combiner path after all, reopening the D172/G_CC_INTERFERENCE lead this session did not have budget to re-examine) or whether the extraction into this repo's `assets/oddtextures.c`-adjacent ROM data diverges from the real cart image for just these 15 offsets. Probes left in tree, both inert by default: `GE_D219RAW` (`src/game/image.c` `texInflateNonZlib`, `port/fast3d/gfx_pc.cpp` `import_texture_rgba16`). **M-141: Steam Deck playtest confirmation, widened symptom scope** — user reports "blood is purple, shots on body are blue/green" (general, not tied to a specific level in the report). Blood/hit-decal color is a new specific instance beyond this row's tracked "PPK explosions"/"particle effects" framing — same RGBA16/particle-color neighborhood as D172/D252, worth keeping in mind together rather than as three unrelated color bugs; whichever gets a real runtime capture first (this row's own `GE_D219RAW` probe, or D252's `GE_D252`) should also check blood decals specifically. |

**M-111 (2026-09-12, live session, 6 build/relaunch cycles): G_LIGHTING lead RULED OUT; texture-cache-collision lead RULED OUT; ROOT CAUSE FOUND — R/B channel swap in the decoded fire particle texture data itself.** Live-captured 4 rounds against a real user playtest (not scripted) with the user confirming explosions were still blue/purple in every capture:
1. **G_LIGHTING probe fired 5000+ times across a full playtest with explosions triggered, 100% `LIGHTING=off`, 0 `D219 SUSPECT` hits.** (First attempt used the original 40-hit cap from the D172 probe, which silently exhausted on ordinary bullet-impact particles before any explosion — fixed to exempt the lit case from the cap and add an always-on periodic counter, confirming the code path really was being hit throughout, not just early. See `port/fast3d/gfx_pc.cpp` `gfx_sp_tri1`'s D172/D219 probe block.) **G_LIGHTING is conclusively not the mechanism.**
2. **Texture-cache-collision lead also ruled out**: extended the same probe to log the actual bound `TEXEL1` texture address/GL-texture-id/byte-count for every particle draw. 7 distinct 448-byte (16×14 RGBA16, matching the DL's own `CALC_LRS(16,14,...)`) addresses observed, each internally consistent — no address aliasing, no size mismatch, no stale-cache-slot signature.
3. **Root cause: dumped the actual decoded fire textures with `GE_TEXDUMP=1`** (extended the existing dumper in `port/fast3d/gfx_opengl.cpp` to exempt 16×14 RGBA images from its own 400-image cap, same masking problem as the D172 cap) — captured all 15 `IMAGE_FIRE_0`..`14` variants as PPM after a live playtest with confirmed-purple explosions. **Every single one of the 15 decoded textures is blue/purple-dominant** (e.g. `fire000`: avg RGB `(102, 67, 138)` — B is the highest channel). Swapping the decoded R and B channels post-hoc on all 15 turns every one of them into a plausible warm fire palette (`fire000` swapped: `(138, 67, 102)`) — consistent across all 15, not a coincidence on one sample.
4. **This is NOT a bug in the shared RGBA16 decoder** (`import_texture_rgba16`, `port/fast3d/gfx_pc.cpp:653-681`) — that function's bit layout (`r=col16>>11, g=(col16>>6)&0x1f, b=(col16>>1)&0x1f, a=col16&1`) is standard N64 RGBA5551 and is proven correct by every other RGBA16 texture in the game rendering correctly; a bug in the shared function would break far more than 15 particle images. Also ruled out: the general C-array byte-order normalization (`gfx_tex_normalize_source`, D71) already runs upstream of every `import_texture_*` call including this one (`gfx_pc.cpp:1001-1008`), so it isn't a missing-normalize gap either. **The wrongness is in the decoded bytes themselves**, i.e. either (a) the ROM-resolved `Globalimagetable` entries for these 15 `IMAGE_FIRE_N` ids point at the wrong ROM offset/wrong data (an extraction/table-mapping bug, not a runtime rendering bug), or (b) these 15 images were authored/stored in a different sub-format than the rest of the game's RGBA16 assets that the shared decoder doesn't special-case. **Do NOT attempt a blanket R/B swap in `import_texture_rgba16`** — it's shared by the whole game and would break every correctly-rendering RGBA16 texture; whatever's wrong is specific to these 15 image ids. **Next, concrete:** audit how `IMAGE_FIRE_0`..`14`'s `Globalimagetable` entries resolve to ROM offsets (`port/src/gimgfixup.c`, `src/game/image.c`, `port/src/romassets_u.s`) and cross-reference against a real ROM dump / reference emulator screenshot of the same texture to confirm whether the extracted bytes are simply wrong-offset (fixable at the table/extraction level, in `port/`, no game-logic change) vs. a genuine N64-side sub-format quirk unique to `oddtextures.c`'s assets. Diagnostic infra left in tree (env-gated, inert by default): D172/D219 probe's uncapped lit-case + periodic summary counter (`gfx_sp_tri1`), and `GE_TEXDUMP`'s 16×14-exempt-from-cap fire dumper (`gfx_opengl_upload_texture`) — both safe to reuse for the follow-up.

**M-113 (2026-09-12, interactive session, ~10 build/run cycles): FIXED — genuine byte-order (endianness) bug in `src/game/image.c`'s runtime texture decompressor, not a data/extraction/combiner issue. Verified with real decoded-texture PPM captures (visual + numeric).** Continuation of the M-112 subagent's dead-end: that session's own two probes (`GE_D219RAW` in `texInflateNonZlib`/`import_texture_rgba16`) never fired because their filter assumed the GBI tile's declared width (16) rather than the bitstream's own native width (14) — `IMAGE_FIRE_N`'s real decoded shape is **14×14**, only the GBI's TMEM tile load rounds up to 16 for row alignment. Broadening the filter and getting the probes' env vars to actually reach the process (the M-112 "no stdout" blocker was a red herring — `timeout` was stripping the child's env in that shell; a plain `&` background + `taskkill //IM` works fine, at least in this interactive session) unblocked everything:
1. **Root cause:** `src/game/image.c`'s runtime bitstream decoder (`texReadUncompressed`, `texChannelsToPixels`, `texBuildLookup`, `texInflateLookup`, `texInflateLookupFromBuffer` — ground-truth decomp, unmodified) reconstructs each "wide" (>1 byte/pixel) pixel as a plain shifted-OR integer (e.g. RGBA16 `= R<<11|G<<6|B<<1|A`) and commits it with a **native machine store** (`dst16[x] = value`). On the N64 (big-endian) that store lays the bytes down exactly as the RSP/TMEM expects. `port/fast3d/gfx_pc.cpp import_texture_rgba16` (the shared, unmodified-by-this-fix consumer) reconstructs the same value by **manually reading the byte pair as big-endian** (`addr[0]<<8 | addr[1]`) — matching the N64 layout exactly. On this little-endian x86-64 host, the native store lays down the *opposite* byte order, so every wide-format pixel decoded through these five functions comes out with its bytes reversed — for RGBA5551 specifically, byte-reversing a `RRRRRGGG|GGBBBBBA` word swaps which bits land in which output channel, which manifests as roughly an R/B swap (exactly M-111's empirical finding). **Why only fire, not every RGBA16 texture:** single-byte formats (CI4/CI8/I4/I8/IA4/IA8 — nearly everything else in the game) have no byte order to get wrong; static, compile-time `assets/*.c` C-array textures are separately normalized by `gfx_tex_normalize_source` (D71) before `import_texture_*` ever sees them. `IMAGE_FIRE_0..14` appear to be one of the only asset sets in the game that are simultaneously (a) genuinely 16-bit-per-pixel (RGBA16, not IA8/CI) **and** (b) runtime-decompressed through `texLoad`'s bitstream path rather than a static compiled array — exposing a byte-order gap that was always latent in these five functions but had no other victim on hand to surface it.
2. **Confirmed empirically, not just by inspection:** dumped raw decoded RGBA5551 words at the exact point `texChannelsToPixels`/`texInflateLookup` write them (`GE_D219RAW=1`, Dam `-level_33`, scripted PPK fire via `GE_INPUTSCRIPT`) — e.g. pre-fix texnum 2111 decoded word `0xf1c2` (R=30,G=3,B=1 under standard RGBA5551 — briefly looks *plausible* red until you realize this is the value BEFORE the byte-swap fast3d applies; the two-stage mis-cancellation is what produced the reported blue/purple, not a one-step swap — see the fix verification below for the real signal). The conclusive test is the rendered/uploaded texture, not the intermediate value in isolation.
3. **Fix (`src/game/image.c`, `#ifdef PORT`/`#else` macro, no `src/` logic changed):** added `PORT_PIXEL16`/`PORT_PIXEL32` — `__builtin_bswap16`/`32` under `#ifdef PORT`, identity under `#else` (so the N64 build is byte-for-byte unaffected, verified by inspection: the macro is a no-op expansion, not a code path change) — and wrapped every wide-pixel (`RGBA32`/`RGB24`/`RGBA16`/`IA16`/`RGB15`) store in the five functions listed above. CI/I4/I8/IA4/IA8 branches (single-byte, no byte order) are untouched. This is the D3x/AGENTS.md#2 narrow ABI/byte-layout exception class: mechanical, semantics-preserving (same numeric color, corrected byte order for the LE host), scoped to exactly the format cases that need it.
4. **Verified, numerically and visually:** `GE_TEXDUMP=1` PPM capture of the 15 `IMAGE_FIRE_N` textures post-fix — average RGB flipped from B-dominant (M-111: `fire000 (102,67,138)`) to R-dominant (post-fix `fire000 (141,127,53)`, `fire003 (147,81,24)`, `fire005 (47,10,2)`); a 10x-upscaled render of one frame and a 6-frame animation strip (frames 000-006, session scratchpad) show a clean, unambiguous fire-sprite animation — bright yellow-white core cooling through orange to red/black ash, exactly the expected in-game fire look, no purple/blue anywhere. Sanity-checked a montage of 12 ordinary (non-fire) world textures from the same Dam capture (clouds, hazard stripes, pipes, grating, concrete) — all render correctly, no regression from the format-gated change. `-level_09` (BUNKER1) and `-level_33` (Dam) both boot and render crash-free post-fix (10s+ each, no `ge007.crash.log`). **Not yet done:** a full 21-level `verify.sh sweep` and a genuine live human playtest confirming the fix in real (non-scripted) gameplay — the scripted-fire + texdump evidence here is strong but the row should stay open until a human eyeballs a real firefight on this build. `docs/porting-notes.md` gets a new entry for this bug class (runtime-decoded multi-byte pixel formats need explicit byte-order handling; static C-array textures already get this via D71/`gfx_tex_normalize_source`, runtime-decompressed ones did not). **M-114 — REGRESSION, reverted.** User playtest after M-113 landed reported 1P muzzle flash rendering blue and the ammo HUD counter rendering pink — colors inverted on textures that were fine before. Root cause: M-113's `PORT_PIXEL16`/`PORT_PIXEL32` swap is applied unconditionally to every call site in `texReadUncompressed`, `texChannelsToPixels`, `texInflateLookup`, and `texInflateLookupFromBuffer` — i.e. to every RGBA16/RGB15/IA16/RGBA32/RGB24 texture decoded through these shared functions, not just `IMAGE_FIRE_N`. This is the exact failure mode already logged above against the sibling function ("do NOT attempt a blanket R/B swap in `import_texture_rgba16` — it's shared by the whole game and would break every correctly-rendering RGBA16 texture"), just repeated one file over. Muzzle flash and the ammo HUD icon evidently decode through this same shared path without needing the swap. **Fix applied:** `PORT_PIXEL16`/`PORT_PIXEL32` reverted to identity (no-op) in `src/game/image.c`; the macro/call-site wrapping is left in place as scaffolding. **D219 is OPEN again** — the fire-particle color bug is unfixed. **Next, concrete, not yet done:** find the actual property that distinguishes `IMAGE_FIRE_N`'s decode from a working same-format sibling (M-112 already narrowed one candidate: FIRE is baked directly into a DL rather than going through the `sImageTableEntry` table that flare/ammo/impact sprites use — confirm whether that DL-embedded path is what needs the swap, or whether it's specifically `compmethod`, per M-112's still-open next step) and scope the macro to that property, not to a hardcoded texnum range. Not runtime-verified this session (static revert only, done to stop an active regression fast) — a human should confirm muzzle flash/ammo colors are back to normal on this build. **M-115 (2026-09-12, delegated session, 3 build/run cycles): fix RE-APPLIED, scoped by GBI pixel format — the M-114 regression resolved by construction, fire fixed again.** Per M-114's brief, found the property that distinguishes `IMAGE_FIRE_N`'s decode from working same-format siblings and scoped the macros to it. (1) **The discriminator is the decoded pixel FORMAT, not compmethod or texnum.** Full offline census of all 2698 ROM image-table entries (`scratch/d219_headers.py`, parsing real bitstream headers with the same MSB-first `texReadBits` contract as `src/game/image_bank.c`; zlib images skipped after the 2-byte `rz` preamble per `decompressdata`): only 94 non-zlib wide-pixel images exist — RGBA16=19, RGB15=14, IA16=1, RGBA32=60. The fire frames span compmethods RLE(4)/LOOKUP(5)/RLELOOKUP(7)/HUFFMANBLUR(8) (confirmed offline AND at runtime), so compmethod cannot scope the fix; a texnum range is a hardcoded landmine. Format IS structural: every 16-bit GBI-size format (RGBA16/RGB15/IA16) is consumed by importers that read the pool as **manually big-endian bytes** (`import_texture_rgba16`: `col16=(addr[0]<<8)|addr[1]`; `import_texture_ia16`: `intensity=addr[0], alpha=addr[1]`), while every 32-bit format (RGBA32/RGB24) is consumed by `import_texture_rgba32`, which does `PD_BE32()` on a **native** u32 load — it expects host-order words. So bswap-at-store is required for exactly the 16-bit family and must be identity for the 32-bit family; M-113's blanket swap hit both (M-114), M-114's full revert broke the 16-bit family again. (2) **Change (`src/game/image.c`, `#ifdef PORT` only, no call-site edits):** `PORT_PIXEL16(x)` = `__builtin_bswap16((unsigned short)(x))`; `PORT_PIXEL32(x)` = `(x)`. Every existing call site already routes by `TEXFORMAT_*` to the right macro, so M-114's scaffolding was already format-scoped per site — only the macro definitions were wrong (both bswapped). 30-line comment at the definition records the importer contract + census. (3) **Runtime verification (Dam `-level_33`, scripted Z-pulse fire via `GE_INPUTSCRIPT`, `GE_D219RAW=1` + `GE_TEXDUMP`/`GE_TEXRAW`):** all 15 fire frames decode at runtime as format=1 (RGBA16) with compmethods 4/5/7/8 exactly matching the offline census (texnum 2085 is genuinely an nc=1 LOOKUP transparent first frame — A=0 everywhere; the visible frames are 2087–2113); the raw pool dumps decode as warm RGBA5551 sprites — mean RGB over opaque texels **(255,238,139)**, R-dominant, correct vertical-blob shape (rows 3–10 of the 16×14 tile); e.g. texnum 2111's pool word is now `0xc2f1`, which the importer reads back as `0xf1c2` = R30,G3,B1 — the true decoded value M-113 recorded pre-fix, i.e. exactly the pool state whose rendered PPMs M-113 proved correct (mean 141,127,53, clean animation). (4) **Regression safety is structural, not empirical:** RGBA32/RGB24 pool bytes are bit-identical to the M-114-revert build (`PORT_PIXEL32` = identity in both), so muzzle flash / ammo HUD / flares / crosshairs — all RGBA32 — render exactly as in the state the user last confirmed fine; no 32-bit path was touched at all. The other non-zlib 16-bit images (texnum 1198–1201 RGBA16, 2430 IA16, 2510–2523 RGB15) were equally byte-reversed before this change and are now corrected by the same importer contract; they do not load during a Dam run, so their on-screen state is owed to the next human playtest in whatever level uses them. (5) **Probes left in tree** (env-gated, inert): `GE_D219RAW` generalized in `texInflateNonZlib` — header line for every non-zlib wide-pixel decode (`format <= TEXFORMAT_IA16`), first-row pixel dump for RGBA16/RGB15/IA16, no texnum filter; plus the pre-existing D172/D219 and `GE_TEXDUMP` fire-dumper probes. **Status: RESOLVED (2026-09-12)** — user playtested this build and confirmed it works (fire correct, no muzzle-flash/ammo-HUD regression); D219 closed. |
| D220 | **1P muzzle flash position shifts upward at higher `Video.FovScale` (M-87, user QA, Jungle/M16).** Likely shares D211's noted residual: the viewmodel draws through `c_perspfovy`/unscaled fovy while the world view is FovScale-scaled (`fr.c:737`), so a flash sprite anchored to viewmodel-space vs. screen-space math can drift as the two diverge more at higher scale. | **ANALYZED (M-91) — the suspected unscaled-viewmodel root does NOT hold; this is correct perspective behavior, likely cosmetic/QoL.** The 1P hand + flash are model nodes in ONE gun DL (`gunfire.c` builds `rwmtx[1..4]` for hammer/flash/billboard nodes) drawn through the single per-frame projection from `viSetupCurrentPlayerView` — which IS FovScale-scaled on PC. So the flash and the muzzle it attaches to share projection and matrix chain: their relative position is geometrically preserved. What moves is the whole camera-anchored viewmodel: under perspective, screen offset from center ∝ 1/tan(fovy/2), so widening the FOV migrates every camera-space point (gun body included) toward screen center — a muzzle sitting slightly below/right of center visibly drifts up-left at FovScale 150. Same family as the M-85 accepted "viewmodel scales down, stays bottom-right anchored". `c_perspfovy`/`currentPlayerSetCameraScale` do stay unscaled (the D211 note) but that feeds frustum-cull/screen↔world math (→ D222), not the flash's render path. **Disposition: human call** — if it bothers at high FovScale, the fix is a QoL viewmodel re-anchor/scale-at-FOV (VIEWMODEL-RESEARCH.md family), not a fidelity bug; no game-logic change either way. Optional headless confirm: Jungle M16 fire at FovScale 100 vs 150, measure flash centroid vs muzzle in PPM pairs (expected: both shift toward center together). |
| D85 | room primary/secondary DL → `texLoadFromGdl` garbage | OPEN (safety-netted; widen/pool fixes landed, geometry now renders) |
| D86 · D87 | modelInitRwData truncated ptr · attract-demo BE `ramromfilestructure` | resolved |
| D88.1–D88.3 · D88.5–D88.6 | `Usetup*Z` header + sub-table width/endian conversion (`d88_emit.py`) | resolved |
| D88.4 | `propDefs` polymorphic record stream not byteswapped → `setupDoor` crash | resolved (`d88_propdefs.py`); layout audit M-20 → **D132** |
| D132 | D88 propDefs layout audit (M-20): converter cursor confirmed vs real PC struct layout for all types the 21 levels emit. DOOR/OBJECT-prefix/VEHICHLE/AIRCRAFT/TANK/AUTOGUN/AMMO/TINTED_GLASS/objective sub-records all MATCH. **Divergence found:** types 14 LINK / 19 SWITCH / 38 LOCK_DOOR / 44 SAFE_ITEM — each has `s32 IndexN` fields sharing a union with a pointer (`LinkRecord.first`/`Index1`), so on PC the field sits in an 8-byte, 8-aligned union slot, but the converter emits it at N64 tight-4-byte-word offsets → `pdef->Index1` reads `Index2`'s value, `Index2` reads 0 → switch-doors / dual-weapons / locked-doors / safes silently fail their validity guard (non-crash; guard failure also prevents the under-sized-record `->next` overflow). | proposed fix (not applied) — see below |
| D157 | **Campaign never unlocks the next level (M-30, user bug — "complete Dam on Agent, Facility stays locked, no save").** — full `## D157` entry at file tail | FIXED (`src/bondtypes.h`, `#ifdef PORT` —… |
| D160 | **Dam level-end cutscene (Bond off the dam / bungee) never plays — cuts straight to the MISSION COMPLETE report (M-30/M-31, user bug D148).** — full `## D160` entry at file tail | DIAGNOSTIC SHIPPED, not root-caused (`c95713f5`, `#ifdef PORT` + `GE_D160=1`)… |
| D161 | **Depot (`-level_30`) ceiling renders as bright-blue speckle + radial "light-ray" streaks converging to a point instead of a dark corrugated roof (M-30/M-31…** — full `## D161` entry at file tail | FIXED (`port/fast3d/gfx_pc.cpp` `import_texture`, `port/fast3d/gfx_opengl.cpp` diag). porting-notes.md §D. |
| D89–D92 | stage-load→frame: stan zero-fill overrun, portal address trunc, chr/AI spawn ptrs | resolved |
| D93–D102 | struct-pun / hardcoded-size pass: player alloc, player.model inline Model, weapon Model pun, master-DL buffer | resolved (D95 pool bump partial) |
| D103–D107 | BUNKER1 viewport height, depth-buffer clear (`G_CLEAR_DEPTH_EXT`), portal near-plane, LOD mip tile | resolved |
| D108–D112 | skeletal models: `d43_emit.py put_f32` byte-reversal bug | resolved |
| D113–D116 | portal BFS ok · matrix chain ok · player raw-offset audit + `gunfire.c` THROW* · HUD text X-mirror | **D114/D116 CLOSED — NOT A BUG (M-33, see D168): the "X-mirror" was an upside-down `GE_PCDUMP` capture misread as mirrored. PPM writer fixed; the game renders correctly on hardware.** |
| D117 | frame nondeterminism root-caused; `GE_DETERM` deferred; `framediff.py` structural. **M-52: `GE_DETERM=1` attempted — boot deadlock fixed, determinism NOT achieved (load-phase fallback still real-time-dependent). See addendum below.** | OPEN (mode attempted, not achieved) |
| D118 | SDL input layer (`port/src/input.c`); M-24 mouse-look rework: mode-aware map (aim mode = analog stick, hipfire = digital C-pitch), `config.c` INI now real. D118b/c FIXED; D118a residual (hipfire pitch only) | resolved (D118a residual) |
| D150 | watch OBJECTIVES / BRIEF page crash: `strcat(buf, langGet(id))` with `langGet` → NULL (unloaded bank) → NULL deref; `str.c` str* builtins elide a plain param NULL check, so guard via asm-laundered `GE_IS_NULL` | FIXED (`src/str.c` `#ifdef PORT`); interactive re-verify pending |
| D260–D265 | v0.2.1 hotfix batch: D260 wheel direction swap (fixed) · D261 watch-menu nav auto-repeat (fixed, feel-check owed) · D262 watch item-preview model missing → root-caused to D264 (fixed at Gfx level, visual check pending) · D263 Facility Ourumov/Trevelyan beat — **resolved: not a bug**, proximity/LOS-triggered as on N64; full chain incl. execution shot verified firing on PC (teleport probe + user playtest) · D264 watch preview renderdata.flags zeroed by cross-global template read on PC (ABI/layout class; fixed) · D236/D265 Surface 1 tree "wall": IA16 importer dimensions transposed (fixed); decode attempt reverted after raw-byte proof that byte-pair is the authored image; root cause = `import_texture_ia4` decoding RDP I2:A2 as I3:A1, inverting the tree strip's alpha ramp → opaque repeating silhouettes; IA4 I3:A1→I2:A2 decode fix applied and user re-tested — **still a wall; reverted, parked** with evidence + suspect list in D265 | partial — see rows |
| D266–D269 | D236/D265 wall, second pass: **the wall is CI8 32×32, not IA** (D269) — band-gated probe finds two draw classes per room stream (OPA oml=0xc4112078 + XLU oml=0xc41049d8), both BILERP/TEXTLOD/2-cycle/**FOG-on**; cards are soft gradient blobs, stripe period ≈ one card per visible room. D266: re-applied the reverted I2:A2 IA4 decode + replaced the hardcoded 0.19 edge quantize with AC-aware handling (uncommitted; census shows zero IA4 draws at boot). D267: IA8 dump infra — the three IA8 images are noise/mast sprites, not the wall. D268: force-alpha A/B proves the blend pipeline works. Ranked suspects now: fog range/application (D218 family) > TEXTLOD/mip sampling > draw density/culling > palette | OPEN — probes live in tree (`GE_D266`/`GE_D266W`/`GE_D267`/`GE_D268_FORCEA`), fixes uncommitted |
| D270 | **Backfilled (M-132) — this label was never given its own row; only code comments and the HANDOFF archive documented it.** D236 wall, between-pass instrumentation toolkit for the fog hypothesis (D269 ranked suspect #1) and general A/B isolation: `gfx_sp_tri1` per-vertex fog min/max attribution; `GE_D270SKIP`/`GE_D270SKIPOML`/`GE_D270KEEPOML`/`GE_D270KILLALL` (drop-by-oml/fmt A/B rules); `GE_D270PIX` (single-pixel draw-order attribution); `GE_D270FOG` (per-frame fog-color-change log); and the `GE_TEXRAW` dump cap raise 400→600 to reach the wall cards' upload index. No independent conclusion of its own — superseded by D271's culling fix and D272's later per-card audit, both of which reused this toolkit's skip/keep hooks. | dead-ended into D271/D272; infrastructure kept in tree |
| D271 | **D236 wall, third pass — suspect #3 (draw density/culling) CONFIRMED as a real over-draw, and the D106 guard was the cause.** At Surface 1's boot camera, 11 portals around room 13 straddle the near plane every frame; their z==0 clip points project to ±1e20–1e26-scale FINITE screen coords. The original D106 guard (`lim=1e5`) mapped any such bound to a full-screen box — but on N64 there is no guard: those finite values flow into `bgRectIntersect` (pure comparisons), which either clamps to the parent box (portal covers view → keep) or yields an empty box (portal projects off-screen → **cull**). 9 of the 11 portals clamp empty on N64; PC was force-drawing those rooms. Fix: guard `lim` 1e5→1e38 (catches only NaN/inf, which this projection can't reach — max ~1e26); finite min>max still hits the original `degenerate` check (N64's own path, D106's real regression case unchanged). Verified: `d106Fallback` 396–781/600 ticks → **0** (Surface 1 + BUNKER1); roomsDrawn 17–18→10–13 at the boot view; per-portal oor-finite dumps confirm every culled portal's finite portion is far off-screen (N64 clamps identically); BUNKER1 room drops traced to the same N64-matching empty clamp. No crashes in long unattended runs. **USER RE-TEST (2026-09-14): NEGATIVE — trees still render as a big wall; user's read: "rather than the texture being mapped correctly"** → suspect #3 was a real over-draw but NOT the wall's root cause; fix kept for N64 fidelity. D236 thread continues — top suspect now per-card S/T / UV windowing (D265 note: all captured cards sampled the same u/v window) | FIXED-but-insufficient (`src/game/bg.c` `#ifdef PORT`, uncommitted) — probes live (`GE_D271`, extended `GE_D104`); wall still open, see D236 |
| D276 | **D236 wall, eighth pass — the true painter of the visible weave, at confirmed weave pixels, is a fully-opaque IA8 blocky-noise texture sampled with a fixed, identical 2×-wrapped UV window and low fog — NOT either of D269's named CI8 classes.** Multi-point pixel attribution + per-vertex fog (D275's `GE_D270PIX` list extension) at 4 pixels picked directly off the reproduced weave screenshot found the D269 OPA/XLU classes absent at 2 of them; the dominant front-most class there (`oml` low bits coincide with the XLU class but `fmt=3`/IA, not CI8) decodes to blocky gray static, matches D267's already-dumped-but-dismissed "seq 113" texture, and reproduces D265's original "all cards sample u:0..64/v:0..16" observation to the exact texel — that lead was reframed away from prematurely in D269's second pass. A third class (`0x0c184b50`, CI8) is present everywhere with erratic z/w (`-0.04..0.99`) and high fog when it does land normally — flagged as the possible real "tree" layer, with its own likely occlusion/billboard bug, meaning D236 may have two separate root causes, not one. | REFRAMED — `oml=0xc81049d8`/`fmt=3` is the primary suspect now, not D269's CI8 pair; `0x0c184b50` unattributed; wall still open, see D236 |
| D279 | **Backfilled — this label was consumed by an undocumented partial session (2026-09-14 ~23:00) that left artifacts in `scratch/` (`d279_run.log`, `d279_imgmap_run.log`, `d279_g110.png`, `d279_g220*.png`, `d279_gray_1500.png`) but no written conclusion; the logs carry no D279-tagged lines, so its exact probe/finding cannot be reconstructed from the tree.** Next available label after it: D280. | undocumented — artifacts only |
| D281 | **`Game.AllUnlocked` (D257) still breaks audio, and can also break right-mouse aim, when enabled before any save has ever been written (v0.2.1 user testing, 2026-09-15) — a distinct regression from D259, which only fixed the all-zero-slot volume case.** User's empirical workaround: complete a level normally first (e.g. Dam on Agent, which writes a real save), THEN enable `All unlocked`; enabling it on a completely fresh install (no save yet) still reproduces both symptoms. Not root-caused this session — D259's `geEepromPatchAllCheats` all-zero-slot seeding was verified against exactly the "fresh eep" case, so either that fix doesn't cover every code path a truly-first-launch toggle can hit (e.g. a race between the F10 toggle taking effect and the very first save-slot creation/read), or there's a second, independent mechanism. The right-mouse-aim symptom is new information with no prior lead — cross-ref D211/D194 (aim-mode sensitivity/mode family) for where a save-state-dependent aim regression could originate. | OPEN — user-reported regression, not root-caused; mitigation is user-facing guidance only (release notes: don't enable until a save exists, treat as experimental). |
| D282 | **QoL ask (v0.2.1 user testing, 2026-09-15): front-end menu navigation (main menu, file select, mission-select map) currently requires the RIGHT analog stick — matches N64 default control schemes but is unintuitive on a modern controller/Steam Deck, where the left stick is the expected primary-navigation input.** Distinct from the F10 options overlay, which is already left-stick/D-pad navigable (`port/src/optionsoverlay.c`). Scope: front-end menu input in `src/game/front.c` (or wherever N64 stick reads are gated by the active `controldef`) — likely wants a port-layer input remap for these menu contexts specifically, not a general control-scheme change. **M-141: root cause found + fixed, port-only.** `front.c`'s menu-nav code reads `joyGetStickX/Y` — the same emulated N64 "Control Stick" channel (`stick_x`/`stick_y`) that in-level gameplay uses. `port/src/input.c`'s gamepad poll routes that channel to the LEFT physical stick by default, but `naturalPitchMode` (D194/D238's SOLITARE-style scheme, **default-on**) reroutes it to the RIGHT stick for continuous look — and since menus share the channel, they silently inherited that reroute too. The poll already computed an unused `padMenuMode` flag (`current_menu != GE_MENU_RUN_STAGE/GE_MENU_INVALID`, previously only gating shoulder-button weapon-cycle edges); moved it earlier and added a `padMenuMode` branch that always sources `stick_x`/`stick_y` from the left stick outside a running stage, regardless of `naturalPitchMode` — matching the F10 overlay's existing left-stick convention (`optionsoverlay.c`). Zero effect on in-level control feel (the branch is inert whenever `current_menu == GE_MENU_RUN_STAGE`). Verified: build green; `-level_09`/`-level_20` unregressed (1100+ frames each, no crash) — bare `-level_XX` launches skip the front end entirely, so this only proves the gameplay path (unaffected) stays clean. **In-menu feel not interactively verified this session** (no physical gamepad available in this environment) — needs a human playtest with a controller to confirm the menu now navigates on the left stick as expected. | FIXED (`port/src/input.c`, `#ifdef` not needed — port-layer-only file), feel unconfirmed. |
| D283 | **Steam Deck first-run preset (D-unnumbered in `port/src/video.c`/`main.c`, `videoApplySteamOSDefaults()`) did not apply for a user who added the game via Steam in Desktop Mode and switched to Game Mode afterward — resolution stayed at some non-native value instead of the Deck's native 1280×800 (v0.2.1 user testing, 2026-09-15).** Root mechanism, confirmed by code inspection: the preset only applies when `getenv("STEAMOS")` is truthy AND no `ge007.ini` exists yet (`main.c:126-129`, before `configLoad()`); once any ini exists, it always wins over the preset on every future launch (by design, so user F10 changes persist). **Leading hypothesis (not verified on real hardware in this environment):** SteamOS's Desktop Mode (a normal Plasma/X11 session) likely does not export the same `STEAMOS` env var that the Game Mode/gamescope session does, so a Desktop Mode test-launch (a natural step when adding a non-Steam game via its file browser) creates `ge007.ini` with ordinary Linux defaults *before* the user ever reaches Game Mode — permanently skipping the preset for that install, since the ini already exists on the first real Game Mode launch. This exact caveat was already known internally (`docs/HANDOFF.md`: "a Deck that already has an ini needs `-fresh` or F10 to pick the preset up") but was never surfaced in user-facing docs (README/release notes) — same class of gap as the stale D230 release-notes reference. **Workaround (lighter than `-fresh`, which also wipes the save):** F10 → Resolution → 1280×800 manually; takes seconds, no data lost. | OPEN — mechanism understood, root cause (Desktop-vs-Game-Mode `STEAMOS` env var difference) unverified on real hardware; user-facing workaround documented in README/release notes. |
| D280 | **D236 wall, ninth pass — fog math VERIFIED N64-faithful (±2 truncation over all 2505 verts of `0xc81049d8`; D276 owed step #1 closed), the IA8 noise class VISUALLY confirmed as the weave's painter (`GE_D278C` collapse: band HF 14.8→2.9, no trees appear without it — first unconfounded test), full census at both confirmed weave points (only 3 real classes; two log entries are probe artifacts), D276's depth-race read INVERTED (`0x0c184b50` tree cards are in FRONT of the noise quad, z/w smaller=closer), `0xc8102078` ruled out as trees (terrain, 1.4M tris), and `0x0c184b50` shown to barely instantiate under deterministic `-level_36` boot (~4 tris vs 720 free-run) with CPU-baked bimodal fog 25/255. **No quick win for trees** — remaining work is CPU-side instantiation/room-streaming, not a port fix. | REFRAMED again — no fix; next steps in detail section, see D236 |
| D275 | **D236 wall, eighth pass — probe upgrades:** `GE_D270PIX` extended from a single point to a semicolon-separated list (each line now tagged `pix=(x,y)` + D274's reliable `addr=`); `GE_D270KEEPOML` gained a "kept N tris" counter so a black A/B capture is diagnosable (tick-misaligned vs. a genuine render-interaction bug). Mechanical groundwork for D276; the class-isolation A/B itself is still owed. | done — probes live, uncommitted; see D276 for what they found |
| D274 | **D236 wall, seventh pass — probe identity (`seq`) proven unreliable both mid-run (cache-hit staleness) and cross-run (shared runtime-arena address reuse); fixed with an unconditional per-slot address stamp.** Resolved the prior pass's open "content vs. decode" question as a bug in this session's own offline analysis (wrong palette byte order — native/little-endian, not big-endian), not a port bug: the port's existing CI palette decode, applied correctly, reproduces a plausible soft gradient on the real, address-confirmed wall-card texture. Confirmed mips are generated (`genmip=1`) and enabled (`tex_lod=1`) for that texture, with correct `GL_REPEAT` wrap — ruling out "mips absent" as the mechanism. Attempted class-isolation A/B was inconclusive (black frame). | `d267_slot_addr` fix live (always-on); content/decode question CLOSED; mip-presence CLOSED; wall still open, see D236/D276 |
| D273 | **D236 wall, fifth pass — user-supplied N64 reference screenshot obtained (discrete tree silhouettes, not a sheet); a pre-existing, undocumented `GE_D273` palette-decode A/B probe found in the uncommitted tree was tested against it and RULED OUT.** Standard-ARGB1555 vs. inherited-shifted CI palette decode: wall pattern visually unchanged either way; unrelated CI8 textures (snow path, viewmodel skin) corrupted by the swap. Not the wall's cause; `GE_D273` stays default-off. Flags a process gap: a live investigative lead sat in the working tree undocumented across a handoff boundary. | RULED OUT (`port/fast3d/gfx_pc.cpp` `palette_to_rgba32`, pre-existing/uncommitted `GE_D273`) — wall still open, see D236/D272 |
| D272 | **D236 wall, fourth pass — per-card S/T audit (HANDOFF next-step #1) RULES OUT "identical UV window" and "converter bug"; the wall cards genuinely sample a wide, per-instance-varying UV range, verified authored in ROM, not injected by `d43_emit.py`.** New `GE_D272` raw per-vertex UV+tile probe (`gfx_sp_tri1`), filtered to D269's exact two oml classes. Both are confirmed CI8 (fmt=2 siz=1, matches D269). Findings: (1) per-instance UV is **not identical** across cards sharing the same bound texture (`seq`) — small (~0.3 texel) jitter for the OPA class, larger (~1.6 texel) jitter for the XLU class — consistent with genuinely distinct authored per-vertex S/T, not a converter collapse; D265's "identical u:0..64/v:0..16" observation traces to a different (non-wall) draw class caught by the old unfiltered band gate, not these cards. (2) The XLU class's U range spans **~143–145 texels — ~4.5× the tile's 32-texel wrap period** (`mask=5,5`); the OPA class stays within roughly one tile width. (3) Read `tools_pc/d43_emit.py`'s vertex emit path (`emit_main_vtx`): `tc[0]/tc[1]` (offsets 8/10 of the 16-byte `Vtx`) go through a plain `be16` byteswap with no scale/offset — **no converter bug possible here**; the ROM must already encode this ~4.5× wrap. Reframes the leading hypothesis: this heavy horizontal wrap is by design (Rare tiling one texture across a long tree-line quad), so the "wall" comes from *how* the port renders that wrap, not from wrong UV data — TEXTLOD/mip generation and GL wrap-boundary sampling for a highly-minified, wide-wrapped CI8 tile is now the top suspect again (D269 hypothesis #2), ahead of fog/palette. | OPEN — probe `GE_D272` added to `port/fast3d/gfx_pc.cpp` (env-gated, uncommitted); wall still open, see D236 |
| D137 | **right-mouse (aim-sight) crash:** — full `## D137` entry at file tail | FIXED (`src/game/gunfire.c` `#ifdef PORT` — `sp54` is `Gfx *`) |
| D119 · D120 | guard-attack `weapons_held[]->chr` pun crash · blood-stain `PointUsage[]` chain hang | D119 fixed · **D120 FIXED (M-30)** |
| D120 | **Blood-decal `PointUsage[]` chain — converter now emits it (M-30).** `d43_emit.py` reserved the opcode-0x18 `PointUsage` region in the layout pass (`add_region(puo, 2*nv)`) but the **emit pass never wrote it** — every entry stayed zero. `chr.c`'s decal walk `index = PointUsage[index]` then cycled 0→PointUsage[0]→0 forever (capped by the `#ifdef PORT` guard at `chr.c:3328`, so decals were missing/wrong rather than a true hang). Fix: `op24_pointusage[puo] = nv` in layout + a byteswap loop (`put_s16(remap(puo)+2k, be16(src, puo+2k))` for `nv` = `numVertices` entries — a negative-terminated `s16` index chain, no remap) in the emit pass. Full regen chain re-run, round-trip validation passes. The `chr.c` guard stays as a safety net (SMALL-FIXES B3). | FIXED (`tools_pc/d43_emit.py`). Interactive verify pending — blood decals on a BUNKER1 firefight should now appear + no hang. |
| D121 | WS1 frictionless per-level boot: bare `-level_XX` injects its `memallocstringtable[]` `-m*` row (`#ifdef PORT` in `boss.c`) | resolved |
| D122 | per-level prop/item model-load crash (Dam/Facility/Runway `modelLoad`/`modelInitRwData`): `d88_propdefs.py` had no handler for 6 ObjectRecord-derived propDef types (47/39/40/45/13/20) → generic arm half-swapped the `[s16 obj][s16 pad]` word → OOB `PitemZ_entries[]` | converter fixed; residual chr/fast3d crashes on those levels are separate |
| D123 | **crash class C1: `chrIsNotDeadOrShot` NULL deref on 6 levels (Dam/Runway/Frigate/Statue/Streets/Cradle).** — full `## D123` entry at file tail | converter fixed (`OBJ_ID_WORDS`); C1 cleared on all 6, residual crashes are fast3d (C2) |
| D125 | **crash classes C3+C6 (Aztec/Bunker2/Surface2):** — full `## D125` entry at file tail | **resolved (M-16b)** — line 374 → 8 NULs; all 21 sidecars regen'd; **Aztec `-level_28` now PASSES** (was C3 CRASH). Bunker2 falls through to a separate DOOR-tail `linkedDoor` layout bug (C3 residual) |
| D126 | **crash classes C3r/C4/C6 (Bunker2 `-level_27` `door7F054FB4` propobj.c:13523, Depot `-level_30` prop.c:902, Surface2 `-level_43` loadobjectmodel.c:393):** — full `## D126` entry at file tail | resolved — `d88_propdefs.py` PROPDEF_PC_BYTES[30/32/33/35]=24 + typed handler; `loadobjectmodel.c` sizepropdef PORT returns 6. Bunker2/Depot/Surface2 now PASS (13→16/21) |
| D124 | **crash class C2: fast3d bad texture pointer.** — full `## D124` entry at file tail | Jungle fixed (`port/src/gimgfixup.c`); Facility → D130 |
| D131 | **crash class C2m (Jungle `-level_37`):** — full `## D131` entry at file tail | resolved — `seg_addr()` restores the module high word for any fallthrough w1 in `[0x40000000, 0x70000000)` (module fixed-based at `0x140000000`, no ASLR; DRAM/KSEG0/segmented/phys all handled earlier).… |
| D133 | **intro render triage (M-20):** — full `## D133` entry at file tail | RESOLVED — not a regression, parked (D75/D76, `GRAPHICS-BACKLOG.md`) |
| D134 | **Frame-2 boot hang / sweep "NO-FRAMES" flakiness — root-caused and fixed (M-22).** — full `## D134` entry at file tail | FIXED (`port/src/libultra.c`) |
| D135 | **Firefight crash — `bgTestHitOnObj` (`propobj.c:8446`) is an unported N64 GBI parser (M-22, WS6 playtest).** — full `## D135` entry at file tail | FIXED (`src/game/propobj.c`, `#ifdef PORT` —… |
| D136 | **Stage-unload crash — `objFree` `obj->prop` reads float garbage (M-22, WS6 Facility playthrough).** — full `## D136` entry at file tail | ROOT-CAUSED → **D139** (it was NOT a layout bug — `cleanupObjects` read the type from the wrong byte on LE and walked off the propDef blob). |
| D139 | **Stage-unload crash root cause + fix (M-23).** — full `## D139` entry at file tail | FIXED (`src/game/cleanup_objects.c`, `#ifdef PORT` —… |
| D140 | **Pause-menu crash (M-23 open → M-27 fixed).** — full `## D140` entry at file tail | FIXED (`src/game/bondview.h`, `src/game/bondview2.c`; `#ifdef PORT`). Exposed the next watch-render crash → **D141**. `-level_09`/`-level_20` unregressed (3/3). porting-notes.md §A. |
| D118d | **Watch-menu list over-scrolls with keyboard (M-27, user bug report).** `game_options_inventory_navigation` / `sub_GAME_7F0A611C` (`options.c`) have a "slam the stick" fast-scroll: a raw *level* check `joyGetStickY(PLAYER_1) < -0x46` / `>= 0x47` that steps `watch_inventory_cursor_pos` by 1 **every frame** it's held. An N64 stick rarely sustains past ~0x46 on-axis so a human taps it; keyboard W/S (and SDL pads pegged at 0x80) sit there every frame → one keypress skips several items (breaks item selection needed to finish the game). Fix: `#ifdef PORT` macro `GE_WATCH_STICK_FAST{UP,DOWN}` → `0` for the port — stick list-nav then goes only through the latched single-step path (`watch_stick_y_pressed_*`, one step per press, existing game mechanism) and the -0x1e..-0x45 smooth-scroll band; D-pad/C-buttons keep edge-repeat. N64 `#else` = verbatim. | FIXED (`src/game/options.c`, `#ifdef PORT`). Build green, `-level_09` unregressed. **Feel to be confirmed in playtest** (probe couldn't reach the INVENTORY watch page headlessly). D118 input family. Latent sibling: `front.c` menu nav uses `joyGetStick*InRange` level checks the same way. |
| D141 | **Watch-menu item-model crash (M-27), exposed by D140.** — full `## D141` entry at file tail | FIXED (`src/game/gunfire.c`, `#ifdef PORT`). Watch page now renders (weapon page verified — mirrored text D114/D116, dark weapon model D75, both parked cosmetic). `-level_09`/`-level_20` unregressed. |
| D150 | **Watch objective/briefing page crash (M-28).** — full `## D150` entry at file tail | FIXED (`src/str.c`, `#ifdef PORT`). Build green (`ntsc-final`); `-level_09` 600+ frames @ 91.67% unregressed. Watch-page repro is interactive — user to re-verify. porting-notes.md §C. |
| D152 | **Mission-failed / objective-failed → permanent black screen (M-28, user bug report — LIVE process inspected, not yet fixed).** — full `## D152` entry at file tail | MITIGATED (`port/src/libultra.c`, M-28, `#ifdef PORT`/port-only).… |
| D152+ | **D152 addendum — static audit + narrow fix (M-31).** — full `## D152+` entry at file tail | FIXED (`src/snd.c`, `port/src/libultra.c`, `#ifdef PORT`). porting-notes.md §D4. |
| D154 | **`bg.c` room-geometry ray/hit-test GBI parser port (M-28, WRITTEN / UNVERIFIED / UNCOMMITTED).** — full `## D154` entry at file tail | **M-30 re-audit + fixes (still playtest-gated for final verify).** Cross-checked the ported G_TRI1 + all 12 G_TRI4 nibble recoveries twice against the N64 `#else` and the raw BE-word bit derivation —… |
| D156 | **Facility outro-cutscene "hang" — the actual fix (M-29, user bug report, 2nd occurrence after D155).** — full `## D156` entry at file tail | GUARDED (`src/game/model.c`, `#ifdef PORT`)… |
| D155 | **Facility outro-cutscene end → "hang" = unclamped `deltaFrames` spiral (M-29, user bug report).** — full `## D155` entry at file tail | FIXED (`src/game/frametiming.c`, `#ifdef PORT` —… |
| D153 | **`-level_09` frame ~900–1400 `0xc0000005` = D117/D134 load flakiness, NOT a bug (M-28).** — full `## D153` entry at file tail | NOT A BUG — closed. Run anchors on a quiet machine. |
| D151 | **Watch data all blank — objectives / mission background / M / Q / Moneypenny (M-28, user bug report).** — full `## D151` entry at file tail | FIXED (`src/bondtypes.h`, `#ifdef PORT` —… |
| D130 | **crash class C2 (Facility `-level_34`, Runway `-level_35`):** — full `## D130` entry at file tail | resolved — stage all 6 N64 fields in a local `f[6]` before writing any. Facility + Runway now PASS (16→18/21). Probes reverted. |
| D164 | **Disclaimer / legal screen only draws the first line (D76 root cause, M-31).** — full `## D164` entry at file tail | **FIX PROPOSED, NOT APPLIED** (`front.c` is owned by another agent this session).… |

| D159 | **Front-end wallet-Bond photo (and other large front-end I4/I8/IA8 textures) render "interlaced" / combed (M-31, user bug RC1 / D149).** — full `## D159` entry at file tail | **FIXED** (`src/game/image.c`, `#ifdef PORT` early-`return` in `texSwapAltRowBytes` —… |

| D165 | **Front-end menu cursor doesn't feel like a mouse pointer (M-31, user bug B5 follow-up).** — full `## D165` entry at file tail | FIXED (`port/src/input.c`; `Input.MenuPointerMode`, `Input.HipfirePitchSpeed` added). Build green, `-level_09` + `GE_STARTMENU=6` smoke crash-free. Feel-check owed (headless can't exercise the mouse). |
| D166 | **Hipfire mouse pitch is digital while yaw is analog — inconsistent feel (user bug B7 / D118a, "cheaper alternative").** — full `## D166` entry at file tail | FIXED (`port/src/input.c`).… |
| D233 | **All levels (intermittent): indoor level geometry invisible near doors — you can see through walls/floors into rooms you can't see into, while world objects (terminals, furniture) in the room still draw (M-106, user QA report; screenshots `docs/img/bugs/20260912-cullingbug-bugged.png` vs `-notbugged.png`).** The geometry-vs-object split localises it to room/portal background visibility or frustum culling in fast3d, not textures/models. Logged in `GRAPHICS-BACKLOG.md`. **Duplicate row — see the full root-cause writeup above (`gfx_sp_tri1` w<0 trivial-reject bug, `port/fast3d/gfx_pc.cpp:1455-1460`).** | **FIXED + VISUALLY VERIFIED (M-107 fix, M-108 verify).** See the other D233 row for full detail. |
| D234 | **Dam: guard towers along the dam structure invisible except the first one (M-106, user QA report).** Repeated instances → per-instance transform decode, shared model/texture resource bound only by the first instance, or cull-volume issue. Same neighbourhood as D233. Logged in `GRAPHICS-BACKLOG.md`. **Duplicate row — see the full entry above for the M-107 investigation + M-108 verification.** | **VISUALLY VERIFIED FIXED (M-108), mechanism unexplained.** See the other D234 row for detail. |
| D235 | **Wrong textures load after dying/restarting a level — not Facility-specific, seen on multiple levels (M-106, user QA report; clarified same session).** Original report ("Facility: wrong textures now") read as a build regression; user clarified the actual trigger is **death/restart** (or a similar level-reload path), and it has recurred on other levels too — this is a state-leak on level re-entry, not a one-level texture-decode bug and likely not caused by #47/#49 at all. Prime suspect: the texture pool / CI-cache (D217/D228 neighbourhood) not being fully reset between a level teardown and its reload — a stale arena base, a leftover palette-cache entry keyed by an address that gets reused with different content, or a texpool index not rewound — surfacing as "wrong texture" only after the reload path runs, never on a fresh level boot. Logged in `GRAPHICS-BACKLOG.md`. | OPEN — needs: does it repro on a *fresh* level boot (rules out reload-state) vs only after death/restart (confirms it); which surfaces; a screenshot. **M-108 static-analysis pass: the leading "stale CI-cache not reset on restart" hypothesis is likely REFUTED, don't re-propose it as-is.** The theory was that `g_TexCacheItems`/`g_TexCacheCount` (`src/game/image.c:38-40`, cleared only in `set_mt_tex_alloc()`, `src/game/initmttex.c:8`) survives a death/restart because the restart path skips `set_mt_tex_alloc`. But tracing the actual call graph: `set_mt_tex_alloc()` runs inside `lvlStageLoad()` (`src/game/lv.c:369`), which has exactly **one** call site (`src/boss.c:485`) inside `bossMainloop`'s outer `while (!done)` loop; that loop re-runs on *every* stage transition, driven by `g_StageNum = g_MainStageNum` (`boss.c:693`) which is set by `bossSetLoadedStage()` (`boss.c:722`) — the same function the front-end's level-select/continue path calls (`front.c:6947`). Nothing in this call graph is level-select-specific vs. restart-specific; both look like they go through the identical full reload (including the `set_mt_tex_alloc` cache reset). Have not yet found the actual death/restart trigger's call site to be 100% certain it doesn't have its own bypass, so this isn't fully closed out — but the single-reload-path structure makes the "cache survives restart" mechanism unlikely as stated. Next: find the actual death-handler call site (search for where a life/continue/retry decision is made, likely `chr.c`/a game-state machine, not yet located) and confirm whether it calls `bossSetLoadedStage`/goes through `bossMainloop`'s reload path same as normal transitions, or has a distinct lighter-weight path — if it's the same path, the real D235 mechanism is still unknown and this CI-cache theory should be dropped. |
| D236 | **Surface 1: 2D billboard trees render as a wall/sheet of tree texture instead of discrete camera-facing tree sprites (M-106, user QA report).** Eight passes so far (D265, D266–D270, D271, D272, D273, D274, D275–D276). Ruled out: IA4/IA8 decode (as originally scoped), the blend pipeline, fog combine (as originally scoped), portal-culling over-draw (D271, real bug, fixed, not the cause), converter/UV-conversion bugs (D272, code-read clean), the CI palette bit-layout (D273/D274, tested twice against real content), "mips absent" (D274, confirmed present/enabled), "identical UV window" *for the CI8 classes* (D272). **Current leading picture (D276, M-134): the actual visible weave at confirmed weave pixels is painted by a different, previously-dismissed class — an IA8 blocky-noise texture (`oml=0xc81049d8`/`fmt=3`) sampled with a fixed, byte-identical 2×-wrapped UV window and low fog, reproducing D265's original "same window" observation exactly (D269's CI8-classes framing had reframed away from that lead prematurely).** A third, unattributed CI8 class (`0x0c184b50`) may be the reference's actual discrete-tree layer and shows an erratic z/w suggesting its own occlusion/billboard bug — D236 may split into two root causes. **Ninth pass (D280, M-135): fog math verified N64-faithful (±2 truncation over 2505 verts), the IA8 noise class visually confirmed as the weave's painter (HF 14.8→2.9 when collapsed), `0xc8102078` ruled out (terrain, 1.4M tris), and the depth-race read inverted — `0x0c184b50` tree cards are IN FRONT of the noise quad (z/w smaller=closer) yet still don't visibly win; under deterministic `-level_36` boot they barely instantiate at all (~4 tris vs 720 in a free run). No quick win: remaining work is CPU-side instantiation/room-streaming, not a port fix.** Distinct from D176(b)'s painted tree-line backdrop. Logged in `GRAPHICS-BACKLOG.md`. | OPEN — ninth pass closed with no fix; see D276/D280 |
| D237 | **QoL ask (M-106, user QA): F10 options overlay — make it scrollable with the mouse wheel, expose more settings into it (e.g. screen-flash / no-hit-flash `Game.NoHitFlash`, and other easily-exposed config keys), and organise rows by category (mouse settings together, video together, etc.).** The overlay (`port/src/optionsoverlay.c`) already renders over the registered `config.c` option list; what's missing is wheel-scroll nav, a broader row set, and grouping headers. Cross-ref `OPTIONS-MENU-PLAN.md` §3 (v1 minimum set) + QOL-INVENTORY "PD has, GE lacks". | OPEN — not designed; port-only |
| D238 | **QoL ask (M-106, user QA): mouse sensitivity rework toward normal PC-shooter feel.** Three parts: (1) a *separate* RMB aim-mode sensitivity in addition to hipfire turn; (2) ideally ONE single master mouse-sensitivity setting driving both, **linked by default**, with an option to set them independently later (today `Input.MouseAimSpeed` and `Input.MouseTurnSpeed` are two independent keys, defaults 16/100 — see `input.c:188-190`); (3) up/down should feel mostly the same as left/right (`Input.MouseYScale` exists but the axis asymmetry in the gain constants `AIM_GAIN`/`MOUSE_TURN_GAIN` + smoothing path is the likely culprit). User's reference for "feels better": GEPD Mouse Injector. Lump with the overall mouse/input fixes (D194/D165/D166 lineage; SOLITAIRE natural-pitch scheme switch may subsume part of this — see D166 M-87 note). | OPEN — not designed; port-only (`port/src/input.c` + `config.c`) |
| D239 | **QoL ask (M-106, user QA): click-to-lock vs always-grab parity.** User: make click-to-lock the default *if it isn't already* — it IS: `mouseCaptureMode = 1` is the shipped default (`input.c:185`). The real content: "always grab" feels bad on the file-select menu, not 1:1 with click-to-lock; if making them feel identical is overly complicated, **drop always-grab entirely and remove the setting**. This re-confirms the existing M-82 user decision to rip out `MouseCaptureMode=0` (see `docs/BACKLOG.md` "M-82 playtest feedback" + D192: mode 0's P-controller path can't reach the outer front-end grid cells — a known mode-0-only defect). Recommendation on this new report: take the user's fallback — remove mode 0 and the `Input.MouseCaptureMode` key, keep click-to-lock only. | CLOSED — executed in the v0.2.0 prep pass (port-only): `mouseCaptureMode` + the config key, `reconcileGrab`'s legacy branch, the `!haveAbs` relative menu path and its now-dead P-controller state (`menuEst/TgtH/V`, `menuCursorStep`, `MENU_P_GAIN`/`MENU_EST_DELTA`/`MENU_MOUSE_TO_PX`), the F10 "Mouse capture" row, and the unused `inputMouseCaptureActive()` all removed from `port/src/input.c` / `optionsoverlay.c` / `port/include/input.h`. Click-to-lock is the only mode; existing ini files setting the old key are ignored. Closes D192 with it |
| D240 | **Player gunshot SFX timing/cadence reportedly does not match the N64 original's rate (M-106, user QA report).** Static trace only (no ROM/build/display in this worktree). Weapon fire re-triggers its SFX through `bondwalkItemGetSoundTriggerRate()`/`bondwalkItemGetSound()`, gated on `g_GlobalTimer` (`src/game/gunfire.c:3246-3286`: `handptr->field_A4C = bondwalkItemGetSoundTriggerRate(var_s1) + g_GlobalTimer`), alternating between two per-hand `ALSoundState` slots (`audioHandle`/`field_A48`, same block) so a fast-firing weapon's report can overlap itself once before being retriggered. `g_GlobalTimer` advances by `g_ClockTimer = speedgraphframes` every tick (`src/game/lv.c:1023,1037`), and `speedgraphframes` is set from `deltaFrames` in `updateFrameCounters()` (`src/game/frametiming.c:64-76`) — the exact tick-delta plumbing already investigated (and partly fixed) under **D155** (catch-up-spiral clamp, same file), **D186** (`Video.FpsCap` sim-coupling clamp), **D193/D209** (AI-locomotion timing keystone — D209 the actual root cause, a 64-bit union-aliasing bug, now fixed and merged), and **D204** (audio pipeline running ~2% below real time forever, fixed by padding silence at the DAC). Since D209 and D204 both merged before this report, the SFX-cadence complaint is either (a) a genuinely distinct, audio-specific residual downstream of those fixes — e.g. a mismatch between the sim-tick-gated *trigger* (above) and the software mixer's own playback clock (`port/src/audio.c:93,155,221-222` — the synth produces samples at a measured ~21600/s against the 22050 Hz output device, the same gap D204 had to compensate for elsewhere) that specifically desyncs how *often* a shot's sample starts rather than the absolute wall-clock rate D204 measured, or (b) a stale pre-D209/pre-D204 build report. Per-weapon rate constants (`AutomaticFiringRate`/`SingleFiringRate`/`SoundTriggerRate`, `src/game/gun.h:60-82`, values in `assets/obseg/gun/*/gunWeaponStat.inc.c`) are plain compiled C literals, not ROM-serialized data, so no pointer-width/ABI reconciliation issue applies here. | **OPEN — not root-caused; static call-site tracing only, no runtime verification possible in this environment.** Confidence LOW-MEDIUM that this is a *new* bug distinct from the D155/D186/D193/D204/D209 timing-keystone family (all of which gate the identical `deltaFrames→speedgraphframes→g_ClockTimer→g_GlobalTimer` chain this SFX trigger consumes). Raise confidence by: (1) confirming the user's build/commit postdates the D209 merge; (2) a `GE_AUDIOTRACE`-style capture logging the tick (`g_GlobalTimer`) and/or wall-clock timestamp of consecutive same-weapon fire-SFX requests during sustained auto-fire (AK47/Klobb, `AutomaticFiringRate=0x03`), diffed against the weapon's rate constant and against N64 reference footage's frame-count between shots. |
| D241 | **Player gunshot SFX goes silent under some conditions — reported on Bunker with the alarm sounding, "or some other factor" — but only for PP7 and AK47, Klobb unaffected (M-106, user QA report, concrete repro).** This is very likely **D207 finally reproduced.** D207 (M-77) root-caused that GE's SFX player has a hard **8-voice cap** (`MUSIC_SFX_SEQ_MAYBE_MAX_SOUNDS`, `src/music.c:76`) and that D206's fix made the alarm klaxon (`ALARM3_SFX`=163, triggered from `handle_alarm_gas_timer_calldamage()`, `src/game/propobj.c:14457`) a genuine **infinite ADPCM loop** that now legitimately occupies one of those 8 voices for the alarm's whole duration; the manual preemption scan (`snd.c:~347-392`) refuses to steal any *owned* LOOPED voice (the M-65 carve-out only exempts *ownerless* loops), so once 8/8 voices are held, a new one-shot request is silently dropped via `sndDisposeSound` (`snd.c:391`). D207's own M-77b session built a `GE_FORCEALARM=1` BUNKER1 harness and could **not** reproduce audible drops (`[D207-DROP]=0`), and explicitly left as "owed": *"needs a real repro from the user: which level, which weapon/situation."* This report supplies exactly that (Bunker, alarm). **Investigated why PP7/AK47 and not Klobb, inconclusive at the code level:** Klobb (`ITEM_SKORPION` — confirmed GE's internal name for the Klobb via `src/game/player.c:576`, `skorpion_stats` in `assets/obseg/gun/skorpion/gunWeaponStat.inc.c`) and AK47 (`ak47_stats`, `assets/obseg/gun/ak47/gunWeaponStat.inc.c`) share the **same** `AutomaticFiringRate=0x03` (fast auto), while PP7 (`wppk_stats`, `assets/obseg/gun/wppk/gunWeaponStat.inc.c`) is semi-auto only (`AutomaticFiringRate=0xFF`) — so a naive "fires faster ⇒ more self-inflicted voice contention" theory does not cleanly explain grouping PP7 with AK47 against Klobb. All three weapons' fire SFX (`GUN_B1_MGUN3_3_SFX`=Klobb, `GUN_B2_HEAVY_SFX`=PP7, `GUN_B4_BOLTACTION_SFX`=AK47 — `src/bondconstants.h:2277-2280`) go through the identical one-shot, priority-0x40 `sndPlaySfx` path (`gunfire.c:3246-3286`); no per-weapon distinction exists in the trigger or priority *code*. A real distinction, if one exists, most likely lives in the **sample bank data itself** (per-sample envelope/decay length — a longer-tailed sample stays `PLAYING` and stealable/held longer, changing its odds in the preemption scan) — that data is ROM-derived binary content, not inspectable without the extracted ROM (`./data/`, absent in this worktree). Confirmed the voice-allocation/preemption code is unmodified `src/snd.c` game logic, not a `port/` mixer bug: `port/src/mixer.c`/`port/src/audio.c` implement DSP synthesis and device output only (per D199), with no SDL `Mix_*`-style channel-priority layer in the port to have introduced a new bug. Per AGENTS.md non-negotiable #2, any real fix belongs in D207's already-designed (reverted, M-77b) last-resort LOOPED-preemption pass in `snd.c`, not in `port/`. | **OPEN — not newly root-caused; high confidence this is D207 reproduced, not an independent bug.** Recommend folding this repro into D207 (or closing D241 as a duplicate once D207 is reopened) rather than tracking separately. Next: re-run M-77's `GE_FORCEALARM=1` harness on `-level_09` with scripted PP7 and AK47 fire (not just generic fire) against Klobb as a control, watching `g_sndAllocatedVoicesCount` / the (currently reverted) `[D207-DROP]` counters to see whether the 8-cap is actually hit for two of the three weapons and not the third — or whether it is really the sample-envelope-length distinction above. No runtime verification possible in this environment (no ROM/display). |
| D242 | **Player-death sequence UX regression vs N64 original: PC requires an LMB click to proceed to the mission-failed/file-select screen; user recalls N64 auto-plays 3 death-camera angles then transitions with no input needed (M-106, user QA report).** **The 3-angle auto-replay and its no-input auto-transition both exist, unmodified, in the decomp — confirmed, not a design difference removed for the port.** `bondviewSetCameraMode(CAMERAMODE_DEATH_CAM_SP)` (`src/game/bondview2.c:901-1010`) calls `pickDeathCameraAngles()` (`bondview2.c:641-782`, picks a fresh random valid camera angle each call) to stage each replay; when it returns 0 — no further valid angle found — the code calls `bossRunTitleStage()` directly, and the source comment at `bondview2.c:1006-1007` literally reads *"...or when the three death replays are finished."* The 3-cycle count is driven by `bondviewAdvanceCameraMode()` (`bondview2.c:1041-1078`): `CAMERAMODE_DEATH_CAM_SP` → `CAMERAMODE_DEATH_CAM_MP` → `camera_mode++`, looping back to `DEATH_CAM_SP` while `camera_mode < CAMERAMODE_SWIRL` (three iterations). Each individual replay's own advance is driven purely by animation/timer state (`bondview2.c:1467-1490`): the `g_CameraAfterCinema` flag that `bondviewAdvanceCameraMode()` acts on is set automatically once `modelGetAnimFrame(bodyModel) >= modelGetAnimEndFrame(bodyModel)`, or after a `camera_transition_timer >= 180.0f` tick-timeout fallback when there is no body model — **no button check in that branch; fully automatic.** **The button-press the user is hitting is a real, intentional N64 skip-ahead shortcut, not a port-added gate:** the same code block (`bondview2.c:1483-1489`) *also* fires the identical advance immediately on `(buttons & ~oldbuttons & (CONT_A | B_BUTTON | Z_TRIG | START_BUTTON))` whenever `bonddead && redbloodfinished && deathanimfinished` — i.e. on real N64 hardware, pressing A/B/Z/Start at any point during a death replay skips straight to the end, by design. **On the port, LMB (mouse fire) is mapped to `GE_CONT_G`, which is Z_TRIG** (`port/src/input.c:664-665`: `(mb & SDL_BUTTON(SDL_BUTTON_LEFT)) ... button \|= GE_CONT_G`; binding table `input.c:351`: `{"Z",GE_CONT_G}`) — exactly one of the four original skip buttons. So an LMB click during the death sequence is GE's own original cutscene-skip input, faithfully wired through the port's control remap, not a port-invented "click to continue" prompt. **What this investigation could not determine** (no build/runtime access in this worktree): whether the *automatic*, no-input path actually completes on PC in a reasonable time. `camera_transition_timer` accumulates via `g_GlobalTimerDelta` (`bondview2.c:1470,1493`), which is set from the same `g_ClockTimer`/`speedgraphframes`/`deltaFrames` chain flagged in D240 above and investigated under D155/D186/D193/D204/D209 — if any residual in that chain stalls or mis-scales specifically in this state, the auto-advance could take far longer on PC than on N64 (or effectively never complete before the player gives up and clicks), which would make an LMB click *feel* mandatory even though the ground-truth logic never requires one. | **NOT A PORT-LOGIC BUG as far as static analysis can tell.** Confidence HIGH that the LMB-click skip path is faithful original behavior (Z_TRIG-mapped, exists unmodified in `bondview2.c`), not a defect to remove. Confidence LOW (untested) on whether the no-click automatic path is actually healthy on PC. Needs a human playtest: die, then **deliberately touch no input** during the death sequence and time whether/how long until the mission-failed screen appears on its own. If it never arrives without a click, the real bug is a `g_GlobalTimerDelta`/`camera_transition_timer` stall in the D240 clock family — cross-reference D240 — not a wrongly-added input gate, so there is still no `#ifdef PORT` skip-gate to remove per AGENTS.md #2. No runtime verification possible in this environment (no ROM/display). |
| D243 | **Dam end-of-level cutscene (the abseil/bungee sequence, D148/D160's subject): the camera violently shakes/"spazzes" during each of the scripted camera-angle cuts (M-107, live `GE_D160=1` trace, two interactive `-level_33` runs + one fresh-save normal-progression run).** **First, corrects D160's own premise: the cutscene is NOT skipped on the current build.** Both debug-loaded runs show exactly the AI-script-scripted sequence firing: `bondviewSetCameraMode(7)` (`CAMERAMODE_POSEND`, the cutscene camera-lock state) called 2-3 times, matching `ai_17`'s three `camera_switch` commands in `UsetupdamZ.c`'s abseil ailist (`assets/obseg/setup/UsetupdamZ.c:2656`+), followed by `bossReturnTitleStage`. So D160's original (b)/(c) hypotheses (AI trigger never fires / puppet renders invisibly) are **refuted on this build** — the mechanism runs; what's broken is what it renders. **Root cause candidate, directly observed:** right in the camera-cut window, fast3d's D146 catch-all (`port/fast3d/gfx_pc.cpp:3037`, "unknown GBI opcode … ending DL") fired a tight burst — same 4 memory addresses repeating every frame for ~10 frames, then clean — with payload bytes `0xfafbfbfcfcfbfdfc`/`0xfcfafbfbfafbfafb`: a **freed/uninitialized-memory fill pattern**, not real GBI commands. This reads as a stale-pointer display-list read, not a decode bug — some segment/scene buffer backing the outgoing camera shot's geometry got freed or recycled while a still-in-flight display list (queued ahead in the port's double-buffered render pipeline — see D134's "pendingGfx" note, `port/src/libultra.c:733-746`, on how many frames of latency the sched round-trip can carry) still referenced it. **Independent evidence of a race, not a fixed-order bug:** the two `-level_33` debug-load runs show the camera_switch(7) calls landing in a *different order relative to* `bossReturnTitleStage` between runs (run 1: three switches, then the stage-return call; run 2: two switches, the stage-return call, *then* a third switch arriving after) — and the D146 corruption burst appeared in run 1 but not run 2 at all, non-deterministically. That pattern (symptom present/absent, call order shuffling) is the signature of a genuine timing race between the AI-list's cutscene commands and the level-exit/report-transition code, not a deterministic logic bug. **User-confirmed on real (non-debug-boot) play, not just the debug harness:** a fresh-save normal playthrough into Dam still hit the camera glitch on this same cutscene — ruling out "artifact of the `-level_33` debug boot" for this specific symptom (unlike the secondary "no report screen, straight to file-select" symptom seen only on the debug-loaded runs, which did NOT reproduce on the fresh-save run and is therefore most likely a `-level_N` debug-boot artifact per the D190 precedent, not filed as its own bug here). | **OPEN — M-109 re-analysis of the same session's raw logs corrects the timing/pattern claims above; the D134/pendingGfx race hypothesis is likely wrong.** Confidence MEDIUM on the correction, LOW-MEDIUM on the new lead. Re-reading `data/playtest_d160.log`/`_run2.log` at the `g_framesRendered` level (not just grepping for `D146`) shows the D146 burst fires **~1600-3900 frames (tens of real seconds) *after*** `bossReturnTitleStage`/the `camera_switch(7)` calls, not during them, in **both** runs — "right in the camera-cut window" and "run 1 only" are both wrong as stated above. This also weakens the D134/`pendingGfx` citation: `osSpTaskStartGo` runs `gfx_run()` synchronously and inline (`port/src/libultra.c:1232-1258`), so nothing from the cutscene's own DL submission is still in flight by the time any teardown code runs — the port doesn't have the async-RSP race the hypothesis above assumes. **New lead:** whatever screen renders in that multi-thousand-frame gap after mission-complete is the more likely site — probably `MENU_MISSION_COMPLETE` (`front.c:7267`, `constructor_menu0D_missioncomplete`) or an adjacent briefing/file-select screen in that chain, same general class as D75 (front-end model absence) / D141 (menu bad-pointer bug), not a cutscene-render race. Of the 4 repeating fault addresses, 3 read as literal zero words and only 1 matches the `0xfafbfbfc…` pattern, arguing against a deliberate poison-fill (none found anywhere in `port/`) and for either genuinely-unwritten memory or a decode landing inside unrelated valid data. Next: an env-gated `front.c` menu-transition trace (`MENU_*` id + `g_framesRendered` on every transition after `bossReturnTitleStage`) plus a `GE_PCDUMP` window over frames ~4500-8000 (not the camera-cut window) to name the exact screen/segment. Cross-ref D160 (this supersedes its stale "cutscene skipped" framing with a "cutscene renders corrupted" one, now further corrected to "corruption isn't in the cutscene render at all") and D173 (Cradle's Bond-floating/spinning cutscene bug — same subsystem family, unconfirmed shared root cause). **M-120: the `GE_D243` trace above was BUILT (statically, i.e. written) on branch `investigate/d243-cutscene-race` (`764a1f26`, pushed, not merged) — `src/game/front.c` only, `#ifdef PORT`, logs the target `MENU_*` name + a local per-tick counter on every `menu_init()` state transition. NOT build-verified: two independent sessions (the dispatched subagent, then this session re-verifying in a fresh worktree) both hit the same reproducible failure — `cmake`'s C-compiler ABI check fails with "Cannot create temporary file in C:\Windows\: Permission denied" when gcc is invoked *through* CMake/ninja, even though a bare `gcc -c` invocation from the same shell, same session, succeeds and correctly resolves its temp dir to `%TEMP%`. This is a genuine, reproducible environment issue distinct from the diff's content (static review of the diff found nothing wrong: enum cases all verified to exist in `bondconstants.h`, brace/`#ifdef` balance checked, follows the existing `GE_D160`/`GE_D178` env-gated trace idiom used elsewhere in the same file) — worth a session with a working toolchain investigating the CMake/ninja-vs-bare-gcc environment discrepancy itself as its own item, separate from D243. Next: build `investigate/d243-cutscene-race` on a session where `./build-pc.sh` actually completes, then run the `-level_33` Dam repro with `GE_D243=1` (and the existing `GE_D160`/`GE_PCDUMP` mechanisms) to correlate the trace against the D146 burst's frame range. **M-141: the M-120 toolchain blocker is EXPLAINED and worked around, GE_D243 is now build-verified.** The "Cannot create temporary file in C:\Windows\" failure isn't specific to CMake/ninja vs. bare `gcc -c` — it's this Claude Code session's `Bash` tool shell (Git-for-Windows Bash, not a real MSYS2 MINGW64 window) not exposing a usable `TMP`/`TEMP` to child processes the way an interactive MSYS2 shell does; a bare `gcc -c` from the same shell happened to not hit the temp-file path M-120 tested, which is why it looked clean while the full build wasn't. Fix: explicitly `export TMP=... TEMP=...` (Windows-style path) alongside the already-known `PATH=/c/msys64/mingw64/bin:/c/msys64/usr/bin:$PATH` fix, both before invoking `./build-pc.sh`/`ninja`/`cmake` — now also recorded in `porting-notes.md` §E as a general session-start gotcha, not just a D243-specific note. Cherry-picked the `GE_D243` trace commit (`764a1f26`) cleanly onto `release/v0.3.0` (branch `investigate/d243-fresh`), fixed one thing the original diff got wrong per this project's own D250 rule (`menu_init()` runs ~once/frame during front-end states — the diff's three `getenv("GE_D243")` calls were uncached; now a single cached `static int s_d243` check), and confirmed it builds clean. **Smoke-tested, not fully exercised:** a bare `-level_33` boot runs 1100+ frames with zero crashes and (correctly) zero `D243:` log lines — bare `-level_XX` skips the front end entirely (per this doc's own M-24-era note), so `menu_init()` never runs; this only proves the probe doesn't regress normal boot, not that it captures anything useful yet. **Still blocked on the same thing M-107/M-109/M-120 all were: an actual interactive playthrough** (main menu → Dam → abseil cutscene → mission-complete → whatever screen the D146 burst fires on, ~1600-3900 frames later) with both `GE_D243=1` and `GE_D160=1` set, which needs a human with input, not available in this session. **Possible connection to D173 worth checking in that same playthrough** (per user, M-141): D173 (floating/spinning Bond puppet in cutscenes, e.g. Cradle/Trevelyan) and D243 (stale-memory DL read some seconds after a cutscene-driven `bossReturnTitleStage`) both sit in the scripted-cutscene/level-transition neighborhood and both ultimately touch "Bond's (or a puppet's) in-world model" — no code-level link found this session (D173's own M-106 pass already ruled out the two most likely struct-aliasing sites, and D243's mechanism looks like a stale/freed buffer read in a *different* subsystem, front-end/menu memory rather than a chr/puppet transform), so treat this as an unconfirmed hypothesis to keep in mind while capturing the next live trace, not an established shared root cause. **Steam Deck playtest confirmation, same M-141 session** — user independently reported "cutscene at end of dam level is badly glitched" without prompting, matching this row's own "camera violently shakes" symptom exactly. Second independent confirmation this is real and reproducible, not a one-off; still needs the interactive `GE_D243=1`+`GE_D160=1` capture this row already calls for. |
| D245 | **IsWater water: two competing patterns with a visible seam that moves as the player looks around (mouse/aim) — "looks almost exactly like skybox" (M-116, user QA report, post-D229-fix build).** The symptom is D227's exact signature class (s16 `Vtx.tc` overflow → rasteriser interpolates between wrapped coordinates → two differently-wrong screen regions + a seam at the wrap boundary; on the sky quad the split ran on the 0–3 diagonal). Water draws through the same port path: `src/game/sky.c` water block (~905–956) — `texSelect(&skywaterimages[WaterImageId])` + `sub_GAME_7F09343C` (two-texunit cross-fade; both tiles re-declared RGBA/16b @ tmem 0, the draw D229 fixed) then a `skyRenderTri` fan over `sp274[]` under `skyPortBeginFan(sp274, s1, FALSE)` — and **the `FALSE` is a documented, deliberate tolerance of exactly this overflow**: the comment above `skyPortBeginFan` (sky.c ~1738) records that the IsWater quad is excluded from D227's tile-shift safety valve because both texunits read the same vertex tc, so a tc rescale would need matched shifts on BOTH tiles and "shifting both makes them sample identically, which flattens the cross-fade"; the 2026-09 decision was that "the water quad keeps the plain 1:1 tc and simply tolerates the overflow on its horizon vertex, exactly as it did before D227." The mechanism is real: `unk20/unk24` are S10.5 units reaching tens of thousands near the horizon; the fold rebases to a multiple of `SKY_TC_WRAP` (2048) but with k=0 an s16 tc holds only ±32767, so any wider span wraps. It was invisible before D229 because every water texel was green mottle garbage anyway; with correct blue texels the wrapped region now reads as a visibly different pattern + seam, and it tracks the view because the fan's S/T depend on camera orientation. (On N64 this draw is a verbatim RDP `G_TRI_SHADE_TXTR` coefficient block with no s16 tc — the overflow constraint is a port artifact of re-emitting as `gSPVertex` batches, same class as D227.) | **OPEN — REOPENED (M-116 live playtest, same session): the `allowShift=TRUE` mechanistic fix below did NOT resolve the symptom.** User report on Frigate post-fix: water still "glitches, kinda goes back and forth between two graphics too quickly," and still "changes depending on your view unlike the sky" (i.e. the sky, using the identical `skyPortBeginFan`/`skyPortEmitTileShift` mechanism post-D227, no longer has this problem, but water still does after getting the same treatment). This means the `allowShift=FALSE→TRUE` change either isn't the actual fix, or there's a second mechanism specific to water's two-texunit cross-fade that the tc-shift alone doesn't address — most likely candidate: the shift only rescales the *shared* vertex tc uniformly, but the tile-0-vs-tile-1 sample-point difference is set by each tile's own `uls`/`ult` (RDP tile-space, not vertex-tc-space) — if per-frame `k` changes granularity of the *quantized* shared coordinate before each tile's independent `uls`/`ult` subtraction, the two tiles' effective sampled texel could drift relative to each other frame-to-frame in a way the static `GE_D227V` trace (single frame, single camera angle) never exercised. Needs a fresh live investigation, not a repeat of the same static-trace approach — a scripted or live multi-frame capture while rotating view, watching both tiles' *decoded* texel positions (not just vertex tc) frame over frame. Do not re-apply the same `allowShift=TRUE` reasoning without new evidence; the superseded mechanism write-up below is unconfirmed-by-playtest and may be wrong. The code change (`allowShift=TRUE` on the water quad) is left in place — it's not shown to be actively harmful, and reverting it isn't obviously better than leaving it, pending the fresh investigation. **Superseded analysis (unconfirmed by playtest, kept for reference):** ~~FIXED (port-layer, `src/game/sky.c`, `#ifdef PORT`) — confirmed the `allowShift=FALSE` tolerance was the cause, and the "flattens the cross-fade" objection was wrong.~~ Changed the water quad's call from `skyPortBeginFan(sp274, s1, FALSE)` to `..., TRUE)`, so it now gets D227's tc-shift safety valve like every other sky fan. Why the old objection doesn't hold: `skyPortEmitTileShift` replays **both** tiles' captured `SETTILE` commands with the identical `k`, so both texunits' decoded S/T are rescaled by the same `2^k` *before* each tile's own (unshifted, tile-native) `uls`/`ult` offset is applied by the RDP — the offset between tile 0 and tile 1's sample points (what drives the cross-fade) is preserved, not flattened, because the shift and the offset act at different stages. **Verified mechanistically:** `GE_D227V=1` on Frigate (`-level_26`) shows a water fan vertex spanning S=-166..299317, T=-6451..288410 in S10.5 units — tens of thousands of texels, the same order of magnitude as D227's pre-fix sky overflow — now getting `kS=4 kT=4` (previously forced to 0), decoding to small in-range tc values (100s, not wrapped). **Verified no regression:** `verify.sh frigate` and `verify.sh "surface 2"` (the two `IsWater` levels) both PASS crash-free post-fix; non-water sky levels (`statue`/`-level_22`) unaffected since only the water call site changed. Screenshot comparison at a fixed scripted-intro camera angle (frame 500, both builds) showed no regression, but that specific angle didn't happen to frame the overflowing part of the fan either before or after — the sinf()-driven cross-fade also makes frame-for-frame screenshot diffs across separate runs noisy (each run's shimmer phase differs), so a same-frame pixel diff wasn't a clean confirmation tool here; the `GE_D227V` trace is the real evidence — **but it turned out not to be evidence the fix actually worked in live play.** Cross-ref D227 (mechanism), D229 (made this visible). |
| D246 | **Tiny pixel strips at the left and right edges of the screen, "like overscan" — look like raw pixels, not 3D graphics (M-116, user QA report).** Uninvestigated. The port renders at native 640×480 (`GE_NATIVE_W/H`; `gfx_current_native_aspect` set in `port/src/video.c:253,570`) and scales to the window on present, so the leading candidate is a scaling/blit rounding artifact: non-integer scale factors leaving 1–2 columns of uncleared backbuffer content at the edges (or an off-by-one in the blit rect). Secondary candidates: an RDP fill-rect/scissor off-by-one in a game DL or the F10 overlay hook (`gfx_run()` seam, D184), or GL edge filtering on the scaled texture. | **OPEN — one candidate ruled out (M-116 follow-up), root cause not found.** Step (1) from the original plan is done: a `GE_PCDUMP` capture (Frigate, native 640×483 buffer, both against the existing save and a fresh EEPROM) shows clean, continuous scene content in the leftmost/rightmost columns — no dead/black/garbage strip at either edge. That means the defect is **not** baked into the native render itself (rules out an RDP fill-rect/scissor off-by-one in a game DL, and rules out D227/D245-style tc issues) — it must be introduced later, specifically in the SDL-window presentation path that `GE_PCDUMP` bypasses entirely (`gfx_run()` reads back the game's own render target directly, never the actual window backbuffer the user sees). Re-examined `gfx_adjust_viewport_or_scissor` (`gfx_pc.cpp:2174`): its floor/ceil rounding is already deliberately anti-gap (comment: "prevents sub-pixel gaps at viewport edges"), so a plain rounding bug there is a weaker lead than it looked; more likely candidates now are the `game_renders_to_framebuffer`-path MSAA resolve `glBlitFramebuffer` (`gfx_opengl.cpp:1252`, only exercised when `Video.MSAA>1` or the window size actually differs from the game's target size mid-frame — a path `GE_PCDUMP`'s headless capture doesn't exercise the same way a live resizable window does) or a genuine sub-pixel/DPI scaling remainder specific to the user's actual window dimensions. **Next, still needs the user (no live window access in this environment):** (1) confirm current `Video.MSAA` setting and window size/aspect at the time the strips were seen; (2) if `MSAA>1`, retest with `Video.MSAA=1` to see if the strips disappear (isolates the resolve-blit path); (3) report whether the strip width changes with window size (fixed 1px = blit-rect off-by-one; scales with window = scale-factor remainder). Cross-ref D247 (separate, now closed — don't conflate the two; D247's letterbox is confirmed-intended and top/bottom only, D246 is left/right and still unexplained). **M-141: Steam Deck playtest confirmation, new platform data point** — user reports "some artifacts left and right of screen, only a sliver... like the fullscreen isn't quite going to full screen" specifically in **Full Screen mode on Steam Deck**. Same symptom description as the original M-116 report (Windows). Checked: the Deck's own first-run preset (`videoApplySteamOSDefaults()`, `port/src/video.c:255`) sets `cfgMSAA = 4` by default — exactly the condition this row already flagged as the leading remaining suspect (the MSAA-resolve `glBlitFramebuffer` path, step 2 of the still-open bisection ask). Next: have the user set `Video.MSAA=1` via F10 and retest on the Deck specifically — if the strips vanish, this row's root cause is confirmed without needing any more code tracing. |
| D247 | **Black bars top and bottom of screen — possible aspect-ratio defect (M-116, user QA report).** Uninvestigated. Context: the port currently renders native 4:3 (640×480); widescreen is a *parked proposal* (`docs/dev/WIDESCREEN-FOV-PLAN.md` — PD-parity native 16:9 render, needs rule-#2 sign-off for its game-code HUD wave), so if the user's window/display is wider than 4:3 and the present path letterboxes to preserve aspect, the bars are current intended behaviour rather than a bug. If they appear in a 4:3 window too, it is a real defect (framebuffer sized wrong vs window, `videoCreateFramebuffer(w,h,upscale,autoresize)` → `gfx_create_framebuffer`). | **NOT A BUG — confirmed faithful, unmodified N64 decomp behaviour, no code change made.** `GE_PCDUMP` captures (both against the pre-existing save and a fresh-EEPROM run, to rule out a stale `Screen` option from earlier test sessions) show the exact same inset on Frigate: real scene content occupies rows ~20–459 of a 640×483 capture (the 483 vs 480 is a separate, harmless 3-row capture-tool rounding quirk, not investigated further), i.e. roughly 220/240 of the vertical field — and `src/game/fr.h:34` defines `VIEWPORT_HEIGHT_DEFAULT_NTSC 220` (240 for the full field), consumed unmodified by `bondviewGetCurrentPlayerViewportHeight()` (`bondview2.c:7858`) as the **default in-game viewport height whenever `PLAYER_OPTION_SCREEN` is `SCREEN_SIZE_FULLSCREEN`** (the decomp's own default value, `options.c:128`) — `VIEWPORT_HEIGHT_FULLSCREEN` (304, no inset) is a red herring name; it's only reachable via the `cameraBufferToggle` branch, not normal solo play. This is GoldenEye 007's well-documented original TV-safe-area letterboxing, baked into ground-truth decomp constants with no `#ifdef PORT` anywhere near it — the N64 game genuinely doesn't fill the full field during normal gameplay. Ruled out a stale/leftover EEPROM `Screen=Wide/Cinema` setting from earlier QoL-testing sessions as an alternative explanation: a fresh EEPROM (default options) shows the identical inset. **No fix needed; this row closes.** If the user's window shows noticeably *thicker* bars than this ~8–9%-per-edge margin, or bars appear on the *sides* instead of top/bottom, that would be a genuinely different (and real) symptom worth reopening under a new label — but "thin bars top and bottom, present at any window size" is correct, original behaviour. Cross-ref D226 (separate, already-tracked HUD-scale QoL ask — unrelated mechanism, don't conflate). |
| D248 | **Port runs a stable 30 FPS instead of 60 FPS game-wide (M-117, user asked to verify FPS after the D209 timing-keystone fixes).** The port's pacemaker (`port/src/libultra.c` `portTickThread`) posts a VI-retrace event at a true 60Hz (NTSC) cadence, so the halving had to be somewhere downstream of that. | **FIXED, root-caused to a genuine pointer-width ABI bug in `src/sched.c` (compiled for the PC build, not excluded).** `osScAddClient()` (`src/sched.c:187-199`) stashes a per-client "present every retrace vs. every other retrace" flag by writing into the *next array slot's* `.next` field (`c[1].next = next`) rather than adding a real struct member — `gfxClient` (`src/init.c:246`) is registered with `next=NULL` (send every retrace → 60Hz) and the audio client `g_AudioClient[0]` (`src/audi.c:441`) with a nonzero sentinel `1` (send every other retrace → 30Hz, its intended cadence). `__scHandleRetrace` (`src/sched.c:334`) reads that flag back with a hardcoded byte offset, `*((s32*)client + 2)` (2×`sizeof(s32)` = 8 bytes). On the N64 (32-bit pointers) `sizeof(OSScClient)` is exactly 8 bytes, so offset 8 correctly lands on `client[1].next` — matching the write. On this 64-bit port `sizeof(OSScClient)` is 16 bytes, so the same hardcoded 8-byte offset instead reads the low 32 bits of `client->msgQ` itself (always a real, non-NULL pointer) — the gfx client's flag reads as permanently "nonzero" and silently falls onto the audio client's every-other-frame path, halving the game's real render/tick cadence to 30Hz although the retrace pacemaker itself was always ticking at the correct 60Hz. Same D122/D126/D132/D209 bug class (32→64-bit pointer-width struct-size growth desyncing a hand-rolled offset), just in the scheduler instead of a ROM-serialized struct. **Fix (`#ifdef PORT`, N64 line kept verbatim under `#else`):** read the real field by indexing the struct (`client[1].next == 0`) instead of hardcoding a byte offset — this naturally scales with pointer width on either platform and is the direct decompilation of what `osScAddClient`'s write side already does. **Verified:** `Video.DisplayFPS=1` overlay (`port/src/optionsoverlay.c`, D213) read a rock-stable "30 FPS" on `-level_09` (BUNKER1) pre-fix and "60 FPS" post-fix, same build/scene, screenshots captured via `GE_PCDUMP`. Rebuilt clean (`ninja -j1`, pre-existing unrelated warnings only); `-level_09` and `-level_20` both ran 12-30s post-fix with no crash (`ge007.crash.log` absent). No game-code files touched beyond this one mechanical read (`src/sched.c` is compiled for the PC target per `CMakeLists.txt`'s `SRC_ENGINE` list, not one of the excluded os/io files) and the audio client's own flag/path is completely untouched, so `#ifdef PORT`-gated audio pacing (D204 tempo work) is unaffected. **Note for `docs/dev/UNLOCKED-FPS-PLAN.md`:** this settles that plan's Phase 0 open question — the ~30fps default was a real port bug, not original N64 behavior or an artifact of retrace-rate tracking sim speed; the plan's default-preset baseline should be re-measured at the corrected 60fps before Phase 1 pacing work resumes. Not yet done: a full `verify.sh sweep` across all 21 levels (only `-level_09`/`-level_20` spot-checked this session) and a live human playtest for feel (input latency, animation smoothness, any level-specific perf regression now that render work happens twice as often per second). |
| D249 | **LOD/culling breaks down on large open levels (Streets, Egyptian named specifically) at *default* FOV — far-away geometry has visible culling issues (M-119, user QA report).** Distinct from **D222** (which is specifically high `Video.FovScale` breaking *near-screen-edge* culling on NPCs/objects — not implicated here, default FOV) and from **D218** (the fog/far-clip *tint* boundary at wide FOV — also not the same symptom). This report is at the stock FOV, on the game's two biggest/most-open outdoor-scale levels, affecting *far* geometry specifically — points at either (a) a portal/room-visibility BFS cutting off too aggressively at range on large multi-room layouts (same neighbourhood as D233/D234's near-camera portal-cull bugs — worth checking whether D233's `gfx_sp_tri1` w<0 trivial-reject fix, which only guarded *behind-camera* vertices, has a sibling far-plane/depth-range case), or (b) GE's own N64 LOD/detail-swap system (mip-tile / simple-model substitution at distance — see D218's still-open question of "does GE have a separate geometry/object LOD-swap distinct from `Visibility.FarFog`") not being ported faithfully, popping or dropping geometry instead of swapping detail level. **Second, separate ask bundled in the same report:** the user would like LOD **improved** in general (presumably distance-based detail transitions look rougher/more abrupt on PC than remembered) but explicitly wants this **gated behind a toggle** so the original N64-accurate LOD behaviour stays selectable ("in case you want old school graphics") — i.e. this is a QoL feature request layered on top of a correctness bug, not one fix. | OPEN — not investigated. Next: reproduce on Streets (`-level_20`) and Egyptian (`-level_08`, confirm level id) at default `Video.FovScale`; `GE_PCDUMP` a long sightline where geometry visibly drops/pops, cross-referencing whatever governs GE's LOD-tile selection (`gfx_lod_tile_offset` family, already touched by D172/D219) and the portal/room-visibility walk (D233/D234 neighbourhood) to determine which of (a)/(b) is actually firing before designing a fix or the requested toggle. |
| D250 | **FPS is inconsistent/low on many levels since the port started actually running at 60fps — drops to 20–40fps intermittently, described as worse than before (M-119, user QA report, follow-on from D248).** Very likely a direct consequence of **D248** shipping: the port's render+tick loop had been silently running at half rate (30Hz) for the entire life of the project (D248, M-118) — the *real* per-level CPU/GPU cost of rendering and simulating this port was always this high, it was simply invisible while D248 halved the work being asked of the hardware every second. D248's own write-up flagged this as an open risk ("any level-specific perf regression now that render work happens twice as often per second") and explicitly left a live playtest / perf pass as un-done. Not yet investigated which side (CPU sim, CPU fast3d command generation, or GPU draw) is the bottleneck, or whether it's uniform or level-specific (the user's report implies specific/intermittent, which fits a per-scene cost — draw-call count, overdraw, or a specific expensive subsystem — rather than a flat doubled baseline cost). **M-120: partial progress, real bottleneck not yet caught.** User report sharpened: ~45fps at 720p on a self-described high-end PC, present even outside of dedicated stress levels — worse than the coarse "many levels, intermittent" framing suggested. Two things ruled out this session: (1) **build flags are not the cause** — `build-pc/CMakeCache.txt` confirms `CMAKE_BUILD_TYPE=RelWithDebInfo` (`-O2 -DNDEBUG`) is the actual default (`CMakeLists.txt:75-76`, unconditional unless a cache already exists), matching what `./build-pc.sh` produces with no extra flags — a debug/`-O0` build is not the explanation. (2) **idle-load is not the bottleneck, at least not in this dev box:** `GE_D193=1` on `-level_09` (BUNKER1, small level) idles at a rock-steady `render=60 sim=60`, frame-render time ~2.5 µs·10^3 (2.5 ms); `-level_29` (STREETS, one of D249's flagged large/open levels) idles at `render=55-60 sim=60-62`, frame-render time 6-7 ms — both comfortably under the 16.7 ms/60fps budget with no catch-up clamp hits. **This does not reproduce the user's 45fps** — either the real cost only shows up under actual gameplay load (AI/combat/particles active, not an empty idle scene — matches D193's own "player fine, AI-driven cost is different" pattern) or it's resolution/GPU-fill-rate-bound in a way this box's small idle test window doesn't exercise (this session could not drive real keyboard/mouse input into the game to generate a loaded scene — no interactive input tooling available headlessly). Also spotted and ruled out in passing: the `D51 vi post` debug log firing every render frame in `port/src/libultra.c:781` is throttled to ~once/sec (`% 60 == 1`), not a per-frame cost — not the culprit, though it's a `TEMP D51` instrument that's been live since early porting and could be removed as cleanup. **M-120 continued, live session with the user driving the game directly — the sim/render split is now confirmed live, and several strong candidates are ruled out, but the actual cost is still unlocated.** Reproduced directly on Dam (`-level_33`): `sim` holds a healthy 59-67/s throughout, `render` sags to 45-50/s, and the CPU-side `frame rendered in Nus` cost (`gfx_run`'s software-RSP interpretation, `port/src/libultra.c:1252`) stays only 6-10ms the whole time — comfortably under the 16.7ms/60fps budget. That leaves a genuine, unexplained ~10-15ms/frame, and a live A/B pass ruled out every obvious GPU-side suspect one at a time, each via the F10 in-game overlay while watching `GE_D193`: **MSAA 2→1 and Anisotropy 16→1 — no change.** **FovScale 125→100 — no change.** **VSync 1→0 (true uncapped rendering) — render still pinned at ~60 max, never higher, and the bad-scene numbers were unchanged** (also tells us the ~60fps ceiling itself is not a vsync artifact — something else already caps the loop at 60, vsync or not). **Resolution 1280×960 fullscreen → 640×480 windowed — no change** (rules out GPU fill-rate/pixel-count entirely; the cost does not scale with pixels at all). **Found and fixed a real, separate bug in passing (did not turn out to be the active cause here):** `sync_framerate_with_timer()` (`port/fast3d/gfx_sdl2.cpp`) computes the remaining frame budget in 100ns units (`qpc_to_100ns`) and passes it straight to `sysSleep()`, which takes **microseconds** — a genuine 10x oversleep whenever that branch fires. Fixed (`left / 10`). But this path only runs when `target_fps` is nonzero (tied to `Video.FpsCap`), and the user's test config had `FpsCap=0`/`VSync=0` — confirmed live that the bug fix made no difference to the reported drops, because the buggy code path was never executing in that config. Still worth landing as a correctness fix for whenever `FpsCap`/`VSync` *is* active. **Added `GE_D250`** (`src/boss.c`, wraps `lvlRender()`, the DL-building/culling/scene-traversal step that runs *before* `gfx_run()` and is separate from it) to check whether DL-building was hiding the missing time — live-tested on Dam: **`lvlRender` averages only 0.5-1.5ms, also not the bottleneck.** So: known costs (`gfx_run` ~6-10ms + `lvlRender` ~1ms) sum to ~7-11ms, real frame period is ~20-22ms (45-50fps) — the remaining ~10-15ms/frame is resolution-independent, GPU-setting-independent, and not in either of the two obvious CPU call sites. Traced the dataflow one level further: the game's main loop (`src/boss.c` `bossMainloop`) blocks on `osRecvMesg(&gfxFrameMsgQ, ..., OS_MESG_BLOCK)` waiting for `OS_SC_RETRACE_MSG`, posted by a dedicated pacemaker thread (`portTickThread`, `port/src/libultra.c`) at a fixed cadence independent of how fast the consumer drains it — meaning the observed render rate is whatever the consumer (game+render loop, one OS thread hop away from the pacemaker) can actually sustain, not a directly-configured cap. **Leading remaining suspect: the emulated N64 inter-"thread" message-passing/scheduler round-trip itself** (`port/src/gesched.c`/`sched.c`, real OS threads + mutex/condvar standing in for the N64's cooperative SP/DP/main threads) — each hop (game thread → scheduler → SP "thread" → done-message back) is a real OS context-switch-and-wake, and Windows condvar/thread wake latency can easily run several ms per hop, unlike the free N64 hardware handoff this code was ported from. **Not yet confirmed** — this needs either a targeted timer bracketing the full scheduler round-trip (same technique as `GE_D250`, just at a different call site) or a real sampling profiler (Visual Studio's own, or Windows Performance Recorder/WPR) attached to `ge007.x86_64.exe` during a live bad-fps moment, which would show this directly without more printf-probe guessing. Ad-hoc per-call timers have now ruled out the two most obvious candidates without finding the cost; a real profiler is the better next tool. **M-120 — ROOT-CAUSED AND FIXED: `getenv()`.** Set up a real profiler as recommended (`tools_pc/perf.ps1`, new sister script to `debug.ps1`, wraps `xperf` CPU-sampling + context-switch tracing). A live WPR/xperf capture on Dam, with Microsoft's public symbols resolved (`-symbols`, downloads system-DLL PDBs), showed the dominant render thread spending only **5.2% of its CPU time in the game's own code** — **43% was in `msvcrt.dll`**, and within that, `getenv()`/`_getenv_helper_nolock` alone accounted for **~50% of the thread's entire CPU time**. Root cause: `port/fast3d/gfx_pc.cpp`'s `import_texture()` (runs on every texture bind) and several per-triangle/per-`G_SETCOMBINE` combine-mode probe checks (`GE_DTEX`, `GE_TEXDUMP`, `GE_TEXRAW`, `GE_D229` ×2, `GE_D172` ×2, plus `GE_TEXDUMP`/`GE_PCDUMP` in `gfx_opengl.cpp`) called `getenv()` **uncached, every single call** — unlike the safe `static int x = -1; if (x < 0) x = getenv(...) != NULL;` pattern already used everywhere else in this codebase (`ft_d193`, `ge_d116`, `dvm`, etc.). On a texture-heavy scene this is tens of thousands of Windows CRT environment-lookup calls per second (each a locked, locale-aware linear scan per Microsoft's msvcrt implementation), completely invisible to both of M-120's earlier timers (`gfx_run`'s "frame rendered in Nus" and the new `GE_D250` `lvlRender` probe) because neither one specifically isolated CRT/library time from the game's own code. **Fix:** cached every uncached `getenv()` site in `gfx_pc.cpp`/`gfx_opengl.cpp` with the established pattern — narrow, mechanical, zero logic change. **Verified:** the exact idle-Dam headless test that was pinned at 45-50fps before the fix now holds a rock-steady 60-61fps, zero dips, with no other change. This also retroactively explains why M-120's earlier resolution/MSAA/vsync A/B pass (all "no change") was so consistent — the `getenv()` cost was a large, *per-call* (not per-pixel) constant that swamped any smaller resolution-dependent GPU cost at every setting tried, so the A/B genuinely couldn't have detected a fill-rate difference underneath it. **User re-tested 1440p/4K after the fix and confirmed it's fine too** — the initial "still runs badly at 1440p+" report was from before this fix landed (a stale-build mixup, same class as the DLL-load-error false alarm earlier in this session); once actually re-tested on the fixed build, high resolutions hold up fine. **D250 fully resolved, all resolutions tested clean.** |
| D251 | **F10 options overlay QoL pass: (a) bottom rows drawn off-screen in the in-game 2D space — the layout assumed a ~320×240 field and hard-fit all rows, but `viSetXY` gives 440×330 on front-end screens (`src/game/front.c:8618`) vs 320×240 in-game (`front.c:8628`), so with 17 rows × 15u pitch + 2 header rows the last rows land at y≈282, off-screen whenever the panel is opened in-level; (b) MSAA row only cycled forward — `rowAdjust()`'s `ROW_MSAA` case clamped its index to [0,3] instead of wrapping, so left-click stuck it at 8x with no way down via clicks; (c) manual % rows (`Video.DrawDistance`/`Video.LodDistance`) showed unconditionally even while their `*AutoFov` toggles were on, where the percent is inert/misleading. | **FIXED (port-only, `port/src/optionsoverlay.c`).** (a) The panel now computes how many rows fit the *actual* current 2D height (`maxVisibleRows()` from `viGetY()`) and scrolls: a visible-row window (`s_scroll`) follows the selection, driven by wheel (`optionsOverlayScroll`) and arrow/D-pad at the list edges; layout, mouse hit-testing (`overlayRowAtY` now returns a visible position) and keyboard nav all index the visible list (`s_visIdx`/`s_visN`, `s_sel` is now a visible-list index), so it fits both 240- and 330-tall spaces with no per-space tuning. (b) `ROW_MSAA` wraps `(idx + dir + 4) % 4` — OFF→2x→4x→8x→OFF in both directions, matching the already-wrapping `ROW_ENUM`/`ROW_RES`; sliders keep clamping (standard for drag-able values). (c) New `Row.hiddenIfOn`/`hidePtr` fields: a row is hidden while the named option is nonzero — `Draw distance %` hides under `Video.DrawDistanceAutoFov`, `LOD distance %` under `Video.LodDistanceAutoFov`; rebuilt every frame in `overlayUpdateVisible()` (selection clamped into range when its row vanishes). Hint line updated to "click / drag / arrows / wheel". **Verified:** compiles clean in the full 241-target build; final link of this session was blocked only by a running game instance holding the exe (re-link pending). Not yet eyeballed live: scroll feel at 320×240, hover/scroll interaction, and the auto/% row appear-vanish transitions. Cross-ref D213 (FPS readout in the same file), D218/D249 (the auto-coupling these rows expose). **Follow-up after live test:** (d) list navigation (wheel + arrows) wrapped at both ends, so scrolling past the bottom made the highlight "repeat" from the top — now clamped at both edges like a normal PC settings list; (e) new `Input.SensLink` toggle (default ON, registered in `port/src/input.c`) links the aim/turn sensitivity knobs: moving either rescales the other to hold the stock 38 (`Input.AimModeSens`) : 50 (`Input.MouseTurnSpeed`) ratio, via a re-entrancy-guarded pair-write in `rowSet()`; (f) the "Mouse aim speed" row now edits `Input.AimModeSens` (the knob the default GEPD aim path actually uses — it previously edited `Input.MouseAimSpeed`, which only feeds the legacy velocity-stick fallback and was inert under default settings). **Second follow-up:** hover-to-highlight removed entirely -- moving the mouse over the panel no longer moves the selection or scrolls; near a list edge the hover->select->scroll-shift->new-row-under-cursor feedback loop made the bottom of the list twitch/echo (the persistent "duplicate row" report). Selection now moves only by explicit click (which also selects the clicked row), wheel, or arrows. **Reassessment of the 4K-fullscreen ghost-row bug:** the presentation-UV-wrap theory is DEAD — `gfxFramebuffer` has zero readers (the textured-quad present never happens); presentation is a full-frame `glBlitFramebuffer` from the same-size MSAA FBO to the default framebuffer (fb index 0 is reserved for the window, `gfx_opengl_init`), which has no UVs and cannot wrap. The overlay DL draws each visible row exactly once, in-bounds, with its own scissor. Local repro (maximized 16:9 window + MSAA=4 + `GE_OPTIONSOVERLAY=1` + `GE_PCDUMP`) shows NO ghost — the FBO/MSAA path alone is not the trigger; the remaining variable is true fullscreen on the user's 3072×1728 display (driver/display-level suspect). Next: user bisection — Fullscreen ON + MSAA OFF (restart) → does the ghost persist? **User follow-up (logged, not fixed):** the 4K-fullscreen ghost is *intermittent* across launches, and the user suspects the `hiddenIfOn` appear/vanish transitions of the Draw/LOD distance % rows (visible-list resize while the panel is open) — unverified; note `overlayUpdateVisible()`+`overlayUpdateScroll()` are re-run together in both emit and input each frame, so a stale-state window would have to come from elsewhere (e.g. the one frame a row count changes). MSAA-off bisection still pending. User: in the sessions seen so far the ghost is present on *first open* of F10 (not only after an auto-toggle flip) — weakens but does not kill the appear/vanish theory (toggles had flipped in earlier sessions before a restart). **v0.2.0-pre playtest (user, latest build):** the symptom has changed — the panel's *bottom row now deterministically duplicates whatever item is currently selected* (e.g. MSAA appears both on its own row and again at the bottom), not just an intermittent 4K/MSAA ghost of the top row; suggests the visible-row window can contain the selected row twice (off-by-one in `overlayUpdateScroll`/`maxVisibleRows` bookkeeping) — investigate post-release. **QoL follow-up:** make the auto-linked % rows visually subordinate to their toggle (indent and/or dimmer value colour) so the pairing is obvious in the panel; hint line also reworded to "select: scroll/click, change: mouse/arrows". |
| D252 | **Particle effects cycle through a rainbow palette (M-124, user playtest of the v0.2.0-pre candidate).** Shooting repeatedly makes the bullet-impact "sparks" slowly progress through the full hue wheel (~10 shots to sweep it); other lingering particle effects — smoke / explosion residue — show the same rainbow colouration instead of their intended grey/orange palette. **Mechanism context (verified this session):** particle tris carry no per-primitive material; their colour comes entirely from the ROM-resident state records `globalDL_0x078 … globalDL_0xa50` (`assets/oddtextures.c`, 15 of them) replayed per particle batch (`explosion.c:894`). Each record binds a **distinct** texture pair — `IMAGE_SMOKE_N` (IA, tile 0 @ TMEM 0) × `IMAGE_FIRE_N` (RGBA16, tile 1 @ TMEM 0x188) under `G_CC_INTERFERENCE` (TEXEL0·TEXEL1), and `explosionRenderPart` uses a fixed all-white default vertex (`explosion.c:215`) — so the intended per-stage colour is whatever those 15 fire/smoke pairs encode, and particles **age through the records** (`var_s2` 14→0 at `frame/flareanimspeed`), so a per-texture decode error surfaces exactly as a hue sweep over time / across shots. **Ruled out this session:** (a) the M-82/D172 tile-offset fix is correctly configured for this path — `gsDPSetTextureLOD(G_TL_TILE)` does NOT set the `G_TL_LOD` bit, so `rdp.tex_lod` stays false (`gfx_pc.cpp:2860`) and `gfx_lod_tile_offset` returns `i`, i.e. TEXEL1 genuinely samples tile 1; (b) the RGBA16/5551 texel decode in `import_texture_rgba16` is correct (big-endian, R=bits15-11 G=10-6 B=5-1 A=bit0). **Prime suspect: the width/height derivation in `import_texture_rgba16`** — `width = rdp.texture_tile[tile].line_size_bytes / 2`, `height = size_bytes / line_size_bytes`. The fire tiles are 16 px wide with a MASK_16 TMEM line (64 bytes), so `width` computes 32 (true: 16) and, with `size_bytes` = the LOADBLOCK's 16×14×2 = 448, `height` computes 7 (true: 14) — the GL texture would be built from the right pixels at the wrong dimensions, scrambling each of the 15 fire sprites into a different wrong-hued smear; that fits "slow rainbow progression" precisely (each aging stage = a different scrambled sprite). **Not yet confirmed** — needs one imported fire tile dumped and compared against the raw `IMAGE_FIRE_N` bytes (`GE_TEXDUMP`/`GE_PCDUMP`, or a targeted probe at the `upload_texture` call). **Per user instruction: documented only, NOT fixed for v0.2.0-pre** — ships as a known issue; pick up afterwards (cheap if the suspect holds: derive width/height from the LOADBLOCK's dtx/dlh rather than the tile line size, or special-case non-32-texel-wide RGBA16 tiles). D172/M-82 lineage. **M-141 update: the width/height suspect is REFUTED by the authored asset data, not just re-tested — the prior "prime suspect" reasoning assumed `line_size_bytes` derives from the tile's `MASK_16` wrap period (64 bytes ⇒ 32 texels), but `assets/oddtextures.c`'s actual `gsDPSetTile` for the RGBA16 **render** tile (tile 1, e.g. `globalDL_0x078:50`) sets `line=4` explicitly (independent of the mask), giving `line_size_bytes = 4*8 = 32` bytes ⇒ `width = 32/2 = 16` texels — and `size_bytes` from the paired `gsDPLoadBlock(...CALC_LRS(16,14,G_IM_SIZ_16b)...)` is 16×14×2 = 448 bytes ⇒ `height = 448/32 = 14`. `import_texture_rgba16`'s existing `width = line_size_bytes/2` / `height = size_bytes/line_size_bytes` formula therefore computes the CORRECT 16×14 for every fire tile, cross-confirmed by the pre-existing `is_fire_shape = (width==16 && height==14)` filter already hardcoded into `gfx_opengl.cpp`'s `GE_TEXDUMP` dumper (D219-era) — that filter only makes sense if 16×14 is what the importer actually produces at runtime. **Do not chase width/height again.** Also checked and rejected: the render tile's `gsDPSetTileSize(1,...,CALC_TILESIZE(56),CALC_TILESIZE(56))` sets a 56×56 texel-coordinate WINDOW for UV-wrapping the 16×14 content against the co-rendered 56×56 smoke texture (`G_CC_INTERFERENCE`, TEXEL0=smoke/TEXEL1=fire) — this is not a usable content-dimension source and switching to it (the naive "use SETTILESIZE instead of line_size_bytes" fix) would have made the bug worse, not better. **GEVR-TRIAGE item E (byte-group-order hypothesis) also does not apply new information** — the original session's decode-order check (still correct: `addr[2i]<<8|addr[2i+1]`, matches the already-fixed D228 IA16 sibling) stands; no runtime capture this session found a byte-order anomaly either (see below). **Root cause is now genuinely open, not "unconfirmed suspect"** — both named hypotheses are dead. Next investigator: the TMEM slot the fire render tile always loads into (`tmem=0x188`/392, `rdp.loaded_texture[]` indexed by tmem, hard-coded identically across all 15 `globalDL_0x078..0x9a8` records) is a plausible next avenue — 15 *different* ROM source images share one fixed TMEM address and one GL-cache key derived from `{texture_addr, fmt, siz, size_bytes, ...}` (`gfx_pc.cpp:~1962`); worth checking whether rapid record-to-record aging ever reuses a stale GL texture object keyed wrong, though this is an unverified new lead, not a confirmed mechanism. **Not runtime-confirmed this session** — added an env-gated diagnostic (`GE_D252=1` in `import_texture_rgba16`, `port/fast3d/gfx_pc.cpp`) that logs computed vs. SETTILESIZE dimensions and dumps raw texel bytes per distinct upload to `scratch/d252_raw_*.bin`, but two 35s/90s headless `-level_37` (Jungle) runs never triggered an explosion/fire draw (no player input available in this environment) — needs an interactive session or a longer/different-level run to actually fire the probe. **User-confirmed live, same session: still reproduces on plain bullet-impact sparks**, not just explosions — this is a much cheaper repro than an explosion (any gunshot against a wall/surface) and should be the target for the next `GE_D252=1` capture, not a scripted explosion. |
| D253 | **Facility (and any level with a FORCECROUCH/LADDER linked special tile) SIGSEGV on Linux/Steam Deck — `stanCheckLinkedSpecialTile` stores to a garbage `outFlags` pointer (D203 root cause).** The B-callback typedef `standTileLocusCallback_B_t` (`bondtypes.h:573`) declares the middle three params as `f32 f32 f32`, so on x86-64 SysV the caller (`sub_GAME_7F0B1DDC`, `stan.c:2228`) passes the 6th arg (`record`) in **`%rdx`** (3rd integer slot; the floats take xmm0–2). But the actual implementation `stanCheckLinkedSpecialTile` was decompiled with an all-`s32` prototype, so it reads `outFlags` from **`%r9`** (6th integer slot) — a stale scratch value left by earlier callees. On the N64 (MIPS o32) both prototypes agree: everything flows through GPRs a0–a5 in order, 6th arg = a5 either way, so the original is fine and the decompilation looked consistent. The mismatch stays dormant until a tile actually hits the FORCECROUCH or LADDER branch and *stores* through `outFlags` → SEGV; that's why it was level-specific (Facility's opening crouch tile) and input-dependent, and why idle repros were negative. The sibling B-callback `stanLocusEdgeIsAboveY` already had the correct float prototype, which is why only this path was booby-trapped. **Fix (ABI-class, semantics-preserving):** declare the three middle params of `stanCheckLinkedSpecialTile` as `f32` in both the forward decl and definition (`stan.c:152`, `stan.c:2407`) to match the typedef; the body never uses them, and on N64 o32 f32 args occupy the same GPRs, so N64 codegen is unaffected. **Verified:** gdb run on WSL (gcc 15) — pre-fix, SIGSEGV at `stan.c:2441` with `r9=0x3` and valid `record=0x409fea00` in the caller frame; post-fix, the same walk route completes with no crash. Side benefit: the FORCECROUCH/LADDER branches were effectively dead-on-arrival on PC (any hit = crash); they now function. **Also observed post-fix (separate, WSLg-only):** a segfault *inside libgallium* during `gfx_opengl_upload_texture` (32×48 RGBA + mipmaps) while walking further — GL-driver-side under WSLg's Mesa; not reproducible on Windows; park as a new finding if it recurs on real hardware. | FIXED — Deck-verified: user got past the crouch tile with the re-cut bundle (a *new*, later crash is D255). |
| D254 | **Linux crash handler destroyed its own logs: `ge007.crash.log` only ever contained the SIGABRT block, never the original SEGV.** The POSIX `crashHandler` (`port/src/crash.c`) finished by calling `sysFatalError()` → `abort()` → SIGABRT, which re-entered the *same* handler (SIGABRT was one of the installed signals); the second pass reopened the log with `"wb"` and overwrote the first pass's SEGV block — faulting PC, signal, everything. The Deck firefight crash (D255) came back as a bare `SIGNAL: 6` abort record for exactly this reason, forcing offline disassembly of the CI binary to recover what the log should have said. **Fix (port-only):** re-entrancy guard (`inCrashHandler`, second pass `_exit(128+sig)` immediately) + terminate directly with `_exit()` instead of `sysFatalError()` (the Windows path already terminated directly for the same reason) + the Linux block now dumps the faulting context: `FAULT ADDR` (`si_addr`) and the full general register set (Rax/Rcx/Rdx/Rsi/Rdi/R8–R11, Rsp/Rbp) from the `ucontext`, so a field crash now names the bad pointer on its face. **Field-verified 2026-09-13:** the D255 Deck crash came back as `SIGNAL: 11` + `FAULT ADDR: 0x10` + full register set — exactly the format designed here; it directly identified the NULL-model case (see D255). | FIXED — field-verified. |
| D255 | **Facility firefight SIGSEGV on Steam Deck (v0.2.0-pre CI bundle, D253 fix confirmed working — user got past the opening crouch tile).** Crash during gunplay; abort-unwind backtrace points at `sub_GAME_7F09BAC4` (`vtxstore.c:158`) ← `sub_GAME_7F09BBBC` (`:220`) ← `lvlManageMpGame` (`lv.c:1302`) ← `bossMainloop` (`boss.c:567`). Disassembling the exact CI binary at the fault PC (`0x200db135`) gives `mov 0x14(%rax),%edi; mov 0x10(%rdi),%rax`: the function walks the active PropRecord list keeping `type == 1` — which is **PROP_TYPE_OBJ, not CHR** (the decompiler cast to the union's first member and misnamed the field `ChrRecord.chrflags`; it is actually `ObjectRecord.model` at `0x14`) — then dereferences `model->obj`. Two port-relevant wrinkles: (a) the load is **32-bit zero-extended** because the decomp types the field through the `CHRFLAG` enum, so a model pointer ≥ 4 GB would truncate to garbage; (b) the loop skips records whose `model->obj` is NULL but *dereferences* `model` unconditionally, so an active object with a NULL/freed `model` faults — plausible in a firefight where destructibles (crates etc.) die. Not reproduced on WSL after equivalent play. **Register capture (D254 format, 2026-09-13):** two more Deck crashes — Bunker mid-firefight, then Frigate with ~3 NPCs alive; log: `SIGNAL: 11 / FAULT ADDR: 0x10 / Rax=0x701e5f54 Rdi=(nil) / PC 0x200db135`. `Rdi` = the record's model value = **exactly NULL** → case (b) confirmed: an active PROP_TYPE_OBJ record with a NULL `model`, dereferenced by the unconditional `mov 0x10(%rdi),%rax`; truncation (a) and stale/garbage pointers are ruled out. Faulting PC matches the recorded disassembly (`sub_GAME_7F09BAC4` = `vtxstore_fix_refs`). Grep-verified: nothing in game code ever writes NULL into `ObjectRecord.model` — `clear_model_obj()` nulls `model->obj` only, and `objInit`/spawn paths only activate a record once model ≠ NULL — so the NULL comes from either a fresh/uninitialized record slot being activated or a destroy-path ordering race (model freed/nulled while the record is still active this tick). **Fix constraint:** any guard lands in game code (`vtxstore.c`) → needs rule-#2 sign-off; `port/` cannot intercept this dereference. Next: (1) decision on whether a NULL-model skip in `vtxstore_fix_refs` qualifies as an ABI-class exception (semantics-preserving defensive guard, PD analogue check); (2) if yes, add it plus a `#ifdef PORT` trace at the deactivation/model-free sites to name the path that leaves the record model-less. **Corrected, deterministic repro (v0.2.1 user testing, 2026-09-15, superseding the initial report): the trigger is Facility's computer terminals — the ones you ACTIVATE (interact with, not shoot/destroy) to open a door for level progression.** User confirms this is consistent/reliably reproducible on those specific terminals, on Steam Deck **and** Linux generally (not Deck-hardware-specific). This is a much stronger lead than the original "sometimes during firefights" — a terminal activation deactivating/freeing its associated object model while the record briefly stays active this tick is fully consistent with the destroy-path-ordering theory above, just via a scripted trigger instead of combat destruction, and gives a deterministic repro instead of a probabilistic one. **Next session should reproduce via terminal activation specifically**, not generic combat — much cheaper to instrument. **M-140: added a diagnostic probe (`GE_D255`, off by default, `#ifdef PORT`) — `vtxstore.c`'s `sub_GAME_7F09BAC4` now skips-and-logs instead of dereferencing when it finds a NULL model on an active OBJ record (identity + state fields), and `propobj.c`'s `objFree` logs every free call for correlation. Attempted a live Windows repro this session (5 terminal activations across 3 Facility launches, one direct `-level_34` boot each time): zero probe hits, zero crashes — Windows does not reproduce this, consistent with the existing WSL-negative result. Two negatives (WSL, native Windows) against one positive (Linux/Steam Deck) points at something genuinely platform-specific (allocator/timing), not a Windows-reachable path — don't spend further Windows session time chasing a live repro.** The probes are harmless (env-gated, zero behavior change when unset) and committed so a Linux/Deck session can set `GE_D255=1` and get the trace directly. **User report, 2026-09-15 (same day, later, real Steam Deck hardware on the published v0.2.1 release — unverified by this session, logging as reported, not re-asserting the activation theory as wrong):** Facility crashes **consistently on terminals being destroyed** (shot), not (only) on activation. Frigate crashed once during this same testing pass, specifics not recorded. Shooting terminals elsewhere/other attempts did not reliably reproduce. This is closer to the *original* pre-M-139 framing (combat/destructibles triggering the NULL-model race) than the M-139 "corrected, activation-only" theory — both are plausible triggers of the same underlying destroy-path-ordering race (any code path that frees/nulls an active OBJ record's model mid-tick), not necessarily contradictory. **Not yet reconciled or verified by a session with hands on the repo** — next Linux/Deck session should test both triggers (activation AND destruction-by-gunfire) with `GE_D255=1` set before revising the "reliable repro" framing again. **Reconciled, same day, further real-Deck testing (v0.2.1 published release): destruction is the trigger, not activation — confirmed across four levels.** Facility: **every** terminal crashes consistently when destroyed (not merely activated). Caverns: terminals near the radio objective crash when they explode — same mechanism, different level, rules out a Facility-specific asset/state bug. Archives: a **similar** crash when a guard's grenade detonates — same destroy-via-explosion mechanism, no terminal involved at all, which narrows the real trigger to "an object dying via explosion damage" in general, not anything terminal-specific. **Negative control:** shooting an entire room of terminals on Frigate did not reproduce it — so it isn't simply "any destructible OBJ record exploding" either; something about the specific object/scene state still gates it. Release notes updated to match (`.github/release-notes.md`, D255 bullet: "consistently crash the game when destroyed"). **Static lead for next session (untested, from this session's code read):** `objFree`'s destroy path (`propobj.c:1017-1039`, entered via `RUNTIMEBITFLAG_REMOVE` from explosion damage, `sub_GAME_7F04424C`/`sub_GAME_7F0442DC`) never explicitly nulls `ObjectRecord.model` — `clear_model_obj()` only nulls `model->obj`. The NULL therefore can't come from that assignment; it must come from either (a) `chrpropFree`'s freed slot being reused by `chrpropAllocate` for a *different*, not-yet-`objInit`'d record before `vtxstore_fix_refs`'s later same-frame pass reaches it (a stale-pointer/reuse race, not a field write), or (b) the active-list unlink (`chrpropDelist`) happening while some other code is mid-iteration over the same list (a modify-during-iterate bug) — explosion damage plausibly spawns new effect props (sparks/debris/smoke) in the same tick that immediately reuse the just-freed slot, which would explain why it's tied to *explosions specifically* and not simple destruction. **Next, concrete:** with `GE_D255=1` set (probes already in tree from this session, ship in any build off `release/v0.3.0` including the local Linux package built this session), reproduce on Facility from the Deck itself (Desktop Mode, launch from a terminal, redirect stdout to a file) and send the log back — the probe already logs the dying record's identity plus every `objFree` call, which should show whether the crashing record's `prop`/`obj` pointer address was recently freed-and-reused this same session.

**ROOT-CAUSED AND FIXED, same day, live Deck session over SSH (M-140).** The `objFree`/destroy-path-race theory above was a red herring — set up a real Deck test (SSH access, key-based auth, WSL-cross-compiled local Linux build deployed and run live against the user's own prepared ROM assets), reproduced the crash directly, and symbolized it on-device against the build's own debug info. **First capture with the `GE_D255` probe active:** the probe's `ObjectRecord.model == NULL` guard did NOT fire, yet the game crashed anyway at `vtxstore.c:199` (`var_v1->RootNode`), `FAULT ADDR: 0x10`, `Rdi=(nil)` — a *different* pointer than the one the probe was checking. **Actual bug:** `sub_GAME_7F09BAC4` (`vtxstore_fix_refs`) reads the record's `Model*` through `((Model*)((ChrRecord*)prop->chr)->chrflags)` — a union pun onto `ChrRecord.chrflags`, which is typed `CHRFLAG`, a **4-byte enum**. On 32-bit N64 this is bit-exact (both the enum and a pointer are 4 bytes, so casting the enum's value to `Model*` reproduces the real stored bytes). On 64-bit PC, `Model*` is 8 bytes but the enum read is still only 4 — the cast reads just the pointer's low 32 bits as an `int` and zero-extends it back to a 64-bit address, discarding the real high bits entirely. This is the **exact D3x pointer-width ABI class** documented in AGENTS.md/porting-notes — a struct field misread across the 32→64-bit pointer-width transition — and it was *already flagged* as suspect (a) in this very row's first paragraph, back in the original v0.2.0-pre investigation, then wrongly deprioritized in favor of suspect (b) (the NULL-model theory) once a live register capture happened to show `Rdi=(nil)` and looked like it confirmed (b). It didn't: `(nil)` is exactly what a 32-bit-truncated read produces whenever the real pointer's low 32 bits happen to be zero — a coincidence that pointed the whole investigation at the wrong field for two sessions. **Proof it was the field, not the object:** this session's `GE_D255` probe read the SAME record through the *correctly*-typed `ObjectRecord.model` field (`Model*`, 8 bytes) and it was **always a valid, non-NULL pointer** — the object's model was never actually missing. **Fix (ABI-class, semantics-preserving, `#ifdef PORT`; `src/game/vtxstore.c`, `sub_GAME_7F09BAC4`):** read the pointer through `ObjectRecord.model` (the same union memory, correctly typed as `Model*`) instead of through the mistyped `ChrRecord.chrflags` alias, at both use sites in the function (the `RootNode` walk and the `modelGetNodeRwData` call). Identical bytes and identical result on N64 (both fields are 4 bytes there); correct on 64-bit PC. The earlier `GE_D255` diagnostic skip-guard (which checked the wrong field and would never have actually fired) is removed — it's no longer needed, the record's model was never really missing. **Verified live, in order:** (1) diagnostic build with the guard checking BOTH fields caught the exact moment of divergence — `ObjectRecord.model` valid, `ChrRecord.chrflags`-cast pointer producing the crash — nailing the mechanism beyond doubt; (2) the real fix (no guard, just the corrected field) held through the *entire* Facility level including multiple terminal destructions (the most consistent prior repro) and through Caverns' radio-objective terminals (the second confirmed level) with zero crashes, live on the user's own Steam Deck hardware over the ROM/assets they already had prepared. Windows rebuild + full 21-level `verify.sh sweep` also run this session as the standard regression gate. | **FIXED** — root-caused (D3x-class 32-bit pointer truncation via a mistyped enum-to-pointer cast, not a destroy-path race) and verified live on real Steam Deck hardware across both confirmed-repro levels (Facility, Caverns), zero crashes; Windows sweep clean. The separate audio-thread crash (D285) is untouched by this fix. |
| D256 | **Saves (and config) silently failed to persist on Linux whenever the game was not launched from the directory containing `data/` — reported as "saves do not work at all" on Steam Deck.** `romdataInit` finds the ROM in `$S/` (CWD `data/`) *or* `$E/` (exe dir) *or* `./`, but `sysResolvePath`'s POSIX `$S` branch (used by `ge007.eep` and `ge007.ini`) only ever produced a CWD-relative path — its "fallback" was `./data/`, identical to the primary. Launch from any other CWD (Steam non-steam-game shortcut, terminal in `~`): ROM loads via the exe-dir candidate, game runs, but every EEPROM/config write hits `fopen(..., "wb")` on a nonexistent `./data/` and fails with a stderr-only "cannot write" — saves never persist, and **every launch also starts with default settings** (F10 changes don't stick), which presents as "draw distance/LOD is low" on levels like Dam where the user had raised `Video.DrawDistance` on another machine. The Windows branch already did CWD-else-exedir; the POSIX branch now mirrors it (`readlink("/proc/self/exe")`) and creates the directory best-effort so first-run writes don't fail. **Verified on WSL:** with exe+`data/` in `/tmp/fb2` and CWD=`/`, config/eeprom/ROM all resolve into `/tmp/fb2/data/` and a 25 s run completes clean; the pre-fix binary in the same layout fails to find the ROM at all. **Caveat:** if the user's Deck launch *was* from the bundle dir (CWD had `data/`), this fix does not explain their save failure — need the save-flow detail (does the memory-card screen appear? slot confirm? "saved" message?) before closing the report. Side note found while testing: a `data/` without the generated sidecars (`pccg-*/`, `pcmodels-*/` from `tools_pc/d69_emit.py`/`d43_emit.py`) starts fine but faults in `load_bg_file` (`bg.c:865`) on level load — the reserve functions warn "stage loads will fault" and return 0; expected degraded mode, not a new bug. | FIXED (path half) — save-flow detail still owed from the user. |
| D257 | **New `Game.AllUnlocked` option (F10; default flipped OFF for v0.2.0 — see D259): skip N64 progression entirely — all 21 solo levels selectable at every difficulty, 007 mode available, and the cheat menu fully populated. Port-layer only; no game-code edits.** Three mechanisms, all in `port/`: (1) `main.c` seeds the two live `LEFTOVERDEBUG` RAM flags after `configLoad()` — `debug_enable_all_levels_flag` (all levels) and `debug_007_unlock_flag` (007 mode; both in `src/game/debugmenu_handler.c`, same class as the existing `GE_UNLOCK_ALL` getenv hook in the getter); (2) the eep shim (`libultra.c geEepromPatchAllCheats`) patches the save block at read time — special-casing `fileValidateSaves`'s exact pattern (block 4, 5×96 bytes) — setting all 20 progression-gated cheat-unlock bits per slot (`unlocked_cheats_1/2/3`; the cheat menu row on mode-select only renders when ≥1 is set, `front.c interface_menu06_modesel`) **and** filling every *empty* stage/difficulty completion-time field with the max time 0x3FF (existing player times preserved); both are then CRC'd per slot with the game's own `fileGenerateCRC` or `fileValidateSaves` would wipe the slots. (3) `optionsoverlay.c` row so users can toggle OFF for faithful progression. **Why the completion-time half is required (found by user repro):** when ANY single-player cheat is active, `g_AppendCheatSinglePlayer` makes mission-select (`get_highest_unlocked_difficulty_for_level`, front.c) accept only levels with status `STAGESTATUS_COMPLETED (3)`, not merely `UNLOCKED (1)` — AllUnlocked's debug-flag path returns 1, so enabling a cheat locked every level ("could only do Silo"). Reading the save as a fully-completed campaign satisfies both branches. Time bit-math mirrors `fileGet/SetSaveStageDifficultyTime` verbatim (10-bit fields, difficulties 0–2; 007 times are virtual). **Pitfall hit during implementation:** the first mirror of `save_data` used `times[80]` instead of `(SP_LEVEL_MAX-1)*4 = 76`, making `sizeof` 100≠96 so the block-read size check silently never matched — now enforced by a compile-time `ge_save_slot_size_check`. **Semantics:** unlocked cheats/completions persist to `ge007.eep` on any subsequent save write (self-consistent; mirrors N64 where completing a level permanently unlocks its cheat). No *active* cheats are enabled (invincibility/all-guns stay OFF); weapons remain per-mission pickups. **Verified:** fresh eep — all 5 slots validate `stored==calc` under `GE_SAVELOG=1`; user playtest: all levels selectable at each difficulty, 007 mode present, cheat menu populated, cheats toggle and work (time-modifying cheats untested by user), mission select intact with all cheats ON. | FIXED — verified by user playtest (PR #79 merged); default flipped OFF for v0.2.0 (D259). |
| D259 | **`Game.AllUnlocked` (D257) breaks audio on a fresh save: with it ON, music and SFX load at volume 0 (silence). v0.2.0 ships with `AllUnlocked` default OFF.** User-observed during v0.2.0 artifact testing ("all unlocked seems to break sound"). **Mechanism:** a fresh `ge007.eep` is zero-filled, so every save slot is all-zero. On N64 such a slot fails `fileValidateSaves`' CRC check and is rewritten as `BLANKSAVEDATA` (`src/bondconstants.h:2148`), whose `music_vol`/`sfx_vol` are `0xFF` (max). With AllUnlocked ON, `geEepromPatchAllCheats` (`port/src/libultra.c`) patches the zeroed slots (cheat bits + completion times) and recomputes a **valid** per-slot CRC, so `fileValidateSaves` accepts them as-is and never calls `fileResetSave`. The volume bytes therefore stay 0, and `fileLoadSettingsForFolder` (`src/game/file2.c:1289-1290`) loads them verbatim: `set_mTrack2Vol((save->music_vol << 7) | (save->music_vol >> 1))` = 0 → silence. With AllUnlocked OFF the zeroed slots fail CRC and reset to `BLANKSAVEDATA`, which is why sound works in the default configuration. **Disposition for v0.2.0:** flip `portAllUnlocked` default to 0 (`port/src/video.c`; `SkipIntro` was already 0) — faithful N64 progression out of the box, F10 *All unlocked* still enables the goodie. The quirk then only affects users who deliberately toggle it ON with a fresh save (recoverable: raising the in-game volume writes a nonzero slot, after which the patch leaves volumes alone). **Deferred fix candidate (port-only, ~6 lines):** in `geEepromPatchAllCheats`, detect all-zero slots and seed `music_vol`/`sfx_vol = 0xFF` before the CRC recompute, emulating what `fileResetSave` would have done; real (nonzero) saves incl. a user-muted one stay untouched. | **FIXED in v0.2.1** — the deferred fix was applied: `geEepromPatchAllCheats` now detects all-zero slots and seeds `music_vol`/`sfx_vol = 0xFF` before the CRC recompute, emulating `fileResetSave`; real saves (incl. a deliberately muted one) stay untouched. **By-ear confirmed by the user** (v0.2.1 test pass: fresh save, AllUnlocked ON — audio plays correctly). |
| D260 | **v0.2.1 (user request): mouse-wheel weapon-select direction swapped.** Wheel-up now selects the *previous* weapon and wheel-down the *next*, matching the in-game wheel menu's on-screen layout (the port had them inverted vs the user's expectation). Port-only: `inputPostWheel()` (`port/src/input.c`) — the notches>0 / notches<0 branches were swapped so scroll-up emits the back-notch and scroll-down the forward-notch. N64 build untouched; no game-code edits. | FIXED — build green; one-line branch swap, trivially reviewable. |
| D261 | **v0.2.1 (user request: "fix watch menu sensitivity for wasd/controller buttons"): holding W/S (or a fully-pressed stick axis) on the watch inventory page crawled one item per tap — no repeat.** Root cause is the interaction of two existing behaviours: keyboard W/S map to `stickY = ±STICK_MAX (0x50)` and SDL pads peg at 0x80, both *outside* the smooth-scroll band (0x1f..0x45) that D118d intentionally left for the port — so a held press only ever hit the latched single-step path (`watch_stick_y_pressed_*`, one step per press edge, no repeat). On N64 the same stick level was the *fast-scroll* band (one step per frame), which D118d removed because digital sources peg there and it became runaway scroll. **Fix (port-only, `#ifdef PORT` in `src/game/options.c`):** a bounded auto-repeat helper `geWatchStickFastStep()` — while the stick is held past the smooth band (≥ 0x46 up / ≤ −0x45 down), step one extra item every `GE_WATCH_NAV_RATE` (6) frames after an initial `GE_WATCH_NAV_DELAY` (15)-frame hold; the first press still goes through the untouched latched path, so a tap behaves exactly as before, and the smooth band + D-pad/C-button edge-repeat are unchanged. Called from `game_options_inventory_navigation()` only. **Verified:** build green; headless smoke (`GE_INPUTSCRIPT` long-W-hold on the inventory page) runs clean. Feel-check (repeat rate) owed to the user — both constants are at the top of the helper for a one-number tune. **User feel-check (v0.2.1 test pass): repeat "feels fine for now", but W/S is still too sensitive overall — desired behaviour is tactile one-press-one-item with no risk of scrolling the whole list on an over-held key; candidate tweaks (not yet done): longer `GE_WATCH_NAV_DELAY`, slower `GE_WATCH_NAV_RATE`, or gating the auto-repeat off for keyboard taps entirely (repeat only for analogue stick). Logged, no code change. | FIXED — feel-check received; W/S sensitivity follow-up logged above. |
| D262 | **v0.2.1 (user report): "3D models missing on watch menu screen item select" — the selected-item 3D preview does not appear on the watch inventory page.** User visually confirmed this session: no 3D model on the item-select page where one should be (default selection is *unarmed*, which has no model — a weapon must be scrolled to with S/DOWN). Headless repro (`GE_INPUTSCRIPT` pulse train + `GE_PCDUMP`, Dam `-level_33`): watch can't open during the intro freeze (`bondviewFrozenMoveBond` zeroes input while `g_CameraMode == CAMERAMODE_INTRO`; freeze ends ~frame 1400–1500 on Dam, and controller reads run faster than sim ticks, so scripts are timed by trial); once open + paged to INVENTORY (R_JPAD), the watch UI renders but the centre of the face is empty for every item. **Root cause found and fixed as D264** (`renderdata.flags == 0` in the watch subdraw pass — the cross-global ModelRenderData template read zeroes on PC; `subdraw()` then gated *every* geometry node, emitting exactly 1 word instead of the model's commands). Residual: with flags restored, the preview emits matrix/light-setup words (71 for item 5) but dump frames still show no visible centre geometry — a possible second layer (LOD/distance gating or fast3d rasterisation of the sub-DLs) remains unverified; user visual re-check pending. | FIXED (primary cause, D264) — pixel-level confirmation pending user run. |
| D264 | **D262's root cause (ABI/layout class, cf. D140/D52): the watch item preview's `ModelRenderData` template is read across two adjacent globals that no longer sit adjacent on PC.** `set_enviro_fog_for_items_in_solo_watch_menu()` (`src/game/gunfire.c`) starts with `ModelRenderData renderdata = *((ModelRenderData *) (&D_80035D00));` — a 64-byte copy that on N64 spans `D_80035D00` (a zeroed u32, .bss) *and* `D_80035D04[] = {1, 3, 0…}` (.data): the effective template is the one documented in the comment above `D_80035D00` in `gun.c` — `basemtx=NULL, zbufferenabled=TRUE, flags=3, rest zero` (`CULLMODE_BOTH` is 0). The PC link places .bss and .data in separate sections, so the read is all zeros: **`flags==0`**, and every geometry emitter in the subdraw node dispatch gates on `renderdata->flags & 1/2` (e.g. `modelRenderNodeDl`) — so `subdraw()` emitted exactly one word (`gSPSegment`) per preview, no matter the item. Probes this session: entry gates all pass (hand has item, model header non-NULL), node tree present (~400 nodes, isLoaded=1), world-path geometry nodes carry valid built DLs — but in the watch pass `flags=0` on every node. **Fix (`#ifdef PORT`, values identical to the N64 ROM bytes):** explicit template — `ModelRenderData renderdata = {0}; renderdata.zbufferenabled = TRUE; renderdata.flags = 3;`. Verified at Gfx level: preview emission went 1 → 71 words (item 5) / 81 (mid-scroll) / 9 (item 47, knife-class). Note for follow-up: the 71-word dump is a repeating {matrix/segment/light} unit with no `gSPDisplayList` in the first 24 words — if the user's re-check still shows no visible model, next suspects are LOD/distance gating of the local watch-space matrices (`modelUpdateDistanceRelations`, cf. 7c48f664) or fast3d handling of the preview sub-DLs (segment 3 / render modes). | FIXED — build green, probes removed; visual confirmation pending. |
| D263 | **v0.2.1 (user report): Facility — "Ourumov not shooting Trevelyan"; the scripted opening beat never triggers. RESOLVED: not a port bug — the beat is proximity/LOS-triggered exactly as on N64, and the full chain (incl. the execution shot) was verified firing end-to-end on PC.** The triage suspects (tick pacing, camera-mode gate, silent anim lookup) were all ruled out at runtime. **The beat, fully mapped** (static decode of `UsetuparkZ.c` — remember every DBYTE/TBYTE in the dump is bswap16/32'd relative to game interpretation): Trevelyan = guard chr 67 (body 9 = Trevelyan_006, pad 131) on ai_18: walks to pad 124, wanders, then loops a sees-Bond check; **only when Bond is in his vision** (range <10k, facing ±110°, fog, random gate ≈1/49/tick, LOS + *same stand tile* via `chrCanSeeBond`) does he jump to ai_19: runs at Bond, alarm + dialogue timers, sets an objective bit; BG list ai_41 (ID 0x1000, auto-started) polls that bit and spawns **Ourumov (body 7) at pad 88 with ai_23** plus a 6-soldier squad; ai_23 renumbers him to chrnum 78 and arms him → ai_22: runs to pad 123, waits for Trevelyan within 0x1400 of it, ~95 s dialogue cascade, then the **execution shot** (`HitChrWithItem` from 78 on 67). It is a multi-minute sequence that *starts* only when the player walks Bond west and down a floor into Trevelyan's room — Bond spawns 8900 units away on the upper floor with LOS blocked (probed directly: `stanTestLineUnobstructed` = obstructed, different stan tiles), identical geometry to N64. **Runtime verification** (`-level_34`, `#ifdef PORT` probes in the AI executor/vision path, all since removed): all 15 BG lists + guards start; Trevelyan alive on ai_18; every vision gate passes except the geometric ones at spawn distance; player movement confirmed working (note for future headless tests: under the port's default SOLITARE/natural-pitch scheme, walking is C-button/D-pad fed — W→C-up — and the analog stick carries pitch, so `GE_INPUTSCRIPT` SUP/SDOWN tokens do *not* walk Bond there; use UP/DOWN pulses). **Decisive test:** a temporary PORT-only teleport probe dropped Bond 65 units from Trevelyan at t≈12 s → seconds later: vision fires (chr 67 → ai_19), objective bit set, `SpawnChr body=7 pad=88` + squad, chrnum→78 on ai_23; ~2½ min later **`HitChrWithItem from_chr=78 target=67 part=8 item=6`** — Ourumov executes Trevelyan. Most likely explanation of the report: the tester never walked Bond into Trevelyan's sightline (or pre-D253 crash masked it). No game-code or port-code change required. **Kept from this session (test-harness only):** `port/src/input.c` — `GE_INPUTSCRIPT` analog-stick tokens were applied *after* the stick was already written out, so SUP/SDOWN/SLEFT/SRIGHT were silently discarded; the script override now runs before the clamp/write. | RESOLVED (not a bug), but refined (v0.2.1 further user testing, 2026-09-15) — beat verified firing end-to-end on PC incl. execution shot; **user playtest confirmed the sequence plays normally.** **Refinement: the exact trigger behavior is significantly sensitive to Bond's precise location and does not fire 1:1 identically to the N64 original across different approach paths/positions** — consistent with the beat's own documented sensitivity (range/facing/LOS/same-stand-tile + a ≈1/49-per-tick random gate, see the mapped mechanism above), so this is likely the same finicky-trigger behavior rather than a new bug, but not confirmed byte-identical to N64 for every approach. **User confirms this does NOT block level progression** either way — logged for awareness, not tracked as an open defect. Probes removed; input-script stick fix kept. **M-141: re-confirmed by user, one more mechanism now documented — the standoff has a hard TIMEOUT fallback, not just the shot-miss/objective-bit gates.** User reported (misread at first as a new freeze) reaching the Ourumov/Trevelyan standoff, waiting without shooting or triggering anything else, and after a couple of minutes Ourumov delivered his lines and executed Trevelyan on his own — **this is by design**: `ai_22`'s label `0x05` OR-gate (four early-exit conditions: two objective bits, a "Bond's shot missed" check, and a third objective bit set by an unrelated guard's surrender elsewhere in the level) is not the only path forward — falling through the whole dialogue-timer cascade (`label 0x42`'s chain of `if_local_timer_greater_than` checks) reaches `label 0x13`, which unconditionally fires `chr_hit_body_part_with_item_damage(0x43, 0x08, 0x06)` (the execution) once enough real time has elapsed, no player action required. Matches this row's own earlier "~2½ min later" measurement. **Do not re-flag "stuck standoff" as a bug without confirming the user waited the full timeout window first.** |
| D265 | **D236 investigation (port-layer, `port/fast3d/gfx_pc.cpp`): the tree cards are IA8/IA4, not IA16 — and `import_texture_ia4` decoded the RDP's I2:A2 texel as I3:A1, inverting the alpha ramp. Fixed; awaiting user visual re-test.** The earlier sessions misidentified the format: a live tri-time tile census (probe in `gfx_sp_tri1`, since removed) shows level_36's tree-card classes are (a) **IA8, line=64B → 64×16** (texdump r107), drawn OPA/bilerp/1-cycle/AC_NONE — *genuinely fully opaque* (A nibble F on 1023/1024 texels; opaque on N64 too) — and (b) **IA4, line=64B → 128×16** (texdump r117), drawn ZMODE_XLU/point/2-cycle with m2a=CLR_MEM/m2b=1MA, i.e. texel-alpha blended. There are no 32-wide IA tiles at tri time; the prior "32×32 tile" reading was wrong, and the whole earlier IA16 byte-pair/I11:A5 debate (r107) was analyzing an IA8 image as byte-pairs. **The wall's root cause:** r117 decoded as I2:A2 is a strip of ~13 discrete conifer silhouettes (trunk+canopy shapes visible in the ASCII render, `scratch/r117_ia4_alpha.png`) with a 3-level alpha ramp {0: 95% of texels, 85, 170} — faint dithered distant trees on N64. But `import_texture_ia4` read each nibble as intensity=`part>>1` (3-bit, SCALE_3_8) + alpha=`part&1`: A2=1 canopy pixels became **fully opaque**, A2=2 edge dithering became **invisible** → opaque repeating tree silhouettes = the "wall". The bug is inherited verbatim from the PD port (`pd_port/port/fast3d/gfx_pc.cpp` has the identical I3:A1 read — latent there, fatal for GE's tree strip). **Fix (applied):** `intensity = part>>2`, `alpha = part&3`, both via new `SCALE_2_8`; this is the hardware-correct RDP IA4 layout. Also cleaned up: the ia16 importer comment now states plain I8:A8 (the byte-pair read *is* correct IA16; the old "r107 proof" was misidentified data), and the earlier part-1 dimension fix stays (harmless/correct for real IA16 tiles). **Caveats/owed:** (1) user visual re-test of Surface 1 trees — headless A/B was in flight when work halted; a cross-run region diff was inconclusive (camera drift between captures); (2) the decode change affects *every* IA4 tile game-wide — regression-sweep other levels for dithered/alpha IA4 art that now renders with correct 3-level alpha; (3) if a hard repeating band remains after the fix, it is the opaque IA8 strip (a), which matches hardware and D236 needs re-scoping. **OUTCOME (user re-test): NEGATIVE — trees still render as the wall with the I2:A2 decode applied; the change was fully reverted (`git checkout` of `port/fast3d/gfx_pc.cpp`, back to the original PD-derived importers) and work is parked.** The census/RM/texture data above stands as evidence for the next attempt: the decode was *necessary* (I2:A2 is the hardware layout and r117 is unambiguously an authored tree strip under it) but *not sufficient* — remaining suspects, in order: (i) fast3d's blending of these XLU/2-cycle cards (texel alpha reaching the GL blend at all — the `texture_edge`/`SHADER_OPT_TEXTURE_EDGE` path forces `a>0.19 ? 1 : discard`, which for a 3-level ramp quantizes to opaque/discard, i.e. still hard-edged), (ii) per-card UV windowing/placement (all captured cards sampled the same u:0..64/v:0..16 window — confirm against N64), (iii) fog/distance. Note for the next session: verify what the N64 actually shows in that region before more port-side changes (the opaque IA8 strip is faithful to hardware either way). | PARKED — reverted to baseline; evidence + suspect list above. |
| D258 | **PAL/JP asset conversion is broken at the source-data level — `scripts/filelist.e.csv` and `filelist.j.csv` carry naming defects, so `gen_romassets.py e/j` (and the d43/d69/d88 emitters' `find_row`) cannot resolve PAL/JP builds. v0.2.0's drop-in ROM feature therefore ships NTSC-U only.** Discovered while scoping issue #6 (drop-in ROM). **PAL:** 31 of the 512 model rows in `filelist.e.csv` have truncated/renamed basenames (an extraction artifact — e.g. `Pbridge_console` ×6 where the resource table says `Pbridge_console1aZ…3bZ`, `Pchrgrenadelauncher` for `PchrgrenadelaunchZ`). **JP:** one font row is mislabeled — `fontBankGothic_fontchartable.bin` @3044276 is named `fontZurichBold_fontchartable.bin` (a duplicate label). Both were proven without a PAL/JP ROM: (a) greedy two-pointer sequence alignment of the 512-row `file_resource_table.inc.c` model order against each CSV's model rows gives an exact 1:1 match (481 by-name + 31 positional for PAL), offsets strictly non-decreasing, and zero size anomalies vs US (all within [0.5×, 2×]); (b) JP font row proven by the contiguity invariant — Gothic kerning @3043600+676 = 3044276 exactly hits the mislabeled row, whose size (8716) equals the US Gothic chartable, while Zurich's own chain is contiguous one row later. **Consequence:** no PAL or JP PC exe has ever been buildable from this tree (latent — only NTSC was ever exercised); the CSV `file` column is a label only (emitters look up by basename, never read the path), so the repair is a pure data fix: rename the 31 PAL model rows + 1 JP font row to the exact table symbols. **Scope decision (v0.2.0):** ship drop-in ROM support for NTSC-U only — `prepare-assets.py`/the frozen `ge007-convert` gate non-US ROMs with a clear message after SHA-1 region detection; the CSV repair + multi-region matrix + per-region exes are deferred to the Part B follow-up. | OPEN — repair path verified, not applied (out of v0.2.0 scope). |
| D284 | **`port/src/random.c`'s MIPS64 `dsll32`/`dsrl32` shift helpers were backwards (modelled as 32-bit shifts, not "shift by n+32 on the full 64-bit register"), desyncing the PC PRNG stream from the N64's from the very first draw, in every release to date (M-138 security/fidelity review; fix decided M-139b).** Fix: `dsll32(x,n) = x << (n+32)`, `dsrl32(x,n) = x >> (n+32)`, no masking — matches real MIPS64 semantics and `src/random.s`/`src/game/chrObjRandom.s` line-by-line. Verified by hand trace + an independent standalone C reproduction (both diverge from the buggy stream from draw 1). Full body below. | FIXED, 21-level sweep clean, no crashes — RNG stream changes from every prior build; existing replays will diverge, savegames unaffected |
| D285 | **A real Steam Deck SIGSEGV caught during D255 live-testing (v0.2.1 published binary, symbolized on-device via `addr2line` against its own debug symbols) is a SEPARATE bug from D255** — `FAULT ADDR: 0x62` inside the **audio thread** (`sndHandleEvent`, `src/snd.c:366`, via `amMain`/`portThreadWrapper`), not D255's `0x10`/`vtxstore_fix_refs` on the render/game-logic thread. Root cause (static, not yet gdb-confirmed): the pool-exhaustion preempt-scan in `sndHandleEvent` reads/walks the shared `D_800243E4` active-`ALSoundState` list with **no `osSetIntMask` lock at all** (checked: absent from the function's entire body), while the list's insert path (`sndSetupSound`, reached from 15 different `src/game/*.c` files' direct `sndPlaySfx` calls, e.g. `explosion.c` — i.e. from whatever thread the caller runs on) **does** lock it — an asymmetric-locking gap, the same D4/D147/D152 "N64 interrupts-off is not a real lock on PC" class. Fits the repro pattern (explosion SFX bursts = high-contention window; Frigate's lower SFX load not reproducing it fits a timing race, not a deterministic trigger). Full body below. | **FIXED (M-141, `src/snd.c`)** — `osSetIntMask` lock added around the preempt-scan, matching `sndSetupSound`'s existing unconditional (no `#ifdef PORT`) pattern in the same file; build green, `-level_09`/`-level_20` unregressed (1400+ frames, no crash, no D152 lock-stuck warnings). Not yet gdb/live-confirmed against a real reproduced crash — see M-141 body. Do not conflate with D255 — different thread, different subsystem. |
| D286 | **User-confirmed real (not system-side) ~8s frame freeze on Caverns, Steam Deck, caught during the same live session as D255/D285 — self-recovered, not a crash.** The game's own hang watchdog (`portHeartbeatCheck`) fired once (no frame for ~8s, `frames=46404`), dumped all-thread state via `crashDumpThreads`, then rendering resumed normally and the session continued to a clean close. Not tied to an explosion/terminal (rules out D255/D285). Not root-caused — no distinguishing log line precedes it; leading unconfirmed guess is a heavy room/asset stream-in stall (Caverns is a large level, cross-ref D250's FPS lineage, though the symptom shape differs). Also fixed in passing: the heartbeat log printed a raw microsecond value with a `ms` label (cosmetic, `port/src/libultra.c`). Full body below. | OPEN — one confirmed occurrence, no repro trigger identified, not root-caused; distinct from D255/D285. Do not chase a repro without a better lead. |
| D287 | **User-confirmed real, non-self-recovering freeze on Windows (M-141, 2026-09-15/16) after ~15-20 minutes with no input, triggered by clicking back into the window.** Caught live under `tools_pc/debug.ps1` (gdb-attached), evidence preserved at `scratch/d287_freeze_gdb.txt`. Last real `frame N rendered` log line: frame 17616. The heartbeat (`portHeartbeatCheck`) then fired 3 times with a **growing** stall (8001ms → 13017ms → 18021ms, `frames=17616` unchanged all 3 times) and then stopped logging entirely, while the VI-retrace pacemaker (`D51 vi post #N`) kept incrementing at its normal ~60 Hz rate for well over 100 more real seconds until the user closed the window — i.e. the pacemaker/OS-timer side is alive, the sim/render side is genuinely wedged, not the D286 shape (which self-recovered after one dump). **Different from D286**: no explosion/level-streaming context (idle + refocus, not combat), and the growing-then-silent heartbeat pattern (vs. D286's single fire-then-recover) points at a different mechanism. **Leading lead, not yet confirmed:** at freeze time, `gfxFrameMsgQ` (`src/init.c`, the scheduler's 32-slot client completion queue, registered via `osScAddClient`) shows `valid=32/32` — completely full — while the VI-retrace queue (8-slot) is not (`valid=1/8`). A full *client* queue with the producer (scheduler) still ticking implies the **consumer** (the main thread's `osRecvMesg(&gfxFrameMsgQ, ...)` in the frame loop) stopped draining it — consistent with the main/render thread blocking on something else (a lock, a GL call, a Windows message-pump stall) during or after a long-idle refocus, not with the scheduler itself stalling. Not yet identified what that something is. A pre-existing, currently-unused diagnostic (`GE_D60`, `port/src/libultra.c:630+`) already watches `gfxFrameMsgQ` receives and looks like the right next probe to enable. **Next, concrete:** reproduce with `GE_D60=1` (and ideally `GE_INPUTLOG=1`) running under `debug.ps1`, leave the window idle+unfocused 15-20 min, refocus, and once frozen `bt full` the actual game/render thread (not just the auto-dump's generic `WaitForMultipleObjects` frames for `shedThread`/`audioThread`) to find what it's blocked on. Cross-ref D134's superficially-similar-sounding-but-different mechanism (a different, 8-slot queue, already fixed) — do not assume the same fix applies here without confirming the actual blocked call site. | OPEN — one confirmed occurrence, real (not self-recovering), not root-caused; `gfxFrameMsgQ` full is a lead, not a confirmed cause. |
| D288 | **Silo intro: a huge, flat-shaded, untextured brown/tan triangle sweeps across the frame during the camera pan down the missile shaft (M-141, user report — "weird triangle drawn on silo intro randomly when the camera pans around the rocket"; previously catalogued as `GRAPHICS-BACKLOG.md` D171, "observed, not investigated" since M-34, never given a full findings.md entry until now).** Reproduced headlessly this session: `GE_PCDUMP=1-2000:10 ./build-pc/ge007.x86_64.exe -level_20` (run backgrounded, not foregrounded — see the environment note below), captured to `./ppm/`. Visual sweep (contact-sheet montages, `docs/img/bugs/D288-silo-stray-triangle-sheet.png` covers frames 571-751; `-closeup.png` is a full-res crop of frame 691) shows: **clean at frames 481-561** (no triangle), **first appears between frame ~562-601**, grows continuously in screen size through frame ~711 (by which point it covers roughly a third of the frame), then **shrinks and is gone by ~731-741** as the scene cuts to the first-person shaft-corridor view. Sampled pixel color is **exactly uniform** across the whole triangle (`RGB(83,62,44)` / `#533E2C` at two different sample points) — no gradient, no texture detail — i.e. this is a flat-shaded (no per-vertex color variance, no texture bound) single triangle, not garbage/noisy vertex-color data. **Reading of the evidence:** the grow-then-shrink-then-vanish arc across ~15 real seconds of camera motion is the classic signature of a genuine 3D triangle in world space that the camera flies close to/through during the scripted pan — not a 2D HUD/UI element (which would stay roughly fixed in screen size) and not obviously a `D146`-class freed-memory read (which reads as noisy garbage bytes, not a clean uniform flat color). The flat uniform color argues this is a REAL, correctly-colored asset triangle whose position/transform is wrong, which points at the same neighborhood as D106/D233 (near-plane / w-clip mishandling in `port/fast3d/gfx_pc.cpp`) rather than a game-code or converter data bug — but this is a read of the visual evidence, not a code-level confirmation; no probe has been run yet. **Not yet identified:** which model/prop this triangle belongs to (nothing in `UsetupsiloZ.c` was grepped for a matching solid-brown material this session), and no `port/fast3d/` tracing was done to catch the actual triangle at draw time. **Environment note (new, useful beyond this bug):** launching `ge007.x86_64.exe` in the **foreground** under this session's Bash tool with `timeout N ...` produced **zero** `ppm/` output despite `GE_PCDUMP` being set and verified exported — the `ppm/` directory was never even created (`GE_MKDIR` never ran or `frames`-gated capture never fired). Backgrounding the same invocation (`... &`, `sleep N`, `kill`) worked immediately and reliably. Not root-caused (didn't block this investigation), but worth remembering for the next session that needs `GE_PCDUMP` from this same tool. **Next, concrete:** (1) grep `assets/obseg/setup/UsetupsiloZ.c` / the missile-shaft model's material table for a solid `RGB(83,62,44)`-ish prim/vertex color to identify the source asset; (2) add a temporary `#ifdef PORT` probe in `gfx_sp_tri1` logging any triangle whose post-transform screen-space bounding box exceeds some fraction of the viewport together with each vertex's `w`, gated to `-level_20` frames ~560-740, to catch the culprit triangle's `w` sign/magnitude at the moment it blows up — same style as the existing `GE_D271`/`GE_D104` probes; (3) cross-check against D233's `w<0` trivial-reject fix and D106/D271's near-plane portal-bound clamp to see if this is a third site with the same underlying class of bug, or a genuinely new mechanism. | OPEN — visually confirmed and precisely bounded (frames ~562-741 of a bare `-level_20` boot), root cause not yet identified; promoted from the long-parked `GRAPHICS-BACKLOG.md` D171 entry with real evidence. |
| D289 | **Dam: a crash in the computer room, not reproducible on a second attempt (M-141, Steam Deck user playtest).** **User's own read, most likely: the same computer-terminal SIGSEGV as D255** (Facility door-opening terminals — root-caused and FIXED this session, M-141, `vtxstore_fix_refs` reading `Model*` through the wrong union field, a 32-bit pointer truncation). D255's mechanism is generic to computer-terminal props, not Facility-specific, so a Dam terminal hitting the same bug is plausible and would already be fixed by the same patch. **Not yet confirmed** which build the user was on when this happened (pre- or post-D255-fix) — if it was pre-fix, this is very likely just D255 and needs no separate investigation; if it recurs on a build that has the D255 fix, that's a new, real, distinct bug worth its own crash log. | OPEN — needs a human re-test specifically on a build with the D255 fix (this session's `release/v0.3.0`, not any earlier build) to see if it still happens. If it does, get `ge007.crash.log` (or a gdb `bt full` under `debug.ps1`) and the exact terminal/action — that's the actionable next step, not further static analysis. |
| D290 | **Watch's in-game "controller" section shows no control-pad graphic (M-141, Steam Deck user playtest).** Likely a missing/failed texture or model bind on one of the watch's equipment-display pages — candidate area `src/game/bondview2.c` (watch display) alongside the front-end options/controller-config screens, not yet located precisely. | OPEN — observed, not investigated. Needs a live repro (open the watch, navigate to the controller-config page) plus a screenshot for comparison against N64 reference. |
| D291 | **Dam: texture glitches when the truck drives past (M-141, Steam Deck user playtest).** No further detail captured (which surface/texture, timing, one-shot vs. every pass). Could be a moving-vehicle-specific instance of a known texture-cache/TMEM-residency class (cf. D124's stale-DL-marker mechanism, a different object family) or something new. | OPEN — observed, not investigated. Needs a live repro with the truck's exact position/timing and, ideally, a screenshot. |
| D292 | **Death animations are glitched — Bond is often sunk halfway into the floor (M-141, Steam Deck user playtest).** **User confirms this is the same broader "Bond's world model is positioned wrong" problem as D173** — not a new/separate mechanism, another surface of the same known-but-under-investigated bug. The umbrella is already tracked: `GRAPHICS-BACKLOG.md` D148 is explicitly "the tracking entry for the whole cutscene-reliability problem," already lists D173 (floating at level start / Cuba credits) and a Cradle instance ("Bond spins in circles" during the Trevelyan-falls cutscene) under it — findings.md's D173 entry has the fullest investigation detail (M-106's two ruled-out D209-style struct-aliasing candidates, LOW-MEDIUM confidence that class is even the right one). Nothing has actually been lost — the write-up just isn't concentrated in one place (split across `GRAPHICS-BACKLOG.md` D148/D173 and `findings.md` D173), which reads as "known but not detailed" unless you go looking. Sunk-into-floor is a third concrete instance of the family (level-start floating, cutscene spin, now death-pose sinking) — worth keeping as one thread when it's next investigated live, since a single root cause (most likely the render-side puppet/joint transform, not collision) could explain all three. | OPEN — same status as D173 (observed, not root-caused). Cross-ref D173/D148 as one investigation thread, not three separate bugs, next time someone has a live session to chase it. |
| D293 | **QoL ask: no discoverable way to quit the game from the UI (M-141, Steam Deck user playtest — "couldn't see a way to exit the game. Maybe build this into the settings/options menu.").** Confirmed in code: the only quit path is `SDL_QUIT` (`port/src/video.c:515`, the OS window-close event / Alt+F4) — there is no Escape-key handler and no "Quit" entry in the front-end menus or the F10 options overlay. Fine on a desktop with visible window chrome; a real gap on Steam Deck / controller-only / fullscreen setups with no easy way to reach the OS window controls. | OPEN — not investigated, low complexity. Candidate fix: a "Quit to Desktop" row in the F10 overlay (`port/src/optionsoverlay.c`) and/or an Escape-key binding in the front-end menu context (`port/src/input.c`), calling the same path `SDL_QUIT` already triggers. Cheap, port-only, good QoL-batch candidate. |
| D294 | **Statue: ground geometry near the large CCCP/hammer-and-sickle monument briefly renders solid black over a wide area, then self-recovers 10-30 seconds later (M-141, Steam Deck user playtest).** **Correction (user, same session): happened at normal ground level during normal forward progression through the level, near the statue/monument — NOT from the elevated ramp vantage point** (the ramp is a nearby landmark used to describe the location, not the trigger condition; a first-pass write-up of this report wrongly framed "wide visibility from elevation" as the likely cause — that framing is retracted). User's settings: max `Video.DrawDistance`/`Video.LodDistance`, high `Video.FovScale` (~130%). **Not an exact match for any existing tracked culling bug, but closely neighbors three:** D222 (high-FovScale culling breakage — but D222 is specifically *NPC/object* culling near screen edges, root-caused and FIXED, not ground/BG geometry); D249 (LOD/culling breakdown on large open levels — but D249 is at *default* FOV on Streets/Egyptian, and doesn't self-recover); D218 (wide-FOV fog/far-clip boundary artifacting — but D218 is a color/tint issue at the frustum edge, not geometry going missing). **Distinguishing features that don't match any of the three:** (a) the drop-out is transient and self-corrects without the player doing anything special, (b) it's ground/BG geometry specifically, not NPCs or props, (c) it happens *within* the draw-distance/FOV settings, not past a boundary. Leading unconfirmed guess, weakened by the correction above but not ruled out: a portal/room-visibility computation (the BFS the D106/D271/D233/D234 family already touches) temporarily excluding a room/cell that contains ground geometry until the walk recomputes on subsequent player movement — the widened-FOV-frustum framing no longer applies (the drop-out wasn't tied to an unusually wide sightline), but a room-transition/streaming-timing trigger at ground level is still plausible and fits the self-recovering, "normal progression" description just as well. This is a hypothesis from the symptom shape, not a code-level finding, and now has less to go on than the first write-up assumed. | OPEN — not investigated. Next: reproduce on Statue (`-level_22`) at the reported settings (`Video.FovScale≈130`, max draw distance/LOD) near the ramp by the CCCP monument; a `GE_D271`/`GE_D104`-style room-visibility trace (`roomsDrawn` count, portal-cull hits) correlated with the black-ground window would confirm or rule out the room-visibility-BFS hypothesis quickly. |

Phase 2 replaced the Phase-1 demo loop with the real `mainproc()` on real OS
threads, compiled GE's real `src/sched.c`, and brought in PD's fast3d software
RSP (`port/fast3d/`). The boot path now runs: ROM map → DRAM reserve → video
init → mainThread (real pthread) → bossEntry → bossInitMainthreadData through
mempool init, VI init, rspInit, joyInit + controller-init timers, stanInit,
gameInit — and currently dies inside `langInit()` (D31).

* **D24 — setjmp/longjmp green threads are unusable on MinGW x64.** The first
  kernel used setjmp/longjmp context switches; under Windows x64 + MinGW the
  longjmp path corrupts FPU/MC state when it interacts with SEH unwinding
  (observed: STATUS_DATATYPE_MISALIGNMENT-class crash inside a resumed
  "thread" with garbage register state). PD's port does not use setjmp at all.
  **Resolved:** `port/src/libultra.c` now runs every game thread as a real
  pthread (`PortThread` side table, 8 MB stacks). Message queues keep the N64
  `OSMesgQueue` layout and get a `PortQueue` side table (mutex + condvar, max
  64 queues); all osSendMesg/osRecvMesg paths lock it. A dedicated tick thread
  posts one VI retrace per frame (NTSC 60 Hz) and services the software timers
  (both OS_MESG_NOBLOCK — the tick thread can never deadlock). `idleThread` is
  intercepted by ID and parked; `mainproc` runs as a real pthread so the host
  main thread is free to pump SDL/OS events (`videoPumpEvents()` in the host
  loop — on Windows WndProc only runs on the window-creating thread). The GL
  context is made current on shedThread per frame via
  `gfx_sdl_make_context_current()`. Verified live: VI retrace flowing, all 4
  controller-init timers fire, bossmq loop completes.
* **D25 — dual-mapped N64 DRAM region.** Game code needs two incompatible
  address forms of the same RAM: `osVirtualToPhysical`/s32 arithmetic wants
  live host pointers that fit in a positive s32, while `offset | 0x80000000`
  rebuilds (bg.c:3184/3322, propobj.c:8578/8699, title.c:476) want a KSEG0
  view at 0x80000000. **Resolved:** one 8 MB backing store mapped twice —
  V1 @ 0x70000000 (all game RAM symbols live here) and V2 @ 0x80000000
  (byte-identical mirror). `port/src/dram.c`: Windows uses
  `CreateFileMappingW(INVALID_HANDLE_VALUE,…)` + two `MapViewOfFileEx(…, base)`
  calls (the documented kernel32 API that maps a section at an exact address;
  `NtMapViewOfSection` fails with STATUS_MEMORY_NOT_ALLOCATED 0xc0000045 and
  segfaults on wrong SECTION_INHERIT values — do not retry it); POSIX uses
  memfd + two MAP_SHARED|MAP_FIXED mmaps. Both views sanity-checked to alias.
  `port/src/dram_syms.s` pins the absolute symbols: `cfb_16` @ 0x70000000
  (0x4B000), `_bssSegmentEnd` @ 0x70050000 (mempool start), tlb block end @
  0x702F4400 (= page_align_down(0x803AB400) − 93×0x2000, the exact N64 value;
  the tlbmanage stub returns it so `mempInit`'s pool size is correct — a NULL
  return here made the mempool spin in `while(1)` at src/memp.c:164).
* **D26 — address-width shims (K0 sign-extension).** N64 K0 addresses have bit
  31 set; passing them through s32 parameters sign-extends to invalid x86-64
  pointers. With the dual map, live RAM sits below 0x80000000 so identity is
  safe: `port/shim/PR/R4300.h` → `PHYS_TO_K0(x) = (x)`; `port/shim/PR/os.h` →
  `OS_K0_TO_PHYSICAL(x) = (u32)((char *)x - 0x70000000)` (GBI w1 words then
  carry small offsets) and `OS_PHYSICAL_TO_K0(x) = (x)`. fast3d's `seg_addr()`
  (gfx_pc.cpp) resolves both forms: full host pointers pass through, values <
  0x800000 get +0x80000000 (landing in V2). Same remap applied to the
  G_MW_SEGMENT handler.
* **D27 — ASLR must be OFF for this exe.** `src/bondgame.h:8` declares
  `extern u32 *_bssSegmentEnd;` (pointer type), so `&_bssSegmentEnd` in game
  code emits a `.refptr._bssSegmentEnd` slot holding the absolute value with a
  BASE relocation. Under ASLR the loader rebases it to runtime_base +
  0x70050000 = garbage, and `mempCheckMemflagTokens` then spins/AVs with
  poolAreaStart ≈ 0xF0xxxxxx. (On MIPS `&abs_symbol` is the symbol value
  itself; x86-64/PE cannot do that.) **Resolved:** MINGW link flag
  `-Wl,--disable-dynamicbase` (note: `--disable-dynamic-base` is NOT
  recognized by GNU ld 2.47); the image now loads at its preferred base
  0x140000000 and relocations are no-ops. `main.c` fails loudly at startup if
  `sysImageBase() != 0x140000000ul` so any future ASLR regression is a clean
  error, not silent DRAM corruption.
* **D28 — ninja stale-object hazard with new shim headers.** If a shim header
  did not exist when a .obj was last built, it is absent from that object's
  .d dependency file and ninja will NOT rebuild the TU after the header
  appears — the old (unshimmed) code silently persists. Observed: boss.c.obj
  still contained `or $0x80000000,%eax` (original PHYS_TO_K0) days after the
  shim landed. **Rule:** whenever anything under `port/shim/` changes, delete
  all `.obj` files and full-rebuild. Verify with `gcc -E` using the exact
  ninja flags (`ninja -C build-pc -t commands <tgt>`) plus objdump of the
  rebuilt object.
* **D29 — `osPiStartDma` must post its completion message on every path.** The
  N64 PI posts the caller's OSMesgPI to mq when a DMA completes; GE never
  inspects the message but always blocks on it (`romReceiveMesg()` →
  `osRecvMesg(OS_MESG_BLOCK)` in ramrom.c:44). The shim did the memcpy
  synchronously and posted nothing → mainThread deadlocked in the first file
  load. **Resolved:** post unconditionally after `piServiceDma()` (also for
  skipped/dropped DMAs) with OS_MESG_BLOCK, posting the caller's OSMesgPI.
* **D30 — crash handler without SEH.** MinGW GCC 16 has no `__try`/`__except`
  and does not recognize `-fseh-exceptions`, so the unhandled-exception filter
  cannot rely on structured exceptions to protect symbolication. **Resolved:**
  `port/src/crash.c` phase 1 writes raw fault info (registers, modules) to
  `ge007.crash.log` without touching dbghelp (SymInitialize/StackWalk64 can
  allocate and re-fault inside the filter); phase 2 does a validated manual
  EBP-chain walk (`-fno-omit-frame-pointer` on all TUs; each frame is [saved
  RBP, return address], stop if saved_fp ≤ fp or outside thread stack limits).
  `TerminateProcess` instead of abort (SIGABRT can re-enter the exception
  machinery). `crashDumpThreads()` (Toolhelp32 snapshot + SuspendThread +
  GetThreadContext + StackWalk64 per thread) is called from the heartbeat
  error dump with TID→name matching; StackWalk64 cannot read DWARF unwind info
  on MinGW, so multi-frame traces depend entirely on the EBP chain. Symbols
  are recovered offline: `addr2line -e build-pc/ge007.x86_64.exe -f -C
  <base+rel>` (the log's "rel" = address − actual load base).
* **D31 — langInit SIGSEGV: `zlib_huft_build` overflows load_resource's frame via the x86-64-grown `struct huft` (RESOLVED).**
  After D24–D29, mainThread reaches `langInit()` and dies on the **first** file
  load (`_fileNameLoadToBank(LnameX_lookuptable[LGUN][…])` → index 670 →
  `fileIndexLoadToBank` → `mempAllocBytesInBank` → `load_resource`). Crash
  signature: RIP = 0x140330003 (inside `.bss`, in `g_Props`), RSP only ~0xD8
  below langInit's entry RSP (a *shallow* chain — the stack pointer itself is
  fine), and the frame region above load_resource's return address is filled
  with 8-byte slots `{u32, 0x00000001}` where each u32 is the **low 32 bits of
  a live .bss pointer**, alternating between `g_Props` (0x14033xxxx) and
  `resource_lookup_data_array` (0x14034xxxx). The crash is a `ret` popping one
  of these truncated pointers as the return address (observed RIP equals the
  table value exactly, high word 0x00000001). Something in the load chain is
  storing 64-bit pointers into s32/int storage (fine on 32-bit MIPS,
  truncating on x86-64) and that storage overlaps a live frame.

  Established facts (all under gdb, `gdb -batch -ex "handle SIGSEGV stop"`):
  * All arguments at `load_resource` entry are valid: ptrdata=0x702aa400
    (mempool), srcfile->hw_address=0x108ed250, rom_size=0x720, source=
    0x702f3ce0. `romCopy` (PI shim memcpy) completes.
  * `zlib_inflate()` itself **completes normally** (`finish` → returns 0,
    rz_wp=3872; the RZ stream is plain deflate after a 2-byte header, so
    endianness is not an issue in the bitstream). At its entry rz_outbuf =
    ptrdata ✓, rz_inbuf = source+2 ✓, and the huft table base (tl) sits inside
    load_resource's 8 KB local `buffer` (rz_hlist).
  * **Root cause (confirmed):** `struct huft {u8 e; u8 b; union{u16 n;
    struct huft *t} v;}` is **8 B on MIPS** (4-byte pointer in the union) but
    **16 B on x86-64** (8-byte pointer, 8-aligned). GE's gzip-1.2.4 inflate
    builds its Huffman tables contiguously into load_resource's fixed local
    `u8 buffer[0x2100]` (8448 B), sized for the 32-bit layout (~1056 entries);
    on x86-64 only ~528 entries fit, so a stream needing more (observed
    `rz_hufts` ≈ 797) writes past the buffer and clobbers load_resource's frame
    — including the return-address slot. The earlier "truncated-pointer table"
    reading was a misattribution of this same overflow (the 8-byte `{u32,0x1}`
    slots are huft records spilling over the frame).
  * Ruled out by code inspection: `mempAddEntryOfSizeToBank` (only touches
    pool pos/prevpos), `fileGetIndex` (reads only), `decompressdata` epilogue
    (returns rz_wp).
  * `load_resource` prologue: `push rbp/rdi/rsi/rbx; mov $0x2128,%eax;
    call ___chkstk_ms; sub %rax,%rsp; lea 0x80(%rsp),%rbp` — RBP is a **fake
    frame pointer** (RSP+0x80); the real return address is at [RBP+0x20E8].
    The compiler reuses pushed-register slots as locals (verified: the saved-
    rbx slot [RBP+0x20D0] legitimately receives `buffer+12`). Watchpoints on
    RBP-relative offsets are therefore easy to misplace — compute from entry
    RSP instead.
  * Hardware watchpoint on the real retaddr slot (set at the first
    instruction, where [RSP] = retaddr): first write changes it from a valid
    .text return address to **`buffer+12`** (a pointer into load_resource's
    own local array), reported RIP at the prologue boundary (`push %rdi`) —
    i.e. either a concurrent write by another thread (hardware watchpoints are
    global) or an 8-byte-attributed access. A `thread apply all bt` at that
    moment showed every other game thread parked (RtlUserThreadStart / sleep
    syscalls; gdb could not unwind their stacks further), so the writer is not
    yet identified.

  **Resolution (fix in `port/`, game code untouched).** Mirrors the Perfect
  Dark port, which replaces its assembly rzip with a real-zlib C impl
  (`pd_port/src/lib/rzip_c.c`: `inflateInit2` + `inflate`). We exclude
  `src/game/decompress.c` + `src/game/zlib.c` from the PC build (CMake
  `list(REMOVE_ITEM SRC_GAME …)`) and add `port/src/rzdecomp.c`, which backs the
  two externally-referenced entry points with host zlib:
  * `decompressdata(src,dst,hlist)` → `inflateInit2(-15)` (raw deflate),
    `next_in = src+2` (GE's RZ = 2-byte header `0x11 0x72` + raw deflate, no
    size field), loop on `Z_OK`, return `total_out`.
  * `rzipGetSomething()` → returns the stored `next_in` (consumed input).
  A generated `port/include/realzlib.h` (from `realzlib.h.in`) `#include`s the
  **absolute** host `<zlib.h>` path, because the game's `src/game/zlib.h`
  shadows `<zlib.h>` on the `-I` path.

  **Verified:** rebuild links clean (238/238); the exe has the new
  `decompressdata`/`rzipGetSomething` and no `zlib_inflate`/`zlib_huft_build`.
  Under gdb, mainThread now boots **past langInit** (all language files load)
  all the way into `bossInitMainthreadData()` (boss.c:233) →
  `initWeaponAnimGroups` — i.e. the original first-file-load SIGSEGV is gone.
  The next blocker is D32.

* **D32 — ROM-serialized structs with embedded pointers have divergent N64/x86-64 layout (OPEN).**
  After the D31 fix, mainThread reaches `bossInitMainthreadData()` and dies in
  `init_weapon_animation_groups_maybe()` (boss.c:233) → `initWeaponAnimGroups`
  → … → `modelAnimReadRootMotionValue` (model.c:914), faulting on
  `desc->bitCount` where `desc = anim->bitDescriptors`.

  **Mechanism.** N64 pointers are 32-bit; x86-64 pointers are 64-bit with
  8-byte alignment. `struct ModelAnimation` (src/bondtypes.h:575) declares
  `ModelAnimBitField *bitDescriptors` and `u8 *bitStream`. In the N64 ROM
  layout these are 4-byte fields at 0x08 and 0x10; on x86-64 the compiler lays
  them out as 8-byte fields — `ptype /o` (gdb) shows bitDescriptors @0x08, a
  4-byte hole, then bitStream @**0x18**, sizeof = **80** (vs 64 on N64). The
  animation data is loaded by `alloc_load_expand_ani_table()`
  (initanitable.c:263) via `romCopy(ptr_animation_table,
  &_animation_dataSegmentRomStart, size)` — raw N64-layout bytes.
  `expand_ani_table_entries()` rebases the **low 32 bits** of
  bitDescriptors/bitStream but (a) leaves the high 32 bits as adjacent-field
  junk and (b) writes bitStream's rebase to offset 0x10, which on x86-64 is
  *not* where the C struct reads bitStream (0x18). Observed under gdb:
  anim = 0x702adad4 (base 0x702ad8c0 + 0x214), `*(u32*)(anim+8)` = 0xc82bd8c0 →
  as a 64-bit pointer the high word is junk → SIGSEGV.

  **Scope.** Any ROM-serialized struct that (i) contains a pointer field and
  (ii) has other fields after it misreads the same way: each 64-bit pointer's
  high word holds adjacent-field bytes, and every post-pointer field shifts by
  the pointer-width delta. ModelAnimation is the first hit; bondtypes.h has ~13
  struct blocks with pointer fields (fewer are actually romCopy'd from ROM).

  **PD ground truth.** PD keeps its ROM-data structs at N64 layout on both
  targets by storing embedded pointers as **u32** — e.g. `struct animtableentry
  { … u32 data; }` (pd_port/src/include/types.h:5124) — and casting to a real
  pointer at the use site. GE's decomp instead uses real pointers in
  ModelAnimation, so it diverges on x86-64.

  **Resolution — Option 1 chosen (user-approved), in progress.** The user
  approved the PD pattern after confirming PD does it (`animtableentry.u32 data`,
  raw-DMA'd then resolved at use sites). Note PD's *exact* mechanism (u32 ROM
  offset + on-demand DMA cache) differs from GE's rebased-DRAM-pointer
  ModelAnimation, so we adopt the **principle** (u32 embedded address, cast at
  use), not PD's code. This required refining non-negotiable #2 (see §H).

  **Part A — struct layout: DONE & proven.** Changed `ModelAnimation`'s two
  pointer fields to u32 (src/bondtypes.h:575): `u32 bitDescriptors` @0x08,
  `u32 bitStream` @0x10; added casts at the only use sites (model.c:913/921).
  `ptype /o` now shows sizeof = **64** with fields at the N64 offsets — layout
  matches. `expand_ani_table_entries()` writes s32 to 0x08/0x10 via its own
  `struct anim_entry`, which now aligns exactly (no change needed there).

  **Part B — pointer-rebase correctness: ROOT-CAUSED (see D33), fix designed,
  implementation pending.** The garbage values are *not* a bug in
  `expand_ani_table_entries`'s arithmetic — the rebase is correct N64 code fed
  wrong bytes, plus one genuine x86-64 ABI stride bug. Both root causes and the
  full fix design are in **D33** below.

* **D33 — ROM file stores structured fields big-endian; animation load needs a per-field endianness fixup + an x86-64 loop-stride fix (root cause of D32 Part B).**

  **Discovery.** The `.z64` ROM file is not a plain image: multi-byte
  *structured* fields (record headers, tables) are stored in **big-endian byte
  order**, while bit-packed data streams are raw. Evidence chain:
  1. `baserom.u.z64` is byte-for-byte identical to `data/ge007.ntsc-final.z64`;
     the ROM header is GE-specific (magic `80 37 12 40` from `src/rom_header.s`,
     BE u32 CRC at 0x10/0x14 written by `tools/n64cksum.c`) — not the standard
     N64 header layout.
  2. Build pipeline: `$(LD)` → ELF → `objcopy -O binary --gap-fill=0xff` →
     RareZip cdata compression (`data_compress.sh`) → `n64cksum`. No byte-swap
     step; the mips64-elf-gcc toolchain (SGI-style BE default) emits BE bytes.
  3. `tools/utils.h` documents the convention: `swap_bytes()` = "convert from
     v64 to z64 ordering", `reverse_endian()` = "convert from n64 to z64
     ordering" — the project's ".z64" is a word-swapped cart image (community
     ".n64" format). Real cart = bswap32(file).
  4. Animation-blob analysis: reading headers as BE u32/u16 yields self-
     consistent values (frame counts, bit widths 0–31, offsets that tile the
     blob exactly); LE reads yield garbage.

  **Per-field transform rule** (applied at load time to the `animation_data`
  segment): u32 fields → bswap32; u16 fields → bswap16; u8 fields → identity;
  bit-packed streams → identity. No single uniform word transform exists (mixed
  field widths).

  **Blob layout** (fully mapped; all 173 entries verified): each C array
  (`animation_table_ptrs1/2` in `src/game/initanitable.c`) is a sequence of
  20-byte records at the `PTR_ANIM_*` offsets. Record fields: +0x00 address
  (entries-segment offset, u32), +0x04 frame count (u16), +0x06 angle bit width
  (u8, used by `modelAnimReadBitsAsU16Angle`), +0x07 loop flag (u8), +0x08
  bitDescriptors blob offset (u32), +0x0C bitsPerFrame root motion (u16, used by
  `sub_GAME_7F06D2E4`: `scaled = unk0C * frame`), +0x0E frame size in bits
  (u16, used by `loadAnimationFrame`: `frameSize = unk0E >> 3`), +0x10
  bitStream blob offset (u32). Tail (+0x14..+0x3C) is unused by any code — the
  effective record size is 20 bytes (= `struct anim_entry`, 5×s32). Interleave
  rule: entry *i*'s payload [descriptors][stream] sits at
  [PTR_ANIM_{i-1}+0x14, PTR_ANIM_i); all regions are disjoint; stream size =
  ceil(frameSizeBits × frames / 8) — exact fit on every entry (fire_standing:
  544 bits × 106 frames → 252 bytes). Descriptor array = `ModelAnimBitField`
  {u16 bitOffset, u8 bitCount, u8 pad, u16 valueOffset} ×4 (6 bytes each); after
  the transform all 692 descriptor arrays pack bitOffsets sequentially from 0.
  Null entries use sentinel **1** (not 0) — `expand_ani_table_entries` skips
  `*var_v0 == 1`.

  **Validation.** Simulated the full transform in Python over all 173 entries:
  0 bad descriptor ranges, 0 non-sequential arrays. Root-motion simulation for
  fire_standing frames 0–3 (widths 6/7/6/0, bitsPerFrame=19) gives smooth
  per-frame deltas (x = 7, 10, 12, 11; y ≈ 1086–1088) — proving the stream
  bytes are identity-correct (no transform). The previously observed garbage is
  reproduced exactly: file bytes `[00 00 01 58]` at fire_standing_fast's bd →
  LE read 0x58010000 + base 0x702ad8c0 = **0xC82BD8C0**, the exact value seen
  under gdb. The "zeros at odd indices" observation was a gdb display artifact —
  the compiled `.data` array is dense (verified by objdump of
  build-pc/ge007.x86_64.exe).

  **Second root cause: x86-64 stride bug in `expand_ani_table_entries`.** The
  loops iterate with `s32** var_v0; var_v0++` — +4 bytes/iter on N64, but
  `s32*` is 8 bytes on x86-64. Verified in the compiled binary (`add $0x8,%rdx`)
  : only even-indexed entries are rebased; odd entries keep raw small offsets
  (e.g. fire_standing, index 1, would stay 0x144 → instant SIGSEGV on use).
  This pattern occurs nowhere else in the compiled set.

  **Fix design (implementation pending — this is the immediate task):**
  1. New port function `romdataFixupAnimationData(u8 *blob, u32 blobSize,
     const s32 *tableA, const s32 *tableB)` in `port/src/romdata.c` (+ decl in
     `port/include/romdata.h`): for each non-null (≠0, ≠1) offset in both
     tables — bswap the 20-byte record header per-field (u32: +0x00/+0x08/
     +0x10; u16: +0x04/+0x0C/+0x0E; u8: +0x06/+0x07 untouched); then transform
     the descriptor array at [bd, bs) step 6 (bswap16 on words 0 and 4 only).
     Guards: skip if bd==0 && bs==0; require bd < bs ≤ blobSize and
     (bs−bd) % 6 == 0.
  2. Call it in `alloc_load_expand_ani_table()` (initanitable.c ~line 265)
     **between** the `romCopy` and `expand_ani_table_entries` — one line +
     `#include "romdata.h"` (`port/include` is already on the include path).
     Mechanical ABI edit per non-negotiable #2; must run *before* expand
     (expand reads/writes the fields as LE after fixup).
  3. Fix the stride in `expand_ani_table_entries()` (initanitable.c:233):
     iterate with `s32 *` instead of `s32**` — mechanical ABI fix, semantics-
     preserving on N64 where it was already correct.

  **Open item.** The entries segment (per-frame joint angles; file offset
  0x124AC0, 0x169EC0 bytes) is left as identity for Phase 1.5 — not consumed at
  boot (only during rendering). Verify visually in Phase 2 when Bond first
  animates.

**D34** `ANIM_DATA_*` address placeholders break on x86-64. On N64 the
`animation_data` segment links at VMA 0, so `(s32)&ANIM_DATA_x` is a small
segment offset; on PC `&ANIM_DATA_x` is a high PE address and the game code's
ubiquitous s32 truncation yields garbage. Audit: all 82 uses are
address-taking only (`(s32)&…`, or through `ANIM_FRAC()` /
`ANIM_FRAC_MUL_FIRST()` in chraction.c/bondview2.c/title.c/
initBondDATAdefaults.c/initactorpropstuff.c) — no data dereferences, no
sizeof, no static initializers. Fix: on x86-64 each `ANIM_DATA_x` is
redefined as an lvalue at `g_pc_animdata_base + offset` where the base is the
PE image address with its low 32 bits zeroed (derived in `romdataInit()` from
a static probe; fail-fast if not 4 GiB-aligned), so
`(s32)&ANIM_DATA_x == PTR_ANIM_x` exactly. Files:
`assets/animationtable_data.h` (guarded macro branch, N64 externs kept under
`#else`), `port/src/romdata.c` (`g_pc_animdata_base` + init),
`CMakeLists.txt` (excludes `assets/animationtable_data.c` +
`animationtable_entries.c` on PC — the arrays are address-only placeholders;
the ROM is the data ground truth via romCopy). Verified under gdb:
`D_80030984 = ANIM_FRAC(ANIM_DATA_walking)` executes; the walking record at
`ptr_animation_table + 0x4018` has entry field `0x10177dcc`
(entries base `0x10124AC0` + `PTR_ANIM_ENTRY_walking`) and frame count 0x25;
after `initWeaponAnimGroups()` the derived globals hold sane floats
(4.59/11.25/16.71/16.19).

**D35** Music sequence table: pointer-width + endianness.
`musicSeqPlayerInit()` faulted in `romCopy()` with a NULL destination:
`tblSegmentSize = sizeof(RareALSeqData) * seqCount + 4` was garbage because
(1) `RareALSeqData.address` was `u8 *` — sizeof 16 on x86-64 vs the ROM's
8-byte records — and (2) the header `seqCount` and entry fields are stored
big-endian in ROM but read as LE (63 → 0x3F00 = 16128). Fix: `address` →
`u32` (`src/music.h`; sizeof 8 on both platforms) with `(void *)` casts at the
three use sites in `src/music.c`; new `romdataFixupMusicSeqTable()` in
`port/src/romdata.c` bswaps `seqCount` + per-entry `address`(u32)/
`uncompressed_len`(u16)/`len`(u16), called after both romCopys in
`musicSeqPlayerInit()`. Verified: the header-only (0x10-byte) copy logs the
expected `seqCount 63 exceeds blob capacity 1` clamp; boot passes the old
fault.

**D36** Music heap / PERMANENT pool sizing on x86-64. `alCSPNew()` SIGSEGV'd
because the music bump heap was oversubscribed: x86-64 libaudio runtime
structs are larger (ALVoiceState 88B, ALSeqPlayer 248B, ALCSPlayer 224B, …)
and the debug-token state (`tokenFind(1,"-level_")==NULL` →
`g_DebugAndUpdateStageFlag=1` → stale `-ml0 -me0` tokens) forces the
fixed-size mempool branch with PERMANENT at 296 KiB. Measured under gdb:
pre-music PERMANENT usage 0x1BCA0 left 189,280 B for music vs a measured init
demand of 0x31660 (shortfall 13,056). Fix (PC-only guards): 
`MUSIC_ALLOCATION_BYTES` 0x2E000 → 0x32000 (`src/music.c`); PERMANENT fixed
branch 296/308 KiB → 320/340 KiB (`src/memp.c`, non-JP/JP). STAGE absorbs the
difference. Verified: boot reaches `bossMainloop()`.

**D37** libaudio bank trees must be re-laid out, not patched in place. The
audio thread SIGSEGV'd in `__initFromBank()` (`b=0xffffffff00000000`):
`ALBankFile/ALBank/ALInstrument/ALSound/ALWaveTable/…` are serialized in ROM
as big-endian scalars + 4-byte packed table-relative offsets, while the x86-64
C structs use 8-byte pointers — so BE values were LE-misread and 4-byte offset
arrays were walked with 8-byte stride, producing wild rebase pointers in
`alBnkfNew()`/`_bnkfPatch*()`. In-place conversion is impossible: each struct's
expanded tail overlaps whatever follows it in ROM (e.g. the ALWaveTable book
slot at +24 lands inside the ALADPCMBook that sits 12 bytes after the
wavetable). Fix: `romdataAudioBankPcSize()` + `romdataFixupAudioBank()` in
`port/src/romdata.c` — a two-pass DFS re-layout into a compact image where
every sub-struct is placed once (8-byte aligned, children before parents) and
each pointer slot stores the sub-struct's NEW offset from the image start,
zero-extended; `alBnkfNew()`/`_bnkfPatch*()` then rebase unmodified
(`ptr + (s32)file`). `ALWaveTable.base` stays an offset into the separate
wavetable data segment — `_bnkfPatchWaveTable()` adds its `table` argument,
verified at runtime (sfx: base+table = 0x102F19A0 =
`_sfxtblSegmentRomStart`). `music.c` PC branch allocates the re-layout size
and fixups after romCopy. Memory: bank images grow +0x1460 (sfx) / +0xA70
(instruments); init-time heap demand measures 0x33530, so 
`MUSIC_ALLOCATION_BYTES` → 0x38000 and PERMANENT → 352/368 KiB (PC only). 
Verified under gdb: runtime tree matches ROM ground truth (SFX instCount=1,
soundCount=261, bendRange=200, sampleRate=22050; INSTR instCount=75; all
env/keymap/wavetable/book/loop pointers valid heap addresses; book
order=2 npredictors=1); 25 s soak with every thread alive. Implementation
bugs caught during bring-up: a double bswap (afRd* already decodes BE),
afWr32 initially wrote BE, and the size pass used an uninitialized visited
array (undercount) — all fixed.

**D38** Implicit function declarations truncate pointer returns on x86-64.
~72 game TUs call ~400 functions with no visible prototype (missing `#include`
of the declaring header). Under C11 an implicit declaration assumes
`int f()`: harmless on N64 (32-bit pointers) but on x86-64 it silently
truncates every pointer return to 32 bits — e.g. `tokenFind()` in
`set_mt_tex_alloc()` returned a low-32-bit "pointer" that faulted in
`strtol()`. Fix: a port-layer prototype shim, `port/include/pc_protos.h`
(398 declarations; one-shot generator `scripts/gen_pcprotos.py` — the committed
header is the source of truth, hand-adjusted after generation), declaring each
function with its TRUE return type and an empty parameter list — purely
additive (only the returned value's width changes). Anchored in
`port/shim/PR/ucode.h`, the LAST include of `<ultra64.h>`: anchoring earlier
(e.g. `gbi.h`) poisons libaudio.h because the bondtypes chain reaches
`snd.h → <PR/libaudio.h>` while ultra64.h is still mid-parse. C-only guard:
C++ fast3d TUs reach this header via `SDL_stdinc.h → <stdarg.h> →
port/shim/stdarg.h`, and pulling the bondtypes/bondconstants chain into C++
brakes on `struct ALSoundState*` in `src/bondtypes.h`; no C++ TU needs a
fix here. C11 gotcha: an empty parameter name list `()` cannot match a true
prototype with default-promoted parameters (`s8`/`u8`/`s16`/…), so the 11
such functions get full prototypes with their TRUE parameter types
(substituting `int` for `s8` is NOT compatible). Also: `src/bondconstants.h`
defines function-like macros `ntohl`/`ntohs` over `CharArrayTo32/16`, which
collide with MinGW `<winsock.h>` in TUs that parse both — neutralized in the
port shim, real host functions provided in `port/src/pc_netorder.c` with
winsock-compatible signatures (`u_long` is 64-bit on LLP64!). No game-TU
edits. Result: build clean, zero warnings; the old `strtol()` crash is gone.

**D39** Globalimagetable rebasing idiom breaks on x86-64. `texReset()` /
`texLoadFromDisplayList()` compute ROM-layout pointers as
`globalbank_rdram_offset + (u32)&sym` with
`globalbank_rdram_offset = (u32)pGlobalimagetable + 0xFE000000`; on N64
`(u32)&sym` = `0x02000000 + off` and the 0x02 base cancels the 0xFE. On PC
`(u32)&sym` is a truncated PE address → garbage DL pointers. Fix (in
`src/game/image_bank.c`, PC branch): an enum of all 49 segment offsets (17
Gfx display lists + 32 sImageTableEntry tables) and
`#define GIMG_OFF(sym) (0x02000000u + g_pc_gimg_off_##sym)` — exactly the N64
`(u32)&sym`; all 49 texReset() sites replaced. No changes needed in
gunfire.c / texLoad / texSelect: all game RAM is s32-safe (V1 view,
0x70xxxxxx < 4 GiB), so the u32 math and pointer casts stay lossless.
Also fixed the segment markers: the CSV asset `Globalimagetable.bin`
(0xAC8) is TRUNCATED — the real linker segment (`ge007.ld`: oddtextures.o
`.data`) spans 0x13F8 = Gfx DLs (0xAC8) + ITE tables (0x930); the CSV's
`rarewarelogo` entry is mis-split (the logo actually starts at ROM
0x29E560). `_GlobalimagetableSegment{Rom}End` 0x1029DC28 → 0x1029E558 in
`port/src/romassets_u.s` + a size override in `scripts/gen_romassets.py`
(regeneration verified stable). All 49 offsets verified against the ROM:
the symbols tile `[0x29D160, 0x29E558)` exactly with no gaps; `gun.c`
ammo IconImage values cross-check. Verified under gdb: all 17
texLoadFromDisplayList diffs correct (0, 0x78, 0x120, …); boot passes
texReset. Known Phase-2 follow-up: `explosion.c`'s
`g_ExplosionDisplayLists[]` is a static table of EXE addresses to Gfx arrays
(16-byte entries on PC) — it must point at the ROM-layout copies in
pGlobalimagetable before explosions render.

**D40** N64-sized BSS placeholder pool overflows with pointer-stride structs.
`initModelHitEntryFreeList()` (initunk_005450.c:38, called from
lvlStageLoad) walks 600 ModelHitEntry records writing next/prev; the
decompiler-emitted BSS chain in objecthandler.c (`char g_ModelHitEntries[0xC];`
…dwords… `char g_ModelHitEntriesPenultimate[0x28]`) reproduces N64's
12 000-byte region (600 × 20 B), but PC ModelHitEntry is 40 bytes → the
free-list init writes 24 000 bytes, a 12 KB .bss overflow that clobbered
`is_ramrom_flag` (writer caught with a hardware watchpoint) → the
demo-replay path in bossMainloop ran with `address_demo_loaded == NULL` →
SIGSEGV at ramromreplay.c:341. Fix: PC branch declares a properly sized pool
(`char g_ModelHitEntries[600 * sizeof(ModelHitEntry)]`); the N64 chain stays
verbatim under `#else`; the sentinel assignment in initunk_005450.c is
computed directly on PC (`entries[LEN-1].prev = &entries[LEN-2]` — N64's
g_ModelHitEntriesPenultimate labels the start of entry 598). Verified: no
more SIGSEGV; boot progresses into GL rendering.

**D41** Cross-thread GL context binding. WGL allows a context to be current
on only ONE thread at a time. The window+context are created and made
current on the host main thread (`videoInit → gfx_sdl_init`), but all
rendering runs on the game's scheduler thread; while main still holds the
context, `wglMakeContextCurrent` from the game thread fails with "The
requested resource is in use" — and this SDL2 build (2.32.10, MSYS2)
swallows it: `SDL_GL_MakeCurrent` returns true on failure / 0 on success
(an inverted int-as-bool ABI artifact; confirmed by a minimal repro and
gdb `$1 = 255`). The silent failure left `wglGetCurrentContext() == NULL`
on the game thread → `glCreateShader` returned 0 → "Vertex shader
compilation failed" with an empty info log. Fix: new
`gfx_sdl_release_context()` (`SDL_GL_MakeCurrent(NULL, NULL)`) called at the
end of `videoInit()` after `set_swap_interval` (the last GL work on main);
the scheduler thread re-binds per frame via the existing
gfx_sdl_make_context_current() in videoStartFrame(). Do NOT branch on the
MakeCurrent return value (unreliable). Verified: frame 1 renders (69.7 ms),
several frames before the next (separate) fault.

**D42** rsp.c task-settings toggle idiom truncates pointers.
`g_gfxTaskSettingsList = (GfxInfo_s*)((u32)list ^ (u32)&g_gfxTaskSettings[0]
^ (u32)&g_gfxTaskSettings[1])` toggles between two adjacent settings structs
by XOR — fine on 32-bit, but on x86-64 the u32-truncated PE addresses XOR to
garbage → the next frame's `((GfxInfo_s*)g_gfxTaskSettingsList)->cfb = …`
(fr.c:458) SIGSEGV'd. Fix: explicit toggle under
`#if defined(__x86_64__)` (`list == &[0] ? &[1] : &[0]` — same semantics,
the list is always one of the two); N64 line verbatim under `#else`.

**D43 (OPEN)** Model-file loading ABI mismatch. Stage object load faults in
`modelPromoteNodeOffsetsToPointers()` (model.c:5688). Model files are
ROM-serialized with N64 layout (ModelFileHeader 24 B, ModelNode 20 B,
4-byte pointer fields) but on PC are read as 8-byte-pointer structs (48 B
each): `load_object_fill_header()` derives RootNode from `numtextures`
at the wrong offset, and the `PROMOTE` rebase (`(u32)var + diff`) then
operates on misaligned fields. Needs a D37-style re-layout of the model
file image (header + switches array + texture table + node tree +
ModelRoData records). Phase-2 project; see §H. Reference implementation:
PD port's `port/src/preprocess/filemodel.c` — same PROMOTE idiom, same vma
0x5000000 (§2.4); adapt + per-field validation, not a drop-in copy.

**D44 (closed 2026-08-22)** Crash-handler Phase 2 backtrace self-faulted before printing.
`crashStackTraceSym()` in `port/src/crash.c` calls
`GetCurrentThreadStackLimits(low, high)` passing the NULL pointer *values*
rather than `&low, &high` (Phase 1's `crashStackTraceRaw` passes them
correctly). The API writes the stack limits to address 0x0 → access violation
*inside the SEH filter* → Windows terminates the process without a second
dispatch. Symptom: `ge007.crash.log` contains only Phase 1 (EXCEPTION/PC/
MxCsr/thread-stack/MODULE lines) and no `BACKTRACE:` section — verified twice
live; this is why every D3x fault so far cost a full gdb launch-mode session.
Fix: one line, `GetCurrentThreadStackLimits(&low, &high)` — applied.
The build already compiles with `-fno-omit-frame-pointer` (CMakeLists.txt:193),
so every crash auto-logs an EBP-chain backtrace to console + ge007.crash.log,
symbolicable offline with addr2line. **Verified live:** the D43 run now logs
`BACKTRACE:` with #00 = `modelPromoteNodeOffsetsToPointers` (model.c:5688) and
#01 = `load_object_fill_header` (objecthandler_2.c:110) — the documented chain,
no gdb needed. Note: frames beyond the true call chain can be stale stack data
(the walk validates fp against stack bounds but not ret_addr against .text;
one such frame observed at #02). addr2line takes the full absolute address
directly (`addr2line -e build-pc/ge007.x86_64.exe -f -C 0x14007a31a`); image
base is still 0x140000000 (re-verified post-rebuild via nm).

**D45 (OPEN — sizing for D43)** Model-file buffers must grow for PC: the
GDL region doubles (8 B Gfx slots → 16 B) and texture-marker expansion in
texLoadFromGdl() emits full RDP setup sequences whose size is data-driven.
Per-file worst-case final size P_final = B_pc + 2×(D_n64−g1) + 16×Σ_markers
(K_t−1), where K_t is the exact worst-case command count per marker type t
(tex.c helpers, maxlod≤8, all texTry* state guards emitting, valid=FALSE —
the realistic branch since D_800483C4 is a gunfire texture index):
type0/LOD 37, type1/DETAIL 46, type2/MIPMAP 36, type3/TILE 18,
type4/TILE_PRESWAP 15 (preamble = PipeSync + gSPTexture ≤ 2; water check ≤ 1).
Marker type = w0&7 of the 0xC0-top-byte slot. Verified over all 512 model
files (build-pc/d43_sizes.py). Texture *pixel* data is unaffected by Gfx
width, so texpool sizes can stay at their N64 values — only the model-file
regions grow. Required buffer edits (all D40-class ABI-forced size constants,
semantics unchanged): gun.c:106 size_item_buffer 0x14820→**0x23000**
(bondview body+head+held-prop chain worst 0x1DB9A; suit pool 0xA0B0 + R
0x18000 = 0x220B0); gun.c:109 D_80032464 0x7530→**0xF000** (GautoshotZ
0xE788); gun.c ITEM_SUIT_LF_HAND R 0xBD70→**0x18000** (Csuit_lf_handZ
0x16F9C) with pool expr size−0x18000; gun.c ITEM_TRIGGER/ITEM_WATCHLASER R
0xAFD0→**0x17000** (GtriggerZ 0x16030); front.c load_walletbond R 0xA000→
**0x17000** (PwalletbondZ 0x1664C; stays below the +0x28000 DL region);
front.c cast screen bufferRemaining 0x18160→**0x1C000** (cast chain worst
0x19CA0 — zbuf at 0x19000+region must not be clobbered); initmenus.c:34 logo
buffer 0x78000→**0x7C000** (texpool 0x19000 + region 0x1C000 + zbuf
440×330×2=0x46B80). Unchanged: title/gunbarrel chain (0x16DF0 ≤ 0x23A00),
front.c logos (≤0x9BCC ≤ 0x3C000), dst=0 fileLoad path (allocates ALL
remaining STAGE space and gives the tail back via
mempAddEntryOfSizeToBank — stage props/NPC bodies are bounded only by live
STAGE headroom, ~2.3 MB). The converter will also assert P_final ≤ R at load
time with a clear error instead of overflowing.
**D46 CORRECTION:** the cast-screen values above are superseded —
front.c bufferRemaining is **0x25000** (not 0x1C000) and the initmenus.c
logo buffer is **0x85000** (not 0x7C000); see D46.

**D46 (RESOLVED — overlap safety + final buffer sizing for D43)**
The GDL expansion in sub_GAME_7F0762E0 is in-place: output starts at the
first-GDL offset and grows rightward; input is read from the mirror copy
texCopyGdls() made at the region tail ([A+R−P_conv+g, A+R)). The initial
gap is R−P_conv and the cumulative output-minus-input excess is ≤
16·Σ(K_t−1) over markers processed so far (monotone; 0xba "texture
already set" skips only reduce it). Since Gfx is 16 B for BOTH input and
output on PC, the no-clobber condition **R ≥ P_conv + 16·M_actual** is
identical to the fit condition — one constraint, not two. (The N64 game
satisfies the same identity with 8-B slots.) The converter therefore sets
`poolRemaining = P_conv = B_pc + 2×(D−g1)` EXACTLY (the pre-expansion
image; markers are expanded at runtime, consuming [P_conv, R)).

**Strict bound from N64 ground truth.** The N64 game works, so per file
8·M_a ≤ R_share_N64 − B_n64 − E. PC slots are 2× wide with the same M_a
(tex.c logic + texture data identical), hence
**P_final_actual(PC) ≤ B_pc + 2×(R_share_N64 − B_n64)** (E cancels).
N64 R_share values (game code): suit 0xBD70, trigger/watchlaser 0xAFD0,
wallet 0xA000, weapons 0x7530, bondview chain 0x14820
(size_item_buffer), cast chain 0x18160 (front.c bufferRemaining).
For chains ΣR_share = bufferRemaining EXACTLY: each subsequent load's R
is the previous file's post-load poolRemaining via
`get_pc_buffer_remaining_value()` (front.c:7809-7825), so the shares
tile the N64 budget with no margin assumption beyond "N64 works".

**tex.c expansion state machine** (for reference / any future
simulation): `sub_GAME_7F0CC4C8()` resets g_TexTileStates[8],
g_TexTileSizes[8] and g_TexLutMode at the START OF EVERY
texLoadFromGdl call → texTry*/texSetLutMode dedup only works WITHIN one
GDL. Per-marker slot counts (lutmodeindex=0): type0/LOD ≤ 2 + 6
(LoadToTmemAddr: SETTIMG+SETTILE≤1+LoadSync+LoadBlock+PipeSync) + 4
(TileFromDef: PrimColor+LUT≤1+SETTILE≤1+SETTILESIZE≤1) + 3×min(maxlod,7)
(TileLods basetile=1) + 3 (CycleType/TxLOD/Detail); type2/MIPMAP ≤ 2 + 6
+ 3×(maxlod + [maxlod==1]); type1 ≤ 2 + 6 (Zero) + 1 (TileSync) + 6 + 4
+ 3×min(maxlod,7) + 3; type3 ≤ 2+6+4+4; type4 ≤ 2+6+4. With LUT formats
the LoadToTmem* helpers double to ≤12 — see next item.

**Texture scan (build-pc/d43_lutscan.py), all 512 files:** 1071 valid
texture-table refs (TID < MAX_TEXTURES=3001; the rest are skipped by
texLoadFromModelFileHeader, and markers whose texnum was never loaded get
tex==NULL → PipeSync only). Image headers parsed via imagelist.u.csv
(order = assets/images.def order): **NO LUT textures** (formats 9-12)
exist in any model file — only format 0 (×1069) and 8 (×2) — so
lutmodeindex=0 universally and the ≤6-slot LoadToTmem bound holds.
maxlod distribution {0: 1063, 6: 7, 7: 1} → texWriteTileLods emits
NOTHING for 99% of markers. The D45 K_BOUND table {0:38, 1:46, 2:36,
3:18, 4:15} is verified a true worst-case bound under these conditions.

**Strict-bound results (build-pc/d43_chainbound.py):** suit_lf_hand
0x1002C ≤ 0x18000 ✓; trigger/watchlaser 0xED20 ≤ 0x17000 ✓;
walletbond 0x105BC ≤ 0x17000 ✓; worst weapon GmapZ 0xE91C ≤ 0xF000 ✓;
bondview chain (worst body spicebond + headbrosnan + autoshot) 0x1CE5C
≤ 0x23000 ✓; **cast chain: rifle 0x240DC / pistol 0x23F24 > 0x1C000 —
SHORT by ~0x8000** (worst: body spicebond B_pc=0x8A14/B_n64=0x7D60 +
head headbrosnan 0x1D90/0x1D50 + rifle autoshot 0x4828/0x3E28 / pistol
wppksil 0x45B0/0x3DC8). The D45 worst-K estimate (0x19CA0) was below the
strict bound because it assumed Rare's N64 buffer had zero margin.

**Resolution (two one-line constant changes, both already PORT-guarded):**
front.c init_menu18_displaycast bufferRemaining 0x1C000 → **0x25000**
(covers 0x240DC); initmenus.c logo buffer 0x7C000 → **0x85000**
(texpool 0x19000 + region 0x25000 + zbuf @ALIGN64(0x3E000) size
440×330×2=0x46B80 → 0x84B80, rounded). MEMPOOL_STAGE headroom is ample
(total stage usage at menu init ≈ 0xD0040 vs ~2.3 MB pool). The converter
should additionally assert `P_conv + 16×Σ_markers(K_t−1) ≤ R` at load
with a clear error (K_BOUND table; provably ≥ actual expansion).

**D47 (RESOLVED — D43 converter contract finalized; capstone endianness fix)**
Session that closed every open question on the model-file converter.

1. **Capstone endianness bug (environment gotcha).** All earlier-session
disassembly of the ROM used capstone's default LITTLE-endian mode against a
big-endian MIPS image → garbage register flow. MUST use
`CS_MODE_MIPS32 | CS_MODE_BIG_ENDIAN`. Compaction + accessors were re-disassembled
cleanly; where BE disassembly is still ambiguous, the byte-matched C in `src/`
is ground truth.
2. **Compaction state values (BE disasm of 0x7F0762E0 + byte-matched C).**
P = entry.poolRemaining, R = entry.rom_remaining, **delta = R − P** (not
P−R); `texCopyGdls(F+G, F+R−P+G, P−G bytes)`; per-GDL count = off_{i+1}−off_i
bytes (or P−off_last); final fileSetSize = ((rep&0xFFFFFF)+0xf)&~0xf.
Both accessors (0x7F0BD11C / 0x7F0BD100) compute `0x80090000 + idx*20` →
**resource_lookup_data_array actually lives at 0x80090000**; the ob.c comment
(0x800888B0) is stale.
3. **memp allocator (src/memp.c).** Bump allocator; `mempAddEntryOfSizeToBank`
rewinds `pool->pos` for the most recent allocation →
`fileSetSize(reallocate=1)` returns the post-compaction tail to the bank.
`load_resource` decompresses INTO ptrdata (=F), reading the compressed source
from the block's TAIL. Fresh dst==0 load: P := S_bank (all remaining stage
bank) → F := alloc(S_bank) → R := S_bank → load_resource sets P := D_N64.
4. **Reload hazard + PC decision.** `fileIndexLoadToBank` takes S_bank only
when poolRemaining==0; fileSetSize leaves P=R=post-compaction size, so a
reload within one stage would alloc(P_old) which can be < D_N64 (source
pointer before the block; decompress past it — latent on N64 too). PC fix:
PORT-guarded reset of `poolRemaining = 0` immediately before
`_fileNameLoadToBank` in load_object_fill_header → every load gets fresh-
bank semantics: staging space = whole remaining bank, steady-state bank
usage identical (fileSetSize rewinds to the same place).
5. **G_VTX w0 encoding (ROM-data proven).** GBI1/gDma1p style:
w0 = 04<<24 | dst<<16 | (16·n), dst = ((n−1)<<4)|v0. All 2826 model-file
instances have v0=0, n≤16 (batches of ≤16 verts — matches G_TRI4's 4-bit
indices). fast3d's `gfx_sp_vertex(C0(0,16)/sizeof(Vtx), C0(16,4), …)` is
correct AS-IS; the converter bswaps w0 only (no field remap). The
F3DEX_GBI_2 `gsSPVertex` block in GE's gbi.h is a red herring — the asset
toolchain emits GBI1 style.
6. **G_TRI4 = standard 4-bit indices.** The "out-of-range seg-5 w1" values
found by scanning are G_TRI4 index data, not addresses (w1 bits 20-31 hold
the last two 4-bit indices). **Opcode-aware remap rule: only {G_VTX=0x04,
G_SETTIMG=0xFD, G_LOADBLOCK=0xF3} carry addresses in w1.** Never remap
TRI4/TRI1/TEXTURE/SETOTHERMODE/CLEARGEO/SETGEO/syncs.
7. **Render-time segment bindings (model.c).** seg5 (COL1) = BaseAddr = F on
every model path; seg4 (VTX) = the record's Vertices base (or a runtime
buffer in dorottex). Hence G_VTX seg4 w1 = displacement from that array →
NO remap needed (array order preserved); G_VTX seg5 w1 = absolute file
offset — 2804/2805 verified to land inside a vertex array at vo+d with
d%16==0 → remap via the unified region map.
8. **Vertex format (all 512 files).** Normal: bswap s16 x,y,z,index,s,t;
bytes @C-F raw (rgba/normal are single bytes). Collision (34,580 verts):
bswap s16 x,y,z,index; LinkedTo u32@8 is ALWAYS 0 or 0x05xxxxxx (node vma);
CollisionRelatedIndex s16@C (range −1..113) + reserved s16@E bswapped.
9. **No embedded texture blobs in any model file** — texconfig TextureID
seg-5 count = 0 across all 512 (the earlier "3 title files" note was the
vestigial logo SETTIMG refs already covered by d43_cover.py). No blob
handling needed in the model converter.
10. **Zero-count vertex arrays.** PexplosionbitZ is the ONLY file with
nv=0 and non-null Vertices (0x98); its GDL uploads 16 verts from there via
absolute seg-5 ref. Rule: emit the array sized up to the next object offset
(→ [0x98,0x198) = exactly 256B here).
11. **GDL tight packing is safe.** fast3d `case G_ENDDL: return` — trailing
junk after ENDDL is never executed at render time. Spans: 2064 tight,
238 with trailing bytes (1904B total). Converter emits up to and including
ENDDL; each PC span = exactly 16·slots → compaction counts are exact.
12. **Reference validator: `build-pc/d43_convert.py`** implements the full
conversion spec (DFS walk with LOD/SWITCH rewiring, region map, per-opcode
record conversion, remap checks) and runs it on all 512 files:
**ALL CLEAN** — every pointer remap resolves, layout invariants hold,
max D_PC = 0xB7E0, max D_PC/D_N64 ratio = 1.31 → staging headroom is
trivially satisfied vs the ~2.3 MB stage bank. This script IS the spec for
the C implementation.
13. **Final converter contract.** Emit `[switches NS×8B LE][texconfigs NT×12B
(bswap TextureID; 0x05xxxxxx would remap — never occurs)]` → nodes+records in
DFS preorder (PC layout via struct assignment; every promoted pointer field
emitted as zero-extended u64 of 0x05|new_off; Primary/Secondary GDL ptrs
remapped but NOT "promoted"; BaseAddr emitted 0) → vertex arrays immediately
after their record (PointUsage = 2×numVertices s16 after op24 CollisionVerts)
→ **GDLs LAST, contiguous, visit order, 16B LE slots**: w0'=bswap32(w0);
w1' = bswap32(raw) EXCEPT for seg-5 of {0x04,0xFD,0xF3} → remap low24 via
the region map; **no LSB set** (ROM convention; fast3d's extended seg_addr
case handles unmarked segmented addresses). Stage at F+R−D_PC, memmove to
F. Set `poolRemaining = D_PC` exactly (never touch rom_remaining).
14. **N64 record field offsets** come from the bondtypes.h comments (which
preserve N64 offsets); all 14 opcodes present in ROM files — 1,2,4,8,9,10,
12,13,15,18,21,22,23,24 — verified against RSZ sizes. PC layout: assign
through the real structs (compiler packs); don't hand-compute.
15. **PointUsage** = 2×numVertices s16 entries, indexed by MAIN-vertex index,
chain terminated by −1 (chr.c:3309-3331).

**D48 (REVIEW — D43 re-plan: offline pre-conversion "Plan B"; process fix)**
Independent review session that audited the D47 handoff against code + ROM
before implementation. Verdict: the remaining task was mis-scoped as a 1:1 C
port of a script that does not emit bytes; a cheaper, lower-risk path exists
that reuses the existing ROM load chain. Plan B is now the default; Plan A
(D47.13 C converter) is the fallback.

1. **`d43_convert.py` does not emit bytes.** It computes the layout/region map
and validates that every pointer remap resolves (512/512 clean, re-run and
confirmed this session), but there is no emission pass in ANY language. The
byte-level contract (PC struct assignment, per-field bswaps, GDL slot writing)
is unimplemented and unvalidated. "Port d43_convert.py 1:1 to C" would have the
next session write ~250-300 lines of new emission code in C and debug it at
runtime (crash → backtrace cycles). The emission must be written once either
way — do it in Python where iteration is seconds, validate offline, then ship
the data.
2. **Plan B: offline pre-conversion through the existing ROM load path.**
Generate 512 PC-layout RZ sidecar files with a Python emit pass; serve them by
patching `file_resource_table[i].hw_address` +
`resource_lookup_data_array[i].rom_size` from the port layer. Verified
mechanical facts (all re-checkable in <1 h):
   - **RZ format is trivially reproducible**: 2-byte header (`0x11 0x72`) + raw
deflate; `decompressdata()` (port/src/rzdecomp.c) skips the header and inflates
with a generous avail_in bound. A sidecar `[0x11 0x72][raw-deflate(PC image)]`
works through `load_resource` UNMODIFIED — and it sets `poolRemaining = D_PC`
automatically (the decompressed size), so no fixup call is needed at all.
   - **`romCopy` on PC is a host memcpy**: src/ramrom.c → `osPiStartDma` → port
shim `piServiceDma` (port/src/libultra.c), gated only by
`romdataCartAddrValid()` (port/src/romdata.c:212). The ROM is VirtualAlloc'd at
CART_BASE 0x10000000 sized romSize — extend the reservation by the sidecar
total, place sidecars at [CART_BASE+romSize, …), extend the validity check.
   - **`file_resource_table` is a plain writable global** (included in
src/game/ob.c:22). ALL 512 model loads funnel through
`load_object_fill_header` (dst==0 → `_fileNameLoadToBank`; custom-buffer
callers in front.c/gun.c/bondview2.c/initmenus.c → `_fileNameLoadToAddr`) —
both read `hw_address` via `load_resource`. No game code dereferences C*/G*/P*Z
symbols directly outside the table (verified by grep; symbols exist only in
the table + romassets_<r>.s markers).
   - **The patch must run AFTER `obInit()`** (called at src/boss.c:179):
obInit computes `rom_size` from adjacent-entry hw_address DELTAS (ob.c:122), so
a pre-obInit patch would corrupt rom_size. Lazy one-shot at the PORT hook site
in load_object_fill_header is simplest — by first model load, obInit has
definitely run.
   - **Footprint**: 512 files = 1,277,088 B compressed total (1.2 MB),
3,289,344 B decompressed N64 total; PC decompressed ≤ 1.31× per file (D47.12).
   - **The indy path is NOT used** (`resource_load_from_indy`, ob.c:56): it is
the N64 host-protocol loader gated by `indy_ready` (src/game/indy_comms.c,
dormant on PC), and its pPayload placement underflows on reload when
poolRemaining == pc_size exactly. Table patching reuses the proven ROM path.
   - **What Plan B eliminates**: the C converter (~300 lines), two-pass staging
+ memmove, staging-space guard, and the "emission bug only visible at runtime"
risk class. What it keeps (all already committed): ABI edits, fast3d seg-5
case, D46 buffer sizing — plus the one-line poolRemaining=0 reset (item 3).
   - **Sidecars are region-specific** (derived from the region ROM); generator
must take the region and write `data/pcmodels-<region>/`.
3. **The poolRemaining=0 reset is STILL needed under Plan B.** fileSetSize
leaves P=R=post-compaction size (ob.c:346-347); a same-stage reload would then
alloc S_bank' = P_old, which can be < round8(compressed)+8 → `load_resource`
hits the `source − ptrdata < 8` branch → poolRemaining=0 → silent load
failure. The D47.4 reset before `_fileNameLoadToBank` covers both plans.
4. **Plan A flaw (if runtime conversion is kept):** D47.13's staging guard
"fail if D_PC > avail" is too weak — the staging region [F+R−D_PC, F+R)
overlaps the live N64 image [F, F+D_N64) whenever S_bank < D_N64 + D_PC, and
emission would then read corrupted bytes. Correct guard: `D_N64 + D_PC ≤
avail`. Low probability in practice (files load at stage start; ~2.3 MB bank
vs ~82 KB worst case), but the stated guard silently misses it.
5. **Process fix:** the 23 d43_*.py investigation scripts (incl. the reference
converter) lived in gitignored `build-pc/` — moved to tracked `tools_pc/`
and committed, so the spec is versioned and reviewable. d43_cover.py's one
internal path reference updated; converter re-run from new location: ALL CLEAN.
6. **Review checklist for the next session** (confirm each item against code +
ROM before executing Plan B; record results as D49): see docs/HANDOFF.md
§Task 1 — eight claims, each with a falsification criterion. If any fires,
fallback to Plan A with the corrected guard from item 4.

**D49 (REVIEW — Plan B verification: D48 checklist R1–R8 all CONFIRMED)**
Independent review session executed the D48.6 checklist against code + ROM
(NTSC `data/ge007.ntsc-final.z64`). **Verdict: 8/8 confirmed — Plan B is
cleared for execution (HANDOFF Task 2); no fallback to Plan A.** Per item:

1. **R1 RZ format — CONFIRMED.** `decompressdata()` (port/src/rzdecomp.c)
skips the 2-byte header and raw-deflates (`inflateInit2(-15)`, avail_in ≤
0x400000). All 512 model files (C*/G*/P*Z in
`assets/obseg/file_resource_table.inc.c`; table↔`scripts/filelist.u.csv`
set identity exact, 512=512) start `0x11 0x72` and inflate cleanly.
Σ round8(compressed) = **1,277,088 B — exactly D48's claimed number**; max
single-file compressed 0x421E; max decompressed/compressed ratio 4.75.
2. **R2 romCopy is a host memcpy with no 0x10C00000 limit — CONFIRMED.**
`romCopy` → `doRomCopy` → `osInvalDCache` (no-op shim) + `osPiStartDma` →
`piServiceDma` (port/src/libultra.c): the only gate is
`romdataCartAddrValid()` (port/src/romdata.c:212), then plain memcpy. The ROM
is VirtualAlloc'd at CART_BASE 0x10000000 sized exactly romSize
(romdata.c:156-159); nothing in the port layer bounds the region, so placing
sidecars at [CART_BASE+romSize, …) needs only a reservation-size extension +
validity-check extension in romdataInit.
3. **R3 the table is the single chokepoint — CONFIRMED.** All 512 model loads
funnel through `load_object_fill_header` (src/game/objecthandler_2.c:89) →
`_fileNameLoadToBank`/`_FileNameLoadToAddr` → `fileIndexLoad*` →
`load_resource`, which read `hw_address`/`rom_size` from
`file_resource_table` + `resource_lookup_data_array`. The C*/G*/P*Z symbols
are `.set` markers in `port/src/romassets_<r>.s` (gen'd by
`scripts/gen_romassets.py`); no game code dereferences them outside the table
(grep: all hits are string literals/comments). Callers audited: bondview2.c,
chr_b.c, ejectedcartridges.c, front.c, gun.c, loadobjectmodel.c.
4. **R4 obInit runs once, before any model load — CONFIRMED.** `obInit()`
is called exactly once at src/boss.c:179 inside one-shot
`bossInitMainthreadData()` (before the infinite main loop); it computes
`rom_size` from adjacent-entry `hw_address` DELTAS (ob.c:122). `rom_size`/
`hw_address` are touched only in ob.c. The D43 crash stack was post-obInit.
A lazy one-shot table patch at the port hook site in
`load_object_fill_header` is therefore safe.
5. **R5 buffers fit — CONFIRMED (quantified).** Fresh dst==0 loads: max
round8(compressed)+8 = 0x4228 ≪ STAGE bank (max 0x24C400 = poolArea 0x2A4400
− me 352 KiB NTSC; ≥ ~1.3 MB even after conservative pre-model stage usage —
BG stan ≤ 0xA3E0 per stage). Worst-case post-compaction size across all 512
files (D_PC + 16·Σ(K_t−1) with K_BOUND {0:38,1:46,2:36,3:18,4:15}) =
**0x16F74** (Csuit_lf_handZ) — 25× headroom vs a fresh bank. dst!=0 callers:
every buffer ≥ 0xF000 > 0x4228; `tools_pc/d43_chainbound.py` re-run against
the CURRENT constants all pass: suit 0x1002C≤0x18000, trigger/watchlaser
0xED20≤0x17000, wallet 0x105BC≤0x17000, worst weapon GmapZ 0xE91C≤0xF000,
bondview chain 0x1CE5C≤0x23000, cast rifle/pistol 0x240DC/0x23F24 ≤ 0x25000
(front.c:7795), logo 0x85000 (initmenus.c:38). `poolRemaining := D_PC` is
automatic (`load_resource` ← `decompressdata` return, ob.c:61).
6. **R6 poolRemaining hazard + reset — CONFIRMED.** Mechanism verified in
current code: `fileIndexLoadToBank` (ob.c:219-247) allocates
`poolRemaining` when non-zero (the post-compaction P left by fileSetSize),
and `load_resource` (ob.c:49-53) needs round8(rom_size)+8 ≤ bytes or it sets
poolRemaining=0 → **silent load failure**. The data CAN trigger it (e.g.
PlegalpageZ: round8(C)+8 = 0xFC8 > D_PC = 0x5F0; its own path is a custom
buffer so it is safe, but the condition exists for any file reloaded in the
same stage). Hazard window = same-stage reloads only: poolRemaining is zeroed
at stage entry/exit (boss.c:415-417 / 639-641), and all consumers of
`get_pc_remaining_buffer_for_index` / `get_pc_buffer_remaining_value` run
post-load+compaction; the PROMOTE walk (sub_GAME_7F075A90) does not touch
lookup data. The addr path is unaffected (bytes = caller's fixed buffer).
The one-line poolRemaining=0 reset before `_fileNameLoadToBank`/
`_FileNameLoadToAddr` in `load_object_fill_header` is **pending** (not yet in
tree) and closes the hazard for both plans.
7. **R7 per-region sidecars + directory matching — CONFIRMED (with notes).**
Region is build-time-fixed: `build-pc.sh` ROMID → CMake ASSET_REGION {u,e,j}
→ `romassets_<r>.s`; the loaded ROM token (`ge007.<romid>.z64` /
`baserom.<r>.z64`) uniquely determines the region and romHeaderValid enforces
the country byte, so romdataInit can derive `data/pcmodels-<region>/` with the
same token logic (baserom.u → ntsc-final, baserom.e → pal-final). Name sets
are NOT identical across regions (filelist: u=512, **e=465**, j=512) — the
generator must map whatever files exist, and the patch loop only patches names
present in the manifest. Notes: (a) romdataInit has **no JP candidate**
(GE007_IS_PAL comes from versioninfo.h.in) — PC cannot boot a JP ROM today,
so JP sidecars are moot until that is added; (b) missing sidecar directory →
warn + continue in ROM-only mode (Task 2 step 2).
8. **R8 emission-spec completeness — CONFIRMED (with one explicitness gap).**
Walked every `ModelRoData_*Record` for the 14 opcodes present in ROM
(1,2,4,8,9,10,12,13,15,18,21,22,23,24) in src/bondtypes.h: every field is
covered by Verified facts + D47.13 (pointer-promotion list D47.5; f32 bswap;
TextureID bswap32; vertex arrays D47.8; GDL slots D47.13). A fresh C probe
(mingw gcc with the EXACT flags from build-pc/compile_commands.json)
re-verified all 15 PC record sizes + ModelNode (0x30) + ModelFileHeader
(0x38) + Vertex (0x10) and every pointer-promotion offset — **no drift** vs
d43_convert.py's PC_REC / the spec. Gap: Verified facts enumerate bswap for
f32, vertex-array s16s and the TextureID u32, but do not explicitly enumerate
the **u16/s16 record scalars** (AnimPart, MatrixIndex, JointID, MatrixIDs,
Group1/2, RwDataIndex, op4 numVertices@0x10, op24 nv/ncv/ModelType/
RwDataIndex). Data check: these hold small BE values (JointID 1–11, nv ≤
73) that raw emission would corrupt (LE read → ×256); all research tooling
reads them BE (`bu16`/`be16`). The rule follows from the D33 per-field
endianness convention (u16 → bswap16), so this is an explicitness gap, not a
missing rule — but Task 2's emit pass must bswap16 every u16/s16 record
scalar (padding/reserved may be zeroed), and the HANDOFF Verified facts
should be amended to say so.

**D50 (RESOLVED — Plan B executed: offline sidecars + port plumbing; boot advances to first model GDL execution)**

Execution session ran HANDOFF Task 2 end-to-end. All 512 NTSC model files
converted offline and served through the existing load path; frames render
and model display lists execute. Sub-items (all `#ifdef PORT`-guarded or
port-layer only; N64 build untouched):

1. **D50.1 Emit pass (`tools_pc/d43_emit.py`, tracked).** Per file:
decompress the N64 image from the ROM (filelist row); build the node map;
run the EXACT `modelIterateDisplayLists` visit simulation (LOD/SWITCH
rewire, BSP splice — d43_gdlorder logic) → gdl_seq; layout `[switches
NS×8][texconfigs NT×12][DFS nodes 48B + records PC_REC + vertex arrays]
…[GDLs packed contiguously in gdl_seq order, 16B per N64 slot]`; byte-exact
emit (bswap32/16; promoted pointers → zero-extended u64 `0x05xxxxxx` VMAs;
GDL Primary/Secondary raw VMAs, NOT promoted; BaseAddr=0); **round-trip
re-parse validation of every field against the N64 source + region tiling
(no gaps)** — 512/512 pass. Compression: `0x11 0x72` + raw deflate level 6
(`zlib.compressobj(6, DEFLATED, -15)`); escalate to 9 only on a dst!=0 fit
violation (none occurred). Cross-checks all pass: per-file dst!=0 buffers
(`round8(C)+8 ≤ buf` AND `D+round8(C) ≤ buf` — inflate overlap safety), G*
hand-weapon worst case vs 0xF000, cast/title/bondview chain cumulative
P_final bounds, totals (Σ round8(C) = 1,277,088 B, matches D49). Output:
**single concatenated image** `data/pcmodels-<region>/pcmodels.bin`
(sidecars at 16-aligned offsets) + `manifest.csv` (`name,offset,size`,
decimal — the C parser uses strtol base 10; file_resource_table.inc.c
order). Deviation from D48's per-file-sidecar sketch: one blob + manifest;
pcmodels.c copies the whole image to `[CART_BASE+romSize, …)` and patches
`hw_address = cartBase+romSize+off`. Regenerate:
`python tools_pc/d43_emit.py [ntsc-final|pal-final|jpn-final|--check-only]`
(needs the region ROM in data/; sidecars are gitignored with data/).
2. **D50.2 Port plumbing (`port/src/pcmodels.c` + `port/include/pcmodels.h`,
new; romdata.c/h extended).** `pcmodelsReserveSize(romImg)` derives
`data/pcmodels-<region>/` from the ROM country byte (+0x3E), parses the
manifest, returns total bytes (0 → warn, ROM-only mode); romdataInit
reserves `romSize + sidecarTotal` at CART_BASE and
`romdataCartAddrValid` accepts the extension; `pcmodelsLoadSidecars(
cartBase, romSize)` copies the blob; `pcmodelsPatchTable()` — one-shot,
called from `load_object_fill_header` (hook block, objecthandler_2.c) after
obInit has run — redirects every manifest row's
`file_resource_table[i].hw_address` + sets
`resource_lookup_data_array[i].rom_size` to the PC compressed size. The same
hook block resets `poolRemaining = 0` for dst==0 loads (closes the D48.3/R6
reload hazard). Boot log: `[INFO] pcmodels: table patched (512 model
entries)`.
3. **D50.3 Language banks (runtime C fixup).** Banks carry a big-endian
offset table; `romdataFixupLangBank(blob, decompressedSize)` decodes it in
place. Called from language.c after each of the 7 langInit loads and lazily
per-id in the `langGetJpnCharPixels` paths (idempotent via poolRemaining
check).
4. **D50.4 Fonts (runtime C fixup).** `load_font_tables` PC branch:
allocate `romdataFontPcSize()` (PC C layout of struct font: kerning[169] +
chars[94] + glyph pixel data), romCopy the N64 size, then
`romdataFixupFont` re-lays out, shifting the pixel block below the expanded
char array; `pixeldata` fields left as relative offsets — the existing
`pixeldata += base` loop promotes them.
5. **D50.5 Legal-screen UB exposure (front.c).**
`constructor_menu00_legalscreen` reads an uninitialized pointer at a lookat
call — on N64 the register happens to hold a readable address and the
result is zeroed by `* 0.0f` anyway; the x86-64 compiler folds the UB read
to NULL → fault. PORT branch seeds it with `legalpage_text_array` (the
value assigned a few lines later); identical lookat.
6. **D50.6 `texCopyGdls` copies only w0 on PC — first model-render crash
(RESOLVED).** `Gfx` is a union whose last member is `long long int
force_structure_alignment`: 8 bytes on N64 (= the whole slot), but on x86-64
the slot is 16 bytes, so `arg1->force_structure_alignment = arg0->…`
copies only the low half. Compaction flow (sub_GAME_7F0762E0):
texCopyGdls mirrors the GDL block `[G,D)` to tail scratch `[B−D+G,B)`, then
texLoadFromGdl reads the scratch and writes expanded output back at
`[G,…)` via full-slot `*(out++)=*(in++)` — propagating the scratch's stale
w1s into the final GDLs. **Byte proof:** PlegalpageZ (NS=0, NT=5,
D=0x2638) loaded at 0x7012EA38; first 64 RAM bytes match the decompressed
sidecar exactly; the executed GDL at file offset 0x2488 has w0s matching
the sidecar exactly (0xB10000BA/0xE7000000/0xFD900000/0xE6000000) but RAM
w1s = 0x50362B58/0x66D73339/0xACAAB819/0x55BDF769 (stale mempool contents;
pad2 dwords also nonzero) vs sidecar w1s 0x0000A898/0/0x050012C8/0 — the
garbage G_SETTIMG w1 → `seg_addr` → OOB read in `import_texture_rgba16`.
Fix: `*arg1 = *arg0;` under `#ifdef PORT` (tex.c). Audit: tex.c is the only
`force_structure_alignment` use in game code (model.c:1510 is an unrelated
local 8-byte union); every other Gfx copy is full-struct. Post-fix:
PlegalpageZ's sub-DLs execute to completion; crash moves on (D51).

**D51 (RESOLVED — font `pixeldata` fixup wrote the pointer at the wrong offsets; frame-5 `import_texture_i8` SIGSEGV)**

The G_TEXRECT tile-0 upload's source was a **font glyph** (I8, loaded via
`gDPLoadTextureBlock`). Root cause in `romdataFixupFont`
(port/src/romdata.c): it wrote each char's pixel pointer at blob offsets
d+20/d+24, but on PC `struct fontchar.pixeldata` is a **u64 at char offset
+24** — the low word landed in padding and only half the pointer was
written, so glyph sources resolved to stale/invalid addresses → OOB read in
`import_texture_i8`. Fix: write the blob-relative offset into d+24 (low)
and zero-extend d+28 (high). Verified: glyphs load, SIGSEGV gone. The old
D51 hypothesis list (seg_addr / tile staleness) was wrong; its GE opcode
facts remain valid reference material (G_IMMFIRST=−65; DMA G_MTX=1 /
G_MOVEMEM=3 / G_VTX=4 / G_DL=6; IMM TRI1=0xBF … ENDDL=0xB8; GE extension
G_TRI4=0xB1, 8-bit packed indices — NOT an address carrier; RDP
pass-throughs G_SETTIMG=0xFD, G_SETCIMG=0xFF, **G_TEXRECT=0xE4 /
G_TEXRECTFLIP=0xE5** — GE-specific values, not libultra's 0x46/0x45;
width-1 stored in the 12-bit field, fmt I=4, G_TX_LOADTILE=7; main DL is
game-built per frame with gSP* macros; segment table via
`gMoveWd(G_MW_SEGMENT, seg*4, base)` → w0=(0xBC<<24)|(seg*4<<8)|6,
w1=data; fast3d's gfx_sp_moveword stores data as-is when ≥ 0x800000 else
+0x80000000).

**D52 (RESOLVED — `osGetCount` tick-rate mismatch → non-deterministic post-frame-2 hang)**

Plain runs sometimes hung after frame 2 (kernel watchdog: "no frame rendered
for 3006180 ms") instead of crashing at frame 5. Root cause:
port/src/libultra.c `osGetCount()` returned **microseconds** (1M/s), but GE's
pacing assumes the N64 RSP counter rate ≈ **46.5525 ticks/µs**:
`MAIN_LOOP_TICK_INTERVAL` = 387,937 ticks (boss.c: NTSC
`INTERVAL_INTER_MATH - 2688U`; PAL
`frameDelay*(CYCLES_PER_FRAME-6450)-(INTERVAL_INTER_MATH-3225)`), and
`waitForNextFrame()` (frametiming.c) waits for
`(elapsed+interval)/775875` ticks (NTSC frame = 775,875; PAL 931,050). At
1M/s the steady-state 16,667 µs between retraces < 387,937 → bossMainloop's
gate never passes → no DL built → hang. Frames 1–2 rendered only because
stage loading took >388 ms of real time. Fix: `osGetCount()` returns
`(u32)(((uint64_t)sysGetMicroseconds()*465525ull)/10000ull)` (rate derived
from GE's own constants: 775875/16666.67µs = 931050/20000µs); wraps every
~92 s like the HW counter. `osGetTime()` still returns µs (osSetTimer/
OSTime). PD's port has the same µs implementation but lacks GE's cycle-based
pacing gate — do not copy it blindly.

**D53 (RESOLVED: model RW-data pool addressing on PC; frame-5 SIGSEGV in `modelInitRwData`)**

Post-D51/D52, the frame-5 SIGSEGV moves to `modelInitRwData`
(model.c ~6131): first BSP node → `movl $0x0,(%rax)` right after
`call modelGetNodeRwData` — writing `visible=FALSE` through NULL. Two
compounding PC layout bugs, both in the D32 class:

- **D53.1 (applied; necessary but not sufficient): `Model.datas` word
  stride.** `RwDataIndex` values are **4-byte word offsets** into the
  RW-data pool (`modelCalculateRwDataIndexes` accumulates
  `len += sizeof(record)/4`; pool = round16(numRecords×4) bytes, allocated
  in modelmgrInstantiateModel(WithAnim)). On N64 `&data[index]` with
  `union ModelRwData **data` strides 4 B; on PC it strides 8 B → every
  non-zero index addresses the wrong record. Fix (3 files, all #ifdef PORT):
  `Model.datas` → `u32 *datas` (bondtypes.h — layout unchanged: single
  pointer field); casts in model.c (`modelGetNodeRwData` local + return
  `(union ModelRwData *)&data[index]`, parent-walk
  `data=(u32*)tmp->RwDatas`, `modelAttachPart`) and propobj.c:7301/14343.
- **D53.2 (root-caused; fix designed, NOT applied): `ModelSlot` /
  `AnimModelSlot` ↔ `Model` type-pun breaks on PC.** The game puns the slot
  structs and Model in both directions: `slot.unk08@8` ↔ `Model.obj@8` (the
  **in-use flag** — `modelInit`'s `objinst->obj = header` marks a slot),
  `slot.unk10@0x10` ↔ `Model.datas@0x10` (RW pool), `slot.unk02@2` ↔
  `rwdatalen@2`. On N64 all pointers are 4 B → offsets agree. On PC there is
  no pack pragma — natural alignment (probe-verified with the exact CMake
  flags, see vsize.c): **Model = 0xE8 B**: chr@8, **obj@0x10**,
  render_pos@0x18, **datas@0x20**, scale@0x28, attachedto@0x30, anim@0x40;
  the slot structs still have unk08@8 / unk10@0x10. (An earlier draft of
  this section said obj@0xC/datas@0x1C — that assumed packed layout and is
  wrong; the log evidence below only fits the natural-alignment offsets.)
  Consequences: (a) `modelInit` writes obj@0x10 — the free check reads
  unk08@8 (low word of Model.chr, never written) → slots are never marked in
  use → every model reuses slot 0; (b) the legal-screen logo model takes
  slot 0, then `update_menu00_legalscreen()` (front.c:1430) →
  `clear_model_obj(logoinst)` writes obj=NULL at @0x10..0x17 — exactly where
  `slot.unk10@0x10` lives; the next instantiation passes
  `modelmgrCanSlotFitRwdata` on stale Model.datas@0x20 (non-NULL) + real
  unk02=20 ≥ 17, then `rwdata = g_ModelSlots[i].unk10` = NULL →
  `modelInit(model, header, NULL)` → fault. **Log evidence** (TEMP D51 trace
  in model.c → d52rw.log):
  ```
  INST model=0x7020ac48 header=0x140141c40 numRecords=0  rwdata=0x7020b0f8 rwdatalen=20 lvreset=0
  INST model=0x7020ac48 header=0x140142700 numRecords=17 rwdata=NULL     rwdatalen=20 lvreset=0
  GND obj=0x7020ac48 datas=NULL idx=0 rwdatalen=-1 op=2 data=NULL res=NULL
  ```
  (same Model address both times = slot reuse; 20 = MODEL_SPARE_RWDATALEN
  spare-slot pool.)
- **D53.2 fix APPLIED and verified.** objecthandler.h under #ifdef PORT:
  both slot structs re-laid out on top of the PC struct Model — `unk02`@2,
  in-use marker `unk08`@**0x10** (low word of Model.obj), pool pointer
  `unk10`@**0x20** (Model.datas), then `char pad28[sizeof(struct Model)-0x28]`
  so each slot is exactly sizeof(Model) = 0xE8 (a full Model, including
  animInit's writes up to PC offset 0xE3, fits). Member names kept; N64
  layout verbatim under #else. model.c heap fallbacks bumped PORT-guarded:
  0x20 → sizeof(struct Model) (non-animated), 0xC0 → sizeof(struct
  AnimModelSlot) (animated). Verified with the d52rw.log probe: both INSTs
  get valid pools (rwdata=0x7020c778, non-NULL), all 17 records initialize,
  frame-5 modelInitRwData crash gone; game runs to ~frame 102 and past the
  Nintendo-logo transition. Note: the bump allocator's `pos` drifts with the
  new sizes → some slots land on 4-byte boundaries; misaligned 8-byte field
  stores are functionally fine on x86-64 (individual field accesses, no
  faults) — accepted without a PORT alignment bump in memp.c.

Size/offset probe: `vsize.c` at repo root (untracked scratch — delete before
commit) prints sizeof + field offsets of Model / ModelSlot / AnimModelSlot /
struct player; compile with the CMake include order (port/shim first, then .,
include, include/PR, src, src/game, src/libultra, port/include) plus
`-DPORT=1 -DAVOID_UB=1 -std=c11`. Re-run after any slot-struct edit.

Environment notes: standalone gcc needs
`/c/msys64/mingw64/bin:/c/msys64/usr/bin` on PATH (cc1 fails SILENTLY without
it — exit 1, zero diagnostics), and `-std=c11` is required (the CMake flag;
under the default gnu23 `typedef s32 bool` in bondtypes.h breaks). Also:
`include/stddef.h`'s body is `#if 0`'d — offsetof/size_t are unavailable in
game TUs (use pointer-difference arithmetic in probes).

**D54 (RESOLVED: cseq ALCMidiHdr endianness; audio-thread SIGSEGV in
`__getTrackByte`)**

After D53, the first music load (M_INTROSWOOSH, seq 44) crashed the audio
thread in `__getTrackByte`. Root cause: a decompressed compact-sequence file
starts with `struct ALCMidiHdr` — 16 **big-endian** u32 trackOffset values +
a BE u32 division. N64 reads them natively; on an LE host alCSeqNew() does
not swap them (track-0 offset 0x44 becomes 0x44000000), builds "valid"
curLoc pointers ~1 GB past the buffer, and the first track-byte read faults.
The rest of the stream is byte-oriented (varlens, MIDI bytes, BE loop offsets
assembled byte-by-byte) so only the 17 header words need fixing. Fix: port-
layer `romdataFixupCseq(u8 *blob)` (port/src/romdata.c — bswap32 of the first
17 u32s; declared in port/include/romdata.h), called under #ifdef PORT after
each of the 3 `decompressdata` calls in musicTrack1Play/2/3 (src/music.c).
Verified: __getTrackByte crash gone; full 60 s run (exit=124 timeout) with no
audio fault, past the Nintendo-logo transition.

**D54b (RESOLVED: synthesizer param-slot sizing; audio-thread SIGSEGV in
`alLoadParam`)**

Next audio crash: `alLoadParam` dereferenced a corrupted free-list slot. Root
cause: alSynInit's "build the parameter update list" allocates
`c->maxUpdates` slots of `sizeof(ALParam)` and the game type-puns those slots
as several AL*Param structs. On N64 all of them are 0x1C bytes (one slot
each); on x86-64 **ALStartParamAlt is 0x28** (two 8-byte pointers: next +
wave) vs ALParam = 0x20 — every start-voice update wrote `wave` into the
neighbouring slot, corrupting the free list. Fix in synthesizer.c alSynInit
under #ifdef PORT: allocate `maxUpdates * sizeof(ALStartParamAlt)` and stride
the init loop by that size (cast each slot to ALParam* for the next-pointer
link); N64 verbatim under #else. Cost +8 B/slot × 0x80 slots = +1 KB vs
MUSIC_ALLOCATION_BYTES — no alHeapAlloc failure observed. Verified:
alLoadParam crash gone; full 60 s run clean.

**D55 (RESOLVED: RLE folder-menu background header endianness; SIGSEGV in
`rle_expand_8bit`)**

After the audio fixes, the game ran to the gun-barrel intro (~frame 654,
~23 s) and crashed in `rle_expand_8bit` (src/game/rle.c:30, the `*dst++`
store). Caller: title.c `sub_GAME_7F008DE4` (the initializeGunBarrelIntro path)
romCopies the asset at
`unknown2` (romassets_<r>.s, NTSC cart 0x102A4D50, size 0x1A580) and RLE-
decodes it into a 0x40400-byte buffer. Root cause: the asset is the title
folder-menu background; its raw ROM header is **big-endian** `01 B8 01 2B`
(w=440, h=299) + 6 pad bytes + a valid RLE stream (decodes to exactly
440×299 = 131560 bytes). rle_expand_8bit reads w/h as **LE** u16s (byte-
matched N64 code): the raw header gives w=47105, h=11009 → remaining ≈
518 MB written into a 256 KB buffer → SIGSEGV. The N64 build embeds this
asset into .data via `assets/romfiles2.s` (`.incbin
"assets/ge007.u.2A4D50.usedby7F008DE4.bin"`) from an extracted .bin whose
header is byte-swapped — title2.c hardcodes 440-wide I8 rows × 299, and this
is the only decodable 440×299 RLE stream in the ROM (a full-file 16-bit swap
was ruled out: it zeroes the first RLE count at +0xC), so only the 4-byte
header differs between raw ROM and the N64 .bin. Fix (port-layer, romdata.c
romdataInit, after the cart-base mapping): bswap32 the first word at
`(u32 *)&unknown2` in place — the image is a writable VirtualAlloc at CART_
BASE — guarded by "only swap if the LE-read w or h > 512" so an already-LE
region copy is a no-op. rle.c/title.c untouched (D37/D54 pattern). Verified:
RLE crash gone; game renders 600+ frames past the gun-barrel background into
the watch intro, where it hits D56.

**D56 (RESOLVED: watch-intro embedded Model/RW-pool raw offsets into struct
player; SIGSEGV in `modelSetScale`)**

Post-D55, the game renders 600+ frames (~10 s into the gun-barrel/watch
intro) then SIGSEGVs in `modelSetScale` (src/game/model.c:778,
`objinst->scale = scale`) with a garbage Model* (crash-log return frame was
corrupted; caller inferred — see below). Prime suspect, and the only code
passing a **raw N64 offset into struct player** as a Model*: `sub_GAME_7F07E7CC`
(bondview2.c:3102-3116, called from bondview2.c:3400 whenever the pause/watch
transition completes — every watch-menu open, and during the gun-barrel
intro):
```
animInit((Model *)((u8 *)g_CurrentPlayer + 0x230), itemheader, (u32 *)((u8 *)g_CurrentPlayer + 0x2ec));
modelSetScale((Model *)((u8 *)g_CurrentPlayer + 0x230), c_item_entries[41].scale * 0.1f);
modelSetAnimation((Model *)((u8 *)g_CurrentPlayer + 0x230), …ANIM_DATA_bond_watch…);
*(s32 *)((u8 *)g_CurrentPlayer + 0x220) = 0;   // = step_in_view_watch_animation
```
Layout facts (probe-verified where noted):
- struct Model (bondtypes.h:1482): 8 pointer fields (chr, obj, render_pos,
  datas [u32* under PORT, D53.1 — note the in-code comment there still says
  "D52", flagged for rename in HANDOFF Task 3], attachedto,
  attachedto_objinst, anim, anim2).
  sizeof_N64 = **0xBC**; sizeof_PC = **0xE8** (vsize probe: chr@8, obj@0x10,
  render_pos@0x18, datas@0x20, scale@0x28, attachedto@0x30, anim@0x40 — note
  the 8-byte pointer alignment padding after the two leading s16s).
- N64 struct player: the **watch Model is embedded at +0x230** (its first
  word is the anonymous s32 `something_with_watch_object_instance`), size
  0xBC, so it ends exactly at +0x2EC where the **RW-data pool** begins. The
  pool region runs to `buttons_pressed`@0x3B4 = **0xC8 bytes** of capacity;
  every field in [0x2EC, 0x3B4) is an anonymous s32 (no pointers → no extra
  PC shift inside the region).
- On PC the embedded Model sits at X = offsetof(struct player,
  something_with_watch_object_instance) (≥ 0x230 + 5×4: cameratile@0x34,
  prop@0xA8, bodyModel@0xD4, autoaim_target_y@0x130, autoaim_target_x@0x140
  are the pointer fields before it) and spans [X, X+0xE8). Remaining capacity
  before buttons_pressed is only 0x184−0xE8 = **0x9C < 0xC8** → the pool does
  NOT fit embedded on PC.
- modelInit stores the pool pointer in `Model.datas` and ALL rwdata access
  goes through `modelGetNodeRwData(model, node)` via `model->datas` — so
  redirecting the pool to separate storage is safe; only sub_GAME_7F07E7CC
  references +0x2EC directly (whole-tree grep).
- The N64 capacity 0xC8 bounds the watch model's real pool size (the game
  works on N64), so a fixed static buffer of that size is safe.
**Applied as designed + verified:** bondview2.c sub_GAME_7F07E7CC under
#ifdef PORT takes the Model by field name —
`Model *watch = (Model *)&g_CurrentPlayer->something_with_watch_object_instance;`
(probe-verified at +0x24C on PC — not 8-aligned; unaligned pointer stores are
fine on x86-64 per D53.2) and hosts the pool in `static u8 watchRwPool[0xC8]`
(N64 capacity; N64 embeds it at player+0x2EC); animInit/modelSetScale/
modelSetAnimation on `watch`; the +0x220 store becomes
`g_CurrentPlayer->step_in_view_watch_animation = 0;`. N64 raw-offset path kept
verbatim under #else. Verified: the watch path no longer crashes — but the
same `modelSetScale` SIGSEGV remained, and an env-gated probe (GE_D56,
logging `__builtin_return_address(0)` in modelSetScale) proved the real
caller was **not** the watch path: `initializeGunBarrelIntro` (title.c)
calling `modelSetScale(NULL, 0.18779343f)` because `setup_chr_instance()` →
`modelmgrInstantiateModelWithAnim()` returned NULL for BODY_Brosnan_Tuxedo —
see D57.

**D57 (RESOLVED: pointer-grown rwdata records overflow the N64-sized spare
pools; `modelmgrInstantiateModelWithAnim` returns NULL → SIGSEGV in
`modelSetScale(NULL, …)`)**

The Brosnan tuxedo's computed PC `numRecords` is **153** words vs the N64-
sized anim spare-pool capacity of **140** (`ANIM_MODEL_SPARE_RWDATALEN =
0x8C`). Cause: two rwdata record structs contain pointer fields and grow 8 →
16 bytes on x86-64 — `ModelRwData_HeadPlaceholderRecord` (ModelFileHeader* +
void*) and `ModelRwData_DisplayList_CollisionRecord` (Vertex* + Gfx*). Since
`modelCalculateRwDataLen()` accumulates sizeof(record)/4 per node, every
HEAD/DLCOLLISION node adds +2 words vs N64. Fix (two parts, both #ifdef PORT):
(1) initunk_005520.c: spare capacities grown with headroom —
`MODEL_SPARE_RWDATALEN 0x14→0x38`, `ANIM_MODEL_SPARE_RWDATALEN 0x8C→0xA8`
(N64 values kept under #else); (2) model.c: in the non-LvResetting branches
of both `modelmgrInstantiateModel()` and `modelmgrInstantiateModelWithAnim()`,
a dynamic slot+pool fallback mirroring the existing LvResetting path
(`mempAllocBytesInBank(sizeof(struct ModelSlot/AnimModelSlot))` + 16-aligned
pool of numRecords words) — the slot is untracked (never reused), acceptable
because with the grown capacities it should not trigger. A u32-field approach
for the two pointer records was rejected: `ModelFileHeader` pointers there are
exe-resident globals (>0x80000000 on PC) and would truncate. Verified:
Brosnan gets a dynamic slot, `modelSetScale` succeeds, game proceeds to
rendering — where it hits D58.

**D58 (RESOLVED: gun-barrel DL — K0 vertex-pointer idiom + 16-byte Gfx
overflow of the N64-sized reservation; SIGSEGV in `gfx_sp_vertex`, then FATAL
"Unknown GBI opcode 0x00")**

Two distinct PC-layout breaks in the same buffer (initializeGunBarrelIntro,
title.c), both found via env-gated probes (GE_D57: per-command VTX/CALL/JMP
log + entry-time hexdump of the barrel DL):
- **Part A (vertex pointer):** title.c passed `barrelDisplayListPtr +
  0x80000000` to sub_GAME_7F01BFF8, which embeds it verbatim in each G_VTX
  w1 (GE's gDma1p writes `(uintptr_t)(v)` — no LSB). On N64 the mempool
  pointer was physical, so +0x80000000 gave the RSP-visible KSEG0 address;
  on PC it is a V1 pointer (0x70xxxxxx, dram.c) and +0x80000000 lands at
  0xF0xxxxxx — unresolvable by fast3d's seg_addr() → SIGSEGV reading the
  vertex array. Fix: rebuild the exact N64 value —
  `(Vtx *)(OS_K0_TO_PHYSICAL((void *)barrelDisplayListPtr) | 0x80000000u)`
  (→ 0x80xxxxxx; seg_addr passes it through to the KSEG0 mirror; segments
  7/8 are never registered, so the unmarked-segment path is skipped).
- **Part B (DL reservation):** on PC `sizeof(Gfx) == 16` — the union's
  trailing `long long` (gbi.h documents it: "except on 64-bit, where it is
  exactly 128 bit"), same class as D50.6. Both writers (`gdl++`) and fast3d
  (`++cmd`) advance by 16, so all game-written DLs are 16-byte-wide — but
  the barrel-DL reservation `bufferSize -= 0x100` was sized for N64's 8-byte
  Gfx. sub_GAME_7F01BFF8 emits 31 Gfx (2×VTX + 28×TRI + ENDDL) = 496 B, so
  slots 16–30 (second TRI batch + ENDDL) overflowed into the RLE region at
  +0x300, and sub_GAME_7F008DE4's expand then clobbered them with image data.
  At render the RSP executed VTX/TRI×14/VTX fine, then hit slot 16 = RLE
  pixels (w0=0x00000001 → opcode 0x00; fast3d has no G_SPNOOP case) → FATAL.
  Fix: reserve 0x200 under PORT. The 0x200 vertex reserve still fits
  (30 Vtx × 16 B = 0x1E0). NOTE for future asset work: any other N64-sized
  reservation for game-written DLs/vertex arrays must be re-checked against
  the 16-byte Gfx / 16-byte Vtx widths (recurring class, cf. D50.6/D53.2).
Verified: barrel DL executes to ENDDL; game proceeds past the gun-barrel
hole into model rasterization — where it hits D59.

**D59 (OPEN — current blocker: SIGSEGV inside an external GL DLL during the
first real model rasterization after the gun-barrel hole)**

Post-D58, the barrel DL runs clean (probe-verified: VTX@+0, TRI×14,
VTX@slot15(+0xF0 in 16-byte form), … ENDDL) and the game crashes shortly
after with EXCEPTION 0xc0000005 at PC 0x7ff8d42a44d3 — inside a DLL loaded
at 0x7ff8d4230000 (offset +0x744d3; almost certainly the OpenGL driver,
not yet confirmed). The crash-log backtrace frame #1 (main+0x227cc0) is a
BSS symbol (`memoryMesgMB`) — garbage stack, no usable caller. Draw path:
gfx_flush() (gfx_pc.cpp:299) → `gfx_rapi->draw_triangles(buf_vbo,
buf_vbo_len, buf_vbo_num_tris)` → gfx_opengl.cpp:813-816
`glBufferData(GL_ARRAY_BUFFER, sizeof(float)*buf_vbo_len, buf_vbo,
GL_STREAM_DRAW); glDrawArrays(GL_TRIANGLES, 0, 3*buf_vbo_num_tris)`, where
`buf_vbo` is the static `float buf_vbo[MAX_BUFFERED*(32*3)]`. Hypotheses:
(a) buf_vbo overflow while accumulating transformed vertices for the first
real model (Brosnan) — check the vertex-append site in gfx_pc.cpp (the
transform loop after ~line 1118) for a missing bounds check against
MAX_BUFFERED; (b) garbage buf_vbo_len/num_tris; (c) bad texture/shader state
on first model draw. Next: env-gated probe logging buf_vbo_len/
buf_vbo_num_tris at every gfx_flush + identify the DLL (PowerShell module
list during a run, or the GL vendor string in the log).

**D59 RESOLVED (sessions G–I).** The "external GL DLL" crash was not a
driver bug: it was an msvcrt.dll `memcpy` faulting on a wild source — the
gun-barrel sub-DL region at `ptr_logo_and_walletbond_DL + 0x200` had been
clobbered by the unbounded RLE write of D64 (below). With D64 fixed, the
barrel renders and the game advances; no fast3d/vertex-buffer change was
needed. The crash handler gained permanent improvements along the way:
FAULT ADDR (ExceptionInformation[1]), a 16-qword STACK@RSP window, and a
module list in `ge007.crash.log` (crash.c + psapi).

**D60 RESOLVED — DMA target validation (port layer).**
`osPiStartDma` (libultra.c) now validates ROM-read targets: the N64 PI can
DMA to any KSEG address, but on PC an unmapped target is a wild memcpy.
`dramHostAddrValid()` accepts DRAM V1/V2 and any host-committed region
(VirtualQuery), so legitimate `.bss`/`.data` targets (e.g.
`ramrom_data_target`) pass while s32-truncated wild addresses are rejected
with a logged FATAL instead of a silent crash. Also added the GE_D60
sidecar-read tracer and GE_D61 per-ROM-read log (`d61dma.log`).

**D62 RESOLVED — OSMesgQueue/OSScMsg layout (port layer).** The shim's
message-queue bookkeeping had to match the PC struct widths: OSMesgQueue is
40 bytes on PC (two OSThread* + 3×s32 + OSMesg*), OSScMsg stays 32 bytes.
Scheduler-thread message flow (retrace/pre-NMI/interrupt/cmd queues in
`os_scheduler`, g_AudioManager frame/reply queues) verified against those
layouts.

**D63 — TEMP diagnostics (to strip).** GE_D63-gated probes in gfx_pc.cpp /
blood_animation.c / front.c / rsp.c tracing the VTX-pool bump pointer
(`g_GfxMemPos`), the gun-barrel sub-DL slot word, dram-branch targets and
rspGfxTaskStart hand-off. Used to prove D64's clobber path and to rule out
VTX-pool overflow; no permanent change.

**D64 RESOLVED — blood RLE sentinel (src/game/blood_animation.c).**
The N64 build places `die_blood_image_end` in the same section directly
after `die_blood_image_1[]`; the RLE decoder's guard `bloodImgNxt <
&die_blood_image_end` relies on that adjacency (only the address is used).
On PC a zero-init symbol lands in `.bss` ~1 MB away, so the guard never
fires and the decoder writes unbounded past the array — it clobbered the
gun-barrel sub-DL at `ptr_logo_and_walletbond_DL + 0x200`, which is what
surfaced as the D59 "GL DLL" crash. Fixed under #ifdef PORT by defining
`die_blood_image_end` as one-past-the-end of the array.

**D65 RESOLVED — `enum HEADS` signed sentinels (src/bondconstants.h).**
The N64 toolchain gave this enum a signed underlying type, so
`HEAD_FIXED == -1` and `head >= 0` guards were real branches. PC GCC 16
picks `unsigned int` for enums whose enumerators are all non-negative
(0xFFFFFFFF > INT_MAX), making every `head >= 0` always true and turning
`c_item_entries[HEAD_FIXED]` into a wild 64-bit OOB read (SIGSEGV in
init_menu18_displaycast). Under PORT the sentinels are now negative
literals (`HEAD_FIXED = -1`, `HEAD_RANDOM = -97`) — identical bit pattern,
signed semantics restored.

**D65b RESOLVED — `enum BODIES` signed sentinel (src/bondconstants.h).**
Same class as D65: ROM tables store 0xFFFFFFFF in `body` fields and the
cast-end check compares `intro_char_table[f].body < 0`; PC's unsigned
underlying type deleted the reset branch, so the cast screen rendered the
terminator entry and `langGet(0)` dereferenced a NULL bank. Added
`BODY_FIXED = -1` under PORT (forces signed underlying type; no existing
value changes).

**D66 RESOLVED — romCopyAligned pointer width (src/ramrom.c/.h) +
ramrom replay truncations (src/game/ramromreplay.c).** The N64 build did
all of `romCopyAligned` in s32; on PC targets live in `.bss` above 4 GiB,
so `(s32)target` truncated (0x1401C6F00 → 0x401C6F00) and the DMA went to a
wild address. PORT version uses uintptr_t throughout and returns `void *`
callers assign straight to pointers. The ramrom replay path had the same
class of `(s32)` truncation on `ramrom_data_target`.

**D67 RESOLVED — struct image_entry layout (src/game/image.h).** The
decompiled field order cannot be right: texLoad() reads
`*(s32*)&entry & 0xFFFFFF` as the data offset (dataoffset must occupy bits
0-23 of word 0) while chrprop.c indexes entries with an 8-byte stride
(sizeof == 8). Under PORT the struct is re-declared with all-u32 bitfields
and `dataoffset : 24` first, so GCC packs it to exactly two words on both
targets and the raw word read is satisfied. The IMAGE() macro initializer
order in image.c is adjusted to match under PORT.

**D68 RESOLVED — Globalimagetable endianness (port/src/gimgfixup.c +
src/game/image_bank.c + src/game/image.c).** The ROM-copied Globalimagetable
segment (texReset) is N64 big-endian, but PC code reads its CPU-interpreted
u32 fields natively: the IMAGESEG-marked G_SETTIMG w1 words
(`IMAGESEG(id) = 0xABCD0000 | id`) and the `sImageTableEntry.index` field of
all 32 table arrays. Unfixed, texLoad computed texnum from byte-swapped ids
(e.g. 52651 for IMAGE_SMOKE_11 = 2106) → out-of-range offsets → a 925 KB
ROM read into the 4000-byte stack compbuffer (FATAL at boot). Fix:
`gimgFixupGlobalimagetable()` bswaps exactly those u32s in place after the
romCopy (17 Gfx DLs walked op-by-op for the AB CD marker; table entry
counts fall out of the D39 symbol layout, 12-byte stride); everything else
in the segment is byte-level (opcodes, single-byte fields, raw pixel blocks
referenced via 0x02xxxxxx segmented addresses) and untouched. Two
consequences handled under PORT: (1) texLoadFromDisplayList's marker scan
now checks bytes 6..7 (CD AB — the LE encoding of 0xABCDxxxx) instead of
4..5; (2) explosion.c executes the *compiled* globalDL_0xNNN shadows via
g_ExplosionDisplayLists[], so `gimgSyncCompiledGlobalDLs()` copies the
texLoad()-patched IMAGESEG w1 values from the ROM copy into those arrays
(command j of the 8-byte ROM DL maps to Gfx slot j of the 16-byte compiled
array; D39 verified them byte-identical). Verified: 137 texLoads with valid
in-range ids (2106, 2084, …), real offsets/sizes from g_Textures, and the
game runs the full ~3.5-minute intro (logo → gun barrel → cast) at ~59 fps
to the first stage load.

**D69 (OPEN — current blocker: BG-file big-endian headers at stage load)**

Post-D68 the game plays through the entire intro and crashes in
`load_bg_file` (src/game/bg.c:830) when loading the first stage (BUNKER1,
"bg/bg_sev_all_p.seg", cart 0x10438660). The header IS loaded correctly —
`obLoadBGFileBytesAtOffset` works: `&fileentry->hw_address[offset]`
evaluates to `hw_address + offset` (a valid cart address; the compiled
absolute asset symbols point into the ROM mapped at the cart base, and the
PI shim memcpys from there). The bug is interpretation: BG-file offsets are
N64 big-endian u32s in segment-0x0F form. Header word 1 in the ROM is
`0F 00 00 14` (BE value 0x0F000014 → file offset 0x14 after
BG_SEG_TO_PTR's `+ 0xF1000000` fold); PC reads it LE as 0x1400000F, so
`ptr_bgdata_room_fileposition_list = header + 0x1400000F - 0xF000000`
lands ~0x5 MB past the stack buffer and `...[1].pPointTableBin` faults
(EXCEPTION 0xc0000005, FAULT ADDR ≈ header + 0x500000F). The whole
stage-load path (bg .seg headers/room tables + Tbg_*_stanZ geometry files)
is riddled with BE u32 fields — the same class as D68 but a far larger
format surface. This is the "next asset type" milestone anticipated in
AGENTS.md. Strategy options: (a) offline per-region conversion of all bg/*.seg
+ Tbg_*_stanZ files into sidecars (the D43/Plan-B pattern; requires fully
decoding GE's BG/stan formats from bg.c/stan.c — note PD's
preprocess/filebg.c describes a *different*, zipped multi-section format;
same family ≠ identical, validate per field); (b) runtime port-layer fixup
after each load (same format knowledge, placed in port/). Either way the
first task is reverse-engineering the formats: header words 0..3 are
pointers (rooms/portals/bgcmds/lights-style tables per the D69 probe:
word1=0x14 room-fileposition list), bg_room_data records carry more
0x0Fxxxxxx offsets (pPointTableBin at record+0x28, see crash disasm), and
stanZ files go through stanDetermineEOF/stanLoadFile. TEMP D69 probe in
ob.c (GE_D69) logs name/index/rom_size/hw_address per BG load.

### G. Phase 2 status snapshot (HISTORICAL — frozen ~D74)

> Stale point-in-time snapshot. **Current status: `docs/HANDOFF.md`.**
> Finding detail: §F index above + §H below.

Done through **D68**: PD fast3d integrated (`port/fast3d/`); GE's real
`src/sched.c` + pthread kernel; dual-mapped DRAM; ROM mapped at cart base;
SDL2 window; full boot chain (D31–D42); **Plan B executed (D50)** — all 512
NTSC model files offline-converted to PC-layout RZ sidecars, served through
the existing load path via a port-layer table patch; runtime C fixups for
language-bank BE offset tables (D50.3) and font re-layout (D50.4);
legal-screen UB seed (D50.5); `texCopyGdls` w1 partial-copy bug fixed and
byte-proven (D50.6). **D51–D58 resolved** (font pixeldata fixup, osGetCount
tick rate, model RW-data pools, cseq BE header, synth param slots, RLE
folder-menu background, watch-intro raw offsets, spare-pool capacities,
gun-barrel DL idiom + reservation). **D59 resolved** (the "GL DLL" crash
was D64's unbounded blood-RLE write clobbering the barrel sub-DL),
**D60–D62 resolved** (DMA target validation, OSMesgQueue layout),
**D63** TEMP diagnostics, **D64 resolved** (blood RLE sentinel adjacency),
**D65/D65b resolved** (HEADS/BODIES enum signed sentinels under PORT),
**D66 resolved** (romCopyAligned + ramrom replay 64-bit pointer width),
**D67 resolved** (struct image_entry N64 layout reconstruction),
**D68 resolved** (Globalimagetable BE→LE fixup for IMAGESEG Gfx words +
sImageTableEntry.index; compiled globalDL shadows synced after texLoad).
The game now boots, plays the intro music, and renders the **entire intro**
(Nintendo logo → gun barrel with Brosnan → cast screen; ~frame 2100, ≈2 min
wall-clock at the current ~20 fps clean-run rate), then crashes in
`load_bg_file` on the first stage load —
**D69, the milestone blocker**: BG-file headers are N64 big-endian
(segment-0x0F offsets) and PC reads them LE. Details in §F/D59–D69.

**D70–D74 (intro-logo pixel work):** D70 env-gated PPM frame capture
(`GE_PCDUMP` → `./ppm/`) for numerical visual debugging; **D71 resolved** —
C-array texture sources (the four rarewarelogo.c RGBA16 images) were
byte-swapped on LE PC (pink/green logo); port-layer per-source bswap in
`import_texture`. **D72 resolved** the UV path (GE always uses authored tc[]
UVs; `lookat_enabled` defaults false). **D73 resolved** — root cause of
D72.3: sinf/cosf `du` double constants are big-endian word pairs, garbage on
LE PC → guMtxF2L emitted −32768 for every sin/cos entry → logo triangles
projected off-screen; DVAL() macro fix in guint.h/sinf.c/cosf.c under PORT.
**D74 resolved** — texture import fallback no longer truncates valid
gDPLoadBlock data (mip chains + sub-tiled textures), TextureCacheKey gains
`size_bytes`, and the VBO path now wraps UVs by tile size for WRAP sub-tiles
(N64 semantics). The logo now renders its four gold letters on the dark-blue
plate (PPM-verified at frame ~555); a final pixel-perfect comparison against
N64 reference footage is still open. **D69 remains the milestone blocker**
(stage load).

**Committed through D74**: D51–D74 fixes; this session's TEMP probes are
stripped, but previously committed TEMP diagnostics (D63 blocks, GE_D71LOG,
and the older D51–D66 leftovers) are still in the tree — strip list in
HANDOFF Task 3. Build is GREEN.

### H. Finding log continued (D32 procedure, D70–D121)

> This section began as a per-session "handoff & plan" and grew into the
> D70–D121 finding log. The **D32 repeatable fix procedure** below is
> permanent reference; the rest is the finding archive (use the §F index).
> **Current status: the README. Current task: `docs/HANDOFF.md`.** Session
> narrative: `docs/dev/HANDOFF-ARCHIVE.md`.

Full paste-ready brief: **docs/HANDOFF.md** (primary thread: D69 BG/stan
stage loading — the milestone blocker; secondary: final pixel check of the
now-rendering intro logo). For D69:
reverse-engineer GE's bg .seg + Tbg_*_stanZ formats from the decompiled
consumers (bg.c, stan.c), then choose offline sidecar conversion (Plan-B
pattern, D43) vs runtime port-layer fixup; PD's preprocess/filebg.c is a
reference for the *approach* only — its BG format is different (zipped
multi-section).
Summary:

**State.** D50–D74 resolved and verified (D73 = sinf/cosf endianness, D74 =
texture import fallback + sub-tile UV wrap); committed through the D74
milestone. The game boots, plays intro music, and renders the entire intro
(logo → gun barrel with Brosnan → cast screen) — the Rareware logo now shows
its four gold letters on the dark-blue plate (D71–D74; PPM-verified at frame
~555, final pixel-perfect check vs N64 footage still open). TEMP diagnostics
from earlier sessions are still in the tree (strip list in HANDOFF Task 3);
visual-debug tooling: `GE_PCDUMP` frame capture. Build is GREEN.

**Thread 1 — D69 (milestone blocker, now primary).** `load_bg_file`
(src/game/bg.c:830) faults on the first stage load (BUNKER1): BG-file header
words are N64 big-endian segment-0x0F offsets; PC reads them LE, so the
room-fileposition-list pointer lands ~5 MB past the stack header buffer.
Full analysis + strategy options in §F/D69. This is the next milestone:
decode the bg .seg + Tbg_*_stanZ formats, convert/fix for PC, get a stage to
load and render.

**Thread 2 — logo final check (low priority).** Compare the rendered logo
frames (ppm/, frame ~550–560) against N64 reference footage; if any letter
is still off, re-add a lightweight triangle/texture attribution probe (the
D74 probes are stripped; the import sizes and sub-tile wrap are verified at
the data level).

**After D69:** get a stage to load + render (bg .seg + Tbg_*_stanZ format
work), then continue the diagnose→fix→verify loop through gameplay; strip
TEMP diagnostics at each milestone; pixel-assert soak (PPM dump +
tools_pc/pixcount.py) once a stage is stable.

**D32 repeatable fix procedure** (apply to any ROM-serialized struct that faults
on a pointer-field read):
1. At the fault, `ptype /o <Struct>` in gdb. If a pointer field's offset/size
   diverges from the N64 offset comment (e.g. an 8-byte pointer where N64 has 4),
   it is this bug class.
2. Change the embedded pointer fields to `u32` in the struct (keeps N64 layout on
   x86-64). Add casts at every use site (`(T *)field`). Document as D3x.
3. Verify the load-time rebase/fixup writes valid **V1** DRAM addresses (< 0x80000000)
   into those u32 fields; if not, debug the fixup (see Part B below).
4. Rebuild and confirm the boot advances past this struct to the next init step.
5. If the struct is a *tree* of packed sub-structs (like the libaudio banks, D37,
   or model files, D43), an in-place BE→LE patch will not fit — the expanded
   8-byte pointer slots overrun the following ROM data. Re-lay it out into a
   fresh compact image and rewrite each pointer slot as the sub-struct's new
   offset (zero-extended), so the existing `ptr + (s32)base` rebase still works.
6. If the fault is NOT in ROM-loaded data but in an address arithmetic idiom over
   exe-resident symbols (`(u32)&sym`, XOR toggles, pointer-delta math — D39/D42),
   guard a PC branch that reproduces the N64 32-bit value exactly (e.g. keep the
   0x02000000 base) or replace the idiom with an explicit equivalent; N64 line
   stays verbatim under `#else`.
7. If two structs are **type-punned** (cast back and forth, e.g. ModelSlot ↔
   Model), verify their layouts still agree on PC — pointer-width changes break
   puns silently: fields the game reaches "through" one struct land at wrong
   offsets in the other (D53.2). Also check any fixed-size allocation that must
   contain the grown struct (the 0x20-byte model heap fallback, D53).
8. If a hang (not a crash) appears with no thread making progress, suspect a
   **timing/pacing gate**: GE's loop gates on `osGetCount()` deltas in N64 RSP
   counter ticks (~46.5525/µs), not µs (D52).
9. If the fault is a store through a pointer built from a **raw byte offset
   into a struct that contains pointers** (`(u8 *)ptr + 0xNNN`, D56): every
   pointer field before 0xNNN shifts +4 on PC (plus 8-byte alignment padding),
   so the offset no longer lands on the intended object. Fix under #ifdef PORT
   via the named field (`&s->field`); if the embedded sub-object has grown
   (sizeof_PC > sizeof_N64) and its trailing companion storage (e.g. a model
   RW pool packed right after it in the struct) no longer fits before the next
   live field, relocate that companion to a static buffer of the N64 capacity —
   safe when all access goes through a pointer stored in the sub-object at init
   (modelInit → Model.datas) and only the one site references the raw offset.

**D70 (dev tooling, TEMP):** env-gated PPM frame capture for numerical visual
debugging: `GE_PCDUMP="first-last[:step]"` dumps the bound FBO to
`./ppm/frame_NNNNNN.ppm` from `videoEndFrame()` (`gfx_opengl_pcdump_enabled()`/
`gfx_opengl_dump_bound_fbo()` implemented in gfx_opengl.cpp). Established the
intro timeline numerically (legal text f20–100, Rareware logo f380–580,
iris/gun-barrel f900–1220, red region f1240–1360, cast f1520+) and confirmed
the user's "pink and green" report at f480 (left half RGB≈(135,78,129), right
half (82,104,0)). Strip per HANDOFF Task 3.

**D71 (RESOLVED):** the Rareware logo rendered as two flat colors — pink-red
and bright green — instead of gold lettering. Root cause: the four RGBA16 logo
images in assets/rarewarelogo.c (`imgRAre_0x0020` etc.) are `u32` C arrays
compiled into the exe `.data`; on LE PC each N64 texel pair is stored as a
little-endian u32, so `import_texture_rgba16`'s big-endian u16 read produced
byte-swapped texels (raw 0xED0F gold → 0x4FCC green + 0x0FED pink — exactly
the two observed colors). All other texture sources are raw N64 BE byte
streams (ROM cart map 0x10xxxxxx, model sidecar 0x10Cxxxxx, KSEG0/V1 buffers
0x70–0x90xxxxxx) and must not be touched. Fix (port/fast3d/gfx_pc.cpp):
`gfx_tex_source_is_c_array()` classifies by address range;
`gfx_tex_normalize_source()` bswaps each u32 once per source into a stable
cached buffer (cache key stays the original address) and `import_texture()`
decodes from it. Verified: exactly two sources normalized in a full run (the
logo image banks); logo now renders gold on dark blue. One-shot per-source log
gated by `GE_D71LOG`.

**D72 (RESOLVED + D72.3 OPEN):** the logo UV path.
- D72.1: removed the PD-inherited normal/lookat-based UV overwrite in
  `gfx_sp_vertex()` — GE always uses authored per-vertex `tc[]` UVs: no GE
  code sets G_TT_BASE/G_TT_CLAMP (only G_TT_NONE), and stage geometry has
  gSPLookAt set every frame yet N64 textures are fixed to surfaces, so the
  normal-derived path is wrong for GE in general.
- D72.2: `rsp.lookat_enabled` now defaults false — N64 boots with RSP memory
  zeroed; no lookat exists until gSPLookAt writes one (the intro logo has
  none). GL UV convention confirmed as U.5 (texel×32) from the LUT/rect path,
  so DL_RAREWARETEXT's identity-scale corner UVs 0x0010..0x03F0 map to full
  32×32 coverage with a half-texel inset.
- D72.3 (OPEN): after D71+D72 the logo still does not appear: DBGTRI traces
  show ALL logo triangles (plate fan + letter quads) project off-screen
  (clip x=y≈−1.6e11, w<0, screen ≈(32776,32776)), yet the frame shows a large
  flat dark-blue (0,0,64) region — rows 150..479, cols 1..639 — filled as a
  PERFECT checkerboard (50/50 pixel parity), i.e. broken rasterization of some
  big triangle drawn after clear_framebuffer_black. Identity not yet
  established (`GE_DBGTALL` all-triangle trace captured; analysis pending).
  Ruled out: matrix-format mismatch — this codebase's guMtxF2L is Rare's
  modified variant writing the interleaved hi16/lo16 s32 Q15.16 packed format
  that gfx_sp_matrix decodes exactly (FTOFIX32=×65536); `D_8002A7D0` is a
  zero-init u32 so `[D_8002A7D0]`==`[0]`; alloc_intro_matrices() runs from
  initmenus. **SUPERSEDED by D73** — the off-screen projection was never a
  matrix-format or UV problem: guMtxF2L's sinf/cosf inputs were garbage on PC.

**D73 (RESOLVED — root cause of D72.3):** GE's `sinf`/`cosf`
(src/libultra/gu/sinf.c, cosf.c) build their double constants through the
`du` bit-union (`{ struct { u32 hi; u32 lo; } word; double d; }`) with
big-endian word pairs `{hi, lo}` (rpi, pihi, and the P[] polynomial tables).
On an LE PC the `.d` read has the two words swapped: `rpi.d` ≈ 2^733,
`pihi.d` ≈ 2^257. Range reduction then overflows: `dn = dx*rpi.d` huge →
ROUND(dn) saturates in cvttsd2si (n = −2^31) → `dx = dx − dn*pihi.d` ≈ 2^288
→ xsq ≈ 2^576 → the result overflows float → sinf/cosf return ±inf/NaN.
FTOFIX32's `(int)(±inf·65536)` is another cvttss2si saturation →
0x80000000 (−32768). Every sin/cos-derived entry of guMtxF2L became −32768,
so the Rareware logo's guRotate matrix was garbage and all its triangles
projected off-screen (the D72.3 symptom; the "checkerboard" was a separate
large triangle rasterizing over the empty logo region). The logo scene is the
first visually verifiable consumer of guMtxF2L sin/cos output in the intro;
other scenes that build matrices via guRotate/guLookAt are affected by the
same bug and were fixed by this change. Fix (narrow ABI exception — constant
interpretation only, algorithm verbatim): `duD()` static-inline + `DVAL(x)`
macro in src/libultra/gu/guint.h under #ifdef PORT re-pack hi/lo into the
correct LE double; the 5 `.d` reads in sinf.c and 4 in cosf.c now use
`DVAL(...)`. The N64 build is untouched (#else branch = original expression).
Only sinf.c/cosf.c use `du` (grep-verified); `fu` (single u32) constants are
endian-safe. Verified numerically: P decodes to the expected perspective
matrix, Rot(−40°Y) gives clean cos=0.766/sin=−0.643, and logo vertices clip
to small finite on-screen NDC values. (tools_pc/mtxtest.c is a standalone
scratch harness for the matrix-convention question that was ruled out along
the way — GE stores all matrices transposed; row-vector ≡ column pipeline.)

**D74 (RESOLVED — final pixel check pending):** with D73 in, logo geometry
was correct but the texture side was still wrong: a dim red-brown blob plus
a flat vertical bar at the right edge, one letter visible at a time. Three
port-layer bugs, all in port/fast3d/gfx_pc.cpp (with gfx_pc.h):
- **Import fallback truncation.** `import_texture`'s old condition
  `(rdp.tex_lod && tile >= rdp.first_tile_index + rdp.tex_detail) ||
  !loaded_texture.addr` overwrote valid gDPLoadBlock tmem data with
  `line_size_bytes * tile.height`. For the logo: letter mip chains 2744 B →
  2048 B (mips 1–5 dropped), and D_02005FF0's 32×32 → 192 B (32×3, stale
  tile.height=3 left by `gsDPSetTileSize(0, 46, 116, 124, 124)` in DL
  D_02004758). Fix: fall back only when `!loaded_texture.addr` (N64 TMem is
  persistent — a populated slot is the faithful source).
- **Cache poisoning.** TextureCacheKey lacked the upload size, so the first
  (truncated) import of an address poisoned every later import. Added
  `size_bytes` to the key (gfx_pc.h + both key initializers + the bucket-only
  aggregate in gfx_texture_cache_delete).
- **Sub-tile UV wrap period.** N64 wraps UVs by the TILE size when a render
  tile is a sub-region of the image; the port wrapped at the full uploaded
  image size (GL_REPEAT). D_02004758's 20×3 tile at offset (11.5, 29)
  therefore sampled row 0 instead of rows 29–31 — the flat bar's color
  matched texel [0,4] of D_02005FF0 exactly. Fix: per-vertex pre-wrap in the
  VBO path — for a WRAP sub-tile (`tex_width2 < tex_width`, where
  `tex_width2 = (lrs−uls+4)/4` is the tile window in texels),
  `u = fmodf(u, tw); if (u<0) u += tw; u += uls/4.0f` (same for v). Known
  limitations: half-texel edge bleed at the sub-tile window edges (GL
  bilinear neighbors may sample just outside [O, O+W)); mirror sub-tiles are
  not handled (the logo doesn't use them; audit if another scene regresses).
Verified: runtime import sizes correct (D_02005FF0 = 2048 B, imgRAre =
2744 B); PPM frame ~555 shows the four-letter RARE band with gold/warm
colors and per-letter segmentation; clean `GE_PCDUMP` runs are stable past
1156 frames. Final pixel-perfect confirmation against N64 reference footage
is still open (see HANDOFF Task 1). All env-gated probes added for D72.3/D74
(GE_DBGUV/GE_DBGTRI/GE_DBGTALL/GE_DBGMAT/GE_D74IMP/GE_D74DUMP/GE_DBGLOAD)
were stripped with this change; the previously committed TEMP D63 blocks and
the GE_D71LOG normalize log remain on the HANDOFF strip list.

**D75 (OPEN: 3D rendering — mispositioned / missing 3D models throughout the
front end).** Post-D74 the Rareware logo is correct (gold letters on dark
blue), but every other intro 3D element is still wrong. Confirmed symptoms
(user observation, 2026-08-28 session L):
- **Nintendo logo** renders but is **positioned wrong**.
- **Gun-barrel intro**: the **James Bond figure is entirely missing** (the
  animated walk-and-shoot character model). Barrel/spiral effect status not
  separately confirmed.
- **Intro credits / cast roll**: the per-character 3D models (each shown
  beside their actor/character name) **do not appear at all** — names draw,
  models do not.
Pattern: 2D/texture and text elements draw; **skeletal/animated character
models never appear**, and non-animated 3D (logos) appears but with a bad
transform. This strongly suggests **(b) below is the dominant bug** — the
animated-model path is broken independently of the matrix sin/cos fix. Two
candidate causes — distinguish before any fix:
- **(a) D73 scope gap.** D73 (`DVAL()` in `src/libultra/gu/guint.h`, PORT-only) was
  documented as fixing *"all scenes using guRotate/guLookAt-derived matrices … not just the
  logo."* If these elements still fail, either their matrix path bypasses that fix or a
  second coordinate defect remains.
- **(b) Separate model/RW-pool path.** The player models specifically may be on the
  `animInit` + embedded raw-offsets-into-`struct player` path (cf. D56), which is *not* the
  guMtxF2L sin/cos path — i.e. an independent bug, not "the same coordinate issue."

Files to check: `src/libultra/gu/guint.h`, `src/game/model.c`, `bondview2.c` (animInit /
modelSetScale sites), `port/fast3d/gfx_pc.cpp`. Verify: determine whether the Nintendo logo
and intro player models build matrices via guRotate/guLookAt (→ 75a) or another transform
(→ 75b); capture `GE_PCDUMP` frames across the logo transition and the gun-barrel/cast
segments to localize.

**D75 ADDENDUM (session M-32) — triage: (a) is DISPROVEN, the defect splits
into two independent bugs, both outside `src/libultra/gu` and `src/game/model.c`.**
Static-only pass (no runtime probe; ~1 build-budget spent on reading + rebase
onto M-30b master which carries D159/D164/D165/D166).

- **(a) D73 scope gap — RULED OUT, high confidence.** The whole
  `src/libultra/gu/` tree is endian-clean post-D73. Audited every file:
  `rotate.c` (`guRotateF`/`guRotate`), `perspective.c` (`guPerspectiveF`),
  `translate.c`, `scale.c`, `ortho.c`, `lookat.c`, `mtxutil.c` (`guMtxF2L`),
  `normalize.c`, `align.c`. Only `sinf.c`/`cosf.c` ever use the `du` bit-union,
  and those 9 reads are `DVAL()`-wrapped under `#ifdef PORT`; every other gu
  file uses plain `float` literals (`3.1415926/180.0` etc.) or pure integer
  bit-packing (`FTOFIX32` + shift/mask — byte-order independent). The Rareware
  logo exercises the *entire* gu matrix pipeline front-end
  (`guPerspective`+`guLookAt`+`guRotate`+`guTranslate`+`guMtxF2L`,
  `title.c load_display_rare_logo`) and renders correctly. So no front-end
  matrix path that goes through gu is broken. **Future sessions: do not
  re-audit gu for D75.** (Recorded in porting-notes.md §C.)

- **Bug 1 — non-animated `logoinst` models (Nintendo logo, GE logo,
  front.c wallets/Bond photo): wrong transform / 180° flip. Almost certainly
  the D114/D116 shared fast3d viewport/MP-matrix mirror — NOT in scope,
  NOT game code.** These do NOT use gu at all: `front.c` builds `basemtx`
  with `matrix_4x4_set_lookat_target(&m, 0,0,3000/4000, 0,0,0, 0,1,0)`
  (`matrixmath.c:643` → `matrix_4x4_set_lookat`), copies it into a
  `dynAllocate`d `render_pos`, calls `model.c subdraw`, then
  `matrix_4x4_f32_to_s32` in place (`front.c:1774/2023/2304/2929`, all
  identical). `matrix_4x4_set_lookat*` / `matrix_4x4_set_projection` /
  `matrix_4x4_f32_to_s32` are pure float/int math, native-LE, byte-identical
  in behaviour to the gu equivalents (D114 already verified this for the
  in-level path and disproved converter-axis / lookat-handedness / F2L as the
  cause). D159 (M-31) fixed the *texture* comb on the wallet photo and its
  addendum explicitly notes the residual **"still 180°-rotated"** — a pure
  two-axis negation, i.e. exactly the symptom D114's write-up predicts for a
  hidden X- or Y-flip in `gfx_calc_and_set_viewport` /
  `gfx_adjust_viewport_or_scissor` / the `MP_matrix` vertex transform
  (`gfx_pc.cpp` ~1122/1739/1771). Confidence the logo/photo transform bug is
  the D114/D116 fast3d flip: **medium-high**. This is a `port/fast3d`
  correctness gap — the fix needed is the one D2 of porting-notes.md gates
  behind a RenderDoc/apitrace capture of one texrect (or the asymmetric
  1-texel-texture experiment). No `src/` change will fix it and static
  tracing it is BANNED (porting-notes.md §D2). Parked exactly as before.

- **Bug 2 — animated character models (gun-barrel Bond, cast-roll models)
  entirely absent: category (b), an independent model-instantiation /
  `render_pos` lifetime bug, NOT a transform bug. Root cause not pinned
  (needs runtime probing).** Key discriminator: **in-level skeletal guards DO
  render as humanoids** (`-level_09`, M-12) through the *identical* joint
  path — `subcalcmatrices`→`modelUpdateMatrices`(`process_02_position` etc.)
  →`drawjointlist`→`modelRenderNodeDl`. So the joint math, `matrix_4x4_f32_to_s32`,
  and `gfx_sp_matrix` decode are all proven. What the front-end animated path
  does *differently* from the working in-level path:
  1. `renderData.mtxlist = dynAllocate(numMatrices << 6)` — the per-frame GFX
     arena (`dyn.c:140`, bump-allocated inside `g_VtxBuffers[0]` /
     MEMPOOL_STAGE emulated DRAM), then `instcalcmatrices`/`subcalcmatrices`
     set `model->render_pos = (RenderPosView*)arg0->mtxlist` (`model.c:2409`)
     and the joint matrices are written there and converted **in place**
     f32→s32 by the caller (`title.c:281`, `front.c:1568` …). D115 MED item
     **#5** already flags this exact mechanism: *"the D102 weapon-model
     `render_pos` is pointed at a `dynAllocate`'d transient arena that on N64
     aliased the persistent `hand->mtxlist`; likely the '1P weapon model
     doesn't draw' cause."* The gun-barrel Bond uses the same
     dynAllocate→render_pos pattern (`title.c:253/257`). If the arena is
     swapped / overwritten between the `f32_to_s32` write and the
     `gSPSegment(3, osVirtualToPhysical(render_pos))` bind in `subdraw`
     (`model.c:5346`) / `drawjointlist`, every joint matrix is garbage →
     model collapses to a point / projects off-screen → "absent".
  2. `chrModelInstance` / `gunModelInstance` are stand-alone `Model`s created
     for the intro (`title.c initializeGunBarrelIntro`), not the in-level
     `chr`/`prop` pool — check they are non-NULL, `obj->numMatrices > 0`, and
     the model file actually loaded (the `pcmodels` sidecar must contain the
     Bond/cast model files; a missing manifest row → `hw_address` unpatched →
     load serves N64-layout or fails).
  3. `modelSetAnimation(chrModelInstance, (ModelAnimation*)((s32)&ANIM_DATA_bond_eye_fire + (s32)&ptr_animation_table->data), …)` (`title.c:220`, and `title.c:542`).
     Looks like a truncation risk but is **probably already covered by D34**:
     the PC branch of `assets/animationtable_data.h` defines `ANIM_DATA_*` as
     an lvalue at `g_pc_animdata_base + offset`, so `(s32)&ANIM_DATA_*` is the
     bare offset, and `ptr_animation_table` points into low (<4 GiB, bit-31-clear)
     emulated DRAM, so the `(s32)` sum and the cast back to a 64-bit pointer are
     lossless. Verify anyway with a probe (cheap), but do not lead with it.
  Confidence this is category (b) and not (a): **high**. Confidence on the
  exact mechanism (1 vs 2): **low** — pick with a `GE_PCDUMP` + a capped
  env-gated probe logging `chrModelInstance`, `obj->numMatrices`,
  `render_pos`, `g_GfxMemPos`, and the first joint matrix at the gun-barrel
  frame. **Lead with mechanism 1** (`render_pos` = transient `dynAllocate`
  arena, D115 MED #5) — it also explains why the *in-level* path works (there
  `render_pos` is a persistent per-chr pool, not the gfx arena). An in-scope
  fix would give the front-end animated models a persistent `render_pos`
  buffer (`#ifdef PORT`) instead of the swapped arena, mirroring the D100/D102
  inline-pool pattern.

- **D144/D146 "seg5+0x9ee4 malformed sub-DL" is a DL-walk desync, not a
  transform bug** — "unresolved matrix pointer then garbage opcodes" is the
  D135 signature (a previous command decoded at the wrong width/offset
  desyncs everything after). seg5 = model COL1/BaseAddr (§11.7), so a G_MTX
  landing on seg5 means the walk is already lost, not that a matrix pointer
  is "unresolved". d43_emit.py's opcode-aware remap (§11.6) deliberately does
  **not** touch G_MTX w1; if a front-end model GDL span is emitted with a
  wrong slot count or a `G_TRI*` / GE tex-macro (`0xba`, cf. D124 Facility)
  is mis-sized, the following `G_MTX` reads garbage. This is
  `tools_pc/d43_emit.py` + `port/fast3d/gfx_pc.cpp` territory — outside this
  task's file scope. Recommend a dedicated converter/fast3d agent dump the
  raw bytes of the `logoinst`/`walletinst` model GDL at `seg5+0x9ee4` with
  `GE_DBG*` and diff the PC span against `build-pc/d43_convert.py`'s
  reference output for that file.

**Net for D75:** (a) closed. Bug 1 (logo/photo transform) folds into the
parked D114/D116 fast3d-mirror item. Bug 2 (absent animated models) is a
real, in-scope lead at `title.c:220` `(s32)&symbol` truncation +
`render_pos`=`dynAllocate` arena lifetime — next session should build with a
`#ifdef PORT` `(uintptr_t)` cast there and a capped probe, not a static pass.
Confidence overall: **medium** (triage solid, no runtime confirmation).

> **M-33 UPDATE — Bug 2 (gun-barrel Bond) is substantially a capture artifact
> (D168), and/or has since regressed away.** With the PPM writer fixed and a
> fresh bare-front-end capture (`GE_PCDUMP="700-1200:25"`, no `-level`), the
> gun-barrel **Bond model renders and animates correctly**: frame ~1075 shows
> Bond walking across the iris, frame ~1150 shows him turn and fire, both
> upright, recognisable tuxedo silhouette, correct position inside the barrel.
> The M-32b probe's "entirely absent — no silhouette" reading was of an
> upside-down capture (an inverted gun-barrel is a mostly-black frame with the
> iris low and Bond hanging inverted from it — easy to call "absent"). The
> `chrModelInstance` being valid with sane matrices, which the probe couldn't
> reconcile with "absent", now makes sense: it was drawing all along.
> **Still genuinely broken:** the **Nintendo logo** — frame ~775 renders it as
> two plain overlapping white ellipsoids shifted left of centre, no wordmark,
> no logo geometry. That is NOT a flip (a flipped Nintendo logo is still a
> Nintendo logo) — it's a real `logoinst` model-transform / geometry bug (Bug
> 1 territory). **Cast-roll models** not captured this pass — status unknown,
> re-check with a `1400-2200` window. Net: D75 shrinks to "Nintendo logo model
> renders as degenerate white blobs" plus the unverified cast roll; the
> gun-barrel half is retired. The M-32b probe text below is left for history.

**D75 Bug 2 — RUNTIME PROBE (session M-32b, `GE_D75=1` in `title.c`
`sub_GAME_7F007F30`, kept env-gated).** Booted the bare front end
(`GE_D75=1 GE_PCDUMP="1-1200:120"`, no `-level`) and captured through the
gun-barrel sequence. Findings:

- **Visually confirmed:** the gun-barrel spiral circles (2D backdrop DL)
  render fine; the animated **Bond chr model AND the gun model are
  entirely absent** — no silhouette, no muzzle. `chrModelInstance`
  (`obj=0x14012da80`, `numMatrices=21`) and `gunModelInstance`
  (`obj=0x140148620`, `numMatrices=1`) are **both non-NULL with valid
  `obj` pointers and sane matrix counts** — so mechanism (2) "model file
  not loaded / manifest row missing" is **ruled out**.
- **`render_pos` is NOT stale / clobbered — mechanism (1) as stated is
  ruled out.** Per-frame `chrModelInstance->render_pos` alternates cleanly
  between `0x700a8f10` and `0x700a3f10` (delta `0x5000` = the double-buffer
  swap). On each frame `render_pos` == the `renderData.mtxlist` base that
  `dynAllocate(21<<6)` returned *that same frame* (`gfxpos 0x700a9490` −
  `0x540` chr − `0x40` gun = `0x700a8f10` exactly), i.e. it points at the
  freshly-`subcalcmatrices`-written matrices, exactly as on N64. It is a
  real low-DRAM (`0x700xxxxx`) address; `osVirtualToPhysical(render_pos)`
  returns it **unchanged** (no D131 truncation). Nothing overwrites the
  arena between the write and the draw within the frame.
- **fast3d logs ZERO DL warnings** during the whole sequence (no
  "ending DL" / unknown-opcode / `fast3d_ptr_ok` / bad-matrix-substitution
  lines) — so Bug 2 is **not** the D144/D146 corrupt-front-end-sub-DL
  family either. The Bond/gun DL is submitted and consumed silently but
  emits no visible geometry.
- Mechanism (3), the `(s32)&ANIM_DATA_bond_eye_fire` truncation, was not
  the probe's focus but the model has valid matrices regardless, so a
  broken anim would at worst freeze a pose, not delete the model.

**Where that leaves Bug 2:** the failure is **downstream of matrix setup**
— inside `drawjointlist` / `modelRenderNodeDl` / `dotube` for these
title-screen `Model` instances: either the joint **vertex / node-DL
segmented-pointer resolution** produces no tris, or the model is
transformed off-screen / to zero scale by `renderData.basemtx`
(`= matrix` from `manipulateGunbarrelAndLogoMatrices`, which folds
compiled DL pointers through `OS_K0_TO_PHYSICAL` at `title.c:121/123/138`
— a D58/D84-class `(u32)`-wrap candidate, though the backdrop DLs it
feeds *do* render). **Next probe:** inside `drawjointlist`/`dotube` at the
gun-barrel frame, log the resolved `vtx`/`nodeDl` pointers and whether
`modelRenderNodeDl` emits any `gSPVertex`/`gSP1Triangle`, plus the
composed `basemtx * render_pos` for joint 0 (is it in the view frustum?).
Confidence the cause is joint vtx/DL resolution vs off-screen transform:
**low** — needs the drawjointlist-level probe. The `render_pos`-arena
in-scope fix hypothesised above is **no longer the lead** — do not land a
persistent `render_pos` buffer, it would not change anything.

**D76 (OPEN: 2D graphics — disclaimer/legal screen only partially drawn). ROOT-CAUSED (M-31) → D164.**
The image-table/D68 hypothesis is **wrong**: the legal screen (`constructor_menu00_legalscreen`,
`front.c:1523`) references **zero** `sImageTableEntry`s — it is 12 lines of `langGet()` font text
(`display_aligned_white_text_to_screen` → `textRender`) plus one 3D `logoinst` model (`subdraw`,
D75 family, parked). The "only 2 lines render" is a **linker-adjacency bug** in the text loop — see
**D164**.

**D77 (RESOLVED — in-level background music was silent; the general audio
pipeline was never at fault).** Original framing ("nothing reaches the PC
speakers") is **stale/wrong** — superseded by the M-78 audio work
(D198–D208): the SDL device opens, the software mixer runs, and menu/intro
music is audibly correct. The actual, narrower symptom (still present until
this fix): **in-level** background music never plays, while the SAME
`musicTrack1Play()` API called directly from the front end (menu/intro)
worked fine.

**Root cause: D6-class missing-`return` bug (see `docs/porting-notes.md`
§D6), NOT a mixer/trigger-plumbing gap.** `sub_GAME_7F0C0BF0()`
(`src/game/mp_music.c`) is a one-line wrapper `{ get_mTrack2Vol(); }` with no
`return` — on N64/IDO the value happened to already sit in `$v0`; at the PC
release build's `-O2` the return register is genuinely undefined. Every
in-level music trigger goes through `set_missionstate()`'s
MISSION_STATE_1/4 cases (the ONLY call path that starts a level's
background track — `lvlStageLoad()` → `sub_GAME_7F0C11FC()` →
`set_missionstate()`), which calls
`musicTrack1ApplySeqpVol(sub_GAME_7F0C0BF0())` /
`musicTrack3ApplySeqpVol(sub_GAME_7F0C0C10())` immediately before
`musicTrack1Play()`/`musicTrack3Play()`. The front-end/menu music
(`front.c`'s several direct `musicTrack1Play(M_INTRO/M_FOLDERS/...)` calls)
never goes through `set_missionstate()` at all, which is exactly why it
worked while level music didn't — different call path, not a different
pipeline. With the garbage/zero volume applied, `musicTrack1Play()`'s own
tail call `musicTrack1ApplySeqpVol(musicTrack1GetVolume())` re-applies the
same zeroed value, so the whole track plays at (effectively) zero gain for
its entire duration.

**Why this was hard to see from a note-level trace alone:** the compact-seq
player keeps working correctly downstream of the bug — `csplayer.c`'s
existing `GE_AUDIOTRACE` `[MUSICNOTE]` probe showed ~170-250 real-looking
note-on events (sane key/velocity) firing continuously through a whole
BUNKER1 capture. The bug is a *track/channel-level gain* fed once per
mission-state transition, not a per-note field, so a per-note trace looks
perfectly healthy; only a trace of the actual mixed output distinguishes
"notes are being processed" from "notes are audible".

**Verification (this session, `chore/golden-bunker1-rebaseline-d215`
worktree, `-level_09`/BUNKER1):**
- Added a temporary, cheap (~1/s, gated on `GE_AUDIOTRACE`) `[AUDIOLVL]`
  peak/RMS probe of the exact buffer `audioSetNextBuffer()` hands to SDL
  (`port/src/audio.c`), writing into the SAME `audiotrace.log` (same
  `geTracePrintf` lock) as the existing `[MUSICNOTE]` line so the two
  interleave in true call order without needing per-line timestamps.
- **Before the fix:** 173 `[MUSICNOTE]` events over a 28 s capture; **every
  single `[AUDIOLVL]` reading (29 samples spanning the full run) read
  `peak=0 rms=0.0`** — literally zero output reaches the mixed buffer for
  the entire run despite continuous note-on activity.
- **After the fix** (`return` added under `#ifdef AVOID_UB`, `#else` keeps
  the N64 body verbatim): same 28 s BUNKER1 capture, 173 `[MUSICNOTE]`
  events, and **every `[AUDIOLVL]` reading from the first note onward is
  strongly non-zero and sustained** (`peak` ~2700–10300, `rms` ~800–2900,
  continuously, not brief bursts) — a categorical before/after difference,
  not a marginal one.
- `tools_pc/verify.sh bunker1`: **PASS**, `frames=3 worst_cell=0` — no
  visual regression (this is a one-line audio-only fix; nothing touches
  rendering).
- Runtime audible confirmation (human ears on real hardware) is still owed
  — the capture above proves samples reach the SDL queue with real
  amplitude, which is the strongest evidence obtainable headless, but a
  human should do a by-ear BUNKER1 pass to close this out completely.

**Fix:** `src/game/mp_music.c`, `sub_GAME_7F0C0BF0()` — `return
get_mTrack2Vol();` under `#ifdef AVOID_UB` (`#else` keeps the N64 body
verbatim, matching the D187/D6 precedent). One line. No game-logic change
(same value the N64 build always effectively returned); ABI/UB-class fix,
not the D3x pointer-width family (no pointer/struct-layout involved here).

**Probe left in the tree:** `port/src/audio.c`, the `[AUDIOLVL]` block,
gated on `getenv("GE_AUDIOTRACE")` (same env var as the pre-existing
`[MUSICNOTE]` probe in `csplayer.c`, not touched this session). Cheap
(~1 peak/RMS scan per second of audio, only when the env var is set) — safe
to leave for future audio work reusing the same trace file. Remove whenever
`GE_AUDIOTRACE`'s other temporary probes are cleaned up.

**Confidence: HIGH.** Root cause matches an already-catalogued bug class
(D6/D187) exactly; the fix is a single `return`; and the before/after
`[AUDIOLVL]` measurement is a direct, categorical (zero → sustained
non-zero) reading of the exact buffer handed to the audio device, not an
inference from upstream state.

**Cross-cutting (Q1 — shared blocker?):** before ordering D75 vs D77, spend one check on whether
any single root cause touches both audio and rendering. Current evidence says **independent**
(audio = libaudio→SDL device; 3D = fast3d RSP emulation; 2D = image-table fixup + texture
import) — default to render-first if no shared cause is found.

**Non-negotiable #2 refinement (applied to AGENTS.md).** The original "game code
compiles unmodified / fix belongs in port/" is too absolute: pointer-width layout
cannot be isolated in `port/` (no hook between the romCopy and the first read).
Refined to "game **logic** is unmodified" with a narrow, documented exception for
mechanical, semantics-preserving ABI/layout changes forced by the 32→64-bit
transition (embedded pointers → u32 + cast at use; PC-guarded pool sizing),
following PD ground truth. No logic/behavior changes; each such edit is logged in
§F/D3x.

**D78 (RESOLVED — StandTile bitfield ABI, `#ifdef PORT` layout exception).**
`StandTile` (`src/bondtypes.h`) declares `u32 id : 24;` immediately followed by
a non-bitfield `u8 room;`. On MIPS/GCC (N64) these share one 4-byte storage
unit (id = bits 31:8, room = the low byte) giving an 8-byte tile header
(id/room word + `mid` u16 + `tail` u16) — the stride `list_of_tilesizes[]`
(0x20…0x58 = `8 + 8*pointCount`) and `stanFillin`'s `link << 3` addressing
both hard-depend on. x86 GCC never lets a non-bitfield member share a
bitfield's storage unit, so the stock declaration compiles to a **10-byte**
header on PC (`room`@4, `mid`@6, `tail`@8 — confirmed via
`offsetof()` probe against the real project headers/flags) — every
`tile->room`/`tile->mid`/`tile->tail`/`tile->points[]` access would silently
misalign, independent of any byte-swapping. `id` is provably dead (no
`.id`/`->id` read or write anywhere in the compiled game code — grep-verified
across `src/`), so under `#ifdef PORT` it is widened to `u8 id[3]` (order
irrelevant, decorative-only) with `room` immediately following as a plain
byte. Verified: this restores the exact N64 stride (`room`@3, `mid`@4,
`tail`@6, `points`@8). Layout-only, no behavior/logic change — same class as
D53.2. This is a prerequisite for D69 (byte-swapping alone cannot fix stan
tile reads if the struct itself is misaligned).

**D79 (RESOLVED — `bg_room_data` pointer-width ABI, `#ifdef PORT` layout
exception).** `bg_room_data` (`src/game/bg.h`) declares `pPointTableBin` /
`pPriMappingBin` / `pSecMappingBin` as `void *`. These are ROM-serialized as
plain 4-byte N64 segment-0x0F offset values and are **never dereferenced**
anywhere in the codebase (grep-verified: every use in `src/game/bg.c` casts
to `(u32)`/`(s32)`/`(u8*) + int` for arithmetic, never `->` or `*`). On
x86-64 `void *` is 8 bytes, silently growing the 24-byte N64 room record to
40 bytes and breaking every `ptr_bgdata_room_fileposition_list[i]` array
index. Fixed under `#ifdef PORT` by declaring them `u32` instead — verified
via `sizeof()`/`offsetof()` probe: `sizeof(bg_room_data)` == 24,
field offsets 0/4/8/12 (matching N64 exactly), so the room table needs no
resizing in the offline conversion, only in-place bswap32. No behavior
change (every existing use site already treats the value numerically); same
class as D53.1/D66.

**D80 (bg `.seg` format spec — converter spec of record).** Header
(`s32 header[0x10]`, only words 0–4 consulted by `load_bg_file`): word0 must
be 0 (bswap32, harmless either way); word1/2/3/4 are `0x0Fxxxxxx`
self-relative offsets (masked `&0xFFFFFF`) to: room-fileposition list,
portal-data-entry table, envdata table (0 = absent), and an optional f32
array (only meaningful if word3 != 0). **Table order in the file is not
index order** — verified across all 34 unique NTSC bg files
(`bg/*.seg` referenced from `levelinfotable`): word3 (envdata) < word2
(portal) in every sample; room-table extent = `[word1, min(word2, word3 if
word3>word1 else word2))`; in every sampled file this divides evenly by 24
(`bg_room_data` record size) and word4 was always 0 (f32-array path
unexercised in this ROM — converter asserts word4==0 and errors loudly if a
future region violates this, rather than silently mishandling it).
- `bg_room_data` (24B): 3× `u32` offset fields (D79) + `coord3d pos` (3×f32).
  All 6 words are plain numeric — blanket bswap32, **no resize** (PC stride
  == N64 stride after D79).
- `bg_envdata_entry_local` (8B, local to `load_bg_file`): `u8 type` + `pad[3]`
  (untouched) + `s32 data` (bswap32). Terminated by `type==0`. **Exception:**
  when `type==ENVIRONMENTDATA_ALT` (100), `data` is not arbitrary — it is
  compared post-rebase against `g_BgPortals[i].offset_portal`
  (`getIndexOfPORTALID`), i.e. it lives in the *same offset space* as
  `bg_portal_data_entry.offset_portal` (the portal point-data blob, below)
  and must receive the identical `+portal_delta` relocation, in addition to
  bswap32.
- `bg_portal_data_entry` (N64 8B: `u32 offset_portal` + 4× `u8`). Unlike
  `bg_room_data`, `offset_portal` **is** dereferenced pervasively elsewhere in
  `bg.c` (`->numPoints`, `->point`, portal/room-visibility walks) — declaring
  it `u32` would require touching dozens of call sites, so it is left as the
  native `bg_portal_entry *` pointer type (no header edit): PC
  `sizeof(bg_portal_data_entry)` is **16B** (`offset_portal`@0 8B,
  `connectedRoom1/2`+`controlbytes1/2`@8-11, 4B pad) — confirmed via probe.
  The **offline converter** (not game code) re-lays the table at 16B/record
  (N portal records + 1 zero terminator record), writing `offset_portal` as
  an 8-byte field: low 4 bytes = bswapped original offset value
  **+ portal_delta**, high 4 bytes = 0. `portal_delta = 8 * (N+1)` (the extra
  bytes inserted by 8B→16B growth). The `bg_portal_entry` point-data blob
  that follows the portal table in the file (target of every
  `offset_portal`/ALT-envdata value) needs **no per-record resize** — PC
  `sizeof(bg_portal_entry)` is 16B, identical to N64 (`u8 numPoints` + `pad[3]`
  + `coord3d point`, no pointer fields) — it is simply relocated by
  `+portal_delta` as a block, with `numPoints`/`pad` copied verbatim and
  `point` (3×f32) bswapped. Net effect: the whole `.seg` file grows by
  exactly `portal_delta` bytes; nothing outside the portal table/blob region
  needs remapping (room table, envdata, and the header's word1/word3 all sit
  *before* word2 and are untouched).

**D81 (`Tbg_*_stanZ` format spec — converter spec of record).** RZ-compressed
(`0x11 0x72` + raw deflate, same scheme as models — decompress/recompress
around the conversion). Decompressed layout: `struct StanPrefixRecord { s32
stanfile; StandTile *ptr_firstroom; }` is dereferenced directly against the
raw loaded buffer (`stanLoadFile`/`stanDetermineEOF` receive the file pointer
itself as `StanPrefixRecord *`), so — same class as D78/D79 — the struct's
PC-compiled layout must match the file's byte layout. N64: `stanfile`@0 (4B)
immediately followed by `ptr_firstroom`@4 (4B pointer). PC: pointer-alignment
forces an implicit 4B pad after `stanfile`, so `ptr_firstroom` compiles to
offset **8**, not 4 — and each subsequent room-offset array slot is a real
8-byte pointer (`stanDetermineEOF`'s `void **roomPtr; roomPtr++` walk and
in-place `*roomPtr = *roomPtr + delta` rebase already use genuine
pointer-width semantics — **no code change needed there**, only the file's
data layout). Converter fix (constant shift, no game-code edit): insert a 4B
zero pad after `stanfile` (array now starts at file offset 8, matching PC
struct layout), and widen every room-offset array slot from 4B to 8B (low
4 bytes = bswapped original file-offset value + `array_delta`, high 4 bytes
= 0; the terminator NULL slot becomes 8 zero bytes). `array_delta = (8 + 8*
(N+1)) - (4 + 4*(N+1)) = 4*(N+2)` where N = room-offset entries before the
terminator. Everything from the old tile-data start to EOF shifts by
`+array_delta` as a block; **tile records need no resize** (D78 restored the
exact N64 8-byte header stride, and `StandTilePoint`/`link` addressing is
already relative to `standTileStart`, computed at runtime — unaffected by
where the tile-data block sits in the file). Per-tile conversion: the 4-byte
id/room word is copied **verbatim** (D78 makes the PC struct byte-identical
to N64 there — no swap needed, it's a byte array not a scalar); `mid.half`
and `tail.half` (s16, top nibble of `tail` = `pointCount` selecting record
size via `list_of_tilesizes[]`: `8 + 8*pointCount`, pointCount 3–10 →
0x20…0x58) are bswap16; each of the `pointCount` `StandTilePoint` entries (8B:
x/y/z s16 + link u16) are bswap16 per field. The **N64-order (still-BE) tail
half must be read to size each record** while walking — same discipline as
the D50 model-node walk. Net: the whole stan file grows by exactly
`array_delta` bytes before recompression; the RZ-compressed sidecar size
(recorded in the manifest, patched into `rom_size`) differs from the N64
compressed size, same as pcmodels (D50) — this is expected and fine, nothing
in the load path assumes N64 compressed size.

**D82 (converter + port wiring — see `tools_pc/d69_emit.py` /
`port/src/pccg.c`).** Implements D80/D81 above: per NTSC bg/stan file
referenced by `levelinfotable`, converts and concatenates into
`data/pccg-ntsc-final/pccg.bin` + `manifest.csv` (`name,offset,size` decimal,
`file_resource_table` order) — same manifest shape as `pcmodels.bin`
(D50/`d43_emit.py`). Port layer (`port/src/pccg.c`, cloned from
`pcmodels.c`): `pccgReserveSize`/`pccgLoadSidecars`/`pccgPatchTable`, wired
into `port/src/romdata.c`'s cart-reservation extension alongside
`pcmodels*`, and `romdataCartAddrValid`/`libultra.c`'s D60 DMA-source bounds
check extended to cover the pccg byte range. One-shot patch call:
`pccgPatchTable()` from the same `load_object_fill_header` hook site as
`pcmodelsPatchTable()` (idempotent, matches every table entry by filename,
rewrites `hw_address`/`rom_size`). Regenerate: `python tools_pc/d69_emit.py
[ntsc-final|pal-final|jpn-final]`; only `ntsc-final` regenerated/verified
this session (PAL/JPN ROMs not present in this environment) — `data/pccg-*/`
is gitignored like `pcmodels-*`.

**D83 (RESOLVED — StandTileHeaderMid/StandTileHeaderTail bitfield ABI,
found during D69 verification).** After D78-D82 landed, a clean run reached
`stanBuildRoomData` (`stan.c:245`) without faulting, but then **hung
forever** (kernel heartbeat: no frame rendered, stuck at the same PC across
repeated snapshots). Root cause: same MIPS-BE-vs-x86-LE bitfield-packing
class as D78, but in a struct D78 didn't touch. `StandTileHeaderTail {
s16 pointCount:4; s16 headerC:4; s16 headerD:4; s16 headerE:4; }` — on
N64/MIPS the FIRST-declared field occupies the HIGH bits (`pointCount` =
top nibble); x86 GCC packs the first-declared field into the LOW bits
(`pointCount` = bottom nibble instead). `tile->tail.hdrTail.pointCount` is
read pervasively (`list_of_tilesizes[]` tile-size lookup used for
navigation, edge walks, `stanBuildRoomData`'s bounds loop) — with the stock
declaration this silently read the wrong nibble on PC. An env-gated probe
(`GE_D69STAN=1` in `stanBuildRoomData`, TEMP, kept) proved it directly:
tile tail=`0x03dc` (N64: pointCount=0, top nibble) decoded to
`pointCount=12` on PC (bottom nibble) — `list_of_tilesizes[12]` is
out-of-bounds (table has 12 entries, 0-11) and happened to read a stray 0,
so `tile` never advanced — infinite loop. Fixed under `#ifdef PORT` by
declaring both `StandTileHeaderMid` and `StandTileHeaderTail`'s fields in
**reverse order**: x86's low-to-high packing then lands each field in the
same bit position MIPS's high-to-low packing does (byte-identical numeric
result, verified via a union/probe against `0x03dc` returning
`pointCount=0`). `StandTileHeaderMid`'s fields (`special`/`r`/`g`/`b`) are
never read via their bitfield names either (only via `.mid.half >> 0xc`
elsewhere in stan.c) so that half of the fix is precautionary. Same
narrow-ABI-exception class as D78; no logic change.

**D84 (RESOLVED — bg.c hand-inlined segment-fold 64-bit-pointer overflow,
found during D69 verification).** With D83 in, `stanBuildRoomData`
completed and the game proceeded into room streaming
(`bgCheckIfRoomModelNeedsLoad` → `bgLoadRoomModelData` →
`bgLoadRoomVtxData`/`bgLoadRoomPrimaryGdl`/`bgLoadRoomSecondaryGdl`), which
then **segfaulted** at a fixed, reproducible fault address
(`0x7104561d`, identical across runs) inside `bgBuildRoomVtxBounds`
(`bg.c:2852`, reading `vtx[i].v.ob[0]`). Root cause: `bgLoadRoomVtxData` /
`bgLoadRoomPrimaryGdl` / `bgLoadRoomSecondaryGdl` each hand-roll the
`BG_SEG_TO_PTR` fold instead of calling the macro:
`offset = (((u8 *)room->pPointTableBin + ptr_bg_data) - ptr_bg_data) +
0xf1000000;` (and the Pri/Sec-mapping equivalents). On N64 this "+base
-base" cancellation is a no-op inside 32-bit pointer arithmetic that wraps
for free. D79 made `pPointTableBin`/`pPriMappingBin`/`pSecMappingBin`
plain `u32` fields (never dereferenced, matching every other use site), but
these three call sites still cast them to `(u8 *)` and did the arithmetic
as real 64-bit pointers: `+0xf1000000` no longer wraps at 32 bits the way
`BG_SEG_TO_PTR`'s explicit `(u32)` cast does, so the computed `offset`
came out roughly 4 GiB too large, corrupting every downstream room-file
read (compressed-data location and size). Fixed under `#ifdef PORT` by
doing the fold as plain `u32` math at all three sites, matching
`BG_SEG_TO_PTR` exactly (`offset = (u32)room->pPointTableBin +
0xf1000000;`, no pointer involved) — same narrow ABI-exception class as
D79/D69's original BG_SEG_TO_PTR fix, no logic change. (Root-caused via an
env-gated probe, `GE_D69BB=1` in `bgBuildRoomVtxBounds`/
`bgLoadRoomPrimaryGdl`, TEMP, kept — confirmed the compressed room-DL bytes
now start with the correct `11 72` RZ magic at the right file offset.)

**D85 (OPEN — room primary/secondary DL binaries decode to garbage after
D84; safety-netted, not crash-fixed at the geometry level).** With D84 in,
the compressed room DL binary loads and decompresses correctly (verified:
`11 72` RZ header at the right offset, plausible decompressed size), but
the **content** `texCopyGdls`/`texLoadFromGdl` produce from it is not a
valid GBI command stream (`GE_D69BB=1` dump: `cmd=00`, `01`, `02`, `52`...
none of these are display-list opcodes actually present in the source
bytes — the raw N64 bytes are untouched by the offline converter (D80: the
whole per-room DL/point-index blob is a byte stream, deliberately left
unconverted, out of scope for this milestone) and `texLoadFromGdl` is the
*same, already-working* model-GDL runtime converter (`bgLoadRoomPrimaryGdl`
calls it identically to the model-loading path) — so either room GDLs use
a BG-specific command/marker convention `texLoadFromGdl`'s marker-expansion
logic doesn't handle, or something upstream of it (compression alignment,
`csize_primary_DL_binary`/`csize_secondary_DL_binary` delta sizing) is
still off. Not yet root-caused; full triage is D75-class 3D-pipeline work,
out of scope for this session. **Crash prevented, not geometry fixed:**
added a `#ifdef PORT` bounds check in `bgBuildRoomVtxBounds` before every
`vtx[i]` dereference (`vtxOff`/`vtxEnd` must fit inside
`usize_point_index_binary`) — a garbage command stream now produces an
empty/degenerate bounding box for that vertex batch instead of an
out-of-bounds read, so a bad room fails to render sanely rather than
segfaulting. Follow-up: decode what `texLoadFromGdl` actually does with
room-specific opcodes (`bgApplyDynamicCCRMLUT`/`ptrDynamic_CC_RM_LUT`/
`DL_LUT_PRIMARY_ADDFOG` suggest room GDLs carry CC/RM-LUT-selection markers
models don't use) and verify `csize_*_DL_binary` sizing end-to-end.

**D86 (RESOLVED — `modelInitRwData` crash was a single truncating pointer
cast in the player's embedded gait/arm model init, unrelated to bg/stan).**
Root-caused with a new env-gated trace (`GE_D86=1`: node-walk trace in
`modelInitRwData` + a load-identity probe in `load_object_fill_header`,
both TEMP, left in place). The trace showed the crash node's low 32 bits
were `(header_ptr & 0xFFFFFFFF) + 0x1E0` with the high 32 bits zeroed, and
that this header was **never** loaded via `load_object_fill_header` (no
matching probe line) — pointing at a statically-embedded model, not a
dynamically-loaded one. `src/game/initplayergaitobject.c:5` does
`player_gait_object_header.RootNode = (int)&player_gait_hdr;` — a
same-width (32-bit) pointer→int→pointer round trip that's a no-op on N64,
but on PC `(int)` truncates the real 64-bit `&player_gait_hdr` to its low
32 bits, and the implicit int→pointer conversion back into `RootNode`
zero-extends it, dropping the executable's load-base high bits (module
maps at `0x140000000`, so the truncated pointer silently loses the
`0x1`). `init_player_gait_object()` runs once from `boss.c:236`, and
`player_gait_object_header` is only used once real gameplay starts
(`initBondDATAdefaults.c:99` `animInit`s the player's gait model) — never
exercised while the game only ever got as far as the intro/cast screens.
Fixed with a `#ifdef PORT` branch in `initplayergaitobject.c` that assigns
the real pointer directly (behavior-identical to the N64 assignment,
ABI-width fix only). Verified: BUNKER1 now loads past this point with a
clean, deterministic repro via `-level_09` (see D88).

**D87 (RESOLVED — attract-mode demo playback (`ramrom_replay_handler`)
crashed on a big-endian `ramromfilestructure` read with no byteswap).**
Found while re-verifying D86: an idle front-end run (no player input)
eventually calls `select_ramrom_to_play()` (`ramromreplay.c`), which picks
a random compiled-in demo blob from `ramrom_table[]` (`ramrom_Dam_1`,
`ramrom_BunkerI_1`, etc. — genuine shipped attract-mode assets, not a
debug-only feature; the debug-menu replay path, `DEB_REPLAYRAMROM`, is
structurally unreachable in this `ntsc-final`-equivalent build since
`DEBUGMENU` isn't defined — confirmed with a `gdb -p <pid>` **attach**
hardware watchpoint on `is_ramrom_flag`, which resolved cleanly and
quickly this session; attach mode works fine for a non-timing-dependent
write, unlike the launch-mode-only guidance logged after the D56 session —
worth a retry next time attach seems useful). `replay_recorded_ramrom_at_address`
loads `ramromfilestructure` via `romCopyAligned()`, a raw byte copy (by
design, D66) from a real ROM-compiled asset — so, like every other
N64-compiled ROM asset, its multi-byte fields are big-endian, and nothing
byte-swaps them on read. A real `size_cmds` of 2 (BE bytes `00 00 00 02`)
read as native LE prints as `33554432` (`0x02000000`); that garbage then
drives the loop bound and pointer arithmetic in
`iterate_ramrom_entries_handle_camera_out`/`ramrom_replay_handler`, which
walks far outside the small `ramrom_blkbuf_2`/`ramrom_blkbuf_3` scratch
buffers and segfaults reading `temp_v0->stick_x`
(`ramromreplay.c:301`/`ramrom.c` callers). Root-caused with a new
env-gated probe (`GE_D87=1`, left in place). Fixed with a `#ifdef PORT`
`ramromFixupEndian()` in `ramromreplay.c`, called once right after the
`romCopyAligned()` in `replay_recorded_ramrom_at_address` (same pattern as
the D54 cseq-header fixup): byte-swaps every multi-byte field
(`u64`/`u32`/enum fields via `__builtin_bswap64`/`32`, `save_data.options`
via `bswap16`); `save_data`'s single-byte fields and the `times[]` byte
array are left alone. The **downstream** per-frame chunks
(`ramrom_seed`/`ramrom_blockbuf`, read via the same `romCopyAligned`
pattern in `iterate_ramrom_entries_handle_camera_out`) are all-`u8`
structs and need no swap. Not BUNKER1-specific — this is a front-end/
attract-mode path that can select any of the 7 demo locations at random;
use `-level_09` (see D88) to skip the front end entirely for deterministic
BUNKER1 testing instead of waiting on/fixing attract mode.

**D88 (OPEN — root-caused, next blocker: per-level `Usetup*Z` "stage
setup" file is raw N64-endian/width ROM bytes read directly through a
PC-widened struct, with no conversion at all).** Found immediately after
D86/D87 while re-verifying BUNKER1 specifically — launch with `-level_09`
(`boss.c:199-339` decodes `-level_XX` into `g_StageNum`, bypassing the
front end/attract-mode entirely for a fast, deterministic repro; NTSC
`LEVELID_BUNKER1 = 9`, and the token's two digit-chars are consumed as raw
ASCII bytes, so `"09"` → `'0'*10 + '9' - 0x210 = 9`) reaches the exact same
crash as the random attract-mode run, immediately and reproducibly:
`proplvreset2` (`prop.c:1306`) segfaults reading
`g_CurrentSetup.pathwaypoints[i1].padID`. `prop.c:1267-1282` loads the
level's `"Usetup<name>Z"` file with `_fileNameLoadToBank` (raw ROM bytes,
**not** run through any PC-layout converter — unlike bg/stan (D69/D80-82)
and models (D43/D50), this asset type has zero PC porting work done on
it) into `local_stage`, then rebases 10 top-level fields
(`pathwaypoints`/`waypointgroups`/`intro`/`propDefs`/`patrolpaths`/
`ailists`/`pads`/`boundpads`/`padnames`/`boundpadnames`, plus nested
`neighbours`/`waypoints`/`ailist` pointers inside the sub-tables) with
`(void *)(((u32) local_stage) + ((u32) local_stage->pathwaypoints))` —
i.e. by reading the *raw file bytes* directly through the live
`struct stagesetup` (`bondtypes.h:4091`), whose 10 fields are declared as
real pointers. This is worse than a plain missing-byteswap bug (cf. D87):
on N64 those 10 fields are 4 bytes each (40-byte header, correctly
self-describing "byte offset from file start" per the code's own
comment), but the PC struct widens every pointer field to 8 bytes (an
80-byte header) — the same class as D79 (`bg_room_data` pointer growth)
— so field N's read doesn't even land on the right *bytes* of the file
past field 0, before even considering that the 4 meaningful bytes it does
read are big-endian. Confirmed no PORT/byteswap handling exists anywhere
in `prop.c` (`grep` for `bswap`/`#ifdef PORT` in the file: zero hits).
**Not fixed this session** — this is format-conversion work at the same
scale as D69 (a whole ROM asset type needs a byte-accurate spec + either
an offline converter sidecar, the established preferred pattern per
AGENTS.md, or a careful runtime fixup pass that parses the raw 40-byte
N64-packed header by explicit byte offset, byte-swaps each field, and
writes the results into the PC-widened `stagesetup` struct — plus the
same treatment for every nested sub-table referenced from it
(`waypoint`/`waygroup`/`PropDefHeaderRecord`/`PathRecord`/`AIListRecord`/
`PadRecord`/`BoundPadRecord`/`pname`, each of which likely has its own
internal offsets/BE fields not yet audited). **This is the actual next
blocker to a rendered BUNKER1 frame** — reachable deterministically via
`-level_09` in well under a minute, no attract-mode wait required.

**D88.1–D88.3 RESOLVED / VERIFIED (2026-08-28, session L).** The
`Usetup*Z` offline converter (`tools_pc/d88_emit.py`, 531 lines) was
written and run in a prior interrupted session; this session verified its
output is correct and consumed at runtime:
- `port/src/pccg.c` `PCCG_MAX_FILES` grown 128 → 256 so the sidecar image
  can also carry the 21 `Usetup*Z` rows (manifest now has them, e.g.
  `UsetuparchZ,3375808,19265`).
- The converter delta-relocates the 8 growing tables (header 40→80B,
  waypoint 16→24, waygroup 12→24, PathRecord 8→16, AIListRecord 8→16,
  PadRecord 44→56, BoundPadRecord 68→80, pname 4→8) and bswaps the s32
  ID/offset fields — same technique as D80/D81, generalized to many
  interleaved regions.
- `src/bondtypes.h` `SetupIntroCamera`: `lang1c`/`lang20`/`prev`
  ROM-serialized pointer-shaped fields kept narrow (`u32`) under
  `#ifdef PORT` so `sizeof` stays 40 and the fixed-stride intro-record
  walk in `bondview_r.c`/`bondview2.c` still matches the 40-byte file
  records; use sites cast `(char *)(uintptr_t)` at each read. Same class
  as D79/D53.1. **Write-before-read verified**: `bondviewLoadSetupIntroSection`
  (`bondview_r.c:276-300`) writes `prev` (list link) and both `lang_ptr`
  members (via `langGet()`) before the only subsequent reads.
- **Verified via `-level_09` + `GE_D88=1` probe**: `proplvreset2` now
  walks the entire pads table correctly — plink name strings (`p1988e`,
  `p12295e`, …) and sane BUNKER1 world coordinates for every pad; the
  crash at `prop.c:1306` (`pathwaypoints[i1].padID`) is **gone**.

**D88.4 RESOLVED / VERIFIED (2026-08-28, session M).** The `propDefs`
polymorphic record stream is now converted N64→PC by the offline sidecar.
`-level_09` no longer crashes in `setupDoor`/`modelLoad` (or anywhere else)
— BUNKER1 loads its full stage setup and renders **1000+ frames
continuously** with zero FATAL/EXCEPTION (only the pre-existing unrelated
`romdataFixupMusicSeqTable` warning).

Key facts established:
- The `propDefs` stream is a flat `s32[]`. Each record's **serialized N64
  word count is fixed per `type` byte across all 21 levels** — verified by
  parsing the getools C sources (`assets/obseg/setup/Usetup*Z.c`), which
  tile every `propDefs` region byte-for-byte against the retail ROM
  (BUNKER1: 206 records, 6477 words, end offset lands exactly on `intro`).
  Table: `PROPDEF_N64_WORDS` in `tools_pc/d88_propdefs.py`.
- Every pointer member inside a serialized record is **`0` in the file**
  (runtime-populated by the `New_*Record` macros). So there is no
  garbage-pointer / delta-relocation problem in the file — only that on PC
  those slots widen 4→8B, growing records that contain pointers and
  changing the `sizepropdef()` walk stride.
- The brief's "narrow the trailing pointers `#ifdef PORT`" idea (D88.1
  pattern) is a **poor fit**: `ObjectRecord.prop`/`.model` are mid-struct,
  and `ObjectRecord`/`DoorRecord`/`GuardRecord` are used in ~350 runtime
  sites — narrowing would touch core gameplay code broadly.
- The `#if 1` branch of `sizepropdef()` was **already N64-correct for every
  type except `OBJ_COPY_ITEM`** (returned 1, real serialized size 3).

Fix (chosen: **converter grows records to native PC layout**, no struct
changes):
1. `tools_pc/d88_propdefs.py` — rewrites each record to its native PC
   struct size: header word (`u16 extrascale`+`u8 state`+`u8 type`) and
   `_mkword` half-pairs byte-swapped independently, scalars `bswap32`'d,
   pointer slots widened to 8 zero bytes (8-aligned), runtime areas
   zero-filled. Per-type PC size = `PROPDEF_PC_BYTES`, sourced from the
   compiler-verified `tools_pc/d88_layoutprobe.c` (`sizeof`/`offsetof`
   against the real port include chain).
2. `tools_pc/d88_emit.py` — feeds the converted stream + its growth into
   the cumulative delta so `intro` and every later sub-table shift.
3. `loadobjectmodel.c` `sizepropdef()` — `#ifdef PORT` branch returns
   `PROPDEF_PC_BYTES/4` so the in-place walk (`prop.c`,
   `loadobjectmodel.c`, `objective_status.c`) matches the emitted stride.

`intro` conversion was already correct (blanket per-word `bswap32`; the
type discriminant is a full `s32`). **Not yet done:** `PROPDEF_PC_BYTES`
for `VEHICHLE`/`AIRCRAFT`/`TANK`/`AMMO`/`DEPOSIT_IN_ROOM` are placeholder
guesses (not used by BUNKER1) — probe them before those levels load.

**D88.5 (WATCH — stan tile name lookups all miss during pad setup).**
With the `GE_D88` probe on, every `stanMatchTileName` call from
`proplvreset2`'s pad loop walks the full ~2599-tile room and returns "no
match". May be benign at load (a NULL `pad->stan` is tolerated by the
reset path), but the stan-id derivation (`stanIdHi`/`stanIdLo` from the
pad `plink` name) could be another victim of a residual endian/width bug.
Re-check after D88.4; do not treat as resolved just because it doesn't
crash.

**D88.5 (RESOLVED — stan tile-name byte-swap in the converter).** With the
`GE_D88` probe on, every `stanMatchTileName` call during pad setup missed
(0 matches / 276 misses). Root cause: `stanMatchTileName` reads a tile's
packed name id through a `StandTilePoint` alias — `(u16)tile->x ==
stanIdHi` and `*((u8 *)&tile->y) == stanIdLo`. D78 left the 4-byte
id/room word as a verbatim byte array on the premise "`id` is provably
dead", which missed this aliased *scalar* read: on little-endian PC the
`(u16)` load of the big-endian id-hi bytes comes back byte-swapped.
Fix (converter, `tools_pc/d69_emit.py` stan path): swap bytes 0–1 of the
id/room word; byte 2 (`stanIdLo`) and byte 3 (`room`, read as
`tile->room`) stay put. Verified 273/273 name matches after the fix; pads
now resolve real stan tiles. Needs sidecar regen. Committed.

**D88.6 (RESOLVED — intro CAMERA `lang1c` is a `u16` pair).**
`SetupIntroCamera.lang1c` is `union { u16 lang_index[2]; u32 lang_ptr; }`
and `bondview_r.c:295` reads `lang1c.lang_index[1]`. `d88_emit.py`'s intro
converter `bswap32`'d the whole word, which swaps element 0 with element 1
— the consumer then read the wrong language-slot id, indexed `g_LangBanks`
out of range and crashed in `langGet` (`language.c:421`) on a NULL bank.
Fix: byte-swap each `u16` of `lang1c` in place; `lang20` (a real `s32`)
still gets `bswap32`. Needs `d88_emit.py --regen`. Committed.

**D89 (RESOLVED — two crashes between stage-load and first frame).**
(a) `init_path_table_links` (`initpathtablelinks.c:144`):
`validationGroupCursors[-3]` is a decomp artifact — a constant negative
index into a 1-element stack array standing in for a plain cursor local.
GCC proves it OOB and emits a trap → SIGILL on PC. Fixed under `#ifdef
PORT` by pointing the name 3 elements into a real 4-element backing buffer
(identical `[-3]` expressions, now in bounds); N64 build keeps the plain
array. (b) `sub_GAME_7F0B0914` (`walkTilesBetweenPoints`): callers like
`domakedefaultobj` pass `&pad->stan`, legitimately NULL when a pad's stan
name doesn't resolve. On N64 the walk reads ~0 for `pointCount` and
returns TRUE via the `crossings==0` early-out; on PC the near-NULL read
faults. Guarded `*tileStack == NULL → return TRUE`. Committed.

**Session-M-2 infra fixes (committed).** Two port-layer gaps that made
`-level_09` a no-op were fixed: (1) `osPiReadIo` was stubbed to 0 so the
cartridge-token read always yielded an empty string — N64 debug switches
were silently ignored and only attract-mode demo playback could reach a
level. Now synthesised from `argv[1..]` (`sysGetTokenString` in
`system.c`, served from the 0xFFB000 range in `libultra.c`). (2)
`pccgPatchTable`/`pcmodelsPatchTable` were called lazily from the first
model load; a direct `-level_XX` boot loads a stage first and
`load_bg_file` read raw big-endian ROM. Moved the one-shot calls to the
end of `obInit()`. NOTE: a bare `-level_09` still needs the per-level
memory args too (`-ml0 -me0 -mgfx100 -mvtx50 -mt700 -ma150` for BUNKER1,
from `boss.c`'s `memallocstringtable`) — the `-level_` branch skips the
default `-m*` string. TODO: auto-inject from `memallocstringtable` in the
port so bare `-level_XX` works.

**D90 (RESOLVED — `stanTileDistanceRelated` zero-fill overran the caller's
stack).** Symptom: after D88.5/D88.6/D89, `-level_09` loaded BUNKER1 and
faulted in `stanIsSpecialBit1Set` (`stan.c:2364`, `arg0 == NULL`) on the
first player collision tick (`bondviewCalcUpdatePlayerCollision` →
`bondviewTrySimpleMovePlayerCollision` → `bondviewTryMoveToStan` →
`stanTileDistanceRelated` → `sub_GAME_7F0B1DDC` → `callbackA(NULL, …)`).
Root cause was NOT the pad→stan resolution (GE_D90 probe confirmed all
159 BUNKER1 pad names resolve and the player spawn pad #102 has a valid
stan). It was `stanTileDistanceRelated`'s N64 "HACK" init loop: it
zero-fills `((s32*)arg4)[0..19]` — **80 bytes** — while
`sizeof(StandTileLocusCallbackRecord)` is 16B. On N64 the 64-byte overrun
landed in adjacent stack scratch; on PC the frame layout differs (and
locals are pointer-widened), so the fill zeroed `bondviewTryMoveToStan`'s
live `sp90` (= `field_488.current_tile_ptr`) right before it was passed
as `&sp90` to the walk. Fix: `#ifdef PORT` clears exactly the 4 record
fields (every consumer only uses those four — cf. `sub_GAME_7F0B21B0`).
Committed. GE_D90 probes left in place (env-gated).

**D91 (RESOLVED — bg portal-descend truncated an array-element address).**
`sub_GAME_7F0B7F84` (both variants, `bg.c`): `i = (s32) &D_800442FC[
portalnum];` then later `*((u8 *) i) = depth;`. The `(s32)` cast drops
the top 32 bits of the array address on PC, so the byte store faulted
during portal occlusion culling. `i` is only used as an `if (i);` no-op
after the cast, so under PORT keep it a plain value and write
`D_800442FC[portalnum] = (u8) depth;` directly. Committed.

**D92 (RESOLVED — two truncated pointers on the chr/AI spawn path).**
(a) `chrAllocate`'s 5th parameter was declared `s32` but both call sites
pass `ailistFindById()`'s `AIRecord *`. The 64-bit pointer was truncated
binding to `s32 arg4`, then forwarded to
`init_GUARDdata_with_set_values`'s `AIListRecord *arg5` → `chr->ailist`
held e.g. `0x40127640` instead of `0x140127640`, and `ai()` faulted on
`(AiListp + Offset)->cmd` at the first AI tick. Param widened to
`AIListRecord *` under `#ifdef PORT` (`chr.c` + `chr.h`). (b)
`Model.unka0` is a 32-bit field that on N64 holds a function pointer —
always `sub_GAME_7F01FC10` (`chr.c:1618` stores `(s32)sub_GAME_7F01FC10`,
the only value the setter ever gets). `model.c` `subcalcpos` calls it
back through a cast → truncated jump target. Widening the field would
shift the rest of `Model`, so under PORT the setter stores a nonzero flag
and `subcalcpos` calls `sub_GAME_7F01FC10` directly. Committed.

**D85 revisited (OPEN — now the live blocker on `-level_09`).** With
D90–D92 in, BUNKER1 loads and ticks all the way to the **first render**,
which immediately hits `sysFatalError("Bad size for RGBA texture in tile
0: 00")` (`port/fast3d/gfx_pc.cpp:967`) — a `G_SETTILE` with `fmt=RGBA
siz=0` (invalid). This is the room-GDL-decodes-to-garbage problem from
D85 (the raw N64 per-room DL/point-index blob is left unconverted, D80),
surfacing in the texture path this time rather than
`bgBuildRoomVtxBounds`. Attract mode's "~2100 frames" never hit this
because those frames were HUD/menu screens, not room geometry. This is
the render-milestone work: decode what `texLoadFromGdl` does with
room-specific opcodes / CC-RM-LUT markers, and verify `csize_*_DL_binary`
sizing (see the original D85 entry above). Interim option if a fresh
session wants to keep moving past it: soften the four `sysFatalError`
"Bad size…" guards in `gfx_pc.cpp` to skip-with-warning (same
safety-net philosophy as the D85 `bgBuildRoomVtxBounds` bounds check) so
the frame renders with placeholder textures instead of aborting.

**D93 (RESOLVED — null-room (room 0) NULL-deref sat in front of the D85
texture wall).** Committed `164d7f99`. On `-level_09` the visible-room
draw list includes room 0, which has no geometry (`csize_*_DL_binary ==
0`), so `bgLoadRoomModelData` never assigns `ptr_expanded_mapping_info`
and it stays NULL. Two consumers then walk it unconditionally:
`bgApplyDynamicCCRMLUT` (called `start=NULL, end=NULL` → the `end==NULL`
sentinel-scan branch derefs address 0) and `bgBuildRoomVtxBounds`
(`while (gdl[i].dma.cmd != G_ENDDL)` on `gdl=NULL`). On N64 address 0 is
readable RDRAM so both walks wander harmlessly; PC page 0 is unmapped.
Fixed with narrow `#ifdef PORT` NULL guards (same safety-net class as the
D85 `bgBuildRoomVtxBounds` vtx-bounds check). `-level_09` now reaches the
documented D85 `sysFatalError("Bad size for RGBA texture in tile 0: 00")`.

**D85 root cause CONFIRMED (was "not yet root-caused").** The
decompressed per-room primary/secondary DL blob is **raw N64 data with
8-byte big-endian `Gfx` slots**, but every PORT-patched consumer
(`texCopyGdls`/`texLoadFromGdl` in `tex.c`, `bgApplyDynamicCCRMLUT`,
`bgBuildRoomVtxBounds`) was patched under `#ifdef PORT` to assume 16-byte
little-endian PC `Gfx` slots — so they stride at 2× the real rate and
read the middle of each N64 command pair as an opcode (the observed
`cmd=00,01,02,52…` garbage), and even at the right stride the `w0/w1`
words are unswapped BE. The model-GDL path is immune only because its
offline sidecar (`tools_pc/d43_*.py`) pre-widens + byteswaps every GDL;
D80 explicitly left the per-room DL/point-index blob unconverted.
**Fix chosen: runtime fixup**, not a sidecar — the blob is RZ-compressed
inside the bg `.seg` and delta-sized (its size is only the offset delta
between consecutive rooms' `pPriMappingBin`), so an offline widen would
force rewriting the whole bg-header offset table / recompression. The
transform is purely mechanical (8→16 byte widen + `bswap32` each word);
room GDLs need no pointer remapping (`G_VTX` seg addresses are resolved
at runtime via `SEGMENT_OFFSET(...) + (u32)vertices`, textures via the
`G_NOOP`+`texnum` marker). Implemented in `bgLoadRoomPrimaryGdl` /
`bgLoadRoomSecondaryGdl` between `bgDecompress` and `texCopyGdls`, plus a
`Vtx` short-field bswap in `bgLoadRoomVtxData` (positions/uv are BE; the
4 rgba `u8`s are fine). Watch the alloc budget — the PC blob is 2× the
decompressed size. Also fix the pre-existing bug at `bg.c:2448`:
`texLoadFromGdl((Gfx *)scratch, (Gfx *)expanded_size, ...)` casts the
size arg to a pointer.

**D85 widen fix IMPLEMENTED (committed `ea8a37a0`/`c732425d` = master
`c732425d`).** `bgWidenRoomGdl()` — in-place back-to-front 8→16 widen +
`bswap32` per word — runs in `bgLoadRoomPrimaryGdl`/`bgLoadRoomSecondaryGdl`
right after `bgDecompress`; the doubled size flows into `texCopyGdls`/
`texLoadFromGdl`/`usize_*_DL_binary`. `bgSwapRoomVtx()` `bswap16`s the 6
leading `u16`s of each `Vtx` in `bgLoadRoomVtxData`. The `bg.c:2448`
size-cast-to-pointer bug is fixed under `#ifdef PORT`. Verified: room
GDLs now decode to real GBI (`GE_D69BB` dump: `E7` RDPPIPESYNC / `BA`
SETOTHERMODE_H / `B9` SETOTHERMODE_L / `FC` SETCOMBINE / `BB` TEXTURE /
`B7` SETGEOMETRYMODE), G_NOOP markers decode to sane texnums. Alloc
budget needs no change (first-load block is `memaGetLongestFree`-sized,
then shrunk from post-widen `used`).

**D85 texture pool FIXED (committed `2b3ee6e7` → master `6f0208d6`).**
`ptr_texture_alloc_start` is declared `struct texpool *` but every use
takes its address and treats the storage *as* a `struct texpool`
(`texInitPool` writes 4 members through `&…`, `texLoad`/`texFindInPool`
read them back). N64: 4-byte members, 16-byte struct, works by layout
luck. PC: 32-byte struct (4 widened pointers) → `texInitPool` smashes
24 bytes of trailing BSS and `->leftpos`/`->rightpos` read back garbage
→ stage pool looks permanently exhausted (`texFreeBytesInBuffer() < 0`)
→ every room `texLoad` bails → `texFindInPool` NULL for every texnum →
`Bad size for RGBA texture` abort. Fix: `#ifdef PORT` define it as a
real `struct texpool` (`image.c:14`, `image.h:98`); all call sites
already `&…` it. Verified: pool fills, ~630 room textures resolve,
FATAL gone. **Latent (not fixed):** `sizeof(struct tex)` is 24 on PC vs
16 N64 (widened `u8 *data` + bitfield align); the pool is a dual stack
(pixels up from `start`, `struct tex` headers down from `end`), so the
stacks collide ~`bytes/3` early. BUNKER1 `-mt700` doesn't hit it;
texture-heavier levels will need a separate header allocation or a
per-level `-mt` bump. (Enlarging the pool alloc starves `MEMPOOL_STAGE`
and hangs `mempAllocBytesInBank`'s OOM `while(1)` — don't.)

**D85 texture wall CLEARED (session M-3, verified).** With sidecars
present and `493c9838`+`6f0208d6` in, `-level_09` **no longer hits
`Bad size for RGBA texture`** — BUNKER1 renders multiple full frames
(`GE_D85GDL` probe: `g_BgNumberOfRoomsDrawn=5`, roomids 9/10/11/15/17,
`b_min/b_max` 0/2, frame GDL advances ~3 KB/frame cleanly). The room-GDL
stream-decode + texture-pool layers of D85 are done. What remains is a
cluster of **non-deterministic** crashes newly reachable now that the
render loop + guard AI actually run in-level (4-run sample: 3 distinct
fault sites). Treat as a fresh crash-chain, not D85:

1. **`bgScissorCurrentPlayerView` frame-GDL overrun** (`bg.c:1355`,
   fault `0x70800000` = 1 byte past the 8 MB emulated DRAM). Hits ~1/4
   runs. The frame GDL write pointer runs off the end — either the PC
   frame-DL buffer is undersized for real room geometry, or a room GDL
   without a clean `G_ENDDL` drives a runaway append somewhere upstream
   (`sub_GAME_7F0B3C8C` double loop over `chrpropsRenderPass` /
   `bgRenderRoomPrimary`). NOTE: an *earlier* "deterministic overrun"
   reading was a red herring — it was `data/` sidecars missing (see the
   `data/` deletion note below), which faults in `load_bg_file:847` on
   raw BE bg-header bytes.
2. **`chrlvInitActAttack` bad pointer** (`chraction.c:1316`, fault
   `~0x40123350`). Hits ~2/4 runs — a guard starting an attack derefs a
   truncated/garbage pointer (`0x401xxxxx` looks like a 64-bit pointer
   with the high word lost, cf. D86/D92 class).
3. **`gfx_sp_matrix` unrelocated segment addr `0x90000000`**
   (`gfx_pc.cpp:1077` via `gfx_run_dl`). Hits ~1/4 runs — a room-GDL
   `gsSPMatrix` w1 carries N64 segment 9 unresolved. Room-GDL matrix
   arg relocation still needed (the widen byte-swaps the words but does
   not remap seg addresses; `G_VTX` seg-addrs are resolved elsewhere via
   `SEGMENT_OFFSET`, but `gsSPMatrix` is not).
4. **`bgTestRayIntersectionInRoom`** (`bg.c:3302` `((u32*)gdl)[1]`,
   `:3383/3388/3521/3526` `*(u8*)gdl`) still reads opcodes/w1 N64-style;
   post-widen the opcode is byte 3 and w1 is `((u32*)gdl)[2]`. Operated
   on garbage before (so "worked"); needs PORT accessors. Hitscan, not
   render — lower priority.
3. **Frame-GDL buffer overrun in `chrpropsRenderPass`**
   (`bgScissorCurrentPlayerViewDefault`, `chrprop.c:569` → `bg.c:1355`,
   write fault ~`0x70800000`). Timing-dependent, masked behind #1.
   Prop/character render path (D75-adjacent), newly reachable.

**D94 (RESOLVED — truncating `(s32)` pointer cast in `chrlvInitActAttack`).**
Committed `63204a27`. `chraction.c:1221`/`1231` compute the firing-anim
table entry as `(s32)arg1[anim_index]->table + (s32)(idx*sizeof(...))` —
`table` is `weapon_firing_animation_table (*)[]`, and the `(s32)` cast
truncates the 64-bit pointer, then the cast back zero-extends, dropping
the `0x1_00000000` module-base bit → `panim_float ≈ 0x4012xxxx` →
`panim_float->anim.anim` faults there when a guard starts an attack
in-level. Replaced both with plain array indexing (`&(*table)[idx]`) under
`#ifdef PORT` — the code's own comments say that's the intent.

**D95 (OPEN — the `-mgfx` master-DL buffer is half-capacity on PC; the
naive fix OOMs `MEMPOOL_STAGE`).** `dyn.c:56` sizes `g_GfxBuffers` from
`-mgfx` (a byte budget from `boss.c`'s per-level `memallocstringtable`,
sized for N64 8-byte `Gfx`). On x86-64 a `Gfx` is 16 bytes, so the master
display list holds **half** the commands for the same budget. Every
render fn appends with a bare `gdl++` and **no bounds check**, so once
BUNKER1 emits real room geometry the list overruns `g_GfxBuffers[1]`/`[2]`,
runs off the stage mempool, and faults writing a GBI command at the top
of the 8 MB emulated DRAM (`0x70800000`) — **non-deterministically**, and
scribbling GBI across DRAM on the way (the run-to-run `Unknown GBI opcode
0x3f/0xffffb9` and `gfx_sp_matrix` seg-9 `0x90000000` faults were all
downstream corruption from this one overrun). Doubling the `g_GfxBuffers`
allocation (`* sizeof(Gfx)/8`) under PORT **stops the overrun** (verified:
`-level_09` then runs 90 s+ / 5000+ VI posts, no crash) **but** the extra
~100 KB starves `MEMPOOL_STAGE` so `zbufAllocate` →
`mempAllocBytesInBank` (`memp.c:204`) spins in its OOM `while(1)` and no
frame ever renders (`kernel heartbeat: frames=0`; mainThread stuck in
`lvlRender`→`viClearZBufCurrentPlayer`→`zbufInit`). Committed as
`70784f80`, reverted `2a506284`, **re-applied `f35eba91`** alongside the
mempool-ceiling fix below.

**`malloc`-the-gfx-buffer is NOT viable** — `osVirtualToPhysical` /
`OS_K0_TO_PHYSICAL` in the port `(u32)`-truncate and subtract
`0x70000000`, so any pointer that flows through them (sub-DL branches
built in the gfx buffer, etc.) must live in the `0x70000000` DRAM window.

**Mempool ceiling fix (`933ba52b`, kept).** `port/src/n64stubs.c`'s
`tlbmanageGetTlbAllocatedBlock` returned the N64-fidelity ceiling
`0x702F4400`, leaving ~5 MB of live mapped DRAM unused below the 8 MB top
(only `animations_frame_buffer` @ `0x707FFD30` up there). Raised to
`0x70700000`; the ~4 MB gain goes to `MEMPOOL_STAGE` (`boss.c:218`). This
+ the re-applied 2x `g_GfxBuffers` removes the OOM hang: `-level_09` now
**renders ~5 frames** (`frame N rendered` logs, VI posts climbing).

The "runaway GDL append at ~frame 5" that D95 chased turned out to be
**two memory-corruption bugs**, both now fixed:

**D96 (`d86ec483`) — prop room-list stack overflow.**
`chrpropUpdateRoomList` + helpers build room lists of up to 7 entries,
then write `prop->rooms[0..n]` + a `0xff` terminator. `PropRecord.rooms`
and `chrpropsRenderPass`'s `s32 sp48[…]` local are both
`PROPRECORD_STAN_ROOM_LEN` = **4**. BUNKER1 patrol guards routinely span
≥4 rooms (`[1a 13 14 10]`, …) → 4 IDs, no terminator inside the array →
`chraiGetPropRoomIds`'s `for (i=0; self->rooms[i] != 0xff; i++)` walks
off the end, overflowing the caller's stack frame → garbage `gdl` →
GBI write fault. Every *other* `chraiGetPropRoomIds` caller already used
`s32[8]`. Fix: `PROPRECORD_STAN_ROOM_LEN` → 8 under `#ifdef PORT`
(`bondconstants.h`), + a defensive bound in `chraiGetPropRoomIds`. N64
unchanged.

**D97 (`2fbcc556`) — `bondviewPlayerTickDamageAndHealth` negative
`damagetype`.** US build (unlike EU/JP) has no low clamp;
`damagetype = (s32)(health*8)` goes negative on a lethal hit → OOB
`g_DamageTypes[]` read → segfault when a guard shoots Bond (~frame 5).
Extended the EU/JP clamp to PORT.

**D98 (`000ed6af`) — `initBONDdataforPlayer` under-allocates the player
struct.** It `mempAllocBytesInBank`s a hardcoded `0x2A80` (N64
`sizeof(struct player)`, `0x2A70` EU). The PC struct is much larger
(pointer fields widened 4→8). The player block sits directly below
`g_GfxBuffers[0]` in `MEMPOOL_STAGE`, so writes past ~offset `0x2A08`
(`bondviewRenderDebugBondView`'s `g_CurrentPlayer->field_2A08 = ft4`,
run every frame since `debug_render_raster` defaults to `DEB_BOND_VIEW`)
scribbled the zbuf-clear `gsDPSetRenderMode`'s `w1` onto master-DL slot
11 → `Unknown GBI opcode 0xffffb9`. Fix: allocate
`(sizeof(struct player) + 0xF) & ~0xF` under `#ifdef PORT`.

**D99 (`253caa23`) — `modelTickAnim` garbage function-pointer call,
FIXED.** `struct Model.animflipfunc` (`bondtypes.h:1640`, "0x98") is
`s32` but `modelSetAnimFlipFunction` (`model.c:2840`) stores a `void *` →
truncated on PC → `((void(*)(void))animflipfunc)()` at `model.c:3534`
jumped to `0x00010100` (`bheadFlipAnimation` at module+0x10100). It is
*only* ever set to `bheadFlipAnimation`, for `g_CurrentPlayer->model`
(`initBondDATAdefaults.c:198`, `bondhead.c:430`). Fix (D92/`unka0`
pattern): under `#ifdef PORT` the field is a bool flag and
`modelTickAnim` calls `bheadFlipAnimation()` directly. `-level_09` now
reaches VI post ~601 (was ~421).

**D100 (`8eaad547`, PARTIAL) — `struct player.model` is an inline
`struct Model`, not a pointer.** Every use is `&g_CurrentPlayer->model`
passed to a `modelXXX(struct Model *)` fn. The decomp splits it as
`Model *model;` + ~45 `s32 field_59C..field_650` (≈ N64
`sizeof(struct Model)` ~0xB8). PC `struct Model` is ~0x2A8, so
`animInit(&model, …)` overran into `field_654` (the gait RW-data pool
`animInit` was *also* handed), `bondheadmatrices`, and the viewport
fields → garbage `model->datas` → the bit-32 fault. Fix (PORT only):
`model` becomes an inline `struct Model`; a dedicated `u32
gaitRwData[256]` at the end of `struct player` replaces the `&field_654`
gait pool; `initBondDATAdefaults.c` points `animInit` there. `sizeof`
grows ~0x2A0 — D98's `sizeof(struct player)` alloc already covers it, and
`field_59C..field_650` are grep-verified dead (only `field_654` was
used). **⚠️ LANDMINE this exposes:** `struct player` has raw
hard-coded-offset accessors above 0x594 that are NOT PORT-adjusted and
are now further off — `gunfire.c:4934-4945` `THROWMTX/THROWPOS` at
`g_CurrentPlayer + handoffset + 0xAD8`, used for grenade/knife throwing.
Those were *already* PC-wrong (PC struct ≠ N64 before D100 too); D100
doesn't regress a working path, but a real `struct player` PC-offset
pass is owed. (`bondview2.c:3165` `+0x230` watch model and `+0x2ec` are
below 0x594 → unaffected.)

**D101 (`b2234f82`) — `sub_GAME_7F06DB5C` `(s32)` pointer stash, FIXED.**
The anim2-blend variant of `modelBuildGroupMatrices` did
`sp1C = (s32)arg2->Parent` then `modelFindNodeMtx(arg1, (ModelNode*)sp1C,
0)` — the exact 32-bit-pointer idiom the sibling was already cleaned up
for. `(s32)` drops the `0x140000000` base of a static `ModelNode` →
`modelFindNodeMtxIndex` deref `0x4012c9c0`. Also stashed `&sp48[sp54]`
(a `RenderPosView*`) the same way and mis-called `g_ModelJointPositionedFunc`
with a 3-arg cast. Under `#ifdef PORT`, full-width pointer work + the real
2-arg callback (mirrors `modelBuildGroupMatrices`).

**D102 (`1b078f6d`) — 1P weapon `Model` / RW-pool `struct hand` overlap,
FIXED.** `gunfire.c` punned the weapon `Model` onto `hand->field_B68` and
its RW-data pool onto `hand->modeldatas` (a fixed run at `field_B68 +
0x20`). N64 `sizeof(struct Model)` ~0xBC keeps `modelInit`'s header
writes clear of the pool; PC `struct Model` is 0xE8, so `modelInit`
writes `objinst->datas` (0x20) **directly onto `&hand->modeldatas`** —
the field and pool base alias, and the first `modelInitRwData` record
write clobbers `datas` (bit 32 set) → the `modelInitRwData` fault
(`0x1_70076514`). `objinst->obj` (0x10) lands on `hand->mtxlist`
likewise. Fix: `struct hand.weaponModel` (inline `struct Model`) +
`weaponRwPool[192]` under `#ifdef PORT`, routed through
`HAND_WEAPON_MODEL`/`HAND_WEAPON_RWPOOL` macros; set
`weaponModel.render_pos` explicitly (N64 aliased `hand->mtxlist`). Same
class as D53.2 (watch) / D100 (player gait model). The `d43` record
layout was ruled out — `GE_D51` proved it was `Objinst->datas` *field*
corruption, not a garbage `RwDataIndex` (`idx` was 0/1, valid).

**RENDER CHAIN CLEAR (session M-3).** With D93–D102 in, `-level_09` boots
BUNKER1 and **renders continuously with no fault** (60 s+, 3500+ VI
posts, full framerate; attract mode also clean). ~46 % non-clear pixels,
content biased to the lower half of the screen. The
stage-load→first-frame→in-level crash chain (the blocker since D69) is
**resolved**. Remaining: **D75** — 3D-model rendering quality
(animated/skeletal character models), a separate track; and the
`struct player` / `struct hand` raw-hardcoded-offset landmine above
(grenade/knife throw). `struct tex` headers 24 B vs 16 B on PC is a
separate open pool-pressure item — a PC memory-budget pass is owed.

**Still open, separate:** `struct tex` headers are 24 B vs 16 B on PC
(texpool-triage) — pool-pressure; a real PC memory-budget pass should
cover it.

**D103 (session M-4) — native-resolution height, FIXED.**
`port/src/libultra.c` `osViSetMode` hard-coded the fast3d native
resolution to the *visible scanline count* (`pal ? 400 : 480`) instead of
the CFB space GE authors its GBI in. fast3d's `SCREEN_WIDTH`/
`SCREEN_HEIGHT` (`gfx_pc.h`) alias `gfx_current_native_viewport`, and
`RATIO_X/RATIO_Y` (`gfx_pc.cpp:53-54`) scale the game's 320×240-space
`gSPViewport`/`G_SETSCISSOR`/2D-texrect coords onto the 640×480 window.
With height 480 the ratios were asymmetric — `RATIO_X = 640/320 = 2.0`
but `RATIO_Y = 480/480 = 1.0` — so every viewport/scissor came out a
220 px band (≈45.8 % of the frame → the "~46 % non-clear" figure),
placed in the upper GL region and shown (post `invert_y`) in the **lower
half of the screen**. `gfx_current_native_aspect` was likewise wrong
(320/480 = 0.67). Fix: recover the real CFB height from
`vm->fldRegs[0].yScale` (GE sets it to `bufy * YSCALE_MAX(0x800) /
SCREEN_HEIGHT_MAX(480)`, `fr.c:417`) with a `pal ? 272 : 240` fallback;
`comRegs.width` already carried `bufx` so width was fine. Verified:
`-level_09` `GE_PCDUMP` frames go from ~46 % / lower-band to **91.7 % /
bbox (0,20)-(639,459)** — the full 10..230 viewport ×2 with correct
letterbox, HUD ammo counter correctly placed top-right. Purely a
transform fix — no per-model work.

**D104 (session M-4) — investigation, folded into D105.** Chased "flat
dark-blue fill + HUD, no room geometry" after D103. Ruled out (probes +
subagent trace): segment 0x0E vtx resolution, room-DL chaining/execution,
per-room scissor (`GE_D104` shows bbox `(1,10)-(319,230)` full-view,
`viewleft/top/x/y = 0/10/320/220`), the view-projection matrix
(`field_10E0` decodes to a sane perspective×lookat — rows ≈ unit-scale
3×3, translate row `518,-22,430,459`), and the room GDL decode (valid
F3DEX2, no embedded `G_MTX`). A temporary fast3d tri counter proved
~160 k room triangles WERE being submitted to GL each frame — yet the
final image was 90 % exactly `(0,16,65)` (BUNKER1's `env->{Red,Green,
Blue}` sky-fill from `skyRender`, `sky.c:326`, the 1P no-clouds
`gDPFillRectangle` path). Geometry drawn, invisible → depth. `GE_D104`
room-render-pass probe left in `bg.c` (capped, gated).

**D105 (session M-4) — depth buffer never cleared, FIXED.**
`zbufClearCurrentPlayer` (`src/game/viewport.c:89`) clears the N64 Z
buffer with the classic "point the colour image at the Z buffer and
`gDPFillRectangle` it with a packed-Z fill colour" idiom. fast3d does
**not** emulate that: `gfx_dp_fill_rectangle` (`gfx_pc.cpp:2281`) bails
when `color_image_address == z_buf_address` on the assumption the depth
clear already happened via `glClear` — but the frame-start clear is
`clear_framebuffer(true, false)` (`gfx_pc.cpp:2855`), **colour only**.
fast3d expects the game to emit `G_CLEAR_DEPTH_EXT` (opcode `0x44`,
`gfx_pc.cpp:2678`) for a real depth clear, and **nothing in GE ever
emits it** (grep-confirmed). Worse, here the two addresses don't even
match (`gDPSetDepthImage(z_buffer & ~0x3F)` vs `gDPSetColorImage(
OS_K0_TO_PHYSICAL(z_buffer))`), so the bail isn't taken and the packed-Z
fill colour gets scribbled onto the real framebuffer instead. Net: every
in-level frame renders against a stale/garbage depth buffer, all ~160 k
room triangles fail the Z test, and only `skyRender`'s background fill
survives. Fix: `#ifdef PORT` branch in `zbufClearCurrentPlayer` emits a
bare `G_CLEAR_DEPTH_EXT` Gfx word and returns, skipping the N64 idiom
(which is a no-op / actively harmful in fast3d anyway). Verified
`-level_09`: frame goes from 90 % flat sky-colour to **2585 distinct
colours, sky down to 16 %** — recognisable BUNKER1: textured walls,
storage racks, floor. 70 s run, crash-free.

**D106 (session M-4) — portal near-plane projection garbage culled the
next room, FIXED.** After D105, many camera angles in BUNKER1 showed a
large flat sky-fill void where an adjacent room should be visible
through a doorway. A `GE_D104` probe (per-frame visible-room list) showed
the portal-visibility BFS returning only 1–3 rooms; a `GE_D106` probe in
`sub_GAME_7F0B7F84` showed the plane-side metric cull (`bg.c:4135/4144`)
and `zfar` (10000, fine) were NOT the cause — portals were dying in
`sub_GAME_7F0B5864`'s screen-space projection. When the player straddles
a portal plane, `sub_GAME_7F0B5528` emits z==0 near-plane clip points and
`transform3Dto2DWithZScaling` (`bondview.c:730`) uses `inv_z = -1e20`,
projecting them to ±1e20-scale coords. On N64 the two clip points
bracket the view symmetrically → the degenerate-box check (`bg.c:1713`)
or the downstream `screensize` clamp yields a full-screen box and the
room is kept. On x86-64 the exact garbage can come back `min > max` on
only one axis (or non-finite), slipping past `bg.c:1713` → portal
dropped → room vanishes → sky-fill through the doorway. Fix: `#ifdef
PORT`, any wildly out-of-range / non-finite projected bound is treated
as degenerate → full `screensize` (matches the N64 outcome). Verified:
per-frame visible-room count 1–3 → 2–4; rooms past near doorways (r11
from r27, etc.) now render. `c_screenleft/top/halfwidth/halfheight/
recipscalex/recipscaley` were all confirmed correct (0/10/160/110/
190.53/190.53), so the projection *fields* are fine — this was purely
the z==0 edge case. `GE_D104` probe (visible-room list) left in `bg.c`,
gated + capped; `GE_D106` probes removed.

**D107 (session M-4) — fast3d sampled an unloaded mip tile for GE's LOD
textures (blurry surfaces), FIXED.** GE room GDLs emit `G_TL_LOD +
G_TD_DETAIL` for mip-mapped textures, but `gfx_detail_textures_enabled`
is false for this port. `gfx_lod_tile_offset` (`gfx_pc.cpp:1258`) then
ran `rdp.tex_lod ? rdp.tex_detail : i` → returned `tex_detail` (1) for
every texel → fast3d sampled GE's first mip (render tile 1). GE loads
the whole mip chain with one `G_LOADBLOCK` to TMEM 0 and points tiles
1..5 at offsets inside it; fast3d keys `loaded_texture[]` by TMEM addr
and has no entry for tile 1's offset, so `import_texture` fabricated a
16×16 from the top-left 256 B of the base image and magnified it across
the polygon — the "blurry brown/grey ceilings & wall panels" in BUNKER1.
Fix: with detail textures disabled, `gfx_lod_tile_offset` always returns
0 (the base render tile — the only correctly-loaded level; matches the
N64 near-surface look). Verified: BUNKER1 ventilation room
ceilings/walls/light panels render crisp. Single-level textures (racks,
room 1) were already fine, unaffected. (Root-caused by a subagent — the
`gfx_detail_textures_enabled=false` + GE's unconditional `G_TD_DETAIL`
contradiction; hi-confidence, matches every symptom + the
room-1-vs-storage-room split.) A residual minor vertical squish from
`G_LOADBLOCK` size including the mip bytes is noted but not fixed.

**Storage-room sky void — root-caused, NOT a visibility bug (session
M-4).** The large void in rooms 27/28/29 is **closed doors**: portals 25
and 26 (the storage-area doorways) carry `controlbytes1 & 1`
(`PORTALFLAG_DISABLED`) at runtime — verified correct: `load_bg_file`
clears bit 0 for every portal at load (`bg.c:1015`), then
`bgToggleDataPortalsContrlBytes1Bit1` (`bg.c:5502`) re-sets it for
doors that start closed. So `sub_GAME_7F0B7F84` correctly refuses to see
through them (`bg.c:4147`). The bug is that **the door prop/model that
should fill each closed doorway does not render** — `chrpropsRenderPass`
(`chrprop.c:482`) *runs* with `n=3..10` props per room (`GE_D96`), but
the frame GDL barely advances (~7 `Gfx` for 9 props), i.e. props are
enumerated but emit almost no geometry. This is the prop/character
model-rendering track (D75(b) + the `struct player`/prop raw-offset
landmines), not portal/visibility work. Portal visibility itself is
healthy post-D106 (room 18 → 8 rooms deep; enabled doorways traverse
correctly). Other residuals: low-frequency stray green/yellow polygons
(1–2 per room — a degenerate vertex, likely one bad entry per room-vtx
table via `bgSwapRoomVtx` or the point-index blob length); front-end
text draws mirrored. Next: prop/door model rendering (unblocks the void
AND the missing weapon), then D75(b) skeletal models.

**D108–D112 (session M-5) — the "props emit no geometry" premise was
stale; the real bug was a converter byte-swap off-by-one, FIXED.**
Re-probing `chrpropsRenderPass` / `chrpropRender` / `modelRenderNodeDl`
(`GE_D96*`) showed props *do* emit complete leaf DLs now (~15k
`nodeDl` EMITs/run, valid `gdl`/`vtx`/`BaseAddr`; fast3d transforms
~300–600 prop verts/frame with sane `w`). Two separate residual bugs
were isolated: (1) in the BUNKER storage area the per-frame visible-room
count drops to 1–2 while `cameramode` goes 1→0 and `g_RoomLoadBudget`
200→3 — but the 200→3 is *intentional* N64 behaviour
(`bgRoomVisibilityRelated`: 0xC8 for intro/swirl cams, 3 for FP) and
`currentPlayerGetProjectionMatrix()` + `g_CurrentPlayer->screensize`
were both verified valid, so this is a portal-BFS under-reach, still
open. (2) **Skeletal characters rendered as a "3D line" / not at all**
(also visible in attract mode): `drawjointlist` → `subcalcmatrices` →
`modelUpdateMatrices` → `process_02_position` → `modelBuildGroupMatrices`
produced `render_pos[]` joint matrices with **sane rotations but garbage
/ 1e27 / NaN positions**. `basemtx` (`camGetWorldToScreenMtxf()`) was a
clean orthonormal lookat; joint *rotations* from the anim bitstream were
fine; but `group->Origin` (`ModelRoData_GroupRecord.Origin`, a
`coord3d`) read as garbage for every joint whose bytes weren't
zero — while the adjacent `JointID`/`MatrixIDs` (u16/s16) read fine.
Root cause: **`put_f32` in `tools_pc/d43_emit.py` had an off-by-one in
its BE→LE byte reversal** — `buf[o:o+4] = src[doff+4:doff:-1]` yields
bytes `doff+1..doff+4` (dropping the value's MSB, pulling in one byte of
the next field) instead of `doff..doff+3`. This corrupted *every* f32
field in converted model rodata (joint `Origin`s, LOD near/far, BSP
planes, bounding radii) for every sidecar-loaded model; the compiled-in
front-end intro models (native-LE C structs, not run through the
converter) were unaffected, which is why the Rareware logo looked right
but every in-level character collapsed. Fix: `src[doff:doff+4][::-1]`.
After regen (`python tools_pc/d43_emit.py ntsc-final`), all 20
`render_pos` entries of a BUNKER guard are coherent (~±60 units around
the body), monitor screens draw their content, and character models
render as recognisable humanoids (still some pose/orientation polish
owed — a guard appears inverted — a separate matrix-handedness item).
`d43_emit.py`'s verify pass only checks pointers/opcodes, not float
values — a float spot-check is worth adding. Probes `GE_D96GATE /
D96DL / D108 / D109 / D110 / D111 / D112 / GE_TRICNT` were all removed
after the fix.

**D113 (session M-6) — "portal-BFS under-reach / storage-room void" was a
STALE premise; there is no BFS bug.** A subagent dumped BUNKER1's runtime
portal table and traced every accept/reject decision of `sub_GAME_7F0B7F84`
(non-EU/LEFTOVERDEBUG copy, `bg.c:4091`; queue = 5-arg `bgQueuePortalTraversal`
`bg.c:3843` — confirmed the compiled ones). With the camera in room 29 the
BFS visits portals 25,26,27,28: portals 25 (r29↔r26) and 26 (r29↔r27)
carry `PORTALFLAG_DISABLED` (`controlbytes1 & 1`) because door props obj
153/154 spawn `openPosition==0` → `doorDeactivatePortal`, and nothing
re-opens them in the fly-through — so they are culled at `bg.c:4138`, a
site that is **not** `#ifdef PORT`-guarded and behaves identically on N64.
Result: a stable 3 visible rooms {29,28,25} (D104 probe agrees), which is
the correct answer for this topology, not "1–2". No s32/pointer
truncation, no `min>max`/NaN reaching an accept flip, no depth-cap
(`D_8004489C=0xF`) or visited-set (`>=9`) early-out firing. `sub_GAME_7F0B5864`
is fine post-D106. **The void = the two closed-door models (obj 153/154)
filling rooms 26/27's doorways not rendering where they should** —
`GE_D96` shows room 29's prop pass emits their leaf DLs, so the door
geometry exists but its world transform puts it outside the doorway,
leaving the wall hole unoccluded → sky. This is the D114 model-transform
track, not visibility. No `bg.c` change. Probes reverted.

> **M-33 CORRECTION (D168): this finding is retracted as a capture artifact.**
> Every `GE_PCDUMP` PPM was written upside-down (`gfx_opengl_dump_bound_fbo`
> did not reverse `glReadPixels` rows). The developer confirms the game renders
> correctly on real hardware. The "shared fast3d mirror" this and the D116
> entries chase never existed — the probes kept coming back clean because there
> was nothing wrong; the analysts were reading inverted screenshots of
> asymmetric content (guards, text, logo) and calling them "mirrored". The PPM
> writer is fixed. The genuinely-open item that remains is D75 Bug 2 (front-end
> animated models absent), which is unrelated. Original text kept below for
> history.

**D114 (session M-6) — character inversion / mispositioned crate:
OPEN, matrix chain verified sane, residual is a shared fast3d mirror
(D75 class).** A subagent statically traced the whole joint-matrix chain
(`chr.c:2564` → `subcalcmatrices` → `modelUpdateMatrices` →
`process_02_position` `model.c:1662` → `modelBuildGroupMatrices`
`model.c:1318` → `matrix_4x4_multiply_homogeneous` →
`bondviewTransformManyPosToViewMatrix` → `matrix_4x4_f32_to_s32`
`matrixmath.c:495` → `gfx_sp_matrix`) and an env-gated `GE_D114` probe
(24 joint builds, 3+ BUNKER guards). Findings: `basemtx` from
`camGetWorldToScreenMtxf()` is a proper orthonormal right-handed lookat
(+Y up, `up ≈ (-0.09, 0.96, -0.26)`, `right = up × forward`), scaled by
`bgGetLevelVisibilityScale()` = 0.1 (also applied to the room path —
expected); joint `render_pos` are coherent and **upright** (head y ≈ 280
> pelvis ≈ 230 > limb ends ≈ 132); view-space z negative (in front,
correct). Converter clean: `put_f32` is now the only float writer and is
correct post-D112; `matrix_4x4_f32_to_s32` is byte-identical to
`guMtxF2L`; `matrix_4x4_set_lookat` / `guPerspectiveF` /
`matrix_4x4_multiply_homogeneous` match N64 and are native-LE. So the
plan's hypotheses (converter axis/sign, lookat handedness, compose
order, F2L) are all **disproven**. Leading residual: a horizontal (±
vertical) mirror in the **shared fast3d viewport / MP-matrix path**, not
the model path — consistent with the open D75 notes ("front-end text
draws mirrored", "Nintendo logo mispositioned"): a flip is only visible
on asymmetric content (humanoid, text, logo) and invisible on a boxy
room. GE bakes a left-handed screen convention into its own transforms
(`transform3Dto2DCoords`, `bondview.c:726`: `screenX = center −
x·invz·scale`); if GE's RSP (`rsp/graphics/gmain.s`) applies an X flip
that `port/fast3d` does not, every model and room is mirrored. Suspects:
`gfx_calc_and_set_viewport` / `gfx_adjust_viewport_or_scissor`
(`gfx_pc.cpp:1739/1771`), `MP_matrix` vertex transform
(`gfx_pc.cpp:1122`). This is a fast3d-correctness gap → write-up, not a
narrow ABI patch. Also: **rebuild with D112+D115 and eyeball a guard
first** — the D115 `gunfire.c` fix stops a per-shot 64-byte scribble into
the inline-Model / bondhead-matrix region that could itself have caused
"inverted". The crate: static-prop placement comes via a per-prop
`basemtx` from the Usetup `PropRecord` position/rotation — check
`d88_emit.py` separately from the character question. Probe reverted;
re-apply snippet in the M-6 agent report.

**D115 (session M-6) — `struct player` / `struct hand` raw-offset audit +
first fix.** Full survey in `docs/dev/AUDIT-M6-player-offsets.md` (10 offset
sites; 3 live HIGH, 1 dead const block, 2 MED, 4 already-correct).
**Fixed:** `gunfire.c:4960-4962` `THROWMTX` / `THROWPOS(k)` / `THROWPREV(k)`
were raw byte offsets — `(u8*)g_CurrentPlayer + handoffset + 0xAD8/0xB08/
0xB48` with `handoffset = handnum * sizeof(struct hand)`. Both the stride
(`sizeof(struct hand)` ≈ 0x968 PC vs 0x3B8 N64, D102) and the base offset
are N64-sized, so on x86-64 the address lands inside the inline gait
`Model` / bondhead-matrix region of `struct player`, and
`matrix_4x4_copy(THROWMTX, …)` scribbles 64 bytes of live render state on
**every shot from a casing-ejecting weapon** (`sub_GAME_7F068508`,
solo only). The offsets are exactly
`hands[handnum].throw_item_pos_related[_prev]` and its translation row, so
the macros now use those field accessors under `#ifdef PORT` (N64 build
keeps the byte-offset macros unchanged). **Still open (MED):** #5 — the
D102 weapon-model `render_pos` is pointed at a `dynAllocate`'d transient
arena that on N64 aliased the persistent `hand->mtxlist`; likely the
"1P weapon model doesn't draw" cause, may fold into D114. #6 — the D56
watch-preview Model pool (`watchRwPool[0xC8]`) is N64-sized and the
inline Model overruns live watch fields. See the audit doc for the full
table and recommended fix order.

> **M-33 CORRECTION (D168): retracted as a capture artifact — see the D114
> correction above and D168.** "Every stage probed is clean, yet the output is
> mirrored" was the tell: the output was not mirrored, it was upside-down (the
> PPM writer never reversed `glReadPixels` rows). No `GE_D116` S-swap / U-flip
> was ever warranted. Probe scaffolding (`GE_D116` in `textrelated.c` /
> `gfx_pc.cpp`) can be stripped. Original text kept below for history.

**D116 (session M-7) — the HUD/menu text mirror (D75 class) is a
per-glyph TEXTURE-space flip, NOT a screen/framebuffer/matrix mirror.**
Rebuilt at D115 (`14b6b432`), build green, captured `-level_09` frames
(`GE_PCDUMP="120-600:60"`, `ppm/frame_0003*.ppm`). Findings:
- The "OBJECTIVE C: FAILED" flash (proportional font,
  `textrelated.c:textRenderGlyph` → per-glyph `gDPLoadTextureBlock`
  `G_IM_FMT_I/G_IM_SIZ_8b` + plain `gSPTextureRectangle`, dsdx `0x400`)
  renders with **string order preserved (O first, D last) but every
  glyph individually horizontally mirrored**. That signature is a
  texture-S reversal per glyph, not a block/screen flip.
- The HUD ammo digits ("83", top-right, upright and correct) use a
  different font path and are **not** mirrored → the bug is specific to
  the proportional-font `textRenderGlyph` load/sample path, not global.
- Static trace found nothing: `gfx_dp_texture_rectangle` / `gfx_draw_rectangle`
  (`gfx_pc.cpp:2161/2086`) assign `uls`→left corner, `lrs`→right corner
  with positive dsdx; `import_texture_i8` (`gfx_pc.cpp:810`) uploads
  linearly; viewport/scissor paths have no X negation; GE's camera
  (`fr.c:694` `guPerspectiveF` + `matrix_4x4_set_lookat` = standard
  gluLookAt) is conventional RH (confirms D114).
- **Consequence for D114:** the unified "shared fast3d screen mirror"
  hypothesis is at least partly wrong — the text mirror is texture-space.
  The inverted-guard / mislocated-crate symptoms are either a *separate*
  defect or the same texture-S flip applied to model textures (an
  asymmetric guard skin flipped reads as "facing wrong way").
- **Next (needs a runtime probe, not static):** dump the glyph texrect's
  tile size / `line_size_bytes` / uploaded texture width vs `curchar->width`
  for one glyph, and the final S texcoords fast3d hands GL. Suspects:
  `gDPLoadTextureBlock` 8b `line`/pad (`(curchar->width+7)&0xF8` load
  width vs real `curchar->width` sample width) interacting with tile
  wrap; or the compiled-in font blob (`assets/obseg/text/LmiscE.h` /
  `chars[]` `pixeldata`) being 32-bit word-swapped by the asset step
  (would reverse 4-texel groups — check whether the mirror is clean or
  chunked at 4px). Probe artifacts: `ppm/frame_000300.ppm` (obj text),
  `ppm/frame_000480.ppm` (ammo).

**D116 probe results (session M-7, overseer-run — supersedes the "specific
to the proportional font" claim above).** Killed a subagent that was
drifting toward a global `GE_D116FLIP` S-swap on every non-flip texrect
(mirrors the whole HUD/3D to "fix" glyphs — wrong). Kept its
`GE_D116`-gated `fprintf` probes in `textrelated.c` (both glyph paths) and
`gfx_pc.cpp:gfx_dp_texture_rectangle`. Built green at D115, ran
`-level_09` + `GE_PCDUMP="200-360:40"`. Data:
- **The ammo digits ARE mirrored too.** `ppm/frame_000320.ppm` top-right:
  "83" renders with each digit individually X-flipped, digit order
  preserved; the clip-count glyph left of the mag icon likewise. So the
  prior session's "ammo digits render correct → bug is `textRenderGlyph`-
  specific" differentiator is **FALSE at D115**. The mirror is NOT
  path-specific. (`textRenderGlyphOutlined` is what the level text uses;
  `-level_09` direct-boot renders no "OBJECTIVE" flash / crosshair at
  all — only the ammo HUD — so the M-7 "OBJECTIVE C: FAILED" note came
  from a different scenario/run.)
- **Rect screen position is correct; only the texture content is flipped.**
  Ammo HUD sits top-right and renders top-right. So this is a texture-U
  reversal per quad, not a screen/framebuffer/viewport flip and not a
  scissor issue.
- **Every stage probed is clean, yet the output is mirrored:**
  - glyph bitmap in memory: correctly oriented — dumped `curchar->pixeldata`
    for 'D' (idx 68, w7/loadw8/h9) row-by-row; stroke on the left, bowl on
    the right, col 7 = wrap-pad of col 6. Not reversed. Rules out (c) a
    font-blob word-swap.
  - `gDPLoadTextureBlock` params: `loadw = (w+7)&0xF8` = 8 (or 16 for w9),
    `line_bytes` = 8, tile `siz=1 fmt=4 cms=2`. Consistent. No evidence
    the load-width vs sample-width interaction reverses anything — rules
    out (b) as the *primary* cause.
  - fast3d texrect: `ul.u = 0`, `lr.u = <positive max>` (e.g. 256 = 8
    texels); `ul.x < lr.x`. Corner↔UV pairing is correct
    (`gfx_dp_texture_rectangle` 2187-2211, `gfx_draw_rectangle` 2113-2136:
    ul=(ulx,uls), lr=(lrx,lrs)). Rules out (a) at the texrect layer.
  - `import_texture_i8` (`gfx_pc.cpp:810`): strictly linear byte copy, no
    row/col reversal.
  - `gfx_opengl.cpp`: `vTexCoord = aTexCoord` pass-through; no U negation
    in the vertex/fragment shader; `cms=2`→`GL_CLAMP_TO_EDGE`.
- **Conclusion:** the U-flip is downstream of everything fast3d computes —
  in the GL vertex-buffer assembly / draw, OR it is a shared transform on
  the rect quad's clip-space X that desyncs from U. This **re-opens
  D114's shared-mirror hypothesis** (relocated from screen-space to
  per-quad U/X space) and retires the D116 "proportional-font-specific,
  texture-S load path" framing. Same mechanism plausibly explains the
  "inverted" guards (mirrored skin) and mislocated door props (D113).
- **Next:** shader/vertex-buffer-level probe — dump the actual per-vertex
  (x, u) pairs in the buffer handed to `glDrawArrays` for one glyph quad,
  and render a 1-texel asymmetric test texture on a known rect to see
  which axis inverts. Confidence the mirror is a real per-quad U/X flip
  (not a capture artifact): **high** — reproduced on two independent HUD
  text paths, rect positions provably correct. Confidence in the
  GL-layer-vs-shared-transform split: **low** — not yet isolated.
- Probes left in tree (all `#ifdef PORT` + `getenv("GE_D116")`, zero-cost
  when unset): `textrelated.c` textRenderGlyph / textRenderGlyphOutlined;
  `gfx_pc.cpp` gfx_dp_texture_rectangle.

**D116 code-trace follow-up (session M-7, overseer, static — no build).**
Traced the rect quad end-to-end through fast3d for the I8 glyph case
(`cms=2`=CLAMP, `tile.uls=0`, `dsdx=1024`, 8px glyph):
- `gfx_dp_texture_rectangle` 2187-2211: `ul={x:ulxf, u:uls=0}`,
  `ur={x:lrxf, u:lrs=256}`, `ll={x:ulxf,u:0}`. `ulxf<lrxf`. Correct pairing.
- `gfx_sp_tri1(ul, ll, ur, is_rect=true)` -> `gfx_sp_tri1` 1497-1572: for
  each vtx `u = v->u/32 - tile.uls/4`; the D74 wrap block is gated on
  `cms & G_TX_WRAP` and G_TX_WRAP==0 so it is a no-op for CLAMP glyphs;
  `is_rect` skips the persp/filter half-texel. Result: `buf_vbo.u` =
  `0/8 = 0.0` for left vtx, `256/32 / 8 = 1.0` for right vtx.
- `gfx_adjust_x_for_aspect_ratio` 1092: `(aspect_ofs*w + x) * aspect_scale
  / aspect_ratio`. `aspect_scale` (1714) is always a positive aspect
  ratio; `aspect_ofs` is a monotonic shift. **Cannot invert X.** Ruled out.
- `buf_vbo` X for a rect = `v->x` verbatim (clip space from
  `gfx_draw_rectangle` 2105-2111), Y optionally `invert_y`-negated (1504),
  X never negated.
So the **entire fast3d 2D texrect -> vertex-buffer path emits a correct,
non-mirrored quad** (x-left<->u=0, x-right<->u=1). Whatever flips U is
DOWNSTREAM of `buf_vbo`: the GL vertex/fragment shader, the ortho/MVP the
backend applies to direct (non-fb) draws, or the sampler. That also means
it is testable without determinism (static per-quad property).
- **Still owed (needs a build — currently serialized behind the
  determinism/framediff agent):** (1) a probe dumping `buf_vbo` (x,u) for
  one glyph tri right before `glDrawArrays`, to confirm the CPU-side
  buffer is non-mirrored as traced; (2) render a 1-texel asymmetric test
  texture on a known screen rect to see which axis GL inverts; (3) check
  whether 3D world geometry is ALSO X-mirrored (bunker is near-symmetric
  — need an asymmetric in-world texture or a guard-facing check) to tell
  a HUD-only (direct-draw) flip from a global one.

**D116 runtime probe part 3 (session M-8, overseer). CONTRADICTION —
every stage from font-bitmap to GL-draw verified non-mirrored at runtime,
yet the on-screen glyphs are unambiguously X-flipped. Investigation
budget-capped; deprioritised (cosmetic, not a playability blocker).**
Confirmed symptom (zoomed `ppm/frame_000300.ppm`): "OBJECTIVE C: FAILED"
(via `textRenderGlyphOutlined`) AND the ammo "83" + spare-clip digit (a
*separate* HUD-number path — `textrelated.c` probes never fire for the
digit indices) both render with word/column order preserved and every
glyph individually horizontally mirrored; screen positions correct.
Runtime probes added (`GE_D116`-gated, `#ifdef PORT`):
- `[D116/vbo]` in `gfx_sp_tri1` after the per-vertex `buf_vbo` writes:
  for the OBJECTIVE glyph quads (left edge, `tile.uls=0 lrs=28`,
  `cms=2`=CLAMP) it prints `vtx0 x=-0.8125 u=0.0` and `vtx2 x=-0.7625
  u=1.0`. **x-left <-> u=0, x-right <-> u=1. The CPU vertex buffer handed
  to GL is correct and non-mirrored.**
- `[D116/i8up]` (since reverted) in `import_texture_i8` dumped the
  uploaded bytes for glyph 'D' — **byte-identical** to `curchar->pixeldata`
  in memory (`00 00 0f 37 34 0b 00 00` ...), and the bitmap is correctly
  oriented (stroke col 1, bowl col 6-7). No flip, no word-swap.
- GL backend: vertex shader is `gl_Position = aVtxPos` (fast3d does all
  transform on the CPU — no GL matrix can flip anything);
  `vTexCoord = aTexCoord` pass-through fragment path; sampler
  `GL_CLAMP_TO_EDGE`/`GL_NEAREST`; `gfx_opengl_copy_framebuffer` has only
  a `flip_y`, no X flip; `G_TEXRECT` decode (`gfx_pc.cpp:2602`) extracts
  `ulx<lrx` correctly; `gfx_adjust_x_for_aspect_ratio` is a positive
  scale+shift (cannot invert).
- `GE_D116TEST` (reverted) replaced glyph I8 textures with a
  left-opaque/right-transparent split; on screen the digit slots came
  back near-uniform white, not half-and-half — **inconclusive** (the
  9-pass outline multi-draw and/or the unidentified digit path's sampling
  window smear the result).
- **Conclusion + confidence:** the flip is real (HIGH — two independent
  text paths, clean zoomed capture). It is NOT in: the font asset, the
  texture upload, the fast3d texrect->vbo math, the GL shaders/sampler,
  the framebuffer blit (each HIGH, runtime-verified). Where it IS: unknown
  (LOW). The contradiction means a stage is being mis-modelled — leading
  candidates now: (a) the render-*tile* setup (`gDPSetTile`/`gDPSetTileSize`
  S params — probes covered the *load* tile) vs fast3d's tile-window
  sampling; (b) a GL-driver-level surprise only a real API trace
  (RenderDoc/apitrace) would show; (c) the dedicated ammo-digit renderer
  (unidentified — not in `textrelated.c`) emitting mirrored S, with
  `textRenderGlyphOutlined` doing likewise via a shared lower-level
  helper. NEXT PERSON: capture one glyph texrect in RenderDoc, or find
  and read the HUD-number path, before touching fast3d again. Do not
  exceed ~30 min without one of those in hand. Probe left in tree:
  `[D116/vbo]` in `gfx_pc.cpp` (`GE_D116`, zero-cost).

**D117 (session M-8) — frame-to-frame nondeterminism root-caused;
`GE_DETERM` fixed-tick mode assessed NOT-narrow, deferred with a design.
`tools_pc/framediff.py` added (structural/tolerant).**

*Root cause — pure frame pacing (variable timestep), NOT PRNG / uninit
state. Confidence: HIGH.*

GE is a variable-timestep simulation. Per rendered frame it advances game
logic by `deltaFrames` = *however many 60 Hz ticks of wall-clock elapsed
since the previous frame*:

- `src/game/frametiming.c:75` `waitForNextFrame()` —
  `nextFrameTime = (osGetCount() - copy_of_osgetcount_value_1 + 387937)
  / 775875` (NTSC: 775875 RSP-counter ticks per 1/60 s), loops until
  `>= frameDelay` (normally 1), then `updateFrameCounters(nextFrameTime)`.
- `src/game/frametiming.c:46` `updateFrameCounters(deltaFrames)` sets
  `speedgraphframes = deltaFrames`; everything downstream (physics, AI,
  animation blends, `g_Vars.lvupdate*`) scales by it.
- `src/boss.c:456-495` main loop: blocks on `gfxFrameMsgQ` for
  `OS_SC_RETRACE_MSG` (posted by the port pacemaker), and only builds a
  new frame once `mainTickElapsed = osGetCount() - copy_of_osgetcount_value_1
  >= MAIN_LOOP_TICK_INTERVAL` (387937). A two-level gate, both levels
  keyed on `osGetCount()`.

On the console `osGetCount()` is the CP0 Count register (fixed CPU rate);
on PC `port/src/libultra.c:88-100` maps it to **real elapsed
microseconds** scaled to 46.5525 ticks/µs (D52 — required so
`waitForNextFrame` doesn't block ~388 ms/frame). So `deltaFrames` tracks
real render time: a frame that took 28 ms advances logic 2 ticks, one
that took 14 ms advances 1. Machine load, GL driver, vsync phase and the
~20 fps clean-run rate all jitter this → two runs of the same build take
different numbers of logic steps to reach "frame N" and their sim
trajectories diverge. Measured with `framediff.py --exact`: two
`-level_09` runs differ **32 % (frame 440) – 69 % (frame 200)** of pixels.

Ruled out as *primary*:
- **PRNG seeding is correct.** `port/src/random.c:22`
  `g_randomSeed = 0xAB8D9F7781280783ULL` (the two `.word`s from
  `random.s`), `g_chrObjRandomSeed` likewise (`:93`); the xorshift is a
  line-by-line port. `randomSetSeed` matches the `.s` (`+1` before
  store). *However* — because AI/animation code calls `randomGetNext()` a
  `deltaFrames`-dependent number of times per frame, the PRNG *stream
  position* still diverges between runs. It is a victim of the pacing
  jitter, not an independent source.
- **Uninitialised state:** not investigated exhaustively, but the
  `--exact` divergence grows smoothly from frame 200→445 rather than
  being present at frame 1, which is the signature of accumulated
  timestep drift, not a per-run uninitialised seed.

*`GE_DETERM=1` fixed-tick mode — assessed NOT NARROW, deferred. Confidence
that it's not narrow: MEDIUM-HIGH.*

The obvious hook — make `osGetCount()` a virtual clock that advances
exactly 775875 ticks per presented gfx frame (`videoEndFrame` in
`osSpTaskStartGo`, `port/src/libultra.c:1187`) — deadlocks the `boss.c`
main loop. That loop only presents a frame *after* the retrace gate
`mainTickElapsed >= 387937` passes, and with a frame-coupled clock
`mainTickElapsed` is 0 until a frame is presented → the first in-loop
frame never renders. `boss.c:442` (`waitForNextFrame()` after
`lvlStageLoad`, before any frame) hangs the same way. Breaking the
seal requires the **VI retrace post itself** to drive the virtual clock
(advance a fixed quantum per `portPostVIEvent`) *and* the pacemaker to be
frame-gated so it can't enqueue >1 retrace per render (else `deltaFrames`
jumps to the queue depth). That is a redesign of the port pacing model
(`portTickThread` / `portPostVIEvent` / `osGetCount`), with real deadlock
risk in loops that pump retraces without presenting — the `boss.c:448`
`NOBLOCK` drain, multi-frame stage loads, the pause menu, `front.c`
menu loops. It is contained in `port/` and env-gatable, but it is not
"a few lines, obviously correct" — it changes retrace/tick semantics, so
per AGENTS.md it is written up here rather than patched.

**Recommended design (for a future dedicated pass):**
1. `GE_DETERM=1` → `portTickThread` stops pacing on wall clock. Instead:
   a global `g_determFrameReady` flag is set by `videoEndFrame`; the tick
   thread posts exactly one `OS_SC_RETRACE_MSG` and advances a virtual
   `g_determTicks += 775875` (931050 PAL) **only** when it sees a new
   presented frame (or when `g_viRetraceMQ->validCount == 0` and no frame
   is pending — to service pre-first-frame / load-screen waits, advancing
   by the same quantum so `waitForNextFrame` sees exactly 1).
2. `osGetCount()` returns `g_determTicks` verbatim in this mode (no
   sub-frame interpolation — `store_osgetcount`/profiling just see 0
   deltas, which is harmless).
3. Seed `g_determTicks = 775875` at `portKernelInit` so the first
   `waitForNextFrame` computes `(775875 + 387937)/775875 == 1`.
4. Leave `osGetTime()` (µs wall clock) alone — audio mixing cadence and
   the heartbeat watchdog should stay real-time.
5. Prove with `framediff.py --exact`: two
   `GE_DETERM=1 GE_PCDUMP=... ` runs must produce ~0 % pixel diff
   (allow `--tol 2` for GL dithering). Then regenerate
   `tools_pc/golden/` with `--update` and switch CI to `--exact`.
Risk to watch: any game loop that calls `waitForNextFrame()` in a context
where no gfx task will be submitted (true loading spinners) — those need
the step-1(b) "no frame pending" fallback or they hang.

*`tools_pc/framediff.py` — DONE, committed. Confidence: HIGH (validated).*

Structural/tolerant by default (no determinism to lean on): per-frame it
computes (a) 16×12 grid-cell mean-RGB delta, (b) whole-frame non-clear
(non-black, pixcount.py rule) pixel-% swing, (c) a 16×16 aHash Hamming
distance; fails the frame if any exceeds its threshold. `--mask
X0,Y0,X1,Y1` (repeatable) drops known-animating regions (HUD) from all
three. `--exact` mode (per-pixel, `--tol`/`--tol-pct`) is there for a
future deterministic build. `--update` refreshes the golden set. Reads
`.ppm` (GE_PCDUMP) and `.png` (golden) on either side; PNG decoder is
built in (stdlib `zlib`; falls back to PIL for odd formats). Validated:
two nondeterministic re-runs of HEAD pass structural (worst cell
dmean 15, phash ≤ 23, non-clear Δ ≈ 0) while `--exact` correctly reports
32–69 %; a deliberately wrong frame pair fails (18 cells, phash 105).
Golden set at `tools_pc/golden/frame_0002{00,320,440}.png` is the D115
baseline — since it is a nondeterministic capture, only structural mode
is meaningful against it today.

**D117 — M-52 addendum: `GE_DETERM=1` implemented per the design above,
tested live (Windows console). Boot deadlock FIXED as predicted; full
determinism NOT achieved. Confidence: HIGH on both findings (live-tested,
not argued).**

Implemented exactly the 5-step design (`port/src/libultra.c`, env-gated,
default off, zero change to the non-`GE_DETERM` path): `osGetCount()`
returns a virtual `g_determTicks` seeded to the region quantum; a
`g_determFrameReady` flag is set at the `osSpTaskStartGo` graphics-task call
site (i.e. by the real "a frame was presented" event, not `videoEndFrame`
directly — same effect, one call site, no `video.c` change needed);
`portTickThread` gets a `GE_DETERM` branch that advances the clock + posts
one retrace only when that flag is set, polling every 200 µs instead of
pacing to the real 16.7/20 ms tick interval.

*Step 1(b) risk, confirmed exactly as written.* Without a load-phase
fallback, `-level_09 GE_DETERM=1` hangs before the first frame: symbolicated
the live hang (not guessed) — `shedThread` blocked in `osRecvMesg`
(`sched.c:256`, `__scMain`) and `mainThread` in `joyCheckStatusThreadSafe`
(`joy.c:166`), both waiting on the scheduler's interruptQ (the same queue
`portPostVIEvent` posts into, per the D134 comment), which never receives a
retrace because no frame has rendered yet to set `g_determFrameReady` —
exactly the chicken-and-egg the design called out ("boss.c:442 hangs the
same way").

*Fallback added, per the design's own suggestion — fixes the hang, breaks
determinism.* Added: when `g_determFrameReady` is false AND
`g_viRetraceMQ->validCount == 0` (nothing left to consume from a prior
tick), advance the clock and post a retrace anyway, so a boot-time waiter
isn't stuck forever. This **does** fix the hang — `-level_09 GE_DETERM=1`
now boots and renders past frame 3000+ instead of hanging. It does **not**
achieve determinism: two independent runs, same build, same command,
`framediff.py --exact` at the golden frame numbers (200/320/440):

```
FAIL frame_000200: 276998/307200 px over tol (90.169%, maxchan=218)
FAIL frame_000320: 85131/307200 px over tol (27.712%,  maxchan=102)
FAIL frame_000440: 47936/307200 px over tol (15.604%,  maxchan=102)
```

That's *worse* at frame 200 than the port's existing wall-clock
nondeterminism without any `GE_DETERM` mode at all (32–69 % per the
original D117 measurement above) — the fallback didn't just fail to fix
determinism, it introduced a new, bigger source of it. Root cause (analysis,
not yet re-tested): the fallback fires whenever the 200 µs poll happens to
observe an empty queue, which during the load phase depends on real
OS-scheduling latency between the polling thread and whatever the load path
is doing — so the *number* of extra ticks consumed before the first real
frame renders varies run-to-run, exactly the same class of divergence D117
already documents, just relocated from "wall-clock frame pacing" to
"wall-clock load-phase tick count." User-observed side effect, consistent
with this: with the fallback active the game visibly runs at a different
apparent frame rate and sped up — because the tick loop is no longer paced
to real 60/50 Hz at all once it starts free-running on empty-queue polls,
it advances (and lets the sim advance) as fast as the poll thread schedules,
not at any fixed rate.

**D117 — M-52 2nd attempt (same session): moved generation from a polled
fallback to a synchronous one. Root-caused precisely why that still isn't
enough. Confidence: HIGH (live-tested, mechanism traced through sched.c).**

Traced the actual consumer chain instead of guessing: `sched.c` `__scMain`
is the **sole** consumer of `g_viRetraceMQ` (`sc->interruptQ`), and every
retrace it processes drives task execution (`osSpTaskStartGo`), `joyPoll()`
(which is what unblocks `joyCheckStatusThreadSafe` too — the other M-52
1st-attempt hang site, for free), and the forward to `mainThread`, all
**synchronously on that one thread** before it asks for the next message.
So replaced the poll-driven tick-thread fallback with generation **inside
`osRecvMesg()` itself**: when `GE_DETERM` is on and the sole consumer of
`g_viRetraceMQ` finds it empty, synthesize the message inline (advance
`g_determTicks`, enqueue) instead of blocking — driven by call sequencing,
not by an independent thread's wall-clock poll.

Result: still boots (no regression on the deadlock fix), and the divergence
profile changed but did not go to ~0%: two runs, `framediff.py --exact`:

```
FAIL frame_000200: 101582/307200 px over tol (33.067%, maxchan=150)
FAIL frame_000320: 116844/307200 px over tol (38.035%, maxchan=161)
FAIL frame_000440: 43617/307200 px over tol (14.198%, maxchan=117)
```

Better than the 1st attempt's 90/28/16% at frame 200, but still nowhere
near the ~0% acceptance target. **Root cause, found by diffing the two
boot logs line-for-line** (not guessed): `boss.c`'s boot sequence has a
genuine, bounded, wall-clock real-time wait —
`for (i = 0; i != MAXCONTROLLERS; i++) { osSetTimer(100ms); osRecvMesg(BLOCK); }`
(`src/boss.c:188-190`, controller detection) — untouched by `GE_DETERM` by
design (`osGetTime()`/real timers are explicitly out of scope). During
that ~400 ms real window, **nothing is yet consuming `__scMain`'s forwarded
retraces** (`mainThread` hasn't reached its own retrace-consuming loop
yet), so the new synchronous-generation code has nothing throttling it —
`__scMain` free-spins through `osRecvMesg` as fast as the OS schedules that
thread, and *how many* virtual ticks accumulate during that fixed 400 ms of
real time depends on real thread-scheduling throughput. Same nondeterminism
class as both prior attempts, relocated a third time: generation-on-ask is
only deterministic when something else deterministically paces the asking,
and right now nothing does during any window where a client isn't actively
mid-frame.

**Disposition:** kept in the tree, env-gated (`GE_DETERM` unset by default,
zero behavior change confirmed via `verify.sh bunker1` before/after across
both attempts — worst cell 12.1–12.5 throughout, within existing noise).
The single-consumer synchronous-generation mechanism in `osRecvMesg()` is
the right shape and should stay — it's strictly better than polling and
fixed a real deadlock class cleanly. What's still missing: a genuine
**rate limit** on `__scMain`'s synthetic-retrace generation, decoupled from
real time — e.g. only synthesize the next one once the client has actually
finished processing the previous forward (a real request/response coupling
to `mainThread`'s progress, not just "the queue looked empty"), with a
separate, explicitly-bounded allowance for pre-first-frame boot code that
legitimately needs to service its own `osSetTimer` waits without the
scheduler racing ahead. That is a genuine design task, not a quick
follow-up — leaving D117 OPEN. `porting-notes.md` not updated (this
addendum's lesson is already exactly what §D117 above says, restated at
one more layer: any unthrottled or wall-clock-throttled generator racing
against a genuinely real-time-bound consumer reintroduces the same
divergence class, no matter how many layers deep you push it).

**D117 — M-52 4th pass: instrumented tracing (`GE_DETERM_TRACE=1`) precisely
localizes what's left — and it is NOT the virtual clock. Confidence: HIGH,
directly measured, not inferred.**

Rather than keep auditing code paths by inspection (two dispatches -- one to
the local delegate, one done directly -- had just spent a round ruling out
the level-load path and the audio-thread shared-globals angle with nothing
to show for it), added a temporary trace: every `g_determTicks`
advance/cap event and the `first_task_run` transition get a monotonic
sequence number + timestamp-free log line (`port/src/libultra.c`, gated by
`GE_DETERM_TRACE=1`, zero cost when unset). Ran `-level_09` twice, diffed
the two trace logs directly instead of guessing.

**Result: the tick mechanism is now proven fully deterministic.** Both runs
are byte-identical (same `ticks=` value at every logged step) all the way
through the `first_task_run` transition — the only difference is *which
spin iteration* it happens on (run 1: spin 4020, run 2: spin 4086 — 66 more
capped no-op spins in run 2, real-OS-scheduling-dependent as expected), but
the **virtual clock value at that moment is identical in both** (1551750 =
`2 × g_determQuantum`). Re-aligning the two logs by that spin-count offset
shows the entire subsequent tick progression is byte-for-byte identical.
The M-52 3rd-attempt fix (cap pre-task ticks to a fixed constant) works
exactly as designed — this is not where the remaining divergence comes
from.

**The actual mechanism, found by reading `sched.c` with the trace numbers
in hand:** every synthesized retrace — capped ones included — is a real
`VIDEO_MSG` that `__scMain` processes via `__scHandleRetrace()`
(`src/sched.c`), which unconditionally does `sc->frameCount++`
(`sched.c:319`) and then, per registered client, gates forwarding on
`(*((s32*)client + 2) == 0) || ((sc->frameCount & 1) == 0)`
(`sched.c:334`) — i.e. some clients only receive every *other* frame's
retrace by `frameCount` parity. The audio client is registered as exactly
this kind of half-rate client: `osScAddClient(&os_scheduler,
&g_AudioClient[0], &g_AudioManager.frameMessageQueue, 1)` (`src/audi.c:437`,
the trailing `1` is the half-rate flag). Since `frameCount` increments on
*every* capped spin — not gated by the virtual clock at all — and the total
spin count before the first real task varies run-to-run by a genuinely
real-time-scheduling-dependent amount, **`frameCount`'s parity at the
moment real gameplay begins is not guaranteed to match between runs**, even
though the clock value is now provably identical. This run pair happened to
land on a same-parity gap (66, even) — a different pair easily could not.
If it lands on opposite parities, the audio client's very first retrace
forward shifts by one frame's phase between runs, and audio-task scheduling
(which shares the same `osSpTaskStartGo`/tick-gating machinery) diverges
from there.

**Why no 5th live-coded attempt this session:** the natural next lever —
don't call `__scHandleRetrace` (and thus don't increment `frameCount`) for
a purely-capped "keep `__scMain`/`joyPoll` alive" spin — reopens the exact
chicken-and-egg the 1st attempt hit: `joyCheckStatusThreadSafe`'s handshake
and `__scMain`'s own unblocking need *something* processed during that
window, and `sc->frameCount` is decomp-internal state with no port-visible
accessor to correct after the fact. A real fix needs `frameCount`'s parity
(or the half-rate audio client's phase) to be made deterministic
independent of the real-time-dependent capped-spin count — a genuine
4th-layer design task, not a variant of any of the first three attempts.
Stopping here deliberately rather than rushing a live guess under momentum,
per the same discipline the earlier attempts were written up under.

**Disposition:** `GE_DETERM_TRACE` tracing kept in the tree alongside
`GE_DETERM`, same env-gating, zero cost when unset — it's the tool that
should be reused for the next attempt rather than re-instrumented from
scratch. D117 stays OPEN. porting-notes.md not updated yet (the
generalisable lesson here — "a per-frame parity/half-rate dispatch gate
fed by an unconditionally-incrementing counter inherits whatever
nondeterminism feeds that counter, even after the counter's *primary*
consumer is made deterministic" — is specific enough to this scheduler
that it's better captured here than genericized prematurely).

**D118 (session M-9) — SDL input layer implemented (Phase 3).**
`port/src/input.c` was an unimplemented stub; input only worked via
`libultra.c`'s `contSnapshotFromKeyboard()` (Z/Space=A, X=B, LCtrl=Z,
arrows=stick, WASD=D-pad — no C-buttons, no mouse, no gamepad).

Implemented a focused `input.c` (NOT a full port of pd_port's 1551-line
module — GE's menu/config code never calls that VK/bind-string API).
Provides `inputInit/inputUpdate/inputDestroy/inputGetNumControllers`
plus two helpers for `libultra.c`: `inputConnectedMask()` and
`inputComputePad(idx, *sx, *sy) -> u16 button`. The N64 button bits are
duplicated as `GE_CONT_*` in `input.c` rather than `#include <PR/os.h>`
(its `u8 errno;` field collides with `<errno.h>`'s macro; libultra.c
only gets away with it via a `#pragma push_macro` dance).

*Binding scheme.* Kbd/mouse (controller 0): WASD or arrows = analog
stick (move/strafe); mouse motion = C-buttons (aim, see bridge below);
LMB/LCtrl = Z (fire); RMB/LShift = R (aim mode); Space/Z/E = A;
X/R/F = B; Q = L; Enter/Tab = Start. Gamepad (SDL_GameController; pad 0
merges into controller 0, pads 1-3 → controllers 1-3): left stick =
stick, right stick = C-buttons (digital, 50 % threshold), RT = Z,
LT = R, A/X = A, B/Y/RB = B, LB = L, D-pad = D-pad, Start = Start.

*Mouse-look → C-button bridge (the subtle part).* GE aims with digital
C-buttons and has no analog-aim hook reachable without editing `src/`.
`inputUpdate()` integrates the relative-mouse delta (× `MouseAimSpeed`/
100) into a per-axis signed accumulator clamped to ±8. Each controller
poll `inputComputePad()` emits the matching C-button while |accum| ≥ 0.5
and drains one unit — so a flick holds the C-button for several frames
(proportional dwell) instead of a single blip. Limitations: still
digital (GE's own accel curve makes turn speed non-linear in mouse
speed); fast flicks saturate at ~8 frames of turn; diagonals limited to
the 8 C-button combos. A real analog path would need an `#ifdef PORT`
hook in `bondview.c` — left as TODO.

*Wiring.* `contSnapshotFromKeyboard()` in `libultra.c` (SI section) now
just calls `inputUpdate()` then marshals `inputComputePad()` output into
`g_contPad[]`/`g_contStatus[]` for all `MAXCONTROLLERS`, and sets
`g_contConnected` from `inputConnectedMask()`. It is still driven by
`osContStartReadData`/`osContStartQuery` (once per game logic tick) — no
new frame hook in `video.c` was needed. `osContGetReadData()` now
`memcpy`s the whole `g_contPad` array (N64 semantics: one pad per
channel) instead of only controller 0 — joy.c passes a
`MAXCONTROLLERS`-long array. `inputInit()` opens the gamepad subsystem +
all connected controllers and enables relative mouse mode.

*Config.* `ge007.ini` `[Input]` `MouseEnabled` (0/1), `MouseAimSpeed`
(1..500, default 50), `MouseInvertY` (0/1), via `configRegisterInt` in a
`PD_CONSTRUCTOR`. `config.c` has no float support so speed is an int
percent. Key/button REBINDING is not wired (would need the pd_port
bind-string system or a new mini-parser) — hardcoded scheme + TODO.

*Status.* Build GREEN (`ntsc-final`). Boots `-level_09` crash-free for
35 s, 900+ frames, `input: ready (mask=0x1, 1 controller(s))` logged.
`GE_INPUTLOG=1` gates a per-poll `sysLogPrintf` of the OSContPad when
button/stick are nonzero (zero-cost when unset, kept). Live input
UNTESTED (headless agent) — needs a human pass: see
`docs/HANDOFF.md`. Confidence: build/wiring HIGH; in-game feel of the
mouse→C-button bridge MEDIUM (aim speed tuning likely needed); gamepad
UNVERIFIED (none present in build env).
TODO: (1) key rebinding; (2) gamepad hotplug (opened only at init today);
(3) optional analog-aim `#ifdef PORT` hook in bondview.c — only needed now
for D118a (fully-analog hipfire pitch) and toggle-aim-scheme correctness.
[M-24: mouse-look rework + real INI parser done — see "Mouse-look rework"
below.]

*Mouse-look rework — M-24 (`port/src/input.c`, port-layer only, no `src/`
change).* Root reading of `bondviewProcessInput`/`MoveData`: GE's aim is
**mode-dependent** — hipfire (`!insightaimmode`) yaw = analog stick-X,
pitch = digital C-up/C-down (stick-Y = move, does not pitch); aim mode
(R held) yaw+pitch = analog stick past ±60 → `(stick-60)/10`, and
C-up/C-down there = crouch/lean/zoom (`bondview2.c:5340,5351`), *not* aim.
Also GE's native pitch is inverted (`U_CBUTTONS → speedVertaDown`, i.e.
C-up looks down, `:5272-5279`). New mapping in `inputComputePad`, keyed on
our own RMB/LShift state as the hold-to-aim proxy:
- **aim mode** → push `stick_x`/`stick_y` into the 61..80 band
  (proportional to per-poll mouse delta × `MouseAimSpeed`/100 × 4), **emit
  no C-buttons**. Yaw and pitch now identical analog feel.
- **hipfire** → yaw `stick_x += dx × MouseTurnSpeed/100 × 6`; pitch =
  digital C-up/C-down on `|dy| ≥ 1.5` px/poll.
- "mouse-down = look down" by default (accounts for GE's inversion);
  `MouseInvertY` flips.
Per-poll deltas, no accumulator (rate device). `ge007.ini [Input]`:
`MouseEnabled`, `MouseAimSpeed` (50), `MouseTurnSpeed` (100), `MouseInvertY`
(0) — **now actually parsed** (`config.c` INI load/save implemented, M-24).

- **D118b (FIXED, M-24)** — mouse-Y inversion resolved by the "mouse-down
  looks down" default above + `MouseInvertY` toggle.
- **D118c (FIXED, M-24)** — aim + mouse-down → crouch. Fixed *without* a
  `src/` hook: in aim mode the mouse now drives the analog stick and emits
  no C-button, so the crouch/lean/zoom mappings are never triggered by
  look input. (Proxy caveat: a toggle-aim control scheme would need a read
  of `g_CurrentPlayer->insightaimmode` since RMB-held ≠ aim state there.)
- **D118a (RESIDUAL, lower)** — in **hipfire only**, yaw (analog) vs pitch
  (digital C-button) still feel different. Aim mode is now fully analog and
  consistent. A fully-analog hipfire pitch would need the `#ifdef PORT`
  `bondview.c` hook (TODO 4). Minor — precise vertical aim in GE happens in
  aim mode, not hipfire.
- **D137 (FIXED)** — right-mouse *crash* (not the crouch bug):
  `gunDrawSight` `s32 sp54` truncated a `Gfx*`; see §F D137.
- Weapon switch on kbd/mouse = the A button (`Space`/`Z`/`E`), same as
  action/use (GE overloads it). No dedicated key; mouse-wheel cycle is a
  candidate quick-win.

**`data/` deletion + recovery (session M-3).** `git worktree remove
--force` on an agent worktree that had a directory *junction*
`worktree/data → main/data` followed the junction and deleted the real
`data/` contents (both `.z64` baseroms + `pccg-*`/`pcmodels-*` sidecars).
`data/` is gitignored so nothing tracked was lost. Recovered:
`cp baserom.u.z64 data/ge007.ntsc-final.z64` (sha1
`abe01e4a…` == the canonical `ge007.u.z64` build hash, so byte-identical
to what was there), then regenerated all three sidecars — `python
tools_pc/d43_emit.py ntsc-final` (pcmodels, 512), `d69_emit.py
ntsc-final` (pccg bg/stan, 52), `d88_emit.py ntsc-final --regen` (Usetup,
→ 73 rows, `pccg.bin` 3605249 B) — all "ALL CHECKS PASSED". **Lesson:
never junction `data/` into a throwaway worktree; copy it or point the
generator's `ROM_PATH` at the repo-root baserom.** The extra
`GoldenEye 007 (U) [!].z64` copy was not restored (unused — runtime and
generators use `ge007.ntsc-final.z64`, falling back to root
`baserom.u.z64`).

**Docs-to-commit reminder (session L).** The D88.1–D88.3 work
(`tools_pc/d88_emit.py`, `port/src/pccg.c`, `src/bondtypes.h`,
`src/game/bondview2.c`, `src/game/bondview_r.c`) plus the `GE_D88` probes
in `prop.c`/`stan.c` have been format-verified and pass `-level_09` up to
the D88.4 crash, but **remain uncommitted** (carried through two
interrupted sessions). Commit in sub-milestones per the usual pattern
once D88.4 is understood: format spec → converter → port wiring → probes.

**D69 status after D78-D88: the ORIGINAL blocker (`load_bg_file` faulting
on first stage load) is fully resolved and verified** — a clean run loads
BUNKER1's header/room/portal/envdata tables and its full ~1066-tile stan
file correctly (spot-checked byte-for-byte against the N64 source via the
`GE_D69STAN` probe: tile room/mid/tail/point values match). The game now
progresses substantially further than before (through room-streaming
setup, past the old D86 model-init crash and the D87 attract-mode crash)
before hitting D88's separate, newly-exposed "stage setup" file format
gap. D85 (room geometry renders wrong, not yet crash-free at the *visual*
level) and D88 (stage-setup file format, unconverted) remain open
follow-ups — **the "loads without
fault" acceptance bar is not yet fully met** (the process still exits via
crash, just much later in the load sequence), but the converter, port
wiring, and every ABI fix identified so far are format-verified correct
and committed.

**Environment reminders.** MSYS2 tools in `/c/msys64/mingw64/bin/` (not on PATH —
prefix `export PATH=…`). Build: `./build-pc.sh ntsc-final`. gdb **launch** mode
is far too slow for timing-dependent crashes (a D56-class crash ~10 s in took
>300 s under gdb to reach 2 frames — DBGHELP symbol loading + the D51
stall-heartbeat thread dumps; don't wait on it): prefer env-gated TEMP probes +
the built-in crash log for reproducible faults. **Correction (D87 session):**
gdb **attach** mode (`gdb -batch -x cmds -p <winpid>`, where `<winpid>` is the
Windows PID from `ps`, 4th column — the game must already be running, e.g.
launched with `nohup ... &`) works fine and is fast, since the process is
already warmed up and running at full speed before you attach; a hardware
watchpoint (`watch *(int*)0xADDR`) caught a global's write in well under a
minute. Useful for "is this global legitimately written, or corrupted"
questions on a long-running, non-crashing process — attach once the process
has been running a while, `continue`, and it'll fire on the very next real
write. Still avoid gdb for the crash itself if the crash is reproducible via
the crash log; symbolicate
offline with `addr2line -e build-pc/ge007.x86_64.exe -f -C <0x140000000+rel>`.
Image base 0x140000000. `load_resource`/many init fns use a fake RBP — compute
stack offsets from entry RSP. The D30 crash handler writes `ge007.crash.log`
with a working Phase-2 backtrace (**D44** fixed) — first stop for any fault;
frames past the true chain may be stale (a corrupted return address outside the
module, as in D56, means unwind depth is limited — confirm callers by code-
path analysis + behavior). Standalone probe compiles need `-std=c11` (without
it `typedef s32 bool` in bondtypes.h breaks under gnu23) and pointer-difference
arithmetic instead of offsetof (include/stddef.h is #if 0'd).

---

**D121 (session M-12, WS1) — frictionless per-level boot.** A bare
`./build-pc/ge007.x86_64.exe -level_XX` with no `-m*` args booted every
level with the default memory pools and OOM-crashed early in the load
sequence; you had to hand-copy the per-level `-ml -me -mgfx -mvtx -mt -ma`
row from `memallocstringtable[]` (`boss.c:101`) onto the command line.

*Root cause.* On N64 the per-stage auto-inject loop in `bossMainloop`
(`boss.c` ~line 425: `tokenSetString(memallocstringtable[i].string)`)
supplies that row, but it is gated on `g_DebugAndUpdateStageFlag`. With
`-level_` present that flag is left to `rmonGetToken()` (`boss.c:178`),
which the PC stub (`port/src/n64stubs.c`) returns 0 for, so the loop never
runs.

*Rejected fix.* Setting `g_DebugAndUpdateStageFlag = 1` when `-level_` is
present (the "mirror `boss.c:199-202`" idea in the old HANDOFF) is **not
behaviour-preserving on PC**: it makes boot render the full title-stage
intro (logos → gun-barrel → cast) instead of loading the level directly.
The flag does more than pick the token row.

*Fix (`#ifdef PORT` in `bossInitMainthreadData`, after the flag's own
`tokenSetString` block).* When `-level_` is present and no `-m` token was
given, look up the stage's `memallocstringtable[]` row (same match the N64
loop does) and `tokenSetString` it — **prefixed with `-level_XX` (and
`-hard N` if present)**, because `tokenSetString` (`src/token.c:41`)
`strcpy`s over the whole token buffer; without the prefix `bossMainloop`'s
`tokenFind(1, "-level_")` returns NULL and it boots the title stage. No
flag change, no control-flow change.

*Verified.* Bare `-level_09` now loads BUNKER1 directly (85–92 % non-clear
frame content from frame 30, no intro), matching the old
`-level_09 -ml0 -me0 -mgfx100 -mvtx50 -mt700 -ma150` repro. Residual
BUNKER1 prop/`modelLoad` crash is intermittent (D117 nondeterminism +
D88.4) and pre-existing — out of WS1 scope. HANDOFF repro line updated to
drop the manual `-m*` list.

**D122 (session M-12) — per-level prop/item model-load crash: propDef
converter missing handlers for ObjectRecord-derived record types.**
`-level_33` (Dam), `-level_34` (Facility), `-level_35` (Runway) crashed
deterministically before frame 1: `modelLoad` (`loadobjectmodel.c:393`,
`PitemZ_entries[modelid].header->RootNode` with an OOB `modelid` — `Rax`
held ASCII) and `modelInitRwData` (`model.c:6249`). BUNKER1 (`-level_09`)
and Silo (`-level_20`) were fine.

*Root cause.* The D88.4 propDef stream converter (`tools_pc/d88_propdefs.py
convert_record`) had per-type handlers for most record layouts but **four
`inherits ObjectRecord` types fell through to the generic arm**: 47
TINTED_GLASS, 39 VEHICHLE, 40 AIRCRAFT, 45 TANK (plus 13 AUTOGUN and 20
AMMO/MultiAmmoCrate, exposed once the first four were fixed). The generic
arm bswap32's N64 word 1 as a single 32-bit value, but for an
ObjectRecord that word is `[s16 obj][s16 pad]` — it needs `_hh_word`
(swap each half in place). Result: `obj` (the model id, read by
`domakedefaultobj` at `prop.c:155` → `modelLoad`) landed in the wrong
half → garbage/OOB index into `PitemZ_entries[]`. The generic arm also
never widened the ObjectRecord's embedded pointer slots (prop*/model* 4→8
B), so every tail field was mis-offset too. BUNKER1/Silo use none of
these six types (verified with a per-level `PROPDEF_N64_WORDS` type
histogram), which is why they were unaffected — the stream *walk* was
self-consistent for all 21 levels (`convert_stream` validates region
length), only the *field contents* of these records were wrong.

*Fix (converter only — `tools_pc/d88_propdefs.py`).* Added `OBJ_TAIL_DESC`
= {type: (ptr_word_set, hh_word_set)} for types 47/39/40/45/13/20 and a
handler that emits the real 144-byte PC ObjectRecord prefix
(`_emit_object_prefix`, which does `obj` correctly) followed by the tail
with pointer members widened 4→8 B and 8-aligned, `_hh_word` for u16-pair
words (vehicle aioffset/aireturnlist; ammo-crate `slots[]`), `_bswap32`
elsewhere. Updated `PROPDEF_PC_BYTES` (39: 176→208, 40: 180→208, 45:
224→248; 47/13/20 already correct) and the matching
`sizepropdef()` `#ifdef PORT` arms in `loadobjectmodel.c` (VEHICHLE
44→52, AIRCRAFT 45→52, TANK 56→62 words; N64 values kept in the trailing
comment / `#else` `#if 1` switch). No game-logic change. Regenerate:
`python tools_pc/d88_emit.py ntsc-final --regen` (21/21, ALL CHECKS
PASSED).

*Verified.* `-level_33/34/35` all get past the model-load chain now (no
`loadobjectmodel.c:393` / `model.c:6249` crash). BUNKER1 unregressed
(83–92 % non-clear frame content, frames 80–200). Silo unregressed
(exercises the new type-20 handler — no crash).

*Residual (separate blockers, out of D122 scope — hand to the level
sweep).* Dam + Runway now crash later in `chrIsNotDeadOrShot`
(`chraction.c:4483`, `self` = an image rodata address — guard/chr setup,
not propDef-related). Facility crashes in `import_texture_i8`
(`gfx_pc.cpp:821`, bad texture pointer — fast3d). Neither is touched by
this fix.

*Tail-layout confidence: medium* for 39/40/45 (VehichleRecord/
AircraftRecord/TankRecord field offsets have "locs need confirming" notes
in `bondtypes.h`; the pointer-word sets and total sizes are derived from
the struct as written and may need a nudge if a level actually drives
tank/vehicle motion). **High** for 47/13/20 (offsets are known) and for
the crash fix itself (the `obj`-field half-swap is unambiguous).

---

**D123 (session M-13) — crash class C1: `chrIsNotDeadOrShot` NULL deref on
6 levels (Dam 33, Runway 35, Frigate 26, Statue 22, Streets 29, Cradle
41). Converter zeroed the vehicle/aircraft `ailist` id. FIXED
(`tools_pc/d88_propdefs.py` only).**

*Symptom.* Bare `-level_33` etc. crashed before frame 1 in
`chrIsNotDeadOrShot` (`chraction.c:4483`, `self->actiontype`), fault addr
0x8 → `self ≈ NULL`. (The log's frame #1 `0x1401296a0` / `chraidata.c:61`
was a stale stack-scan hit, not a real frame — misled the original
triage.)

*Real backtrace (gdb):* `objTick` (`propobj.c:5504`) →
`ai((PropDefHeaderRecord *)poTruck, PROP_TYPE_OBJ)` →
`ai()` sees `Entityp->type == 39` (PROPDEF_VEHICHLE) so sets
`VehichleEntityp`, `AiListp = VehichleEntityp->ailist`, `ChrEntityp =
NULL`. `AiListp` resolved to `m_AimAtBond` (`GAILIST_AIM_AT_BOND` == id 0).
Its first opcode is `AI_TRYFireOrAimAtTarget` →
`actor_aim_at_actor(ChrEntityp=NULL,…)` → `chrIsNotDeadOrShot(NULL)`.

*Root cause.* `prop.c:1764` does `pdef_veh->ailist =
ailistFindById(pdef_veh->ailist)` — it reads the **pre-populated integer
AI-list id** out of the ROM record (VehichleRecord.ailist @ N64 0x80;
Dam's is `0x040a` = 1034, a per-level list). But D122's `OBJ_TAIL_DESC`
put word 32 in `ptr_word_set`, whose handler emits **8 zero bytes** for
the widened 4→8B slot (correct for `path*`/`Sound*`, which `prop.c`
overwrites with 0 — but wrong for `ailist`, read before overwrite).
Result: `ailistFindById(0)` → `isGlobalAIListID(0)` true → global list id 0
= `GAILIST_AIM_AT_BOND` → a vehicle runs a CHR-only aim list with a NULL
chr. `AircraftRecord.ailist` (`prop.c:1786`) is identical. BUNKER1/Silo
emit no vehicle/aircraft propDef → unaffected. All 6 crashing levels have
a vehicle or aircraft prop.

*Fix (converter only).* New `OBJ_ID_WORDS = {39:{32}, 40:{32}}` in
`tools_pc/d88_propdefs.py`; the `OBJ_TAIL_DESC` walk now emits
`_bswap32(w)` into the low 4 bytes of the 8-aligned slot for id words
(high 4 stay zero), so the u16/s32 id survives into the widened pointer
field. Words 32 removed from the 39/40 `ptr_word_set`. Struct size
unchanged → no `sizepropdef()` edit. TANK (45, `collision*`) and AUTOGUN
(13, `unkC4/unkC8/beam`) left as ptr/zero — verified genuinely
runtime-populated (`prop.c:1712` / `setupAutogun` `prop.c:694`). Regen:
`python tools_pc/d88_emit.py ntsc-final --regen` (21/21 ALL CHECKS
PASSED).

*Verified.* `chrIsNotDeadOrShot` crash gone on all 6. Dam 83%, Statue
80%, Frigate 68–90%, Cradle 77% non-clear frames, no crash log.
BUNKER1 (91.7%) + Silo (91.7%) unregressed. Runway and Streets now crash
**later** in fast3d (`palette_to_rgba32`/`import_texture_i4`
`gfx_pc.cpp:851`; `gfx_sp_matrix` `gfx_pc.cpp:1046`) — class C2/other
track, not C1. Frigate/Statue/Cradle/Dam reach a frame cleanly.

*Confidence: high* — root cause reproduced in gdb, fix is a 1-field
converter change matching the exact `prop.c` read pattern, and the id
value (0x040a) matches the N64 `UsetupdamZ.c` Vehicle record tail.

**D124 (session M-13) — crash class C2: fast3d handed a garbage texture
pointer on Facility (`-level_34`) and Jungle (`-level_37`).** Symptom:
early deterministic crash — Facility `import_texture_i8` `gfx_pc.cpp:821`
fault `0x72181ee8`; Jungle `gfx_tex_normalize_source` `gfx_pc.cpp:644`
fault `0xabcd0824` (= `IMAGESEG(0x824)`). Two *different* root causes:

*Jungle — FIXED.* `gimgSyncCompiledGlobalDLs()` (`port/src/gimgfixup.c`,
D68) is meant to copy the texture pointers that `texLoadFromDisplayList()`
patched into the ROM copy of the Globalimagetable back into the compiled
`globalDL_0xNNN` shadow arrays (`assets/oddtextures.c`) that `explosion.c`
executes via `g_ExplosionDisplayLists[]`. Its slot-detect test was
`p[6]==0xCD && p[7]==0xAB` — the *post-fixup marker* — but by the time it
runs `texLoad()` has **already replaced every IMAGESEG w1 in the ROM copy
with a real pointer**, so that test never matches and the sync copied
*nothing*. The compiled arrays kept their link-time `IMAGESEG(id) =
0xABCD0000|id` words; the first explosion/smoke/particle DL that
Facility/Jungle render feeds `0xABCDxxxx` straight into fast3d
`seg_addr()` (falls through to `return (void*)w1`) → `import_texture`
deref. The 7 "working" levels are only working because nothing triggered
an explosion-DL draw inside the ~22 s capture — **the bug was latent in
every level.** *Fix:* detect the slot from the **compiled** array instead
(`dst[j]` is a `G_SETTIMG` whose w1 is still `0xABCDxxxx`), then copy the
ROM copy's resolved `*(u32*)(p+4)`. `port/src/gimgfixup.c` only; no
converter/sidecar change. Verified: Jungle now renders ~300 frames
(91.7% non-clear) before a *separate* downstream crash (`gfx_sp_matrix`
`gfx_pc.cpp:1046`, bad G_MTX seg addr in the same explosion-DL stream —
those DLs also carry unconverted `G_MTX` words; follow-up, D75/matrix
family). BUNKER1 (83%) + Silo (91.7%) unregressed.

*Facility — NOT fixed (separate class).* Fault `0x72181ee8` is a bogus
`G_SETTIMG` w1 (`fmt=4 siz=2 w=0`, prev cmd an unconsumed `0xba` GE
tex-macro opcode) in a DL living in V1 DRAM — a **model/prop GDL from the
runtime `texLoadFromGdl()` relocation path** in `sub_GAME_7F0762E0`
(`objecthandler_2.c:82`), not a room GDL and not the global bank.
`GE_C2` probe of `texLoadFromGdl` shows its model-GDL calls (second batch,
`src≈0x706a****`) write to **non-16-aligned `dst`** (e.g.
`0x701eac01`, `0x701eb7d5`) — the `replacementgdl`/`name`(=srcsize)
offset arithmetic in `objecthandler_2.c` still mixes N64 8-byte and PC
16-byte `Gfx` strides, so converted commands land mid-slot and a
`G_SETTIMG` w1 reads as garbage. This is the open D80/D82/D83
"model/room GDL runtime conversion not verified" area — shared infra
touched by every level; out of C2 budget, left for a dedicated pass.
BUNKER1/Silo/Jungle don't hit it because their visible prop models
happen to convert cleanly; Facility's do not.

*Probes:* none left in tree — `GE_C2` scratch prints removed after
root-cause. *Files touched:* `port/src/gimgfixup.c` (fix),
`docs/porting-notes.md`, `docs/internals.md`, `docs/dev/LEVEL-STATUS.md`.
*Confidence:* Jungle fix **high** (mechanism proven with a `[C2sync]`
trace showing 32 slots going `abcd08xx -> 700e****`, verified render).
Facility diagnosis **medium** (probe evidence strong; exact off-by in
`objecthandler_2.c` not yet pinned).

**D125 (session M-14) — crash classes C3 + C6: the converted `propDefs`
blob in RAM does not match `tools_pc/d88_propdefs.py`'s offline output;
the `sizepropdef()` walk drifts, so every `pdefIndex`-keyed lookup
(`setupDoor` → `linkedDoor`, `weaponAssignToHome`, `modelLoad(modelid)`)
gets the wrong record.** NOT root-caused; investigation cut for a usage
reset. Affects Aztec (`-level_28`, `propobj.c:13523/:13601`), Bunker2
(`-level_27`, same `door7F054FB4`), Surface2 (`-level_43`,
`loadobjectmodel.c:393` `PitemZ_entries[modelid].header`).

*Evidence.* Instrumented the `proplvreset2` propDef walk
(`prop.c:1865`, `#if defined(PORT)` `getenv("GE_C3")` print of
`pdefIndex / phead->type / sizepropdef(phead) / phead / g_CurrentSetup.propDefs`)
and the `setupDoor` `linkedDoor` resolve (`prop.c:1206`). On `-level_27`
(`UsetupsevbZ`, `setup_text_pointers[27]`):
  - `g_CurrentSetup.propDefs = 0x701ce814`.
  - Runtime walk: idx0 `type=35` (WATCH_MENU, correct), then idx1–4
    `type=0`, idx6 `type=112`, idx15 `type=26`, … — i.e. after the first
    record the blob is **zeros / garbage**, but `sizepropdef` still
    strides the phantom `type=35` at idx0/5/10 forward 16 B each.
  - By the first real DOOR the walk has counted **147** records where the
    offline converter (`convert_stream(UsetupsevbZ)`) puts it at **46**
    (verified: offline output = 346 records, DOOR at idx 46, self-
    consistent with `PROPDEF_PC_BYTES` == `sizepropdef()` PORT strides —
    every type cross-checked against `d88_layoutprobe` sizes, all match).
  - `setupDoor` then gets `arg2=147` for a stream-index-46 door;
    `linkedDoorOffset(-1) + 147 = 146` → `setupGetPtrToCommandByIndex(146)`
    returns a `type=0` record → `door->linkedDoor` garbage → fault
    walking the linked-door list in `door7F054FB4`. Surface2's C6 is the
    same mechanism one step earlier: a drifted `pdefIndex`/`modelid`.

*What is NOT the bug (ruled out).* (a) `d88_propdefs.py` `convert_stream`
itself — self-test passes, offline walk of all 3 crashing levels'
converted output is clean, DOOR/linkedDoorOffset values correct, stream
tiles `[propDefs, intro)` exactly (`end=OK` for all 21). (b)
`sizepropdef()` PORT strides — every `case` equals
`PROPDEF_PC_BYTES[type]/4` and equals the compiler `sizeof` of the widened
struct (checked via `d88_layoutprobe`). (c) DOOR converter layout —
`linkedDoorOffset@144`, `linkedDoor@216`, `unkcc@224`, size 296 all match
`d88_layoutprobe`; tail pointer-widening walk is correct end-to-end. (d)
The per-type histogram diff (brief's suggested approach): the 3 crashing
levels emit no propDef type absent from the 12 passing levels — TINTED_GLASS
and AUTOGUN appear in passing Caverns/Control too. **This is not a
forgotten-type converter gap like D122/D123.**

*Best hypothesis (M-14) — DISPROVEN (session M-15).* M-14 guessed the
converted `propdefs_pc` bytes were not landing in the sidecar at the
relocated `propDefs` offset (a `d88_emit.py` region-tiling / delta-reloc
bug, lines ~308-338/471-476). The M-13 overseer note refined this to a
pass-1 delta / region-`end` mismatch (`pd_end != pd_start + _n64len`).
**Both are wrong.** New diagnostic `tools_pc/d125_check.py` decompresses
the *emitted* `data/pccg-ntsc-final/pccg.bin` entry for every Usetup*Z,
reads the relocated `propDefs` header field, and byte-compares that slice
to a fresh `convert_stream()`:

  - **All 21 levels: `MATCH`** (Aztec/Bunker2/Surface2 included). The
    sidecar's propDefs blob, post-RZ-roundtrip, is byte-identical to the
    converter output, placed exactly where the relocated header field
    points.
  - Instrumented `d88_emit.py` (D125_DEBUG, since reverted): for all 21,
    `tiled_pd_end == H[intro]` exactly, `_n64len == pd_end - pd_start`
    exactly, the region after `propdefs` in the sorted list is always
    `intro`, and the growth `pclen - n64len` fed to `cum` is correct.
    There is **no** region-end/delta bug at `d88_emit.py:308-343`.
  - Full re-cross-check of `sizepropdef()` PORT strides vs
    `PROPDEF_PC_BYTES/4`: all 28 types match (as M-14 already found).

**Therefore the offline pipeline (converter → emit → RZ compress) is
correct end-to-end. The "idx0 correct, then zeros" RAM signature M-14
observed must originate at or after runtime load** — candidates, in
rough order: (a) `decompressdata()` truncation — `port/src/rzdecomp.c`
inflates with `Z_FINISH` and a 4 MiB `avail_out`, loops while `Z_OK`;
if the deflate stream for these larger converted files hits `Z_BUF_ERROR`
or `Z_STREAM_END` mid-output the `while (ret == Z_OK)` exits early and
the tail of `dst` stays zero (needs a probe: log `ret` + `produced` vs
expected decompressed size for `UsetupsevbZ`); (b) the STAGE bank alloc /
`mempAllocBytesInBank` giving a buffer that is fine for the N64-size file
but the converted file is ~15-25% larger — check `mempGetBankSizeLeft`
at the `_fileNameLoadToBank(strResource, …, 256, MEMPOOL_STAGE)` call
(`prop.c:1271`) vs the decompressed size; (c) the `prop.c:1280-1308`
rebase loop or `langLoadToAddr` (`prop.c:1274`, runs right after the
load, same bank) overwriting the tail of the setup file. M-14's own
runtime dump had `g_CurrentSetup.propDefs = 0x701ce814` (`&15 == 4`).

*Next step.* Probe `decompressdata()` for `UsetupsevbZ`: print `ret`,
`produced`, and the first 64 bytes at `dst + 16` (offline record 1
start). If `produced` < the offline decompressed file size → truncation
(fix the inflate loop / `avail_in`). If `produced` is correct but the
bytes are still zero → something overwrites it post-load (bisect
`langLoadToAddr` / the rebase loop). `tools_pc/d125_check.py` stays in
tree as the offline-side regression guard.

*Probes:* **none left in tree** — the two `#if defined(PORT)`/`GE_C3`
scratch prints in `src/game/prop.c` (lines ~1206 and ~1865) and the
temp scan scripts `tools_pc/c3_scan.py` / `tools_pc/c3_doors.py` were all
reverted/deleted. *Files touched:* only `docs/internals.md` (this
entry) + `docs/dev/LEVEL-STATUS.md` (one line). No code, no converter, no
sidecar change — **sidecars do NOT need regen**, tree builds clean at
`f2beae4b` + M-13 uncommitted set.

**D125 addendum (session M-16) — ROOT CAUSE FOUND: `d88_emit.py:374`
4-byte literal into an 8-byte slice; per-record buffer shrink; out-of-range
leaf writes insert at the current end → boundpad name blob drift.**

*Crash mechanism (Aztec, `-level_28`), fully traced.* Door idx 226
(`objid=307`, `pad=33`) is left with `model=NULL`/`prop=NULL` by
`setupDoor()`: its boundpad's plink name in the sidecar is `"8d2"`
(original ROM: `"p138d2"`) → `stanPackId()` rejects it (valid = `p`/`q`
+ decimal ≤32767 + letter + optional digit 0-7; probe line `D88
stanMatchTileName id=8d2 hi=ffff lo=ff` is the failure sentinel) →
`init_pathtable_something()` returns 0 with `*tile_stack=NULL` →
`getposstan()` nonzero → `setupDoor` sets `door->prop=NULL`, skips
`doorInit()` → later tick `door7F054FB4()` (`propobj.c:13601`) derefs
`door->model->obj`; `Model.obj` sits at PC offset 0x10 (widened `chr`
ptr) → fault addr 0x10, Rcx=0. Linked-door pair: idx225 (`pad=32`, valid)
has `linkedDoorOffset=1` → idx226. All 178 regular pads resolve fine; only
boundpad names are corrupt.

*The bug.* `tools_pc/d88_emit.py` `emit_pad()` line 374:
`out[dst_o + 0x30:dst_o + 0x38] = b"\x00\x00\x00\x00"` — the stan zero-fill
assigns a **4-byte** literal to an **8-byte** slice. Python silently allows
in-range short-RHS slice assignment, so each of Aztec's 179 pads + 148
boundpads shrinks the output `bytearray` by exactly 4 B (measured: after
header 0x14d1c → after pads 0x14a50 → after boundpads 0x14800). Later
verbatim leaf writes (the plink string blob, relocated to ~0x147EC+) then
target offsets at/past the shrunken tail; for those, Python clamps the
slice to an empty range **at the current end** and inserts the data there —
each string lands at the buffer tail instead of its relocated offset,
drifting earlier with every write. A simulation of exactly this semantics
reproduces the observed sidecar names byte-for-byte: pad30 `p139d2`
(intact), pad31 `139d2`, pad32 `37d1`, **pad33 `8d2`**, pad34 `2`,
pad35 `1`, pad36 `` (empty).

*Why it was missed.* No error is raised; the final file is only
`(n_pads+n_boundpads)*4 − re-extension` bytes short of `total_new`
(Aztec: 85276 vs 85145) because later leaf writes re-extend the buffer;
`d125_check.py` verified only the propDefs slice. M-14's "propDefs zeros in
RAM" signature was a misread — M-15 already proved sidecar propdefs
byte-correct, and M-16's post-load RAM dump matched the sidecar byte-for-
byte (85,145 B), so runtime load is clean too.

*Fix (pending).* Line 374 → `b"\x00\x00\x00\x00\x00\x00\x00\x00"` (8 NULs).
Audit of every other `out[` write in the file: all remaining slice/RHS
pairs are width-matched — this is the only one. **All 21 Usetup*Z sidecars
are affected to varying degrees** (shrink = 4×(pads+boundpads) per level;
which names drift depends on where the blob lands vs the shrunken tail) —
regen all: `python tools_pc/d88_emit.py ntsc-final --regen`, then re-run
`tools_pc/level_sweep.sh` (expect C3 Aztec+Bunker2 to clear; currently-
passing levels can only improve, but the sweep is the proof).

*Probes in tree (temporary — remove after fix verified):* `GE_D125`
env-gated prints in `port/src/rzdecomp.c` (decompress ret/produced),
`src/game/prop.c` (setup header + post-load 0x15000 dump to
`d125_setupdump.bin` + post-`setupDoor()` door state), `src/game/propobj.c`
(`door7F054FB4` NULL-model detect+skip, `objInit` failure print, plus temp
`#ifdef PORT` stdio/stdlib includes). Scratch: `tools_pc/d125_convtest.py`
(instrumented converter copy — source of the shrink evidence),
`tools_pc/d125_check.py` (**keep** — offline regression guard),
`tools_pc/d125_inflate_test.c`, `d125_offs_tmp.c`, `d125_t.c`,
`d125_run.log`, `d125_setupdump.bin`, `build-pc/d125conv/`.

**D125 addendum (session M-16b) — FIX LANDED.** `d88_emit.py:374` changed
to 8 NUL bytes (+ a warning comment). Offline proof (Aztec, throwaway
`d125_proof.py`): the fixed converter's boundpad plink names 30–36 read
`p139d2 p139d2 p837d1 p138d2 p85a2 p1377b1 p504d2` (= original ROM), 147
boundpad records recovered (M-14's "46" was the corrupt read — no second
bug). `d43/d69/d88 --regen` → 21/21; `d125_check.py` → 21/21 propDefs
still MATCH. Direct re-test: **Aztec `-level_28` PASSES** (91.0%, 900+
frames, was C3 CRASH at `propobj.c:13601`); Bunker1/Silo unregressed.
**Bunker2 `-level_27` still crashes** — moved to `door7F054FB4`
`propobj.c:13536` (`var_s1->openPosition`, `var_s1` = `0xffff…` from
`linkedDoor`). Separate bug: the DOOR-tail converter in `d88_propdefs.py`
(`DOOR_TAIL_PTR_WORDS` / `linkedDoorOffset` word 32) vs the compiled PC
`DoorRecord` layout — `linkedDoor` is resolved from `linkedDoorOffset`
(`prop.c:1204`), a read-before-write int id like D123's `ailist`; Bunker2
has linked double-doors, Aztec's are singletons so it dodged it. Folded
into the converter write-audit (C3 residual). Sweep: 12 → 13/21.
All M-16 probes reverted, scratch files deleted, `d125_check.py` kept.

**D126 (session M-17) — objective-subrecord `->next` pointer growth desyncs
the propdef walk (C3r Bunker2 + C4 Depot + C6 Surface2).** The Bunker2
`door7F054FB4` crash was NOT a DOOR-tail layout bug (the M-16b hypothesis).
Instrumenting the `proplvreset2` first-loop walk showed `pdefIndex` drifting
~100 ahead of the offline `convert_stream` record index; a byte dump proved
the RAM propDefs matched the sidecar exactly *at load* but had record N+1's
header zeroed *by the time the walk reached it*. Cause: four objective
sub-record types — `criteria_picture` (30), `criteria_roomentered` (32),
`criteria_deposit` (33), `setup_objective_text` (35) — are
`{ s32 x N; T *next; }`. `set_parent_cur_obj_photograph/enter_room/
deposited_in_room` and `setup_briefing_text_entry_parent` (`objective.c`)
write `arg0->next` unconditionally while the setup walk visits each record,
building a runtime linked list. On N64 `next` is a 4-byte word at the end of
a 16/20-byte record — no spill. On x86-64 the pointer widens to 8B and the
compiler 8-aligns it at offset 16, so the struct is 24B; the 8-byte
`->next` store lands at record+16..24, clobbering the *following* record's
`PropDefHeaderRecord` (type byte at +0x13). Walk desyncs → command indices
drift → `setupDoor`'s `setupGetPtrToCommandByIndex(linkedDoorOffset + arg2)`
returns the wrong record → `door->linkedDoor` chain walks into non-door
memory → crash. Depot (`prop.c:902` tile-walk room) and Surface2
(`PitemZ_entries[modelid]`, the C6 "D122 continuation") were the same
desync landing on different downstream derefs. Fix:
`tools_pc/d88_propdefs.py` `PROPDEF_PC_BYTES[30/32/33/35] = 24` + a typed
handler emitting the leading s32 words then an 8-byte zero `next`;
`src/game/loadobjectmodel.c` `sizepropdef` PORT returns 6 for those types
(N64 `#else` kept). Regen'd all 21 sidecars, `d125_check.py` 21/21 MATCH.
**Bunker2 `-level_27`, Depot `-level_30`, Surface2 `-level_43` now PASS**
(91.5% / 79.8% / 70.4%); Bunker1/Silo/Aztec/Archives/Egypt/Train
unregressed. Sweep 13 → 16/21. Generalisable quirk appended to
`porting-notes.md` §A. Remaining crashes: C2 Runway/Facility (model-GDL
align), C2m Jungle (`G_MTX`), C5 Control (BG portal), C7 Surface1
(`sndSetupSound`).

**D133 (session M-20) — intro "mostly black" is the D75/D76 steady state, not a
regression.** The M-18/M-19 handoff "Next task" listed "intro renders mostly
black" as a regression against the M-17 handoff line "the entire intro renders —
logos → gun barrel → cast".

Method: `git worktree add ../ge007-m17 9ec6121e` (COPY `data/`, never junction —
see the data-dir-junction-hazard memory), `./build-pc.sh ntsc-final`, capture
`GE_PCDUMP="20-900:20"`, `tools_pc/pixcount.py` per frame. Compared against the
integrator's HEAD (`0b5f5d1a`) capture.

| frames | M-17 `9ec6121e` | HEAD `0b5f5d1a` |
|---|---|---|
| 20–100 (legal screen) | non-clear **6677 (2.17%)**, bbox (50,265)-(589,446) | non-clear **6677 (2.17%)**, same bbox |
| 120–360 | 0–16 non-clear px (near-black) | 0–16 non-clear px |
| 380–400 (centred logo) | 20.5k–22.1k (6.7–7.2%), bbox ~(253,142)-(394,336) | 18k–24k (6–8%), bbox ~(250,138)-(390,340) |

The legal-screen frame is a static 2D framebuffer; a **pixel-identical**
non-clear count on two independently-built binaries is proof the render path is
unchanged. 2D/text/texture layers draw on both; the animated character-model
layers (Nintendo-logo transform, gun-barrel Bond figure, cast models) never
appear on either — this is **D75** verbatim ("skeletal/animated character models
never appear; non-animated 3D appears with a bad transform"). The M-17 "entire
intro renders" line was aspirational, not a measured coverage state.

Bisect not needed (state identical at the older commit). No cheap `#ifdef PORT`
regression exists; the fix is the parked D75 3D-pipeline work
(`docs/dev/GRAPHICS-BACKLOG.md`). `-level_09` re-verified PASS (690+ frames, 91.7%)
so the capture harness is sound.

Worktree removed (`git worktree remove --force ../ge007-m17` — safe because
`data/` was copied, not junctioned). Tree restored to `master`. Confidence:
**high** (pixel-identical legal-screen coverage; symptom is a verbatim match to
the open D75 description).

Secondary (M-20, not investigated per scope): Facility `-level_34` boot-crashes
`frames=0` PC `0x1400c3b77` (addr2line unresolved); Jungle `-level_37` renders
frames 1–2 then kernel-heartbeat hangs (`frames=2`). Both claimed PASS in
M-18/M-19 (D130/D131). Machine was only lightly loaded. Needs a clean-machine
re-verify (`docs/dev/LEVEL-STATUS.md`) before calling it a regression.

**D131 (session M-19) — Jungle `-level_37` C2m crash: `osVirtualToPhysical()`
truncates a compiled-symbol pointer inside a GBI matrix command.** Jungle
renders ~300 frames then AVs in `gfx_sp_matrix` (`gfx_pc.cpp:1046`,
`int32_t int_part = addr[…]`) with `addr = 0x401c68e0`. Caller is the `G_MTX`
case of `gfx_run_dl` → `seg_addr(cmd->words.w1)`.

Probe (`GE_MTXPROBE`, reverted) dumped the faulting DL — a well-formed model
render sequence in a per-frame DRAM GDL: `gSPMatrix(PROJECTION,LOAD)` w1 =
`0x700d8070` (valid), `gSPMatrix(MODELVIEW,LOAD)` w1 = `0x401acbe0` (bad),
`gSPMatrix(MODELVIEW,MUL)` w1 = `0x401c68e0` (bad, the fault), `gSPVertex`
w1 = `0x700da350` (valid). All lens `0x40`, opcodes sane — the walk is *not*
desynced, only the modelview-matrix pointers are wrong.

`0x401c68e0` == `(u32)(0x1_401c68e0)` — a compiled-module address with the
`0x1_00000000` high word dropped (module fixed-based at `0x140000000`, cf.
D94 `chraction.c:1243` comment). Source: `explosionRenderPropSmoke`
(`explosion.c:1501`) does
`gSPMatrix(gdl++, osVirtualToPhysical((void*)&dword_CODE_bss_8007A100),
G_MTX_NOPUSH|G_MTX_MUL|G_MTX_MODELVIEW)` — `&dword_CODE_bss_8007A100` is a
compiled `.bss` matrix; `osVirtualToPhysical()` is `(u32)(uintptr_t)va`
(`libultra.c:1207`), so the store truncates. The `[-1]` LOAD matrix is the
room matrix from `applyRoomMatrixToDisplayList` (same truncation path).
Why the projection matrix is fine: `get_BONDdata_field_10E0()` returns a
runtime `0x70xxxxxx` pointer that survives the `u32` cast. Why
`gSPDisplayList(&globalDL_0xNNN)` is fine: the port's `Gwords.w1` is a
64-bit `uintptr_t` (`include/PR/gbi.h:1730`) and `gDma1p` stores the full
pointer — only `osVirtualToPhysical` truncates.

~30 sites pass a compiled matrix/vtx symbol through `osVirtualToPhysical`
(`explosion.c`, `glass.c`/`glass2.c`, `blood_animation.c`, `bondview2.c`) —
each latent until that effect first draws, which is why only Jungle (an
early scripted explosion) tripped it in the no-input sweep.

**Fix** (`port/fast3d/gfx_pc.cpp` `seg_addr`): at the final fallthrough,
if `w1 ∈ [0x40000000, 0x70000000)` restore the module high word
(`(mod_hi | w1)`, `mod_hi` read from this TU's own load address). Everything
legit that reaches the fallthrough is either DRAM V1 (`≥0x70000000`), KSEG0
(`≥0x80000000`), an LSB-set / low-nibble segmented address, or a
sub-`0x800000` physical offset — all handled in earlier branches — so the
range is unambiguous. Narrow, ABI/pointer-width class (same as D94/D3x).
Verified: Jungle renders to logic-frame 2400+ crash-free at 91.7% non-clear;
`-level_20` (Silo) + `-level_24` (Archives) unregressed. **Sweep 18→19/21.**

Not fixed this session: **`-level_09` (BUNKER1) now crashes at boot**
(`frames=0`, PC `0x1400066fc`, fault `0x00a2fc68` — a stack pointer
`0x0fa2fc68` with the top nibble masked, i.e. another 28/24-bit segment-mask
applied to a real pointer). Reproduced on a **clean `git stash` checkout of
master `0a4b3bae`** → pre-existing regression, not caused by D131 (a
fast3d-only change cannot affect a pre-first-frame boot crash). Contradicts
the M-18 handoff's "`-level_09` unregressed" claim; likely fallout from a
D126/D128 propdef/portal change or an environmental sidecar/ROM mismatch.
Next session: `git bisect` D125→HEAD against `-level_09`. Silo also shows an
intermittent hang in the intro fly-down-to-Bond cinematic (user report;
not always reproducible — ran 300 frames clean here).

---

**D130 (session M-18) — Facility + Runway C2 crash was `romdataFixupFont`,
NOT the model-GDL relocation.** `-level_34` (and `-level_35`) fault in
`import_texture_i8` (`gfx_pc.cpp:821`) on `loaded_texture.addr = 0x72181ee8`,
an un-mapped wild pointer. gdb showed the fault DL is a 2D `G_TEXRECT` block
(`gfx_dp_texture_rectangle` → `gfx_sp_tri1(is_rect)`), with the wild pointer
baked into a preceding `G_SETTIMG` w1. Traced back through `GE_C2GDL` probes:

- `texLoadFromGdl` **never copies a `G_SETTIMG` (0xFD)** on these levels, and
  `texWriteLoadToTmemAddr/Zero` are **never called** — so the M-14 addendum's
  "model-GDL relocation writes non-16-aligned dst" hypothesis is wrong. A
  probe in `sub_GAME_7F0762E0` also confirmed `gdl` from
  `modelIterateDisplayLists` is still a **segmented `0x05xxxxxx` value**
  (`modelPromoteNodeOffsetsToPointers` does NOT promote DL Primary/Secondary),
  so the `Switches + (x & 0x00ffffff)` idiom is correct as written.
- The wild `G_SETTIMG` is emitted by `gDPLoadTextureBlock(gdl,
  curchar->pixeldata, G_IM_FMT_I, G_IM_SIZ_8b, …)` in
  `textRenderGlyphOutlined` (`textrelated.c`) while rendering the level-title
  string "Chemical Warfare Facility #2" — specifically glyph `#` (and `"`).
- `curchar->pixeldata` for BankGothic/ZurichBold glyph indices **0, 1, 2** is
  corrupt after `load_font_tables`: glyph 1 `.width = 0x01000000`, glyph 2
  `.pixeldata = 0x020002e8` (→ `+= font_base` → wild).

Root cause: `romdataFixupFont` (`port/src/romdata.c`) re-lays-out the N64 24B
`fontchar` array into the PC 32B array **in place** — backward over glyphs,
forward over the 5 leading `u32` fields per glyph. PC−N64 stride = 8, so
`d = s + 8*i`; for `i ∈ {0,1,2}` `d` overlaps the 24-byte source read span
and an early `*(u32*)(d+4k) = bswap(*(u32*)(s+4k))` write clobbers a later
`k`'s source. Glyph 1 (`Δ=12`): k=3/k=4 read bytes k=0/k=1 already
overwrote → `width = bswap(bswap(index)) = index = 0x01000000`. Glyph 2
(`Δ=20`): k=0 writes `s+20`; line 507 then reads `s+20` for the pixeldata
offset → `o = index = 0x02000000` → `o >= pixStart` → remapped to
`pcPixOff + (0x02000000 − pixStart) ≈ 0x020002e8`. Confirmed exactly against
a raw-bytes dump. On N64 the array is read-only rodata — never re-laid-out —
so no bug there; this is purely a PC-relayout aliasing bug.

Fix: stage all six N64 fields into a local `u32 f[6]` before writing any
field of `d`. (A previous iteration's write can never touch `s(i)` — for
`j > i`, `d(j)` starts past `s(i)`'s end — so a full pre-read is safe.)
`GLYPH_IDX` clamp in `textrelated.c` was tried and reverted: the crashing
string has no control byte, and a negative-index guard is a separate latent
concern (logged in porting-notes.md §A).

**Verification:** `-level_34` 0 crashes / ~14 runs (was ~50–100%); `-level_35`
0 crashes / 4 runs (was ~3/4). `-level_09`/`-level_20`/`-level_24` unregressed
(91.3% / 91.7% / 90.8% non-clear). `-level_37` (Jungle) unchanged — still the
separate C2m `gfx_sp_matrix` explosion-DL `G_MTX` crash at ~frame 300 (D75
family). Sweep **16 → 18/21**. All probes reverted; only
`port/src/romdata.c` (11 lines) changed. Confidence: **high** (exact
byte-level match, deterministic value, 2 levels fixed, no regression).

**D124-Facility addendum (session M-14 — partial, NOT fixed, out of time).**
*(Superseded by D130 — the model-GDL hypothesis below was wrong.)*
Re-instrumented with a `GE_C2GDL` probe in `sub_GAME_7F0762E0`
(`objecthandler_2.c`) + a bad-`G_SETTIMG` catcher in fast3d
(`gfx_pc.cpp` G_SETTIMG case). Both probes **reverted** — tree is clean.
Findings:

1. *Exact crash cmd.* `-level_34` faults in `import_texture_ia16`
   (`gfx_pc.cpp:757`, via `import_texture` :955) — NOT `import_texture_i8`
   as the old note said. The offending command in the relocated model GDL
   (dList base `0x7007b870`, cmd at `0x70080490`) is:
   `w0=0xfd900000` (`G_SETTIMG`, fmt=7/rgba? siz per bits, w=0),
   `w1=0x72181ee8`. `seg_addr(0x72181ee8)` returns it unchanged (top byte
   0x72 is not a segment) → deref of unmapped DRAM → AV. Preceding slot is
   `0xba 00 0e 02` (GE tex-load macro, w1=0) then a `00000000` slot.
2. *`Gfx` stride confirmed.* `Gwords{ uintptr_t w0; uintptr_t w1; }` →
   on x86-64 `w0`@+0 (8B), `w1`@**+8** (8B), `sizeof(Gfx)==16`. The probe
   dump (8-byte granular) shows w0 at slot+0 and w1 at slot+8, i.e. the
   command is genuinely `{0xfd900000, 0x72181ee8}` — the converter wrote a
   real 16-byte slot, it is not a half-slot artefact.
3. *`w1` is a bad `tex->data`.* This `G_SETTIMG` is emitted by
   `gDPSetTextureImage(gdl++, .., tex->data)` inside
   `texWriteLoadToTmemAddr` / `texWriteLoadToTmemZero` (`tex.c`). So the
   **texture-pool** entry `tex` returned by `texFindInPool()` has a
   garbage `.data` (should be a `0x05xxxxxx` seg-5 ref, resolved later by
   fast3d `segmentPointers[5]`). i.e. the fault is (at least partly) in
   the `texLoadFromModelFileHeader` → `texLoad` texture-pool path, driven
   off `objheader->Textures`, **not** solely the `texLoadFromGdl`
   command-stream copy. This partially contradicts the original
   D124-Facility hypothesis.
4. *Alignment smoking gun.* `GE_C2GDL` probe shows **every** model file's
   `objheader->Switches` base has `((uintptr_t)Switches & 15) == 1`
   (odd!). `mempAllocBytesInBank` (`memp.c`) does **zero** alignment —
   `allocation = pool->pos; pool->pos += bytes;` — so a single earlier
   odd-sized bank alloc leaves the STAGE pool cursor permanently odd and
   every subsequent model file loads at an odd address. `delta`
   (`romremaining - pcremaining`, the scratch-relocation offset) is also
   not 16-aligned (`&15` = 3, 7, 11 across files). `sub_GAME_7F0762E0`
   then forms `dst = Switches + (replacementgdl & 0x00ffffff)` and
   `src = Switches + gdloff + delta`; observed `dst&15` = 5, 7, 9, 13.
   The N64 code assumes the file base is Gfx-aligned (8 on N64). The
   converted model GDLs the sidecar carries also have per-DL offsets that
   are only 4/8-aligned (`gdloff0 & 15` = 4, 8, 12), so even a 16-aligned
   base would not make every DL 16-aligned.
   - *Note:* an odd/misaligned base is self-consistent between the
     converter writer (`Switches+off`) and the fast3d reader (seg-5 =
     `Switches`, `+off`), so it does **not by itself** corrupt command
     *content* on x86 (unaligned `uintptr_t` loads are tolerated). The
     bad `tex->data` (#3) must come from an *offset* miscalc, likely the
     model texture-blob offset assuming N64-sized (8B-Gfx) GDL extent
     while the PC sidecar's GDLs are pre-expanded to 16B — the texture
     data then sits at a stale/too-large offset (`0x72181ee8` is ~0x1.9M
     past the bank end `~0x706a8000`), OR `texLoad` reads a `textures[i]`
     descriptor at a wrong stride off the odd `objheader->Textures` base
     (`(u8*)filedata + sizeof(ModelNode*)*numSwitches`, D43/D45 PORT line).
   - *Next step:* probe `texLoad` (`image.c`) — dump `textures[i]`
     `TextureID` / offset / resulting `tex->data` for `-level_34`, and
     dump `objheader->Textures`, `numtextures`, `numSwitches`, and the
     sidecar's texture-section layout. Compare against `d43_emit.py` /
     `pcmodels` sidecar builder: does it emit model texture blob offsets
     in N64 (8B-Gfx) or PC (16B-Gfx) GDL-extent terms? That converter is
     the likely fix site (offline sidecar per the D43/D69/D88 rule),
     with a fallback `#ifdef PORT` 16-align of the STAGE pool cursor /
     the model file base in `mempAllocBytesInBank` or
     `load_object_fill_header`.
5. *Files touched this session:* none committed; both probe edits
   (`src/game/objecthandler_2.c`, `port/fast3d/gfx_pc.cpp`) reverted via
   `git checkout`. Docs only: this entry + `docs/dev/LEVEL-STATUS.md`.
   *Confidence:* crash cmd + `Gfx` stride **high**; `tex->data`-is-the-bad-
   value **high**; root cause of *why* `tex->data` is bad **low–medium**
   (two candidate mechanisms, neither proven).

---

## D186 — `Video.FpsCap` throttles the whole sim, not just presentation (M-39 diagnosed, M-42 clamp landed)

**Symptom (M-39, user report):** window title read "GoldenEye 007 - 1 fps";
the game had become a slideshow. Root cause was `Video.FpsCap = 10` in the
(gitignored) `data/ge007.ini` `[Video]` section — almost certainly dinged down
via the F10 options overlay (D184) during testing; ESC-close calls
`configSave()` so it persisted.

**Mechanism.** `gfx_set_target_fps(cfgFpsCap)` feeds `sync_framerate_with_timer()`
in `port/fast3d/gfx_sdl2.cpp`, which does `sysSleep(left)` + a `sysCpuRelax()`
busy-wait. That wait runs **inline on the game's scheduler thread**:
`src/sched.c __scExec` → `osSpTaskStartGo` (`port/src/libultra.c:1119`) →
`gfx_run` → `gfx_sdl_swap_buffers_begin` → `sync_framerate_with_timer`, and only
*after* it returns does `osSpTaskStartGo` post `OS_EVENT_SP`/`OS_EVENT_DP`. While
that thread is parked in the pacing wait it is not delivering VI-retrace events
either, so every game thread blocked on `osRecvMesg(retraceQ)` stalls with it —
the sim is dragged down to the render-cap rate. On console the RSP/RDP are
separate silicon and VI retrace is unconditional, so a "render cap" cannot slow
the CPU; there is no N64 analogue. Verified M-39 (`-level_09`, fixed 35 s
window): `FpsCap=10` → 5 frames; `FpsCap=30` → 5; `FpsCap=60` → frame ~300;
`FpsCap=0` → frame 900+.

**Landed (M-42, clamp only — `port/fast3d/gfx_sdl2.cpp` + `port/src/video.c`):**
any `0 < FpsCap < 30` is treated as `0` (uncapped) with a one-line warning, at
both the config-load site (`videoInit`, so a bad ini is normalised and persists
sane on the next `configSave()`) and the fast3d chokepoint (`gfx_sdl_set_target_fps`,
which also covers the live path from the F10 overlay). Caps ≥ 30 are unchanged.
Verified: `-level_09` framediff 3/3 within threshold at the default `FpsCap=0`
(golden-identical); with `FpsCap=10` the sim now runs to VI post ~1740 in a
30 s window (≈ the `FpsCap=0` rate) instead of ~300; `FpsCap=60` still paces
(~56/s, no warning).

**Still OWED (the real fix):** move the frame-pacing wait off the scheduler
thread so *any* cap only drops/duplicates presented frames without blocking
retrace delivery (e.g. pace in the host event-pump thread, or gate the swap
without sleeping the RSP path). Until then a low cap is refused rather than
supported. Same timing-compensation family as D117 / D134 / D155.
porting-notes.md §E.

---

## D183 — M-36 "Family A" texture line/pitch shear: DISPROVEN on `-level_36` (M-36)

**Brief:** fix the "diagonal grey static / comb interlacing" attributed to
`import_texture_*` assuming `full_image_line_size_bytes == line_size_bytes`
(`docs/dev/M-36-TRIAGE.md` Family A; bugs D176(b) Surface walls, D182(2)
file-select spiral). Repro: `-level_36` Surface, headless `GE_PCDUMP`.

### 1. The pitch hypothesis does not fire on Surface — measured, not argued

`full_image_line_size_bytes != line_size_bytes` can only be produced by
**`gfx_dp_load_tile`** (`gfx_pc.cpp:2166`, a windowed sub-rect load).
`gfx_dp_load_block` (`:2139`) sets `line = full = size_bytes` unconditionally,
so the two are equal by construction.

GE loads *every* texture through `gDPLoadBlock`: `texWriteLoadToTmemZero` /
`texWriteLoadToTmemAddr` (`src/game/tex.c:492,617`) emit
`gDPSetTextureImage` + `gDPSetTile(…, line = 0, …)` + `gDPLoadBlock` and never
a `gDPLoadTile`. Confirmed at runtime: `GE_DTEX` (now prints `line=`/`full=`
and a `<-- STRIDED` marker) over a Surface run — **0 of 64** logged imports
strided; the full `GE_TEXDUMP` set — **0 of 166**. So no importer on this level
ever does the flat-read-of-a-strided-source that the hypothesis needs.

### 2. The importers' width/height/pitch math is *verified correct*

New probe `GE_TEXRAW=1` (with `GE_TEXDUMP=1`) writes the raw source bytes
handed to each importer to `texdump/rNNN_f<fmt>_s<siz>_<w>x<h>.bin`. Scanning
each dump's vertical neighbour-difference over every candidate row pitch
4…128 finds the true pitch of the source image:

| texture | fmt/siz | tile | pitch fast3d uses | best-scoring pitch | vdiff |
|---|---|---|---|---|---|
| `r030` | IA8 | 54×54 | 56 | **56** | 0.22 (smooth) |
| `r021` | CI8 | 32×32 | 32 | **32** | 0.72 (smooth) |
| `r031` (the wall) | IA8 | 32×32 | 32 | none — flat ~3.7 at *every* pitch | — |

The importers pick exactly the pitch the data is stored at. (The 54→56 case is
`texAlignIndices`/`texChannelsToPixels` 8-byte row alignment,
`src/game/image.c:340,1740` — handled correctly today.)

Also re-tested and re-confirmed: applying `texSwapAltRowBytes`' odd-row u32-pair
swap to these dumps makes vdiff **worse** (0.22 → 1.57 on `r030`). **D159's
`#ifdef PORT` no-op is correct; do not re-enable it.**

### 3. What the Surface walls actually are

The cliff/perimeter wall texture is `GE_TEXI` fmt=3 siz=1 = **IA8, 32×32,
block size exactly 1024 B (no LOD chain)**. Its source bytes are a
high-frequency grey noise field: intensity nibble mean 6.5, full 0–15 range,
horizontal neighbour diff 2.98 / vertical 4.06 (uniform random ≈ 5.3); alpha
nibble is a constant 15. There is no pitch, no row swap and no bit-depth
reinterpretation that turns it into a coherent image — the *data itself* is
noise-like. It is decoded and uploaded faithfully; the on-screen "static" is
that texture tiled at ≈1 texel per pixel on the wall faces.

Two filtering theories were tested and **both failed** (temporary
`GE_MINFIX` probe in `gfx_opengl.cpp`, since reverted):

- forcing mip minification when the game point-samples
  (`gfx_opengl.cpp:752` `linear_filter ? … : GL_NEAREST`) — **no pixel change**;
  the walls are already `linear_filter == true`.
- forcing `mipmaps = true` on every sampler apply — **no change to the walls**,
  and it *regressed* the 2D HUD ammo digits to white blocks.

So the walls already get `GL_LINEAR_MIPMAP_LINEAR`. Not a filter bug.

### 4. Where D176(b) actually goes next (unresolved)

Remaining candidates, in order:

1. **The wrong texture is bound to those faces.** The next decisive step is ROM
   ground truth: decode Surface's texnums offline (the
   `TEXTURE-GLITCH-ANALYSIS.md` §7 toolchain) and check whether *any* Surface
   texture is this 32×32 IA8 noise field, or whether the rock wall should be a
   different, larger, structured texture. If it exists in ROM as-is, D176(b) may
   be **not a bug at all** (GE's Surface cliffs are a mottled grey rock) and the
   user report is really about *tiling density*.
2. **UV / tiling scale** — the texture repeats far too densely across each wall
   face. That is RC3/`Video.WrapFix` territory (D167) plus the `shifts`/`shiftt`
   the game sets per LOD tile in `texWriteTileLods` (`tex.c:596` passes
   `shifts = shiftt = lod`), which fast3d stores but the LOD-tile selection
   (D107 "always base tile") may be mismatching.
3. **D182(2)** (file-select spiral after re-entry) is *not* covered by this
   finding — it was never reproduced here, and it is the one symptom that could
   still be a real `gfx_dp_load_tile` strided case (front-end `texLoadFromGdl`
   paths do use windowed loads). The defensive fix in §5 would cover it if so.

**M-37 UPDATE (inconclusive, ~8-min timebox):** the §7 offline decoder + logs
live *outside* this checkout — a full offline ROM decode was not achievable.
Static findings: Surface's bg is `bg/bg_sevx_all_p.seg` (`bg.c:201`); the wall
texture is a `u16` image ID *embedded in the bg GDL stream*, not a static
per-level texnum list, so it cannot be grepped — needs a runtime `GE_DTEX`/
`GE_TEXRAW` correlation. `assets/images.def` has no rock/cliff names (only
`STATIC_NOISE 0x389`); Surface-named entries are mission-select thumbnails.
Leaning **"real texture / tiling-density" (candidate 2), low-med confidence** —
D183's own measurement (constant full alpha 0x0F + full-range noise intensity,
zero residual structure at any pitch) is the signature of an intentional grey
rock-detail tile, not a mis-decoded structured image. Decisive next steps:
(i) grep the *converted* `bg_sevx_all_p` GDL for opcode `0xC0` (`G_SETTEX`) —
settles the one remaining "wrong bind" scenario (stale tile state, TEXTURE-
GLITCH-ANALYSIS §2 RC1) offline; (ii) headless `-level_36` `GE_DTEX/TEXRAW=1`
logging the image ID bound to the cliff tris, look it up in `g_Textures[]`,
read + decode that ROM range per §4. Scratch note: M-37 session scratchpad
`D176b-rom-groundtruth.md`.

### 5. What shipped (`port/fast3d/gfx_pc.cpp` only)

- **De-stride in `import_texture()`**, before the format dispatch: when
  `full_image_line_size_bytes > line_size_bytes`, compact the strided rows into
  a contiguous scratch buffer and hand every importer a row-packed image with
  `line == full`. This is the correct behaviour the seven `SUPPORT_CHECK
  (full_image_line_size_bytes == line_size_bytes)` asserts (`gfx_pc.cpp:689,
  710, 739, 765, 791, 818, 866`) merely assert against — and those asserts
  compile out in the release build, so today the wrong read is silent. Only the
  CI8 importer strides correctly on its own (`:893`); it becomes a no-op stride
  after this change. **Strictly a no-op whenever `full == line`, which is
  166/166 loads on Surface and every `gDPLoadBlock` in the game** — hence
  golden-safe. Unverified as a *fix* (nothing in the current repro exercises it);
  kept because it is provably-correct and removes a latent silent-corruption path.
- **`GE_DTEX`** now prints `line=`/`full=` and flags `<-- STRIDED (pitch shear?)`.
- **`GE_TEXRAW=1`** (alongside `GE_TEXDUMP=1`) dumps raw importer-input bytes.

### 6. Verification

- `-level_09` golden framediff: **3/3 within threshold** (200/320/440;
  worst dmean 10.4, phash 21/1/4) — unchanged.
- `-level_36` before/after: **visually identical** (as expected — the de-stride
  never fires there). The wall static is unchanged and remains open.

*Confidence:* "Family A pitch shear is not the Surface bug" — **high**
(runtime-counted, 0/166). "The wall texture is decoded correctly" — **high**
(independent offline pitch scan of the raw importer input). "The shipped
de-stride is golden-safe" — **high**. "What D176(b) really is" — **low**;
needs ROM ground truth for that texnum.

---

## D176(a) — Surface black sky: env data is correct, the cloud-sky emit renders nothing (M-36, partial)

Surface (`-level_36`) sky is solid **black** where the N64 shows a warm
sunset cloud gradient. Headless `-level_36` + `GE_D176=1` probe
(`src/game/bgfog.c` `fogLoadLevelEnvironment`, `src/game/sky.c` `skyRender`):

- **The environment data is right.** `fog_tables[]` matches `Id=36`,
  `sizeof(EnvironmentRecord)=92`, `Sky.Clouds=1`, `Sky.RGB=96,96,128`,
  `CloudRGB=240,120,30` (the warm sunset), `CloudRepeat(skyheight)=10000`,
  `SkyImageId=0`, `IsWater=0`. So this is **not** a serialized-struct
  byte-order bug like D178 — `fog_tables[]` is a compiled-in C initializer
  and every field reads sane.
- **`skyRender` runs the cloud path** (not the `!Clouds` flat-fill early
  return). Corner probe: `eye=(-467.7, 374.2, -7190.5)`,
  `WaterConcavity=7.0`; the four screen-corner unproject rays come back
  `c0/c1.y ≈ +62` (`skyIsScreenCornerInSky` → TRUE) and `c2/c3.y ≈ -13.5`
  (→ FALSE). So the corner-classification switch value is
  `(1<<3)|(1<<2)|0|0 = 12` — the "top half of the screen is sky" case,
  which is correct for that camera.
- **Nothing draws.** Case 12 → `s1=4`, the sky region is built from the
  edge-vertex path and handed to `sub_GAME_7F097388` (pure-C project +
  perspective divide → screen coords) then `skyRenderTri` / `skyRenderFull`
  with `texSelect(&skywaterimages[SkyImageId=0], …)` and a
  `SHADE,ENV,TEXEL0,ENV` combine. The output is pure black, not even the
  `env->Red/Green/Blue` (96,96,128) fill — so the sky polygons are either
  degenerate after projection, culled, or the `skywaterimages[0]` bind is
  failing in a way that kills the primitive.

**Not yet root-caused.** Next steps (needs ~1–2 h / a dedicated pass):
1. Probe `sub_GAME_7F097388`'s output `unk28/unk2c` (screen x/y ×4 subpixel)
   for the 4 verts — are they on-screen and non-degenerate, or all clamped
   to one edge? That splits "projection math wrong" from "emit path wrong".
2. Check `skywaterimages` (`src/game/image_bank.c:284`,
   `globalbank_rdram_offset + GIMG_OFF(s_skywaterimages)`, PC offset
   `0xFB4`) actually resolves to valid `sImageTableEntry` records on PC, and
   that entry `[0]` (the Surface sky texture) loads — a bad global-image-bank
   offset (D69/pccg family) would give `texSelect` a junk texture.
3. fast3d-side: trace whether the sky tris reach `gfx_sp_tri` at all and
   with what verts/CC. `skyRenderTri` is a 500-line subdivided-tri emitter —
   a PC vtx/DL bug there is plausible (cf. D75 model-transform family).
4. Cross-check another cloud-sky level (Dam `-level_33` exterior, Statue
   `-level_22`) — if they're also black it's the emit path; if only Surface,
   suspect that level's sky image / `skywaterimages[0]`.

`GE_D176=1` probe left in tree (`#ifdef PORT`, env-gated, inert):
`src/game/bgfog.c` + `src/game/sky.c`. Separate from D176(b) (the tree/rock
"grey static", Family A / a different investigation).

### D176(a) — M-37 UPDATE: ROOT-CAUSED (static analysis, high confidence)

The emit path is not "wrong" — **the PC software RSP deliberately discards it.**
`port/fast3d/gfx_pc.cpp:2901`:

```cpp
case (uint8_t)G_RDPHALF_1:
case (uint8_t)G_RDPHALF_2:
case (uint8_t)G_RDPHALF_CONT:
    // on N64 skyRender uses these to render some types of skies and skybox water
    // by issuing low-level ucode commands G_TRI_FILL and G_TRI_SHADE_TXTR
    // the port renders the sky in a different manner
    break;
```

`skyRenderTri` / `skyRenderFull` build their geometry **exclusively** as
`gImmp1(gdl++, G_RDPHALF_1 / G_RDPHALF_CONT / G_RDPHALF_2, …)` pairs — a bespoke
packed triangle-raster stream (command byte in bits 24–31 = `G_TRI_SHADE_TXTR`/
`G_TRI_FILL`, `0x00800000` backface flag, subpixel Y coords, then S15.16 edge X +
dX/dY slopes via `sub_GAME_7F094298`, then per-vertex RGBA shade + S/T/W tex-coord
gradients, `G_RDPHALF_2` terminator) that GE's modified RSP ucode (`gmain.s`)
interprets. There is **no** `gSPVertex`/`gSP*Triangle` fallback and **no**
fill-rect on the textured case-12 branch — only `viSetFillColor(env->RGB)` sets
FILL state that nothing consumes. → dropped → solid black.

Explains every prior observation: `!Clouds` levels use a plain `gDPFillRectangle`
(→ work today); `Clouds` levels (Surface, Statue `-level_22`, Frigate `-level_29`,
Dam exterior) go through the RDPHALF stream (→ black). The projector
`sub_GAME_7F097388` is pure float math off endian-clean matrices — not the bug.

**No PD-port shortcut:** `pd_port/port/fast3d/gfx_pc.cpp:2512` has the identical
no-op + comment and PD's `sky.c` emits the same stream; no port-side sky
replacement exists anywhere in the PD tree. The comment is aspirational in both.

**Fix (Phase 2, ~1 session, genuine new fast3d code):** decode the RDPHALF stream
in `gfx_pc.cpp` — accumulate words, on a `G_RDPHALF_1` carrying a tri opcode start
a primitive, synthesise screen-space `LoadedVertex[3]` (bypass model/proj
transform like the 2D `gSPTextureRectangle` path; screen XY = value/4 − viewport;
shade + ST from integrating the slopes over the vertex delta), rasterise with the
current combiner/tile, close on `G_RDPHALF_2`. Full stream spec + the alternative
`#ifdef PORT` sky.c option in the M-37 scratch note.

Cheap first step: headless `-level_22` / `-level_29` — confirm also black (proves
emit path, not a Surface asset).

### D176(a) — M-42: `-level_22` headless repro CONFIRMED

`-level_22` (Statue Park, night) at frame ~360, **bare boot, no input** — the
spawn/intro camera already faces the horizon: the entire upper half of the
frame is **pure black** `(0,0,0)` where N64 shows a moonlit sky; the park
pillars and ground render fine, HUD present. Confirms the defect is the shared
cloud-sky emit path (RDPHALF stream dropped), not a Surface-specific asset.
`-level_22` frame 360 is the cleanest headless verification target for the
RDPHALF-decoder fix — no `GE_INPUTSCRIPT` needed.

Note: `-level_36` / `-level_43` (Surface) bare-boot intro cameras point **down
at the terrain**, so a no-input capture there does not frame the sky. `-level_29`
(Streets, night) top-of-frame reads dim `(29,24,22)` not pure black — needs a
visual check to distinguish "dark night sky drawn" from "partial".

### D176(a) — M-43 UPDATE: full stream decode + two implementation paths (static analysis)

The M-37 scratch note (`D176a-sky-rootcause.md`) was never committed. This is the
recovered, complete spec, re-derived from `src/game/sky.c` and `include/PR/gbi.h`.
No build, no run — analysis only. **Implementation is still owed and is a full
session; this entry exists so the next pass starts from the format, not a
grep.**

**What the stream actually is.** `skyRenderTri` (`sky.c:1480`, textured tris),
`skyRenderFull` (`sky.c:1976`, textured quads = 2 tris) and the untextured
`G_TRI_FILL` path emit a **verbatim N64 RDP triangle command**, byte-for-byte,
chopped into 32-bit halves and carried as `gImmp1(G_RDPHALF_1 / _CONT / _2, word)`
pairs. GE's modified RSP ucode (`rsp/graphics/gmain.s`) does nothing clever with
them — it copies the words into a DMEM scratch buffer and DMAs the assembled
command straight to the RDP as a `G_TRI_FILL` (0xc8) or `G_TRI_SHADE_TXTR` (0xce)
edge-walked triangle. So "decode the stream" == "own an RDP triangle rasteriser
with edge + shade + S/T/W coefficient planes." fast3d has no such thing (it is an
RSP-level vertex/`gSP*Triangle` interpreter); `pd_port` has the identical no-op.

**Word-by-word layout** (all values are S15.16 or int/frac-split 16.16 unless
noted; `sub_GAME_7F094298(f)` = clamp ±32767.9 then `(s32)(f * 65536)`, i.e. a
float→S15.16 fixed convert):

1. *Edge header* — 2 words:
   - `w0 = (opcode<<24) | (backface ? 0x00800000 : 0) | (s32)YL`
     where opcode is `G_TRI_SHADE_TXTR` or `G_TRI_FILL`, backface = `sp444 < 0`
     (signed double area after the 3-way `unk2c` vertex sort), `YL = sp47c->unk2c`
     (lowest vertex, screen-Y × 4, subpixel).
   - `w1 = ((s32)YM << 16) | (s32)YL_mid` → actually `(s32)sp480->unk2c << 16 |
     (s32)sp484->unk2c` = `YM<<16 | YH` (mid, high). N64 order is YL/YM/YH; here
     the code sorts so `sp484 ≤ sp480 ≤ sp47c` on `unk2c`, then emits
     hi=sp484, mid=sp480, lo=sp47c.
2. *Edge slopes* — 3 word-pairs, `(X.16.16 , dXdy.16.16)` for the L, M, H edges:
   - `(7F094298(sp480->unk28 * 0.25), 7F094298(sp384))` — XL, DxLDy
   - `(7F094298(sp410),               7F094298(sp394))` — XH, DxHDy
   - `(7F094298(sp408),               7F094298(sp38c))` — XM, DxMDy
   `sp384/38c/394` are the edge inverse-slopes (`dx/dy`) clamped to ±1877;
   `sp408/sp410` are the H/M edge X starting values back-stepped to the YH
   scanline (`sp37c` = sub-pixel fraction of `sp484->unk2c * 0.25`).
   `*0.25` because `unk28/unk2c` are stored ×4.
   **`skyRenderTri` returns here if `!textured`** (G_TRI_FILL: header + edges
   only).
3. *Shade coefficients* (textured only) — 8 word-pairs = the RDP shade DMEM
   block: RGBA base, DrDx/DgDx/DbDx/DaDx, DrDe.., DrDy.. each as an int-parts
   word then a frac-parts word (`(v & 0xffff0000) | (next & 0xffff0000) >> 16`
   packs two ints; the matching `<<16 | (next & 0xffff)` packs the fracs). Values:
   `sp210[0..3]` = colour at origin, `sp290[0..3]` = d/dx, `sp2b0[0..3]` = d/dy,
   `sp230[0..3]` = d/de, all run through `7F094298`. Colours are the per-vertex
   `SkyRelated38.rgba` barycentrically solved to a plane via `sp440`
   (= 1/doubled-area) and the `sp3c8..sp3d4` edge deltas.
4. *Texture coefficients* (textured only) — 8 word-pairs, same int/frac packing,
   for S, T, W: `sp210[4..6]`, `sp290[4..6]`, `sp2b0[4..6]`, `sp230[4..6]`, each
   scaled by `sp190` (an LOD/overflow clamp: `1/max(perspective-corrected
   gradient magnitude / 1024, 1)`). The final pair is emitted with
   `G_RDPHALF_2` (not `_CONT`) — **that is the stream terminator.**

Total data words: **2 + 6 + 16 + 16 = 40** for `G_TRI_SHADE_TXTR` (20 RDPHALF
pairs), **2 + 6 = 8** for `G_TRI_FILL`. The 2-word "header" is exactly the raw
N64 RDP triangle command's word0 (`[cmd 8b][backface 1b @23]...[YL 14b]`) +
word1 (`[YM 14b][YH 14b]`) — matches the canonical hardware format; the shade
and texture blocks are the standard 16-word int-halves-then-frac-halves RDP
coefficient layout (cross-checked M-43 against `include/PR/gbi.h` immediate-word
packing conventions — the low-level RDP tri macros themselves are RSP-internal
and not in this tree's `gbi.h`).

`skyRenderFull` is the same but solves the plane over 4 corner verts and emits
the quad as two RDP tris sharing the coefficient blocks; its own header pair also
ends the first tri and a second header starts the second.

**Vertex data available upstream** (before the DL): `sub_GAME_7F097388`
(`sky.c:1398`) fully projects each sky corner and writes `SkyRelated38`:
`unk28` = screen X ×4, `unk2c` = screen Y ×4 (minus `WaterConcavity*4`),
`unk30` = screen Z (0..0x7fff), `unk34` = 1/w, `unk20/unk24` = S/T, `rgba` =
per-vertex colour. **This is a complete screen-space vertex.** The projection is
pure float math off endian-clean matrices — not the bug, and reusable as-is.

---

**Path A — RDP triangle rasteriser in `port/fast3d/gfx_pc.cpp` (port-only, "correct").**
At `case G_RDPHALF_*` (`gfx_pc.cpp:~2901`): accumulate words into a small buffer;
when a `G_RDPHALF_1` carries a `0xc8`/`0xce` opcode start a command, close on
`G_RDPHALF_2`. Then either (a) feed a real edge-walk rasteriser writing into the
current render target with the active combiner/tile, or (b) **invert the plane
equations** — you have YH/YM/YL, the three edge X-at-Y and slopes, so recover the
3 screen (x,y); evaluate the RGBA and STW planes at those 3 points to get 3
`LoadedVertex` (already-projected: set `.x/.y` from screen coords mapped to
NDC-ish like the `gDPTextureRectangle` path, `.z` from a fixed sky depth, `.w`
from `unk34`), then hand to the existing `gfx_sp_tri` / GL path with a forced
"2D, no model matrix" flag. (b) is a few hundred lines and reuses the whole
existing combiner/texture pipeline; (a) means a new software rasteriser. Prefer
(b). Cost: ~1 session, genuinely new code, verifiable headless.

**Path B — swap the emit in `src/game/sky.c` behind `#ifdef PORT` ("cheap").**
`skyRenderTri` / `skyRenderFull` already hold 3–4 fully-projected screen-space
`SkyRelated38` verts. Under `#ifdef PORT`, skip the whole RDPHALF block and emit
a normal `gSPVertex` of screen-space `Vtx` (XYZ from `unk28/2c/30` ÷4, ST from
`unk20/24`, RGBA from the struct) + `gSP1Triangle` / `gSP2Triangles`, with a
`G_TEXTURE`/combiner setup mirroring what the RDP path implied
(`SHADE,ENV,TEXEL0,ENV`, `texSelect(&skywaterimages[SkyImageId], …)` already runs
just before). fast3d already eats screen-space verts (that is the front-end / HUD
path). ~30–60 lines, no new fast3d code. **Cost: a few hours.** Tension with the
"no game-logic edits" rule — but this is a pure N64-RSP-idiom substitution (same
class as the `G_TRI4` and dynamic-lighting `#ifdef PORT`s already in `src/`), it
changes no behaviour on N64 (`#else` keeps the stream verbatim), and it is
exactly the "if a game file seems to need a behavioural change, the fix belongs
in `port/`" *narrow* exception for a hardware idiom that cannot be isolated in
`port/` without reimplementing the RDP. Document as D176(a) in §F if taken.

**Recommendation:** Path B first — it lights up every cloud-sky level
(Surface `-level_36`, Statue `-level_22`, Frigate `-level_29`, Dam exterior) for
a few hours' work and de-risks Path A by giving a visual ground truth. Path A
stays the "right" long-term answer if Path B's screen-space verts show
seams/precision artefacts vs. the RDP's subpixel edge walk. Either way the
verification target is **`-level_22` (Statue, night) frame ~360, no input** — the
M-42 clean headless black-sky repro.

**Cheap first step (unchanged, still not done):** headless `-level_22` /
`-level_29` — confirm also black, proving it is the shared emit path and not a
Surface-only asset.

### D176(a) — M-46: Path B implemented (`src/game/sky.c`, `#ifdef PORT`)

Path B is in. `src/game/sky.c` only; N64 path kept byte-for-byte under `#else`.

**What changed.** A new file-scope `#ifdef PORT` helper `skyPortRenderPoly(gdl,
SkyRelated38 **v, nverts)` and an early `#ifdef PORT` return in both
`skyRenderTri` (3 verts, `arg1..arg3`) and `skyRenderFull` (4 verts,
`arg1..arg4`), placed right after the `sp444/sp488 == 0` degenerate-area guard
and `1.0f / area` so all shared early-outs still run. The helper:

- `dynAllocateVertices(n)` + two `dynAllocateMatrix()` from the vtx pool
  (same lifetime idiom as the bondview2 watch-gauge ortho overlay and
  explosion.c billboards).
- `guOrtho(proj, l, l+w, t+h, t, -32768, 32768, 1)` where `l/t/w/h =
  getPlayer_c_screen{left,top,width,height}()` — maps the already-projected
  pixel coords 1:1 into clip space, Y flipped (screen-down → NDC-up); identity
  modelview. `gSPMatrix` LOAD PROJECTION then MODELVIEW. `bg.c` reloads the
  perspective projection at the top of `bgLevelRender` (lines 638/689/…), so
  leaving ortho loaded after `skyRender` is safe.
- Per vertex: `ob = {unk28, unk2c, unk30} * 0.25` (the `*4` subpixel scale the
  RDP path used); `tc = {unk20, unk24} * 32` (standard N64 S10.5); `cn = rgba`
  from the struct.
- `gSPClearGeometryMode(G_LIGHTING|G_CULL_BOTH|G_FOG)` +
  `gSPSetGeometryMode(G_SHADE|G_SHADING_SMOOTH)`, then `gSPVertex` +
  `gSP1Triangle(0,1,2)` (tri) / `gSP2Triangles(0,1,3, 0,3,2)` (quad, winding
  irrelevant with culling off). Combiner (`SHADE,ENV,TEXEL0,ENV`), env colour
  and `texSelect(&skywaterimages[SkyImageId/WaterImageId], 1,0,2)` are already
  emitted by the caller immediately before — the helper does not touch them.

**Verified.** `src/game/sky.c` compiles clean under the real PC build flags
(`gcc.exe … -DPORT=1 -std=gnu11 -Wall`, exit 0, no new warnings) with the
worktree's `build-pc/` CMake configuration. Full link + run NOT done: the
worktree has no `data/` ROM so assets can't be extracted here. Every symbol the
helper uses (`dynAllocate*`, `guOrtho`, `guMtxIdent`, `gSPVertex`,
`gSP{1,2}Triangle{,s}`, `osVirtualToPhysical`) is already linked by the compiled
set, so link is expected clean.

**Owed to a human (cannot headless-verify):**
1. Eyeball `-level_22` (Statue, night) frame ~360, bare boot no input — the
   M-42 black-sky repro. Expect a rendered moonlit sky in the upper half.
   Also spot-check `-level_36` (Surface), `-level_29` (Streets).
2. **Untested risk — texture scale.** `tc = unk20/unk24 * 32` is the
   conventional guess; `unk20/unk24` come out of `sub_GAME_7F097388` as
   `unk0c * (65535/65536)` with `unk0c ≈ worldX*0.1` for clouds. If the sky
   texture is tiled wrong / smeared, the fix is localised to the two `tc[]`
   lines in `skyPortRenderPoly` (multiplier, or an explicit `gSPTexture`
   scaling override) — no other code moves.
3. **Untested risk — ortho Z / depth.** `ob[2] = unk30*0.25` in `[0,8192]`
   under an `[-32768,32768]` ortho; sky renders first into a cleared buffer so
   depth ordering should not matter, but if the sky Z-fights or is culled,
   drop `ob[2]` to 0.
4. Seam/precision check vs the N64 RDP subpixel edge-walk (the Path A
   motivation) — only relevant once it renders at all.

**Confidence:** MODERATE. The emit-path substitution is sound and matches the
established screen-space-overlay idiom; compile is clean. The two numeric
unknowns (texture scale, and whether fast3d's viewport transform composes with
this ortho exactly as the projection math intends) need one visual check.

### D176(a) — M-82: PR #18 rebased; Defect 1 (texcoord) fixed by static analysis

- **PR #18 rebased onto `main`** (was 5 days stale, `docs/porting-notes.md`
  conflict resolved). Builds clean; `bunker1` verify PASS, `worst_cell`
  unchanged (sky.c change is `#ifdef PORT` and only touches cloud-sky levels).
- **Defect 1 root-caused, no visual iteration needed.** The M-46 "untested
  risk — texture scale" is resolved:
  - Cloud tile = `s_skywaterimages[0]` = `IMAGE_CLOUDS_GRAYSCALE`, **64×64
    IA8, `G_TX_WRAP`** (`assets/oddtextures.c:646`). It *tiles* — S/T are not
    [0,1) fractions.
  - GE's own screen-space textured-quad idiom is `tc = width << 5`
    (`src/game/glass2.c:678`), so the patch's `× 32.0f` **is the right scale**.
  - `unk20`/`unk24` are texel coords reaching ±20 000 near the horizon (M-47),
    so `(s16)(S × 32)` **overflows** → the streak.
  - **Fix (committed):** per-primitive phase-fold — subtract the same
    multiple of the 64-texel tile period from every vertex of one triangle
    before scaling. Preserves inter-vertex deltas (hence the interpolated
    texture); keeps `tc` in s16 range. Exact for any triangle whose S/T span
    is < ~1000 texels — i.e. everything except triangles straddling the
    horizon line, which remain Defect 2.
- **Defect 2 (coverage / horizon-straddling tris) still open** → adaptive
  tessellation, denser toward the horizon.
- **Owed:** one `-level_22` visual check — see `docs/dev/D176a-SKY-NOTES.md`
  "Verification ask".

### D176(a) — fidelity pass: "scrolls too fast / tiles too visibly" ROOT-CAUSED + FIXED (linear vs. perspective-correct tc interpolation)

M-84/M-87 playtest found Path B partially working: where the sky renders it
"scrolls too fast and repeats too visibly" (tiling period reads too short).
Root-caused with a `GE_D176SCROLL` probe (temporarily added to
`skyPortRenderPoly`, removed after use — same throwaway-probe pattern as
`GE_D176`):

- **`skyPortRenderPoly` placed every sky vertex under a plain pixel-space
  `guOrtho` projection**, i.e. every vertex got clip `w = 1`. fast3d's normal
  GPU pipeline then interpolates `tc` (and every other varying) *linearly in
  screen space* across the primitive — correct only when every vertex shares
  the same `w`.
- **They don't.** `unk20`/`unk24` (S/T, texel units) are computed upstream
  (`sub_GAME_7F097388`, `sky.c:1401`) from the *world-space* cloud-plane
  intersection, with **no perspective divide applied** — exactly like the
  real RDP's `G_TRI_SHADE_TXTR` coefficient block, which carries `S·w′`/`T·w′`/`w′`
  triples (`docs/dev/D176a-SKY-NOTES.md` wire-format section) and relies on
  the RDP's own perspective-correct texture unit to divide back out per
  pixel. `SkyRelated38.unk0c` (a field Path B wasn't reading before this fix)
  already carries that raw camera-space `w` (`sky.c:1450`,
  `arg5->unk0c = sp68[3]` — the same `w` sub_GAME_7F097388 itself divides by
  to get screen X/Y, so it's directly reusable).
- **Measured on `-level_22` (Statue), frame ~330:** one sky quad's four
  corners had camera-space `w` = 5469/5469 (top edge, near) vs.
  207300/207257 (bottom edge, at the horizon) — a **38x ratio** — over an S
  span of ~30000 texels. Linear interpolation smears that whole 30000-texel
  delta evenly across every screen pixel of the quad; perspective-correct
  interpolation concentrates it almost entirely into the last few pixels
  next to the horizon-grazing vertices (hand-computed check: naive linear
  interpolation at the screen-space midpoint gives S≈-9090, true
  perspective-correct gives S≈-884 — an order of magnitude apart). The
  linear version reads exactly as reported: a uniformly dense, aliased
  "hatching" over the *entire* visible sky instead of a gentle gradient that
  only gets busy right at the horizon.
- **Fix (`src/game/sky.c`, `skyPortRenderPoly`, `#ifdef PORT`):** replaced
  the `guOrtho` load with a hand-built projection matrix that gives each
  vertex its *real* camera-space `w` (from `unk0c`) instead of `w=1`, while
  keeping every vertex's screen-space `x,y` position bit-identical (the
  matrix is constructed so `clip.xy = ndc.xy * w` and `clip.w = w`, so the
  GPU's perspective divide recovers the same `ndc.xy` regardless of `w` —
  only the *interpolation* across the primitive changes). `Vtx.ob[]` is
  `s16` and can't hold raw `w` (up to ~2×10^5 near the horizon), so each
  vertex pre-divides by a per-primitive `wScale` (`max|w| / 30000` when that
  exceeds 30000, else 1) and the projection matrix multiplies back by
  `wScale` for `clip.x/y/w`; `ob[2]` (unused for depth — the sky draws first
  into a cleared buffer, M-46) carries `w/wScale` through to the matrix's
  w-column. This rides the exact same fast3d vertex/triangle/GLSL pipeline
  every other textured draw in the game already depends on for correct
  perspective (`gl_Position = aVtxPos;`, `port/fast3d/gfx_opengl.cpp:320` —
  default GLSL `varying` interpolation is perspective-correct whenever
  `w != 1`); no `port/fast3d` changes.
- **Verified:**
  - `tools_pc/verify.sh bunker1` — PASS, `frames=3`, no crash (non-sky level,
    confirms no regression).
  - `-level_22` (Statue) and `-level_41` (Cradle) headless `GE_PCDUMP`
    captures, before/after, eyeballed. The intro camera pan isn't frame-exact
    between runs (small wall-clock-driven timing drift in the pre-gameplay
    camera — a separate, out-of-scope-for-this-task timing question, not the
    D193/D209 gameplay-tick class which is already fixed), so an exact
    per-pixel diff isn't reliable, but the *qualitative* pattern change is
    unambiguous and reproduces every run: before, uniform parallel aliased
    streaks over the whole sky; after, a proper perspective-converging fan —
    coarse near the top (near vertices), converging to fine detail only near
    the horizon edge (far vertices) — the textbook look of a textured plane
    correctly extending to a horizon. Cradle (`-level_41`) sky at default FOV
    renders a normal blue sky with cloud streaks, not black.
  - Not verified: an actual side-by-side against real N64 footage (none
    available in this environment), and whether the residual "busy" look at
    the very top of Cradle's sky (still fairly dense) is acceptable or needs
    Defect 2's tessellation on top — see below.
- **Interaction with Defect 2 (coverage / horizon-straddling tris, still
  open):** this fix corrects the interpolation *mode* (linear → perspective)
  but does not add tessellation. A primitive whose w-ratio between vertices
  is this extreme will still show real (now correctly-shaped, but still
  dense) detail concentrated at its horizon edge — finer subdivision near
  the horizon (Defect 2) is what would spread that out further and is still
  the right follow-up for full visual quality. Defect 2 itself (vertical
  coverage / horizon-straddling triangles needing adaptive tessellation,
  and the separate FOV-coupled black-sky-past-frustum-edge symptom on Cradle
  at high FovScale, D218/D222-adjacent) is **unchanged/still open** — not
  attempted this pass.
- **Confidence:** HIGH on root cause (confirmed numerically: measured 38x
  `w` ratio + hand-computed order-of-magnitude interpolation error) and on
  the fix being a real, correct improvement (reproducible qualitative
  before/after pattern change, matches the mathematically expected
  perspective-converging shape). MODERATE on "fully resolves the reported
  symptom to N64 parity" — no N64 reference capture to diff against in this
  environment; a human should eyeball `-level_22` and `-level_41` and say
  whether the remaining density right at the horizon reads as acceptable or
  still warrants Defect 2's tessellation work.
- **Porting-notes quirk added:** `docs/porting-notes.md` §D — screen-space
  "already-projected" GBI overlays (HUD/front-end idiom) are only safe to
  place under an identity/orthographic (`w=1`) transform when every vertex
  in the primitive really does share one depth/scale; a primitive whose
  per-vertex data was computed via a *world-space* perspective projection
  upstream (sky, and potentially other N64 RDP-immediate idioms) needs its
  real `w` preserved through to the GPU, or every non-positional varying
  (texture coords, vertex colour) interpolates wrong across the primitive.

---

## D177 — Ladders non-functional: `count`/`rooms` land in the high half of a widened pointer (M-36)

**FIXED (M-36).** Climbing was completely dead — `MoveBond`
(`src/game/bondview2.c`) probes for a ladder tile every frame via
`stanTileDistanceRelated(&curLocus)` and gates the whole ladder-collision
path on `stanGetLocusCount(&curLocus)`, which was always 0 on PC.

Two compounding pointer-width ABI bugs (same class as D79/D90), all in the
stan navigation layer — **no game logic touched**:

1. **The ladder signal was written into the wrong half of a pointer.**
   `stanCheckLinkedSpecialTile` receives the caller's
   `struct StandTileLocusCallbackRecord` typed as `s32 *outFlags` and does
   raw `outFlags[0] = 1` (FORCECROUCH) / `outFlags[1] = 1` (LADDER). The
   record's first member is `s32 *rooms` — 4 bytes on N64, **8 on PC** — so
   on x86-64 `outFlags[1]` is the *upper* 4 bytes of `rooms`, and `count`
   (which moved from +4 to +8) never gets written. `stanGetLocusCount()`
   reads `record->count` → still 0 → ladder path never runs. FORCECROUCH
   survived by luck (LE low half of `rooms` is at +0, and
   `stanGetLocusField0` truncates it back). Fix: `#ifdef PORT` cast to the
   real struct and write `->rooms` / `->count` by name.
2. **`curLocus` was too small to hold the record.** It was declared
   `struct move_bond_temp_struct` — a 2-word "placeholder while matching"
   (`bondview.h`). 8 bytes on N64 = exactly the record; on PC the record is
   larger, so `stanTileDistanceRelated`'s field writes (the D90 zero-fill)
   and the fixed `count` store overflowed the local. Fix: `#ifdef PORT`
   declare it as the real `struct StandTileLocusCallbackRecord`.
3. **`stanGetMoveBondCollisionTiles` would have AV'd on the first climb.**
   Once the ladder path actually runs it calls
   `stanGetTileOrderedPointWorldPos(…, (coord3d *)((s32)coords + off))` —
   `(s32)coords` truncates the 64-bit `&bondCollision` stack pointer → the
   four quad corners written through a garbage address. Fix: `PORT_PTRADD`
   macro (`uintptr_t` arithmetic), added at the top of `stan.c`.

Touched: `src/game/stan.c` (macro + `stanCheckLinkedSpecialTile` +
`stanGetMoveBondCollisionTiles`), `src/game/bondview2.c` (`curLocus` type).
All `#ifdef PORT`, N64 path kept verbatim under `#else`. Candidate root-cause
came from an M-36 subagent; parts 1/3 were its work, part 2 (the local size)
found on review. **Verified:** build links clean; `-level_09` framediff 3/3;
`-level_36` (Surface, has mandatory ladders) boots crash-free.
**Interactive climb test — DONE (M-37): user confirmed climbing works
in-game.** Merged to `main` via PR #5 (`955349db`).

porting-notes.md §C (pointer-width in ROM/record structs).

---

## D132 — D88 propDefs layout audit (M-20)

Static analysis only (no build, no game run). Goal: prove
`tools_pc/d88_propdefs.py`'s per-type cursor matches the real compiled PC
struct layout (8-byte pointers, 8-byte alignment) for every `PROPDEF_*`
record type any of the 21 solo `Usetup*Z` levels actually emits, and that
the handler set is total.

### Handler totality

Authoritative per-level type histogram: walk each level's `propDefs`
region with `d88_propdefs.PROPDEF_N64_WORDS` as the stride (the same
table `convert_stream` uses; it walks all 21 to an exact region-end
match, and `d125_check.py` confirms offline emit == converter for all 21
— so the stride table and the encountered type set are sound).

Union of all `type` bytes emitted across the 21 levels:

    1 2 3 4 5 6 7 8 9 10 11 12 13 14 17 18 19 20 21 22 23 24 25 26 27 28
    30 32 33 34 35 36 37 38 39 40 42 43 44 45 46 47 48

Never emitted (in the enum, absent from every shipped solo level):
15 (DEBRIS), 16, 29 (OBJ_DEPOSIT), 31 (OBJ_NULL), 41.

`tools_pc/d88_propdef_scan.py` is **stale / unreliable** — its private
`WC` + `SIZEOF_N64` guess table desyncs mid-walk and reports spurious
`unmapped type NOTHING`. `d88_propdefs.convert_stream` is the ground-truth
walker. (Scan tool left as-is; a follow-up could point it at
`PROPDEF_N64_WORDS`.)

Every emitted type has a non-generic handler **or** is provably safe
under the generic arm (header `_hdr_word` + `_bswap32` of the remaining
words, no pointer widening, no sub-word packed field):

- Generic-safe: 2 (DOOR_SCALE, `s32 Scale`), 23 (OBJECTIVE_START — see
  below), 24 (OBJECTIVE_END, header only), 25/26/27/28 (the meaningful
  field is the `s32 ObjRefID` at 0x4 that `get_status_of_objective` reads
  via the `MissionObjectiveRecord` cast — the `u16 unk4` in the stub
  structs is not how it is accessed, so `_bswap32` is correct), 34
  (OBJ_COPY_ITEM, 3 scalar words), 46 (CAMERAPOS / `CutsceneRecord` —
  `coord3d pos; f32 theta; f32 verta; s32 pad`, **no pointer**, 7 words,
  `PROPDEF_PC_BYTES[46]=28` correct).

### Per-type layout comparison (audited types)

Legend: N64 word -> PC byte offset; check = converter cursor lands on the
same PC offset the compiler would.

**ObjectRecord prefix** (shared by 3/5/12/17/36/42/43 and the head of
1/4/6/7/8/10/11/13/20/21/39/40/45/47). N64 0x80 / 32 w -> PC **144 B**.
Field-by-field: header@0; obj/pad `_hh_word`@4; flags@8; flags2@12;
prop*@16(+8); model*@24(+8); Mtxf mtx@32..96; runtime_pos@96..108;
runtime_bitflags@108; collisiondata*@112(+8); projectile*@120(+8);
maxdamage@128; damage@132; shadecol@136 (verbatim); nextcol@140;
sizeof **144**. Converter `_emit_object_prefix` cursor lands on every one
of these. MATCH.

**DOOR (1)** — `DoorRecord`, tail N64 w32..w63. The C3r Bunker2 suspect.
PC struct = **296 B** (`PROPDEF_PC_BYTES[1]=296`, `sizepropdef` PORT
`return 74`).

| field | N64 off | PC off (compiler) | converter | ok |
|---|---|---|---|---|
| linkedDoorOffset s32 (read-b4-write id) | 0x80 | 0x90 | 0x90 `_bswap32` | yes |
| maxFrac..maxSpeed 5xf32 | 0x84..0x98 | 0x94..0xa8 | same | yes |
| doorFlags u16 / doorType u16 | 0x98 | 0xa8 (`_hh_word`) | 0xa8 | yes |
| keyflags..doorOpenSound 3xu32 | 0x9c..0xa4 | 0xac..0xb4 | same | yes |
| frac..speed 5xf32 | 0xa8..0xb8 | 0xb8..0xc8 | same | yes |
| openstate s8/unkbd s8/calcopacity s16 | 0xbc | 0xcc (byte-pattern) | 0xcc | yes |
| TintDist s32 | 0xc0 | 0xd0 | 0xd0 | yes |
| CullDist s16/soundType s8/fadeTime60 s8 | 0xc4 | 0xd4 (byte-pattern) | 0xd4 | yes |
| linkedDoor* | 0xc8 | 0xd8 (8-al) +8 | 0xd8 +8 | yes |
| unkcc* (Vertex*) | 0xcc | 0xe0 +8 | 0xe0 +8 | yes |
| bbox (u32 + `struct bbox` 24B = 28B, 7 w) | 0xd0 | 0xe8..0x104 | 0xe8..0x104 | yes |
| openedTime u32 / portalNumber s32 | 0xec / 0xf0 | 0x104 / 0x108 | same | yes |
| openSoundState* | 0xf4 | 0x110 (8-al, 4B pad) +8 | 0x110 +8 | yes |
| closeSoundState* | 0xf8 | 0x118 +8 | 0x118 +8 | yes |
| lastcalc60 union s32/f32 | 0xfc | 0x120 | 0x120 `_bswap32` | yes |
| **sizeof** | 0x100 | **0x124 -> pad 296** | 296 | yes |

**Conclusion: the DOOR converter cursor matches the real PC layout
exactly.** `linkedDoorOffset` lands at PC 0x90 and is emitted as an int
id (not zeroed). The C3r Bunker2 residual was fixed by **D126** (objective
sub-record `->next` growth desyncing the walk so `linkedDoorOffset+arg2`
resolved to the wrong record) — not a DOOR-tail cursor bug. Bunker2
currently PASSES, consistent with this.

**VEHICHLE (39)** — `VehichleRecord`, PC **208 B** (`sizepropdef` PORT
`return 52`). ailist@w32 = int id (`OBJ_ID_WORDS`, D123);
aioffset|aireturnlist@w33 = `_hh_word`; path@w41, Sound@w43 = ptr.
Cursor: id 144->152, hh 152->156, 7xf32 156->184, path 184(8-al)->192,
nextstep 192->196, Sound 200(8-al, 4 pad)->208. MATCH (208).

**AIRCRAFT (40)** — `AircraftRecord`, PC **208 B**. Same shape; path@w43,
Sound@w44. Cursor lands path 192(8-al)->200, Sound 200->208. MATCH.

**AUTOGUN (13)** — `AutogunRecord`, PC **248 B** (`sizepropdef` PORT
`return 62`). 17 scalar tail words 144->212, then unkC4*/unkC8*/beam*
(w49/50/51) 216(8-al, 4 pad)->224->232->240, is_active 240->244, unkD4
244->248. MATCH (`PROPDEF_PC_BYTES[13]=248`).

**AMMO / MultiAmmoCrate (20)** — `MultiAmmoCrateRecord`, PC **200 B**.
Tail = `slots[13]` of `{u16 modelnum; u16 quantity}` -> 13 `_hh_word`
writes 144->196, pad->200. `AMMOTYPE_GLOBAL_MAX == 13` matches
`PROPDEF_N64_WORDS[20]=45` (32+13). MATCH (`sizepropdef` PORT `return 50`).

**TANK (45)** — `TankRecord`, converter **248 B** (`_emit_object_prefix`
+ collision*@w32 + 23 `_bswap32` words). The `TankRecord` struct as
declared in `bondtypes.h` is **53 words**, but `PROPDEF_N64_WORDS[45]=56`
(getools/ROM) — the struct's `//s32 unk88..` comments imply a 3-word gap
the C declaration omits, so the *field offsets* past `rect` are
unconfirmed. However: the only pointer is `collision`@w32 (correctly
widened + kept in slot), the stride is self-consistent with `sizepropdef`
PORT (`return 62` = 248), and TANK appears only twice (Depot x1, Runway
x1). Walk integrity is intact; only tank-field *semantics* (non-crash)
could be off. **Confidence medium**; acceptable until a tank level is
played.

**TINTED_GLASS (47)** — `TintedGlassRecord`, PC **168 B**. 5 `s32` tail
words 144->164, pad->168. No pointers. MATCH (`sizepropdef` PORT
`return 42`).

**Objective sub-records 30/32/33/35** (D126) — re-confirmed: each is
`{header; s32 payload x N; T *next}`; converter emits header + N
`_bswap32` words + 8 zero bytes for `next` at offset 16, total 24 B.
`criteria_deposit` (33) has N=3 (5 N64 words); the rest N=2 (4 words).
MATCH (`sizepropdef` PORT `return 6`).

**OBJECTIVE_START (23)** — `MissionObjectiveRecord` ends in
`WatchMenuObjectiveTextRecord *nextentry` @0x10, so PC `sizeof` is 24,
but `PROPDEF_N64_WORDS[23]=4` (nextentry is not serialized) and
`sizepropdef` PORT `return 4` (16 B stride). This is **safe**: nothing
writes `MissionObjectiveRecord.nextentry` — the briefing-text linked
list (`ptr_last_briefing_setup_entry_type23`, `objective.c:54`,
`objective_status.c:88`) is a chain of `struct watchMenuObjectiveText`
(type-35 records) via *their* `nextentry`, despite the misleading global
name. The type-23 record is only ever *read* (`get_status_of_objective`
walks it as `ObjRefID`@4 / `TextID`@8 / `MinDificulty`@0xc). Uniform
16-byte stride, no pointer store -> no D126-class overflow.

### Divergence found — types 14 / 19 / 38 / 44 (union-with-pointer index slots)

`LinkRecord` (14), `SwitchRecord`=`LinkRecord` (19), `LockDoorRecord`
(38), `SafeObjectRecord` (44) each declare their index fields as a
**union with a pointer**:

    typedef struct LinkRecord {
        inherits PropDefHeaderRecord;                       // 0, 4 B
        union { struct PropRecord *first;  s32 Index1; };   // PC: off 8 (8-al, 8 B)
        union { struct PropRecord *second; s32 Index2; };   // PC: off 16
        struct LinkRecord *next;                            // PC: off 24
    } LinkRecord;                                           // PC sizeof = 32

On PC the union is 8 bytes / 8-aligned, so `Index1` sits at **byte 8**
(4 B pad at byte 4), `Index2` at **byte 16**, `next` at **byte 24**.
`LockDoorRecord` is the same shape (sizeof 32). `SafeObjectRecord` has
three such unions (`item`/`safe`/`door`) + `next` -> `Index1`@8,
`Index2`@16, `Index3`@24, `next`@32, sizeof **40**.

The converter's handler for `(14, 19, 38, 44)` does:

    out[0:4] = _hdr_word(src[so:so + 4])
    for i in range(1, n64w):
        out[4 * i:4 * i + 4] = _bswap32(src[so + 4 * i: so + 4 * i + 4])

i.e. it lays the N64 index words at PC bytes **4, 8, 12** — the N64
tight-4-byte packing. Result on PC:

| read | converter put | compiler expects | effect |
|---|---|---|---|
| `pdef->Index1` (byte 8) | N64 word 2 = **Index2** | Index1 | wrong id |
| `pdef->Index2` (byte 16) | (never written) = **0** | Index2 | always 0 |

`PROPDEF_PC_BYTES` also under-sizes 14 & 19 & 44 (24 vs real 32 / 32 / 40;
38 is coincidentally 32).

**Runtime effect (`prop.c` `proplvreset2` walk):** `PROPDEF_SWITCH`
reads `index1 = pdef_switch->Index1` / `index2 = pdef_switch->Index2`,
resolves `doorA`/`doorB`, and only if
`doorA && doorA->prop && doorB && doorB->type==PROPDEF_DOOR && doorB->prop`
writes `pdef_switch->first/second` + calls `initSetLevelLoadPropSwitch`
(which does `arg0->next = ...`). Because the indices read wrong, that
guard **fails**, so the pointer stores (and the would-be `->next`
overflow past the under-sized record) never happen. Same for
`PROPDEF_LOCK_DOOR` (`pdef_lock_door->door/lock` + `->next`),
`PROPDEF_SAFE_ITEM` (`pdef_safe->item/safe/door`), and `PROPDEF_LINK`
(`propweaponSetDual`, also `guna && gunb` guarded).

**Net: non-crashing but silently broken** — switch-activated doors, dual
(left+right) weapon pickups, padlocked doors, and safe/safe-item links do
not initialise on PC. This explains why Streets (`UsetuptraZ`, 20x
LOCK_DOOR) and Aztec/Dam (SWITCH) still PASS the load+no-crash sweep.
Below crash work in priority, but a WS6 objective-playthrough blocker.

Affected levels: LINK — Caverns, sevb. SWITCH — Archives, Aztec, Dam.
LOCK_DOOR — Dam, sevx, sevxb, **Streets x20**. SAFE_ITEM — Archives,
Depot, sevb, sevx.

**Not previously caught:** D125's stride re-check only asserted
`sizepropdef == PROPDEF_PC_BYTES/4` (internal consistency), never
`PROPDEF_PC_BYTES == real compiler sizeof`. D122's totality note flagged
`[u16|u16]` half-swap but not the `union{ptr; s32}` index-slot case.

### APPLIED (M-20, commit pending) — ABI/layout only, `#ifdef PORT`

Fix below applied verbatim. `tools_pc/d88_propdefs.py`: `PROPDEF_PC_BYTES`
`14/19 -> 32`, `44 -> 40`; the `(14,19,38,44)` handler now emits each
`Index{k}` into the low 4 B of the 8-aligned slot at PC `8 + 8*k`
(`nidx = {14:2,19:2,38:2,44:3}`). `loadobjectmodel.c sizepropdef()` PORT
switch: `LINK/SWITCH/LOCK_DOOR -> return 8`, `SAFE_ITEM -> return 10`,
`TAG` stays `6`. Regen chain run. Verified: `-level_09` framediff 3/3
PASS; `-level_20` behaviour byte-identical to a freshly-built pre-D132
baseline (the `frame_000320` phash delta is pre-existing stale-golden /
Silo slowdown, NOT a regression — confirmed by building the baseline);
Archives (`-level_25`, SWITCH+SAFE_ITEM), Streets (`-level_29`, 20x
LOCK_DOOR), Dam (`-level_33`, SWITCH+LOCK_DOOR) all load + render +
no-crash. Sweep runs this pass were heavily flaky (level_09 itself
0-framed on a loaded machine) — pure D117/watchdog noise, no crashes.

### Original proposed fix (as written pre-apply)

`tools_pc/d88_propdefs.py` — `PROPDEF_PC_BYTES`: `14: 24 -> 32`,
`19: 24 -> 32`, `44: 24 -> 40` (38 stays 32). Replace the
`(14, 19, 38, 44)` handler:

    if type_byte in (14, 19, 38, 44):  # LINK / SWITCH / LOCK_DOOR / SAFE_ITEM
        # D132: each Index{1,2,3} field shares a union with a pointer, so on
        # PC it lives in the LOW 4 bytes of an 8B/8-aligned slot at
        # PC offset 8 + 8*(N-1); the record ends in a *next the setup walk
        # writes.  The N64 image packs the indices as tight 4-byte words;
        # emitting them there put Index1 where the compiler reads Index2.
        nidx = {14: 2, 19: 2, 38: 2, 44: 3}[type_byte]
        out[0:4] = _hdr_word(src[so:so + 4])
        for k in range(nidx):
            w = src[so + 4 * (k + 1): so + 4 * (k + 1) + 4]
            out[8 + 8 * k: 12 + 8 * k] = _bswap32(w)   # low 4B, LE; high 4B = 0
        return bytes(out)

`src/game/loadobjectmodel.c` `sizepropdef()` `#ifdef PORT` switch: move
`PROPDEF_LINK` / `PROPDEF_SWITCH` out of the `return 6` group and give
`LINK`/`SWITCH`/`LOCK_DOOR` -> `return 8` (32 B) and `SAFE_ITEM` ->
`return 10` (40 B). `PROPDEF_TAG` stays `return 6` (its `ID`/`OffsetToObj`
are plain `u16`/`s16` at 0x4, not in a pointer union; handler already
correct).

After applying: `d88_emit.py --regen` (all 21 sidecars), then the full
`d43 && d69 && d88 --regen` chain, then re-verify `-level_09` / `-level_20`
golden + a Streets/Dam/Archives load.

### Confidence per type

| type(s) | verdict | confidence |
|---|---|---|
| ObjectRecord prefix (3/5/12/17/36/42/43) | matches | high |
| 1 DOOR | matches (incl. linkedDoorOffset @0x90) | high |
| 39 VEHICHLE / 40 AIRCRAFT | matches | high |
| 13 AUTOGUN | matches | high |
| 20 AMMO/MultiAmmoCrate | matches | high |
| 47 TINTED_GLASS | matches | high |
| 30/32/33/35 objective sub-records | matches (D126) | high |
| 23 OBJECTIVE_START | 16B stride safe (nextentry never written) | high |
| 2/24/25/26/27/28/34/46 (generic) | safe under generic arm | high |
| 4 KEY / 7 MAGAZINE / 21 ARMOUR / 8 COLLECTABLE | prefix + plain tail; widely used in passing levels | medium-high |
| 45 TANK | struct offsets unconfirmed (53 vs 56 w); stride self-consistent, walk intact, non-crash | medium |
| **14 LINK / 19 SWITCH / 38 LOCK_DOOR / 44 SAFE_ITEM** | **DIVERGENT** — union index slots at wrong PC offset; silently non-functional (non-crash) | fix high confidence |

### Files touched

Docs only: this subsection + the §F index rows (D88.4 status, new D132
row). No code changed; the proposed diff above is not applied. No temp
probe scripts left.

## D185 — fresh `d69_emit.py` run yields a game-crashing `data/pccg-<region>/` (M-38)

**Status: OPEN — alpha-release blocker.** (Add to the §F index.)

**Symptom.** `python tools_pc/d69_emit.py ntsc-final` in this repo writes
`data/pccg-ntsc-final/pccg.bin` = 3346895 B (52 sidecars, 1856-B manifest),
prints `ALL CHECKS PASSED`, and `-level_09` / `-level_20` then segfault before
frame 1 (`EXCEPTION: 0xc0000005`, faulting return address is an ASCII
model-name string). The pre-migration working sidecar
(`C:/Users/james/Source/Repos/007/data/pccg-ntsc-final/pccg.bin`, 3604378 B,
2421-B manifest) drops into this tree and `-level_09` runs crash-free.

**Isolation done (M-38).**
- Not the romdata M-38 change: baseline (stashed) crashes identically.
- Not the build: same exe, only the `data/pccg-ntsc-final/` bytes swapped.
- `tools_pc/d69_emit.py` is byte-identical between the public repo and the
  pre-migration repo `007` (only a doc-path comment differs).
- The migration commit `52539d10 "Prepare for public release"` normalised the
  committed d69 inputs from CRLF to LF: `scripts/filelist.u.csv` (−812 B, one
  CR/line) and `assets/obseg/file_resource_table.inc.c`.
- **Restoring CR to both inputs did NOT restore the 3604378-B output** → line
  endings are not the (whole) cause. Some other part of the d69 input closure
  regressed in the history rewrite, or the working sidecar predates a real
  d69/​input change and 3346895 is a separate second bug.
- `d43_emit.py` is unaffected (its `find_row` tries `name`, `name[1:]`,
  `name[:-1]`, `name[1:-1]` — tolerant of a mangled basename; `d69`'s
  `fl_by_base.get(basename)` is exact).

**Impact.** `docs/building.md` tells users to run `d69_emit.py`; M-38's
`prepare-assets.py` does the same. Both currently produce a broken game on a
clean checkout. The committed `data/` dirs (carried from pre-migration) hide it.

**Next.** Diff the sidecar *name* set: pre-migration `manifest.csv` vs a fresh
run's — find which bg/stan entries d69 now drops, trace through `find_row` →
`file_resource_table.inc.c` → `filelist.u.csv`. Fix so a fresh run reproduces
3604378 B / 2421-B manifest; re-verify `-level_09`/`-20` crash-free. Do NOT
regenerate `data/pccg-ntsc-final/` before the fix (it holds the working copy).

### D185 — RESOLVED (M-38): not a d69 regression, a missing d88 step

`d88_emit.py` **appends** the 21 `Usetup*Z` per-level stage-setup files
(object placement, AI opcode streams, pads/nav) to the same
`data/pccg-<region>/pccg.bin` + `manifest.csv` that `d69_emit.py` writes
(52 bg/stan rows → 74 after d88). The M-38 `prepare-assets.py` first shipped
ran only d43 + d69, so every level loaded with no setup data → instant
segfault. The "working" pre-migration sidecar was simply a complete
d43+d69+d88 output; nothing regressed in d69 or its inputs.

Fix (commit `4491db3f`): `prepare-assets.py` runs d43 → d69 → `d88 --regen`;
`bundle-win.sh` vendors `d88_emit.py` + its only local import
`d88_propdefs.py`. Verified: fresh 3-script run from an assembled bundle is
byte-identical to the known-good `pccg.bin` (3604378 B) / `pcmodels.bin`;
`-level_09` + `-level_20` boot crash-free. The CRLF→LF observation on
`filelist.u.csv` / `file_resource_table.inc.c` is real but a red herring for
this crash (d69 tolerates it; output unchanged).

## D200 — `-level_09` segfault: u32 reverb delay indices negate into +4GB offsets on 64-bit (fixed)

**Symptom.** `-level_09` reliably segfaulted within seconds of audio starting, in
`_filterBuffer` (`src/libultrare/audio/reverb.c`) — `d->lp` (a pointer into the AFX
delay array) resolved to garbage like `0x664d59be`.

**Root cause.** `ALDelay.input` / `ALDelay.output` are `u32`, and reverb.c computes
back-references into the delay ring as `&r->input[-d->output]` (and variants). On
N64, `ptrdiff_t` is s32, so `-160` (the u32 constant `0xFFFFFF60`) wraps to −160 and
the pointer arithmetic is correct. On PC x86-64, `ptrdiff_t` is s64: the u32 negation
**zero-extends** to +4294967136 samples, so `out_ptr` flies forward ~8GB — verified
exactly: `0x706d30a0 (r->input) + 0xFFFFFF60·2 = 0x2_706D2F60`, the observed wild
address. The subsequent `aSaveBuffer`/`aLoadBuffer` then write 13–52KB of wild data,
corrupting the delay array, `r->base`, and adjacent heap blocks — a self-propagating
loop that grows each frame until it hits unmapped memory.

**Fix (ABI/layout only, D3x class).** Five sites in `src/libultrare/audio/reverb.c`
cast to `(s32)` before negation so the offset sign-extends: `alFxPull` (`in_ptr`,
`out_ptr`), `_loadOutputBuffer` ×2 (`out_ptr`), plus the ramalign line, which uses
`(s64)out_ptr & 0x7`. Perfect Dark ground truth (`src/lib/naudio/n_reverb.c`) does
exactly this: `(s32)-d->input`, `(s32)-d->output`, `(s32)-(d->output - d->rsdelta)`,
`(intptr_t)out_ptr & 0x7`. No game logic changed; behaviour on N64 is identical.

**Verification.** `-level_09` now runs 120 s+ crash-free (previously segfaulted in
seconds); the delay pointers advance correctly (`out_ptr = input − 320B`, +0x140/frame).
A temporary DRAM guard window over the FX region, registered with the software mixer
(`port/src/mixer.c`), confirmed zero out-of-bounds writes after the fix. All
instrumentation removed; only the five casts remain in the diff.

**Sibling-risk audit.** Grep of `src/libultra/audio/*.c` + `src/libultrare/audio/*.c`
for negative pointer indexing: reverb.c is the only file with the pattern (its one
positive index, `&r->input[d->output]`, is safe — positive u32 zero-extension is
correct).

**Lineage.** D3x pointer-width reconciliation; same family as D177 (widened-pointer
high half) and D191 (32-bit field read as 64-bit).

## D202 — Phase-3 audio: silenced PPK plays "slap" not gunshot, PLUS runtime mixer corruption (piling up / glitching / eventual silence) (M-52; reclassified M-59; REOPENED M-60 on reference clip; M-61/M-62 closed the data/index/decode chain and recommended close; M-63 REOPENED again; M-65 PARTIALLY resolved — voice-leak cascade root-caused + guarded, `sndCreatePostEvent` un-stubbed; **M-66 ROOT CAUSE ESTABLISHED for the audible stuck door loop — sound 203 behaving exactly as ROM + ground-truth code specify: faithful N64 behaviour / original quirk; M-66b DISPOSITION C IMPLEMENTED + MEASURED — port-side expiration fades ownerless infinite loops out ~2.5 s after start, awaiting user by-ear pass**)

**Symptom (M-52 playtest, first listening pass on the Phase-3 software mixer,
post-D200).** In-level (Bunker1, Dam — not menu/intro, which is music-only and
correct): (a) weapon fire consistently plays the WRONG but always-the-same
clip per weapon (e.g. Bond's silenced PPK plays what sounds like a melee/punch
hit, every single shot — fully deterministic, not intermittent garble); (b)
level music never plays — a looping SFX-like sound plays in its place from
level start.

**Ruled out, with hard evidence (do not re-investigate these):**
- `soundIndex -> ALSound*` resolution (`snd.c sndPlaySfx`, `soundBank->instArray[0]->soundArray[soundIndex]`):
  traced live with a `GE_AUDIOTRACE` probe against real gameplay (`-level_09`,
  repeated PPK fire). `soundIndex=46` (confirmed via `bondconstants.h`'s SFX
  name table = `46_GUN_SILPPK_A_SFX`) resolves to the same `ALSound*`/`ALWaveTable*`
  every time; 13+ distinct indices captured across one session, all distinct,
  all valid.
- D37's bank re-layout (`port/src/romdata.c romdataFixupAudioBank`): preserves
  `soundArray`/`instArray` order; not a re-ordering bug.
- The `ALWaveTable.base`/`.book` pointer chain: traced all the way to
  `alLoadParam(AL_FILTER_SET_WAVETABLE, ...)` (`src/libultra/audio/load.c`) —
  the exact same `table`/`base`/`book` addresses `sndPlaySfx` resolved are the
  ones wired into the physical decode filter. No swap between resolution and
  voice setup.
- `ALStartParamAlt` (40B on x86-64) vs the generic `ALParam` pool slot (32B):
  already fixed pre-M-52 (D54, `synthesizer.c:127-140`, `#ifdef PORT` sizes
  each slot for the largest variant). Re-confirmed still in place, not a
  regression.
- Per-voice synth processing (`synthesizer.c`'s client-list handler loop) is
  strictly sequential — one client's full setup+decode completes before the
  next starts. Ruled out global-mixer-scratch (`sCtx`/`sAdpcmTable`/`sVol` in
  `port/src/mixer.c`) cross-talk between concurrently active voices.
- `ALSndpEvent` (snd.c-local union, force-cast to `ALEvent*` through
  `alEvtqPostEvent`/`alEvtqNextEvent`, which copy `sizeof(ALEvent)` bytes)
  vs `ALEvent` (libaudio.h): measured directly with a throwaway compile —
  both are exactly 32 bytes on x86-64. Not a truncating-copy bug.
- `sndCountAllocList`'s `(ALEventQueue *)&D_800243E4` cast (a real, compiler-flagged
  `-Warray-bounds` OOB read — `D_800243E4_s` is 24B, aliased as 40B `ALEventQueue`):
  confirmed genuine but functionally inert — its output (`numFree`/`numAlloc`)
  is computed and immediately discarded in `sndHandleEvent`, never used.

**Not yet ruled out — pick up here:**
1. **`aADPCMdecImpl`'s decode math** (`port/src/mixer.c`, this session's new
   Phase-3 code, never verified against real N64 audio output). Structurally
   matches the standard VADPCM order-2 predictor algorithm on inspection
   (9 input bytes -> 16 output samples, intra-frame reverb-of-already-decoded-
   samples loop) but no reference trace/audio diff exists yet.
2. **The raw PCM bytes embedded in the exe at the resolved ROM offset being
   wrong at the source** (asset-embedding/segment-placement issue, unrelated
   to anything traced above) — would need a byte-level compare of the
   PC-embedded `_sfxtblSegmentRomStart` segment against the source `.z64`.

**Key diagnostic fact:** the wrong sample is 100% deterministic per weapon
(same wrong clip every shot, confirmed by user) — this argues against a
stateful/interleaving bug (already the least-likely explanation given the
sequential-processing proof above) and toward either (1) or (2) above, both
of which would consistently mis-decode/mis-fetch the same bytes every time
for the same `ALWaveTable`.

**Diagnostic tooling added this session (still in the tree, `#ifdef`/env-gated,
harmless if left in, should be stripped once root-caused):**
- `src/snd.c` `sndPlaySfx`: `GE_AUDIOTRACE=1` env var → appends resolved
  `soundIndex`/`sound`/`wavetable`/`base`/`len`/`type`/`flags`/`book` to
  `audiotrace.log` (cwd-relative, unbuffered).
- `src/libultra/audio/load.c` `alLoadParam` `AL_FILTER_SET_WAVETABLE`:
  same env var → appends the filter/table/base/len/book actually wired into
  a voice to `audiotrace_wire.log`.

**Also noticed, not yet a confirmed bug:** `alLoadParam`'s
`a->memin = (s32) a->table->base;` (`load.c:382,438`) truncates a real 64-bit
`u8 *` to `s32` — currently harmless only because the ROM is deliberately
mapped at a low, sub-2GB virtual address (`0x10000000`, 12MB window per the
boot log), so no real address in range overflows 32 bits. Fragile; worth a
`(s64)`/`uintptr_t` widening pass regardless of whether it's D202's cause.

**M-53 update — both of the two remaining suspects now independently checked,
both look clean (static analysis only, no listening pass possible this
session):**

1. **`aADPCMdecImpl` decode math (`port/src/mixer.c:148-192`) — compared
   line-by-line against the local Perfect Dark PC port's scalar reference
   (`pd_port/port/src/mixer.c:333-352`, the `#else` non-SIMD path).**
   Identical: same nibble unpack (`(((*in >> 4) << 28) >> 28) << shift` /
   same for the low nibble), same `prev1`/`prev2` history read from `out[-1]`/
   `out[-2]`, same accumulator (`tbl[0][j]*prev2 + tbl[1][j]*prev1 +
   (ins[j]<<11)` plus the `k<j` intra-frame correction term), same `>>11`
   scale-down + clamp, same 16-sample/32-byte state carry via
   `memcpy(state, out-16, ...)`. No discrepancy found. This was PD's
   ground-truth non-vectorized fallback, i.e. the reference the SIMD paths
   are themselves checked against — a strong match.
2. **`_sfxtblSegmentRomStart`'s ROM offset (`0x102F19A0`,
   `port/src/romassets_u.s:2948`) — cross-validated three independent ways,
   all agree:**
   - `scripts/filelist.u.csv` row 28: `3086752,797360,assets/music/sfx.tbl` →
     `0x10000000 + 3086752 = 0x102F19A0` exactly.
   - Row 27 (`sfx.ctl`, `3063264,23488`) is contiguous with row 28:
     `3063264 + 23488 = 3086752` — matches `music.c:686`'s assumption
     (`size = &_sfxtblSegmentRomStart - &_sfxctlSegmentRomStart`, i.e. ctl
     immediately precedes tbl in ROM with no gap).
   - `ge007.ld:164-165` (N64 ground truth, untouched) places
     `sfx.ctl.o (.data)` immediately followed by `sfx.tbl.o (.data)` in the
     `musicfiles` segment — same adjacency, independently, from the linker
     script rather than the CSV scanner.
   - Read the real bytes from `data/ge007.ntsc-final.z64` at both offsets
     (this session, ad hoc Python): `sfx.ctl` decodes as a plausible
     `ALBankFile` (`bankCount=1`, one bank offset `0x5ba0` inside the 23488B
     ctl blob); no structural red flag. Byte-level content correctness
     beyond that needs a decode+listen pass, not available in this session.

   Net: the segment-offset math is corroborated by 3 independent sources
   (scanner CSV, N64 linker script, raw ROM read) and is very unlikely to be
   a simple wrong-base bug. Downgrading this suspect's priority.

**Also re-confirmed clean, not previously called out explicitly:** the ADPCM
book coefficients (`ALADPCMBook`, referenced off `ALWaveTable.waveInfo.
adpcmWave.book`) go through `afFixupBook()` (`port/src/romdata.c:704-735`),
which correctly `afRd16`/`afWr16` byte-swaps every predictor coefficient
during the D37 bank re-layout — not just the offset/pointer fields. And
`_bnkfPatchWaveTable()`'s `w->base += table;` (`bnkf.c:140`) is `uintptr_t`
end-to-end post-D201, so no truncation there either.

**Net effect: the M-52 write-up's two recommended suspects are now the
LEAST likely explanations, not the most likely.** The bug is probably
further downstream, in a stage neither this nor the M-52 session traced:
**the resample/pitch pipeline** (`aResampleImpl` / whatever computes a
voice's playback rate from `ALKeyMap`/`unityPitch` — `synallocvoice.c`,
`seqplayer.c`, `synport.c`/equivalent — not yet audited this session). A
wrong pitch/rate calculation would (a) make a correctly-decoded SFX sample
sound like a completely different sound if pitch-shifted far enough to be
mistaken for one, matching "PPK sounds like a melee hit" better than a
literal wrong-sample-content theory would, since sample *content* tracing
(D202 original session) already came back clean; and (b) explain "music
never plays, an SFX-like loop plays instead" if the same mis-pitched/
mis-selected-voice bug applies to sequence-player note events too, not just
raw `sndPlaySfx()` calls. **Recommended next step:** trace one PPK-fire note
event's computed pitch/rate value (another `GE_AUDIOTRACE`-style probe, this
time in the resample-rate / `ALKeyMap` lookup path) against what the N64
build would compute for the same key/velocity, rather than re-auditing the
decode or offset math further. Not yet attempted — no build/run done this
session (static-analysis-only pass); still needs the real listening-pass
sign-off per the standing M-52 note.

**M-53 cont. (build+run, same session) — user rebuilt off the M-53 doc-only
commit and confirmed no behavior change (expected — no code changed yet);
then supplied two new concrete data points that reframe symptom (b) and
partially (a):**

1. **"the looping sound [that plays instead of music] is actually a door
   sound in the attract screen getting looped and continuing to play."**
   This is not a wrong-sample or wrong-pitch bug — it's confirmed to
   literally be a real, correctly-identified SFX (a door slide-loop) that
   fails to stop. Traced the mechanism live with `GE_AUDIOTRACE`
   (`-level_09`, scripted door-interact input): a single player action
   fires a **chained** `sndPlaySfx` sequence — soundIndex 202
   (`METAL_SLIDE_OPEN_SFX`) → 204 (`METAL_SLIDE_LOOP_SFX`) → 203
   (`METAL_SLIDE_CLOSE_SFX`) — via the `do…while` chain in `sndPlaySfx`
   (`snd.c:927-928`, each `ALSound`'s `keyMap->velocityMin`/`keyMin` fields
   double as a "next soundIndex" link; this chain mechanic is stock N64
   design, not a port bug).
   - The game-code side that's supposed to **stop** a stuck door loop —
     `door7F053B10()` (`propobj.c:12835`, marked `//#MATCH`, i.e. decompiled
     byte-exact to the N64 binary, so its logic is ground truth and NOT a
     candidate for the bug itself) — calls `sndDeactivate(door->
     openSoundState)` whenever the door's *open* or *close* sound is still
     playing. `door->openSoundState` is already a real `ALSoundState *`
     in `bondtypes.h:3042` (not a narrowed/truncated field — checked, this
     is NOT another D3x pointer-width bug).
   - So the bug is downstream of this game-logic call: either (a)
     `sndDeactivate`'s posted `AL_SNDP_DEACTIVATE_EVT` (delta 0) isn't
     being serviced/dequeued reliably by the PC event-queue/mixer tick, or
     (b) (more likely given the user specifically saw this originate **in
     the attract screen** and persist **past it, into gameplay**) whatever
     N64 does to reset/clear all playing SFX at an attract-demo → real-game
     mode transition isn't happening equivalently on PC, orphaning the
     loop's voice with no `DoorRecord` left alive to ever call
     `sndDeactivate` on it again. **Not yet located** — haven't found the
     mode-transition audio-reset call (candidate: `sndDeactivateAllSfxByFlag`
     family, `snd.c:996`, or an equivalent full-stop at level-load).
2. **"ammo pick up is one of the knife sounds."** Checked against
   `bondconstants.h`: `PICKUP_AMMO_SFX`=234, `PICKUP_KNIFE_SFX`=233 — adjacent
   indices. Combined with the M-53 PPK report (`GUN_SILPPK_A_SFX`=46 sounding
   like `PUNCH1_SFX`=47, also adjacent but in the OPPOSITE direction), this
   does **not** fit a simple constant index-shift theory (would need to be
   the same sign both times) — leaves either (a) two unrelated single-entry
   bugs, or (b) the "wrong sound" being a badly-mispitched *correct* sample
   that a listener reasonably mistakes for a neighboring, timbrally-similar
   entry (a hard door clank at the wrong pitch could pass for a knife sound;
   a gunshot pitched down for a punch). Not resolved this session — would
   need an actual decoded-audio comparison (dump the decoded PCM for a
   known-bad play and listen/compare band energy against the two
   candidates), which the environment can produce (`GE_AUDIOTRACE` gives the
   exact `base`/`len`/`book` addresses) but wasn't attempted here.

**Still OPEN. Build/run environment confirmed available this session**
(`./build-pc.sh ntsc-final` after `export PATH=/c/msys64/mingw64/bin:$PATH`,
`GE_INPUTSCRIPT` with `"A"`=`GE_CONT_A`, `"Z"`=`GE_CONT_G` per
`port/src/input.c:311` — note **"Z" triggers the door/action interact in
this control scheme in `-level_09`, not weapon fire**; hadn't yet found the
right button/level combo to headless-repro an actual gunshot before running
out of session time). **Next step:** locate and instrument the attract→game
(or menu→level) audio-reset call to confirm/deny the orphaned-voice theory
for symptom (b); separately, dump+listen-compare decoded PCM for one
mis-sounding SFX to settle symptom (a)/(2)'s index-vs-pitch question.

**Follow-up (same session) — checked the level-transition "deactivate all"
sweep, found it's flag-filtered, which narrows the orphaned-voice theory
to a specific, already-catalogued bug family.** `lv.c:1443/1473` calls
`sndDeactivateAllSfxByFlag_1()` at level transitions → `flag=1`
(`SOUND_FLAG_FINAL_IN_SEQUENCE` only) → `sndDeactivateAllSfxByFlag`
(`snd.c:996`) only touches sounds whose `unk3e` has **every** bit in `flag`
set (`(item->unk3e & flag) == flag`). A door's loop-tail sound is very
unlikely to carry `FINAL_IN_SEQUENCE` (it's a middle link in the
open→loop→close chain, not the end), so this sweep would **not** catch an
orphaned door loop even on real N64 — meaning N64 must instead rely on the
attract-mode demo's scripted input naturally executing its "close door"
press before the demo ends, letting `door7F053B10()` clean up the loop
normally. **This reframes the theory:** if the attract-mode demo's
canned input is driven by a fixed *frame count* but the PC build's
sim-time-per-frame differs from N64's (the same `g_GlobalTimerDelta`/
`g_ClockTimer` wall-clock-vs-sim-time family already root-cause-pending
for **D193**, the AI-locomotion-too-slow bug), the demo could get cut off
before its scripted door-close input fires on PC even though it always
does on N64 — orphaning the loop's voice with no code path left to stop
it. **If true, D202(b) (the stuck-loop symptom) may not be a distinct
audio bug at all — it may be a second visible symptom of D193's root
cause once that's found**, not something to fix in the audio code. Worth
confirming/ruling out before spending more time in `snd.c`/`propobj.c`.

**M-54 update — checked the D193-linkage theory's timing mechanics
directly; it's weaker than M-53 framed it, and got two direct
experimental data points instead.**

1. Read `frametiming.c`/`libultra.c`'s timing chain end to end.
   **The ramrom-demo path (`iterate_ramrom_entries_handle_camera_out`,
   `ramromreplay.c:355`) does NOT go through `waitForNextFrame()`** — it
   calls `updateFrameCounters(ramrom_blkbuf_2->speedframes)` directly
   (`ramromreplay.c:400`), feeding the **recorded** N64 speedframes value,
   not a value derived from PC wall-clock elapsed time. The D155
   catch-up clamp (`FRAMETIMING_PORT_MAX_CATCHUP`) lives inside
   `waitForNextFrame()` and is irrelevant here. The call site
   (`boss.c:502-533`) gates *when* this fires on real elapsed ticks
   (`mainTickElapsed >= MAIN_LOOP_TICK_INTERVAL`) but that gate
   self-throttles correctly even if `gfxFrameMsgQ`'s retrace messages
   back up during a stall (checked: `osCreateMesgQueue(&gfxFrameMsgQ,
   ..., 32)`, `init.c:236` — 32-deep, so a stall *can* queue up dozens
   of retrace messages) — the very first drained message after a stall
   recomputes a big `mainTickElapsed` and fires exactly once, which
   immediately refreshes `copy_of_osgetcount_value_1`, so every
   subsequent backlogged message in the same burst sees a near-zero
   elapsed and does nothing. **No runaway/duplicate advance mechanism
   found.** This weakens (doesn't fully rule out) the D193-linkage
   theory: the demo's sim-time progression looks insulated from PC
   real-time hiccups by design, not just by luck. The D193 tie-in is
   still plausible in principle (if the *count* of real retrace ticks
   over the whole demo differs from N64's, e.g. via `osGetCount()`'s
   scaling formula, D117/D134) but the specific "backlog causes runaway
   catch-up" mechanism I'd hoped to point to isn't there. Downgrading
   this from "leading theory" to "one of several still-open".
2. Extended the existing `GE_AUDIOTRACE` probes: `sndPlaySfx` now also
   logs the `newState` pointer per `soundIndex` link (not just the
   `ALSound*`), and `sndDeactivate` now logs every call with its `state`
   arg (`snd.c`, both `#ifdef PORT`/`getenv("GE_AUDIOTRACE")`-gated,
   still uncommitted). Ran two headless captures
   (`GE_AUDIOTRACE=1 GE_D87=1 ./build-pc/ge007.x86_64.exe`, no level arg,
   no input, 100s then 240s, idling at the front end so the >=30s idle
   timer — `front.c:2493`, `MENU_TIMER >= 1801` at NTSC 60fps — kicks
   attract mode on). Findings:
   - **`sndDeactivate` DOES fire and DOES work** when called: trace
     shows `sndPlaySfx: soundIndex=260 -> newState=...706f4e18` followed
     later by `sndDeactivate: state=...706f4e18 (non-null)`, and that
     same `ALSoundState*` slot gets legitimately reused for a new
     `soundIndex` right after. This rules out "the event-queue doesn't
     service `AL_SNDP_DEACTIVATE_EVT` reliably" (branch (a) from the
     M-53 writeup) as a *general* mechanism — deactivation is not
     structurally broken.
   - Both captures only ever logged front-end/menu-click `soundIndex`
     values (258, 260, 111, 232, 109) — despite `GE_D87` confirming
     `iterate_ramrom_entries_handle_camera_out`/`ramrom_replay_handler`
     WERE firing repeatedly (159 D87 log lines in the 240s run, i.e. the
     attract-mode camera flythrough demo was genuinely playing), **no
     door-chain soundIndex (202/203/204) appeared in either capture.**
     Root cause of the miss, found by reading `ramromreplay.c:631-634`:
     which of the 14 `ramrom_table[]` demo files plays is
     `randomGetNext() % i` — **randomly selected per attract cycle**, not
     a fixed sequence. A short capture has no guarantee of landing on a
     demo that triggers a door open/loop at all (most of the 14 clips
     — Dam/Facility/Runway/Silo/Frigate/Train camera paths — may not
     pass a metal sliding door in-frame). Neither capture is long enough
     to be conclusive either way for symptom (b); this needs either (a)
     a much longer unattended capture (several attract cycles, each
     clip's `totaltime_ms` likely 30-90s, so plausibly 10+ min for good
     odds of hitting a door-bearing clip), or (b) a temporary
     `PORT`-gated env-var override pinning `ramrom_table` selection to a
     specific known door-bearing index for fast, deterministic repro
     (not yet added).
   - Separately confirmed the `romdataFixupMusicSeqTable: seqCount 63
     exceeds blob capacity 1` `[ERROR]` log line seen at boot is a
     **known, benign, single-fire false alarm** — it's the deliberate
     D35 header-only peek (`music.c:731-736`, buffer intentionally sized
     16 bytes to decode just the `seqCount` field before the real,
     correctly-sized allocation+decode at `music.c:738-742`) — confirmed
     it fires exactly once per run and the real table load that follows
     doesn't. Not a new lead; noting it here only so a future session
     doesn't re-flag it as one.

**M-54 cont. — got the live repro, and it's NOT a door.** Re-ran headless
with a `GE_INPUTSCRIPT` pressing A every second for the first ~20s (to
clear the Rare/Nintendo/legal/cast-intro screens fast) then idling, 280s
total. Landed on the `ramrom_Train` attract demo (confirmed: `GE_D87`
shows `iterate_ramrom_entries_handle_camera_out`/`ramrom_replay_handler`
actively firing — real attract-mode canned-input playback, not the boot
cinematic, which is a separate unrelated state machine in `front.c`).
`audiotrace.log` (full, unedited):

```
soundIndex=64 -> newState=706f4db0 flags=2   (TRAIN_GO_SFX, first voice)
soundIndex=64 -> newState=706f4f50 flags=2   (TRAIN_GO_SFX, second voice)
...
sndDeactivate: state=706f4e80 (non-null)
sndDeactivate: state=706f4ee8 (non-null)
```

`soundIndex=64` = `TRAIN_GO_SFX` (`bondconstants.h`), `flags=2` =
`SOUND_FLAG_LOOPED` — the **only** looped sound requested in the entire
run. **Neither `706f4db0` nor `706f4f50` ever appears in a
`sndDeactivate` line, anywhere in the log** — confirmed by grepping the
full file for both pointers. Meanwhile two *other*, non-looped states
(`706f4e80`, `706f4ee8`) DO get swept and deactivated, at exactly the
point the demo cycle ends and the sequence rolls back to the boot logos
(`RARELOGO_SFX`/`RARELOGO_FAINT_SFX`, indices 258/260, reappear
immediately after). **This is a direct trace-level confirmation of the
M-53 "Follow-up" theory**: whatever cleanup sweep runs at the
demo-cycle/mode-transition boundary deactivates some voices but
specifically skips the looped one(s) — consistent with
`sndDeactivateAllSfxByFlag`'s `FINAL_IN_SEQUENCE`-only flag filter
(`snd.c:996`, see M-53 above) never matching a genuinely-looped ambience
voice. Reframes the bug from "door-specific" to **"any `SOUND_FLAG_
LOOPED` voice started during a `ramrom` attract demo has no code path
that ever stops it once the demo cycle moves on"** — the door report and
this train report are the same bug, different SFX.

**New candidate mechanism for why N64 doesn't audibly show this**:
`src/libultra/audio/synallocvoice.c`'s `_allocatePVoice`/`alSynAllocVoice`
implement real N64 hardware **voice-stealing** — a small fixed pool of
physical `PVoice`s (`ALSynConfig.maxPVoices`, set from
`MUSIC_SYN_CONFIG_MAX_P_VOICES` in `music.c:772`, unchanged decompiled
value) where a new voice request silently steals and ramps out the
lowest-priority already-playing physical voice once the pool is full
(lines 100-131). On real hardware this would eventually silence an
orphaned looped voice anyway, once enough *other* sounds compete for
the same small voice pool — even though its `ALSoundState` bookkeeping
never gets told to stop. **Not yet checked**: whether the PC port's
mixer (`port/src/mixer.c`) and the synth's free/alloc/lame-list
machinery actually enforce the same `maxPVoices` cap end-to-end, or
whether something in the Phase-3 PC mixer path effectively gives every
voice a free physical channel (never triggers stealing), which would
make an orphaned loop audible forever on PC even in scenes where N64
would eventually have silently stolen it away. This is the most
promising next-step lead — cheaper to check (audit `mixer.c`'s voice
dispatch / count) than adding more ramrom-timing probes.

**Still OPEN.** Next session, in priority order: (1) audit whether
`port/src/mixer.c` honors `maxPVoices`/voice-stealing identically to
`synallocvoice.c`'s original algorithm — if it doesn't, that's the fix,
and it would explain both the door and train loop reports without
touching `snd.c`/`propobj.c`/`front.c` game logic at all; (2) if voice
stealing checks out fine, look at whether N64 has some OTHER explicit
stop call for these ambience loops that isn't in the `Dxx`-traced code
paths yet (e.g. a `lvlManageMpGame`-adjacent full-audio-reset at
attract-cycle boundaries not yet located). Symptom (a)/(2)
(wrong-sample-per-weapon) still untouched — still needs the
decoded-PCM listen-compare. The `GE_AUDIOTRACE` trace additions to
`snd.c` (now covering `sndPlaySfx`'s `newState=`/`flags=` and every
`sndDeactivate` call) are left in place, uncommitted, for reuse.

**M-55 — priority-(1) voice-stealing audit done; result is negative, and
narrows the bug further.** Static/read-only session (local Qwen `triage`
dispatch for the grep/citation legwork, spot-checked line-for-line
against the real files before trusting it).

- **`port/src/mixer.c` doesn't need to honor `maxPVoices` — it never
  participates in voice allocation at all.** Grepped it for
  `maxPVoices`/`pAllocList`/`pFreeList`/`pLameList`/`numPVoices`: zero
  hits. `mixer.c` (~436 lines) only implements the low-level `aXxx`
  RSP-acmd DSP commands (envmixer/resample/etc `Impl` functions) that
  `synallocvoice.c`'s already-allocated `PVoice`s get pointed at — the
  lame-list/free-list/steal-lowest-priority bookkeeping (D202 M-54's
  suspect) lives entirely in **unmodified ground-truth**
  `src/libultra/audio/synallocvoice.c` (`alSynAllocVoice`/
  `_allocatePVoice`, lines 27-131, byte-identical to decomp) plus
  `src/libultra/audio/synthesizer.c`'s `alSynNew` (`numPVoices =
  c->maxPVoices`, `music.c:772`'s `MUSIC_SYN_CONFIG_MAX_P_VOICES`,
  unchanged). **The M-54 "maybe the PC mixer gives every voice a free
  channel" theory is ruled out — there's no PC-specific voice-pool code
  to have that bug in.** Since this pool logic is unmodified decomp
  code, its steal behaviour on PC is provably identical to N64's.
- **Priority (2) has an answer too, and it's negative in a useful way.**
  Traced whether the `ALSynth` driver (or anything else audio-wide) gets
  reset/reinitialized at the attract-demo → boot-logo transition:
  `alInit` (`src/libultra/audio/sl.c:28-33`) is a **hard singleton**
  (`if (!alGlobals) { alSynNew(...); }`) called exactly once, transitively
  from `bossEntry()` → `musicSeqPlayerInit()` (`src/boss.c:312`,
  `src/music.c:648`) → `amCreateAudioManager()` (`src/music.c:781`) →
  `alInit()` (`src/audi.c:374`/`378`) → `alSynNew()`. The only teardown,
  `alClose()` (`src/audi.c:481`), fires solely on audio-thread
  `MAIN_QUIT_MESSAGE` (process exit), not at a demo boundary. The actual
  attract-demo stop path, `stop_demo_playback()`
  (`src/game/ramromreplay.c:604-616`), touches only joystick-playback
  bookkeeping (`joySetPlaybackFunc`/`joySetContDataIndex`/
  `ramrom_demo_related_3`/`is_ramrom_flag`) — **no audio call of any
  kind.** All of `sl.c`, `audi.c`, `boss.c` here are unmodified
  ground-truth, so **N64 has exactly the same absence of a
  demo-boundary audio reset** — confirming there is no missing-on-PC
  system-reset bug to find in this direction either.
- **Net effect: both of M-54's priority leads are now closed, and
  neither explains the bug.** This pushes the M-53 "Follow-up" theory
  back to the front — a genuinely `SOUND_FLAG_LOOPED` voice has *no*
  code path that stops it at a `ramrom` attract-cycle boundary on
  **either** platform; whatever the real fix is, it is a **pre-existing
  N64 behavior being newly exposed**, not a PC regression, and per rule
  #2 (game logic is ground truth) is very unlikely to be something we
  should "fix" by adding a stop call the original game never had. The
  live-repro data (M-54 cont.) already showed this exact voice audibly
  looping forever in the PC build; whether real N64 hardware's `PVoice`
  stealing (now confirmed byte-identical logic) actually does silence it
  in practice depends on whether some *other* sound eventually claims
  that priority slot during/after the demo cycle — untested acoustically
  either way.

**Still OPEN. Next session:** stop auditing the voice-pool/reset code —
both candidate mechanisms are now cross-validated clean/absent on both
platforms. The two live options are (a) accept this as a real, harmless-
on-N64-in-practice edge case (the orphaned voice gets stolen once enough
other SFX compete for the small `PVoice` pool during normal gameplay,
just not during an idle unattended attract-mode capture) and downgrade
D202(b) priority accordingly, or (b) do the acoustic check directly —
run a longer capture (10+ min, several attract cycles) or a temporary
`PORT`-gated `ramrom_table` index override to force a door/train demo
reliably, then listen past the demo boundary to hear whether the loop
in fact does eventually cut out once other SFX start competing for
voices (would confirm (a) and close D202(b) as WONTFIX/expected).
Symptom (a)/(2) (wrong-sample-per-weapon, PPK→PUNCH1) is still fully
untouched and is now the higher-value remaining D202 thread — needs the
decoded-PCM listen-compare, not more voice-pool tracing.

**M-56 — D202(a) got the decoded-PCM listen-compare, live user confirmation,
and the index/pointer/ROM-data chain is now independently proven byte-exact
against the real ROM (not just "ruled out" by code audit). Root cause is
narrowed to the actual decode/mix runtime.**

- Added a temporary `GE_AUDIODUMP=1` probe (`port/src/audio.c`,
  `audioSetNextBuffer`) that raw-dumps every mixed s16 stereo buffer
  reaching the SDL device to `audiodump.raw` — the true final output, same
  path as what plays in-game. Ran a headless repro (`-level_33` Dam,
  `GE_INPUTSCRIPT` holding aim+fire for 8 trigger-pulls starting ~10s in),
  converted the dump to WAV, located the transients via a 50ms-window RMS
  scan, and sent the clip to the user. **Exact repro command** (run from
  repo root, `export PATH=/c/msys64/mingw64/bin:$PATH` first):
  ```
  rm -f audiotrace.log audiodump.raw
  GE_AUDIODUMP=1 GE_AUDIOTRACE=1 GE_INPUTSCRIPT="600:R,Z;601:R,Z;650:R,Z;651:R,Z;700:R,Z;701:R,Z;750:R,Z;751:R,Z;800:R,Z;801:R,Z;850:R,Z;851:R,Z;900:R,Z;901:R,Z;950:R,Z;951:R,Z" \
    timeout 30 ./build-pc/ge007.x86_64.exe -level_33
  ```
  `audiotrace.log` (per-call soundIndex/pointer trace) and `audiodump.raw`
  (raw s16le stereo @ 22050Hz, no header — wrap with Python's `wave` module
  or `ffmpeg -f s16le -ar 22050 -ac 2 -i audiodump.raw out.wav`) land in the
  repo root. **User confirmation: "it's playing
  the ricochet sound effect and the slapping melee attack sound effect on
  each gunshot"** — i.e. two distinct wrong-but-real game sounds, not
  silence/noise/a single mislabeled clip.
- The `GE_AUDIOTRACE` probe (already in tree) on that same run showed the
  fire event genuinely requests **soundIndex=46 (`GUN_SILPPK_A_SFX`)**
  paired with **soundIndex=122 (`CART_SPENT_SFX`, the shell-casing sound —
  `gunfire.c:5582`)** — both legitimate, correct call sites for this
  weapon/action, resolving to non-null, structurally-valid
  sound/keyMap/wavetable/book pointers. This is a **live, real-gameplay
  confirmation** of what M-52 had only checked structurally.
- **Independently re-parsed the raw ROM bytes from scratch in Python**
  (no reuse of any project code — a from-first-principles walk of
  `ALBankFile → ALBank → ALInstrument → ALSound → ALWaveTable` against
  `data/ge007.ntsc-final.z64` at the `_sfxctlSegmentRomStart`/
  `_sfxtblSegmentRomStart` addresses) to check the PC port's custom
  `port/src/romdata.c` re-layout tool (`romdataAudioBankPcSize`/
  `romdataFixupAudioBank`, the `afFixup*` family — hand-written PORT code,
  not decomp, and the one part of this chain not previously scrutinized
  this closely). **Result: exact match.** Index 46's wavetable resolves to
  `base=0x103128A0 len=2242`, index 122's to `base=0x10351DF8 len=2196` —
  **both bit-for-bit identical to the live runtime trace.** Also pulled
  each sound's ADPCM book (order/npredictors/coefficients) and `ALKeyMap`
  directly from ROM: indices 46/47/122 all have distinct, non-aliased book
  coefficients (not a copy-paste/dedup collision), though 46 and 47
  (`PUNCH1_SFX`) do share an identical `keyBase=54/detune=50` keymap —
  coincidental (both are attack-transient sounds tuned the same), not by
  itself an index confusion.
- **This closes out the index/pointer/ROM-offset chain for good** — it is
  now proven correct by independent reconstruction, not just "audited and
  believed clean" (M-52's original phrasing). Since each sound's raw
  ROM-embedded ADPCM data is objectively distinct (different books), the
  wrong-sounding output can only be happening in the **decode or mixing
  runtime itself** — i.e. the actually executed `port/src/mixer.c` path,
  not the data feeding it.
- **New leading suspect, not yet tested**: **D199's flagged risk** — the
  Phase-3 mixer's `aSetBuffer` **persistent-DMEM-context inference**
  (macro-swap `aXxx`→`aXxxImpl` execution with no RSP disassembly to
  verify against, `findings.md` D199) — if two voices active in the same
  audio frame (e.g. the continuous Dam wave-ambience loop already seen
  looping throughout this capture, soundIndex 65/67, plus the fire SFX)
  end up sharing/clobbering the same inferred DMEM decode-context instead
  of each getting its own, one voice's ADPCM decoder state could bleed
  into another's output — producing exactly this "correct data in, wrong
  but real-sounding-like-something-else data out" symptom, deterministically
  per concurrent-voice pairing. This is a hypothesis, not yet verified.

**Still OPEN. Next session:** audit `port/src/mixer.c`'s `aSetBuffer`/DMEM
context handling for whether concurrently-active voices (not just
back-to-back `Impl` calls for the *same* voice) get isolated decode state —
trace which physical DMEM buffer/context each of the fire SFX and the
continuous ambience loop actually uses per frame during the same repro
(`GE_AUDIODUMP`+`GE_AUDIOTRACE`, `-level_33`, aim+fire script above already
gives a clean repro). Do NOT re-check the ROM offset/pointer/index chain
again — independently re-derived byte-exact this session, closed for good.

**M-57 — the M-56 concurrent-voice-DMEM hypothesis is FALSIFIED by a live
opcode-level trace, not just re-audited. A new (currently dormant)
pointer-truncation bug was found in the same family as D198/D201, and a
full sibling audit was produced, but NEITHER explains D202 yet — still
OPEN.** Interactive session, build+run environment confirmed working
(`export PATH=/c/msys64/mingw64/bin:$PATH` from repo root).

- Added a temporary `GE_MIXERTRACE=1` probe (`port/src/mixer.c`) that logs
  every `aSetBuffer`/`aLoadADPCM`/`aADPCMdec`/`aResample`/`aEnvMixer` call
  with its DMEM addresses and state pointer, in call order, to
  `mixertrace.log`. Reran the exact M-56 repro
  (`GE_AUDIODUMP=1 GE_AUDIOTRACE=1 GE_MIXERTRACE=1 GE_INPUTSCRIPT=... -level_33`)
  and read the trace around the soundIndex 46/122 fire event (cross-referenced
  via `audiotrace.log` and the `alLoadParam` wire probe's `audiotrace_wire.log`,
  both already in tree from M-52/M-56).
- **Result: every voice's full chain — `aLoadADPCM` → `aSetBuffer`(decode) →
  `aADPCMdec` → `aSetBuffer`(resample) → `aResample` → `aSetBuffer`(main) →
  `aSetBuffer`(aux) → `aEnvMixer` — runs to completion, strictly sequentially,
  before the next voice's `aLoadADPCM` starts.** No interleaving between
  voices was observed anywhere in ~62k logged opcodes spanning the fire
  event. The single shared `sDmem`/`sAdpcmTable`/`sCtx` globals are safe
  under this call pattern for exactly the same reason the real RSP's single
  physical DMEM is safe: one voice's full pipeline finishes and its result is
  accumulated into the persistent `AL_MAIN_L_OUT`/`AL_AUX_*` buffers before
  the scratch region (`AL_DECODER_IN`=0) is reused by the next voice. **D199's
  "concurrent-voice DMEM clobber" risk is closed, negative** — falsified by
  direct trace, not just re-reasoned about.
- Also noticed `a->table->len` visibly changes (2242 → 2241) between the
  first and later `alLoadParam` wire-probe log lines for the *same* shared
  `ALWaveTable`. Traced this to `load.c:396`
  (`a->table->len = ADPCMFBYTES * ((s32)(a->table->len/ADPCMFBYTES));`) —
  **unmodified N64 ground-truth code**, rounds the table length down to a
  whole-ADPCM-frame multiple. It mutates the shared table in place but is
  idempotent (rounds to the same value on every call after the first) and
  behaves identically on N64. **Not a bug — ruled out.**
- **New finding (not yet root-caused as D202's cause, likely a separate
  latent issue): `src/libultra/audio/load.c` lines 111, 173, 259, 320, 382,
  438 all do `f->memin = (s32) f->table->base + ...` / `a->memin = (s32)
  a->table->base` — narrowing a real 64-bit `u8 *base` pointer
  (`ALWaveTable.base`, `include/PR/libaudio.h:217`) into the `s32 memin`
  field** (matches N64 ground truth's struct layout, `synthInternals.h`).
  This is the same truncation class as the already-fixed D198
  (`K0_TO_PHYS`) and D201 (`bnkf.c` relocation offsets), but **currently
  dormant in this session's build**: the observed `ALWaveTable.base` values
  (e.g. `0x103128a0`) sit comfortably under 4GB (this exe loads its ROM-data
  blob at a low, non-ASLR'd address), so the `(s32)` truncation followed by
  `osPhysicalToVirtual`'s `u32`→pointer zero-extension round-trips exactly —
  no data corruption observed in this run. Flagging as a **hardening item**,
  not a confirmed D202 cause; would need a build/environment where the
  ROM-data blob's heap address exceeds 4GB to prove it live. Not fixed this
  session (out of scope for this investigation; needs its own D-number if
  picked up).
- Delegated a read-only sibling-audit of this cast pattern across the whole
  audio subsystem to the local Qwen `repo-mapper` agent (grep-only, no
  edits, checked against real files). Full hit list, confirmed genuine
  pointer-truncation sites beyond `load.c`:
  - `src/music.c:879,1073,1266` — `(u8*)((t3 + (s32)thing.seqData) -
    trackSizeBytes)` — same class, narrows a seq-data pointer before a ROM→RAM
    copy. Worth checking next alongside `load.c`.
  - `src/libultra/audio/heapinit.c:26` — narrows the audio heap `base`
    pointer to `s32` but immediately masks with `& AL_CACHE_ALIGN` (only
    alignment low-bits survive) — low risk in practice, still worth a
    `uintptr_t` pass for hygiene.
  - Every other `(s32)`/`(int)` cast in `src/libultra/audio/*.c` and
    `src/libultrare/audio/*.c` (`resample.c`, `reverb.c`, `env.c`,
    `seqplayer.c`, `csplayer.c`, `save.c`, `synthesizer.c`, `drvrNew.c`)
    narrows a genuine small integer (byte/sample counts, pitch ratios,
    indices) — not pointer truncation, nothing to fix.

**M-58 — priority (1) from the M-57 write-up (decoded-PCM-per-voice-slot dump)
executed. Result: address resolution AND the ADPCM decode math are BOTH
independently verified bit-exact correct for a live fire-SFX occurrence.
D202 is NOT in `aLoadBuffer`'s DMA/address plumbing and NOT in
`aADPCMdecImpl`'s decode arithmetic. Still OPEN — the bug is downstream of
decode or in voice/output routing.**

- Extended the `GE_MIXERTRACE` probe (env-gated, zero cost when unset)
  three ways: (a) `port/src/mixer.c` `aLoadBufferImpl`/`aLoadADPCMImpl`/
  `aADPCMdecImpl` now log the actual DMEM source bytes, the loaded ADPCM
  book coefficients, and the first decoded 16-sample frame (skipping the
  16-sample history lead-in) into the *same* `mixertrace.log` as the M-57
  opcode trace; (b) `src/libultra/audio/load.c` `_decodeChunk` logs
  `[DMAREQ] filter=<ALLoadFilter*> memin=<addr>` right before calling
  `f->dma(...)`, and `alLoadParam`'s existing `AL_FILTER_SET_WAVETABLE` case
  now also emits a `[BINDTABLE] filter=<ptr> <- table=... base=...` line into
  `mixertrace.log` (previously only in the separate `audiotrace_wire.log`)
  so table-bind, DMA, and decode events for the same physical voice filter
  interleave in one chronologically-ordered file instead of needing
  cross-file correlation by line count (a real mistake made mid-session:
  comparing `wavetable->base` from one run's `audiotrace.log` against
  `aLoadBuffer` addresses from a *different* run's `mixertrace.log` produced
  a spurious "addresses never match" false alarm — resolved by getting both
  probes into one file in one run); (c) `src/audi.c` `amDmaCallback` now
  logs `[DMAHIT]`/`[DMAMISS]` with the input ROM-side address, matched
  buffer's `startAddr`, and the returned staging-buffer address.
- Reran the exact M-56/M-57 repro
  (`GE_AUDIODUMP=1 GE_AUDIOTRACE=1 GE_MIXERTRACE=1 GE_INPUTSCRIPT=... -level_33`,
  see M-56 for the exact command). Picked one live fire event
  (`soundIndex=46`, `wavetable->base=0x103128a0`) and followed its physical
  voice filter (`ALLoadFilter*`) end to end in one linear read of
  `mixertrace.log`:
  - `[BINDTABLE] filter=...706d7230 <- table=...706c4cc8 base=0x103128a0`
  - `[DMAREQ] filter=...706d7230 memin=0x103128a0` — **exactly matches** the
    bound table's own base. No cross-voice/cross-heap address confusion.
  - `[DMAMISS] addr=0x103128a0 ... ret=0x706dd490` — legitimate cold DMA
    (N64-analogous ROM→staging-buffer copy via `port/src/libultra.c`'s
    `piServiceDma`, which correctly `memcpy`s from the cart-mapped `.z64`
    region at `srcPA=0x103128a0`, a real, valid cart address under
    `romdata.c`'s mapping — not a stray heap pointer as first suspected
    mid-session before the single-file trace fix above).
  - `[LOADBUF] dram=0x706dd490 ... bytes=900de113224092ecde90f602424ae1bf` —
    the exact ADPCM-encoded source bytes now loaded into DMEM.
  - `[ADPCMDEC] ... book0=-881` and the dumped book coefficients match
    `sndPlaySfx`'s own `book=00000000706c4ca0` trace for soundIndex 46
    (`+8` byte print-pointer offset, same object) — the correct predictor
    table for this sound, not a stale/aliased one.
  - `[PCMOUT] book0=-881 first-decoded-frame[0..15]=0,-1536,-2475,-1165,...`
- **Independently hand-decoded the same 9 input bytes against the same
  dumped book coefficients using the textbook VADPCM algorithm** (order-2
  predictor, `header byte -> shift+tableIndex`, nibble sign-extension,
  cumulative predictor sum), entirely outside this codebase, no reuse of
  `mixer.c`'s logic. First three samples worked by hand: `0`, `-1536`,
  `-2475` — **bit-exact match** against the C implementation's dumped
  output. This independently confirms `aADPCMdecImpl`'s math is a correct
  VADPCM decoder, not just internally self-consistent.
- Also noticed (not a bug, recorded for the record): `sAdpcmTable` is a
  single static global reused across sound switches without zeroing unused
  predictor slots, so predictor indices beyond a sound's own `npredictors`
  hold stale coefficients from a previous sound's book load. Harmless in
  practice — a sound's own ADPCM-encoded stream only ever emits
  `tableIndex` values within its own encoded `npredictors` range (confirmed
  here: soundIndex 46's book is `bookSize=32` bytes = 1 predictor via
  `2*order*npredictors*ADPCMVSIZE`, and its stream's header nibble was `0`,
  the only valid index) — but it is fragile precisely because it depends on
  the encoder never emitting an out-of-range index; a future corrupted/
  malformed bank could read garbage predictor coefficients silently. Not
  pursued further this session (not shown to relate to D202).
- **This closes out the entire `aSetBuffer`→`aLoadBuffer`→`aLoadADPCM`→
  `aADPCMdec` chain for the cases actually traced.** Combined with M-56
  (ROM data/index chain closed) and M-57 (DMEM concurrency closed), the
  first three links of the pipeline — request routing, ROM data integrity,
  and decode — are now all independently verified correct for at least one
  live fire-SFX instance. The mid-session false alarm (see above) is a
  useful lesson: **always put every new correlated probe into the same
  linearly-ordered trace file**; cross-file correlation by line count
  across separate runs is not valid even when the repro is "the same"
  command, because per-run nondeterminism (already catalogued, D117) shifts
  line counts and, this session, briefly looked exactly like a real bug.

**Still OPEN. Next session, in priority order:** (1) the remaining untraced
links are **resample → envmix → final-mix routing**, and **physical-voice
(PVoice) allocation/identity** — whether the correctly-decoded PCM for one
logical `ALSoundState` ends up accumulated into the *wrong* voice's/output's
persistent envelope-mixer volume-ramp state (`sVol`/`saved->t[]` in
`aEnvMixerImpl`) or gets attributed to the wrong output buffer, which would
produce exactly "correct audio computed, wrong audio heard." Trace the same
way as this session (one physical filter/voice followed end-to-end in one
`mixertrace.log`, `[ENVMIX] state=...` pointer identity checked against which
logical sound it's supposed to belong to) rather than assuming and re-deriving.
(2) If that's clean too, suspect PVoice allocation/identity higher up (the
synth's voice-slot assignment, independent of DMEM/decode). (3) Only after
(1)/(2): consider fixing the `load.c`/`music.c` pointer-truncation family as
a hardening pass (own D-number) — still not shown to cause D202. Do NOT
re-check the ROM offset/index/pointer chain (M-56), DMEM/`aSetBuffer`
concurrent-voice sharing (M-57), or the decode arithmetic/addressing
(M-58, this session, verified bit-exact) — all closed.

**M-59 — likely resolution: D202 is NOT a decode/mixer bug. The "ricochet +
melee slap" audio the M-56 listening test heard is a SEPARATE, entirely
legitimate SFX request made by unmodified ground-truth game logic on every
shot in this specific repro, because the scripted aim direction hits a wall/
prop, not open air. Recommend a follow-up open-air listening test before
resuming the M-58 priority-1 envmix/PVoice trace.**

- Two more concrete checks, continuing directly from M-58's decode-chain
  verification (which stays valid and closed): (a) independently confirmed
  the actual DMEM-loaded ADPCM byte string M-58 dumped for soundIndex 46
  (`900de113224092ecde90f602424ae1bf...`) is byte-for-byte identical to the
  **raw, untouched `data/ge007.ntsc-final.z64` file** at the corresponding
  cart offset (`0x03128a0`, i.e. `base 0x103128a0` minus `CART_BASE
  0x10000000`) — read directly with a throwaway Python script, no project
  code reused. Traced why this address is meaningful at all:
  `_sfxtblSegmentRomStart` (`port/src/romassets_u.s:2948`) is a linker
  constant equal to the literal cart address `0x102F19A0`, used as-is by
  `alBnkfNew`'s `table` argument (`src/music.c:703`) → `_bnkfPatchWaveTable`
  (`src/libultra/audio/bnkf.c:140`, `w->base += table`) — i.e. GE's
  wavetable sample data is **never relaid-out or copied by
  `port/src/romdata.c`'s bank-structure fixup at all**; `base` ends up
  pointing straight into the same live-mapped `.z64` image `port/src/audi.c`
  reads cart data from generally (`docs/dev/findings.md` cart-mapping notes,
  M-56). This closes the last theoretical gap in the M-56/M-58 chain: ROM
  data integrity is now verified at the raw file-byte level, not just
  structurally.
  - With the entire request→ROM-offset→DMA→decode chain now proven correct
    end to end, (b) re-examined the *complete* `audiotrace.log` from the
    M-58 repro (`-level_33` Dam, aim+fire `GE_INPUTSCRIPT`) rather than
    filtering to just soundIndex 46/122, and cross-referenced every other
    soundIndex against `src/bondconstants.h`'s `"NN_NAME_SFX"` string table
    (which encodes the index directly in the name — e.g. `"46_
    GUN_SILPPK_A_SFX"`). **Every single `soundIndex=46` (fire) request in
    the trace is immediately followed by a `soundIndex` from `{23, 24, 25,
    37}`** — `"23_RICO_6_TAJ_A_SFX"`, `"24_RICO_6_TAJ_B_SFX"`,
    `"25_RICO_6_TAJ_C_SFX"`, `"37_RICO_5_C_SFX"` — i.e. **ricochet/bullet-
    impact sounds**, cycling per shot, plus the already-known `122` =
    `"122_CART_SPENT_SFX"` (cartridge eject).
  - Found the call site: `src/game/gunfire.c`
    `recall_joy2_hits_edit_detail_edit_flag()` (line ~2270, unmodified
    ground-truth decomp function, original obfuscated name) is the bullet
    hit-detection handler. For a hitscan that strikes a non-character prop
    (`prop->type != PROP_TYPE_CHR && != PROP_TYPE_VIEWER`, i.e. scenery/
    wall, not an enemy), it unconditionally does:
    `ricochet_sounds_small_copy = ricochet_sounds_small;
    sndPlaySfx(g_musicSfxBufferPtr,
    ricochet_sounds_small_copy.arr[randomGetNext() % 20], sound_state);`
    — **every bullet that hits a wall/prop plays a randomly-chosen ricochet
    SFX from a 20-entry table, by design, on real N64 too.** This is not a
    port artifact; it is exactly what GoldenEye does when you shoot a wall.
  - **Conclusion**: the M-52/M-56 repro's fixed aim direction
    (`-level_33`, default spawn facing, held-down aim+fire) has the player
    shooting directly into a nearby wall/prop the entire time, so *every*
    trigger pull legitimately plays fire (46) + cartridge-eject (122) +
    a random ricochet impact (23/24/25/37) simultaneously — three real,
    correct SFX per shot, not one wrong one. The M-52 playtest's original
    verbal description ("sounds like a melee/punch hit") and the M-56
    listening test's ("ricochet sound effect and the slapping melee attack
    sound effect") both match a ricochet impact's percussive/metallic
    character far better than a coincidental decode bug that happens to
    always produce two *other real, nameable* game sounds. No corrupted or
    misattributed audio has been found anywhere in the chain despite three
    full sessions (M-56/M-57/M-58) of independent, bit-exact and byte-exact
    verification at every single link — address resolution, ROM data
    integrity, book/coefficient selection, ADPCM decode arithmetic, and now
    the request pattern itself.
  - **This does not yet prove the gunfire (46) sound itself is
    perceptually correct/prominent** — only that it is being computed
    correctly and is not what's "wrong" in what the tester heard. It's
    possible the fire SFX is simply quiet/subtle by design and gets
    acoustically masked by the louder ricochet transient, which would be
    expected N64 behavior too, or (lower probability, unverified) there is
    a genuine separate volume/mix-balance issue between simultaneous
    voices. Not fixed or further diagnosed this session — needs a
    clean listening test to settle it (see below).

**Recommended next session (before resuming the M-58 priority-1 envmix/
PVoice trace, which may now be unnecessary): re-run the exact same
`GE_AUDIODUMP` capture but with an `GE_INPUTSCRIPT` aim direction that
points at open air / the sky instead of a wall (or simply hold fire without
the aim-lock at a fixed wall-facing spawn), confirm via `audiotrace.log`
that NO `ricochet_sounds_small`-family soundIndex fires, and have the user
listen to that clip specifically. If the fire (46) + cartridge (122) sounds
alone are now clearly audible, correct, and recognizable as gunfire, D202
should be closed as a repro artifact (aim-at-wall), not a bug — with a note
that the original M-52 playtest report may itself have simply been the
player firing at nearby geometry during normal play, which is entirely
expected. Only if the open-air clip STILL sounds wrong should the M-58
priority-1 envmix/PVoice-identity trace be resumed.

**Follow-up done same session (M-59 cont.):** turned the character with
`GE_INPUTSCRIPT="1:SRIGHT;400:SNONE;600:R,Z;..."` (sustained analog-stick
turn before the aim+fire sequence) and reran the `GE_AUDIOTRACE`+
`GE_AUDIODUMP` capture on `-level_33`. Result: the *first* shot in this run
requests **only** `soundIndex=46` (fire) + `122` (cartridge) — no
`ricochet_sounds_small` index at all — confirming the turn moved the aim
off whatever wall/prop the original fixed-facing repro was hitting. The
*second* shot (moments later, once the turn brought the aim back toward
nearby geometry) requests `46` + `122` + `23` (`RICO_6_TAJ_A_SFX`) again,
giving a same-run A/B pair. Located both transients in `audiodump.raw` via
a 50ms-window RMS scan (clean shot ≈ t=11.0–11.7s, ricochet shot ≈
t=11.7–12.6s), exported both as WAV (`scratchpad/
d202_clean_shot_no_ricochet.wav` / `scratchpad/d202_shot_with_ricochet.wav`,
not committed — scratch), and sent both to the user for a direct listening
comparison. **Awaiting user confirmation** on whether the ricochet-free
clip sounds like correct, recognizable gunfire — that answer settles
whether D202 closes outright as a repro artifact or the envmix/PVoice trace
needs to resume.

**M-59 result: user says BOTH clips are the ricochet/slap sound, just at
different volumes — the "clean" clip is not clean.** This falsifies the
"aim-at-wall repro artifact" theory as originally framed. Investigated
further:

- The two exported clips both came from `dumppos`-adjacent events with
  soundIndex 46 and a ricochet index at the **same audio-frame position**
  in this run's `audiotrace.log` (a new `dumppos=<byte offset into
  audiodump.raw>` field was added to the existing `GE_AUDIOTRACE` probe —
  `port/src/audio.c` now exposes `audioDumpBytePos()`, an exact running
  byte counter, so future clip extraction can key off a logged sample
  position instead of an RMS-scan guess). Confirmed the fire (46) and
  ricochet requests land at the *same* `dumppos`, i.e. they are requested
  in the same tick and their audio overlaps in time rather than being
  sequential — so the RMS-peak-picking method used for the M-59 clips could
  not actually isolate a ricochet-free window; both exported clips likely
  had some ricochet content, just at different relative volume (distance-
  dependent ricochet gain, or a different ricochet variant per shot).
- To get a **guaranteed** ricochet-free sample, decoded `soundIndex 46`
  **entirely offline**: read `data/ge007.ntsc-final.z64` directly at ROM
  offset `0x03128a0`, len 2242, ran the same hand-verified VADPCM decode
  used in M-58 (order-2, single predictor, coefficients from the M-58
  trace dump) with no live game/mixer involved at all — zero possibility of
  a simultaneous ricochet voice. Rendered to `scratchpad/
  d202_soundindex46_isolated.wav`, sent to the user. **User confirmed: "yeah
  that is the slap effect that in the original game plays when you do a
  melee attack."** This is unambiguous — soundIndex 46's actual ROM sample
  content genuinely is (or matches) the melee-slap SFX, independent of any
  mixing/overlap question.
- **Traced the request chain one level further up, past `sndPlaySfx`, to
  find out why soundIndex 46 is requested for the silenced PPK's fire at
  all.** `bondwalkItemGetSound(item)` (`src/game/gun.c:1345`) returns
  `get_ptr_item_statistics(item)->Sound` — a `u16 Sound` field on the
  per-weapon `WeaponStats` struct (`src/game/gun.h:87`, doc comment:
  "Sound effect played when gun is shot. There are 261 sound effects, or
  0 - 105h."). `get_ptr_item_statistics` (`gun.c:721`) returns
  `gitem_structs[item].item_weapon_stats` — **not a runtime ROM-converted
  pointer**; `gitem_structs[]` (`assets/obseg/gun/gunModelFileRecord.inc.c`)
  is a **static, compiled C data table**, decompiled directly from the
  ROM's `.data` segment (each `GUNFILERECORD(...)` entry pulling in a
  `GUNSTATS(name)` initializer). The silenced PPK's is
  `assets/obseg/gun/wppksil/gunWeaponStat.inc.c`
  (`WeaponStats wppksil_stats = {1.0, 11.0, ..., 0x2E, ...}`, tagged with a
  decompilation RAM-address comment `//D:800326C4`) — **`.Sound` is
  positionally the 14th field, value `0x2E` = decimal **46**.** This is
  unmodified, hand-decompiled ground truth, not a port-side conversion —
  there is no PC-specific fixup in this chain at all (unlike the wavetable
  data, this table is a plain compiled literal, identical on N64 and PC by
  construction).
- **Conclusion: on real N64 hardware, firing the silenced PPK requests
  the exact same soundIndex 46 this port does.** Every link — decompiled
  weapon-stat data, SFX-bank index resolution (M-56/M-59), ROM sample
  bytes (M-58/M-59), and ADPCM decode (M-58) — is now independently
  verified correct and, more importantly, **identical to what N64 hardware
  would do with the same ROM**. If soundIndex 46 truly sounds like a melee
  slap and not a silenced pistol, that is either (a) the original,
  shipped game's actual designed audio (silenced firearms can have a soft,
  percussive "thwack" character that isn't unreasonable to mistake for a
  melee sound, and reusing one SFX asset across two unrelated actions to
  save ROM space is a common technique of this era) — in which case there
  is **no bug anywhere in this port to fix**, this is how GoldenEye 007
  actually sounds; or (b) a very rare decompilation transcription error in
  `wppksil_stats`'s `.Sound` field specifically — but this project's
  entire premise is byte-verified matching against the real ROM's `.data`
  segment (README/AGENTS.md), so an error surviving in a byte-matched
  build would be surprising, though not provably impossible without
  independently re-reading the literal ROM bytes at this struct's true RAM/
  ROM address (not attempted this session — would need the N64 build's
  linker map to translate the `//D:800326C4` RAM comment to a ROM file
  offset).
- **Not fixed, not committed — this is a research conclusion, not a code
  change.** No `port/`, `src/`, or `assets/` change is warranted unless (b)
  above is confirmed; per AGENTS.md, `assets/obseg/gun/wppksil/
  gunWeaponStat.inc.c` is decompiled ROM data (ground truth) and must not
  be "corrected" speculatively.

**Recommended definitive tie-breaker (next session, if the user wants to
keep pursuing this rather than close it):** run the **same ROM** in a
known-good N64 emulator (or, if available, real hardware) and fire the
silenced PPK. If it also sounds like the melee-slap effect there, D202 is
closed for good as a genuine original-game characteristic (or a very
deep, pre-existing GoldenEye quirk) — not a decompilation or port bug, and
no further engineering time should go into it. If the emulator's silenced
PPK sounds like a normal muted gunshot, the bug is almost certainly the
rare decompilation transcription error in (b), and the fix would be a
single-field data correction in `gunWeaponStat.inc.c` guarded by the
project's existing byte-match verification (which would need to still
pass — if it doesn't, this table is confirmed byte-correct after all and
the "bug" is upstream of Bond's silenced-PPK path in some other way not
yet identified).

**M-60 — user explicitly rejects the "ground truth, not a bug" conclusion**
("I know for a fact its the wrong sound playing in the PC port
specifically") and supplies a real reference clip instead of an emulator
A/B: `scratchpad/B00I00S2D.wav` (untracked/scratch), a genuine recording of
the silenced PPK's real fire sound, dropped directly by the user (source
not yet stated — presumably real hardware/video capture, not this repo).
Measured: mono, 16-bit, 16 kHz, 3264 samples, 0.204 s. Its 10-bucket RMS
envelope (`12645, 11548, 15152, 16172, 14006, 12766, 15106, 15669, 8280,
6608, 8032, 9409, 9812, 8849, 7745, 6819, 4299, 1477, 496, 176` over 20
buckets) is **sustained** — energy stays in the 6000-16000 range for
roughly the first 80% of the clip's duration before decaying in the last
~20% — structurally different from `d202_soundindex46_isolated.wav`'s
envelope, which peaks early (`11209, 19213, 17387, 16284, 11296, 10047,
6116, 3801, 2238, 1059` over 10 buckets) and decays smoothly and quickly
from the first bucket onward. This is consistent with the user's "slap"
identification (a single sharp decaying transient) vs. a real gunshot's
more sustained buzz/reverb character, and is independent corroborating
evidence — not proof by itself, since envelope shape alone doesn't rule out
a resample/pitch difference distorting the comparison (see below).

**Re-opened the investigation on the user's instruction; do not close D202
as WONTFIX.** Two concrete threads were started this session, neither
finished:

1. **The debug-name label itself is suspicious independent of any
   listening test.** `bondconstants.h`'s soundIndex 46 is *named*
   `"46_GUN_SILPPK_A_SFX"` — a name that specifically claims to be the
   silenced-PPK gunshot — yet its decoded ROM content is what the user
   twice identified as the melee-slap effect (which has its own, different,
   adjacent indices: `47_PUNCH1_SFX`/`48_PUNCH2_SFX`/`49_PUNCH3_SFX`). A
   correctly-decoded sound whose content doesn't match its own debug label
   points at a **sound-bank navigation/indexing bug** (wrong bank, wrong
   instrument, or an off-by-one somewhere in the ALBank -> ALInstrument ->
   ALSound walk) rather than at "GoldenEye's SFX design reuses one asset
   for two things" (M-59's charitable reading (a)). `src/snd.c`'s
   `sndPlaySfx` doc comment is explicit about the resolution path:
   `soundBank->instArray[0]->soundArray[soundIndex]` — i.e. `soundIndex` is
   an index into **instrument 0's** sound array, not a bank-global flat
   index. Worth checking directly: does `romdata.c`'s bank re-layout
   (`afCtx`/`_bnkfPatchWaveTable`, `port/src/romdata.c` ~L582-780) walk
   instruments/sounds in the same order the ROM's `ALBankFile` stores them,
   or could the PC conversion silently shift indices by one somewhere in
   that re-layout pass? Not yet checked line-by-line against the ROM bytes
   for THIS specific bank/instrument.
2. **Started `scratchpad/rom_sfx_decode.py`** (untracked, WIP, not yet
   working correctly) — a fresh, from-scratch offline ROM bank parser
   (independent of the live game) intended to decode a whole *range* of
   `soundIndex` values (44-50, 107, extensible) directly from
   `_sfxctlSegmentRomStart`/`_sfxtblSegmentRomStart` and compare each one's
   envelope/waveform against the new `B00I00S2D.wav` reference, to find
   which soundIndex (if any) actually matches the real gunshot — which
   would prove a specific off-by-N in the bank walk rather than relying on
   ear-only comparisons. **This script's ADPCM predictor loop is currently
   WRONG and must not be trusted as-is** — it stubs out the intra-subframe
   accumulation term with a dead `if False` branch. Confirmed the *exact*
   correct algorithm by re-reading `port/src/mixer.c`'s `aADPCMdecImpl`
   (L185-239, ground truth, this is what the live port actually runs):
   table shape `sAdpcmTable[predictor][2][8]`; per output sample `j` (0-7)
   within a subframe, `acc = tbl[0][j]*prev2 + tbl[1][j]*prev1 +
   (ins[j]<<11) + sum_{k=0}^{j-1} tbl[1][j-k-1]*ins[k]`, `ins[]` being the
   8 sign-extended-then-shifted nibbles of that subframe, `prev1`/`prev2`
   the last two output samples (carried from the previous subframe/frame).
   **Next session: fix `rom_sfx_decode.py`'s decode loop to match this
   exactly** (the M-58/M-59 "isolated decode" script that the user's slap
   identification was based on was verified byte-exact against this same
   algorithm in a prior session, so that specific result stands — only the
   new range-scanning script is unverified), then run it across a wider
   soundIndex range and diff each envelope against `B00I00S2D.wav`.
3. **Also flagged, not yet resolved:** the M-59 "correctpitch" resample
   experiment (`scratchpad/d202_soundindex46_correctpitch.wav`, pitch=23849
   from the M-58 live trace) produced *fewer* output samples (1459) than
   input (3984) at a pitch ratio ~0.364. If pitch < 1.0 means "advance less
   than one input sample per output sample" (slow down / stretch), output
   sample count should be `input / ratio` (more samples), not `input *
   ratio` (fewer) — the direction looks inverted. This resample script's
   formula needs re-deriving against `port/src/mixer.c`'s actual
   `aResampleImpl`/`sResampleTable` before its output is trusted either
   way; it has NOT been sent to the user and should not be used as evidence
   yet.

**Status at handoff: D202 is OPEN, user-confirmed-wrong-in-PC-port, actively
being re-investigated.** Do not re-close as "ground truth, not a bug" —
that conclusion was explicitly rejected by the user with a real reference
recording now in hand (`scratchpad/B00I00S2D.wav`). No code changes made
this session (diagnostic-only), nothing to revert.

**M-61 — both M-60 threads resolved (one closed, one fixed+run); the
decisive question is now "what byte is actually in the original ROM", and a
new logical constraint makes that the keystone fact.** Diagnostic-only
session again; no game/port code touched. Results:

1. **Thread 1 (bank-navigation off-by-N) CLOSED — negative.** Line-by-line
   review of `port/src/romdata.c`'s bank re-layout (`afCtx` /
   `_bnkfPatchWaveTable`, ~L582-780) confirms the PC conversion walks
   `ALBankFile -> ALInstrumentFile -> ALSoundFile` in exact ROM order with
   no index shift; combined with M-56's byte-exact runtime trace, PC
   soundIndex N resolves to the same ROM sample as N64 soundIndex N. The
   "wrong bank / off-by-one in the walk" hypothesis is dead. Do not
   re-litigate it.
2. **Thread 2 (offline decoder) FIXED and VALIDATED.**
   `scratchpad/rom_sfx_decode.py`'s `adpcm_decode()` was rewritten to match
   `port/src/mixer.c`'s `aADPCMdecImpl` exactly. Three bugs fixed: (a)
   tbl0/tbl1 indexing was swapped, (b) the intra-subframe accumulation term
   `sum_{k<j} tbl[1][j-k-1]*ins[k]` was missing, (c) wrong `ncoef` count.
   Validation: full-bank scan of all 261 sounds (gun.h L85: "There are 261
   sound effects, or 0 - 105h") decoded cleanly; **index 45
   (`45_DROP_GUN_SFX`) is the ONLY strong match for the user's reference
   clip `B00I00S2D.wav` (envelope correlation +0.999)**; index 46
   (`46_GUN_SILPPK_A_SFX`) decodes to the early-peak fast-decay "slap".
   Candidates saved as `scratchpad/romsfx_44..60.wav`. The decoder is now
   trustworthy for future soundIndex lookups.
3. **Fire path traced end-to-end in game code — no runtime remapping of the
   index exists.** `src/game/gunfire.c` L3196/L3200:
   `sndPlaySfx(g_musicSfxBufferPtr, bondwalkItemGetSound(var_s1), ...)` →
   `src/game/gun.c` L1343-1345: `bondwalkItemGetSound()` returns
   `get_ptr_item_statistics(item)->Sound` — the raw u16
   `WeaponStats.Sound` field (offset 38; layout confirmed from
   `src/game/gun.h`: 7×f32, s32 AmmoType, s16 MagSize, u8 AutoFireRate,
   s8 SingleFireRate, u8 ObjectsShootThrough, u8 SoundTriggerRate, u16
   Sound, ptr cartridge, …). The silenced PPK's fire sound IS
   `wppksil_stats.Sound`, used verbatim. (Unsilenced PPK uses 0x6B=107.)
4. **Build wiring: both N64 and PC builds compile the same initializer from
   source.** `src/game/gun.c:135` includes the aggregate
   `assets/obseg/gun/gunWeaponStats.inc.c`, which defines `wppksil_stats`
   inline at `//D:800326C4` with `.Sound = 0x2E` (it only #includes the
   per-weapon files for fist/knife; everything else is inline). The
   per-weapon file `assets/obseg/gun/wppksil/gunWeaponStat.inc.c` carries a
   second, IDENTICAL copy (also `0x2E`, same `//D:800326C4`). The PC build
   exclusion list covers only `assets/obseg/{bg,brief,setup,stan}`
   (CMakeLists ~L473), so gun.c's compiled data is identical in both
   builds. **The decompiled source is internally consistent: 0x2E in both
   copies.**
5. **KEY LOGICAL CONSTRAINT (new, this is the keystone):** the project's
   premise is that the N64 build byte-matches the original ROM. Compiled
   data cannot silently diverge — if `gunWeaponStat.inc.c`'s `.Sound`
   were a transcription error, the built ROM would differ from the
   original at exactly those 2 bytes and the match would fail. Therefore
   **0x2E almost certainly IS in the original ROM**, i.e. real N64 hardware
   also requests sound index 46 for the silenced PPK. That contradicts the
   reference clip matching index 45 at +0.999. Exactly one of these must be
   false:
   (a) the byte-match verification does not actually cover this data
       segment (check the verify tooling/output!);
   (b) `B00I00S2D.wav` is not an in-game capture from real N64 hardware —
       its filename looks like a media-library asset ID, provenance was
       never stated; if it's e.g. film/marketing audio, the "N64 plays 45"
       premise collapses and M-59's "faithful reproduction" conclusion
       returns WITH evidence;
   (c) N64 runtime resolution of index 46 differs from ROM-order index 46
       (Thread 1 closed *ordering* in the PC re-layout, not real-hardware
       ALBank content — lower priority given (a)/(b));
   (d) the +0.999 envelope correlation with index 45 is misleading (room
       reverb tail can inflate envelope matches).
6. **Direct ROM-byte read of `wppksil_stats.Sound` FAILED — do NOT repeat
   the same searches (they looped 40+ times this session, all 0 hits):**
   - Any pattern containing a decompiled float literal is unreliable: the
     literals are decimal approximations that don't round-trip to exact
     IEEE-754 bits (e.g. `-20.799999` → `c1a66666`, zero occurrences in
     ROM). Integer-only tail patterns (`AmmoType=1, MagSize=7, 0xFF, 0x10,
     …`) also returned 0 hits — either the byte layout at that ROM location
     differs from assumption or the data sits where the scan didn't look.
   - The hand-computed RAM→ROM mapping produced ROM offset `0xC11934` >
     file size `0xC00000` — WRONG. Note `ge007.ld` puts `.csegment` at ROM
     `0xC00000` (= exactly EOF), so `wppksil_stats` (vaddr `0x800326C4`)
     cannot live there; the correct vaddr→ROM mapping must come from the
     N64 build's map file/symbol table, not arithmetic.
   - No N64 `.map` found at maxdepth 3 (only the PC CMake tree
     `build-linux/`). Search DEEPER: `assets/obseg/Makefile.*`, `ld/`,
     `dist/`, or any N64 ELF in the repo; `nm`/`objdump` on it yields the
     exact vaddr+section, and the ROM offset then follows from the linker
     script.
7. **Next session, in priority order:**
   1. **Establish what byte the original ROM holds at
      `wppksil_stats.Sound`** — find the N64 build map/ELF (search deeper
      than maxdepth 3) and read the 2 bytes. This single fact splits the
      tree: ROM=0x2D → transcription error (then audit why byte-match
      verification missed it); ROM=0x2E → data is ground truth, and the
      discrepancy lives in reference-clip provenance (b) or N64 runtime
      resolution (c).
   2. **Check the byte-match verification tooling** — does it compare data
      segments or code only? Locate the last full-verify output.
   3. **Ask the user for `B00I00S2D.wav`'s provenance** (one question; see
      (b) above).
   4. Cheap corroboration: live PC trace logging the soundIndex actually
      passed to `sndPlaySfx` on PPK fire (expected 46) — confirms the PC
      side behaves exactly per its data.

**Status at handoff (M-61): D202 OPEN.** Threads 1+2 from M-60 are closed
(negative / fixed). No code changes; only `scratchpad/rom_sfx_decode.py`
modified (now correct + validated). Do not re-run the failed ROM byte-
pattern searches — go for the N64 map file instead.

### M-62 — item (2) resolved without a build: byte-match verification DOES cover this exact struct, and its checksum was captured after the source's last edit. `.Sound = 0x2E` is very likely ground truth. Pivoting to (b)/(d).

Static/read-only session. Skipped the N64-map/ELF route (item 1) — no MIPS
toolchain is installed in this environment (`mips-linux-gnu-*` absent,
`docs/building.md` says extraction needs `binutils-mips-linux-gnu`; building
the full N64 target to get a map file is a real toolchain-install task, not
attempted this session) — and went straight for the cheaper item (2) instead.

**Found: this repo's own CI (`.github/workflows/ci.yml`) never builds the
N64 target at all.** `validate`/`linux-build`/`windows-build` are all PC-port
jobs; there is no job invoking the N64 `Makefile` or
`scripts/test_files.sh`. So "byte-matches US/EU/JP ROMs" (AGENTS.md's
opening claim) is inherited from the upstream `n64decomp/007` project's own
history, not continuously re-verified here — worth knowing, but a separate
concern from D202.

**The verification tooling itself does cover `.data`, and gun.o specifically:**
`scripts/test_files_readme.md` — `test_files.sh` extracts `.text`/`.code`/
`.bss`/`.data`/`.rodata` from every `.o` under `src/`, `src/game`, and the
four asset dirs, and diffs each against a checked-in known-good md5
(`scripts/ge007.u-test_basis.csv`). Confirmed by direct grep:
`scripts/ge007.u-test_basis.csv:946`: `0289c36e967840bba7f6fbd026dda001,.data,build/u/src/game/gun.o`
— `gun.o`'s `.data` section (which is where `wppksil_stats` lives, via
`gun.c`'s `#include` of `gunWeaponStat.inc.c`) has a recorded checksum.
This directly falsifies alternative (a) from M-61 — the verification does
cover this exact data.

**Timing check, so the checksum can't be stale relative to a since-edited
source:** `git blame -L946,946 scripts/ge007.u-test_basis.csv` →
`cc14d64e7 "for england james?"` (2026-08-16, upstream sync commit).
`git log -- assets/obseg/gun/wppksil/gunWeaponStat.inc.c` → last touched
`9fbe1fd5` ("sync. run make forceextractassets..."), which is an ancestor of
`cc14d64e7` in this file's history. **So the checksum was captured AFTER
the last edit to this exact source file** — it isn't testing stale content.
Re-read the file directly: `.Sound` is still `0x2E` (14th positional field,
`assets/obseg/gun/wppksil/gunWeaponStat.inc.c:3`), unchanged.

**Conclusion (not a rebuild-confirmed pass, but strong circumstantial
evidence):** the upstream decomp project's own generated fingerprint of a
known-good `gun.o` build covers this exact struct, was captured against
this exact (still-current) source content, and this repo's non-negotiable
rule #2 (game logic/decomp source frozen, PC-port work never touches
`src/game/gun.c` or `assets/obseg/gun/**`) means nothing has drifted since.
**Actually running `test_files.sh` to get a live "pass" line still requires
installing a MIPS toolchain and doing a real N64 build — not done this
session; flagged as the one remaining way to make this airtight rather than
strongly-inferred.** Barring that, alternative (a) is now the *least* likely
of the four M-61 alternatives, not the most. Combined with M-61's already-
closed alternative (c) (traced fire path, no runtime remapping), the two
live leads are now **(b) reference-clip provenance** and **(d) envelope-
match reliability** — both point away from a port bug and toward a
misidentified reference clip.

**Next session, in order:**
1. **Ask the user the one open question: where did `scratchpad/
   B00I00S2D.wav` come from?** (Real N64-hardware capture of this exact
   weapon on this exact ROM region? An emulator? A sound-effects library
   asset, given the filename reads like a media-asset ID, not a capture
   log?) This is the cheapest possible next step and was never actually
   asked — do it before any more code/data spelunking.
2. If the user can't confirm hardware provenance: run the same ROM in a
   known-good N64 emulator (the M-59 handoff's original recommended
   tie-breaker, never done), fire the silenced PPK, listen directly —
   settles (b)/(d) without needing the reference clip at all.
3. Only if both of those come back still contradicting `.Sound = 0x2E`:
   install a MIPS toolchain (`binutils-mips-linux-gnu` equivalent) and
   actually run `scripts/test_files.sh` for real, to convert this
   session's strong-inference close of alternative (a) into a hard pass/fail.
4. Cheap corroboration, still not done: live PC trace of the soundIndex
   passed to `sndPlaySfx` on PPK fire (expected 46) — confirms PC-side
   behavior matches its own data, independent of what the data *should* be.

**Status at handoff (M-62): D202 OPEN**, but the center of mass has shifted
firmly away from "port bug in the data/decode chain" (four independent
sessions, M-56 through M-62, now corroborate `.Sound = 0x2E` end to end)
and toward "reference clip may not be what it was assumed to be." No code
changes this session — read-only investigation, one `findings.md` write-up.

**Same session, follow-up: asked the user directly where `B00I00S2D.wav`
came from. Answer: "Not sure / found online"** — the user does not know its
provenance and it was sourced from the internet, not recorded from real N64
hardware or a verified emulator run against this ROM. This closes item (1)
of the M-62 next-session list and directly confirms alternative **(b)**
from M-61 (reference clip isn't a verified in-game N64 capture) as the true
explanation — an unverified internet clip was compared against a
byte-verified ROM data table and, unsurprisingly, didn't match.

**Reassessment: five independent sessions (M-56 through M-62) have now
verified every link in the actual chain** — soundIndex resolution, bank
re-layout order, the ALWaveTable/ALSound pointer chain, the ADPCM decode
math (hand-verified bit-exact against the C output), the ROM segment
offset (verified 3 ways), and now the weapon-stats source data itself
(byte-match checksum, captured post-edit) — **all say the silenced PPK
correctly requests and plays soundIndex 46 exactly as real N64 hardware
would.** The only unverified link left is the thing D202 was reopened over:
an internet-sourced reference clip of unknown origin. That is no longer a
credible basis to keep this open as a port bug.

**Recommendation: close D202 as NOT-A-BUG** (same conclusion M-59 reached,
now on much stronger evidence) **unless the user wants the belt-and-braces
emulator A/B** (fire the silenced PPK on the actual ROM in a known-good N64
emulator, listen directly) — cheap, and the only remaining way to fully
rule out a genuine transcription error without installing a MIPS toolchain
and running `scripts/test_files.sh` for a hard pass. Do not spend further
session time on the data/decode/index chain — it is exhaustively verified.

### M-63 — SUPERSEDES the M-62 "recommend close": user reports the bug is broader than a single wrong sample — real-time mixer/voice corruption, not a data-table issue. REOPENED for real, live repro attempted. Root cause NOT found; one concrete anomaly banked for next session.

**User's report (verbatim symptom, not the M-60 reference-clip framing):** the
wrong ("slap") sound always plays on the silenced PPK; the correct suppressed
shot sometimes plays **too** (both heard together, not one-or-the-other);
and over a play session sounds "pile up and spirally glitch out," eventually
leaving **no audio except a looping sound (e.g. Bunker's door) stuck playing
since level start.** This is categorically different from "wrong sample
selected" — it describes runtime mixer/voice-lifecycle corruption. The M-56
through M-62 chain (which verified the index/data/decode path bit-exact) does
**not** cover this; it only proves soundIndex 46 is the *correctly requested*
sound, not that the mixer plays it *cleanly* once requested. Reopening.

**Build+run environment IS available this session** (`export
PATH=/c/msys64/mingw64/bin:$PATH`; MSYS2 MINGW64 gcc/cmake/SDL2 all present).
No MIPS toolchain (confirmed absent again, unrelated to this thread).

**False-alarm detour (banked so it isn't repeated): a "crash in `load_bg_file`
(`bg.c:865`) on every `-level_09` boot" is NOT a real bug** — it was this
session's own test-setup gap. A throwaway `build-pc/data/` with only the
`.z64` copied in (no `pccg-ntsc-final/`, `pcmodels-ntsc-final/`) leaves
`file_resource_table` pointing at raw big-endian ROM instead of the converted
PC sidecars (`obInit()`'s own comment describes exactly this failure mode);
the two `[WARN]` lines ("pcmodels.bin not found", "pccg.bin not found —
stage loads will fault") were sitting right in the log and should have been
read before chasing a gdb backtrace. Copying the real `data/pccg-ntsc-final/`
+ `data/pcmodels-ntsc-final/` (+ the repo's real `ge007.eep`/`ge007.ini`)
into `build-pc/data/` makes it load clean, 60s+, no crash, both with and
without the uncommitted audio-diagnostic files. **Next session: always mirror
the full `data/` dir into `build-pc/data/`, not just the ROM, before treating
any level-load failure as real.**

**Repro built:** `GE_MIXERTRACE=1 GE_AUDIOTRACE=1 GE_AUDIODUMP=1
GE_INPUTSCRIPT="600:R,Z;601:R,Z;690:R,Z;691:R,Z;...;3210:R,Z;3211:R,Z"`
(30 fire pulses at ~90-frame/~0.75s spacing) `timeout 60
./ge007.x86_64.exe -level_09`, run against the real save
(`data/ge007.eep`) copied into `build-pc/data/`. Ran clean 60s, no crash.
Logs preserved: `scratchpad/d202-m63/{audiotrace.log,mixertrace.log,
audiodump.raw}` (mixertrace.log is 25MB — grep it, don't read it whole).

**Checked and RULED OUT this session, from the trace:**
- **Voice-pool exhaustion / leak:** only **16 distinct `newState` voice-slot
  pointers** were used across all 153 `sndPlaySfx` calls in the 60s window
  (highest reuse count: 20), and `sndDeactivate` count (159) tracks
  `sndPlaySfx` count (153) almost 1:1. The pool is being recycled
  continuously, not growing unbounded — the "voices never freed" theory
  (the D199/M-53–M-55 lineage) does **not** reproduce in this repro.
- Firing the silenced PPK produced exactly the expected `soundIndex=46`
  requests (58 of them, one burst per scripted trigger pull, occasionally
  2 back-to-back — plausibly dual-wield, not investigated further) with no
  other unexpected soundIndex interleaved in the same burst.

**New anomaly found, not yet explained — the most concrete lead for next
session:** `soundIndex=109` (`bondconstants.h:2546`,
`"109_GUN_B4_BOLTACTION_SFX", //used for AK47`) fires **94 times in 60
seconds** despite the player never touching an AK47 in this repro (silenced
PPK only). A chunk of those requests hit the *same* voice slot
(`newState=00000000706f4fb8`) at a suspiciously uniform **2880-byte spacing
in the mixed output stream** — at 22050Hz stereo 16-bit that's exactly
**720 samples / ~32.6ms between re-triggers**, four-in-a-row
(`dumppos=688576,691456,694336,697216,700096`). ~30 rounds/sec is far
faster than any plausible automatic-weapon cadence in this game (real
full-auto weapons here fire far slower) — this reads like something is
re-issuing `sndPlaySfx` for the same logical sound every audio tick/frame
instead of once per actual game event, which would exactly explain "sounds
piling up." **Not yet confirmed as related to the user's PPK-slap complaint
specifically** (soundIndex 109 isn't the PPK's sound) — could be a
*second*, independent SFX-retrigger bug that happens to share the same
mixer, and its presence would still explain general "audio feels broken."

**Not yet reproduced: the user's exact reported chain** (piling
up → spiraling glitch → eventual total silence except one stuck loop). This
60s/58-shot synthetic burst didn't collapse the mixer — either it needs a
longer, more realistic session (real human play, many minutes, mixed
NPC/player/ambient audio denser than this script), or the soundIndex=109
retrigger-storm anomaly above is the seed that, given more real playtime,
compounds into what the user described (worth designing a longer repro
around forcing that specific retrigger condition first, cheaper than a raw
extended random playtest).

**Next session, in order:**
1. Find what actually calls `sndPlaySfx(109, ...)` (search `src/game/*.c`
   for callers keyed to soundIndex 109 / `GUN_B4_BOLTACTION_SFX`) and
   determine whether the ~32.6ms retrigger cadence is a real per-shot
   automatic-fire loop (legitimate, if unusually fast) or a bug re-firing
   once per audio buffer/tick instead of once per game event. This is the
   single most concrete, checkable lead from this session.
2. If (1) doesn't explain it: build a much longer repro (5+ minutes,
   `GE_INPUTSCRIPT` walking + intermittent firing near guards, or just ask
   the user to reproduce with `GE_MIXERTRACE=1 GE_AUDIOTRACE=1
   GE_AUDIODUMP=1` set and hand back the logs) aimed specifically at
   reaching the "eventual total silence except one stuck loop" end state,
   then diff voice-slot reuse counts/timing against this session's healthy
   60s baseline to see where it diverges.
3. Listen to `scratchpad/d202-m63/audiodump.raw` directly (raw s16 stereo
   22050Hz PCM) for audible garbling/overlap around the soundIndex=109
   retrigger-storm timestamps — not done this session (no audio playback
   attempted, log-analysis only).
4. Do NOT re-litigate the index/data/decode/ROM-offset chain (M-56–M-62,
   exhaustively verified) — this session's finding is orthogonal to that
   one; D202 is reopened on a *different* mechanism than the M-60 reference
   clip.

**Status at handoff (M-63): D202 REOPENED, root cause not found.** The
M-62 "recommend close" no longer stands — the user's actual complaint is a
runtime mixer/voice issue, not (only) a data-correctness question. One
concrete, unexplained retrigger anomaly (soundIndex 109, ~32.6ms cadence)
is banked as the top lead. No code changes this session — build+run+trace
only; the pre-existing uncommitted diagnostic files (`port/src/audio.c`,
`port/src/mixer.c`, `src/audi.c`, `src/libultra/audio/load.c`, `src/snd.c`)
were exercised, not edited.

### M-65 — Root-caused the "eventual silence" half: an ownerless `SOUND_FLAG_LOOPED` SFX permanently consumes one of only 8 voices. Guard measured. The "wrong/multiple sounds" half gains a second, independent cause (`sndCreatePostEvent` was stubbed out, killing ALL distance attenuation). The audible stuck loop is still NOT reproduced.

**User's report this session:** "wrong sound playing or multiple sounds playing
when there should be one", and "on bunker if the attract shows the door open,
it plays a sound and the sound gets stuck looping forever until you close the
game/level", later refined to "the double door sound by the computer room still
infinitely loops after they begin playing" and "a good chance it just happens
when the level starts if you get that attract intro with them opening".

#### (1) CONFIRMED + MEASURED: the voice leak that produces "eventual silence"

Chain, each link verified rather than argued:

- `soundIndex=203` (`METAL_SLIDE_CLOSE_SFX`) has `envelope->decayTime == -1`.
  Verified at the **source bank level**, not just at runtime, with a new
  `GE_BANKDUMP` probe walking `instArray[0]->soundArray[i]` in the ROM-layout
  blob: `BANKSRC inst@0x005778 soundIndex=203 sound@0x0046B8 env@0x000008
  decay=-1`. `GE_KEYMAPDUMP` additionally dumps every source `ALEnvelope`
  (`ENVSRC src@0x000008 raw=00000000 FFFFFFFF 000007D0`). **Genuine ROM data,
  faithfully converted** — 19 of the SFX instrument's 261 sounds are
  deliberately sustained this way.
- `decayTime == -1` sets `SOUND_FLAG_LOOPED` (`sndSetupSound`, via
  `decayTimeFlag`). A looped sound **never posts `AL_SNDP_STOP_EVT`**
  (`snd.c` `AL_SNDP_DECAY_EVT` skips it), *and* the preemption scan
  **explicitly refuses to steal looped voices** (`!(iterState->unk3e & 0x12)`).
  So the only release is `sndDeactivate`, which needs an owner.
- `doorPlayCloseSound0/1` play it as `sndPlaySfx(..., METAL_SLIDE_CLOSE_SFX,
  NULL)` — **no `pendingState`**, so nothing stores the state and nothing can
  ever deactivate it. (NB the decomp names mislead: `doorStartClose` calls
  `doorPlayOpenSound1`; `doorFinishOpen`/`doorFinishClose` call
  `doorPlayCloseSound0`/`1`. The *Open* pair own their slide loop via a slot;
  the *Close* pair are unowned terminators.)
- `maxSounds` is **8** (`MUSIC_SFX_SEQ_MAYBE_MAX_SOUNDS`, `music.c:76`).

**Measured** (115 s headless `-level_09`, new `[VOICE+]`/`[VOICE-]` probes
pairing every `g_sndAllocatedVoicesCount++` with its decrement):
**7 of the 8 voices permanently held, every one of them `sound=706c8e48`
(soundIndex 203) with `flags=7` (`LOOPED|PLAYING`)**, never released. The
64-entry *state* pool stays healthy (18 slots recycling, never NULL) — which is
why this is silence, not a crash: later `sndPlaySfx` calls get a state but no
voice and are dropped via `sndDisposeSound`.

**Guard (port-side deviation, `src/snd.c` preemption scan).** Allow reclaiming
a voice only when it is `SOUND_FLAG_LOOPED`, not `SOUND_FLAG_RETRIGGER`, and
`state->state == NULL` — i.e. provably ownerless, so nothing could ever have
stopped it. Owned loops keep their existing protection exactly.

| metric (115 s `-level_09`) | before | with guard |
|---|---|---|
| `g_sndAllocatedVoicesCount` | **pinned at 8** | oscillates 1–8 |
| voices acquired over the run | 171 | **629** |
| residual stuck 203 voices at exit | **7** | **1** |

User confirmed the symptom improves ("still happens but less cascading
perhaps"). **This is a game-logic change (AGENTS.md rule #2) justified by
measurement, not by ground truth — it is NOT known why the N64 does not leak
here.** Keep it flagged as a documented deviation, not a silent fix.

#### (2) FIXED, independent: `sndCreatePostEvent` was stubbed out, removing all distance attenuation

`sndCreatePostEvent` returned immediately under `#ifdef PORT` (D138), from when
"libaudio is Phase-3 parked" and nothing drained
`g_sndPlayerPtr->evtq->allocList`, so every positional tick appended an item
that was never consumed and `alEvtqPostEvent`'s ordered insert went O(n^2)
until the kernel-heartbeat watchdog tripped on Facility ambience.

**That premise expired when the Phase-3 audio thread landed (D198-D201)** —
`amMain` now drains the queue every audio frame. Every caller posts type 8 =
`AL_SNDP_VOL_EVT` (`bondview2.c`, `chrai.c`, `gunfire.c`, `propobj.c`), so the
stub had been silently removing **all distance-based volume**: every sound
played at full volume regardless of range. That is very likely a large part of
the "wrong sound / multiple sounds / too loud" impression, independent of (1).

Stub removed. Re-measured: `rt=0.996`, `drop=0`, frame times 91-329 us, zero
watchdog trips over 115 s — the O(n^2) regression does not return.

#### (3) STILL OPEN: the audible endless door loop is NOT reproduced

Not reproduced in any headless run. Everything that could make a sound audibly
repeat was measured and excluded — see the ruled-out list below.

#### RULED OUT this session (measured — do not re-walk)

- **Wave-level looping.** New `[WAVELOOP]` probe: `loop == NULL` on **every**
  sound in the SFX bank. No sample can repeat itself.
- **Decoding past the end of a wave.** New `GE_PULLTRACE` `[PASTEND]` assertion
  in `alAdpcmPull` (fires when `f->sample` exceeds the wave's total samples
  while `nOver == 0`): **0 hits**. `aClearBuffer` zero-fill works; an exhausted
  voice goes silent.
- **Stale / accumulating software DMEM.** `GE_DMEMWIPE` zeroes all 4 KB of
  `sDmem` at the exact frame boundary (`alMainBusPull`'s `AL_MAIN_L_OUT`
  clear is the first opcode of every frame). Audio still 99 % non-silent and
  unchanged — no audio state carries across frames in DMEM.
- **Reverb feedback self-oscillating** (plausible because `aPoleFilterImpl`,
  the damping in the feedback path, is an unimplemented no-op — D199).
  `GE_NOWET` mutes the reverb send entirely: signal unchanged.
- **The `AL_SNDP_PLAY_SFX_EVT` retrigger machinery.** 0 retrigger posts in the
  user's trace. (Also independently confirms D204's closure of the M-63
  `soundIndex=109` anomaly: measured here as 184 plays, **owned 184 / orphaned
  0** — a properly-owned repeated sound, not a retrigger storm.)
- **Doors oscillating.** ~4 open/close cycles per door per 110 s across 4 doors;
  203 played ~once per door state change. Normal, not thrashing.
- **Doors orphaning their loop because both slots are busy.** 0 occurrences
  (`[DOORSND] ... ORPHAN-both-slots-busy` never fired). The orphan is real but
  arises differently: the call site passes `NULL` by design.
- **`ALKeyMap` conversion.** `GE_KEYMAPDUMP` dumps all 220 source keymaps from
  the ROM-layout blob; runtime values match, including entries that are
  legitimately all-zero in the fields Rare repurposed.
- **`DoorRecord` 64-bit overlap.** `sizeof(DoorRecord)=0x128`, observed
  allocation stride exactly `0x128`; `openSoundState@0x110`,
  `closeSoundState@0x118`. No overlap.
- **The `ALSoundState`-pointer-to-pointer pun** in `sndPlaySfx`
  (`pendingState->link.next = nextState`, relied on by `doorPlayOpenSound0/1`
  and `audioPlayFromProp`). Survives the 64-bit transition — `ALLink.next` is
  at offset 0 and pointer-sized on both. New `[SLOTWRITE]` probe confirms the
  slot receives the state.
- **`ALWaveTable::len` drifting at runtime** (observed 6939 -> 6940 etc.).
  Benign: `alLoadParam` rounds `len` down to a whole number of 9-byte ADPCM
  frames on first bind, so the pre-round value is simply seen once.

#### Corrections banked (mistakes made this session — do not repeat)

- **`romdataFixupMusicSeqTable: seqCount 63 exceeds blob capacity 1` is
  BENIGN**, not a music failure. The function is called twice and the first
  call deliberately passes a 16-byte header slice (capacity `(16-4)/8 = 1`);
  the full-size second call succeeds. Music loads and plays.
- **A continuous ~1400 RMS "drone" in `GE_AUDIODUMP` captures is the
  background music, not a stuck sound.** It was briefly mistaken for the bug
  because only *SFX* events carry `dumppos` timestamps, so music looks like
  sound with nothing playing it. Confirmed music by pointer range: the heavily
  rebound wavetables (`706cbaf8`, `706ccfe8`) lie outside the SFX bank's
  `706c3e70`-`706ca430` range and never appear in any `sndPlaySfx`; they are
  driven by `csplayer.c:535` (`alSynStartVoiceParams`). The ~28 k
  `AL_FILTER_SET_WAVETABLE` binds per 60 s are ordinary note-ons, and
  `f->sample` resetting per note is correct.
- **Multiple `audiotrace.log` files exist** (repo root vs `build-pc/`) and a
  shell whose cwd shifts between commands will silently analyse the wrong one.
  Always use an absolute path, and check `ls -la` timestamps before trusting a
  log's provenance.

#### Diagnostics added (all `#ifdef PORT` / env-gated; remove when D202 closes)

| env var | file | what it prints |
|---|---|---|
| `GE_AUDIOTRACE` | `src/snd.c` | `[ENVELOPE]`, `[WAVELOOP]`, `[VOICES]`, `[VOICE+]`/`[VOICE-]`, `[SLOTWRITE]`, `[RETRIGGER-POST]`, keymap fields on every `sndPlaySfx` |
| `GE_AUDIOTRACE` | `src/game/propobj.c` | `[DOORSND]` — every door sound state change, both slots, and an `ORPHAN-both-slots-busy` verdict |
| `GE_KEYMAPDUMP` | `port/src/romdata.c` | `KEYMAP` + `ENVSRC` — source ROM bytes of every `ALKeyMap` / `ALEnvelope` |
| `GE_BANKDUMP` | `port/src/romdata.c` | `BANKSRC` — the source bank tree per `soundIndex` |
| `GE_PULLTRACE` | `src/libultra/audio/load.c` | `[PASTEND]` — asserts no voice decodes past its wave |
| `GE_DMEMWIPE` | `port/src/mixer.c` | zeroes software DMEM at the frame boundary |
| `GE_NOWET` | `port/src/mixer.c` | mutes the reverb send |

**Caution:** the `[VOICE+]`/`[VOICE-]` probes are written from both the game
thread and `amMain` to one unbuffered `FILE*`, so occasional lines are torn.
Parse defensively (filter to well-formed hex) or add a lock before relying on
exact counts.

#### Next step

Needs a user capture at the failing door (computer-room double doors, or a
level start whose attract intro shows them opening) with **both**
`GE_AUDIOTRACE=1` and `GE_AUDIODUMP=1`. The dump allows isolating the stuck
sound numerically and separating it from music without listening, and
`dumppos` pins it to the exact door event.

**Status at handoff (M-65): D202 PARTIALLY resolved.** The "eventual silence"
cascade is root-caused and mitigated (measured); the missing distance
attenuation is fixed; the audible stuck loop is still not reproduced.

### M-66 — the audible stuck door loop IS root-caused: it is faithful N64 behaviour (an original-game quirk), not a port bug

The user capture M-65 asked for arrived (`build-pc/audiodump.raw`, 31,905,536 B
= 361.7 s stereo s16 @22050; `build-pc/audiotrace.log`, same run, `-level_09`
level start whose attract intro shows the computer-room double doors). The
stuck loop is **sound 203 (`METAL_SLIDE_CLOSE_SFX`) doing exactly what the ROM
tells it to do**, and every link of the chain is ground truth:

1. **ROM data:** wavetable loop = `(start=2471, end=7719, count=-1)`;
   `src/libultra/audio/load.c` implements `count` with the explicit comment
   "-1 is loop forever" (decrements only when > 0). Envelope:
   `attack=0 decay=-1 release=2000 aVol=127 dVol=127`.
2. **Game code (byte-matched decompilation):** `doorPlayCloseSound0/1`
   (`src/game/propobj.c`) plays 203 fire-and-forget with a `NULL` owner state.
   The `[SLOTWRITE]` probe confirms the close-sound state is **never stored in
   any door slot**, so the door's stop-sounds path can never reach it, and no
   other code calls `sndDeactivate` on it. Nothing in the game will ever stop
   it except level exit.
3. **sndp (ground truth):** the preemption scan (`src/snd.c` ~line 321) skips
   any voice with flag `0x12` (`SOUND_FLAG_LOOPED|SOUND_FLAG_RETRIGGER`).
   `AL_SNDP_DECAY_EVT` schedules the release→stop envelope "except for a looped
   sound" (`src/snd.c:509`) — a looped voice gets no stop event at all.
   With `decayTime=-1`, `sndCreateSound` sets `state->priority = 0x41`
   (vs `0x40` normal), and libultra's `_allocatePVoice`
   (`src/libultra/audio/synallocvoice.c`) steals only voices with
   `priority <= new voice priority` — so a 0x41 leaked loop is unstealable by
   any ordinary 0x40 SFX or music note.
4. **Pool geometry:** GE's soft limit is 8 (`maxSounds =
   MUSIC_SFX_SEQ_MAYBE_MAX_SOUNDS`, `src/music.c:76`); the physical PVoice pool
   is 24 (`MUSIC_SYN_CONFIG_MAX_P_VOICES = 0x18`, `src/music.c:67`). Leaked
   loops therefore don't deadlock the synth, but they do eat GE's 8-slot soft
   limit — which is where M-65's guard bites.

**Consequence (identical on N64):** every metal-door close leaks one looping
voice that persists until level exit. During attract idle no new sounds arrive,
so nothing ever evicts them and they play forever — exactly the user's report:
"the double door sound by the computer room still infinitely loops after they
begin playing". Every code path in this chain is shared original source
(`snd.c`, `propobj.c`, libultra audio); none of it is port-specific. **This is
an original-game quirk, faithfully reproduced — rule #2 says leave it.**

#### The capture, timeline (corrected)

`dumppos` is the dump-file byte offset; time = `dumppos/88200` s.

- t=0–21 s: level intro — heavy SFX, door cycles on five doors (…7d2c, …7c04,
  …79b4, …7adc, …80a4). 203 plays at **1.3 / 3.3 / 11.3 / 17.2 / 21.0 s**.
- t≈39.1 s: door …80a4 closes → 203 (state …5020). Then a long quiet stretch:
  the same door's next close comes at **t≈360.7 s** (another 203, state
  …4f50, still looping when the capture ends at 361.7 s). The attract demo
  keeps running the whole time — it just gets sparse.
- Leaked 203 voices alive by t≈40 s: four (states …4db0 from 1.3 s, …50f0 from
  17.2 s, …4e80 from 21.0 s, …5020 from 39.1 s). Two earlier ones WERE
  reclaimed — but only because M-65's guard fired under pool pressure:
  `VOICE- snd8e48/st4e18` at t=14.8 s and `st4fb8` at t=16.4 s, each in the
  same block as a new-sound play (the preemption path). Without the guard both
  would persist and the 8-slot limit would fill by ~t=21 s.
- Steady state from t≈39 s: RMS ≈1560, strong periodicity at **327.0±0.1 ms**.

#### The 327 ms mystery (solved — it was the pitch)

The loop segment is 5248 samples; at the bank rate (22050) that is 238 ms, not
327. But the voice does **not** play at unity pitch: keymap `keyBase=54`,
`detune=50`, and `DEFAULT_SETUP_PITCH_SHIFT = -0x1770 = -6000` (`src/snd.c:20`)
give `alCents2Ratio(54*100 + 50 - 6000) = alCents2Ratio(-550)` = **0.72783**.
Loop period at playback = 5248 / (22050 × 0.72783) = **327.007 ms** — matching
the observed 327.0±0.1 ms exactly. The spectral fingerprint agrees: FFT peaks
of the decoded loop segment at 355.3 / 371.4 / 694.4 / 710.6 Hz × 0.7278 →
259 / 271 / 506 / 518 Hz vs observed steady-state lines 256.7 / 272 / 504.7 /
513.8 Hz. Cross-correlating the **pitch-stretched** loop template against the
t=100 s window gives r = 0.835 (the unpitched template gave ~0.10 — which is
why every earlier template test "failed"; see corrections below).

#### Corrections to M-65 (banked)

- **M-65's correction "the continuous ~1400 RMS drone in captures is the
  background music, not a stuck sound" is wrong (or at best incomplete).** The
  steady-state signal's spectral lines and 327 ms period are the leaked 203
  loop, pitch-shifted as above; r=0.835 template match. Music is present in the
  mix too (M-65's pointer-range argument that *those particular* rebound
  wavetables are music-driven still stands) — the steady state is a composite,
  but the *stuck* component is 203, and it dominates the fingerprint.
- **"The audible stuck loop is still not reproduced"** — it is now: user
  capture + complete mechanism, every link grounded in ROM data / byte-matched
  code / libultra source.
- The earlier "full library correlation: best 0.28, sfx203 only 0.10" was an
  artifact of correlating **unpitched** decodes against pitched playback. Do
  not re-run template tests without applying `keyBase/detune/shift` → ratio.
- M-65's guard comment "only 8 voices exist" is imprecise: there are 24
  PVoice plus the 8-slot GE soft limit; the binding constraint for SFX is the
  soft limit, and the guard operates on exactly that boundary.

#### Guard status and decision pending (user)

M-65's ownerless-loop preemption guard (`src/snd.c` ~line 325) fired twice in
this capture and demonstrably prevents pool exhaustion — but it can only fire
when a *new* sound arrives and the limit is hit. During attract idle nothing
arrives, so the audible loop persists with or without the guard. Options:

- **A) Pure fidelity:** remove the guard; document D202 as "faithful
  reproduction of an original quirk". Consequence (also faithful): after 8
  leaked loops in one level session, new SFX are silently dropped until level
  exit — that is what N64 does.
- **B) Keep the guard (status quo, recommended):** a small, precisely scoped,
  documented port-side deviation. Prevents the SFX-deadlock; the idle loop
  remains audible exactly as on N64.
- **C) Extended deviation:** cap/expire ownerless loops in the port layer to
  silence the idle loop. Fixes the user's complaint but is a larger departure
  from ground truth than B.

Diagnostics: with the root cause established, the M-65/M-66 probe set
(`GE_AUDIOTRACE` lines, `[DOORSND]`, `[SLOTWRITE]`, `GE_PULLTRACE`,
`GE_DMEMWIPE`, `GE_NOWET`, `GE_BANKDUMP`, `GE_KEYMAPDUMP`) may be removed once
the A/B/C decision lands and D202 closes.

**Status (M-66): D202 root cause ESTABLISHED — the audible stuck door loop is
sound 203 behaving exactly as the ROM + ground-truth code specify; faithful
N64 behaviour, original-game quirk. Disposition: user chose C (M-66b below).**

### M-66b — disposition C implemented and measured: ownerless infinite-loop SFX now expire (fade out) on PC

The user chose **C** (extended deviation) as part of the PC port's audio
implementation. M-65's guard stays (slot reclamation under pool pressure);
this adds the audibility fix.

**Placement — why `src/snd.c` and not `port/src/mixer.c`.** The predicate that
distinguishes a *leaked* loop from a legitimate one is **ownership**
(`soundState->state == NULL`), which only sndp sees. The mixer iterates
ALVoice/PVoices with no ownership field; scoping there would force either
hardcoding region-fragile SFX-bank DRAM pointer ranges or capping *all*
infinite-loop voices (music included). snd.c already carries the two documented
D202 port-side deviations (M-65 guard, D138 un-stub), all `#ifdef PORT` — every
D202 port-side change stays in one auditable place. N64 build: zero diff.

**Mechanism.** New PC-only event `AL_SNDP_PORT_EXPIRE_EVT` (reuses the
`AL_SNDP_UNUSED_13_EVT` = 1<<13 slot, defined under `#ifdef PORT` in
`src/snd.h`; no shared-enum change). In PLAY_EVT, immediately after a voice
starts, if the state is LOOPED + FINAL_IN_SEQUENCE + not RETRIGGER +
`state->state == NULL` **and** the wave carries an ADPCM loop with `count == -1`
(finite loops self-terminate in load.c and must not be cut short), post the
expire event at `D202_EXPIRE_DELAY_US` (2 s). Handler: re-validate the full
predicate at fire time, then ramp volume to 0 over `D202_EXPIRE_FADE_US`
(500 ms) and post END_EVT — the stock fade/teardown path. Constants are in
**ALMicroTime = microseconds** (the event queue is µs-driven: `DELTA_1_MS` is
1000, `DELTA_33_MS` is 33333).

- **Not a reused STOP_EVT:** the stock STOP ramp uses the envelope's
  releaseTime — ~2.7 ms for sound 203 (release 2000 / pitch ratio 0.7278) —
  which would hard-cut the loop (click). The dedicated event gives a clean
  500 ms fade.
- **Stale-event safety:** `sndDisposeSound` already removes *all* pending
  events for a state (`sndRemoveEvents(..., 0xffff)`), and any rebind disposes
  first — so an expire can never fire against a new binding. The fire-time
  predicate re-check is belt-and-braces (also no-ops if the state was rebound
  to the *same* sound).
- **Sequence-walk safety:** the type is added to the `isEventForSingleSound`
  mask under `#ifdef PORT`, and the post-site requires FINAL_IN_SEQUENCE, so
  the handler's do-while can never walk into an unrelated alloc-list neighbor.
- **Owned loops are untouched** (`state != NULL` fails the predicate) — they
  behave exactly as on N64. Retrigger sounds (managed by PLAY_SFX/DEACTIVATE
  chains) are excluded.

**Files:** `src/snd.h` (PORT-only event define), `src/snd.c` (two constants,
post-site in PLAY_EVT, handler case, mask token, guard-comment cross-ref).
New `[EXPIRE]` line under `GE_AUDIOTRACE` (see GE-ENV-PROBES.md).

**Measured.** 134 s headless `-level_09` run with `GE_AUDIOTRACE=1
GE_AUDIODUMP=1`. Trace: every sound-203 play posts `[EXPIRE]`; each leaked
voice gets its VOICE- ~2 s after its VOICE+; run ends at
`[VOICES] allocated=0 / max=8` (no survivors). Spectral check (per-1-s
Goertzel at 256.7 Hz, the strongest 203-loop line × 0.7278):

| capture | 256.7 Hz line |
|---|---|
| user pre-fix (361 s) | **flat at max from t≈39 s to end** — the complaint, verbatim |
| post-fix (134 s) | transient ~2–4 s bursts per door close, **full silence between** (gaps at t=60–62, 85–87, 94–98, 110–112, 128–131) |

The stuck infinite loop is gone; each close now rings ~2 s + 0.5 s fade, as a
door clunk naturally would. Scripts + archived user capture:
`scratchpad/d202-m66/` (`check_dump.py`, `line_track.py`,
`user-capture-audiodump.raw`, `user-capture-audiotrace.log`).

**Remaining:** user by-ear pass on the new binary (window is open — probes are
still in). On sign-off: remove the D202 probe set (`GE_AUDIOTRACE` additions,
`[DOORSND]`, `[SLOTWRITE]`, `GE_PULLTRACE`, `GE_DMEMWIPE`, `GE_NOWET`,
`GE_BANKDUMP`, `GE_KEYMAPDUMP`) and close D202.

**Status (M-66b): disposition C IMPLEMENTED + measured — ownerless
infinite-loop SFX fade out ~2.5 s after starting on PC; faithful N64 behaviour
otherwise. Awaiting user by-ear verification, then probe removal + close.**

---

### M-67 — Static analysis exhausted: full-bank data scan clean, DSP audit complete, reference clip REFUTED as an in-game capture; new `GE_VOICEDUMP` per-voice probe added for the decisive runtime A/B (this session)

**1. Full-bank SFX data scan — all 261 sounds clean.**
`scratchpad/sfx_bank_scan.py` walks every sound's wavetable/book/data in the
ROM: no out-of-range ADPCM table indices, all books valid (order=2, npred 1–4),
all data within segment bounds, all type-0. The old decoder's silent clamp
(`if tableIndex >= npred: tableIndex = 0`) had been masking this check — with
it removed as a hard error, the whole bank passes. No stale/corrupt-table
hypothesis is possible; the data side of "wrong sample" is closed.

**2. DSP audit complete (vs PD ground truth).** ADPCM scalar decode
(nibble→ins, predictor loop, clamp), the resample algorithm + 256-entry table,
the envelope mixer, and the pitch path (`alCents2Ratio` pure/correct; keymap
formula `keyBase*100 + detune - 6000`, snd.c:803/807) all match the PD port
line-for-line. The `A_LOOP` (flags&2) branch in `aResampleImpl` is **dead code
in GE**: the resampler filter's `first` is only ever 0 or 1 (load.c:461/530);
A_LOOP(2) is never passed to aResample. Not a bug — closed.

**3. The reference clip is refuted as an in-game capture.**
`scratchpad/ref_match_all.py`: `B00I00S2D.wav` does NOT waveform-match any of
the 261 SFX decodes (best single CC = 0.19 ≈ noise; index 45's waveform CC is
0.048), and `ref_match_mix.py` shows no PPK mixture (46+122+ricochet) matches
either (best 0.09). M-61's "+0.999 with index 45" was purely an
**envelope-correlation artifact** (two percussive sounds with similar decay
envelopes). The "N64 plays index 45 for the PPK" premise is dead; the clip's
provenance is unknown and must be asked of the user.

**4. M-63 final-mix cross-check was inconclusive, not exonerating.**
`scratchpad/m63_crosscheck.py`: door events (dumppos 23296/54976, index 203)
show CC≈0 vs offline decode, but those segments are dominated by music
(broadband RMS 2000–8000, no SFX periodicity). CC≈0 in a music-masked final
mix proves nothing about which sample the voice played. A per-voice dump (item
5) or a music-off capture is required.

**5. New probe: `GE_VOICEDUMP=1`** (`port/src/mixer.c`, temporary, D202 diag).
Writes each voice's resampled mono stream — the `aEnvMixer` input, i.e.
post-resample/pre-envelope, per-voice, no music/reverb masking — to
`voicedump.raw` as records `[u32 stateAddr][u32 nSamples][u64 us][s16 × n]`.
The `[WIRE]` lines (load.c `alLoadParam(SET_WAVETABLE)`) and the first
`[AUDIOTRACE] sndPlaySfx` line now also carry `t=<µs>` (sysGetMicroseconds)
so all three streams correlate in time. Offline matcher:
`scratchpad/voicedump_match.py` (each record vs all 261 pitch-resampled
decodes). Self-test: `scratchpad/d202-m67/` (25 s `-level_09`, 5.7 MB dump,
clean parse, records match their sources once the item-6 direction fix is
applied).

**6. Two tooling bugs found in the self-test (my scripts, not the game):**

- **Resample direction was inverted in every offline model so far.** The
  hardware playhead advances `pitch/16384 = ratio` input samples per output
  sample (`aResampleImpl`: `pitchAccumulator += pitch<<1; in += accum>>16`),
  i.e. `y[k] = x[ratio·k]` — ratio < 1 plays SLOWER and lower. The correct
  offline model is `resample_linear(native, RATE*ratio, RATE)`, not
  `(native, RATE, RATE*ratio)` as used in M-63's cross-check and the first
  matcher pass. **All prior CC results at pitch ≠ 1 are suspect** (the M-63
  door check doubly so).
- **WIRE bases are absolute; bank offsets are segment-relative.** The
  in-memory wavetables hold `base = SFXTBL_START(0x102F19A0) + bank offset`;
  matching WIRE entries against raw bank offsets finds nothing.

**7. GE's "music" is an SFX-sequence system with its own sample set.**
`sndNewPlayerInit(ALSeqpSfxConfig*)` (snd.c:168; config in src/music.c:823,
`MUSIC_SFX_SEQ_MAYBE_MAX_SOUNDS = 8`): music notes are sequenced SFX played
through the same 8-voice alSynth pool. The WIRE log shows ~8 music samples
whose bases sit at ROM 0x103ed1e8–0x10407860 — 240–350 KB **past** the last
SFX-bank sample — with wavetable structs allocated contiguously after the SFX
bank's. Music notes do not pass through `sndPlaySfx`, so they are invisible to
the `[AUDIOTRACE]` request probe: persistent voices in a voicedump are usually
music, and their "best match" against the 261-SFX bank is coincidental
(music samples reuse field recordings; free-ratio sweep found e.g. CC 0.92 vs
index 122 at transposed pitch). Any per-voice analysis must subtract music
first (or use the timestamped WIRE assignments to identify each voice's sample).

**8. Open anomaly parked for the next session: sound 232 (PICKUP_GUN_SFX)
requested at `-level_09` start produced no audible audio.**
Self-test trace: `sndPlaySfx(232)` at dumppos=0 → VOICE+ (count 1/8) →
VOICE- within ~1 frame; the WIRE-assigned filter was recycled to music
immediately; no env state in the dump matches decode(232) (max CC 0.48 ≈
noise, even with the corrected resample direction). Either the level script
stops it instantly (faithful N64 behaviour — needs an N64 A/B to confirm)
or something disposes it early on PC. Not chased this session.

**Status (M-67): static analysis has hit diminishing returns — every pipeline
stage (index resolution, bank data integrity, ADPCM, resample, envelope,
pitch, loop mechanism) is verified against ROM + PD ground truth, and the
reference clip is refuted. The remaining questions (PPK/door/armor "wrong
sound" by ear) require a user runtime A/B capture with the new
`GE_VOICEDUMP` probe: `GE_AUDIOTRACE=1 GE_AUDIODUMP=1 GE_VOICEDUMP=1`,
reproduce each scenario, send `audiotrace.log` + `audiotrace_wire.log` +
`voicedump.raw`. Also owed: provenance of `B00I00S2D.wav`, and the item-8
sound-232 anomaly.**

### M-68 — Decisive runtime A/B (user captures, 3× level_09): all three complaint classes play the exact ROM samples at the exact ROM pitches; no port bug found (this session)

Three interactive user captures (`audiodebug.ps1 -Play -Trace -Dump
-VoiceDump`), analysed with full-slide correlation of `voicedump.raw`
per-voice records against `decode_sound(idx)` resampled at the keymap ratio
(correct direction, M-67 item 7: `y[k] = x[ratio·k]`). A new `[EVT] t=µs
type=N state=P` probe (PORT-gated, `sndHandleEvent` top) plus a
`[STOP-EVT] … deltaUs=N` probe (release-ramp site) were added this session;
event types decode per `src/snd.h`: 1=PLAY, 2=STOP, 8=VOL, 64=DECAY,
128=END.

**1. Doors — correct.** `METAL_SLIDE_OPEN` 202: CC **0.957** @ ratio 0.500;
`METAL_SLIDE_CLOSE` 203: CC **0.906** @ 0.728; `METAL_SLIDE_LOOP` 204: CC
**0.993** @ 0.728. Event flow is textbook: PLAY → DECAY (attack=0, fires
immediately) → self-stop at the decay-scheduled STOP (204: +154 ms =
112197 µs ÷ 0.7278; 202: ~450 ms) → release ramp (`[STOP-EVT] deltaUs` =
releaseTime ÷ ratio, e.g. 1373/3999 µs) → END → dispose.

**2. PPK shot — correct.** `GUN_SILPPK_A` 46: CC **0.90–0.97** @ ratio 0.728
across four separate shots (full-slide, any offset). Combined with M-59
(46 + 122 CART_SPENT + a ricochet is exactly what the ROM scripts for a PPK
hit), the user's "gunshot plays broken glass/slap/ricochet" is the original
design, reproduced faithfully.

**3. Body armor pickup — correct, full length.** `ARMOUR_COLLECT` 81:
voice lifetime measured **833 ms** from `[EVT]` timestamps (PLAY → decay-
scheduled STOP), against ROM prediction 469649 µs ÷ 0.5612 = **837 ms**;
release ramp 3563 µs; CC **0.828** vs decode(81) @ ratio (window contains
other concurrent voices, hence <0.95).

**The M-67-era "instant death" metric was a line-count artifact.** Log lines
≠ time: a voice stopped by the door script's `stop-sounds`
(`door7F053B10`, `src/game/propobj.c` — stock logic, identical on N64) when
the player shuts a door quickly produces only a few log lines between birth
and death even when it plays its full ~20 ms. With µs timestamps the STOPs
land exactly at their scheduled times; no voice is disposed earlier than its
decay/stop schedule in any of the three captures.

**13 `[EXPIRE]` events in the final capture:** the M-66b guard expired 13
ownerless `METAL_SLIDE_CLOSE_SFX` (203) loops, each at exactly 2.5 s (2 s
delay + 0.5 s fade). This remains the single outstanding PORT deviation;
whether the N64 lets that loop run until level exit is the open by-ear A/B
(see M-66b).

**Status (M-68): D202's "wrong sample" hypothesis is closed negative for all
three user complaint classes — PC plays exactly what the ROM specifies,
verified per-voice at runtime. Remaining: (a) N64 A/B on door-close-loop
duration to confirm or retract the M-66b cap; (b) the parked sound-232
anomaly (M-67 item 8); (c) once the user confirms by ear, remove the D202
probe set (`GE_AUDIOTRACE`/`GE_AUDIODUMP`/`GE_VOICEDUMP` probes in
`src/snd.c`, `src/libultra/audio/load.c`, `port/src/mixer.c`,
`src/game/propobj.c`) and decide the M-66b guard's fate.**

### M-69 — Full-corpus exact match (scripted `-level_09`, 320 SFX requests): 286/320 OK after tooling fixes; the remaining failures are *dropped allocations*, not corruption (this session)

New offline tool `scratchpad/exact_match.py` (gitignored scratchpad; helpers
`rom_sfx_decode.py`, `m63_crosscheck.py`) matches **every** `sndPlaySfx`
request in a full trace to its voice and slides the ROM-decoded PCM
(resampled at the keymap ratio, M-67 item 7 direction) against that voice's
`GE_VOICEDUMP` stream. Corpus: `scratchpad/exact_run/` (scripted `-level_09`,
61 input entries, `GE_AUDIOTRACE` + WIRE + `GE_VOICEDUMP`) — **320 requests**.
Chain: request → its own `[EVT] type=1` PLAY (states are recycled across
requests, so the wire window must anchor on *that* PLAY) → WIRE record with
the same wavetable in [play−10 µs, play+50 ms] → decoder index → env-state
slot address → voicedump records for that slot.

**Result: 286/320 OK (CC ≥ 0.85).** The initial 241 OK / 35 MID / 44 BAD split
was almost entirely tooling artifacts, not audio:

1. **`GE_VOICEDUMP` records are rendered ahead — the µs stamp is the END of
   the block.** Each record's n samples end at its stamp, spaced 1/22050 s.
   Treating stamps as start offsets misplaces content by up to one block and
   produced a phantom "stale content" in pv20's slot (sound 29 looked ~180 ms
   early). With per-sample record times, sound 29's onset lands at **+263.8 ms
   vs its WIRE timestamp +264 ms — exact**. Cross-slot correlation confirmed
   the dump attribution was never wrong (max CC vs neighbouring slots ≤ 0.076;
   slot content is unique per voice).
2. **`slide()` needed step-1 refinement around the coarse best hit.** Coarse
   stepping misses sharp onsets: sound 29 measured CC 0.590 at the coarse
   argmax and **0.970 one sample later**. (Also fixed a bug: the best-offset
   tracker was only updated in the last coarse loop, so refinement searched
   around offset 0.)
3. **Clip the reference at voice death.** Requests whose voice is cut by a
   DEACTIVATE/STOP after PLAY (e.g. idx=109, killed at +131 ms) cannot match
   the full ROM decode (CC 0.016); clipping the ref to the live window gives
   CC **0.999** at offset +0 ms. The matcher now parses `[EVT]` per state and
   clips accordingly.

**The remaining ~34 failures are not corruption:**
- **"NO WIRE AT PLAY" / DROPPED (no PLAY evt):** the request was processed but
  `alSynAllocVoice` never gave it a voice (no `[VOICE+]`, no DECAY) — e.g.
  t=567204976 idx=46. Leading suspect: the 8-voice pool is exhausted under
  scripted rapid fire and low-priority sounds are dropped by design (`snd.c`
  threads `state->priority` into `config.priority`; `alSynAllocVoice`'s steal
  path has a ~512-sample delay). **Not yet verified — the next step.**
- **pv10 mismatches:** consecutive requests share wavetables, so the wire
  window can pick up the *next* request's wire and attribute the wrong slot.
  Tighten: require the WIRE to fall between this PLAY and the next PLAY on the
  same wavetable.
- Log-interleaving gotcha: AUDIOTRACE lines from multiple file handles
  interleave and corrupt some lines (a missing/corrupt PLAY line sends the
  matcher to its wide-window fallback).

**Status (M-69):** every request that *got* a voice plays the exact ROM sample
at the exact ROM pitch — this extends M-68's per-class verification to the
whole corpus. D202 remains "no port bug found"; open items are M-68's plus:
(a) confirm the dropped requests are pool-exhaustion drops that N64 would drop
identically (check `[VOICES] allocated=N / max=8` around each drop;
`src/snd.c:916`, `alSynAllocVoice`); (b) fix the shared-wavetable wire
attribution; then M-68's by-ear + N64 A/B, probe removal, close.

### M-70 — Decisive full-corpus exact match on a clean re-capture (209 requests, scripted `-level_09`): every voiced request plays the exact ROM sample; D202's wrong-sample hypothesis is refuted (this session)

The prior 316-request corpus was contaminated mid-capture (pause menu hit),
so a fresh 75 s scripted capture was taken (`GE_AUDIOTRACE=1 GE_VOICEDUMP=1`,
`-level_09`, `scratchpad/exact_run3_script.txt` fire burst) and the matcher
(`scratchpad/exact_match.py`) was fixed in three places before the decisive run:

1. **Death-clip both sides.** Clip *stream* as well as *ref* at the voice's
downstream death event (STOP/END/DEACTIVATE). M-69 only clipped `ref`; when a
slot is reused ~200–400 ms later, the unclipped 1.5 s stream window is mostly
the *next* occupant's samples.
2. **`slide()` sparse-tail bug (the big one).** The offset search was dense
only up to 9000 samples (~408 ms) and used step = span/40 beyond that. Delayed
voicing — `sndPlaySfx` queue wait, observed up to **+1.2 s** between request
and PLAY — pushes the onset into the sparse zone, where the narrow CC peak
falls between grid points: idx=134 (onset +532 ms) scored **CC 0.154** in the
matcher but **0.973** under a full-range step-4 scan of the identical data.
   Fixed: step-1 near the window start, step-4 across the *full* span, ±8
   step-1 refine around the best hit. idx=136 went 0.285 → ≥0.85 the same way.
3. **PLAY search window 600 ms → 2 s**, catching two idx=175 instances voiced
   at +1196 ms (their WIREs land with the PLAY, so pairing is unambiguous).

**Result — 209 requests:**

| bucket | count | disposition |
|---|---|---|
| OK (CC ≥ 0.85) | **200** | exact sample at exact keymap pitch |
| MID (0.6–0.85) | 4 | all idx=109, CC 0.744–0.811 — see below |
| NO WIRE AT PLAY | 5 | 1 matcher artifact + 4 never voiced — see below |

- **The 4 MID (idx=109):** manual verification on the tight `[play, death]`
   window with a ratio sweep gives **CC 0.946 at the nominal keymap ratio**
   (0.7236 = keyBase 54, detune +40, −6000 shift) — correct sample *and*
   pitch. The residual is measurement, not audio: the voicedump records are
   stamped with **wall-clock time at mix time**, and the scheduler's catch-up
   delivers them in ~32 ms bursts of ≤160-sample records with clustered
   stamps, so bisect windowing on those stamps is fuzzy by tens of ms — and
   these voices die young (killed at +67…+131 ms), leaving little clean
   signal to average over.
- **NO WIRE ×5:** idx=109@610342131 was in fact voiced — its WIRE lands at
   **+133 ms** after PLAY, outside the matcher's 50 ms wire window (all other
   idx=109 instances match fine). The other four (idx 8/40/123/202) show no
   voicing evidence anywhere: no `VOICE+` on their `newState` after t0, and no
   WIRE for their wavetable within 2–3 s *in any state*. **Not pool
   exhaustion** — `[VOICES] allocated` was 0–5/8 at each drop (run max 7/8).
   Mechanism unverified; consistent with game-logic queue/priority timing in
   code identical to the N64 build, and orthogonal to D202 (no samples were
   played, so nothing could be wrong).

**Verdict:** across 204 voiced requests spanning ~30 distinct sound indices,
**zero** play a wrong sample. Combined with M-68 (per-class runtime A/B) and
M-66 (stuck door loop = faithful N64 quirk), the D202 "wrong sample / mixer
corruption" hypothesis is refuted: what the user heard is the original design
(overlapping SFX, short decays, quick-stop truncation, the ownerless infinite
door loop).

**Status (M-70):** D202 data-closed. Remaining before final close + probe
removal: user by-ear A/B on the M-66b expiration behaviour (does the stuck
drone now fade out ~2.5 s after a door close, vs N64's forever-drone). Probes
(`GE_AUDIOTRACE`, `GE_VOICEDUMP`, `[WIRE]`, `[EVT]`, `[VOICE±]`, `[SLOTWRITE]`,
`[ENVELOPE]`, `[VOICES]`) stay in place until that pass is done; they are
`#ifdef PORT`-gated and env-var-off by default.

### M-71 — User by-ear pass on M-66b: PASS. D202 ready to close. New complaint class split out as D205 (this session)

**By-ear verdict (user, PC vs N64 side-by-side):** "the metal door sound
does not loop forever now" — the M-66b ownerless-infinite-loop expiration
behaves as intended; Disposition C validated. That was the last gate on
D202: probe removal + close is all that remains (see HANDOFF).

**New user complaint class (NOT D202 — split out to D205).** With M-70's
per-voice exact-match result standing (every voiced request plays the exact
ROM sample), the user reports by ear, and confirms against an N64 A/B that
the PC is *not* faithful: (1) general gunfire has a slap/glass-shatter
layered over the correct gunshot — audible on PC, silent on N64 for the
same shot; (2) an explosion plays a scream-like sample (the soldier-death
voice); (3) armor pickup plays a wrong sample. This session exonerated the
two remaining static suspects: the D37 bank converter offline (`scratchpad/
banktest/`: real `romdataFixupAudioBank` harness + `diff_bank.py` → all 261
entries map identically N64→PC, 0 mismatches) and the D154/D135 GBI raycast
parsers + volume law (static re-audit, ABI-correct). The remaining
explanation is that PC **requests different or extra sound indices** than
N64 in the same scenario. See D205.

### M-72 — Fourth symptom reported (unarmed melee → Klobb shot idx 106); controlled `GE_AUDIOTRACE` runs exonerate every request path; mechanism (b) now the leading suspect (this session)

User added a fourth D205 symptom: Bond's unarmed slap attack plays
Klobb's shooting sound (`GUN_B1_MGUN3_3_SFX`, idx 106) instead of the slap
effect. Also fixed the test environment: the "flaky /GS crash" of earlier
attempts was `SDL2.dll` missing from PATH (it lives in
`C:\msys64\mingw64\bin`; the exe is `-mwindows`, so a failed DLL load just
dies with 0xC0000105 and no console). Launchers now prepend it.

Controlled traces (`build-pc/d205_*.ps1`, gitignored): PPK fire at
`-level_09` requests exactly {46, one of the 20-entry ricochet table,
122} per shot — no glass/slap layer; unarmed fire-presses at levels 01–08
(all start unarmed) request only 105 PUNCHING_AIR. Static: punch hit =
constant {47,48,49}, idx 106 is data-only (`skorpion_stats.Sound`), and the
player fire-sound path cannot run for a fist — so Bond's own punch cannot
request 106; an NPC guard firing its Skorpion can. No symptom reproduces in
micro-scenarios → the layering must come from **extra events during real
gameplay** (mechanism b). Next: user `GE_AUDIOTRACE=1` capture at the exact
failing moment (splits per-symptom), then controlled explosion + armor.
Full detail in D205.

## D207 — **Alarm klaxon starves combat SFX (surfaced by D206)** (M-77, root-caused, fix designed & not applied)

### How D206 caused it

`ALARM3_SFX` (SFX_ID 163) is played by `handle_alarm_gas_timer_calldamage()`
(`src/game/propobj.c:14457`) whenever `alarmIsActive()`, owner `&ptr_alarm_sfx`.

- **Pre-D206:** resolved to on-disk bank slot **163** — a **one-shot** sample
  (`decayTime` finite, no ADPCM loop). Played once, released its voice. The
  alarm was audibly the *wrong* sound but cost the voice pool nothing.
- **Post-D206:** correctly resolves to slot **162** — the real klaxon:
  `envelope.decayTime == -1` (→ `SOUND_FLAG_LOOPED`, `sndSetupSound`
  `snd.c:~836`), wavetable `adpcmWave.loop.count == -1` (infinite ADPCM loop).
  It now holds a voice for the whole time the alarm is up.

### Why it dominates

The SFX player has **8 voices** (`MUSIC_SFX_SEQ_MAYBE_MAX_SOUNDS = 8`,
`src/music.c:76`; the synth has 24 physical voices but the SFX soft cap is 8
and is what `snd.c:328` gates on).

`sndSetupSound` sets `state->priority = (decayTime == -1) + 0x40` → the alarm
gets **0x41**, one-shots get **0x40**.

When `g_sndAllocatedVoicesCount == 8`, `snd.c` PLAY_EVT does **not** call
`alSynAllocVoice` (which *would* do libultra priority-stealing); it runs the
manual preemption scan at `snd.c:~347-392`. That scan:
- refuses to steal any `unk3e & 0x12` voice (LOOPED | RETRIGGER) — the M-65
  carve-out only exempts *ownerless* loops (`state == NULL`), and the alarm
  is **owned**;
- ignores priority entirely — it steals the first non-looped PLAYING
  non-PREEMPT voice it finds;
- if it finds none → `sndDisposeSound(soundState)` → **the new SFX is
  dropped** (`snd.c:391`).

So an active alarm removes one of the 8 voices, and in a firefight (gunshots +
ricochets + impacts + enemy fire + ambient door/machine loops) the remaining
non-looped voices are all either already PREEMPT-marked from the same burst or
themselves looped → gunshots drop.

### Measured (BUNKER1, `-level_09`)

Temporary probe `GE_FORCEALARM=1` (added to `handle_alarm_gas_timer_calldamage`,
`#ifdef PORT` / `getenv`) pins `alarm_timer` so the klaxon stays up, plus a
scripted `GE_INPUTSCRIPT` firing Z every 8 reads:

| condition | SFX requests | dropped (no voice) |
|---|---|---|
| no alarm | 52 | 1 (~2%) |
| alarm active | 137 | 28 (~20%) |

Dropped indices were gunshot/impact/ricochet sounds (29, 30, 41, 90, 36, 109…).
`ptr_alarm_sfx` stayed stable while active and was cleared on
`alarmDeactivate()`; the alarm voice was freed and the pool recovered — the
**"doesn't recover after the alarm stops" half of the user report was NOT
reproduced** in any trace (light BUNKER1 combat; may need a heavier level, or
it is perceptual).

### The 8-cap and the scan are faithful decomp

`MUSIC_SFX_SEQ_MAYBE_MAX_SOUNDS = 8` is unmodified game code; the manual
preemption scan is stock `sndHandleEvent` (M-65 modified it in place). Upstream
GE has the same. So this is not purely a port regression — it is partly the
open **D202** voice-pool-margin problem (ambient infinite-loops holding voices
longer than they should on the port) becoming audible because the alarm now
legitimately consumes a slot.

### Designed fix (PORT, not yet applied)

In the `snd.c` preemption scan, add a **last-resort second pass**: after the
normal pass (non-looped victims + the M-65 ownerless-loop carve-out) finds
nothing and `limitReached` is still true, and **the incoming `soundState` is
itself not LOOPED**, run one more pass that will preempt a stoppable LOOPED,
non-RETRIGGER voice (the alarm klaxon; also the D202 door loop as a bonus).

Self-healing: preempting the alarm posts `AL_SNDP_END_EVT` → `sndDisposeSound`
→ `sndUnlinkClearSound` clears `state->state->link.next`, i.e. sets
`ptr_alarm_sfx = 0`; the very next frame `handle_alarm_gas_timer_calldamage`
sees `ptr_alarm_sfx == 0` and re-issues `sndPlaySfx(ALARM3_SFX, &ptr_alarm_sfx)`,
which re-acquires a voice (or WAIT_VOICE-retries at 33 ms). Net effect: the
klaxon briefly yields to a gunfire burst and comes right back — gameplay audio
wins, which matches player expectation.

Alternatives weighed and rejected for now: (a) raise `maxSounds` > 8 on the
port — game-code change, rule #2 risk, changes mix loudness; (b) properly fix
the D202 ambient-loop lifetimes — the correct fix but the hard open problem.

### M-77b — designed fix tried, hypothesis not confirmed, reverted

The last-resort LOOPED-preemption pass was implemented in the `snd.c` scan,
plus `[D207-DROP]` counters at both no-voice `sndDisposeSound` sites and a
`[D207-YIELD]` trace on the new pass. BUNKER1 test (`GE_FORCEALARM=1` +
`GE_INPUTSCRIPT` firing every ~10 controller reads for ~85 s, alarm cycled
on/off/on):

- **`[D207-DROP]` = 0** — not a single SFX actually dropped.
- **`[D207-YIELD]` = 0** — the new pass never even ran; pass 1 always found a
  non-looped voice to preempt.
- Per-frame voice count: **1** during quiet alarm, spikes to 3–5 under fire,
  back to **0** after every alarm-off. No leak, no pegged pool.
- The earlier "≈20 % dropped" number was bogus: it counted requests whose
  `[VOICE+]` came more than 40 log-lines later. Cross-checking "states that
  never got a voice at all" = **0 / 136**. Non-looped sounds get 2 `unk38`
  retries then the scan, which nearly always places them a frame or two late —
  audible as slightly loose timing, not silence.

So on BUNKER1 the alarm does **not** starve the pool. The change was reverted
(`snd.c` now carries only the D206 `-1`). Either BUNKER1 combat is too light to
reach 8/8, or the user's symptom is a different mechanism.

### Owed next session — needs a real repro from the user

The BUNKER1 forced-alarm harness cannot reproduce this. Get from the user:
**which level, which weapon/situation**, and a `GE_AUDIOTRACE=1` capture (or a
short clip) of the failing alarm. Then decide between:

1. **Genuine 8/8 starvation on a heavy level** — Facility / Silo / Statue
   firefights have many more guards firing. Soak one with `GE_FORCEALARM` +
   sustained scripted fire (or the user's own capture) and watch for
   `[D207-DROP]` > 0 / `count=8` pinned. If real, the reverted last-resort
   pass is the fix — re-apply it.
2. **A leak on a specific alarm-stop path** — the `alarm_timer` auto-timeout
   (`CHROBJ_GAS_TIMER`) vs `AI_AlarmOff` vs `deactivate_alarm_sound_effect`
   from a cutscene. Trace `g_sndAllocatedVoicesCount` across each.
3. **Perceptual** — the klaxon is a *continuous* infinite loop (`slot 162`,
   no gaps). A real GE alarm pulses. Check: is `ALARM3` supposed to retrigger
   (keyMap velMax=7 was noted but `SOUND_FLAG_RETRIGGER` needs `keyMax & 0xf0`,
   and keyMax=4 → not set)? And compare slot-162 playback **volume** to N64 —
   if the drone is just mixed loud it will mask everything without dropping a
   single voice.

**Temp probe kept:** `GE_FORCEALARM` in `handle_alarm_gas_timer_calldamage`
(`src/game/propobj.c`, `#ifdef PORT` / `getenv`).


## D206 — **PC played every SFX one bank slot too high: `ALInstrumentAlt_s.soundArray` shifts offset 12→16 at 64-bit** (M-74 found, M-76 FIXED)

**This is the root cause of D205's melee→Klobb and armor-pickup symptoms, and
of D202's original "silenced PPK plays a slap" report.** It was invisible for
eight sessions because *the verification tooling shares the bug* — see
"Why M-68/M-70 passed" below.

### The evidence

The user supplied an independently-produced, verified rip of GE's sound
effects (`B00I00S<hex>.wav`, 186 files). Filenames are **hexadecimal** bank
indices. Cross-checked by PCM frame count against `scratchpad/rom_sfx_decode.py`
decoding the ROM's `sfxctl` bank:

```
offset -2:   0/184 matches
offset -1:   0/185 matches
offset  0: 186/186 matches   <-- rip index == bank index, exactly
offset +1:  21/185 matches
offset +2:  16/185 matches
```

So the rip is ground truth for *what sound lives at each bank index*, and our
decoder's indexing agrees with it. The user then identified three by ear:

| user's statement | rip file | = bank index | game requests (from `audiotrace.log`) |
|---|---|---|---|
| correct body-armour sound | `S50` | **80** | — |
| what the PC port actually plays for armour | `S51` | **81** | **81** (`ARMOUR_COLLECT_SFX`) |
| the actual Klobb sound | `S69` | **105** | **105** (`PUNCHING_AIR_SFX`, punch whiff) |

Both are exactly `requested index → played bank[index]`, and in both cases the
**correct** sound sits at `index - 1`.

### The structural confirmation

- ROM `sfxctl` bank: `bankCount=1`, `instCount=1`, **`soundCount=261`**,
  `percussion=NULL` → valid `soundArray` indices are **0…260**.
- `SFX_ID` (`src/bondconstants.h`) has **262** members, beginning
  `NOTHING_SFX = 0`, which `snd.c:992` explicitly treats as "no sound"
  (`if (soundIndex == 0) return NULL`).
- **262 enum members − 1 sentinel = 261 real IDs (1…261), mapping exactly onto
  bank slots 0…260.** The arithmetic only closes with a `-1`.
- Corollary bug: the last ID, `BIG_CLANK_SFX` = 261, currently indexes
  `soundArray[261]` — **one past the end of a 261-entry array.**

`src/snd.c:1001` is unmodified decomp code:

```c
sound = (soundBank->instArray[0]->soundArray[soundIndex]);   /* needs - 1 */
```

### Independent confirmation from D202's own reference clip

`scratchpad/B00I00S2D.wav` — the silenced-PPK reference clip that drove D202
from M-60 to M-67, and whose provenance the user could only describe as "not
sure / found online" — **is a file from this same rip**: `S2D` = **bank 45**
(3264 frames, identical). And `wppksil_stats.Sound = 0x2E` = **46**.

So the silenced PPK's ROM data value 46 must resolve to bank **45**, and the
user's clip *is* bank 45. **M-61 found precisely this** ("index 45 is the only
strong match for the user's reference clip, envcorr +0.999; index 46 is the
'slap'") and it was argued away over four sessions — M-67 "refuted" the clip as
a genuine capture on a waveform-correlation test. The +0.999 was right. The
"slap" heard on every silenced PPK shot is `bank[46]`, which under the correct
mapping is `PUNCH1_SFX` — **literally a slap.**

### Why M-68/M-70's "exact match" passed anyway

`scratchpad/exact_match.py` verifies a runtime voice against
`rom_sfx_decode.py`'s decode of **the requested index**, and that decoder walks
`soundArray[i]` with the same missing `-1`. `scratchpad/banktest/diff_bank.py`
likewise compares an N64-layout walk to a PC-layout walk using the same
indexing on both sides. **Every tool built to check the mapping reproduced the
mapping.** M-68/M-70 proved the port plays *the sample it asked for, at the
right pitch* — self-consistency — and never tested whether the index asked for
was the right one. Method note added to `porting-notes.md` §E.

### What it explains

- **D202** — "silenced PPK plays a slap": requests 46, plays `bank[46]` =
  `PUNCH1`; correct is `bank[45]`. ✔
- **D205 (4)** — melee plays the Klobb gun: whiff requests 105, plays
  `bank[105]` = **the Klobb sample**; correct is `bank[104]`. ✔ user-confirmed
- **D205 (3)** — armour pickup wrong sample: requests 81, plays `bank[81]`;
  correct is `bank[80]`. ✔ user-confirmed
- **D205 (1)** — slap/glass layered over gunfire: consistent (a gunshot ID
  landing on the adjacent impact/punch sample), not yet individually confirmed.
- **D205 (2)** — explosion → scream: **NOT explained.** Explosion IDs 169–183
  map to bank 168–182, all still explosions. Stays open; M-73's reading
  (guards killed by the blast screaming, ±1 s) remains the likely answer.

### RESOLVED (M-76) — pointer-width layout shift in `ALInstrumentAlt_s`

The M-75 reframe below was on the right track (candidate B — a PC-side layout
mismatch, not a code or data `-1`). The exact mechanism:

`sndPlaySfx` (`snd.c:1001`) reads the on-disk sound bank — a **standard**
`ALBankFile`/`ALInstrument` (verified: `rom_sfx_decode.py` finds `soundCount`
= 261 at instrument offset 14, standard `ALSound*` table at offset 16) —
through GE's cast alias:

```c
struct ALInstrumentAlt_s {      /* src/snd.h:165 */
    s32 unk0;   /* off 0  */
    s32 unk4;   /* off 4  */
    s32 unk8;   /* off 8  */
    ALSound *soundArray[1];     /* N64: off 12 (4-byte ptr) | PC: off 16 (8-byte ptr, 8-aligned) */
};
```

- **N64:** `soundArray` at struct offset **12** — it overlaps the on-disk
  instrument's `bendRange`(12) + `soundCount`(14) words. So `soundArray[0]`
  is those two `s16`s (a bogus pointer, but `SFX_ID 0 = NOTHING_SFX` is
  short-circuited at `snd.c:992` and never dereferenced), `soundArray[1]` is
  on-disk table entry `[0]`, and in general **`soundArray[N]` == on-disk
  entry `[N-1]`**. GE's `SFX_ID` enum and every `WeaponStats.Sound` byte are
  therefore **1-based** into the sound table, by design.
- **PC:** 8-byte pointers + 8-byte alignment insert 4 bytes of padding after
  `unk8`, so `soundArray` sits at offset **16**. `port/src/romdata.c`
  `afFixupInst` rebuilds the bank to exactly this layout (16-byte header,
  `8*soundCount` bytes of 64-bit pointers from offset 16, entry `i` = on-disk
  entry `i`). So PC `soundArray[N]` == on-disk entry `[N]` — **one slot
  higher than N64** for every SFX.

This is the D3x class (a ROM-serialized struct whose pointer-width field
shifts the layout at 64-bit). **Fix:** `src/snd.c:1001`, `#ifdef PORT` branch
→ `soundArray[soundIndex - 1]`, N64 line kept verbatim under `#else`. The one
read site covers both the initial lookup and the retrigger chain (the loop
recomputes `soundIndex` from `keyMap->velocityMin`, same 1-based space). No
game-logic change; `romdata.c` untouched (its layout is internally correct —
it just wasn't the layout the N64 struct alias assumes).

### Open before fixing — where the fix belongs (M-75 reframe) — SUPERSEDED by M-76 above

**M-75:** upstream `n64decomp/007` `src/snd.c` byte-matches the real ROM,
carries no `NON_MATCHING`/`GLOBAL_ASM` guard on `sndPlaySfx`, and its index
line is identical — `soundArray[soundIndex]`, **no `-1`**. Blame puts
`snd.c:1001` in the 2022 upstream import, not this fork. So the earlier
hypothesis "(i) real ROM does `soundArray[soundIndex-1]`, decomp dropped it"
is almost certainly **dead**: a matching decomp would contain the `-1`. The
N64 runs `soundArray[N]` for ID `N` and that game sounds correct.

That forces a reconciliation. Candidates, cheapest test first — full decision
gate and per-candidate byte-level experiments in
`docs/dev/notes/D206-IMPL-BRIEF.md`:

- **(A)** the rip `B00I00S<hex>` is keyed by `SFX_ID`, not by 0-based bank
  slot → there is **no engine bug**, the PC audio path is faithful, and
  D205 / D202's slap need another cause. Re-audit whether
  `scratchpad/rom_sfx_decode.py`'s 0-based `sound_offsets` walk actually
  models the runtime `soundArray` (leading NULL? offset-table vs `soundCount`
  mismatch? sfxctl not a plain `ALBankFile`?). Cheapest — do first.
- **(B)** PC-only off-by-one in the *converted* bank — `port/src/romdata.c`
  `afFixupInst` drops/adds one `soundArray` entry vs the ROM. Test: dump PC
  runtime `soundArray[81]` wavetable `base`/`len` (existing `GE_AUDIOTRACE`
  probe at `snd.c:1002`) and byte-compare to `rom_sfx_decode.py` bank 80 vs
  81. PC-side fix, rule-#2 clean.
- **(C)** PC-only off-by-one in the `soundIndex` reaching `sndPlaySfx`
  (request path / miscompiled or region-mismatched enum). Test: `GE_AUDIOTRACE`
  already logs `soundIndex=` at the call — confirm armour logs `81`, not `82`.
- **(D)** genuine data-side: ROM `SFX_ID` constants + `WeaponStats.Sound`
  bytes are each +1 vs a 0-based bank. Only reachable if A/B/C all fail.
  ~532 enum refs + debug-name table + weapon-stats asset bytes — **not** a
  legal port edit under AGENTS.md rule #2; stop and escalate to the user
  (possible "wontfix / faithful").

M-62's caveat still applies: this repo never builds or byte-verifies the N64
target, so "it byte-matches" cannot by itself settle (D).

### Verification done (M-76)

- Build clean (`build-pc.sh ntsc-final`).
- BUNKER1 (`-level_09`) `GE_AUDIOTRACE=1`, 60 s: crash-free, 446 `sndPlaySfx`
  calls. Every one of ~90 distinct requested `soundIndex` values resolves to a
  wavetable whose `len` matches `rom_sfx_decode.py` slot **`soundIndex - 1`**
  (0/90 match slot `soundIndex`). Includes `ARMOUR_COLLECT_SFX`=81 → slot 80,
  `GUN_B4_BOLTACTION_SFX`(AK47)=109 → slot 108, `GUN_SILPPK_A`=46 → slot 45.
- `-level_09` golden framediff (`200-440:120`): **3/3 within threshold** on two
  consecutive runs (an audio-path change cannot affect rasterisation; ritual).

### Still owed

User by-ear A/B vs N64: armour pickup, unarmed melee whiff, silenced PP7 fire.
If clean → **D202 closes outright**, D205 shrinks to symptom (2) (explosion→
scream) alone. Then remove the temporary D202 `GE_AUDIOTRACE` / `[VOL]` /
`[DISTVOL]` probe set (checklist in HANDOFF).

## D205 — Wrong/extra SFX requested on PC vs N64: explosion→scream, armor pickup wrong sample, slap/glass layered over gunfire, melee→Klobb shot (M-71/M-72)

**Symptoms (user by-ear; N64 A/B confirms non-fidelity):**
1. General gunfire: a slap/glass-shatter plays *over* the correct gunshot on
   PC; silent on N64 for the same shot.
2. Explosion: plays a scream-like sample (the soldier-death voice) instead
   of / over the explosion.
3. Armor pickup: wrong sample.

**Exonerated — do not re-investigate:**
- Playback chain: M-68/M-70 exact match — every voiced request plays the
  exact ROM sample at the exact ROM pitch (204/204 voiced, zero wrong-sample).
- D37 bank converter: offline harness runs the real `romdataFixupAudioBank`
  on the same ROM bytes and diffs N64 vs PC images entry-by-entry —
  261/261 identical (base/len/type/keyMap/env), 0 mismatches
  (`scratchpad/banktest/harness.c` + `diff_bank.py`; gitignored, local-only).
- D154/D135 GBI raycast parsers ABI-correct; volume law
  (`sub_GAME_7F0537B8`: d≤200→max, sqrt falloff to 5000, linear to 6000)
  and suppression logic faithful (M-71 static audit).

**Therefore: PC requests different or extra sound indices than N64 in the
4. Unarmed melee: Bond's slap attack plays Klobb's shooting sound
   (`GUN_B1_MGUN3_3_SFX`, idx 106) instead of the slap effect (M-72).

**M-72 (controlled runtime traces, this session):** environment note first —
the earlier "flaky /GS crash" was an investigator artifact: `SDL2.dll` lives
in `C:\msys64\mingw64\bin`, which is NOT on the system PATH and not copied to
`build-pc/`; launches without it die instantly (exit 0xC0000105), with or
without any env vars. Prepending mingw64/bin to PATH in the launcher
(`build-pc/d205_*.ps1`, gitignored) gives clean 40–60 s runs. Results under
`GE_AUDIOTRACE=1`: (1) PPK fire at `-level_09` (unarmed start; first Z picks
up the spawn PPK, idx 232): every shot requests exactly {46 GUN_SILPPK,
one of 27/28/39/40 — all inside the 20-entry `ricochet_sounds_small` table,
gun.c:415 — so `rnd1 % 20` is IN RANGE, not OOB, 122 CART_SPENT}. No glass
(70), no slap (47–49) layering. (2) Unarmed melee at levels 01–08 (all start
unarmed): fire-press requests ONLY 105 PUNCHING_AIR (whiff); punch-HIT path
is the compiled constant `punch_sounds` = {47,48,49} (gunfire.c:2338), whiff
= 105 (chrprop.c:1465). No 106 anywhere. (3) idx 106 is referenced by NO
code constant — it enters only via `WeaponStats.Sound` (`skorpion_stats.
Sound = 0x6A`) through `bondwalkItemGetSound`; the player fire-sound path
(gunfire.c:3192–3200) runs only in `GUN_ANIM_STATE_FIRE` and a fist sets
PUNCH state with `.Sound = 0`, so Bond's own punch cannot request 106 — an
NPC guard firing its Skorpion (`chraction.c:5588`) can. **Conclusion:
all four symptoms' request paths are faithful constants/tables, and none
reproduces in controlled micro-scenarios; the leading suspect is now
mechanism (b) — extra/different EVENTS during real gameplay (a guard firing
as you punch → 106 layered on the slap; an NPC scream at the blast; a second
surface hit adding glass/slap to a shot).**

**M-73 (user combat capture) — ROOT CAUSE WITHDRAWN by M-74 below.** The
session analysed the user's full-combat `GE_AUDIOTRACE=1` capture
(`build-pc/audiotrace.log`) and concluded that guards were stuck
re-triggering weapon fire sound idx 109 (`GUN_B4_BOLTACTION_SFX`,
AK47/Spectre) "256x over 29.7 s with no corresponding bullets", pinning it on
`stanTestLineUnobstructed` guard->Bond LOS returning clear on converted
collision geometry. **M-74 re-counted the same file and every load-bearing
number was wrong** — see below. The code walk M-73 produced is still accurate
and worth keeping as reference: `chrlvFireWeaponRelated` (`chraction.c:6530`)
passes `phi_a2 = sp27C || sp278` to `sub_GAME_7F02BFE4` (:6902), so a
non-bullet auto tick still plays the weapon sound, gated by
`SoundTriggerRate`/`field_178[hand]` (and by `CHRHIDDEN_FIRE_TRACER` 0x80,
cleared once per `chrlvTriggerFireWeapon`); sustained attack needs
`seen_bond_time >= g_GlobalTimer - CHRLV_SEEN_RECENT_CHECK` (:6593),
refreshed by `setSeenBondTimeToNow` behind LOS queries (:3917/:3972,
`chrCanSeeBond` :3945). That mechanism is real. What M-74 shows is that
**nothing in the capture indicates it ever misfired.**

**M-74 (re-count of the SAME capture — M-73's evidence does not hold, this
session):** counting only the canonical one-per-call line
(`sndPlaySfx: t=... soundIndex=...`) in `build-pc/audiotrace.log`:

| M-73 claim | Actual |
|---|---|
| idx 109 requested **256x** | **128x** — the count was doubled |
| "~11 guards each re-triggering" | 128 / 29.73 s = **4.31/s aggregate**, ~0.4/s per guard |
| "sustained roar, gate blown open" | AK47 `SoundTriggerRate` = `RATE_AK47` = 4 `g_GlobalTimer` ticks (`assets/obseg/gun/gunWeaponStats.inc.c`) => design ceiling **15/s per guard**. Observed is **~6x UNDER** the weapon's own rate — `field_178` is not even the binding constraint |
| "no corresponding bullets/impacts anywhere near that rate" | In the same 29.7 s window: **~30 ricochet / wall-hit requests** (idx 19, 20, 22-27, 29, 31-33, 35-37, 40, 41), **69 x2** (flesh hit), **123-132** (12 body-falls), **134-147** (14 guard yelps). Guards were firing *and* hitting |
| "only ~32 non-109 requests in the window" | ~150 |

Also corrected:

- **The 128-vs-256 gap is the double-log artifact D204 already documented.**
  D204 corrected M-63's "idx 109 fired 94 times in 60 s" with exactly this:
  "the count is doubled by the probe (it logs two lines per call)". M-73
  re-made the same mistake on the same sound index. Count `sndPlaySfx: t=`
  lines, never bare `soundIndex=` matches.
- **The monotone index runs are ground-truth round-robin, not a broken
  selector.** The capture contains strictly increasing sweeps
  `134,135,...,147` (GET_HIT_MALE) and `123,124,...,132` (BODY_FALL), one
  request each, never repeating. This looks exactly like a `random() % n`
  gone wrong and was checked specifically: `chraction.c:2454-2470` selects
  `male_yelps[male_guard_yelp_counter]` and post-increments modulo 25 — a
  deliberate cycling counter in byte-matched game code. Faithful. Do not
  re-open.
- The trace file spans 1321.85 s wall (288 requests); the actual firefight is
  ~40 s of it (`t` in 715754-715800 s). M-73's "29.7 s" is the 109 span only.

**Net: the 109 traffic in that capture is normal, in-spec guard combat.** The
LOS / converted-geometry root cause is **withdrawn** — it was inferred from
counts that do not survive re-derivation, and M-73's Next steps 1-3 (probe
`setSeenBondTimeToNow`, decode pccg stan in Python, replay guard->Bond rays)
would have been a full session spent on a phantom. The only genuinely open
observation left from that capture is that the firefight lasted a long time,
which is a gameplay-fidelity question for an N64 A/B, not evidence of an
audio bug.

**Where that leaves D205.** Every layer upstream of presentation is now
proven correct: bank conversion (261/261 identical, M-71), requested indices
(182 explosion x1, 81 armor x1, 46 PPK, 47/48/49 punch hits, 105 whiffs — all
correct and in range, here and in M-72's controlled runs), and sample+pitch
(M-68/M-70 exact match). That leaves exactly one untested layer between "the
right sample" and "what the user hears": **spatial presentation — per-voice
volume, pan and concurrency.** It fits all four complaints in one shape: if a
sound's apparent *position* is wrong, a guard firing behind you is by ear
indistinguishable from "a gun sound came out of my punch", a nearby guard's
death scream reads as "the explosion screamed", and a ricochet reads as "a
slap over my gunshot". The user's ear reports *what and where*; only *where*
has never been verified. It is also the youngest code in the stack —
`sndCreatePostEvent` was stubbed out entirely (D138) until M-65 un-stubbed
it, so every distance/pan post-event has been live for only a few sessions
and has never been A/B'd against N64.

**Original M-71 framing (kept as background; DEMOTED by M-74 — every index
observed in the user's capture was correct, so "PC requests a different
index" is no longer the leading shape of the bug):** PC requests different or
extra sound indices than N64 in the same scenario. Two mechanisms, ranked:
(a) **Converted level/setup data** — sndID / explosion-type / item-sound
    fields in the pccg sidecars differ from ROM. Explosions play
    `sndPlaySfx(g_musicSfxBufferPtr, sp44->sndID, NULL)`
    (`src/game/explosion.c:285`) with `sp44->sndID` =
    `g_ExplosionTypes[type].sndid`; the *type* comes from level data.
    Explosion SFX are idx 169–183 (EXPLOSION_2A_SFX=169 … EXPLOSION_9_SFX=
    183, bondconstants.h). A scream is outside that range, so either the
    type index is wrong (→ a non-explosion sndID) or a second event fires.
    Armor: idx 81 verified exact in M-68 — a "wrong" armor sample means a
    different index is requested in the user's scenario (or a layered
    second sound).
(b) **Extra hit events** — converted collision geometry makes bullet
    raycasts or explosion proximity checks hit surfaces/NPCs the N64
    misses. The slap+glass-over-gunshot signature is exactly the
    non-penetrating-OBJ-hit path: `sub_GAME_7F064720`
    (`src/game/gunfire.c`) plays `HIT_BULLET_GLASS_SFX` (idx 70) for every
    non-penetrating hit; ricochet tables classify AFDM (27–30) = prop only,
    GBU/TAJ/RICO_4 (19–26, 31–33) = wall only.

**M-74 step 1 — static audit of the presentation layer: H-A FALSIFIED,
H-B narrowed to "faithful but shallow". Done this session.**

The pan machinery is complete and faithful end to end:
`src/snd.c:461/530` computes `pan = clamp(soundState->pan + sound->samplePan -
AL_PAN_CENTER)` → `alSynSetPan` (`synsetpan.c`, posts `AL_FILTER_SET_PAN`) →
`src/libultrare/audio/env.c:400-402` turns pan into independent L/R targets via
the `eqpower` table (`ltgt = volume*eqpower[pan]`, `rtgt =
volume*eqpower[LEN-pan-1]`) and emits `aSetVolume(A_LEFT|A_RATE)` /
`(A_RIGHT|A_RATE)` → `port/src/mixer.c` `aSetVolumeImpl` decodes A_LEFT
(0x02) into `volCur/volTgt[0|1]` and `aEnvMixerImpl` applies `vol[0]`/`vol[1]`
to `dry[0]`/`dry[1]` independently (:488-496). Nothing is dropped or collapsed
to mono. **The port's stereo path is correct.**

**But GE never uses it.** `AL_SNDP_PAN_EVT` (`snd.h:41`) is handled at
`snd.c:527` and **posted by nothing anywhere in the tree** — grep returns the
enum and the handler, no callers. So `soundState->pan` keeps the
`AL_PAN_CENTER` set at `snd.c:842` for every SFX's entire life, and the only
pan input is the static per-sound `samplePan` from the bank. Dumped from the
converted bank image this session: **252 of 261 sounds have `samplePan == 64`
(= `AL_PAN_CENTER`)**; the 9 exceptions are 4, 24, 44, 74, 84, 104, 123.
Every sound in the user's capture is centred — 46, 47, 48, 49, 81, 105, 106,
109, 182 all `samplePan = 64`.

**⇒ GoldenEye does not spatially pan SFX at all, on N64 or PC.** There is no
pan to get wrong. H-A is dead; do not re-open it, and do not write the
"pan varies with bearing" probe (M-74 step 2 as originally drafted) — the
answer is known and it is "never, by design".

Field-coverage gap in M-71's bank exoneration closed at the same time: the
converter writes `samplePan`/`sampleVolume`/`flags` at `+24/+25/+26`
(`port/src/romdata.c:890`), correct for the PC layout (3 × 8-byte pointers
before them), and `scratchpad/banktest/diff_bank.py` already diffs them as
`panvol` — re-run this session, **261/261, 0 mismatches**. (M-71's prose
listed only base/len/type/keyMap/env; its actual coverage was broader,
including a 24-byte sample-data fingerprint.) The volume formula
(`snd.c:453`, `:565`) also has no PC-side divergence: peak intermediate is
127 × 32767 × 127 ≈ 5.3e8, inside s32 on both platforms, and `vol` is `s16`
(`snd.h:90`) which holds the 32767 the `[VOL]` probe reports.

**What this leaves.** GE's only spatial cue for an SFX is distance volume, and
that law is shallow: the capture's `[DISTVOL]` lines show flat max out to
~200 units, then 0.60 of max still at dist 1794. So a guard firing across the
room plays **near-full volume, dead centre**, at a per-sample volume
comparable to the player's own actions (idx 109 `sampleVolume` = 90; punch
47/48 = 100, 49/105 = 110). **That is faithful, and it is also exactly the
perceptual condition that generates all four of the user's reports** — with no
spatial cue, a correct sample played at correct volume from across the room is
by ear a sound at your own position. This reframes D205: the complaints are
fully consistent with *correct* sounds overlapping, and the only remaining
question is whether the PC generates **more** such overlaps than the N64
(H-D). Nothing on the PC side can settle that; it needs the N64 A/B.

**Next (cheapest first; re-ranked by M-74 — M-73's list is withdrawn):**
0. ~~User capture~~ — DONE (M-73), but see M-74: it does NOT show a guard
   re-trigger bug. 182 and 81 were both requested correctly and exactly once.
1. ~~Static audit of the presentation layer~~ — **DONE (M-74 step 1 above).
   H-A falsified: the port's stereo path is correct, but GE never pans SFX at
   all (no `AL_SNDP_PAN_EVT` poster exists; 252/261 bank sounds are
   `samplePan == 64`). Bank `panvol` re-diffed 261/261 clean. Do not re-open.**
2. ~~Pan probe run~~ — **cancelled by step 1: pan cannot vary, by design.**
3. ~~The PC side is now exhausted~~ — **SUPERSEDED by D206.** The premise
   ("requested indices are correct") was wrong: the indices are correct as
   *values* but are resolved one slot too high, which D206 root-causes. What
   survives from this step: the pan audit (H-A dead) and the bank `panvol`
   re-diff. Original text: bank, playback, requested indices,
   sample+pitch, pan and per-sound volume are all verified faithful. The only
   remaining hypothesis is **H-D** — that the PC generates *more* overlapping
   combat events than the N64 — and nothing measurable on the PC alone can
   settle it. **The N64 A/B is now the gating step, not an optional
   corroboration.**
4. **User-side, batched, held until 1-3 report:** (a) ROM samples 106
   (Skorpion/"Klobb") vs 109 (AK47) as WAVs via `scratchpad/rom_sfx_decode.py`
   for the user to settle the Klobb identification by ear — their by-ear
   identification stands either way, this decides *which* sound, not whether
   they heard it; (b) N64 A/B of one moment: punch a guard while others fire —
   does the same combat sustain on N64?
5. After root-cause + fix: user re-test of melee / explosion / armor pickup
   vs N64.

**Ranked hypotheses after M-74:**
- **H-A (top) — pan/spatialisation wrong or absent in the mix.** Untested end
  to end; youngest code in the stack (D138/M-65 lineage).
- **H-B — distance law right, scale wrong.** The capture's `[DISTVOL]` lines
  do follow the documented curve (flat max to ~200, falloff after) but only
  reach 0.60 of max at dist 1794. Never compared to N64 loudness at the same
  distance.
- **H-C — voice-pool masking.** 8-voice soft limit vs 11 guards + music. Weak:
  M-70 measured alloc <= 5/8.
- **H-D — real content difference (PC guards engage where N64 guards don't).**
  Possible, but after M-74 it has *no* supporting evidence. Gated behind the
  N64 A/B.

**Probes added this session (live under `GE_AUDIOTRACE`, catalogued in
GE-ENV-PROBES.md):** `[VOL] t=… state=… rawVol=… playing=…` (`src/snd.c:
556`, per-VOL-event value) and `[DISTVOL] pos=(…) player=(…) dist=… vol=…`
(`src/game/propobj.c:12780`, the distance-attenuation funnel
`sub_GAME_7F053894`).

## D204 — Audio pipeline runs ~2 % below real time forever: GE's AI feedback loop has a 3 ms setpoint that OS jitter clears, so the DAC is padded with silence (M-64, FIXED + measured)

**Found while reviewing the D202 audio state, not by chasing D202's stated
symptom.** D202 itself is NOT resolved by this — see "What this does and does
not explain" below.

### The architecture (established, not assumed)

GE's audio is a closed-loop, DAC-backpressure-driven producer:

- `amMain` (`src/audi.c:433`) blocks on `OS_SC_RETRACE_MSG`. `__scMain`
  (`src/sched.c:334`) forwards a retrace to a client either every frame or,
  when the client registered with a non-zero flag, every *second* frame. The
  audio client registers with `1` (`audi.c:441`), so it wakes at **30 Hz**.
- Per wake, `amHandleFrameMessage` (`audi.c:531`) sizes the next block:
  `frameSamples = (u16)((g_FrameSize - (osAiGetLength() >> 2) + 16 + 0x25) & ~0xf)`,
  then clamps *below* at `g_MinFrameSize`.
- NTSC constants: `outputRate=22050`, `FRAMES_PER_FIELD_AS_POW2=1`,
  `MAYBE_FRAME_RATE=60` → `g_FrameSize=736`, `g_MinFrameSize=720`,
  `g_MaxFrameSize=789`. `info->data` is allocated exactly `g_MaxFrameSize*4`
  = 3156 bytes (`audi.c:388`).
- So `osAiGetLength()` is a **±64-sample trim**, and the loop only asks for
  more than 720 once the reported length falls under ~69 frames — **3 ms**.

### The defect

720 samples per 30 Hz block is **21600 samples/s against a 22050 Hz device**
— the nominal rate is structurally 2 % short, and the design depends on those
under-69-frames top-up blocks (784 samples) to make it back.

On N64 a 3 ms setpoint is fine: AI is double-buffered and the VI interrupt is
exact. On PC, `osAiGetLength()` maps onto the SDL queue, and ordinary OS
scheduling jitter empties a 3 ms cushion before the loop reacts. The queue
hits zero, SDL pads playback with silence, and **that lost time is
unrecoverable** — a queue-*depth* reading can never tell the regulator it has
already fallen behind. Measured, sustained, indefinitely:

```
GE_D204_OLD=1  rt=0.980  q=0..416   (q reaches 0 repeatedly)
```

`rt` = produced audio seconds / wall seconds. 0.980 is exactly 21600/22050.

### The fix (F5, port-side only — `src/audi.c` untouched, rule #2 clean)

`port/src/audio.c`'s `audioGetAiLengthBytes()` subtracts a target cushion
(`AUDIO_TARGET_FRAMES` = 1024 frames ≈ 46 ms) before reporting, so "the queue
holds the cushion" reads to the game as "the queue is empty". The loop then
tops up while ~46 ms of slack remains and settles just above the cushion
instead of oscillating into starvation. audi.c's own control law is unchanged.

```
after (same repro)   rt=1.000  q=368..672  drop=0   (q never reaches 0)
```

Also landed, all in `port/src/audio.c` / `port/src/libultra.c`:

- **F1** — `osAiGetLength()` now reports the residue of a *single* buffer
  (`min(queued, lastBufferBytes)`), matching AI_LEN_REG, instead of the whole
  SDL queue depth. **Robustness only: measured A/B shows zero behavioural
  change.** See the falsified hypothesis below.
- **F3** — `Audio.QueueLimit` default 8192 → 2880 frames (372 ms → 130 ms
  worst-case latency), and a full queue now logs a dropped block instead of
  discarding it silently. **Caveat: an existing `data/ge007.ini` pins
  `QueueLimit = 8192`, so this default only reaches fresh configs.**
- **F4** — invariant guard refusing any block larger than `g_MaxFrameSize*4`.
  Believed unreachable; costs one compare, versus a silent heap overrun.
- **`GE_D204=1`** — audio-health monitor, one line per 5 s: `rt=` real-time
  ratio, `q=` queue depth, `drop=`, `max=` largest block vs its allocation.
  Deliberately cheap (~30 clock reads/s) so it can be left on for a whole
  playtest. **`GE_D204_OLD=1`** restores pre-fix behaviour for in-binary A/B.

### Hypothesis raised and FALSIFIED — do not re-open

I predicted the u32 subtraction in `audi.c:531` wraps once the queue exceeds
`g_FrameSize + 0x35` = 789 frames, truncates to u16 (e.g. 58128), and drives a
74× heap overrun of the 3156-byte `info->data`. **The queue really does exceed
789 routinely — measured high-water 2064 frames over 5 minutes — and the wrap
really does happen, but it is harmless.** `frameSamples` is declared
**`s16`** (`audi.c:145`), so the wrapped value lands negative and audi.c's own
`(s32)frameSamples < (s32)(s16)g_MinFrameSize` clamp catches it, yielding a
nominal 720-sample block. Max block observed is 3136 bytes against the
3156-byte allocation, in both old and new modes, across every run. There is no
overrun. This is why F1 measures as a no-op.

### Two M-63 conclusions corrected

1. **`soundIndex=109` firing "94 times in 60 s with no AK47 present" is a red
   herring — closed, negative.** The count is doubled by the probe (it logs two
   lines per call); 47 real calls. All of them fall in tight bursts *after*
   t≈5.4 s, interleaved with ricochet indices 27/28/30/35/39/41 — **guards
   returning fire**, correct behaviour once the player starts shooting in
   Bunker. And the "suspiciously uniform ~32.6 ms cadence" is not a retrigger
   artifact at all: 32.6 ms **is** the audio-block quantum (720 frames × 4
   bytes = 2880 bytes), so *every* `dumppos` in the entire trace is quantised
   to it. Several sounds landing in one block read as "4-in-a-row".
2. **M-63's implied "audio produced at 13 % of real time" (708608 dump bytes
   = 8.03 s over a 60 s run) was a measurement artifact of its own probe.**
   That run had `GE_MIXERTRACE=1`, whose unbuffered per-opcode `fprintf`
   produced a 25 MB log and slowed the process enough to starve the audio
   thread. The identical repro without it produces 58.0 s of audio in 60 s.
   **Lesson: `GE_MIXERTRACE` is not safe for any timing-sensitive measurement**
   — that is exactly why `GE_D204` was written to be cheap. M-63's
   recommendation to have the user replay with `GE_MIXERTRACE=1` set would
   have induced the very starvation it was looking for.

### What this does and does not explain

**Does:** a permanent low-level stutter/gap artifact in all PC audio — ~2 % of
playback was silence, in every level, for every user, from the first frame.

**Does not:** D202's reported chain (wrong sound on the silenced PPK, correct
sound sometimes heard simultaneously, sounds piling up and glitching over a
session, eventual near-silence except one stuck loop). Nothing in this
session's headless runs degraded: a 5-minute `-level_09` soak held `rt=1.000`,
`drop=0`, no voice-pool growth. D202 stays OPEN. It needs a real playtest —
now instrumentable with `GE_D204=1` (plus `GE_AUDIOTRACE=1`, and **not**
`GE_MIXERTRACE`), where `rt` falling or `q` pinning at `queueLimit` with
`drop=` climbing would localise it immediately.

### Verification

Reproduce with **`.	ools_pcudiodebug.ps1 -AB -Fire`** (added this session;
`-Play -Trace` is the instrumented interactive run for D202). `-level_09`,
scripted PPK fire, MSYS2 MINGW64 build, full `data/` mirror. In-binary A/B via
`GE_D204_OLD` (same executable, no build-to-build variance):
`rt` 0.980 → 1.000, `q` min 0 → 368, `drop` 0, `max` 3136/3156 unchanged.
Not yet verified by ear — the artifact is a ~2 % silence rate, so a human
listening pass on a real playthrough is still owed before calling it closed.

### Harness gotchas found while building `audiodebug.ps1` (both cost a debug cycle)

1. **`Start-Process -RedirectStandardOutput` does not capture this game's
   stdout.** `CMakeLists.txt:577` sets `WIN32_EXECUTABLE TRUE` (`-mwindows`),
   so `ge007.x86_64.exe` is a GUI-subsystem binary with no console of its own;
   the redirect yields a 0-byte log while the output lands on the parent
   console. A shell that opens the file itself before `CreateProcess` (cmd's
   `>`, or a bash redirect) does work.
2. **MSYS2 `bash.exe` launched from PowerShell arrives with a stripped
   environment.** Neither variables inherited from PowerShell nor the script's
   own `export` reach the child process — verified with a minimal
   `export GE_PROBE=hello; env | grep GE_PROBE` probe, which prints under
   `sh script.sh` from MSYS but nothing when the same script is run by
   `powershell -Command "& bash.exe script.sh"`. Every `GE_*` probe silently
   read as unset, which looks exactly like "the probe is broken". `cmd.exe`
   inherits normally, so `audiodebug.ps1` generates a `.bat` and runs that.

### M-78b — user repro on Bunker: symptom is specific and permanent

Natural Bunker alarm trigger (spawn, go forward into the room ahead, shoot
one of the two guards → the other trips the alarm), then sustained firefight.
User by-ear result:

- Most SFX keep playing during the klaxon.
- **Eventually the *player's* gunshots stop being emitted — and stay muted
  even after the alarm ends.** Enemy gunfire and all other SFX continue fine
  during and after.

This is NOT generic pool starvation (that would drop random sounds
transiently). It is specific to the player-weapon fire sound and it is
*permanent* — a stuck-state bug, not a transient voice shortage.

**Structural lead (static, unconfirmed — needs the capture).** The player
fire sound is played through a **two-slot double buffer** in
`bondfirefunc` / `gunfire.c:3178-3202`: `handptr->audioHandle` and
`handptr->field_A48`. Each trigger interval: `sndDeactivate` whichever slot
is currently *playing*, then `sndPlaySfx(..., &slot)` into whichever slot is
*NULL* (`audioHandle` first, else `field_A48`). If **both** slots are ever
left non-NULL-but-not-playing at once, the `else if (field_A48 == 0)` guard
fails and **no further player gunshot can play, ever** — matches the symptom
exactly.

A slot is cleared only by `sndUnlinkClearSound` (via `sndDisposeSound`)
following the state's back-link `state->state->link.next`. `sndDeactivate`
is **asynchronous** — it posts `AL_SNDP_DEACTIVATE_EVT` to the 64-slot SFX
event queue (`MUSIC_SFX_SEQ_CONFIG_MAX_EVENTS = 0x40`). If that queue is
saturated (heavy firefight + the alarm's self-retrigger posts + every
ricochet/impact/yelp posting PLAY/PLAY_EVT/STOP/DEACTIVATE), `alEvtqPostEvent`
**drops the event silently** → the deactivate never runs → the state never
disposes → the slot never clears. Once both slots are stuck, permanent mute.
The `[VOICES] ... evtq=N/64` probe already in `snd.c` will show queue
saturation; `[SLOTWRITE]` + `[VOICE+]/[VOICE-]` + `sndPlaySfx t=` trace the
two player-gun slots. All three counts (8 voices / 64 states / 64 events) are
byte-match stock GE — so if this is the mechanism it is a port *timing* bug
(the SFX event queue drained too infrequently vs N64's RCP cadence, D204
family), not a value to bump.

**Next:** read the user's `build-pc/audiotrace.log` — confirm (1) `evtq`
approaches 64, (2) a `[AUDIOTRACE] sndDeactivate: state=X` with no following
`[VOICE-] state=X` / dispose, (3) two `[SLOTWRITE] slot=<audioHandle>` and
`slot=<field_A48>` with no clear afterwards. Then the fix is port-side: drain
the SFX event queue more often, or (cheaper, D202-family-consistent) a
bounded self-heal in the double-buffer — if a slot has pointed at a
non-playing state for > N ms, force-clear it.

## D217 — M-88 defect ① root-cause narrowing: NOT the model sidecar; the runtime TLUT-address path (session dispatched to investigate a `tools_pc/d43_*.py` fix; investigation concludes the fix does not belong there)

**Dispatch brief for this session** asked for a root-cause + fix of D217
defect ① (wrong palette bytes arrive for model-DL CI textures) confined to
`tools_pc/d43_*.py`, working theory: the model-DL/sidecar's embedded
palette-source pointer is a fixed 4 bytes / 2 entries too low. **That theory
is refuted** and the fix does not belong in the converter at all — see
below. Per the brief's own instruction ("if your investigation concludes
otherwise, stop and write up why instead of editing game code"), no code
was changed; this is a write-up only.

### Why `tools_pc/d43_emit.py` cannot be the site of this bug

Traced the full model-texture reference chain from the ROM to the draw call:

- A model's texture table (`ModelFileTextures[]`, the "texconfig" block
  `d43_emit.py` emits at `8*NS .. 8*NS+12*NT`) stores each texture as a
  **numeric `TextureID`** (`src/bondtypes.h:836-847`), not an embedded
  pixel/palette blob. `d43_emit.py:508-516` passes this `u32` straight
  through byte-order-swapped (BE→LE) — **the numeric value is never
  remapped or altered**. Verified by reading the emit loop directly; there
  is no code path in `d43_emit.py` that touches texture *content*.
- `texLoadFromModelFileHeader` (`src/game/image.c:2522-2536`, decomp ground
  truth) resolves each `TextureID < MAX_TEXTURES` via `texLoad()` against
  the **global, per-game (not per-model, not per-level) image table**
  (`Globalimagetable`). That table's BE→LE fixup lives in
  `port/src/gimgfixup.c` (D68) — a `port/src/` runtime file, not a
  `tools_pc/` offline converter. `tools_pc/d43_lutscan.py` (read-only
  analysis script, confirmed by re-reading it end to end) independently
  shows every model texture reference resolves through this same global
  `imagelist.u.csv`-indexed table — there is no per-model or per-level
  texture/palette blob for `G*`/`P*`/`C*` model files to have a "sidecar
  offset" bug in.
- Actual pixel+palette bytes are decoded **at runtime, once per level load**
  by `texInflateZlib`/`texInflateNonZlib` (`src/game/image.c`). For the
  zlib/CI-native path (`texInflateZlib:169-331`), the palette is decoded
  from the bitstream (`texReadBits`, endian- and pointer-width-agnostic)
  and **appended after the pixel data in the same `tex->data` allocation**
  (lines 320-326), with `tex->unk0a = numcolours - 1` (`image.c:238`)
  recording the palette's position. None of this is touched by any
  `tools_pc/d43_*.py` script — it is pure runtime C, identical instructions
  on both platforms.
- The actual `G_LOADTLUT`-equivalent commands are **synthesized at runtime**
  by `texWriteLoadToTmemAddr` (`src/game/tex.c:494-554`), which computes a
  hardware-accurate (S,T)-style byte-offset split (`uls`/`ult`) into
  `tex->data` using `tex->unk0a`, and issues them via the *same*
  `gDPSetTextureImage(gdl++, tex->gbiformat, depth, 1, tex->data)` call used
  for the pixel LOADBLOCK moments earlier (`tex.c:503` and `tex.c:519`) —
  there is no separate "model-DL TLUT source pointer" for a converter to
  get wrong; the address is a live pool pointer computed fresh every level
  load, in game code, not ROM-serialized data.

Conclusion: **the "4-byte / 2-entry leading pad" theory in the original
D217 row does not have a mechanism** — nothing in the offline pipeline ever
places a palette 4 bytes off from where the runtime decoder expects it,
because the offline pipeline never places a palette at all for these
textures.

### What the runtime probe actually shows (empirical, this session)

Built successfully in this worktree (`./build-pc.sh ntsc-final`, clean
build, `data/` symlinked from the main checkout — see VERIFY below) and
reproduced the exact byte values quoted in the original D217 row using the
**existing, unmodified** `GE_TEXDUMP=1` env probe already in
`port/fast3d/gfx_pc.cpp:1066-1075` (no code changes, read-only env-gated
diagnostic that ships in the tree):

```
BUNKER1  (-level_09,  GE_PCDUMP=200-440:120):
  GE_TEXI[125] fmt=2 siz=0 palfmt=32768 size=8 tile=1x1 pal[0..3]=0001 18c7 18c9 2107
  GE_TEXI[185] fmt=2 siz=0 palfmt=32768 size=8 tile=1x1 pal[0..3]=0001 0800 0840 0842

FACILITY (-level_34,  GE_PCDUMP=300-400:20):
  GE_TEXI[157] fmt=2 siz=0 palfmt=32768 size=8 tile=1x1 pal[0..3]=1a00 0000 0001 0843
```

`fmt=2 siz=0` = CI4, `tile=1x1` = a flat single-texel fill (matches "grip =
solid colour block"), `pal[0]=0001` on BUNKER1 (black, matches the row's
claim exactly) vs `pal[0]=1a00` on FACILITY (matches the row's claimed
green `0x1A00` exactly, byte-for-byte). This independently confirms the
symptom is real and reproducible with zero code changes — good news for
whoever picks up the actual fix.

**Where the bug most likely actually lives (not verified further — out of
this session's scope to touch):** `port/fast3d/gfx_pc.cpp:2202-2237`
(`gfx_dp_load_tlut`) computes the palette source address as
`rdp.texture_to_load.addr + pitch*ult + uls` where
`pitch = rdp.texture_to_load.width + 1`, and `rdp.texture_to_load.width`
comes from the immediately preceding `gDPSetTextureImage` call. GE's
`texWriteLoadToTmemAddr` (`src/game/tex.c:503,519`) always issues that
`SetTextureImage` with a **literal width argument of `1`** regardless of
the texture's real dimensions — real N64 hardware LOADTLUT addressing uses
this "width" field as a DRAM line-stride for the S,T→byte-offset
computation, so fast3d's software model of that computation is only as
good as its handling of the `width=1` special case for a texture whose
real pixel-data length (`len`, hence the palette's real starting byte) can
be much larger than what a `pitch=2` u16-stride implies. This is consistent
with the finding row's "different bytes per level for the same tile" —
if this pitch/offset math is wrong, the byte read for a `count=1` TLUT load
would land in whatever the shared texpool arena happens to hold near
(but not at) the intended 2-byte palette tail, which is exactly the
level-dependent "reads adjacent garbage" signature already logged. This is
a **hypothesis with strong circumstantial support, not a confirmed
mechanism** — I did not instrument `gfx_dp_load_tlut` further because doing
so would require editing `port/fast3d/gfx_pc.cpp`, explicitly off-limits
for this dispatch.

### Confidence

- **High confidence** the bug is NOT in `tools_pc/d43_*.py` (structural
  proof: the converter never sees texture/palette bytes for model-DL CI
  textures; traced the complete reference chain from ROM to draw call).
- **Medium confidence** the mechanism is the `gfx_dp_load_tlut` /
  `texWriteLoadToTmemAddr` width=1/pitch interaction described above — this
  is the most concrete remaining hypothesis but was not instrumented or
  bisected this session (would require touching the off-limits file).
- Re-affirms the M-86 finding that ② (the CI cache-key content-hash fix,
  `fix/d217-model-texture-palette` / `f6fc0399`) is real, separate, and
  unaffected by any of the above.

### For whoever picks this up next

A follow-up dispatch with `port/fast3d/gfx_pc.cpp` (specifically
`gfx_dp_load_tlut`) and/or `src/game/tex.c` in scope (the latter under the
narrow ABI-only exception if a genuine platform difference is found, or
`#ifdef PORT` in `port/` if the fix is a fast3d-side hardware-model
correction) is the natural next step. Suggested probe: temporarily log
`rdp.texture_to_load.width`, `pitch`, `uls`, `ult`, and the resulting `base`
pointer's offset from `tex->data` inside `gfx_dp_load_tlut` for the
`tile=1x1`/`count=1` case on both BUNKER1 and FACILITY, and compare against
what `texWriteLoadToTmemAddr`'s `len`/`tex->unk0a` actually were for that
texture (a second probe point in `tex.c`, already has a `GE_D85TEX`
env-gate precedent to extend). If `base` does not land exactly 2 bytes past
the real end of the decoded pixel data, that confirms the pitch/width=1
special case as the mechanism.

### VERIFY — what was and wasn't completed this session

- **Build:** clean (`./build-pc.sh ntsc-final` from MSYS2 MINGW64, this
  worktree, `data/` symlinked read-only from the main checkout since this
  worktree had none — no writes to the main checkout's `data/`).
- **BUNKER1:** ran `tools_pc/verify.sh bunker1` — PASS, frames=3, no crash.
  (Note: the golden-frame-stem comparison itself reports a pre-existing,
  unrelated harness mismatch — "golden frame stems don't match --dump
  window" — even though the requested and golden windows are numerically
  identical (200-440:120); this looks like a `verify.sh`/`framediff.py`
  stem-formatting bug, not a regression from this session's zero code
  changes. Not investigated further — out of scope, no code was touched
  either way, and `worst_cell=0` with no crash is the relevant signal here.)
- **FACILITY:** ran the game directly (`GE_PCDUMP=300-400:20`,
  `GE_TEXDUMP=1`) rather than through `verify.sh` (to keep the temp capture
  dir instead of having it auto-deleted) — captured frames 300-400 cleanly,
  no crash, and confirmed the green `pal[0]=1a00` grip palette entry is
  still present (expected — no fix was applied). Did not do a pixel-eyeball
  of the PPM captures beyond confirming they exist; the log-level TLUT
  value match (`1a00` = the documented green) is the decisive signal.
- No `tools_pc/d43_*.py` file was modified; `git status` in this worktree
  shows only the two docs edits (this section + the D217 table-row pointer)
  plus a local `data` symlink (not tracked/committed) used for the build.
## D176(a) — M-93: human playtest of PR #43 (debug build) — Cradle default-FOV black sky CONFIRMED FIXED; Statue scroll speed improved but still off; NEW residual split out as D227 (sky renders as two independently-moving halves)

**Method:** user's own debug build (`build-linux-dbg`), branch
`land/d176a-sky-perspective-scroll` checked out. Sampled Statue
(`-level_22`) and Cradle (`-level_41`), plus a Bunker1 sanity pass.

**Confirmed fixed:** Cradle's sky at default FOV is no longer black — matches
the PR's claim exactly.

**Confirmed improved, not fully fixed:** Statue's cloud scaling reads
noticeably better than before (the PR's perspective-correct-`tc`
interpolation fix is doing real work), but the user's read is it's **still
too fast**. Not yet re-measured numerically post-verdict; could be residual
under-correction in the fix itself, or could be downstream of the D227
split-sky issue below making the whole thing harder to judge cleanly.

**Confirmed still-open, as expected:** at **max FOV**, Cradle's black-skybox
artifact reappears. This matches the PR body's own disclaimer verbatim
("the separate FOV-coupled black-sky-past-frustum-edge symptom (D218/D222
-family) ... unchanged/still open — not attempted here") — not a surprise,
not a regression, just confirmation the known gap is real and reproducible.

**New, not previously documented:** the sky visibly renders as **two
independently-moving halves** on both Statue and Cradle — user's words,
"like there's two different sky processes running at once roughly taking
half the sky each," each half scrolling differently. Split out as its own
row, **D227**, rather than folded into D176(a), since it's a distinct visual
signature (a seam + differential motion between two regions) rather than a
scroll-rate or color/black issue. Not established whether this pre-dates
PR #43 (the fix could have just made it easier to notice once the gross
scroll-speed error calmed down) or whether it's a widescreen/aspect-ratio
rendering artifact rather than a logic bug — user explicitly flagged the
latter as plausible and unconfirmed pending a side-by-side against real
N64 output at the same camera angle.

**Sanity pass:** Bunker1 (non-sky level) — clean, no regression, matches
the PR's own headless `verify.sh bunker1` PASS claim.

**Disposition:** PR #43's fix is real and should still merge — it closes
the confirmed black-sky sub-symptom on Cradle and measurably improves (if
not fully fixes) the Statue scroll-rate symptom, with no regression
elsewhere. D176(a) stays open for the residual scroll-speed-still-too-fast
report and defect 2 (horizon-band tessellation, pre-existing, unchanged).
D227 is a new, separate open item pending investigation.

## D227 — M-94: visual confirmation + a concrete mechanistic hypothesis for the "two sky processes" seam — per-primitive `wScale` quantization in D176(a)'s fix, not yet confirmed or fixed

**Method.** Headless capture, `-level_22` (Statue), `GE_PCDUMP` at frames
300/330/345/350/355/360, `land/d176a-sky-perspective-scroll` built and run
directly (not through `verify.sh`, to keep the PPM frames instead of having
them auto-deleted). Converted to PNG and cropped to the sky band with PIL for
close inspection.

**What the capture shows.** Frame 345's sky band has a clearly visible
crease: the cloud streak pattern on the right portion of the frame breaks
continuity from the pattern on the left, with a faint diagonal artifact line
right at the boundary — this is a single static frame, so it rules out "just
normal camera pan across a busy texture" as the explanation for at least this
instance. Screenshot: `docs/img/bugs/d227-sky-seam-statue.png`.

**Hypothesis.** D176(a)'s fix (`skyPortRenderPoly`, `src/game/sky.c`)
computes a per-call `wScale = maxAbsW / 30000` from the vertices passed to
*that one call*. But `skyRenderTri`/`skyRenderFull` issue **up to 5 separate
`skyPortRenderPoly` calls per frame**, one per triangle in a 5-vertex sky fan
(`sky.c:900-920` and `:1372-1393`) — so each triangle in the fan can pick a
different `wScale` depending on its own 3 vertices' camera-space `w`
(`unk0c`).

The position math itself is exact for any `wScale > 0` (re-derived
algebraically: `clip.x = ob[0]·wScale = (ndcX·w/wScale)·wScale = ndcX·w`,
independent of `wScale` — this was already verified when the D176a fix was
first reviewed this session). So this is not a bug in the fix's core
logic. The likely mechanism is **quantization**: `Vtx.ob[]` is `s16`, so
`ndcX·wv` (where `wv = w/wScale`) rounds to a 16-bit integer *at a
granularity that depends on `wScale`*. Two triangles sharing an edge each
call `skyPortRenderPoly` independently, each picking its own `wScale` from
its own 3 vertices — so a vertex that is logically shared between two
triangles can round to two slightly different quantized `ob[]` values
depending on which triangle's call computed it, cracking the shared edge by
a small but visible amount. That reads exactly like "two regions, different
pattern" without requiring two draw calls, two cloud layers, or any
screen-half-specific logic — a single quantization granularity mismatch
between neighboring primitives is sufficient.

**Not yet confirmed.** This explains the static crease. It does **not** yet
confirm the user's "each half moves at a different rate" observation — that
would require isolating camera motion from sky-texture scroll (a
static-camera, sim-time-only comparison), which this session's captures
didn't attempt. The user's original alternate theory (a widescreen/aspect-
ratio rendering artifact, unrelated to game logic) is also not ruled out —
it could coexist with or be independent of the quantization crease.

**Fix direction, if this hypothesis holds:** compute one `wScale` shared
across the *entire* sky fan for the frame (max `unk0c` over every vertex in
every poly about to be drawn, not just the 3 vertices of one
`skyPortRenderPoly` call), then pass it into each call instead of
recomputing locally. Every triangle's vertices would then quantize against
the same scale, and shared edges should match exactly. Not attempted this
session — flagging the direction for whoever picks this up, along with the
static-camera test needed to confirm before spending a build cycle on it.

**Housekeeping:** capture files (`ppm/`, a `ppm2/` scratch dir, and the run
logs) were cleaned up after pulling the one evidence screenshot; nothing
else from this session's captures is left in the tree.

**M-94 addendum — static-camera confirmation, same session.** `-level_22`'s
fixed level-title cutscene ("Statue Park, St. Petersburg") holds a genuinely
static camera (verified: the foreground statue/pedestal region differs by
only mean ~3.0/255 between frames 30 and 90 — lighting/AA noise, no
detectable pan/zoom) while the sky region above it differs by mean ~15.9/255
over the same two frames. **This confirms the sky animates over time,
independent of camera motion** — it is not merely a static geometric crease
that only *looks* like motion under a moving camera. That said, the diff
magnitude was non-monotonic across the sampled frames (30→45 measured larger
than 30→90), which doesn't cleanly fit a simple constant-rate linear scroll
either — worth another look with a denser, longer static-camera sample
before concluding *how* it's animating. Does not distinguish between "one
sky animating uniformly past a static quantization seam" (this session's
leading hypothesis) and "two separately-animating regions" (the user's
original read) — either is still consistent with a real per-frame diff in
the sky region. Capture files cleaned up after taking these measurements;
nothing added to `docs/img/bugs/` beyond the one screenshot already
committed.

## D229 — M-101: Frigate/IsWater-level water renders green and pulses blue<->green. Pre-existing (reproduces on main), mechanism partly identified, NOT root-caused

**Reported (user playtest, M-101):** on Frigate the water "turns green
randomly and goes between looking right and glitching green", later refined to
**"it changes from blue to green on an interval like pulsating"** and, at one
point under a bad intermediate build, "only green". Also reported in the same
message: green weapon textures on Frigate and Surface 2 — **that half was a
stale-branch artifact, not a bug** (see below).

**The gun half: not a bug.** `land/d176a-sky-perspective-scroll` was 2 commits
behind `main`, and those 2 commits were exactly the D217 green-texture fixes
(`7278baad` = #47 texpool 8-byte alignment, `f15a12b9` = #49 CI texture cache
keyed on palette content). The user was play-testing a branch build that
predated both. Merged main in (`f8822805`); user confirms **"gun is good on
frigate"**. No new defect. Worth remembering as a process point: a D217-family
symptom report against a feature branch should check branch-vs-main first.

**Pre-existing, and not caused by the D227 work — proven by A/B.** Built plain
`main` (none of the D227 changes) and captured `-level_26` frame 340: the
water is the **same green**. The only difference is tiling density — main
shows it 32x over-tiled (the D227/M-100b unit bug), the branch shows it
smooth. Green in both. So D229 is independent of D227.

**Mechanism, as far as it is established.** The water quad is `sky.c`'s
`IsWater` block (`skyRender` block1; Frigate confirmed `IsWater=1` via
`GE_D176`, which also answers M-99's ranked item 1 — that path IS textured and
does go through `skyPortRenderPoly`). Its draw state comes from
`sub_GAME_7F09343C` (`src/game/unk_092E50.c:278`):

- tile 0 and tile 1 are both `G_IM_FMT_RGBA`/`G_IM_SIZ_16b`, `line=4`,
  **both at TMEM 0** — i.e. the same texture — with `masks = maskt = 5`
  (32-texel wrap).
- tile 1 is offset: `gDPSetTileSize(1, uls=90, ult=150, lrs=0, lrt=0)`, so it
  samples the same image shifted by 22.5 / 37.5 texels.
- 2-cycle combine `gDPSetCombineLERP(TEXEL1, TEXEL0, PRIM_LOD_FRAC, TEXEL0,
  ...)` with cycle 2 `= COMBINED * SHADE`, and the LOD fraction is
  **`gDPSetPrimColor(..., sinf(flt_CODE_bss_80079E88) * 127.0f + 128.0f, ...)`
  — an animated sine.**

**That sine is the "pulsating".** The effect cross-fades TEXEL0 and TEXEL1 over
time; the user seeing blue<->green on an interval means **one texunit resolves
correct (blue) and the other resolves green**, with the sine fading between
them. This is the single most useful clue in the report and it came from the
user, not from a capture — a static frame cannot see it (frame 340 samples
green on *all three* builds tested, because the sine phase happens to land
there, which is why single-frame diffs were useless here).

**Ruled out:** the shade path. Water vertex colour is
`skyChooseWaterVtxColour` = `env->RGB + env->WaterRGB * (1 - c/255) * (1 - a)`;
measured on Frigate `env RGB = 16,48,96` and `WaterRGB = 255,255,150`, so the
shade ranges from dark blue to pale yellow and **can never be green**. Measured
water pixels are ~`(5, 70..175, 20..37)` — R and B crushed, so the green comes
from the texel side, not the shade.

**Not yet established:** why one texunit decodes green. Both tiles point at the
same TMEM with the same format, so a plain "wrong texture" explanation does not
fit; the difference between them is only the uls/ult offset and whatever
fast3d does with tile 1's inverted window. One concrete oddity found and *not*
yet shown to matter: `uls/ult/lrs/lrt` are `uint16_t` and
`gfx_dp_set_tile_size` computes `width = ((lrs - uls) >> 2) + 1`, so tile 1
gets `((0 - 90) >> 2) + 1 = -22` stored into a `uint16_t` = **65514** (height
likewise 65500). `tex_width2` goes negative by the same arithmetic, which as it
happens makes the `g_wrap_fix` sub-tile-window branch skip (`tex_width2 > 0`
fails) and fall through to the mask-based 32-texel wrap — which is the correct
period — so this may be harmless. It has not been proven either way.

**FOV context (M-101):** the user's playtest ran with **FOV turned up**
(`Video.FovScale`, PR #36's slider, range 50-150). Every headless capture in
this entry ran at the ini's `FovScale = 100`, i.e. **default**, and still
reproduces the green — so raised FOV is *not required* to trigger D229. Whether
it worsens it is untested (deliberately: the config file was open in the user's
editor and not worth racing). Worth checking when this is picked up, because
FOV-coupled sky/frustum coverage is a known separate symptom family
(D218/D222), and the black bands visible at the edges of the Frigate captures
are likely that rather than D229.

**Next steps:** dump what fast3d actually decodes for each texunit on this draw
(the existing `GE_D172` multitex probe does not fire here — it filters on
`tile1.tmem != tile0.tmem`, and these are equal, so it needs widening);
capture a *sequence* across the sine period rather than one frame, since the
defect is temporal; and check `skywaterimages[WaterImageId=2]`'s real format
against the `RGBA/16b` that `sub_GAME_7F09343C` re-declares.

**D227 deliberately decoupled from this.** The M-100b safety valve now takes an
`allowShift` argument and the IsWater quad passes `FALSE`, leaving that path
byte-identical to its pre-D227 behaviour. Reason: the water quad is a
two-texunit draw sharing one set of vertex tc, so a tc rescale must be matched
by a tile shift on *both* tiles — and shifting both makes them sample
identically, which flattens the very cross-fade the effect is built on. That
was observed live: an intermediate build that shifted both tiles turned the
water from pulsating into **constantly green**. Neither variant is right, so
the water quad keeps plain 1:1 tc and tolerates the overflow on its horizon
vertex exactly as before. D227's fix stands on its own for the non-water
levels; D229 is a separate defect needing its own fix.

**ROOT CAUSE FOUND AND FIXED (port-layer, `port/fast3d/gfx_pc.cpp`).** The water
image (`skywaterimages[2]` = texnum 1509, `IMAGE_WATER_BLUE`, 32×32, level 6) is a
ZLIB+CI8 image: the texture pool holds raw 8-bit palette indices (verified:
the dumped pool block is 1400 bytes, every byte ≤ 23 — exactly the 24-colour
blue palette's index range; the full mipmap chain 32²+16²+8²+4²+2² plus row
padding). `texSelect` (`othermodemicrocode.c`, "Perfect Dark texSelect") loads it
with `gDPSetTextureImage(CI, 16b)` + `gDPLoadBlock` — the 16b "size" is only the
fast3d 4KB-per-block convention (lrs counts 16-bit units: 700 × 2 = 1400 bytes of
8-bit indices). On N64, GE's custom RSP ucode expands the indices through the
TLUT into real 16-bit pixels in TMEM at load time — which is why
`sub_GAME_7F09343C` can then re-declare TMEM slot 0 as RGBA/16b and sample blue.
fast3d has no such expansion: `import_texture_rgba16` read pairs of index bytes
as 16-bit texels, where g = 2·idx mod 32 dominates and r/b ≤ 2 → the green
mottle. The "pulsing" is the pre-existing `sinf()`-driven `PRIM_LOD_FRAC`
cross-fade between the two tiles (both mis-decoded, at different UV offsets),
not a second bug.

**Fix:** track the source format of each TMEM load (`texture_to_load.fmt` was
previously dropped; now recorded into `LoadedTexture.src_fmt` by
gfx_dp_load_block/gfx_dp_load_tile) and, in `import_texture`, when a tile is
declared RGBA/16b but the slot's last load came from a CI source, route the
import through `import_texture_ci8` (palette expansion via `rdp.palette`) — the
port-layer equivalent of the ucode's expand-at-load. Fires only on the genuine
format disagreement; real RGBA16 textures are untouched. Verified live: Frigate
water renders blue and the user confirms "looks better". Closure still wants the
standard verification ritual (sweep + by-eye on Surface 2, the other IsWater
level).

## D227 — M-100b: FIXED. The real cause is a unit error (`unk20`/`unk24` are S10.5, not texels); the s16 overflow was its symptom, and so are the over-tiling and the "scrolls too fast" complaint

**This supersedes the fix design in M-100 below.** M-100's measurement and its
overflow diagnosis were right; its *explanation* of why the numbers were so
large was not. Neither tessellation nor the tile shift turned out to be needed.

**The unit bug.** M-82 ("D176(a) Defect 1") read `unk20`/`unk24` as texel
counts and therefore baked `tc` as `(S - fold) * 32` to convert texels into
`tc`'s S10.5 units. They are not texel counts. `skyRender` sets them as
`worldpos * 0.1f` (the `sp4b4[].unk0c`/`unk10` assignments) and the N64 path
hands them to the RDP's coefficient block unscaled (`skyRenderTri`:
`S * unk34 * sp368`), so they are **already in the RDP's native S10.5
1/32-texel units**. The stray `* 32` made every sky texture coordinate 32x too
large, which produced three symptoms tracked until now as separate bugs:

1. **D227 itself.** Spans of ~30,000 *texels* against an `s16` `tc` holding
   +/-1024 -> 7 of 8 coordinates overflowed and wrapped -> starburst / "two
   sky processes".
2. **Over-tiling** — ~466 repeats of the 64-texel cloud across one quad
   instead of ~15.
3. **"The N64 sky is nearly static, ours races"** (M-96), measured at
   ~7 px/tick in M-98. `g_SkyCloudOffset` advances 1 unit/tick, which is
   1/32 texel — not 1 texel. **The M-96 scroll complaint and D227 are the same
   bug**, which is why M-98's re-derivation of the tick path kept coming back
   clean: the accumulator was never wrong, its unit was.

**The fix:** `tc = (S - fold)`, 1:1, with the fold rebasing to a whole tile
period (`SKY_TC_WRAP` = 64 texels x 32 = 2048 units). The ~30,000-unit span
then fits an `s16` unaided — which is also why the original N64 code needs no
tessellation here: it never had a range problem to solve.

**How the intermediate wrong turn read on screen, for the record.** Building
the tile-shift version first (correct overflow handling, still-wrong units)
removed the starburst and the diagonal split completely but left the sky a
dense fine moire — the *intended* mapping at 32x the intended scale. That is
what identified the unit error; the user independently flagged the same thing
on sight ("background looks like it tiles too small... sky i mean").

**Verified** (`-level_41` Cradle frame 350, the M-98 starburst repro):
`docs/img/bugs/d227-cradle-fixed-m100.png` — soft cloud bands, correct scale,
correct perspective compression toward the horizon, no starburst, no diagonal
split, no moire; matches the character of the emulator reference. `GE_D227V`
confirms `kS = kT = 0` (no shift needed) and every `tc` in range, max 31,513
of 32,767. `-level_22` Statue frame 345: clean horizon cloud band, no seam.
`tools_pc/verify.sh bunker1`: **PASS**, `worst_cell = 10.32` (the known
stale-golden-baseline number, unaffected by sky work). Surface1 frame 345 has
almost no sky in frame at that camera, so it is not evidence either way — the
user's own play-test screenshot remains the Surface evidence.

**What was kept, and why.** M-95's shared `wScale` and M-98's shared fold both
stay — they were real bugs, and per-fan consistency is still what keeps a
fan's shared vertices bit-identical between its primitives. The tile-shift
machinery (`skyPortPickShift` / `skyPortCaptureTile` / `skyPortEmitTileShift`)
also stays as a **safety valve**: the observed span is 31,513 against a 32,767
limit, only ~4% of headroom, so a camera that exceeds it is not far-fetched,
and the mechanism degrades gracefully instead of overflowing. It is not
untested code — the intermediate build above ran it at `k = 5` and it worked
correctly (coherent mapping; the wrong scale there was the unrelated unit
bug). It uses the RDP's own tile `shifts`/`shiftt`, which `port/fast3d`
already implements (`gfx_pc.cpp` ~1783), so it needs no fast3d change.
`skyPortCaptureTile` replays the exact `G_SETTILE` that `texSelect` emitted
with only the shift fields patched, rather than rebuilding it, so it cannot
drift from `texSelect`'s format/line/mask derivation.

**Still owed:** a human play-test. Every previous D227 "fix" (M-95, M-98)
passed a static headless check and then failed live, so this does not count as
closed until it is played — though unlike those two, this one has a mechanism
that explains every reported symptom, and the user has eyeballed the fixed
frame. Also worth re-measuring the scroll rate against M-98's ~7 px/tick
baseline: it should now be ~0.2 px/tick.

## D227 — M-100: ROOT CAUSE FOUND AND PROVEN — `s16` texture-coordinate overflow. Not a crack, not view-dependence, not a quantization mismatch. Fix not yet implemented (design decision open).

**Status: root cause proven by direct measurement + a confirming diagnostic
build. D227, D176(a) "Defect 2", and the M-98 "Cradle tower starburst" are
all the same defect.** M-95 (shared `wScale`) and M-98 (shared S/T fold) each
fixed a real but *second-order* bug; neither could have fixed this one.

**The unblocking evidence was a human screenshot pair** (user supplied, same
level / weapon / ammo count / near-identical framing, Surface):
`docs/img/bugs/d227-surface-emulator-ref.png` (1964 emulator — soft, blobby,
near-isotropic clouds) vs `docs/img/bugs/d227-surface-starburst-ours.png`
(our build — clouds smeared into long streaks radiating from a point, with
sharp *radial* boundaries). A radial starburst is not a crack between two
mis-quantized vertices, which is what M-94→M-98 had been chasing.

**M-99's "blocked, needs a human co-session" conclusion was wrong, and cost a
session.** The defect reproduces headlessly and *statically* — M-98 had
already captured it at Cradle frame 350
(`docs/img/bugs/d176a-cradle-tower-starburst.png`) and even confirmed frames
350/380 pixel-identical, but filed it as "pre-existing Defect 2, not D227."
That misclassification is what made D227 look view-dependent-and-unreachable.
Note the Cradle capture matches the *original* D227 report almost word for
word: left half starburst, right half normal horizontal streaks, hard
boundary between them — "each half of the screen shows a different sky
pattern."

**Measurement (new `GE_D227V` probe, `skyPortRenderPoly`, env-gated in the
same style as `GE_D176`).** `-level_41`, the frame-350 camera. The sky is one
`nverts=4` quad, corners TL/TR/BL/BR:

| i | screen x,y | w (`unk0c`) | S (`unk20`) | T (`unk24`) |
|---|---|---|---|---|
| 0 | 0.0, 10.0 | 6457.0 | 236.7 | 489.3 |
| 1 | 319.8, 10.0 | 6457.0 | -658.1 | -122.8 |
| 2 | 0.0, 212.9 | 215228.8 | 1422.2 | 29806.0 |
| 3 | 319.8, 212.9 | 215277.7 | -28408.5 | 9406.1 |

All four w are **positive** — no vertex behind the camera, no sign change, so
no projective pole (a hypothesis worth ruling out; it is ruled out).

The span across this single primitive is **29,830 texels in S and 29,929 in
T**. `Vtx.tc` is S10.5 (1/32-texel units) in an `s16`, so one primitive can
represent at most **±1024 texels**. The fold (`floor(min/64)*64`) rebases the
minimum vertex but cannot compress a span 29x larger than the entire
representable range. Computing `(S - foldS) * 32` for these four vertices
(`foldS = -28416`, `foldT = -128`):

| i | tc S exact | → `s16` | tc T exact | → `s16` |
|---|---|---|---|---|
| 0 | 916,886 | **-618** | 19,754 | 19,753 ok |
| 1 | 888,253 | **-29,252** | 166 | 166 ok |
| 2 | 954,822 | **-28,218** | 957,888 | **-25,152** |
| 3 | 240 ok | 240 | 305,091 | **-22,589** |

**Seven of the eight texture coordinates overflow and wrap to essentially
arbitrary values.** The rasteriser then interpolates between wrapped
garbage — which is exactly a starburst.

**Why it reads as "two halves that differ, split by a line."** The quad is
emitted as `gSP2Triangles(0, 1, 3, 0, 3, 2)` — split on the **0-3 diagonal**.
Vertices 0 and 3 are precisely the two that each keep *one* coordinate in
range, so the two triangles rasterise different mixes of wrapped values,
divided by that diagonal. This also explains the reported view-dependence
without any view-dependent mechanism: S/T are functions of view, so the
wrap point moves as you look around, and the garbage pattern changes with it.

**Confirming diagnostic (decisive).** Rebuilt with `tc` baked at 1-texel
instead of 1/32-texel granularity (32x the range, deliberately *wrong*
texture scale, nothing else changed), same `-level_41` frame 350:
`docs/img/bugs/d227-cradle-tcscale-diagnostic.png`. The starburst and the
two-halves split are **completely gone**; the sky is smooth and coherent and
structurally matches the emulator reference. The texture is over-zoomed, as
expected for a deliberately wrong scale. **s16 tc overflow is the whole
mechanism** — there is no residual second defect hiding behind it.

**Consequences for the existing record:**
- D176(a) "Defect 2" (adaptive tessellation of wide-span sky triangles) is
  not a separate cosmetic nicety — it *is* D227, and it is the dominant sky
  defect on every affected level.
- The `skyPortRenderPoly` code comment predicted this exactly ("Triangles
  whose S/T span exceeds ~1000 texels ... still need Defect-2's finer
  tessellation") but the severity was underestimated: it is not a
  horizon-only artifact, it corrupts the entire sky quad.
- M-95's and M-98's shared-`wScale`/shared-fold fixes remain correct and
  should stay: once spans are small enough for tc to be in range, per-fan
  consistency is exactly what keeps sub-primitive boundaries seamless.
- The M-96 "sky scrolls extremely fast" complaint and M-98's measured
  ~7 px/tick are plausibly the *same* bug (wrapped tc makes the apparent
  pattern move at a rate unrelated to the true 1 texel/tick scroll), but
  that is not yet confirmed — recheck after the fix.

**Fix not implemented — one design decision is genuinely open.** The fix is
to subdivide until each sub-primitive's S/T span is under ~960 texels, with
per-sub-primitive folds (folds differing by multiples of 64 texels are
seamless under `GL_REPEAT` at tile width 64, which is what makes subdivision
safe). Two things are already worked out:

- **Exact subdivision math.** Interpolate `(1/w, S/w, T/w, colour/w)`
  bilinearly in the quad's `(u,v)` and divide through — exact for any quad,
  because affine-in-screen functions commute with affine combinations whose
  weights sum to 1 (true of bilinear weights). A triangle is handled as a
  degenerate quad with the `v=1` edge collapsed, also exact.
- **Where to place the cuts.** Along an edge, `1/w` is affine in screen, so
  `S = (a + b·(1/w))/(1/w) = b + a·w` — **S is linear in w**. Subdividing
  *uniformly in w* therefore gives uniform S/T steps, and the cut count is
  just `ceil(span / threshold)`. Uniform-in-*screen* subdivision is badly
  wrong here: on this quad's right edge, half the screen distance carries
  807 texels and the other half carries 26,943.

**The open decision is the vertex budget.** `dynAllocateVertices` is an
unchecked bump allocator over a per-frame pool of `-mvtx50` = 50 KB (3,200
`Vtx`), shared with the whole scene; overrunning it corrupts memory silently.
A pure-tessellation fix needs ~32 divisions per axis at this camera, i.e.
~1,200-1,300 verts (~20 KB, ~40% of the pool) even with vertex sharing and
2-row `gSPVertex` batches (fast3d's `MAX_VERTICES` is 128, so batching is not
the constraint). Options:
- **A. Pure tessellation.** No port/fast3d change; costs ~40% of the frame's
  vertex pool on sky alone. Self-limiting via `dynGetFreeVtx()` with
  graceful degradation (slightly-over-threshold cells in extreme views).
- **B. Widen the texcoord transport in `port/fast3d/`** so tc carries more
  range, then only ~4x4 subdivision is needed (~100 verts). Cheap at
  runtime; touches the RSP emulation, which is port-layer and therefore
  permitted, but is a wider blast radius.
- **C. Hybrid** — a modest fast3d widening plus modest tessellation.
- **D. RDP tile shift (found last, and it dominates A/B/C — recommended).**
  The RDP already has a native mechanism for trading texcoord precision
  against range: the tile descriptor's `shifts`/`shiftt`. **`port/fast3d/`
  implements it fully, including the left-shift case** (`gfx_pc.cpp` ~1783:
  `shifts <= 10` divides, `shifts >= 11` multiplies by `1 << (16 - shifts)`).
  So emit tc at `32 / 2^k` units instead of 32 and set `shifts = 16 - k` to
  multiply it back — **2^k more range, no tessellation, no fast3d change, no
  vertex-budget cost, and an authentic RDP idiom rather than a port-only
  hack.** At `k = 5` tc is 1 texel per unit with a range of ±32,767 texels,
  which covers the measured 29,830-texel span outright — that is exactly the
  M-100 diagnostic build that already demonstrably removes the artifact, with
  the scale corrected by the tile shift instead of left wrong. Precision cost
  is a ≤0.5-texel rounding *at the vertices only* (≤0.8% of the 64-texel
  tile); interpolation between them is float in the GPU. Implementation care
  needed in two places: emitting the shift on the sky's tile without
  disturbing other draws (an extra `gDPSetTile` after the existing
  `texSelect`), and choosing `k` per draw from the measured span (`k = 0`
  when the span is already small). A light 2x2 subdivision can still be added
  on top for spans beyond ±32,767 texels if any camera produces them.


**Human playtest (same session, after M-98's fold-sharing fix was pushed):**
the seam is still visible on Statue, and — new data point, sharper than
anything in M-93/M-96 — **it visibly changes as the mouse/player view
moves.** Also reported: "still pretty bad" on Cradle, and Surface has the
same issue. M-98's fix was real (verified via a same-frame diff) but is
evidently not the dominant mechanism live play is hitting — the same
pattern as M-95's wScale fix before it.

**Ruled out this pass:** the cross-quad water/cloud texture-scale mismatch
theory (block1's below-horizon quad uses raw `unk0c`, block2's cloud quad
uses `unk0c*0.1f`, and each computes its own independent `wScale`/fold via
separate `skyPortBeginFan` calls — a plausible source of a horizon-tracking
seam). Checked via the existing `GE_D176` env probe: **both Statue and
Cradle report `IsWater=0`.** For `IsWater==0`, block1 is a flat
`gDPFillRectangle` (`sky.c` ~874-897) — no texture, no scroll, can't be the
source of a texture-pattern seam. This theory is **not ruled out in
general** — only on the two `IsWater==0` levels actually tested. It's
untested on any `IsWater==1` level (Frigate/Runway/Depot/Surface1/2 are
candidates — check `GE_D176=1` first, cheap).

**Blocked on: headless camera-pan repro doesn't work.** Tried reproducing
"changes with view" via `GE_INPUTSCRIPT` — both sustained analog-stick
(`SRIGHT`, held 90 ticks) and repeated C-button pulses (`CRIGHT`/`CUP`)
produced **zero visible camera rotation** across captured frames (identical
look direction/HUD framing throughout). The real in-game mouselook almost
certainly injects yaw/pitch through a path `GE_INPUTSCRIPT` doesn't reach —
it only emulates the N64 controller (buttons + one fake stick), and
turning during solo FPS play is evidently driven by something else
entirely (real SDL mouse deltas, likely, given D194's mouselook work).
**Any view-angle-dependent visual bug needs a human playtest to chase until
this gets a debug hook** (e.g. an env-gated `GE_FORCEVIEW=yaw,pitch` set
once per frame, or finding the actual mouse-delta injection point in
`port/src/input.c` and driving it directly instead of going through the
N64-controller emulation layer). Don't re-attempt `GE_INPUTSCRIPT`
stick/button tokens for camera turn next time — confirmed dead end.

**Ranked candidates for next session, in order of promise:**
1. Cross-block independent-fan state on an `IsWater==1` level (see above,
   not yet ruled out) — cheapest to check, `GE_D176=1` first.
2. Frame-to-frame popping at the S/T fold boundary (not a same-frame
   spatial crack) — `skyPortRenderPoly`'s fold recomputes every frame from
   the scrolling min-S, jumping by 64 texels whenever it crosses that
   boundary; never independently verified that fast3d's real texture wrap
   makes that jump invisible for this combiner. **Doesn't need view
   rotation to test** — a static-camera `step:1` capture bracketing a
   known fold-crossing tick (from M-98's ~7px/tick rate) would show it
   directly. Best headlessly-testable option if no human is available.
3. The clip-shape switch (`skyRender`'s 16-case corner-mask switch) handing
   block2 a differently-*shaped* fan frame to frame as screen corners
   cross the horizon, not just moved vertices — could read as "the pattern
   changes as you look around" without a same-frame crack. Not
   investigated.

**Concretely blocked on:** a screenshot or precise repro detail (level,
look angle/pitch, screen location) from a human. Full narrative + the
environment gotchas hit this pass: HANDOFF.md "M-99" (local, this
session — condense into this row if HANDOFF gets rotated before someone
picks this up).

## D176(a) — M-98: Cradle "tower starburst" — concrete screenshot of the already-known, still-open Defect 2 (horizon-band tessellation)

**Context:** checking the Cradle side of PR #43 (never done in M-97, per its
own "not done this session" note) for D227 regressions from the fold-sharing
fix below. It wasn't regressed, but turned up a much more visually dramatic
pre-existing artifact than anything previously screenshotted for this level.

**What it looks like:** near Cradle's spawn, looking up the suspension
tower/catwalk, roughly the left half of the sky renders as a sharp
"starburst" of radiating streaks converging toward a point near the top of
frame, directly adjacent to normal horizontal cloud streaks on the right
half. Screenshot: `docs/img/bugs/d176a-cradle-tower-starburst.png`
(`-level_41` frame 350).

**Confirmed pre-existing, not caused by this session's work:** rebuilt with
`src/game/sky.c` reverted to its pre-M-98 (and pre-M-95, since the revert
target predates both) state, same camera/frame — pixel-identical to the
post-fix capture. Also confirmed **static**, not the D227 "two halves move
at different rates" symptom: frame 350 and frame 380 (30 ticks apart, same
static camera) are pixel-identical.

**This is Defect 2, not a new bug.** The D176(a) finding row has carried
"Defect 2 (horizon-band coverage/adaptive tessellation) ... unchanged, still
open" since M-90, and `skyPortRenderPoly`'s own code comment already
predicts exactly this shape of failure: "Triangles whose S/T span exceeds
~1000 texels (only those straddling the horizon) still need Defect-2's
finer tessellation." Looking almost straight up the tower is precisely the
camera angle that produces one sky fan triangle spanning from near-camera
to horizon-grazing texel coordinates — the >1000-texel-span case the fold
logic already flags as unhandled. The perspective-correct interpolation
(M-90) gets the *shape* of the gradient right end-to-end, but with only 1-3
triangles per fan and no subdivision, a single triangle covering that much
of the texture's S/T range still shows visible linear-interpolation
artifacts across itself — reading as this radiating streak pattern rather
than smooth cloud, because the fold-and-interpolate math treats the whole
huge span as one linear ramp.

**Not fixed this pass** — tessellation (subdividing wide-span sky triangles
before the fold/perspective step) is a materially bigger change than the
fold-sharing fix below, and out of scope for "did Cradle regress." Logged
here with a concrete repro + screenshot so Defect 2 has hard evidence
instead of the prior "M-37: inconclusive, leaning tiling-density" note.

## D227 — M-98: seam residual FIXED (fold-sharing, same class as M-95's wScale-sharing) + scroll-rate quantitatively measured

**Picking up exactly where M-96/M-97 left off:** re-derive the intended N64
sky scroll rate and check the port's actual per-frame tc delta, rather than
re-attempting the seam in isolation.

**Ruled out first: the tick rate.** `skyTick()` (`src/game/sky.c`, verbatim
N64 logic, no `#ifdef PORT`) does `g_SkyCloudOffset += g_ClockTimer` once per
logic tick, wrapping at 4096. `g_ClockTimer` is set from `speedgraphframes`
(`frametiming.c`'s `updateFrameCounters`), which D193/D204 already
established runs at ~1/tick under normal load with a 6-tick catch-up clamp.
If the global tick rate were wrong, movement/AI/animation would all be sped
up too, not just the sky — that's not the report. Also checked
`sub_GAME_7F097388`'s call sites (`sky.c:862,1351`) for a second,
compounding time-based scale — both pass constant `65535.0f` args, not
anything frame-dependent. So the accumulator side is clean; per AGENTS.md
rule 2 (game logic is ground truth), any real bug has to be in the
`#ifdef PORT` D176(a) Path B substitution, not the decompiled game code.

**Found: M-95 only shared half of what needed sharing.** `skyPortBeginFan()`
(added by M-95) computes one `wScale` across the whole fan so shared-edge
vertices quantize their *position* consistently. But `skyPortRenderPoly`'s
S/T fold (`foldS`/`foldT`) — the `floorf(minS/64)*64` that keeps texel
coordinates in `s16` range — was still being computed **per call**, from
only that triangle's own 2-4 vertices, exactly the bug class M-95 fixed one
field over and never touched here. Two triangles in one fan can each
independently floor a shared vertex's S/T to a different multiple of 64
texels; the *visual* result is supposed to be identical (`GL_REPEAT` has
period 64), but each triangle bakes `(S - fold) * 32` into its own `s16`
`Vtx` buffer, and two folds 64 texels apart produce tc values exactly 2048
apart (`64 * 32`) — an s16-precision jump easily large enough to crack at
the shared edge. This is why M-96's human playtest still saw the seam after
M-95: M-95's fix was real (closed *a* crease, per its frame-345 diff) but
not the dominant one live play was hitting.

**Fix:** extended `skyPortBeginFan()`/`skyPortEndFan()` to also compute one
shared fold across the whole fan (`s_skyFanFoldS`/`s_skyFanFoldT`,
`s_skyFanFoldSet`), threaded into `skyPortRenderPoly` the same way as
`wScale` (falls back to the old per-call computation if unset — same
defensive shape as the wScale fallback). `src/game/sky.c`, `#ifdef PORT`
only, no game-logic touch.

**Verified:**
- Build clean (`./build-pc.sh ntsc-final`), no new/duplicate symbols.
- `-level_22` frame 345 (same frame M-95's before/after screenshots used) —
  fresh capture shows continuous diagonal cloud streaks with no visible
  crease at the ~x=1000/1280 location that's clearly cracked in the
  pre-M-95 reference screenshot (`docs/img/bugs/d227-sky-seam-statue.png`).
  No new screenshot committed this pass (didn't diff against an M-96-era
  "still broken" capture, since none was saved) — human eyeball still owed
  per the usual rule for visual fixes landed without display access.
- `verify.sh bunker1`: PASS, `worst_cell=10.3` — unchanged from before this
  fix, matching M-95's note that this number is the already-known stale
  golden-baseline issue (owed a re-baseline, unrelated to sky.c), not a
  regression.

**Scroll-rate measurement (the actual M-96/M-97 ask):** captured 11
consecutive frames (`GE_PCDUMP=345-355:1`, static camera, Statue spawn) and
cross-correlated a sky-band row (y=60) frame-to-frame. Result: a highly
consistent **~7 px/tick horizontal shift at 640px capture width** (346→347
through 354→355 all landed at 6-7px, one outlier at the 345→346 transition
likely a cutscene/fade boundary). At 60 ticks/sec that's a full 640px
screen-width sweep roughly every 90 ticks (~1.5s) — hard quantitative
confirmation of "extremely fast," not just a subjective read, and a
concrete number for whoever chases the magnitude next.

**Not resolved this pass: *why* that magnitude is what it is.** Since the
accumulator (1 texel/tick) and the vertex geometry that turns a texel delta
into screen pixels (`sub_GAME_7F097388`) are both unmodified N64 logic, a
genuine wrong-magnitude bug (if the rate really is wrong relative to N64,
which still isn't confirmed absent a hardware/emulator reference) would
have to live in how the D176(a) Path B substitution maps a given per-vertex
S/T value to on-screen texels — e.g. camera-to-cloud-plane distance, FOV,
or a scale factor implicit in the `skyPortRenderPoly` perspective matrix —
not in the scroll-rate accumulator itself. Didn't chase this further this
pass; next session should compare the recovered per-pixel S gradient
(`tc[0]` delta between adjacent screen columns within one frame, not
frame-to-frame) against what the `130.0f`/`arg4`-driven N64 edge-setup math
in the `#else` branch of `skyRenderTri` computes for the same triangle, to
find whether the *shape* of the perspective correction (M-90) also carries
the right *magnitude*, independent of the wrap-rate question addressed here.

**Still open / human-gated:** the seam-close and the scroll-speed data
point both need a live eyeball before this can come off draft — same as
every other rendering fix landed without display access this session.
Cradle side of PR #43 still never checked (Statue only, again, this pass).

## D227 — M-96: human playtest — seam/crack STILL PRESENT after the M-95 fix; the "static-crease" theory alone does not explain the reported symptom

**Method:** user's own build, this branch (`land/d176a-sky-perspective-scroll`
at `b1733eda`, which includes the M-95 `wScale`-sharing fix). Launched
directly into Statue (`-level_22`), looked at the sky.

**Result: sky renders (the original D176a black-sky bug is fixed — sky is
visibly present, not black), but the seam/crack is still there.** This
directly contradicts the M-95 commit's "closes the visible seam" claim.

**New scroll-speed data point (user, same session): the sky should be
*almost static* on the original N64 game, and on this build it currently
scrolls *extremely fast* — a stronger complaint than M-93's "measurably
better but still reads too fast."** This bears directly on D176(a) itself
(the perspective-correct tc-interpolation fix this whole PR is about), not
just the D227 seam residual. Two live symptoms reported together on the
same Statue session: (a) scroll rate way too fast, (b) the seam. Worth
treating as possibly related — if the effective per-frame tc delta is
wrong by a large factor, that same wrong-magnitude value could plausibly
also drive the per-vertex/per-primitive divergence read as a "seam" (the
D227 M-94 hypothesis assumed *quantization* was the cause of the seam, but
never ruled out a *scroll-rate* mismatch between primitives as an
alternative explanation for "two things moving at different rates"). **Not
yet investigated — next session should re-derive the intended N64 scroll
rate from `sky.c`/the ROM env data and check what the port's actual
per-frame tc delta computes to, rather than assuming D176(a)'s fix already
has the right magnitude just because it's no longer solid black.**

**What this means for the M-94/M-95 diagnosis.** M-95's write-up was explicit
that it verified only the *static per-primitive `wScale` quantization* half
of M-94's hypothesis (via a same-frame before/after screenshot diff), and
flagged the *"each half moves at a different rate"* half of the original
user report as **not re-verified** — that measurement (M-94's static-camera
diff) was never repeated against the M-95 build. This playtest is that
missing verification, and it says the crease is still visible in live play.
Two non-exclusive readings:

1. The screenshot-diff comparison (single static frame) wasn't sensitive to
   whatever actually causes the live-play seam — i.e. the M-95 fix may have
   closed *a* crease (real, per the frame 345 before/after) without being
   *the* crease the user is seeing, or without being the dominant one.
2. There is a second, independent seam-causing mechanism — most likely the
   still-unconfirmed "two halves animate at different rates over time" theory
   from the original M-93/M-94 report, which is a *temporal* divergence
   (rates, not one-frame geometry) that a single-frame diff cannot detect at
   all.

**Status: D227 NOT resolved. PR #43 stays in draft — do not merge on the
strength of the M-95 fix alone.** Next investigation should target the
temporal/rate theory specifically: capture the same fixed camera position
across a range of frames (not just one) on both sky "halves" and measure
whether their texture-coordinate scroll rate actually diverges over time,
per M-94's original (but never re-run post-M-95) methodology.

## D227 — M-95: FIXED (the static-crease part) — one shared `wScale` across the whole sky fan instead of one per triangle

**The fix.** `src/game/sky.c`, `#ifdef PORT` throughout, three additions:

- `skyPortBeginFan(SkyRelated38 *v, s32 n)` — scans `n` verts' `unk0c` for the
  max absolute camera-space `w` and stores `wScale` in a new file-static
  `s_skyFanWScale`. Forward-declared near the top of the file (its first use
  is ~600 lines before its definition).
- `skyPortEndFan(void)` — resets `s_skyFanWScale` to `0.0f` (the "unset"
  sentinel).
- `skyPortRenderPoly` now checks `s_skyFanWScale > 0.0f` first; if set, uses
  it directly instead of recomputing `maxAbsW`/`wScale` from just its own
  2-4 verts. Falls back to the original per-call computation if unset —
  `skyRenderTri`/`skyRenderFull` have no callers outside this file (checked
  by grep), but this keeps a safe default instead of silently reusing a
  stale value if that ever changes.
- Both fan-drawing blocks (the `sp274[5]`/`sp94[5]` local arrays,
  `src/game/sky.c`) call `skyPortBeginFan(array, s1)` once right after
  populating the array — before dispatching to their 1-3
  `skyRenderTri`/`skyRenderFull` calls — and `skyPortEndFan()` once after
  the last one. For the `sp94[]` block specifically: computed from `unk0c`,
  which the position-override code for the full-quad special case (`s1==4`,
  the `sp94[0].unk28 = ...` assignments) does not touch, so the order is
  safe regardless of whether that override path runs.

**Why this was the right diagnosis.** `skyRenderTri`/`skyRenderFull` draw up
to 5 triangles per frame from a *shared* pool of up to 5 vertices (built
once, then fan-triangulated across 1-3 calls). Each call used to compute its
own `wScale` from only its own 2-4 verts. The position math is exact for any
`wScale > 0` (re-derived algebraically in M-94), so this was never a logic
bug — but `Vtx.ob[]` is `s16`, and two triangles sharing a vertex could each
independently pick a different `wScale`, quantizing that logically-identical
shared vertex to two different 16-bit values depending on which triangle's
call computed it. That's a crack at the shared edge — exactly the visible
"seam" from M-94's screenshot.

**Verification.**
- `./build-pc.sh ntsc-final`: clean build, no new/duplicate symbols.
- Full `verify.sh sweep`: **21/21 PASS**, no crash/regression, both with and
  without this change (isolated by `git stash`/`stash pop` and re-running).
- `bunker1`'s non-zero `worst_cell` framediff is unaffected by this change —
  reproduced identically (~64) on a pre-fix build. Confirmed pre-existing:
  the stale golden-baseline issue already noted in the M-83/M-84 memory
  (owed a re-baseline), unrelated to `sky.c`.
- **Direct same-frame before/after comparison, `-level_22` frame 345:**
  `docs/img/bugs/d227-sky-seam-statue.png` (pre-fix — visible kink in the
  cloud streaks around x=1000/1280 in the cropped+2x-upscaled screenshot)
  vs. `docs/img/bugs/d227-sky-seam-statue-fixed.png` (post-fix, same exact
  frame — streaks flow continuously through that region, kink gone).
- **Cradle checked too** (`-level_41`, the other level named in the M-93
  report): the large black rectangular gap visible there is confirmed
  **pre-existing and unrelated** — reproduced identically on a pre-fix
  build capture at the same frame. That's the already-documented
  FOV-coupled black-sky-past-frustum-edge symptom (D218/D222 family, "not
  attempted" per the D176(a) PR body) — a different defect, not touched by
  this fix, not chased further here.

**Not re-verified post-fix:** the "each half moves at a different rate"
half of the original user report (M-94's static-camera diff measurement
was not repeated against this build) and the aspect-ratio-artifact
alternate theory (never independently confirmed or ruled out either way).
The crease visibly closing at a fixed resolution/aspect is reasonable
evidence the crease itself was code-side, not a display-scaling artifact,
but this is not a substitute for a human eyeball pass — owed before merge,
consistent with every other rendering fix landed this session without
interactive display access.

**Generalizable quirk for `porting-notes.md`:** when a decompiled routine
fans a single shared vertex set out across multiple separate perspective-
correction / projection-matrix-building calls (one call per triangle
instead of one call for the whole primitive group), and the fix stores a
derived per-call scale factor into a low-precision (`s16`) vertex field,
each call's independently-chosen scale can quantize a *logically shared*
vertex differently — producing a visible crack at the shared edge even
though the underlying position math is exact for any scale value. Fix:
compute the scale once for the whole shared-vertex group, not per call.

## D217 — M-89: the `gfx_dp_load_tlut` width=1/pitch hypothesis is REFUTED (it computes correctly by design); new lead pointed at the runtime texpool arena + image-table conversion, not the model sidecar

A local-Qwen triage dispatch (read-only, no edits) proposed hardcoding
`pitch = 1` in `gfx_dp_load_tlut` (`port/fast3d/gfx_pc.cpp:2214`,
currently `pitch = rdp.texture_to_load.width + 1`), on the theory that N64
LOADTLUT addressing never strides by the SetTextureImage width and the
port's `+1` computation was wrong. **Verified by hand and refuted before
applying it — do not repeat this fix:**

- `gDPSetTextureImage`/`gSetImage` (`include/PR/gbi.h:3097-3111`) pack the
  RDP command word as `_SHIFTL((width)-1, 0, 12)` — i.e. the **raw** command
  field the port reads back (`C0(0,10)` at the `G_SETTIMG` handler,
  `gfx_pc.cpp:2786`) is `width - 1`, not `width`. `texWriteLoadToTmemAddr`
  (`tex.c:503,519,629,643`) always calls `gDPSetTextureImage(..., width=1,
  ...)`, so the raw field is `1-1=0`, and `pitch = 0 + 1 = 1` **already**,
  for every model-DL TLUT load in the game. The proposed patch is a no-op
  for this call site — it changes nothing about the reported bug, because
  the code it "fixes" already computes the value it wants.
- Continuing the derivation with `pitch=1` confirmed: `gfx_dp_load_tlut`'s
  `base = addr + pitch*ult + uls` reduces, for every call from
  `texWriteLoadToTmemAddr`'s `gDPLoadTLUT07(uls, ult, tex->unk0a + uls,
  ult)`, to **`base = addr + len`** regardless of the internal `uls`/`ult`
  split (`tex.c:530-541`; the split exists only to keep both RDP command
  fields inside their 10-bit range when `len` is large — it cancels out in
  the `pitch=1` address math). `addr` is `tex->data` (the `SetTextureImage`
  dram pointer) and `len` is `texGetDepthAndSize`'s halfword-converted
  pixel-data length. This is exactly the documented on-disk/in-arena layout
  from `texInflateZlib` (`image.c:320-326`: palette bytes appended
  immediately after `totalbytesout` bytes of pixel data) — **the addressing
  scheme is self-consistent and correct by inspection**, assuming `len`
  (computed from `tex->width`/`height`/`depth`/`maxlod` via
  `texGetLineSizeInBytes`) truly equals the pixel-byte count the decoder
  actually wrote (`totalbytesout` before the palette-append loop). Also
  confirmed `tex->unk0a = numcolours - 1` (`image.c:238`) exactly explains
  the `count = width*height = unk0a+1` term in `gfx_dp_load_tlut` — matches
  the M-85 `GE_D217Z` "count=1/28/125/172/193" observations precisely (a
  **genuine 1-color palette** for the grip tile, not a partial/truncated
  load of a larger one).
- **New question this raises, not yet answered:** if the addressing math is
  correct and `numcolours=1` is a real, intentional 1-entry palette for
  this tile, then the SAME `texturenum` decoded from the SAME ROM bytes
  should produce the SAME single color on every level — it does not
  (`0001` on BUNKER1 vs `1a00` on FACILITY). That points away from a
  decode-math bug entirely and back toward either (a) **the bytes handed to
  the decoder differ per level** (i.e. an image-table/asset-resolution
  issue — but a *different* one than the already-ruled-out model-DL
  `TextureID`/`d43_emit.py` path; `texLoad` sources compressed bytes from
  `g_Textures[texnum]` offsets into `_imagesSegmentRomStart`, a wholly
  separate table/segment from the per-model `Globalimagetable` that
  `port/src/gimgfixup.c`/D68 already fixes — **not yet audited for this
  bug**), or (b) **the shared texpool arena serves a stale/aliased
  `tex->data` pointer** for what should be a fresh per-level decode (the
  arena is a two-ended allocator, `struct texpool` `leftpos`/`rightpos`,
  reused across many textures per level per `texLoad`/`image.c:2414-2519`;
  if `bytesout` returned by `texInflateZlib`/`NonZlib` for the LOD-limited
  or shrink-paletted paths (`image.c:277-318`, the `forcenumimages >= 2`
  branch with its `if (totalbytesout + imagebytesout > 0x800) { maxlod = j;
  break; }` early-exit) doesn't exactly match what `texGetDepthAndSize`
  later assumes was written for a *different* texture that landed nearby,
  a later texture's palette read could land on an adjacent texture's
  bytes — this would explain "same weapon, different levels, different
  wrong colors" as a level-dependent *decode order/adjacency* effect, not
  a fixed byte offset).
- **Explicitly ruled out this session (in addition to the M-88 sidecar
  ruling):** `gfx_dp_load_tlut`'s pitch/address computation itself
  (`gfx_pc.cpp:2202-2237`) — confirmed correct by hand-derivation against
  the GBI packing macros, not just "not yet disproven."
- **Confidence:** HIGH that the proposed `pitch=1` patch is a no-op and
  must not be applied/merged. MEDIUM on the two new leads above — neither
  was instrumented or built this session (time-boxed to the math
  derivation); the game-fed vs. arena-staleness question is the right next
  probe (an env-gated log of `g_Textures[texnum]` offset + decoded byte
  count + `pool->leftpos`/`rightpos` delta, on the exact FACILITY repro
  frame, would disambiguate directly).
- **For whoever picks this up next:** do NOT re-propose changing
  `gfx_dp_load_tlut`'s pitch formula without first checking whether the
  bytes arriving at `texInflateZlib`'s `src` parameter (i.e. what
  `g_Textures[g_TexNumToLoad]`/`_imagesSegmentRomStart` resolves to) are
  actually identical across BUNKER1 and FACILITY for this `texturenum` —
  that is the one load-bearing assumption ("same weapon, same ROM bytes")
  this whole investigation has taken for granted but never directly
  checked with a probe.

## D217 — M-89 cont.: decisive probe — decode is proven correct+identical across levels; the parked defect ② fix does NOT fix the visible symptom; new leading theory is a stale/dangling `tex->data` arena pointer, not a decode or cache-key bug

Ran the actual empirical test the M-89 section above called for, rather than
delegating it further. **Result answers the open question decisively and
rules out both of this row's previous leads.**

**Probe added** (`src/game/image.c`, `texInflateZlib`, `#ifdef PORT` +
`getenv("GE_D217TEX")`, follows the `GE_TEXDUMP`/`GE_D85TEX` pattern):
logs `texnum`/`format`/`numcolours`/`palette[0]`/`palette[1]` **straight off
the decode bitstream**, before any arena/pointer math — the most direct
possible check of "does this texture number decode to the same bytes on
both levels."

**Result:** ran BUNKER1 (`-level_09`) and FACILITY (`-level_34`) headless
with `GE_D217TEX=1`. `texnum=1608` is the only `numcolours=1` texture common
to both logs, and its `pal0` value (`0001`, black) matches the previously
documented BUNKER1-correct value exactly — confirming this is the PP7 grip.
**`texnum=1608` decodes to `pal0=0001` on BOTH levels, identically, every
time it's called.** This directly disproves the M-89 "different ROM bytes
per level" open question (lead a is dead — the decode is correct and
level-invariant) and also rules out any remaining decode-math theory: the
bug is not in `texInflateZlib`, `texGetDepthAndSize`, or `gfx_dp_load_tlut`
(already shown correct M-89) — it is **strictly downstream of a provably
correct decode**.

**Then tested the parked defect ② fix (`fix/d217-model-texture-palette`,
`f6fc0399` — the CI-texture-cache-key content-hash) directly against this
exact FACILITY repro**, since "correct decode, wrong render" is precisely
the signature that fix targets (a stale cache `HIT` serving a different
material's already-decoded palette). **Built it, ran `-level_34`, captured
frame 380: the grip is still solid green.** The already-written, previously
untested-against-this-repro defect ② fix does **not** fix the visible
symptom. This means defect ① and defect ② are **not the same underlying
bug** — they were right to be tracked separately, and defect ②'s fix should
still land (real, separate, still needs its 21-level sweep) but will not
close this row on its own.

**New leading theory (untested, most concrete lead yet):** since a fresh
decode is proven correct, and the fast3d-side cache-key fix doesn't help,
the render-time GBI command for this material must be reading a `tex->data`
pointer whose *arena contents have changed* between when it was decoded and
when the DL that references it is actually drawn — i.e. a **stale/dangling
pointer into the `texpool` arena**, not a decode or cache-lookup bug. GE's
`texpool` (`image.c` `texLoad`/`texInitPool`, two-ended `leftpos`/`rightpos`
allocator) is reused across many textures per level; if it resets or the
model-node/GDL that references texture 1608 caches a raw pointer/segment
address resolved once (e.g. at weapon-pickup or level-load time) rather
than re-resolving through `texFindInPool` every time it draws, then any
*later* texture decode that reuses that same arena address (plausible on a
texture-heavy level like FACILITY, less likely on the lighter BUNKER1) would
silently overwrite the grip's palette bytes with whatever the next tenant's
content is — explaining both "same weapon, different wrong color per level"
(depends on what else that level's arena happened to decode nearby/after)
and "clean on the lighter level" (less arena churn/reuse pressure).

**Confidence:** HIGH the decode itself and the fast3d/cache-key layer are
now both cleared (empirically, not just by inspection). MEDIUM-LOW on the
stale-arena-pointer mechanism — plausible and consistent with all evidence
so far, but not yet probed directly.

**For whoever picks this up next:** probe whichever code path holds the
`Gfx*`/model-node reference to this material's `SetTextureImage` address at
*draw* time (`texWriteLoadToTmemAddr`'s caller chain from the weapon
model-DL walk, not `texLoad` itself) and log the `tex->data` pointer value
+ a fresh read of the two bytes at the computed palette offset **at draw
time**, alongside a second log of what `texInflateZlib` most recently wrote
to that same address (extend the `GE_D217TEX` probe with an address field
to correlate). If the drawn address's live content no longer matches what
was last decoded there, that confirms the stale-arena-pointer theory
directly and narrows the fix to wherever the model DL caches that pointer
instead of re-resolving it. Probe left in tree (`GE_D217TEX`, env-gated,
inert by default).

**VERIFY performed this session:** `tools_pc/verify.sh bunker1` — PASS,
no regression from the (inert-by-default) probe. `-level_09`/`-level_34`
`GE_D217TEX=1` headless captures — decisive result above. Built and ran
the parked defect ② fix directly against `-level_34` frame 380 — still
green, screenshot evidence in this session's scratchpad.

## D217 — M-90: defect ① ROOT CAUSE FOUND + FIXED — texpool arena base misaligned on PC; the palette sits 4 bytes early inside every weapon-pool allocation while LoadTLUT still reads the N64 offset

**Verdict: the stale-arena-pointer theory (M-89 cont.) is REFUTED, and the
real root cause is found, fixed, and verified.** Arena bump allocations are
never overwritten in place (`leftpos` only advances), so a drawn address's
content cannot "change" after its decode. What actually happens: the palette
**was never at the address the draw reads**.

**The mechanism (all hand-derived, then probe-confirmed):**

1. `texAlignIndices` (`src/game/image.c`) pads each CI index row to an
   **absolute** 8-byte boundary: `outptr = (u8 *)(((u32)outptr + 7) & ~7)`.
   The palette is then written at `dst + imagebytesout`, and the whole
   allocation's size (`texInflateZlib`'s return) depends on where `dst` sits
   in memory.
2. Model-DL TLUT loads address the palette at `tex->data + len*2` bytes
   (`texWriteLoadToTmemAddr`: `uls = len` from `texGetDepthAndSize`, and the
   port's `gfx_dp_load_tlut` computes `base = settimg_addr + pitch*ult + uls`
   in 16-bit units — correct RDP semantics, re-verified). `len` is a
   **formula** that only matches the real layout when rows start 8-aligned.
3. The invariant that makes (1) and (2) agree: **`leftpos` stays 8-aligned
   for life** — pool base 8-aligned, header `+= 8`, every inflate size a
   multiple of 8 (checked: zlib aligns at return; non-zlib per-image). On
   N64 the memp bump allocations feeding these pools happened to be
   8-multiples, so bases were aligned and absolute ≡ relative row padding.
4. **PC broke the invariant.** `mempAllocBytesInBank` (`src/memp.c`) is a
   plain bump allocator with no per-allocation alignment; PC pointer bloat
   (D36/D37-class struct growth) means some STAGE-bank allocation sizes are
   no longer 8-multiples, so by the time `initBondDATA.c:238` allocates the
   weapon buffers, the pool position is off. Probe-confirmed: on FACILITY
   both weapon-pool bases land **%8==4** (stage pool aligned), and *every*
   decode in the misaligned pool has `dst%8==4`.

**The smoking gun (probe data, `GE_D217TEX` extended with a
`D217TEX-SIZE` line at `texInflateZlib`'s return):** pre-fix FACILITY,
`texnum=1608` (PP7 grip, 1x1 CI4, `numcolours=1`, `pal0=0001`) decodes to
**size=8**: image = 1 byte + absolute-align padding = **4 bytes**, palette
at `dst+4..5`. The draw's count=1 TLUT reads `dst+8` (the N64 layout: image
8, palette at +8) — which is exactly the **next tenant's 8-byte header slot**:
`texnum=26` stored as an LE s16 = bytes `1a 00` → `PD_BE16` → **`0x1A00` =
the green `(24,65,0)`**. Permanent (arena never rewrites it), per-level and
per-weapon (the "next tenant" texnum differs → any RGBA5551 value: the green,
blue, white, orange of the M-87 playtest matrix).

**This single mechanism also explains M-86's long-unexplained tell:** every
*larger* model TLUT showed `raw[0]==raw[1]==0x0000` with real colour from
index 2. In a %8==4 pool the whole palette sits **4 bytes (2 entries) early**,
so a load at the N64 offset reads two zero-pad entries then the real palette.
The "leading zero pad / pointer 4 bytes too low" suspicion was literally
true — just on the *decode* side, not the draw side.

**Fix (landed):** `texInitPool` (`src/game/image.c`, `#ifdef PORT`): round
the arena base up to 8 bytes before publishing `start`/`leftpos`; `end`
stays at the original buffer limit (pool loses ≤ 7 bytes). A no-op on N64
(bases were already aligned — the game works there, which is itself the
proof the invariant held). Chosen over aligning inside
`mempAllocBytesInBank` (would shift *every* allocation — far larger blast
radius, and D36/D37 note the STAGE bank is nearly full). This is a
semantics-preserving ABI/layout edit in the AGENTS.md narrow-exception class
(32→64-bit transition broke an implicit layout invariant); no game logic
touched.

**Verification:**
- `./build-pc.sh ntsc-final` clean; no new symbols (local var + expression).
- FACILITY `-level_34`: all 188 decodes now `%8==0`; `texnum=1608` size=16,
  palette at +8; draw probe reads `addr=<base+8> pal0=0001 count=1` (black,
  the decoded value) — zero `pal0=1a00` events. PPM frames 250–450: **0**
  exact-green `(24,65,0)` pixels (pre-fix: ~1300 in frames 300–410) and 0
  strong-green (`g>90, r<70, b<70`) across all 21 frames.
- BUNKER1 `-level_09` (the clean control): still clean — 183/183 decodes
  aligned, 0 CHANGED draw events, 0 green pixels in the capture window.
- DEPOT `-level_30` (had green pre-fix): 229/229 aligned, 0 exact/strong
  green in the capture window.
- Full 21-level `verify.sh sweep`: **21/21 PASS**, no CRASH/REGRESSION/
  STALLED (`sweep_d217.log`, repo-root scratch).

**Status change:** defect ① = **FIXED, pending the sweep + a human
playtest eyeball** (the M-87 matrix weapons: PP7/PPK/AK47/M16 grip+body,
watch face, NPC-held guns). Defect ② (CI texture-cache content-hash key,
branch `fix/d217-model-texture-palette` @ `f6fc0399`) remains a real,
separate, parked fix — still owes its own sweep before merge; M-89 already
proved it does not fix the visible symptom.

**Probes:** stripped before commit — the fix ships clean. The committed
M-89 decode probe (`GE_D217TEX` in `texInflateZlib`) remains as left by that
session; if the playtest eyeball surfaces a residual, re-add the
`D217TEX-SIZE`/`D217TEX-DRAW` probes exactly as documented above.

## D217 — M-92: human playtest eyeball of PR #47 (debug build) — PP7/PPK/watch CONFIRMED FIXED; AK47 and NPC/world-model weapons have a DISTINCT, still-open residual

**Method:** user's own debug build (`build-linux-dbg`, their preferred manual
test flow, not the headless `verify.sh` capture), `fix/d217-texpool-alignment`
checked out. Sampled Dam, Facility, Jungle, Frigate.

**Confirmed fixed by this PR:** PP7/PPK (grip + body) clean on every level
tested; Bond's watch face clean where it was previously green. Matches the
probe/pixel-count evidence in M-90 — the %8==4 texpool-alignment fix is
real and closes the exact symptom it targeted.

**NOT fixed — residual, still open:**
- **AK47** — both **1P viewmodel and 3rd-person world model** still show
  white (missing-texture) patches.
- **NPC-held weapon world models** — white patches on parts that vary per
  gun (magazine, receiver, etc. depending on weapon), **consistent across
  guns/levels** (not intermittent). Frigate was called out as especially bad.

**Reading this against the M-87 raw notes above:** the M-87 matrix already
recorded AK47 white-patch symptoms (viewmodel) pre-fix; M-92 shows the
texpool-base-alignment fix (M-90) did not touch that mechanism at all — it
only fixed the PP7/PPK/watch "wrong solid colour from a misread palette
slot" failure mode. AK47 + NPC-world-model "white/missing" is a different
failure signature (white usually means "texture never bound / fell back to
a default," not "bound the wrong palette entry" — cf. the D161 family this
row already flagged). **Likely a second, distinct D217 mechanism** — not
yet root-caused. Candidates to check first, cheapest to most expensive:
whether the AK47/NPC weapon model DLs route through a *different* texpool
(not the one `texInitPool`'s 8-align fix touches — note `gun.c:930-950`
hand multiple distinct offset/pool pairs per weapon, and `bondview2.c:504`/
`initmttex.c:15` are separate `texInitPool` call sites entirely from the
per-weapon ones); or a genuinely separate "texture never loaded for this
DL" bug (D161-class G_TT_NONE / missing-TLUT-emission) rather than a
misaligned-read one.

**Status:** PR #47 fix is real and should still merge — it closes a
confirmed sub-defect (PP7/PPK/watch) with no regressions. **Do not close
D217 defect ① on this merge** — reopen/retitle as the AK47 + NPC-world-model
white-texture residual once root-caused, since the M-87 matrix already
described it as a separate colour-vs-white failure mode from the start.

## D221 — M-91: SELECT FILE mouse hit-box offset — root cause is a compiler stack-layout dependency, not a coordinate/origin mismatch; fixed by passing explicit `{min,max}` pairs (`#ifdef PORT`)
**Symptom (M-87 user QA):** on the SELECT FILE screen (default
`MouseCaptureMode=1`, free OS cursor), clicking directly over a visible
file wallet does nothing; you have to click *below* it to select.

**The hit-test.** `interface_menu05_fileselect` (`front.c`) loops the four
folders, gets each wallet model's world-space extents
(`modelGetXYExtents(walletinst[n], &xmax, &xmin, &ymax, &ymin)` — four
separate f32 locals) and projects them with
`projectRectCornersTo2D(&folderpositions_camspace[n], &xmin, &ymin,
&folderbbox.right, &folderbbox.left)`. The function's signature takes
`struct coord2d *` **pairs** — `{f[0]=min, f[1]=max}` — and projects
`arg1->f[0]`, `arg1->f[1]` (horizontal) and `arg2->f[0]`, `arg2->f[1]`
(vertical). Passing `&xmin` therefore *silently reads the stack float
physically after `xmin`* as "xmax", and `&ymin` reads the float after
`ymin` as "ymax". The decompiled C only reproduces the N64 binary's
behavior if the N64 compiler happened to lay the four locals out with each
max adjacent to its min (reverse declaration order) — which it evidently
did, since file select works on N64.

**The PC build breaks that assumption.** Probe (`GE_D221`, temp,
removed before commit) logged the actual adjacency in the MinGW build:
`((f32*)&xmin)[1] == ymax` (declaration-order locals: `xmax,xmin,ymax,ymin`) and
`((f32*)&ymin)[1] == 2.25734`, later `nan` — an unrelated stack slot. So the
projected hit box used `(x=xmin, x=ymax-as-X)` horizontally (a valid but
over-wide band) and `(y=<garbage/NaN>, y=ymin)` vertically — a ~14 px strip
at the *bottom edge* of each wallet (f=0: `[164.8, 179.0]` vs the true
folder extent `[36.7, 179.0]`), and NaN in other frames (the band check
`down <= v && v <= up` is false against NaN → unselectable that frame).
A ~14 px bottom-edge strip is exactly "hovering over the file does nothing;
click below it works." The horizontal over-width is why left/right felt
fine. The other caller, `objective_status.c:525`, passes genuine
`struct coord2d` locals (`sp7C`/`sp74`) and was never affected.

**The fix (game code, ABI/layout exception per AGENTS.md rule 2).** In
`interface_menu05_fileselect`, under `#ifdef PORT`: declare
`struct coord2d xminmax, yminmax;`, fill them from the four extents, pass
those to `projectRectCornersTo2D`. The N64 path keeps the original call
byte-for-byte. This is the same class as the D3x pointer-width edits:
the decompiled source encodes a memory-layout invariant the PC toolchain
does not honor; the edit makes the C express what the N64 binary already
did, changing no logic or values on either platform.

**Verification (headless, controlled).** Title screen → any button reaches
file select (`GE_INPUTSCRIPT="300:A;900:A"`). The default cursor sits at
(154.7, 127.8), inside folder 0's *correct* box but **outside** its broken
band (y=127.8 ∉ [164.8, 179.0]). Pre-fix: A-press there never selected
(the probe kept printing for the whole 60 s run — 4392 lines). Post-fix:
same script, folder 0's box is `[x 96.5..214.9] × [y 36.7..179.0]`, the
A-press selects folder 0 and the file-select loop exits (probe output
stops after ~59 frames; no crash/BUG lines). Identical cursor position and
button press on both sides — only the box changed.

**Status:** fixed in `front.c` (PR pending), probe stripped. Human
eyeball owed: normal mouse play through file select (hover highlight,
select/copy/delete options) — the headless test proves the hit-test fires
where the visible wallet is, but the highlight/selection *feel* is a human
check. D192 (mode-0 clamp range) is a separate bug and remains open.

## D228 — M-97: AK47 white missing-texture patches survive D217's texpool-alignment fix — split out as a distinct defect (white ≠ wrong-colour)

**Method.** PR #47's branch (`fix/d217-texpool-alignment`) merged up to date
with `main`, rebuilt clean, launched directly into FACILITY (`-level_34`) —
the level the D217 log names as the sharpest repro. Human eyeball across the
M-87 weapon matrix.

**Result — a clean split:**

| Symptom (M-87 matrix) | After PR #47 |
|---|---|
| PP7/PPK grip + hammer green | **FIXED** — correct colour |
| Bond watch face green | **FIXED** |
| Green cast on hands / 3rd-person world model | **FIXED** — none seen |
| **NPC-held AK47 white patches (magazine/receiver)** | **STILL PRESENT** |
| General world/HUD textures | fine, no regressions, no crash |

This is the second independent human playtest to land on exactly this split
(M-92 was the first) — and it is the split M-90's write-up predicted in
advance, which is good evidence it is a real boundary between two mechanisms
rather than an artifact of one incomplete fix.

**Why these are different mechanisms, not degrees of one.** D217's root
cause (M-90) is *a wrong palette being bound*: the texpool arena base landed
`%8==4`, so every palette in the affected pools sat 4 bytes early and
`G_LOADTLUT` read the neighbouring tenant's header bytes as colour — the
tile decodes fine, it just decodes against the wrong 512-byte TLUT, which
necessarily produces a **wrong colour** (the documented green `0x1A00`).
A **white** patch is a different signature entirely: it means the tile is
sampling a texture that is missing, never uploaded, or cleared — not one
that is present with wrong contents. Aligning the arena base cannot fix a
texture that was never decoded into the arena in the first place, which is
precisely what the playtest shows.

**Why the label split.** Leaving this inside D217 would mean the row can
never close cleanly: its own stated mechanism *is* fixed and sweep-verified,
and carrying an unrelated open failure mode under the same label is how a
finding log rots (cf. the D170→D193 subsumption that had to be untangled
later). D228 takes the white-texture residual; D217 closes on the green
mechanism it actually describes.

**Next (not started).** Run `-level_34` with `GE_D217TEX=1` / `GE_TEXDUMP`
while an AK47-carrying guard is on screen and determine which of two cases
holds:

1. the AK47's affected tiles **never reach `import_texture`** — a decode
   that never ran, pointing at the weapon-model texture *load* path rather
   than anything palette-related; or
2. they reach it with **empty/zeroed source bytes** — a decode that ran
   against a cleared or exhausted texpool slot, pointing back at the
   two-ended `leftpos`/`rightpos` allocator in `image.c` `texLoad`.

Useful control: the AK47 is not a level-start weapon on Facility, so its
textures load later than the PP7's — compare against a level where the AK47
*is* an early/starting weapon. If it renders clean there, load ordering /
pool exhaustion is implicated directly. D217 · D161 lineage.

## D266 — M-125: re-applied the two reverted D265 port fixes (I2:A2 IA4 decode + AC-aware edge handling); new tri-time census shows **zero IA4 draws from the boot camera** — the D265 "IA4 tree strip" class does not reproduce at the wall

Re-applied, on top of the committed baseline: (a) `import_texture_ia4` I3:A1→I2:A2 decode (`intensity = part>>2`, `alpha = part&3`, new `SCALE_2_8`) — hardware-correct RDP IA4 layout, necessary per D265; (b) `gfx_opengl.cpp` edge path: the inherited hardcoded `if (texel.a > 0.19) texel.a = 1.0; else discard;` replaced with alphacompare-keyed semantics — AC_DITHER (`opt_noise`) → noise-dithered discard; AC_THRESHOLD → discard only fully-transparent; AC_NONE → discard below 0.5, survivors keep alpha for blending (the old line forced every survivor opaque, defeating ALPHA_CVG_SEL); the plain `opt_alpha && opt_noise` dither line is now gated off when edge is also set. `gfx_pc.cpp`: `texture_edge` no longer set for AC_DECAL tiles. Both are port-only; identical latent lines exist in the PD port (PD's art never exercises them — PD has **no billboards at all**, full audit this session).

New env-gated census probe (`GE_D266`, in `gfx_sp_tri1`) logging per-draw tile fmt/line-width/RM class for alpha-bearing draws. **Decisive result: over 1.5M tris from the Surface 1 boot camera, ZERO IA4 draws.** The dominant alpha classes are IA8 line=32B/64B XLU 2-cycle cards. D265's census (class (b) = "IA4 128×16 tree strip") either came from a different camera position or was misattributed — the visible wall band is not drawn by IA4 at boot (see D269 for what actually draws it).

**State: both fixes remain in-tree, UNCOMMITTED, with the probes.** They are necessary-but-unverified for the wall (the wall turned out to be a different texture class); the I2:A2 decode still owes a game-wide IA4 regression sweep before any commit. Nothing user-visible changed at the wall — consistent with D265's negative re-test.

**LANDED (M-140, R2 extraction, `fix/d266-d271-extraction` off `release/v0.3.0`):** the two real fixes only (IA4 I2:A2 decode + AC_DECAL-excluded texedge) extracted out of the parked D236 probe tree, stripped of every D267–D280 debug probe (census counters, PPM dumpers, per-tri logging — none of it behavior-affecting, but none of it belongs in a landed commit either). Full 21-level `verify.sh sweep` clean (20/21 PASS on first pass; the 21st, `dam`, came back NO-FRAMES/pixcount-0 on the sweep run but reproduced PASS on an immediate solo re-run — a load-sensitivity flake per the D117/D134 class, not a regression from this change). Build: 242/242, clean link. The D267–D280 probe code + everything else in the D236 investigation stays parked on `park/d236-probes-m139` for whenever the tree-instantiation thread resumes.

## D267 — M-126: IA8 dump probe + per-tile-slot upload-sequence tracking — the three IA8 images in play are an opaque noise card and two sparse mast sprites; none of them is the wall band

Probe infra (`GE_D267`, `gfx_pc.cpp`): PPM dumper + alpha histogram for 32px-wide IA8 strips in `import_texture_ia8`; per-tile-slot upload sequence tracking (`d267_seq` counter + `d267_slot_img[128]` stamped in the `import_texture()` dispatch) so tri-time probes can name which uploaded image a draw uses. Caveat discovered: `gfx_texture_cache_lookup()` returns early **before** the stamp on cache hits (key = fmt+addr+size+pal), so a draw's logged imgseq can be stale — the actual image is the earlier import that populated the cache (traceable via the GE_TEXI census log).

Results (level_36, 15 s): three distinct IA8 images — seq 113 (32×32, **fully opaque**, 1023/1024 texels at A=F; blocky noise pattern), seq 165 (32×64, sparse mast/pole sprite: 1344 transparent / 592 opaque / ~112 mid-alpha), seq 170 (32×64, another sparse sprite). Dumps in `scratch/d267_ia8_tile0_seq*.ppm`. None of these matches the horizon stripe band → the wall is not an IA8 card either. Combined with D266: **the entire IA-focused framing of D236/D265 was examining the wrong texture class.**

## D268 — M-127: force-alpha A/B test proves the texel-alpha blend pipeline works end-to-end — the wall is NOT a broken blend path

`GE_D268_FORCEA=0|15` in `import_texture_ia8` forces every IA8 texel alpha to 0 (transparent) or F (opaque). Three captures at frame 730 (`scratch/d268/{normal,alpha0,alphaF}_730.ppm`): the horizon band region vanishes with alpha=0 and becomes a solid bright block with alpha=15. The GL blend state, texel-alpha path, and shader generation all composite correctly — whatever renders the wall does so with *correct* blending of *something*. (Side note: the user's 9/14 BSOD 0x119 was HWiNFO64/nvml.dll, not the game — game log clean.)

## D269 — M-128: **the wall is CI8 32×32, not IA at all** — band-gated all-format probe identifies two draw classes per room stream (OPA + XLU), both BILERP/LOD/2-cycle/**FOG-on**; the cards are soft gradient blobs, and the stripe period ≈ one card per visible room

`GE_D266W` v4 (`gfx_sp_tri1`, env-gated): tracks **all** draw classes whose projected vertices fall in the top 3–9% of the viewport (the measured stripe band), logging fmt/line-width/oml/omh/zmode/cyc/blender/imgseq + first-vertex screen position. Results per room stream (imgseq cycling ~174→179→264→269→282…, shared ROM addresses 0x70173840 / 0x70173de8):

- **Class A (OPA):** CI8 32px, oml=0xc4112078, omh=0x0099ac20, first vertex screen(1.30, 0.07) — off-screen right
- **Class B (XLU):** CI8 32px, oml=0xc41049d8 (m1=CLR_IN·A_IN / m2=CLR_MEM·1MA = standard straight-alpha), omh=0x0099ac60, zmode=XLU, first vertex screen(0.61, 0.08)

**omh decode (GE gbi.h field layout: TEXTFILT@12, TEXTLUT@14, TEXTLOD@16, TEXTPERSP@19, CYCLETYPE@20):** both classes are **BILERP + TEXTLOD-on + perspective-corrected + 2-cycle**, and **bit 13 = FOG enabled**. The band is ~19 repeats across a 1957px frame (period ≈101px) ≈ one card per visible room in the row.

Texture content (`GE_TEXDUMP`/`GE_TEXRAW` + offline decode of `texdump/r*.bin`): the CI8 cards are **soft vertical gradients / sparse blobs** (index histograms concentrated at low indices, e.g. idx 6–14; ASCII renders show gradients and small clumps, not tree silhouettes). NCC correlation of every dumped card against the frame-730 capture at scales 1/2/4: nothing above 0.35 — the cards are not visible as rigid copies (consistent with bilerp magnification + fog wash).

**Reframed hypotheses, ranked:**
1. **Fog.** Both classes have FOG on. On N64 these distant cards are presumably washed toward the sky colour into a smooth pale band; if the PC's fog range (`rsp.fog_mul`/`fog_offset`; per-vertex fog computed at `gfx_pc.cpp:1591`, `fog_z = z*winv*rsp.fog_mul + rsp.fog_offset`) or its shader application differs, the raw dark gradient cards show through as hard stripes. D218 family ("fog/far-clip tint"). **Cheapest test:** extend the band probe to log per-vertex fog for class A/B tris (if vFog≈1 but screen shows dark → fog not applied in shader; if vFog small → range mismatch), or env-gate a fog disable and compare.
2. **TEXTLOD / mipmap sampling.** N64 samples mip levels when minifying; check whether the port generates GL mipmaps for these tiles (tex_lod path) and which min/mag filter is used — and first measure the *projected quad extent* (the probe only logs one vertex; if cards are minified rather than magnified, mips matter directly).
3. **Draw density / portal culling.** PC may draw cards from rooms N64 culls → denser wall. Compare total band-tri count vs expected visible-room count.
4. **Palette (D217 family).** GE_TEXRAW now also writes `.pal` (512 B = 256 RGB555) for CI formats; the aborted run never produced them. Decode class A/B cards as RGB and check colours against the N64 ref (`docs/img/bugs/d227-surface-emulator-ref.png`: smooth bright band + organic clusters centred x≈0.35–0.75).

**Owed next (concrete):** (1) re-run `rm -rf texdump && mkdir texdump && GE_TEXDUMP=1 GE_TEXRAW=1 GE_D267=1 timeout 15 ./build-pc/ge007.x86_64.exe -level_36` and confirm `.pal` files land; (2) decode `texdump/r173..r282_f2_*` CI8 cards + palettes offline to RGB PNGs for the user's eyeball (this model cannot view images — programmatic analysis only; a vision-capable session could do the N64-ref comparison directly); (3) add projected quad extent (min/max screen x over the 3 vertices) + per-vertex fog to the `GE_D266W` band probe — one more run decides hypothesis 1 vs 2.

## D271 — M-129: **the wall, third pass — suspect #3 (draw density / portal culling) confirmed as a real over-draw; the D106 guard was the cause.** Near-plane-straddling portals produce huge *finite* bounds that N64 clamps/culls in `bgRectIntersect`; the PC guard mapped them all to full screen, force-drawing whole rooms N64 never submits. Guard narrowed to non-finite-only; culling now matches N64

**Context.** D269's ranked suspects for the Surface 1 tree wall ended with "#3 draw density / portal culling — PC may draw cards from rooms N64 culls → denser wall". This pass tests exactly that, using the boot camera (`-level_36`), where the wall is most prominent and the scene is (nearly) static.

**Instrumentation (`#ifdef PORT`, `src/game/bg.c`, all env-gated, kept in tree).**
- `GE_D271` in `sub_GAME_7F0B5864` (the portal-bounds builder): per-point capture of every clip point's `z` and its `transform3Dto2DWithZScaling` projection; on each D106-guard hit, dumps the portal number, connected rooms, flags, scale, and all points (`D271 d106 …` / `D271 ptN …` lines). A second probe logs **finite** out-of-range bounds that now flow through to `bgRectIntersect` (`D271 oor-finite …`, budget 8) so their clamp outcome can be checked against N64 semantics.
- `GE_D104` print line extended with `hasPortals=… ordDegen=… d106Fallback=…` — the two counters separate the *ordinary* degenerate case (genuinely edge-on portal, min>max on a real axis — expected, N64 does the same) from the D106-fallback case.

**Findings.**
1. It is not one portal: at the Surface 1 boot camera, **11 different portals around room 13** straddle the camera near plane every frame (a clip point lands exactly at z==0; `transform3Dto2DWithZScaling`'s `inv_z = -1e20f` sends it to ±1e20–1e26-scale screen coords). The D106 guard (`lim=1e5`) fired on **every frame** (~396–781 hits per 600-tick print window) — occlusion culling was effectively defeated for that viewpoint.
2. **The N64 semantics the guard got wrong:** there is no such guard in the N64 code. The huge values are *finite*, and `bgRectIntersect` is pure min/max comparisons: a bound pair that brackets the screen clamps to the parent/screen box (portal covers the view → room kept); a pair whose finite portion falls off-screen yields an **empty** box (room culled). The original guard collapsed both cases to "full screen" — correct for the first, **over-inclusive for the second**.
3. Per-portal check of the over-included case (`D271 oor-finite`): every culled portal's finite portion is far off-screen — e.g. Surface 1 p54 (r22) min.x=997 (>319, entirely right), p68 (r8) max=(-282,-391) (entirely left/below), p73 (r14) min.x=500; BUNKER1 p28 (r25) y-range 339.85..2.9e24 (entirely below the 10..230 screen). With identical f32 values N64 produces the identical empty clamp → **N64 culls these rooms from these viewpoints; the old guard was drawing them.**

**Fix (one constant, `src/game/bg.c` `sub_GAME_7F0B5864`, `#ifdef PORT`).** D106 guard `lim`: `100000.0f` → `1e38f`. Rationale: this projection can reach ~1e26 at most, so 1e38 catches only non-finite values (NaN/inf), which are genuinely unsafe to let through (NaN poisons the downstream comparisons). Finite out-of-range bounds now flow into `bgRectIntersect` exactly as on N64. The finite min>max case is untouched — it still hits the original `degenerate` check (the N64 code path itself), so D106's actual regression (BUNKER1 sky-fill voids from *discarded* portals) stays fixed.

**Verification.**
- `d106Fallback`: ~396–781/600 ticks → **0** on both Surface 1 (`-level_36`) and BUNKER1 (`-level_09`); `ordDegen` unchanged (ordinary edge-on portals still map to full screen, as on N64).
- roomsDrawn at the Surface 1 boot view: **17–18 → 10–13**; per-room boxes now narrowed correctly (e.g. r8 = right sliver `246,10-319,230`, r12 full-screen via a genuine bracketing clamp) instead of ~18 full-screen boxes.
- BUNKER1: roomsDrawn 4→3 (r6 state, lost r5) and 3→1–2 (r29 state, lost r25/r28). Traced per portal above — the losses are the N64-matching empty clamps, i.e. moves *toward* N64 fidelity, not regressions; the D106 void case (discarded min>max portals) is untouched.
- Long unattended runs (`timeout 90+`): no crashes on either level; intro camera progresses normally to later states (curRoom=18, roomsDrawn=4).
- **Pixel-level A/B — inconclusive, owed to the user.** Frame captures at matched `curRoom` state (frame 1500, room-13 state: `scratch/d271_1500_sbs.png`, baseline left / fixed right) show a 14.7% changed-pixel difference — but a same-build/same-state pair differs by 14.8% with an *identical* spatial grid (top-heavy band = animation/camera micro-motion at this viewpoint, not the fix). The boot view is not static enough for programmatic A/B; the user's eyeball on the tree wall is the deciding check.

**OUTCOME (user re-test 2026-09-14): NEGATIVE.** With the fix in place the trees still render as one big wall — and the user's read is that **the texture itself isn't being mapped correctly**, not merely an over-draw/alpha problem. So suspect #3 was a real N64-divergent over-draw (kept — it is a fidelity fix, do not revert) but it is NOT the wall's root cause. The D265 note gains weight: *"all captured cards sampled the same u:0..64/v:0..16 window — confirm against N64"* — if every card samples the same region of a shared texture, the result is exactly a repeating sheet. Top suspect for the next pass: **per-card S/T coordinates / UV windowing** (model DL vertex S/T, tile setup foldS/foldT/tcShift, or `d43_emit.py`'s S/T conversion), then TEXTLOD/mip sampling, then palette.

**LANDED (M-140, R2 extraction, `fix/d266-d271-extraction` off `release/v0.3.0`):** just the one-constant fix (`lim = 100000.0f` → `1e38f`), stripped of the `GE_D271`/`GE_D104` debug probes (per-portal capture, hit counters) — none of it changes behaviour, none belongs in a landed commit. Checked against **D249** (distant-geometry drop-out on Streets/Egyptian, per M-139's R2 overlap concern) before landing: no overlap — D249 is a far-plane/LOD symptom on a different code path (open, unrelated), this fix is a near-plane portal-straddling guard; D249 stays open. Full 21-level `verify.sh sweep` clean post-extraction (see D266's landing note for the one dam/NO-FRAMES flake, confirmed non-regression on retry).

**Owed / next.**
1. **Per-card S/T audit** — dump the actual vertex S/T + tile for each wall card at draw time (extend `GE_D266W`) and check whether distinct cards carry distinct S/T windows; compare against the N64 ref (`docs/img/bugs/d227-surface-emulator-ref.png`: smooth bright band + sparse organic clusters, not a continuous sheet).
2. Regression sweep of other levels' room counts is cheap insurance (`GE_D104=1` spot-checks); BUNKER1 already re-verified above.
3. Probes are budget-limited and env-gated; keep until the user check lands, then remove with the D266–D269 probe set if the wall closes.

## D272 — M-130: **the wall, fourth pass — per-card S/T audit (D271's owed next-step #1).** Distinct cards carry distinct (not identical) UV windows; the XLU class wraps its CI8 tile ~4.5× horizontally; traced to genuinely-authored ROM vertex data, not a `d43_emit.py` conversion bug — hypothesis shifts back to TEXTLOD/mip handling of a heavily-wrapped, highly-minified tile

**Context.** D271's fix (portal-culling guard) was verified but the user's re-test was negative: "trees are a big wall rather than the texture being mapped correctly." HANDOFF's next-steps for this thread ranked **per-card S/T / UV windowing** as the top suspect, driven by a D265 note that "all captured cards sampled the same u:0..64/v:0..16 window." This pass builds the audit HANDOFF asked for and answers the decisive question it posed: do distinct card instances carry distinct S/T windows, or identical ones?

**Instrumentation (`#ifdef PORT`, `port/fast3d/gfx_pc.cpp`, `gfx_sp_tri1`, env-gated `GE_D272`, budget 200 lines, kept in tree).** A new raw per-triangle dump, filtered to D269's exact two wall-card classes by low-24-bit `other_mode_l` (`0x112078` OPA / `0x1049d8` XLU — matching regardless of the top-byte pipeline bits, which vary between otherwise-identical-looking draws sharing the same blend equation). For each matching tri inside the measured wall band, logs: `oml`/`omh`, the bound texture's `seq` (image identity), tile `fmt`/`siz`/`line`/`width`×`height`/`mask`/`shift`, the tile's own UV window (U10.2 → texels, `/4`), and all three vertices' UV (vertex ST is s10.5 → texels, `/32`).

**Findings.**
1. **Both classes confirmed CI8** (`fmt=2 siz=1`), matching D269 exactly — the earlier unfiltered run of this same probe (before the oml filter was narrowed) picked up an unrelated `fmt=3` (IA) class sharing the same low-24-bit blend bits under a different top byte; that class is NOT one of D269's wall cards and is excluded by construction once the filter checks the full pair of known `oml` values.
2. **UV windows are NOT identical across instances of the same class/texture.** OPA class (`seq=4`, 8 samples): first-vertex U ranges 5.25→5.56 (spread ≈0.3 texel), consistent quad shape (u≈0..32, v≈0..11) each time but never byte-identical. XLU class (`seq=168/169`, multiple samples): U range spans ~143.16–144.81 (spread ≈1.6 texel) across different physical card instances. This is genuine per-vertex variation, not a converter collapsing every card onto one hardcoded window — **D265's "identical window" observation did not come from these cards**; it most likely came from a different draw class caught by the old unfiltered/undifferentiated band probe (D266W has no oml filter and this screen band is crossed by several unrelated classes — see the `fmt=3`/`oml=0x0f0a4000`/`0x00504340` families logged in this session's first (unfiltered) D272 run, several of which *do* show identical `(0,0)` UV across every instance — degenerate/filler geometry, not wall cards).
3. **The XLU class's U range is ~143–145 texels wide — ~4.5× its own tile's 32-texel wrap period** (`mask=5` ⇒ `1<<5=32`). The OPA class stays within roughly one tile width (u≈0..32). This is a large, consistent multi-wrap pattern, not noise.
4. **Traced to `tools_pc/d43_emit.py`'s vertex emit path — no converter bug found.** `emit_main_vtx` (the function that writes every model's `Vtx` array) copies `tc[0]`/`tc[1]` (the S/T fields, offsets 8/10 of the 16-byte N64 `Vtx`) via `put_s16(q+8, be16(src, o+8))` / `put_s16(q+10, be16(src, o+0xA))` — a plain big-endian→little-endian 16-bit byteswap, **no scale, no offset, no reinterpretation**. Every other field in the same struct (`ob[3]`, `flag`, `cn[4]`) goes through the identical plain-copy pattern. There is no code path here that could inflate, rescale, or corrupt a small authored S/T into a ~4.5×-wrap value — the ~144-texel figure must already exist in the ROM's ilbytes at that vertex. **This rules out the converter as the source of the wide wrap** (cf. the D108–D112 `put_f32` byte-reversal precedent this thread was checking against — that bug class is not present here).

**Reframed hypothesis (supersedes D271's "top suspect").** Since the wide UV wrap is real, ROM-authored data (Rare's own design: tile one CI8 texture ~4.5× across a long background quad), the wall is not a windowing/converter bug. The candidate mechanism shifts to *how* the port renders a heavily-wrapped, heavily-minified (TEXTLOD-on) tile: GL's `GL_REPEAT` + whatever mip level(s) OpenGL selects/generates for this tile vs. the RDP's own per-scanline TMEM address wrap + LOD blending. D269's hypothesis #2 (TEXTLOD/mip sampling) is now the leading suspect again, ahead of fog range/application and palette.

**Owed / next.**
1. Measure the actual on-screen minification factor for these cards (projected quad extent vs. tile texel count) — if the quad is only a handful of screen pixels wide while sampling 144+ texels, GL's automatic mip selection is likely picking a coarse level, and any difference from the RDP's own LOD blend range would show up exactly as "hard repeating pattern" vs. "soft distant band."
2. Check whether `glGenerateMipmap` (`gfx_opengl.cpp:748`) is actually invoked for these specific CI8+`GL_REPEAT` tiles, and whether the generated mip chain wraps correctly at the tile's `1<<mask` period rather than at the raw 32×32 upload (should be identical for a power-of-two tile, but confirm — this is the one open question RC3/D167 didn't fully close).
3. Cross-check against the N64 reference (`docs/img/bugs/d227-surface-emulator-ref.png` / `mupen64/`) to confirm whether the intended look really is "soft blended repeats" (mip-driven) rather than a different UV entirely — do not assume without a direct look.
4. Probe is budget-limited (200 lines) and env-gated (`GE_D272`); keep until the wall closes, then remove with the D266–D272 probe set.

## D273 — M-131: **N64 ground-truth screenshot obtained; palette-decode A/B (pre-existing, undocumented `GE_D273` probe found in the uncommitted tree) tested against it and RULED OUT** — the standard ARGB1555 palette layout does not change the wall's appearance at all, but corrupts unrelated CI8 textures (path, viewmodel skin tone)

**Context.** The user supplied two reference screenshots (`docs/img/screenshots/gepd-surface.PNG` — 1964 N64 emulator, GEPD edition ROM, same Surface 1 corridor; `docs/img/screenshots/port-surface.PNG` — our PC port, same location) for a direct side-by-side the thread had been missing (D271/D272 both flagged "no direct N64 look yet" as an open risk). **N64 reference: discrete conifer silhouettes (trunk+canopy visible) lining both sides of a snow corridor, blending softly into a dusk sky — sparse, not a sheet.** **Our port: a continuous, uniform, fine-grained diagonal herringbone/houndstooth weave covering the same wall area, flat-topped, no tree silhouettes at all** (`scratch/d273_base_1500.png`, `-level_36` frame 1500, matches the user's `port-surface.PNG` framing).

**Also found this session (not written up by whoever added it): an existing, uncommitted `GE_D273` A/B probe already in `port/fast3d/gfx_pc.cpp`'s `palette_to_rgba32`** (env-gated, default off, part of the pre-session diff — not authored this session, and not mentioned in HANDOFF.md or findings.md before now). Its comment hypothesizes the inherited PD `RGBA16`/CI-palette decode (`a=bit0, r=15:11, g=10:6, b=5:1`) is one bit off the N64 hardware `ARGB1555` TLUT layout (`a=15, r=14:10, g=9:5, b=4:0`) that GE's RSP ucode actually uses for CI palette expansion, and that for one sampled wall-card palette entry (`0x0011`) this flips the AC_NONE alpha-test outcome from discard (sparse, N64-correct) to keep (opaque sheet, the wall bug) — a plausible, concrete mechanism, matching the HANDOFF's long-parked "side suspect: palette" line, but never actually tested against a live frame.

**Test performed this session.** Captured matched-tick frames (`GE_PCDUMP=1500-1500`, `-level_36`) with and without `GE_D273=1`.
- **Wall region: visually unchanged.** Same herringbone pattern, same density, same silhouette (none). A pixel diff of the wall crop shows a non-trivial mean delta (~28/255), but this is consistent with D271's already-documented run-to-run boot-camera/room-streaming jitter (~14–15% baseline noise between two *identical*-build runs at this same viewpoint) — not a content change; eyeballing both frames side by side shows the identical weave.
- **Everything else got worse.** The snow path (also CI8-textured) turned into a garish cyan/magenta moiré, and the first-person hand/viewmodel skin tone rendered teal-green instead of flesh tone — both regressions exactly where you'd expect if the "standard" ARGB1555 reading is in fact the wrong one for these particular palette sources, or the reinterpretation is simply not selective enough (the probe applies unconditionally to every CI4/CI8 palette lookup, `ignoring palette_fmt` per its own comment, which is a blunt instrument for testing one specific texture's TLUT).

**Conclusion.** The palette-bit-layout swap is **not** the wall's root cause: it does not visibly touch the wall at all, while breaking unrelated known-good CI8 content. `GE_D273` should stay default-off; do not enable it or pursue this specific decode-layout change further for the wall. The underlying question the probe's comment raised (is GE's RSP CI-palette expansion really "standard" ARGB1555 regardless of `palette_fmt`, per `gmain.s`) may still be worth resolving someday for its own sake, but it is decoupled from D236 now.

**Process note.** This probe existed in the working tree with a full, specific, plausible root-cause theory and was never mentioned in `docs/HANDOFF.md`'s CURRENT TASK or `docs/dev/findings.md` — a live investigative lead sat undocumented across at least one handoff boundary. Whenever picking up an in-flight investigation, `git diff` the suspect files (not just what the day's handoff narrates) before assuming the handoff is a complete inventory of uncommitted work.

**Owed / next.**
1. Get an actual reference capture from `mupen64/` (this session used a third-party 1964-based build the user had on hand, GEPD edition — a second, known-clean N64 emulator source would strengthen the ground truth) for a pixel-level comparison, not just an eyeball read.
2. D272's TEXTLOD/mip hypothesis is still the live lead — this session only closed off the palette branch. Proceed with D272's owed next-steps (minification measurement, confirm `glGenerateMipmap` actually fires for this tile).
3. Given the N64 reference shows genuinely discrete tree shapes (not merely "softer" versions of the same repeating pattern), also keep open the possibility that the port is sampling entirely wrong texture content for this tile (not just filtering it differently) — the herringbone weave does not resemble D269's offline-decoded "soft gradient / sparse blob" content at all, which is itself a discrepancy worth reconciling before committing further to the mip-filtering theory.

## D274 — M-132: **seq/addr identity is unreliable across runs (both are runtime-arena, not stable ROM, identifiers) — fixed with an unconditional per-slot address stamp; the "content vs. decode" discrepancy D273 flagged is RESOLVED as an offline-analysis bug in this session, not a port bug — mips confirmed generated + enabled for the real wall-card tile, narrowing the live suspect to mip-blend *quality*, not presence**

**Context.** An external review of the D273 write-up (this session) flagged that all three passes analyzing "the wall cards" (D269, D272, and a `gfx_pc.cpp` code comment) reported different `seq` ranges, and correctly identified the cause: `d267_slot_img[tile]` (D267) is stamped in `import_texture()` *after* the texture-cache-hit early return (`gfx_texture_cache_lookup`, line 1131) — on a cache hit the stamp is never refreshed, so it can read back the identity of whatever last triggered a real upload into that TMEM slot, not the texture actually bound for the current draw. The review recommended anchoring identity on the ROM source address instead.

**Fix (`port/fast3d/gfx_pc.cpp`, `#ifdef PORT`).** Added `d267_slot_addr[128]`, a companion array stamped **unconditionally**, immediately after `orig_addr` is known and *before* the cache-lookup early return — so it always reflects the real source address of whatever is currently bound to that tile slot, cache hit or miss. Threaded into every D272-family log line (`D272[...]`, `D272S`, `D272U`, `D272P`) as `addr=%p`.

**This fix immediately surfaced a second, independent identity hazard: the runtime address itself is not a stable cross-run identifier either.** Two hard confirmations from live runs against the same build/level:
1. **Same `seq`, different `addr`, same run.** Isolating the XLU wall class alone (`GE_D270KEEPOML=1049d8 GE_D272=1`), `seq=65` was logged against **two different addresses** (`0x701587c8` ×161, `0x70155738` ×17) in one 15 s capture — direct proof the seq stamp is stale mid-run, not just across runs.
2. **Same `addr`, different declared format, different run.** `0x70155738` (one of the two addresses `seq=65` produced) showed up in a *separate* run (with `GE_TEXDUMP=1`) as `GE_TEXI[4] fmt=3 siz=1 tile=54x54` (IA8) — a completely different format/size than the CI8 32×32 the tri probe recorded for it moments earlier in a different run. GE's texture/model pipeline decodes into a shared runtime arena (the M-88 lesson already in `docs/porting-notes.md`: "GE's model and image pipelines resolve almost everything through small numeric IDs into shared, runtime-decoded pools"); load-order timing (already documented as non-deterministic across runs by D271) evidently reuses the same arena address for different content run-to-run. **Neither `seq` nor the raw runtime `addr` is a stable identifier across separate process runs — only within a single run, and even then only outside a cache-hit window for `seq`.** D269's recorded "shared ROM addresses 0x70173840/0x70173de8" should be read as *that run's* addresses, not persistent identifiers to search for in a fresh capture.

**Practical consequence for this pass:** identify and decode the real wall-card texture **within one run**, correlating `D272U`'s `addr=` (confirmed CI8, confirmed actively drawn as a wall card in that same run) against that same run's `GE_TEXI[N]`/`texdump/rNNN_*.bin(.pal)` dump by address match, not by `seq`. Done: `GE_TEXDUMP=1 GE_TEXRAW=1 GE_D270KEEPOML=1049d8 GE_D272=1` in one combined pass (per the review's efficiency suggestion) → `GE_TEXI[11] addr=0x701587c8 fmt=2 siz=1 tile=32x32`, matching `D272U`'s live ci8-upload log for the same address in the same run → `texdump/r011_f2_s1_32x32.bin(.pal)` is unambiguously the real, currently-drawn wall-card texture.

**Decoding it exposed a bug in this session's own offline analysis, not a port bug.** First decode attempt (`struct.unpack('>256H', ...)` — big-endian) produced multi-hued rainbow noise under *both* candidate palette layouts (port's shifted decode and the standard ARGB1555 from D273) — which looked like a genuine new "wrong content" finding until traced back: `rdp.palette` is an in-process `uint16_t[256]` array, and `GE_TEXRAW`'s `.pal` dump (`fwrite(rdp.palette, 1, 512, pf)`) writes it as **raw native-endian bytes** — little-endian on x86-64, not the big-endian the ROM's own on-disk TLUT would be. Re-decoding with `'<256H'` (little-endian) immediately produced a coherent image under the port's existing default (shifted) decode: a smooth gray/brown diagonal gradient wedge, closely matching D269's original "soft vertical gradients" description. The **same ARGB1555 alternate that D273 already ruled out for the live scene** also decodes this specific real texture into a broken, speckled result (mostly-transparent with scattered noise along the edges) — a second, independent confirmation that D273's conclusion was correct and this session's earlier "content vs. decode" open question (from the D273/D236-sixth-pass HANDOFF entry) is now **resolved: there is no decode discrepancy.** The port's existing palette decode is correct for this texture; D269's "soft gradient" description was accurate.

**Mip/LOD state confirmed nominal for the real tile** (closes D272's owed next-step #2 for this specific texture): `D272U[4] ci8 upload 32x32 genmip=1` (mip generation *is* requested at upload) and `D272S ... addr=0x701587c8 ... cms=0 cmt=0 tex_lod=1` (LOD/mip sampling *is* enabled at draw time; `cms=cmt=0` = `G_TX_NOMIRROR|G_TX_WRAP` = `GL_REPEAT`, as expected for the ~4.5× wrap D272 measured). **Rules out "mips are simply missing/disabled" as the mechanism.** The remaining candidate narrows to *quality*, not *presence*: whether GL's automatic trilinear mip chain, sampling a small (32×32) diagonal-gradient tile repeated ~4.5× within a small screen-space quad, actually reproduces the RDP's explicit `TEXTLOD`/lod-frac blend closely enough — a smooth gradient wedge tiled many times, if under-blended at the wrap seams, is a plausible mechanism for a regular diagonal herringbone/houndstooth artifact (the wedge's own diagonal edge, repeated, is exactly that shape).

**Attempted, inconclusive: the two-class moiré A/B** the review also asked for (render with `GE_D270KEEPOML` isolating first the OPA class, then the XLU class, to see whether the weave is a single-class tiling artifact or an interference beat between the two overlapping classes). `GE_D270KEEPOML=112078` at the boot-camera frame 1500 produced a **fully black frame** — not just "wall cards missing," *nothing* drew, including geometry that should be unrelated to the wall (this needs its own triage: either the low-24 match genuinely excludes everything in view at this exact tick, or the keep-list interacts with something upstream of the wall class in a way that blanks the frame). Not investigated further this pass; flagged as owed, not blocking — the palette/mip findings above stand independent of it.

**Owed / next.**
1. **Debug the `GE_D270KEEPOML` all-black result** before relying on it — log which `oml` values actually cross the screen at the target tick (the existing unfiltered `GE_D266W` can do this) and confirm `112078`/`1049d8` are really present, then redo the class-isolation A/B.
2. **Mip-blend quality, not presence, is now the question.** Compare a forced-coarse-mip render (LOD bias, or temporarily forcing `GL_NEAREST_MIPMAP_NEAREST`) against the current trilinear result at the wall — if a coarser/blurrier mip visibly smooths the weave into something closer to the N64 reference's soft band, that confirms the mechanism and points at LOD selection/bias tuning as the fix; if it doesn't change materially, mips are a dead end and the remaining candidates are GL_REPEAT-wrap seam continuity across mip levels, or something in how the two overlapping (OPA+XLU) classes composite.
3. Regenerate `docs/dev/findings-index.csv` (`tools_pc/gen_findings_index.py`) — done this pass, but re-run again after any further findings.md edits.
4. **Process note, separate from the palette-decode-byte-order bug above**: `D270` (the skip/keep-oml hooks, the `D270P`/`D270KEEPOML`/`D270SKIPOML` machinery, and the `tdc` cap raise 400→600) has never had its own findings.md row — only code comments and HANDOFF-archive mentions. Same undocumented-lead class D273 flagged. A one-line backfill row is owed whenever someone next touches that thread; not done here to stay focused on D236. **Now backfilled as its own index row, see below.**

## D275 — M-133: probe upgrades for the eighth pass — `GE_D270PIX` extended from one point to a list, `GE_D270KEEPOML` gained a "kept N tris" counter (an external review's step 1/2 groundwork)

**Context.** A second external review of the D274 write-up proposed a four-step eighth pass: (1) multi-point pixel attribution + fog state inside the weave to name the actual front-most painter and close D269's never-done "log per-vertex fog" step; (2) an aligned class-isolation A/B with a visible "kept" counter so a black capture can be told apart from a genuine render bug; (3) an LOD-quality A/B; (4) stay alert to the framing risk that the confirmed wall-card content (a gradient wedge, D274) cannot by itself explain the N64 reference's discrete conifer silhouettes — a second, unattributed draw class is likely responsible for those, and D236 may split into two separate defects. This entry covers the mechanical probe work for steps 1–2; D276 covers what step 1 actually found.

**Changes (`port/fast3d/gfx_pc.cpp`, `#ifdef PORT`, both still env-gated/opt-in).**
- `GE_D270PIX="x0,y0[;x1,y1;...]"` (was single-point `"x,y"`): up to 8 normalized screen points, semicolon-separated. Each `D270P[ptN][...]` line now also carries `pix=(x,y)` (which point) and `addr=%p` (D274's reliable per-slot address, not `seq`).
- `GE_D270KEEPOML`: was silent. Now logs `D270KEEPOML: kept tri #N (oml24=0x...)` for the first 4 kept tris and every 5000th thereafter — a black capture with a nonzero kept-count is a genuine render-interaction bug; a black capture with `kept==0` at that tick is tick misalignment or an empty keep-set, not a probe failure.

**Offline groundwork before the live run:** picked 4 candidate weave pixels directly from `scratch/d273_base_1500.png` (the confirmed reproduction of the user's reported wall) by eyeball + sampled RGB to confirm they land in the gray weave band, not the snow floor or the black letterbox bars: `(0.05,0.10)`, `(0.25,0.15)`, `(0.85,0.12)`, `(0.50,0.06)` (the last one turned out to be sky, not wall — a miss, harmless, see D276).

**Attempted, still open:** the class-isolation A/B itself (review's step 2) was not completed this pass — D276's step-1 finding reframed the whole "which class is the wall" question first, which the review itself said should gate everything else ("step 1 first because it's the only step that can invalidate the current framing"). Re-run step 2 with the *correct* class identified in D276, not the D269-named OPA/XLU pair.

## D276 — M-134: **the true painter of the visible weave, at the confirmed weave pixels, is a fully-opaque IA8 blocky-noise/static texture (`oml=0xc81049d8` low bits, `fmt=3 siz=1`) sampled with a fixed, identical UV window and comparatively low fog — NOT either of D269's named CI8 classes, which don't even appear at two of the four sampled points.** D265's very first "all cards sample the same u:0..64/v:0..16 window" observation matches this class exactly; D269 reframed away from it prematurely.

**Method.** Ran the D275-extended `GE_D270PIX` at the 4 confirmed weave pixels with tick-aligned captures (`GE_PCDUMP=1300-1700:25`, 17 frames, all visually confirmed non-black/weave-consistent — ruling out tick misalignment as an explanation this time), logging every triangle covering each point across the whole 35 s run, then grouped by `oml` and summarized `z/w` and fog ranges per point offline.

**Findings.**
1. **At the two points confirmed inside the weave band** (`(0.05,0.10)`, `(0.25,0.15)`), **neither D269-named class (`0xc4112078` OPA, `0xc41049d8` XLU) appears at all.** The dominant, front-most (`z/w≈0.98–0.998`, excluding the sky quad's construction-mandated `z/w=0`), consistently-drawn class is `oml=0xc81049d8` (note: **low bits `0x1049d8` match the XLU class's low-24 bits, but the full oml/top-byte and — critically — `fmt=3` (IA) differ**; this is the same false-friend class D272/M-130 already had to filter out of its budget) with **average fog 39.5–63.3 out of 255** — mostly-unfogged, so its raw texel color reaches the screen largely unmodified.
2. **Decoded that exact texture within the same run** (`GE_TEXI[106] addr=0x70118d68 fmt=3 siz=1 tile=32x32` → `texdump/r106_f3_s1_32x32.bin`, IA8: high nibble intensity, low nibble alpha per `import_texture_ia8`): **a fully-opaque (alpha nibble mostly 0xf), high-frequency blocky gray static/noise pattern** — visually nothing like D274's gradient wedge, and nothing like a tree. This is almost certainly the same texture D267 already dumped and dismissed ("seq 113, 32×32, fully opaque, 1023/1024 texels at A=F, blocky noise pattern... none of these matches the horizon stripe band") — D267's dismissal was a single-instance visual judgement; it did not test what this content looks like *tiled*.
3. **Every single instance of this class, across every screen position sampled in an earlier unfiltered capture (`scratch/d274_run.log`), uses the byte-identical UV window** `u:0..64, v:0..16` (a 2× horizontal wrap of the 32-wide tile, sampling only its top half vertically) — **this is D265's original "all captured cards sampled the same u:0..64/v:0..16 window" observation, verified to the exact texel, and it belongs to this IA8 noise class, not the CI8 classes D269 and D272 spent three passes on.** D269's second-pass reframing away from D265's own lead was premature: it correctly proved the CI8 classes it found don't share one window, but never established those CI8 classes were the ones D265 was describing in the first place.
4. **A third class, `oml=0x0c184b50` (CI8, `fmt=2 siz=1`) — the "C-cards" the reviewing pass explicitly flagged as unattributed and a candidate for the N64 reference's actual discrete tree silhouettes — is present at every sampled point, with `z/w` ranging wildly from `-0.0381` to `0.9940`** (most other classes stay in a tight 0.96–1.00 band). Where it does land in the normal distant-geometry range, its fog is **high** (up to 255, i.e. fully fog-washed — plausible tree-fading-into-haze behavior) — but the erratic near-camera/behind-camera excursions suggest it is not reliably reaching the screen where it should, consistent with the review's flagged risk: a second, separate class may be responsible for the reference's actual tree shapes, and it may be missing or occluded rather than merely mis-filtered.

**Reframed conclusion.** The visible "wall/weave" the user reported is most directly explained by: a tiled, high-frequency, mostly-opaque IA8 noise texture (2× UV wrap, fixed window, low fog) reaching the screen essentially raw. D274's gradient-wedge CI8 texture was real and correctly analyzed, but is very likely a secondary/different layer, not the one producing the dominant visual artifact. The `0x0c184b50` CI8 class remains an open, unattributed thread that may be the reference's actual "trees" and may itself have a missing/occlusion bug — do not assume D236 has one root cause.

**Not yet done:** confirm this IA8 class's own UV-wrap-vs-fog behavior against the N64 reference directly (does N64 also wrap this noise 2× with low fog, or should the fog be much higher here, matching the `0c184b50` class's 193–255 range, which would wash it to near-invisible as intended); the class-isolation A/B (D275's still-owed step 2) should now isolate `0xc81049d8`/`fmt=3` vs. `0x0c184b50`, not the D269 CI8 pair. The mip-blend-quality experiment (D274's owed next-step, this pass's proposed step 3) is now lower priority — the newly-identified content (blocky noise, not a smooth gradient) makes a filtering-quality explanation less necessary than it seemed against D274's wedge; a wrong/insufficient fog blend is now at least as plausible a single-mechanism explanation and is cheaper to test (it's already logged, no new probe needed).

**Owed / next.**
1. **Fog-correctness check for the `0xc81049d8`/IA8 class**, now that D269's owed "log per-vertex fog" step is finally done: is `39.5–63.3` average fog the *correct* N64 value for this vertex position (compute what `fog_mul`/`fog_offset`/world-z *should* give and compare), or is it a computation bug specific to this draw path? If N64's own fog math would also produce a low value here, the wall is arguably N64-faithful in isolation and the real bug is elsewhere (e.g. the `0c184b50` occluder not painting over it); if the port's fog is wrong, that is a direct, cheap fix.
2. **Attribute and fix-scope the `0x0c184b50` CI8 class's erratic `z/w`** (`-0.0381` to `0.9940` at a single screen point) — likely candidate for a billboard/orientation bug (camera-facing quad computed with a bad or momentarily-invalid basis), separate from the noise-texture fog question. If this class is meant to draw the discrete tree shapes, understand why it doesn't reliably reach these pixels before concluding it's simply "the wrong layer."
3. Re-run the class-isolation A/B (D275) on the newly-identified classes, with the D275 kept-counter, at the D275-established tick-aligned window (`GE_PCDUMP=1300-1700:25`) so a black frame is diagnosable.
4. Regenerate `docs/dev/findings-index.csv` after this edit (done).

## D280 — M-135: **D236 wall, ninth pass — fog math VERIFIED N64-faithful (owed step #1 closed), the noise class VISUALLY confirmed as the weave's painter, and the "trees" question reframed: `0x0c184b50` is in FRONT of the noise quad (depth-race read inverted), its fog is CPU-baked bimodal 25/255, and it barely draws at all under deterministic `-level_36` boot — no quick win for trees; the remaining work is CPU-side instantiation/room-streaming, not a port fix**

1. **Fog math verified (D276 owed step #1, closed).** Computed the expected per-vertex fog from the game's own constants: RSP fog ramp 0..154 → `fmul=32000`, `foff=-31744`; the noise class's z/w ∈ [0.992, 1.0] maps to expected fog ≈ 26–28. Compared against the port's actual per-vertex fog for all **2505 vertices** of `0xc81049d8`: mean abs deviation ±2 (integer truncation only). D276's "low fog" is therefore the *correct* N64 value — on hardware this quad would also reach the screen largely unfogged. **Fog is not a bug; no port-side fix exists here.**

2. **A/B visual confirmation of the painter (D276 owed step #3, closed).** `GE_D278C=1` (collapse `0xc81049d8` to a point) + `GE_PCDUMP=1500-1500:1`: the weave band's high-frequency energy drops **14.8 → 2.9**, mean 108 → 121; the band becomes a smooth bright haze (the fully-fogged CI8 cards behind it). The pixel-attribution-only evidence from D276 is now corroborated visually — `0xc81049d8`/IA8 is definitively what paints the visible weave. **No trees appeared with it removed** — the first clean test of that question, unconfounded by the noise quad (all prior A/Bs, D268/D273, were behind it).

3. **Full class census at the confirmed weave pixels** (`(0.05,0.10)`, `(0.25,0.15)`): only three real classes reach either point — `0xc81049d8` fmt=3 IA8 (the noise painter), `0x0c184b50` fmt=2 CI8 (720 tris; fog bimodal 25/255, **gfog=0** → CPU-baked per-vertex fog), and `0xc8102078` fmt=2 CI8 (69 tris). Two further "classes" in the logs are probe artifacts: `0x00552048` (z/w=0, dl=1 placeholder) and `0x00504340` (z/w=-1, behind camera).

4. **Depth-race correction — D276's "occluder" framing was inverted.** N64 z/w: *smaller = closer*. At pt0, **278/356** `0x0c184b50` tris are CLOSER than the nearest noise tri (z/w < 0.9908); at pt1, 310/356. The tree cards are IN FRONT of the noise quad, not behind it. The open question is therefore "why don't the closer textured cards show?", not "does the noise win the depth race?".

5. **`0xc8102078` ruled out as a tree layer.** `-level_36` + `GE_D270KEEPOML=c8102078`: **1.4M triangles** fill the lower half of the screen (y ≥ 0.35) — terrain/ground, not conifers.

6. **The tree class barely draws under deterministic boot.** `-level_36` + `GE_D270KEEPOML=0c184b50` + `GE_TEXRAW`: ~3–4 tris/frame at tick 1500, zero visible pixels, empty `texdump/` (no CI8 upload). Meanwhile the non-`-level_36` fog run logged 720 covering tris. **`0x0c184b50`'s instantiation is strongly run-to-run dependent and `-level_36` does not pin it** — the room-streaming state that feeds this class varies even with the boot flag.

**Conclusion: no quick win for trees.** The visible weave is the IA8 noise class (N64-faithful fog, confirmed painter — arguably N64-faithful *in isolation*). The discrete conifers of the reference come from `0x0c184b50`, which (a) barely instantiates in deterministic runs and (b), when it does instantiate, sits in front of the noise yet still doesn't visibly win. Making trees appear requires understanding this class's CPU-side placement/instantiation (room streaming / billboard emission) — multi-step game-logic investigation, not a port-layer fix.

**Owed / next.**
1. Single instrumented run: `-level_36` + `GE_D270PIX` (both weave points) + `GE_TEXRAW` + `GE_PCDUMP=1400-1600:1`; correlate per tick — does a fog=25 `0x0c184b50` tri cover the point AND is the pixel dark?
2. Log `curRoom`/`roomsDrawn` (D271 infra) alongside `GE_D270PIX` to correlate `0x0c184b50` presence with room-streaming state.
3. Decode `0x0c184b50`'s CI8 texture offline — needs a run in which it actually uploads (TEXRAW has not captured it yet).

## D284 — M-140: `port/src/random.c`'s MIPS64 `dsll32`/`dsrl32` shift helpers were backwards — modelled as 32-bit shifts instead of "shift by n+32 on the full 64-bit register" — desyncing the PRNG stream from N64 from the very first draw, in every release to date

**Origin: M-138's security/decomp-fidelity review** flagged `random.c` as documented bit-exact to the N64 but not actually verified as such; a forked review agent's numeric trace and a hand trace of `src/random.s` in that same session both found the same divergence. Held open pending the user's fix-vs-disclose call; **M-139b: fix it.**

**Root cause.** `dsll32 rd,rs,n` / `dsrl32 rd,rs,n` are real MIPS64 instructions that shift the **full 64-bit register** by `n+32` — they exist specifically to encode a 32–63 shift amount in a 5-bit immediate field (`dsll`/`dsrl` alone only cover 0–31). The original port modelled them instead as ordinary 32-bit shift-by-`n` operations, masked to 32 bits after: `(x << n) & 0xFFFFFFFF` / `(x >> n) & 0xFFFFFFFF`. Both the shift amount (missing the implicit `+32`) and the masking (real hardware does not truncate — the destination register keeps whatever the 64-bit shift produces) were wrong. This is a "bit-exact-looking but wrong" bug: it compiles, runs, and produces a deterministic, self-consistent stream — it is just not *the N64's* stream, from the very first call.

**Verified two ways.**
1. **Hand trace of `src/random.s`** (`randomGetNext`, `randomGetNextFrom`, and the identical transform in `src/game/chrObjRandom.s` — three call sites, same bug, same fix) against the corrected semantics.
2. **A standalone C reproduction** (compiled and run outside the game, this session) comparing the buggy helpers against the corrected ones from the same seed (`0xAB8D9F7781280783`, `random.s`'s initial value): first five draws diverge completely (buggy: `0x783006, 0x83006830, 0x6830068, ...`; corrected: `0x40ec37cf, 0x630aedd7, 0x1f58071e, ...`) — confirms the bug is real and the fix changes the actual output stream, not just the code path.

**Fix (`port/src/random.c`, mechanical, `#ifdef PORT` not needed — this file only exists on PC, the N64 build uses `src/random.s`/`src/game/chrObjRandom.s` directly and is untouched):**
```c
static inline u64 dsll32(u64 x, int n) { return x << (n + 32); }
static inline u64 dsrl32(u64 x, int n) { return x >> (n + 32); }
```
(was `(x << n) & 0xFFFFFFFFULL` / `(x >> n) & 0xFFFFFFFFULL`). No other line in the file changes — every call site (`randomGetNext`, `randomSetSeed`, `randomGetNextFrom`, `chrObjRandomGetNext`, `chrObjRandomSetSeed`) already mirrors `random.s`/`chrObjRandom.s` line-by-line using these two helpers plus plain `<<`/`>>` (which already matched `dsll`/`dsrl` correctly — only the `32`-suffixed pair was wrong). One side effect worth flagging: `g_randomSeed`/`g_chrObjRandomSeed`/the `randomGetNextFrom` caller-owned seed now genuinely carry values beyond 32 bits between calls (matching the real `sd $a0, g_randomSeed` full-64-bit store) — callers only ever consumed the `u32`-truncated return value, so this is inert for them, but don't "simplify" the stored seed back to 32 bits.

**Verified:** 242/242 clean build; full 21-level `verify.sh sweep` PASS, no crashes (loot/AI RNG touches every level, so a sweep is the right breadth check here, not just one level). Not (and cannot be) verified: the corrected stream against a live N64/emulator capture of `random.s`'s real output — the hand trace + independent numeric reproduction is the practical ceiling of verification available in this repo without ROM-attached hardware/emulator tooling.

**User-facing note:** this silently changes every RNG-derived sequence (loot placement/rolls, AI behavioural variance, anything seeded from `randomGetNext`/`randomGetNextFrom`) from the first draw, in every build going forward — savegames remain compatible (the seed is not part of save state), but **existing replays recorded against the old stream will diverge** (replay state persists `g_randomSeed` itself, so a replay recorded pre-fix and played back post-fix, or vice versa, will not reproduce the same run). Called out in the v0.3.0 release notes per M-139c.

## D285 — M-140: a real Steam Deck SIGSEGV during D255 live-testing turned out to be a SEPARATE, previously-undocumented bug — an unprotected shared-list read in `sndHandleEvent`'s preempt-scan (`src/snd.c:366`), classic D4-class main-thread-vs-audio-thread race, case identified but NOT fixed

**Context.** While setting up a live Deck test session for D255 (SSH access, `GE_D255=1`), found a fresh `ge007.crash.log` on the user's Deck (mtime inside the user's active testing window, confirmed by the user as "any crashes from the last hour or two on that build is my testing") from the **published v0.2.1 binary**, which still had its own debug symbols (`with debug_info, not stripped`). Symbolized directly on-device with `addr2line` against the exact crashing binary — not a rebuild, not a guess:

```
SIGNAL: 11  FAULT ADDR: 0x62
#00 crashStackTrace                       port/src/crash.c:332
#01 sndHandleEvent                        src/snd.c:366          <- fault site
#02 sndPlayerVoiceHandler                 src/snd.c:253
#03 _timeToSamplesNoRound                 src/libultra/audio/synthesizer.c:330
#04 amHandleFrameMessage                  src/audi.c:551
#05 amMain                                src/audi.c:452
#06 portThreadWrapper                     port/src/libultra.c:387
```

**This is a different crash from D255.** D255's known signature is `FAULT ADDR: 0x10`, PC inside `vtxstore_fix_refs` (render/game-logic thread, via `bossMainloop`). This one is `FAULT ADDR: 0x62`, entirely inside the **audio thread** (`portThreadWrapper` → `amMain`), a different subsystem and a different thread. The user's testing session (explosions/terminal-destroy repro) apparently hit *this* bug at least once, not necessarily D255 — the two may have been conflated in earlier reports because both surface as a SIGSEGV during the same kind of combat/explosion testing.

**Root cause (static, code-read only — not yet gdb-confirmed on a live crash).** `src/snd.c:366` is `iterState->unk3e` inside `sndHandleEvent`'s pool-exhaustion preempt-scan:
```c
ALSoundState *iterState = (ALSoundState *) D_800243E4.node.prev;
do {
    s32 ownerlessLoop = (iterState->unk3e & SOUND_FLAG_LOOPED) && ...   // line 366: the fault
    ...
    iterState = (ALSoundState *) iterState->link.prev;
} while (limitReached && iterState != NULL);
```
`D_800243E4` (`struct D_800243E4_s D_800243E4 = { {NULL, NULL}, NULL };`) is the sentinel head of the active-`ALSoundState` linked list; `node.prev` is genuinely `NULL` when the list is empty (verified: `sndDisposeSound`, `snd.c:868-875`, resets `node.next`/`node.prev` to `NULL`-adjacent state as the list empties; `sndSetupSound`, `snd.c:802-824`, is the only inserter, always under a lock — see below). This do-while dereferences `iterState` **before** its own `!= NULL` check ever runs (do-while always executes the body once) — safe *only* if the caller can guarantee the list is non-empty. The scan is gated on `limitReached` (the 8-voice pool is full), which *should* imply at least 8 live nodes — so a `NULL` here means the list state and the pool-full accounting have desynced.

**Prime suspect: unprotected concurrent access, the established D4/D147/D152 race class** (`docs/porting-notes.md` §D4 — "N64 `osSetIntMask(OS_IM_NONE)` is their only mutual-exclusion primitive; on PC the 'audio interrupt' is a real preemptible thread, so a no-op shim = no mutual exclusion = concurrent linked-list mutation" — already fixed generally via a process-wide recursive mutex, D4/D152). Checked directly this session: `sndSetupSound` (the `D_800243E4` list-insert path, reached via `sndPlaySfx`, which **15 different `src/game/*.c` files call directly** including `explosion.c` — i.e. from whatever thread the caller runs on, not queued through the audio thread) wraps its list mutation in `osSetIntMask(OS_IM_NONE)`/restore (`snd.c:807,827`). **`sndHandleEvent`'s entire body (`snd.c:263-668`) — including the preempt-scan that reads/walks/mutates the same `D_800243E4` list — never calls `osSetIntMask` at all.** A writer (`sndSetupSound`, potentially running on the main game thread via a direct `sndPlaySfx` call) is lock-protected; this reader (`sndHandleEvent`, running on the audio thread) is not — a classic asymmetric-locking gap, exactly the shape that lets one thread observe another's list mutation mid-update.

**Fits the reported repro pattern.** Explosions synthesize a burst of near-simultaneous SFX triggers (explosion + debris/glass + metal), i.e. a burst of `sndPlaySfx` calls in a tight window — the highest-contention case for a race that needs the audio thread's preempt-scan (only entered when the 8-voice pool is *exhausted*, itself a busy/high-SFX-load state) to overlap a main-thread insert/remove. This matches the user's report of explosions/terminal-destruction reproducing it and Frigate (presumably lower simultaneous SFX load in that test) not reproducing it — a timing race, not a deterministic trigger, fits both observations.

**NOT fixed — two open questions before it should be.**
1. **Is the missing lock a genuine N64-decomp-fidelity gap, or ground-truth-faithful?** Unlike D4's already-fixed cases, this would be a *new* `osSetIntMask` call site in decompiled game code (`src/snd.c`), not a shim-implementation change. If the real N64 assembly at this function's address (comment header cites `9630 70008A30`-style N64 addresses elsewhere in this file — check `sndHandleEvent`'s own) already has an interrupt-mask pair around the equivalent scan that the decompiler dropped, restoring it is a decomp-fidelity fix (expected, not a rule-#2 case). If N64's real code genuinely never masks interrupts there (plausible if the real audio ISR's preempt-scan only ever runs from within the ISR itself, inherently serialized against the main CPU by hardware, with no real race possible on N64), then adding a new lock here is the same class of exception D4 already established (a PC-only concurrency fix, not a behavior change) — but that determination needs an MIPS-disassembly check against the real ROM, not a code read of the decomp alone.
2. **Not yet reproduced/confirmed via gdb on a live crash** — this is a strong static lead from one real crash's symbolized backtrace plus a code read, not a confirmed root cause. The next Deck session should `gdb attach` (per D253's established recipe) while reproducing via a burst of simultaneous SFX (an explosion is the natural trigger) and confirm `D_800243E4.node.prev == NULL` (or a torn/partially-updated node) at the fault, and check the exact other thread's state at that moment if possible.

**Do NOT conflate with D255** — different fault address, different thread, different subsystem, different code path. Both may be independently real; testing one does not rule out the other, and a fix for one will not affect the other.

**Next, concrete:** (1) settle question 1 above (N64 disassembly check or accept the D4-class port-layer exception); (2) if a lock is warranted, wrap `sndHandleEvent`'s preempt-scan (and audit the rest of the function for the same gap against every other `D_800243E4` touch point in the file) in `osSetIntMask(OS_IM_NONE)`/restore, matching `sndSetupSound`'s existing pattern; (3) live-verify via gdb on a real reproduced crash, not just the static read.

**M-141: FIXED, per the accepted D4-class port-layer exception; not yet live-verified.** Question 1 is settled by precedent rather than a fresh N64 disassembly read: `sndSetupSound` (the paired writer) already takes `osSetIntMask(OS_IM_NONE)`/restore completely unconditionally — **no `#ifdef PORT` at all** — around the identical `D_800243E4` list mutation, in this same file, with no N64-behavior caveat noted anywhere near it. Since `osSetIntMask` is genuine N64 API (not a port shim call added out of nowhere) and the port's own `osSetIntMask` implementation (`port/src/libultra.c`, "Interrupt mask -> global recursive lock", D147/D152) is a **process-wide recursive** critical section — safe to nest inside `alEvtqPostEvent`/`alSynSetVol` calls without deadlock risk — adding the identical lock/unlock pair around `sndHandleEvent`'s preempt-scan follows the exact same already-accepted shape as its sibling writer, unconditionally, no `#ifdef PORT` needed (matches the surrounding code's own style). Audited the rest of `sndHandleEvent` (`snd.c:263-680`) for other `D_800243E4` touch points per the open item above: **the preempt-scan is the only one** — nothing else in the function reads or writes that list. Fix: `src/snd.c`, wraps `ALSoundState *iterState = ...` through the closing `} while (limitReached && iterState != NULL);` in `OSIntMask d285Mask = osSetIntMask(OS_IM_NONE);` / `osSetIntMask(d285Mask);`. **Verified:** build green; `-level_09`/`-level_20` boot 1400+ frames each with zero crashes and zero `D152: osSetIntMask lock stuck` warnings (which would fire if this new lock ever wedged). **Not verified:** a `GE_INPUTSCRIPT`-driven rapid-fire stress test (modelled on D219/M-90's scripted-Z-pulse approach) was attempted to force the 8-voice pool exhaustion this scan requires, but the script's own parse-confirmation log line never appeared in this run's output — the scripted-input harness did not visibly activate for a reason not root-caused this session (worth a quick look next time `GE_INPUTSCRIPT` is used: confirm it still fires on current `main`/`release/v0.3.0` before trusting a "no crash" result from it as a real stress test). No live Deck/gdb reproduction attempted this session (no Deck SSH session active). Given the D4/D152 precedent's exact-shape match, confidence in the fix is HIGH; confidence that it's been exercised under real load is LOW — the honest bar met here is "correct by precedent, build/sweep-clean," not "reproduced-then-fixed." Next: an interactive session (ideally the same live Steam Deck setup as D255/M-140, explosion-heavy combat, e.g. Facility/Frigate) should confirm the crash no longer reproduces.

## D286 — M-140: user-confirmed, real (not system-side) ~8s frame freeze on Caverns, Steam Deck, same live session as D255/D285 — self-recovered, not a crash; case not yet root-caused

**Context.** Caught live during the same Deck SSH session used to fix D255: the game's own hang watchdog (`portHeartbeatCheck`, `port/src/libultra.c`) fired once, logging `kernel heartbeat: no frame rendered for 8008866 us (frames=46404); state:` (unit bug in the log line, see below — the raw value is *microseconds*, printed with a `ms` suffix and no `/1000`; fixed this session, `port/src/libultra.c`) followed by a per-thread/per-queue state dump and a `crashDumpThreads` backtrace of the watchdog thread itself (`portTickThread` → `portHeartbeatCheck` → `crashDumpThreads`, `port/src/libultra.c:266/315`, `port/src/crash.c:458` — this is the watchdog reporting on *itself*, a diagnostic dump, not a fault). **User confirmed by direct observation: this was a real, visible freeze** (not a display/system hiccup — the Deck's own WiFi had one unrelated sleep/wake event earlier in the session, so system-side causes were a live alternative hypothesis, now ruled out by the user's own eyes on the screen), happening "pretty early" in a Caverns playthrough, not tied to an explosion or terminal (ruling out D255/D285 as the cause — this is a third, distinct issue).

**Outcome: self-recovered.** VI-post/frame-render logging resumed normally within the same log stream immediately after the watchdog fired, and the session continued (the user later closed the game normally, no crash log, all 5 game threads still `started=1`/`exited=0` at the time of the dump — nothing was permanently deadlocked). This is a genuine ~8s *stall*, not a hang or a crash.

**Not root-caused.** No distinguishing log line immediately precedes the stall (just routine `D51 vi post` lines) — nothing points at a specific subsystem yet. Caverns is one of the larger/more complex levels (cross-ref the D250 "FPS inconsistent" lineage, though D250's own symptom was sustained sub-60fps, not a multi-second total stall, so likely a different mechanism — a one-time heavy room/asset stream-in stall is the leading guess, unconfirmed). A single occurrence with no repro trigger identified; **do not spend session time trying to force a repro** without a better lead. If it recurs, the next capture should note (a) exact in-level location/trigger if any pattern emerges, (b) whether `Rooms`/`roomsDrawn` or a texture/model load was in flight via existing debug logging, and (c) whether it's Caverns-specific or level-general (test other large/open levels — Streets, Egyptian — the D249/D250 neighbourhood).

**Also fixed in passing (cosmetic, no behavior change):** the heartbeat log's unit label — `now - g_lastFrameUs` is computed in microseconds but was printed as `%llu ms` without converting; now divided by 1000 so the log actually reads correct milliseconds. | OPEN — one user-confirmed real freeze on Caverns (~8s, self-recovered, no crash), not root-caused, no repro trigger identified; distinct from D255/D285. Log-unit cosmetic fix landed alongside.

---

## Entries migrated from §F table (2026-09-15)

> Migrated verbatim from the §F index table by `migrate_bucket_a.py` (ctx-opt P1, bucket A): topic/status cell
> content moved unchanged; the table row now carries a one-line
> summary + pointer. Nothing deleted.
## D137 — right-mouse (aim-sight) crash:

right-mouse (aim-sight) crash: `gunDrawSight` (`gunfire.c:6235`) `s32 sp54` holds a `Gfx*` the whole function (`sp54 = *gdl`; `texSelect(&sp54,…)`/`display_image_at_position(&sp54,…)` take `Gfx**`). As N64 `s32` it is 4 B, so `*(Gfx**)&sp54` reads 4 B of the adjacent stack float as the pointer high word → `_g = 0x03e4fca0_70081220` → wild DL write in `texSetRenderMode` (`othermodemicrocode.c:177`) the moment the crosshair raises. §A.

**Status (from §F table):** FIXED (`src/game/gunfire.c` `#ifdef PORT` — `sp54` is `Gfx *`)

## D123 — crash class C1: `chrIsNotDeadOrShot` NULL deref on 6 levels (Dam/Runway/Frigate/Statue/Streets/Cradle).

crash class C1: `chrIsNotDeadOrShot` NULL deref on 6 levels (Dam/Runway/Frigate/Statue/Streets/Cradle). D122's `OBJ_TAIL_DESC` zeroed the widened `VehichleRecord/AircraftRecord.ailist` slot (w32), but `prop.c:1764/1786` reads a pre-populated int AI-list id there → `ailistFindById(0)` → `GAILIST_AIM_AT_BOND` → `ai()` runs a CHR aim list with `ChrEntityp==NULL`

**Status (from §F table):** converter fixed (`OBJ_ID_WORDS`); C1 cleared on all 6, residual crashes are fast3d (C2)

## D125 — crash classes C3+C6 (Aztec/Bunker2/Surface2):

crash classes C3+C6 (Aztec/Bunker2/Surface2): **root cause found (M-16)** — `tools_pc/d88_emit.py:374` assigns a 4-byte literal to an 8-byte slice (`out[dst_o+0x30:dst_o+0x38] = b"\x00…"`, the stan zero-fill in `emit_pad`); every pad/boundpad record silently shrinks the output bytearray by 4 B, and once the tail falls below the boundpad plink string blob, later verbatim leaf writes hit Python out-of-range slice semantics (insert at current end, not relocated offset) → boundpad names drift/truncate (Aztec pad33 `p138d2`→`8d2`) → `stanPackId()` reject → `getposstan()` NULL stan → `setupDoor` leaves door `model=NULL` → crash `propobj.c:13601`. M-14's "propDefs zeros in RAM" was a misread; sidecar propdefs were always correct (M-15) and post-load RAM matches the sidecar byte-for-byte (M-16)

**Status (from §F table):** **resolved (M-16b)** — line 374 → 8 NULs; all 21 sidecars regen'd; **Aztec `-level_28` now PASSES** (was C3 CRASH). Bunker2 falls through to a separate DOOR-tail `linkedDoor` layout bug (C3 residual)

## D126 — crash classes C3r/C4/C6 (Bunker2 `-level_27` `door7F054FB4` propobj.c:13523, Depot `-level_30` prop.c:902, Surface2 `-level_43` loadobjectmodel.c:393):

crash classes C3r/C4/C6 (Bunker2 `-level_27` `door7F054FB4` propobj.c:13523, Depot `-level_30` prop.c:902, Surface2 `-level_43` loadobjectmodel.c:393): the objective sub-records `criteria_picture` (30), `criteria_roomentered` (32), `criteria_deposit` (33), `setup_objective_text` (35) each end in a `T *next` list pointer that the setup walk (`set_parent_cur_obj_*` / `setup_briefing_text_entry_parent`) writes unconditionally. On PC that pointer widens 4→8B and lands 8-aligned at offset 16 → struct is 24B/6w (N64 16/20). `d88_propdefs.py` emitted them at N64 size via the generic arm → the runtime 8-byte `->next` write clobbered the *next* record's header → propdef walk desynced, command indices drifted ~100, `linkedDoorOffset + arg2` resolved to the wrong record → door `linkedDoor` chain walked into garbage.

**Status (from §F table):** resolved — `d88_propdefs.py` PROPDEF_PC_BYTES[30/32/33/35]=24 + typed handler; `loadobjectmodel.c` sizepropdef PORT returns 6. Bunker2/Depot/Surface2 now PASS (13→16/21)

## D124 — crash class C2: fast3d bad texture pointer.

crash class C2: fast3d bad texture pointer. **Jungle** (`0xabcd0824`): `gimgSyncCompiledGlobalDLs()` slot-detect keyed on the post-fixup marker, which `texLoad()` had already overwritten → compiled `globalDL_0xNNN` explosion DLs kept link-time `IMAGESEG` words → latent on every level, tripped by the first explosion-DL draw. **Facility** (`0x72181ee8`): see **D130** — the model-GDL-relocation hypothesis (M-14 addendum) was WRONG; real cause was `romdataFixupFont`

**Status (from §F table):** Jungle fixed (`port/src/gimgfixup.c`); Facility → D130

## D131 — crash class C2m (Jungle `-level_37`):

crash class C2m (Jungle `-level_37`): fast3d `gfx_sp_matrix` AV on a wild matrix pointer `0x401c68e0` (`gfx_pc.cpp:1046`) ~frame 300, when the first explosion/smoke prop renders. `explosionRenderPropSmoke` builds `gSPMatrix(gdl++, osVirtualToPhysical((void*)&dword_CODE_bss_8007A100), …MODELVIEW)` (+ `applyRoomMatrixToDisplayList`). `osVirtualToPhysical()` is a `u32`-returning shim (`libultra.c:1207`) → truncates the compiled `.bss` symbol's `0x1_00000000` module high word → w1 = `0x40xxxxxx` → `seg_addr()` returns it raw. Same class as D94 (chraction 32-bit ptr truncation), but in a GBI DL word. The projection matrix in the same DL is fine (`get_BONDdata_field_10E0()` is a runtime `0x70xxxxxx` ptr, fits u32); the `gSPDisplayList(&globalDL_0xNNN)` refs are fine (port `Gwords.w1` is 64-bit `uintptr_t`, no macro truncation). ~30 `osVirtualToPhysical(<compiled matrix/vtx symbol>)` sites exist (`explosion.c`, `glass*.c`, `blood_animation.c`, `bondview2.c`) — all latent until that effect first draws.

**Status (from §F table):** resolved — `seg_addr()` restores the module high word for any fallthrough w1 in `[0x40000000, 0x70000000)` (module fixed-based at `0x140000000`, no ASLR; DRAM/KSEG0/segmented/phys all handled earlier). Jungle renders to frame 2400+ clean; `-level_20`/`-level_24` unregressed. **18→19/21.** (`-level_09` has a separate pre-existing boot crash, see below.)

## D133 — intro render triage (M-20):

intro render triage (M-20): the "intro renders mostly black" item in the M-18/M-19 handoff is **NOT a regression** — it is the D75/D76 parked-cosmetic steady state. M-17 (`9ec6121e`), whose handoff claimed "the entire intro renders — logos → gun barrel → cast", was built in a scratch worktree and captured with the same `GE_PCDUMP="20-900:20"` window: coverage is **pixel-identical** to HEAD `0b5f5d1a` (legal screen 6677 non-clear px / 2.17% on both; centred-logo frames ~7% on both; near-black between). 2D/text/texture layers draw; animated character-model layers (Nintendo-logo transform, gun-barrel Bond, cast models) never appear = D75 exactly. No cheap regression to fix; parked. Secondary: Facility `-level_34` + Jungle `-level_37` re-verify FAILED this pass (boot crash / frame-2 hang) — flagged in LEVEL-STATUS for a clean-machine re-check, not investigated (scope).

**Status (from §F table):** RESOLVED — not a regression, parked (D75/D76, `GRAPHICS-BACKLOG.md`)

## D134 — Frame-2 boot hang / sweep "NO-FRAMES" flakiness — root-caused and fixed (M-22).

**Frame-2 boot hang / sweep "NO-FRAMES" flakiness — root-caused and fixed (M-22).** `osSpTaskStartGo` (port/src/libultra.c) posted the SP/DP task-done events with `OS_MESG_NOBLOCK` into the scheduler's 8-slot `interruptQ`, which the 60 Hz VI pacemaker thread also fills with `VIDEO_MSG`. A gfx task runs the whole frame synchronously on the sched thread (fast3d); the first two frames take 30-80 ms, so the pacemaker queues several retraces meanwhile. Once the queue is full the RSP-done post is **silently dropped** -> `__scMain` never clears `sc->curRSPTask` -> the client never gets `OS_SC_DONE_MSG` -> `bossMainloop`'s `pendingGfx` never clears -> permanent stall (log: `frame 2 rendered`, then `kernel heartbeat ... frames=2`, retrace queue pinned at `valid=8/8 ret=-1`). Reproduced 2 of 3 `-level_09` boots on an idle machine, i.e. most of what M-13..M-21 wrote off as D117 nondeterminism / machine load. `OS_MESG_BLOCK` is NOT the fix (the done event is posted from the sched thread, the queue's only consumer -> self-deadlock): `portPostEventForce()` drops the OLDEST message (always a stale retrace) and retries, and `portPostVIEvent` now reserves 2 slots. Verified: 6/6 `-level_09` boots reach frame 600 with 0 heartbeats (pre-fix 1/3).

**Status (from §F table):** FIXED (`port/src/libultra.c`)

## D135 — Firefight crash — `bgTestHitOnObj` (`propobj.c:8446`) is an unported N64 GBI parser (M-22, WS6 playtest).

**Firefight crash — `bgTestHitOnObj` (`propobj.c:8446`) is an unported N64 GBI parser (M-22, WS6 playtest).** The object bullet-hit test walks a model's display list command stream with hardcoded 8-byte-N64-`Gfx` byte/word indices (`*(s8*)gdl`, `((u32*)gdl)[1]`, `((u8*)gdl)[5..7]`, the G_TRI4 nibble reads, the backward `G_SETTIMG` scan). On PC the model GDL is the 16-byte `Gfx` layout (`d43_emit.py`, `{u64 w0; u64 w1}`), so every access reads the wrong bytes → the walk desyncs on the first `G_VTX`/`G_TRI1` → `v->coord.x` deref off the end of the DL → AV (`0x709ae02a`, a near-miss past a live DRAM ptr = the §B signature). Triggered by `propobjFindHit` every time a shot ray resolves a hit on an object model (guards' dropped weapons, attached objects, shootable props) — invisible to `level_sweep.sh` (no weapon fire). Two sibling faults seen the same session (`texSelect`/`texSetRenderMode` wild `gdl`, `objFreeEmbedmentOrProjectile` `prop->obj`) are likely the same corrupted-object cascade. PD `bgTestHitOnObj` (`pd_port/src/game/bg.c:3635`) is the fully-ported ground truth.

**Status (from §F table):** FIXED (`src/game/propobj.c`, `#ifdef PORT` — reads `gdl->words.w0/.w1` shifts). **Second fault, same walker:** after the parser fix the Facility firefight then hit `0xc0000094` (int div-by-zero) at `objHit` `propobj.c:9717` `randomGetNext() % impact_sounds->thing2_len` — the ported texnum-recovery (`*(s16*)(phys(w1) − 8)`) returned a bogus positive `texturenum` (`(u32)` cast truncated a 64-bit `words.w1`; the N64 "texnum 8 bytes before the texture data" layout isn't guaranteed for the port's converted/replaced GDLs) → `g_HitTypeSounds[texnum & 0xf]` picked an entry with `thing2_len == 0`. Fixed: the PORT texnum branch now always returns `-1` (→ `g_HitTypeSounds[0]` = `isnd_default`, `thing2_len == 1`, safe). `texturenum` only flavours the bullet-impact sound/decal, so generic-hit is an acceptable degradation (park with D77). **Follow-up:** the BG room-geometry sibling `sub_GAME_…` (`bg.c:~3373-3646`, walks the D85-widened `ptr_expanded_mapping_info`) has the identical unported parser pattern — latent, fires on shooting walls/floor; not yet ported.

## D136 — Stage-unload crash — `objFree` `obj->prop` reads float garbage (M-22, WS6 Facility playthrough).

**Stage-unload crash — `objFree` `obj->prop` reads float garbage (M-22, WS6 Facility playthrough).** After D135 cleared the firefight crashes the level is completable; then on teardown `lvlUnloadStageTextData` → `cleanupObjects` (`cleanup_objects.c:42`) → `objFreePermanently` → `objFree` (`propobj.c:994`) → `objFreeEmbedmentOrProjectile(obj->prop)` faults at `:888` (`prop->obj` deref). Real gdb `bt`: `obj` = `0x701bd730` (valid) but `obj->prop` = `0x3e567400405c30c0` — two packed `f32` (~0.21, ~3.44), i.e. a `coord3d` read as a pointer → `ObjectRecord.prop` (N64 `/*0x10*/`, after `inherits PropDefHeaderRecord` + `s16 obj; s16 pad; u32 flags; u32 flags2;`) lands at the wrong PC offset for this object's propDef type, OR the record is undersized and setup's `->prop` store never happened / was clobbered. Same family as D122/D132 (`tools_pc/d88_propdefs.py` per-type record sizing vs the compiled `ObjectRecord`/subtype layout). Type not yet identified — needs a gdb run with a breakpoint on `objFree` to read `obj->type` (`cleanupObjects` frees WEAPON/AMMO/MAGAZINE/COLLECTABLE/MONITOR/RACK/AUTOGUN/HAT/ARMOUR/GAS_RELEASING/VEHICHLE/AIRCRAFT/GLASS/SAFE/TANK/TINTED_GLASS via this path). Mission completes; crash is before the next briefing loads.

**Status (from §F table):** ROOT-CAUSED → **D139** (it was NOT a layout bug — `cleanupObjects` read the type from the wrong byte on LE and walked off the propDef blob).

## D139 — Stage-unload crash root cause + fix (M-23).

**Stage-unload crash root cause + fix (M-23).** `cleanupObjects` (`cleanup_objects.c`) walks `g_CurrentSetup.propDefs` with `(u8)obj[0]` as the record type — correct only because the header word `[u16 extrascale][u8 state][u8 type]` is **big-endian** on N64, so `type` is the low byte. On little-endian PC the low byte is `extrascale`, so `(u8)obj[0]` is never `PROPDEF_END` → the `while` runs off the end of the blob, and the `switch` dispatches on garbage → `objFreePermanently()` on records that aren't `ObjectRecord`s → `obj->prop`/`obj->model` read adjacent float data (`0x3e567400405c30c0`) → AV in `objFreeEmbedmentOrProjectile` (`propobj.c:888`). `sizepropdef()` and every other propDef consumer already use `pdef->type` (struct member, offset 3) — only this one spot used the raw byte cast. Not a `d88_propdefs.py` layout issue at all.

**Status (from §F table):** FIXED (`src/game/cleanup_objects.c`, `#ifdef PORT` — `CLEANUP_PDTYPE(o)` = `((PropDefHeaderRecord*)o)->type`). **UNVERIFIED** — the fast teardown test (pause→abort) hit D140 first; `-level_09` unregressed. porting-notes.md §C.

## D140 — Pause-menu crash (M-23 open → M-27 fixed).

**Pause-menu crash (M-23 open → M-27 fixed).** `bondviewRenderWatch` (`bondview2.c:8604`) → `bondviewTransformManyPosToViewMatrix(g_CurrentPlayer->field_23C, objheader->numMatrices)` with `field_23C == NULL` → `matrix_4x4_copy(src=0x0)` AV. Root cause (NOT "cache never allocated"): `something_with_watch_object_instance` (N64 player +0x230) is a **`struct Model` (0xBC) + its RW-data pool (0xC8) punned into a 0x184-byte hole** (`field_234..field_3B0`); three of those "fields" are Model members the game reads by name — `watch_scale_destination`==`.scale`, `pause_watch_related_adjust`==`.animframe1`, **`field_23C`==`.render_pos`**. `render_pos` IS populated (by `instcalcmatrices` via `subcalcmatrices` at `:8524`), but on x86-64 `sizeof(struct Model)` grows (4→8B ptrs) so `field_23C` no longer overlaps `.render_pos` → reads Model padding (~0) → NULL. Same class as D100 (gait Model) / D102 (weapon Model) / D56 (this very watch Model's pool). Fix: `#ifdef PORT` makes `something_with_watch_object_instance` a real inline `struct Model` + `u32 watchRwPool[192]` in `struct player` (`bondview.h`); `bondview2.c` redirects the 3 named reads to the member via `GE_WATCH_{ANIMFRAME,SCALE,RENDERPOS}` macros (N64 `#else` = the verbatim field). The D56 branch of `sub_GAME_7F07E7CC` drops its `static u8 watchRwPool[0xC8]` for the inline pool.

**Status (from §F table):** FIXED (`src/game/bondview.h`, `src/game/bondview2.c`; `#ifdef PORT`). Exposed the next watch-render crash → **D141**. `-level_09`/`-level_20` unregressed (3/3). porting-notes.md §A.

## D141 — Watch-menu item-model crash (M-27), exposed by D140.

**Watch-menu item-model crash (M-27), exposed by D140.** Pressing Start → watch page renders → `set_enviro_fog_for_items_in_solo_watch_menu` (`gunfire.c:1620`) → `modelGetNodeRwData(&model, root)` with `root` garbage → AV reading `root->Opcode` (`model.c:478`). `gunfire.c:1720-1743` walks `bodymodel->Switches[]` (a `ModelNode*` array) with **raw byte offsets** `((u8*)Switches) + j + 0x48` / `+ 0x5c` and `j += 4` — assumes 4-byte N64 pointers. `0x48/4=18`, `0x5c/4=23`, `k=j>>2=0..4` → intent is `Switches[18+k]` and `Switches[23+k]`. PC's 8-byte stride makes the raw math read misaligned garbage → bogus non-NULL `ModelNode*`. porting-notes.md §B (raw hardcoded stride into a pointer array, cf. D128). Fix: `#ifdef PORT` uses `bodymodel->Switches[18 + (j>>2)]` / `[23 + (j>>2)]`.

**Status (from §F table):** FIXED (`src/game/gunfire.c`, `#ifdef PORT`). Watch page now renders (weapon page verified — mirrored text D114/D116, dark weapon model D75, both parked cosmetic). `-level_09`/`-level_20` unregressed.

## D150 — Watch objective/briefing page crash (M-28).

**Watch objective/briefing page crash (M-28).** Opening the watch → BRIEF (any subpage) or OBJECTIVES page → AV in `strcat` (`str.c:25`, `Rdi/src = 0`). The pages assemble their display text with `strcpy`/`strcat(buf, langGet(id))` (`options.c:3940-4068`, `front.c:3538-3562`, `propobj.c`). `langGet()` returns **NULL** on PC for a string bank the menu flow never loaded (D129/D143 guards — briefing/objective banks in the in-level watch flow), so `strcat(buf, NULL)` derefs NULL. On N64 those ids always resolve so the decomp never guards. Twist: `str.c`'s `strcpy/strncpy/strcat` are `__nonnull__` **builtins** to GCC, so a plain `if (src == NULL)` on the parameter is optimised away as provably-dead (verified: `-Og` elided it, no `test` in the disassembly). Fix: `#ifdef PORT` NULL guards that launder the pointer through an empty `__asm__` (`GE_IS_NULL()` / `ge_launder_ptr()`) so the check survives — NULL src → treat as empty string (blank text, not a crash), NULL dst → return. Same philosophy as the D143 textRender/textMeasure NULL guards, one layer down. Inert on N64.

**Status (from §F table):** FIXED (`src/str.c`, `#ifdef PORT`). Build green (`ntsc-final`); `-level_09` 600+ frames @ 91.67% unregressed. Watch-page repro is interactive — user to re-verify. porting-notes.md §C.

## D151 — Watch data all blank — objectives / mission background / M / Q / Moneypenny (M-28, user bug report).

**Watch data all blank — objectives / mission background / M / Q / Moneypenny (M-28, user bug report).** Not a crash (D150 fixed that); every line renders empty. The watch text lives in `propDefs` record types **35** (`WatchMenuObjectiveText`) and **23** (`ObjectiveStart`). In the setup stream each record's 3rd word is a **plain 32-bit language slot id** (see `assets/obseg/setup/Usetup*Z.c` `propDefs[]`: `_mkword(0,_mkshort(0,35)), menu, 11281, 0`). The decomp structs `struct watchMenuObjectiveText` / `struct objective_entry` decode that word as `u16 reserved; u16 text;` and read `text` at offset `0xA` — which only works on the N64 because the id sits in the **low 16 bits of a big-endian word**. On LE, `tools_pc/d88_propdefs.py` `_bswap32`'s the whole word (correct for a 32-bit field), moving the id to the low bytes, so `text` (u16 @0xA) reads `0x0000` → `langGet(0)` → NULL → `get_ptr_text_for_watch_breifing_page` / `get_text_for_objective` return NULL/blank. This is the "briefing/objective text renders blank" side-effect left open under D143 / D150. The per-level lang bank itself IS loaded in-level (`langLoadToAddr(langGetLangBankIndexFromStagenum(stageId))`, `prop.c:1274`).

**Status (from §F table):** FIXED (`src/bondtypes.h`, `#ifdef PORT` — `text` is a full `u32` covering the whole word in both structs; N64 `u16 reserved; u16 text;` kept under `#else`). No converter change, no sidecar regen. `nextentry`/`difficulty`/`unkD` unchanged. porting-notes.md §C. Build compiles clean; interactive re-verify pending (open the watch in any level).

## D130 — crash class C2 (Facility `-level_34`, Runway `-level_35`):

crash class C2 (Facility `-level_34`, Runway `-level_35`): `import_texture_i8` AV on a wild texture pointer (`0x72181ee8`), baked into the HUD glyph DL by `gDPLoadTextureBlock(gdl, curchar->pixeldata, …)` when the level-title string ("Chemical Warfare Facility #2") renders `#`/`"`. Root cause: `romdataFixupFont` (`port/src/romdata.c`) converts the N64 24B fontchar array to the PC 32B array **in place**, forward-field, backward-glyph. PC−N64 stride = 8, so for glyphs 0/1/2 `d` overlaps `s` by less than the 24B read span and an early field write clobbers a later field's source mid-`for k<5` loop → glyph 1's `width` = `bswap(index)` = `0x01000000`, glyph 2's `pixeldata` = `pcPixOff+(index−pixStart)` ≈ `0x020002e8` → `+= font_base` → wild → fast3d AV. On N64 the array is read-only rodata, never re-laid-out, so no bug there. NOT the model-GDL relocation (`sub_GAME_7F0762E0` / `texLoadFromGdl`) the D124-Facility addendum suspected — `gdl` there is still a segmented `0x05xxxxxx` value so the `& 0x00ffffff` masks are correct, and `texLoadFromGdl` never copies a `G_SETTIMG` for these levels.

**Status (from §F table):** resolved — stage all 6 N64 fields in a local `f[6]` before writing any. Facility + Runway now PASS (16→18/21). Probes reverted.

## D152+ (a) — D152 addendum (M-31):

D152 addendum (M-31): static audit of every compiled-audio `osSetIntMask(OS_IM_NONE)` — **all balanced**, no unbalanced early-return. Real fixes: `sndSetSfxSlotVolume` now holds the mask across its list walk (matches its twin `sndDeactivateAllSfxByFlag`) + `sndApplyVolumeAllSfxSlot` batches the slot loop under one recursive hold (kills the fade-out lock-acquire storm); `portThreadWrapper` calls `imThreadExitRelease()` on thread exit (kills the "transient thread acquired `OS_IM_NONE` and died" leak + the pthread-id-reuse re-wedge). Steal-lock kept as backstop.

**Status (from §F table):** FIXED (`src/snd.c`, `port/src/libultra.c`, `#ifdef PORT`) — fade-out repro playtest-gated, not headless-verified

## D168 — `GE_PCDUMP` / F12 PPM captures were vertically flipped — the entire source of the bogus D114/D116 "HUD/text X-mirror" (M-33, developer-confirmed).

**`GE_PCDUMP` / F12 PPM captures were vertically flipped — the entire source of the bogus D114/D116 "HUD/text X-mirror" (M-33, developer-confirmed).** `gfx_opengl_dump_bound_fbo` (`port/fast3d/gfx_opengl.cpp`) wrote `glReadPixels` output straight to a P6 PPM. GL framebuffer origin is bottom-left; PPM P6 is top-row-first — so every capture (and the `tools_pc/golden/` set, and every screenshot pasted into the finding log since M-6) was upside-down. On real hardware / an actual screen the game renders correctly (developer confirmed). The successive D114→D116 "shared fast3d mirror" investigations — each of which found *every probed stage clean yet the output "mirrored"* — were reading an inverted capture and pattern-matching upside-down asymmetric content (text, guards, the Nintendo logo) as "mirrored". Fix: emit PPM rows bottom-to-top. `tools_pc/golden/*.png` flipped in place to match (see `tools_pc/golden/README.md`); regenerate from a real run when convenient. **Not runtime-verified this session** (no ROM / toolchain in the migrated tree — see the migration note) — needs a fresh capture to confirm text reads normally. Confidence: **high** (mechanism is unambiguous; developer has hardware confirmation the screen is correct). porting-notes.md §D2.

**Status (from §F table):** FIXED (`port/fast3d/gfx_opengl.cpp`); D114 / D116 reclassified as capture-orientation artifacts (below).

## D157 — Campaign never unlocks the next level (M-30, user bug — "complete Dam on Agent, Facility stays locked, no save").

**Campaign never unlocks the next level (M-30, user bug — "complete Dam on Agent, Facility stays locked, no save").** The type-23 objective record's `MinDificulty` is a big-endian `s32` at word 3; on N64 the 0–3 value sits in its low byte == offset `0xF`, and `struct objective_entry.difficulty` (`bondtypes.h`) reads `s8` @0xF. `d88_propdefs.py` `_bswap32`'s that word (correct — it IS a 32-bit field), moving the value to offset `0xC` on LE → the `s8`@0xF read yields **0 for every objective**. `get_difficulty_for_objective()` → Agent(0) for all → `objectiveIsAllComplete()` on Agent evaluates objectives that should be difficulty-gated out (Dam "Neutralize all alarms" = Secret Agent, "Install covert modem" = 00 Agent) → they're never complete → `bossReturnTitleStage` sees `objectiveIsAllComplete=0` → `end_of_mission_briefing()` (the `fileUnlockStageInFolderAtDifficulty` → EEPROM write) never runs. The D151 sibling — same struct, the field D151 said it "left unchanged". Confirmed via GE_SAVELOG playtest + a direct ROM dump of Dam's records (word 3 = 1/2/2/0).

**Status (from §F table):** FIXED (`src/bondtypes.h`, `#ifdef PORT` — reorder `objective_entry`'s post-swap tail so `difficulty` reads @0xC, `unkD` @0xE; N64 layout under `#else`; sizeof unchanged 16). Also fixed `tools_pc/dump_objectives.py` (same byte-offset bug — read `buf[off+12]` = BE high byte = always 0 → showed "Agent" for every objective) + regenerated `docs/dev/LEVEL-OBJECTIVES.md`. **User-confirmed:** Dam Agent → `objdiff=1/2/2/0` → `objectiveIsAllComplete=1` → `fileWriteSave` with the completion time; Agent checkmark shows in mission-select and persists across kill+restart. New `GE_SAVELOG=1` / `GE_UNLOCK_ALL=1` `#ifdef PORT` env diagnostics added along the way. 21/21 sweep PASS. porting-notes.md §C.

## D160 — Dam level-end cutscene (Bond off the dam / bungee) never plays — cuts straight to the MISSION COMPLETE report (M-30/M-31, user bug D148).

**Dam level-end cutscene (Bond off the dam / bungee) never plays — cuts straight to the MISSION COMPLETE report (M-30/M-31, user bug D148).** Progression is unaffected: `bossReturnTitleStage` fires, the report + next briefing load, EEPROM save lands (D157). So the *transition* works; only the scripted cutscene step is skipped or silently fails. Static trace: Dam's end sequence is `UsetupdamZ` ailist `ai_24` (id 0x1004, gated `if_bond_in_room_with_pad(0x4a01)` — Bond reaches the exit) → lock controls, `bond_set_locked_velocity` (auto-walk off edge), wait `if_bond_y_pos_less_than` (Bond falling) → fade → `if_objective_all_completed` branch → `camera_switch`/`jump_to_ai_list` into the abseil ailist `ai_17` (id 0x0412: `screen_fade_from_black`, `music_xtrack_play`, `guard_play_animation(0xb100…)` = rappel anim on a puppet chr, three `camera_switch` w/ fades, `exit_level`). `camera_switch(TAG,…)` resolves a `CutsceneRecord` (propDef type 46, records 322/324/326 paired with Tags 10/11/12) via `tagGetCommandIndex` + `setupGetPtrToCommandByIndex` (a **command-index count walk over the propDef stream** using `sizepropdef()`). CutsceneRecord itself is all-scalar 28 B, correctly converted (`d88_propdefs.py` 46→28). **Leading hypotheses, unverified:** (a) the index-count walk desyncs because some earlier propDef type Dam emits (Vehicle 321 — D122 "medium confidence" tail; TintedGlass) has a wrong PC `sizepropdef` stride → `camera_switch` gets a bogus `cdef`/`TagIndex<0` → cutscene silently skipped, `ai_17` falls through to `exit_level`; (b) the trigger chr for `ai_24`/`ai_25` isn't spawned / `if_bond_in_room_with_pad` never true and the level ends via a different path; (c) D75/anim-model family — the puppet `guard_play_animation` rappel anim doesn't render (third-person Bond model absent, like the gun-barrel/cast models) so the cutscene "plays" invisibly but the user perceives it as skipped (less likely — user reports an instant cut, not a blank pause).

**Status (from §F table):** DIAGNOSTIC SHIPPED, not root-caused (`c95713f5`, `#ifdef PORT` + `GE_D160=1`): trace points in `bossReturnTitleStage` (entry + caller), `bondviewSetCameraMode` (mode/stage/`g_IntroSwirl`/intro-anim-idx/caller), and AI cmds `EndLevel`/`exit_level`, `CameraLookAtBondFromPad`, `CameraSwitch` (prints `cdef` ptr / `TagIndex<0` / tag `NOT FOUND`). **Next step:** user plays Dam to the exit with `GE_D160=1` (via `debug.ps1`); the trace answers (1) does `CameraSwitch`/`CameraLookAtBondFromPad` execute at all → trigger fires? (2) if yes, is `cdef` a sane pointer → propDef-walk / converter bug (a), same family as D122/D132/D157; (3) if the cam cmds never run → ai_24 trigger / chr-spawn bug (b). If it's (c), `bondviewSetCameraMode` shows POSEND being set + control returns normally. This is likely a WRITE-UP-then-fix (propDef stride) rather than a hack. Confidence in hypothesis (a): medium. **M-106 static re-audit (no build/runtime access this session — see note below): hypothesis (a) independently RE-CONFIRMED RULED OUT.** `CutsceneRecord` (`src/bondtypes.h:3785-3792`, propDef type 46) is **all-scalar** — `PropDefHeaderRecord` (4B header) + `coord3d pos` (12B, 3×f32) + `theta`/`verta`/`pad` (4B each) = 28B exactly, zero pointer/union fields, so it cannot exhibit the D122/D126/D132/D209 pointer-width-growth pattern at all (nothing to widen); this matches `porting-notes.md`'s already-recorded M-31 conclusion (`§ "D160 — a propDef command-index walk desync..."`) that the stride cross-check is clean for all 48 types. Traced the `AI_CameraSwitch` path end-to-end (`chrai.c:3990-4013`): `tagGetCommandIndex`/`setupGetPtrToCommandByIndex` resolve `cdef`, then `gBondViewCutscene = cdef` and `bondviewSetCameraMode(CAMERAMODE_POSEND)`; the render side (`bondview2.c:1558-1590`) reads `gBondViewCutscene->pad`/`pos`/`theta`/`verta` directly off that pointer with no intermediate copy/struct-punning step that could introduce a *new* aliasing bug — this leg of the pipeline is clean. **Did not get far on hypotheses (b)/(c)** (AI trigger-fire vs invisible-puppet-render) — no interactive session / no build toolchain available in this worktree (`gcc`/`cmake` not on `PATH`; MSYS2 MINGW64 environment not active), so the `GE_D160=1` Dam-to-exit playthrough this row has needed since M-31 still has not been driven. **Confidence update: hypothesis (a) — HIGH (ruled out, independently re-derived, not just cited); (b)/(c) — unchanged, still need the live trace.** No code changes made (nothing to fix once (a) is ruled out); this is a documentation-only confirmation pass. See also D173 below — if the Cradle "Bond floating/spinning" symptom filed under this label is the puppet-render family (D75/D156), closing D173 first may explain the residual `#4 GOTO` here too. **M-107 — the live trace finally happened; (b)/(c) both REFUTED, real root cause moved to D243.** `GE_D160=1` on `-level_33` (Dam) plus a fresh-save normal playthrough all show the AI trigger firing and all three `camera_switch` commands executing (`bondviewSetCameraMode(7)` × 2-3, then `bossReturnTitleStage`) — the cutscene is NOT skipped and the puppet is NOT invisible; it renders, but corrupted (violent camera-shake artifact during each angle cut). Root-caused no further than "stale/freed display-list memory read, likely a render-pipeline race" — full writeup and evidence under **D243**, which now owns this bug going forward. This row's own title ("cutscene never plays") is superseded; kept for history.

## D161 — Depot (`-level_30`) ceiling renders as bright-blue speckle + radial "light-ray" streaks converging to a point instead of a dark corrugated roof (M-30/M-31…

**Depot (`-level_30`) ceiling renders as bright-blue speckle + radial "light-ray" streaks converging to a point instead of a dark corrugated roof (M-30/M-31, user bug B2 — `docs/dev/TEXTURE-GLITCH-ANALYSIS.md`).** NOT filtering (3-point+trilinear+aniso didn't touch it), NOT RC2 mip-contamination, NOT the RGBA5551 decode. Root-caused with a `GE_TEXDUMP` PPM+param dump (`gfx_opengl.cpp` / `gfx_pc.cpp`): the offending surface is **one texture, `fmt=CI siz=8b, 16×16`**, uploaded with **`rdp.palette_fmt == G_TT_NONE` (0)** while every other CI texture that frame has `palfmt=0x8000` (`G_TT_RGBA16`). GE draws the Depot roof with the TLUT **disabled** (`gsDPSetTextureLUT(G_TT_NONE)` — a legit N64 idiom: a CI-siz tile with no LUT feeds the raw 8-bit TMEM texel straight into the colour pipe, i.e. it behaves as a plain **I8** texture). fast3d's `import_texture()` ignored `palette_fmt` and always did the palette lookup for `fmt==CI`, indexing a **stale `rdp.palette`** left over from a previous texture → the 16×16 image came out as pure blue/cyan/magenta noise; the "radial rays" were that high-frequency noise aliasing on the steeply-receding ceiling plane (which is also why filtering couldn't help — the data was garbage). Fix: in `import_texture()`, `fmt_eff = (fmt==G_IM_FMT_CI && rdp.palette_fmt==G_TT_NONE) ? G_IM_FMT_I : fmt`, and dispatch on `fmt_eff` (CI4→I4, CI8→I8). Narrow — only changes surfaces that are currently 100% garbage. Verified: Depot corridor + control-room ceilings render as clean dark/grey industrial roofs; `-level_09`/`-level_30`/`-level_34` sweep PASS, unregressed. `GE_TEXDUMP` (env-gated PPM dump of every uploaded texture + a `fmt/siz/palfmt/palidx` log line) kept as an inert diagnostic. Confidence: **high** (before/after screenshots + the palfmt=0 evidence).

**Status (from §F table):** FIXED (`port/fast3d/gfx_pc.cpp` `import_texture`, `port/fast3d/gfx_opengl.cpp` diag). porting-notes.md §D.

## D152 — Mission-failed / objective-failed → permanent black screen (M-28, user bug report — LIVE process inspected, not yet fixed).

**Mission-failed / objective-failed → permanent black screen (M-28, user bug report — LIVE process inspected, not yet fixed).** User killed Trevelyan in Facility (Ctrl = fire, no crouch bind) → failed the objective → screen faded to black and never came back; process stayed alive under `debug.ps1`. `gdb.txt` (kept running): rendering stops dead at frame 12600 (`frames=12657` frozen across 206 heartbeat dumps / ~7 real min), VI pacemaker still posts. **Deadlock on `s_imLock`** — the D147 recursive-mutex that backs `osSetIntMask`. Both consumer threads are blocked in `pthread_mutex_lock(&s_imLock)` and neither owns it: `mainThread` = `sndSetScalerApplyVolumeAllSfxSlot` → `sndApplyVolumeAllSfxSlot` → `sndSetSfxSlotVolume` → `alEvtqPostEvent` (`event.c:82`) → `osSetIntMask(OS_IM_NONE)` (`libultra.c:1232`); `amMain`/audioThread = `amHandleFrameMessage` → `alAudioFrame` → `sndPlayerVoiceHandler` → `alEvtqNextEvent` (`event.c:47`) → `osSetIntMask(OS_IM_NONE)`. `sndSetScalerApplyVolumeAllSfxSlot` is the **mission-failed audio fade-out** (ramps `g_sndSfxVolumeScale` → 0 each frame, `AL_SNDP_RELEASE_EVT` posted per active `ALSoundState` per frame) — so the fade-to-black *is* the intended fail transition, and it hammers the event queue hard enough that the `osSetIntMask` lock is left owned by a party not in the 3-thread dump (a leaked unbalanced `OS_IM_NONE` — libaudio has early-return mask paths — or a transient thread that acquired it and exited; `gdb.txt` shows `New Thread`/`exited` churn every few k frames). D147's "the decomp never blocks while holding OS_IM_NONE → cannot deadlock" assumption is **falsified** by this path. Same family as D147 / porting-notes.md §D4.

**Status (from §F table):** MITIGATED (`port/src/libultra.c`, M-28, `#ifdef PORT`/port-only). `osSetIntMask` is now a **self-healing** logical lock: `OS_IM_NONE`/`OS_IM_ALL` acquire/release a process-wide recursive critical section tracked by `s_imHeld`/`s_imOwner`/`s_imDepth` under a short bookkeeping mutex `s_imMx` + condvar. Normal (balanced, microsecond) audio critical sections behave exactly as the D147 recursive mutex did. A waiter that blocks **> 2 s** (`OS_IM_STUCK_NS`) concludes the holder leaked it, logs `LOG_ERROR` with the stale `s_imOwner` + the stealing caller's return address, and **steals** the section — so a leaked/unbalanced `OS_IM_NONE` (or an acquire-and-exit transient thread) can no longer wedge the game forever; worst case is a ~2 s audio hiccup + a log line naming the leak. Builds clean. Still OPEN: the exact leaking call site (find it from the `D152: ... stealing from owner=` log next time it fires) and the proper narrow fix (dedicated `ALEventQueue` lock, or restore-on-thread-exit). Regression check deferred — `-level_09`/`-level_20` both flaky under the current machine load (D117/D134), and clean `master` itself has a separate nondeterministic `-level_09` frame ~900-1400 `0xc0000005` (see below).

## D152+ (b) — D152 addendum — static audit + narrow fix (M-31).

**D152 addendum — static audit + narrow fix (M-31).** Audited every `osSetIntMask(OS_IM_NONE)` in the *compiled* audio code (`event.c` `alEvtqNextEvent`/`alEvtqPostEvent`/`alEvtqFlush`/`alEvtqFlushType`, `csplayer.c` `__CSPRepostEvent`, `synaddplayer.c` `alSynAddPlayer`, `snd.c` `sndRemoveEvents`/`sndSetupSound`/`sndDeactivateAllSfxByFlag`): **all balanced on every path** — `alEvtqPostEvent`'s lone early `return` (`event.c:86`, freeList empty) restores the mask first. The §F D152 guess "libaudio has unbalanced early-return mask paths" is **not** borne out. Two real defects found + fixed `#ifdef PORT` (N64 verbatim `#else`): **(1)** `sndSetSfxSlotVolume` (`snd.c`) walks the live `ALSoundState` list and posts `AL_SNDP_RELEASE_EVT` to the shared `ALEventQueue` **without holding `OS_IM_NONE`** — unlike its structural twin `sndDeactivateAllSfxByFlag`, which wraps its identical walk. On PC that's an unguarded walk racing `amMain` *and* one lock acquire/release per matching sound; the mission-failed fade (`sndSetScalerApplyVolumeAllSfxSlot`→`sndApplyVolumeAllSfxSlot`→ this, every frame, every active sound) turns it into a per-frame lock-acquire storm on `mainThread` = exactly the §F dump's "hammers the event queue hard enough that the lock is left owned." Fix: hold `OS_IM_NONE` once across the whole walk (nested `alEvtqPostEvent` hits the recursive depth++ fast path, no new contention window); also wrap `sndApplyVolumeAllSfxSlot`'s slot loop so the whole volume update is one recursive hold (1 real acquire/frame instead of `SFX_SLOT_COUNT × active-sounds`). **(2)** The actual leak matches the §F dump's *other* hypothesis — "a transient thread acquired `OS_IM_NONE` and exited" (`gdb.txt` shows `New Thread`/`exited` churn). `port/src/libultra.c` `portThreadWrapper` now calls new `imThreadExitRelease()` after the thread entry returns: if that thread still owns `s_imHeld`, release the orphan immediately (+ `LOG_ERROR` naming it). This removes the wedge *and* the ~2 s steal hitch, and stops the **re-wedge** that occurs when a new host thread reuses the dead thread's pthread id and `imAcquire` mis-detects recursion (`s_imOwner == self`). The M-28 steal-lock stays as the last-resort backstop for a genuinely unbalanced *same-thread* path (none found, but cheap insurance). Build green (240/240); `-level_09` frame 300 renders 91.7% non-clear, crash-free; `-level_20` boots crash-free (frame-count flaky under load per D117/D134, no `D152:` log). Fade-out repro is **playtest-gated** — not reproducible headless, so not end-to-end verified; user should replay the mission-fail and confirm no black screen + no `D152:` steal/orphan log lines.

**Status (from §F table):** FIXED (`src/snd.c`, `port/src/libultra.c`, `#ifdef PORT`). porting-notes.md §D4.

## D154 — `bg.c` room-geometry ray/hit-test GBI parser port (M-28, WRITTEN / UNVERIFIED / UNCOMMITTED).

**`bg.c` room-geometry ray/hit-test GBI parser port (M-28, WRITTEN / UNVERIFIED / UNCOMMITTED).** `bgTestRayIntersectionInRoom` (`bg.c:3331`, called by `bgTestBulletHitBackground` on every shot that resolves against background geometry) is the D135 sibling flagged since M-23: raw `((u8*)gdl)[k]` / `((u32*)gdl)[i]` byte/word indexing into what is now the 16-byte PC `Gfx` room DL (D85 `bgWidenRoomGdl` + `texLoadFromGdl`) → desync → OOB `vtxbase[idx]` / walk-off AV the first time you shoot a wall or floor. Ported `#ifdef PORT` (N64 verbatim under `#else`): `vtxoff`/`vtxbase`/`op` via the `.dma` view (matches the already-ported `bgBuildRoomVtxBounds`); G_TRI1 indices from `(w1>>16/8/0)&0xff /10`; the 12 G_TRI4 nibble extractions from `(u32)gdl->words.w0/.w1` (each mapped from the N64 byte/halfword layout — table in the code comment, cross-checked twice); both `texturenum` recoveries → `-1` (the KSEG0 `*(u16*)(w1-8)` deref is invalid for the port's converted GDLs, same as D135; texnum only picks the impact decal/sound, parked D77). Builds clean. **Verification blocked** — a no-input `GE_PCDUMP` capture never fires a bullet so it never calls this function; needs a real firefight into a wall on an idle machine (this session's box was thrashed → D153/driver crash on every run).

**Status (from §F table):** **M-30 re-audit + fixes (still playtest-gated for final verify).** Cross-checked the ported G_TRI1 + all 12 G_TRI4 nibble recoveries twice against the N64 `#else` and the raw BE-word bit derivation — **G_TRI1 and G_TRI4 are correct**. **Two bugs fixed:** (1) `vtxoff` — the port read `gdl->dma.par & 0xf`, but the PC `Gdma_le` shim (`port/shim/PR/gbi.h`) maps `.par` to bits 0-23 of word0 (the packed *length*, always 16-aligned) not the N64 byte-1 params field at bits 16-23, so vtxoff was silently forced to 0 (breaks any G_VTX batch with `v0 != 0`, i.e. multi-batch rooms → wrong `vtxbase[idx]` → OOB). Now `((u32)gdl->words.w0 >> 16) & 0xf`. (2) **Sibling ported** — `bgTestBulletHitBackground` tail (`bg.c:~3841`, the post-hit G_SETTILE back-scan for `tileformat`/`tilesize`) was itself an unported raw `((u8*)gdl)[0]`/`[1]` + 64-bit `words.w0 << 11 >> 30` parser over the same widened DL; wrapped `#ifdef PORT` (opcode from bits 24-31, byte1 from bits 16-23, `(u32)`-forced 32-bit width for the tilesize shift). **Diagnostic:** `GE_D154=1` → `osSyncPrintf` of the parsed cmd/vtx-index stream (call header + per-TRI1/TRI4 `idx`) for the first 64 invocations, for the user/integrator to eyeball. Build green (`ntsc-final`, 242/242). Runtime unverifiable in a worktree (no `data/`); still needs a firefight-into-wall on an idle machine + `-level_09` framediff. Confidence: HIGH on the mechanical correctness (bit-for-bit vs N64), MEDIUM that nothing else in the walk still bites. porting-notes.md §B (D135 corollary).

## D156 — Facility outro-cutscene "hang" — the actual fix (M-29, user bug report, 2nd occurrence after D155).

**Facility outro-cutscene "hang" — the actual fix (M-29, user bug report, 2nd occurrence after D155).** After D155 shipped the user re-ran the playthrough and it hung again at the end of Facility — same stack (`modelSetAnimFrame2WithChrStuff` ← `modelTickAnim` ← `chrUpdateAnim` ← `chrTick` ← `playerTick` ← `propsTick` ← `lvlRender` ← `bossMainloop`) but this time **`frames=12962` frozen CONSTANT across ~200 heartbeat dumps / minutes** — a true infinite loop, not the D155 slow-catch-up spiral. Root: the `while (1)` at `model.c:3131` in `modelSetAnimFrame2WithChrStuff` steps the root-motion accumulation one anim frame at a time from `framea` (`model->animframe1`) to `frameb` (`frame`, accumulated in `modelTickAnim` as `frame += playspeed * speed` per tick). If `frameb` arrives **NaN or blown up** — a cutscene anim whose blend/transition math produces a huge/non-finite `model->speed`/`playspeed`/`timespeed` (`model.c:3436` `/ model->unkb0`, `:3477` `/ model->timespeed`, both guarded `> 0` but not against tiny values) — then `floorFloatToInt(frameb)` yields garbage, `endframe - curframe` ≈ 2^31, and the loop (each iteration: `sub_GAME_7F06D3F4` joint transform + `cosf`/`sinf` + `modelConstrainOrWrapAnimFrame`) runs effectively forever. **Likely deeper cause (unconfirmed, needs the repro):** these speed/blend fields are `Model` struct members; if this model is punned/inline (D100/D140 family) and `sizeof(Model)` growth misaligned them, they read garbage exactly when a cutscene sets an unusual anim transition.

**Status (from §F table):** GUARDED (`src/game/model.c`, `#ifdef PORT`): (1) in `modelSetAnimFrame2WithChrStuff`, if `frameb` is non-finite or `|frameb| >= 1e6` snap it to `framea` (loop then does ~0 iterations; the `modelConstrainOrWrapAnimFrame` at :3211 still clamps `model->frameb` to a valid frame); (2) in `modelTickAnim` after the tick loop, if `frame`/`frame2` went non-finite/absurd fall back to the pre-loop `model->animframe1`/`animframe2` so the NaN can't poison model state every frame. `!(x > -1e6f && x < 1e6f)` catches NaN too. Anim frames are 0..few-hundred so 1e6 is a safe ceiling. Verified: build green; `-level_09` frame 200 unregressed (guards never fire in normal play). **User to re-verify the Facility outro.** If it still hangs, the NaN source is upstream — dump `model->speed/playspeed/timespeed/unkb0` in `modelTickAnim` during the cutscene. porting-notes.md §E.

## D155 — Facility outro-cutscene end → "hang" = unclamped `deltaFrames` spiral (M-29, user bug report).

**Facility outro-cutscene end → "hang" = unclamped `deltaFrames` spiral (M-29, user bug report).** User watched the Facility outro cutscene; at the end the game froze (process alive, kernel-heartbeat dumps). Symbolicated `mainThread` stack from the port's own hang dump: `modelConstrainOrWrapAnimFrame` ← `modelTickAnim` ← `chrUpdateAnim` ← `chrTick` ← `playerTick` ← `propsTick` ← `lvlRender`, sampled there across *every* heartbeat dump. Root cause: `waitForNextFrame()` (`src/game/frametiming.c`) sets `nextFrameTime = (osGetCount() − prev + 387937) / 775875` and passes it **unclamped** to `updateFrameCounters` as `deltaFrames`. `osGetCount()` is wall-clock on the port (D117), not VI-locked, so a real-time stall — the asset load at the cutscene→debrief boundary, made far worse by the user running a local LLM on the same box — makes `nextFrameTime` balloon to hundreds/thousands. That becomes `speedgraphframes` → `g_ClockTimer`, which drives `modelTickAnim(model, g_ClockTimer, 1)`'s `while (numticks-- > 0)` loop once per character per render, plus dozens of `for (i = 0; i < g_ClockTimer; i++)` sim loops (`bondhead.c`, `bondview2.c`, `explosion.c`, `options.c`…). One render then does thousands of anim ticks × N characters → multi-second frame → heartbeat trips, and the slow frame feeds an even bigger delta next time → unrecoverable spiral. The N64 was physically VI-bound and never produced `deltaFrames > ~2`. Timing-compensation class, cf. D117/D134.

**Status (from §F table):** FIXED (`src/game/frametiming.c`, `#ifdef PORT` — clamp `nextFrameTime` to `FRAMETIMING_PORT_MAX_CATCHUP` = 6). After a hitch the sim just resumes at real-time-ish pace instead of spiralling, exactly as the console did when it dropped frames under load. N64 path verbatim. Confidence: high (stack + arithmetic both point here). porting-notes.md §E.

## D153 — `-level_09` frame ~900–1400 `0xc0000005` = D117/D134 load flakiness, NOT a bug (M-28).

**`-level_09` frame ~900–1400 `0xc0000005` = D117/D134 load flakiness, NOT a bug (M-28).** Reproduced 6/6 while the machine was under heavy concurrent build+run load; on a **quiet** machine `-level_09` ran past frame 2400 with 0 heartbeats, no crash. Same false-positive class the docs warn about every session (M-13..M-22, M-25 Silo ~frame 300 freeze). The crash PC was in a system DLL (a `memcpy` off a buffer) and the frame varied run-to-run — both consistent with the port's host-scheduling jitter under load (`osGetCount` is wall-clock, variable-timestep sim), not a deterministic defect. `romdataFixupMusicSeqTable: seqCount 63 exceeds blob capacity 1` is an EXPECTED benign warning on the header-only first `romdataFixupMusicSeqTable` call (music.c:736; the full-size second call at :742 is silent) — not related. **Lesson: don't regression-test while builds/other runs are in flight.**

**Status (from §F table):** NOT A BUG — closed. Run anchors on a quiet machine.

## D164 — Disclaimer / legal screen only draws the first line (D76 root cause, M-31).

**Disclaimer / legal screen only draws the first line (D76 root cause, M-31).** `constructor_menu00_legalscreen` (`src/game/front.c:1523`) renders the 12-entry `legalpage_text_array[]` with `legal_text_end = (struct legal_screen_text *)&legalscreen_MRD;` then `do { render; ptr++; } while (ptr < legal_text_end);`. This is a **linker-adjacency assumption**: on N64 `.data` the two file-scope globals are emitted in source order so `&legalscreen_MRD == legalpage_text_array + 12` and the loop count is exactly the array length. mingw/GCC reorders them — in `build-pc/ge007.x86_64.exe` `legalscreen_MRD` links at `0x1401305e0`, **0x60 bytes before** `legalpage_text_array` at `0x140130640` (verified via `nm`). So `legal_text_end < legal_text_ptr` from the start; the `do/while` runs the body **once** (entry 0, "TWYCROSS BOARD OF GAME CLASSIFICATION" — long + CENTER_ALIGN so it wraps to ~2 visual lines = the "classification line + one below" the user saw), then exits. The other 11 legal lines never render. NOT a D68 image-table bug — the legal screen references **zero** `sImageTableEntry`s; it is pure `langGet()` font text via `display_aligned_white_text_to_screen`→`textRender`, plus the 3D `logoinst` model (`subdraw`, absent — D75 family, parked). The D68 fixup and the `GE_IMGT` avenue in the brief are both dead ends here.

**Status (from §F table):** **FIX PROPOSED, NOT APPLIED** (`front.c` is owned by another agent this session). One-line `#ifdef PORT` in `constructor_menu00_legalscreen`: `legal_text_end = legalpage_text_array + ARRAY_COUNT(legalpage_text_array);` (N64 `&legalscreen_MRD` cast under `#else`). Same linker-layout class as the D75 loop-bound / D142 enum family. Confidence: **high** — mechanism confirmed by symbol addresses.

## D159 — Front-end wallet-Bond photo (and other large front-end I4/I8/IA8 textures) render "interlaced" / combed (M-31, user bug RC1 / D149).

**Front-end wallet-Bond photo (and other large front-end I4/I8/IA8 textures) render "interlaced" / combed (M-31, user bug RC1 / D149).** The user's menu showed a garbled venetian-blind Bond photo on the folder / mode-select screen; earlier the same model DL walked into garbage (`D146` opcode spam) — that crash is already contained by the D144/D146 guards, so what's left is purely the texture. Root-caused with `GE_DTEX`: the photo is a **65×65 I4** (also seen: 65×65 I8, 95×32 IA8) with a full mip chain; `import_texture_i4` uploads it `line_size_bytes*2` (= padded stride) wide and `size_bytes/line` tall — geometry fine. The comb is **`texSwapAltRowBytes` (`src/game/image.c:2191`)**: it pre-swaps every **odd** texture row in 8-byte (one-`u32`) groups to compensate for the N64 RDP's odd-line TMEM address XOR (bit 2) during 4-byte-word `gDPLoadBlock` loads. fast3d does **not** emulate that XOR (no odd-row handling anywhere in `gfx_pc.cpp`), so the pre-swapped odd rows stay scrambled in the uploaded GL image. It's called on *every* level of *every* texture on *every* path (§4), but only bites where the texture is large enough and viewed 1:1 to see the 8-texel comb — glaring on the ~big front-end photos, ~invisible on small HUD glyphs / distant in-level surfaces (which is why it was written off as "front-end model" work for many sessions). Same class as the K0-fold / interrupt-mask hardware-quirk shims — belongs in `port/`, not game logic.

**Status (from §F table):** **FIXED** (`src/game/image.c`, `#ifdef PORT` early-`return` in `texSwapAltRowBytes` — fast3d wants a plain linear image; N64 body kept verbatim under it). Verified: wallet-Bond photo renders as a clean recognisable grayscale portrait (still 180°-rotated — D75 model-transform family, parked); `-level_09` frame 200 coverage 91.62 vs 91.66 golden (D117 noise) and a visual frame is clean (tiled floor / monitors / world-map screen all correct). Confidence: **high** (GE_DTEX + before/after screenshots). The residual over-tall-mip (RC2) and non-PoT wrap (RC3) issues are separate and still parked; this only removes the row-swap. porting-notes.md §C.

## D165 — Front-end menu cursor doesn't feel like a mouse pointer (M-31, user bug B5 follow-up).

**Front-end menu cursor doesn't feel like a mouse pointer (M-31, user bug B5 follow-up).** M-30 added a "menu pointer mode" that fed mouse *velocity* into the stick axes, but `front.c frontUpdateControlStickPosition` *integrates* the stick as a velocity into a screen-pixel cursor position (`cursor_h_pos += (stickx*0.075 ± 0.5) * g_GlobalTimerDelta`, ±5 deadzone, ±70 clamp, cursor clamped into the ~320×240 virtual rect − 20px margin). Velocity-in → velocity² cursor. **Fix (port-only, no `front.c` edit):** `input.c` now runs a P-controller — it keeps its own estimate `menuEst{H,V}` of where the game cursor is, integrated each poll with the *exact same recurrence* as `frontUpdateControlStickPosition` (`menuCursorStep()`), accumulates a `menuTgt{H,V}` target from raw mouse motion (clamped to the same rect), and emits `stick = clamp(MENU_P_GAIN·(target−est), ±70)`. `MENU_P_GAIN = 6.0` → per-poll error decay ≈ 0.45 (`0.075·6`), so no overshoot and ~8-poll settle; stick falls under the game's ±5 deadzone within ~1px of target so the cursor parks with zero jitter/drift. The estimate re-syncs to the true cursor whenever the target is held at a screen edge (both clamp). `Input.MenuPointerSpeed` scales the mouse→target mapping (bump toward 150–200 if the sweep feels short — `est` steps once per controller poll ≈ 2×/rendered-frame vs `front.c`'s 1×/frame, a rate mismatch absorbed by this knob). `Input.MenuPointerMode` (0 = old velocity, 1 = pointer, **default 1**) is the escape hatch. `GE_INPUTLOG` prints an `menuptr est=/tgt=/eff=/stick=` trace line for tuning. With no mouse motion `target==est==centre` → `eff=0` → emitted stick unchanged → prior behaviour exactly.

**Status (from §F table):** FIXED (`port/src/input.c`; `Input.MenuPointerMode`, `Input.HipfirePitchSpeed` added). Build green, `-level_09` + `GE_STARTMENU=6` smoke crash-free. Feel-check owed (headless can't exercise the mouse).

## D166 — Hipfire mouse pitch is digital while yaw is analog — inconsistent feel (user bug B7 / D118a, "cheaper alternative").

**Hipfire mouse pitch is digital while yaw is analog — inconsistent feel (user bug B7 / D118a, "cheaper alternative").** In hipfire GE takes yaw from the analog stick-X but pitch only from digital C-up/C-down, so `input.c` emitted one C-button per poll once `|mouseΔY|` crossed a fixed threshold — an on/off step. **Fix (port-only):** emit the C-button as a *duty-cycled pulse* whose firing frequency scales with mouse-Y speed — `duty = |dyLook|·(HipfirePitchSpeed/100)/HIP_PITCH_FULL` (clamped ≤1), phase-accumulate, fire when the phase wraps. Fast mouse → `duty≈1` → C-button every poll (solid hold, as before); slow mouse → sparse taps proportional to speed. `HIP_PITCH_FULL = 6.0` px/poll for a solid hold; `Input.HipfirePitchSpeed` (10–500, default 100) tunes it. Aim-mode (RMB, already fully analog) untouched. No mouse motion → no pulse → prior behaviour.

**Status (from §F table):** FIXED (`port/src/input.c`). SMALL-FIXES B7. **M-87: likely SUPERSEDABLE — see the "SOLITAIRE natural-pitch" note below.** This pulse hack was necessary because the port forces a control scheme (HONEY, the 1.1 default) whose hipfire pitch is digital-only. GE's own **SOLITAIRE (1.2) / GOODNIGHT (1.4)** schemes don't have that constraint at all (see D194) — under those, the *entire* primary analog stick is free for continuous, unclamped pitch+yaw ("natural pitch/turn"), because movement is redirected onto digital C-buttons instead. If the port switches its forced `controldef` to SOLITAIRE/GOODNIGHT (also the right move for D223's wheel-cycling design and D194's mouse-feel ask), this whole digital-pulse emulation becomes unnecessary — mouse-Y would drive the stick's Y axis directly and continuously, same as X already does, with **zero digital-button emulation for pitch at all**. Don't remove D166's code without first confirming the scheme switch; keep as a fallback for whichever legacy path still needs it.

## D169 — Front-end mouse pointer can't reach the outer cells of the mission/level-select grid — only an inner ~3×3 selectable (M-33, developer bug report).

**Front-end mouse pointer can't reach the outer cells of the mission/level-select grid — only an inner ~3×3 selectable (M-33, developer bug report).** The D165 pointer P-controller in `port/src/input.c` is hard-coded to a **320×240** virtual field (`MENU_CURSOR_HI_H`/`_HI_V` = 300/220, `MENU_CURSOR_MID_H`/`_MID_V` = 160/120, `input.c:145-149`), but GE's front end runs at **440×330** (`front.c:8570` `viSetViewSize(440,330)`; default cursor home 220/165 `front.c:285`). `frontUpdateControlStickPosition` clamps the real cursor to `[20,420]×[20,310]` (`front.c:1195-1217`); the mission-select hit-test grid spans x 73…352 / y 62…270 with outer split points 317 / 235.5 (`front.c:516,519,3180,3194`). The port clamps its pointer *target* and *estimate* to 300/220 (`input.c:553-564`) → the stick zeroes out once both saturate → the game's `cursor_h_pos`/`cursor_v_pos` park at ≈300/220, short of the two outer columns and the bottom row. File-select / mode-select / main-menu are unaffected because their hit targets (centred folder boxes, the `x=126` mode list, the `x=106` difficulty list) all sit inside the 320×240 sub-box. Not the D118d `joyGetStickY`-threshold class — this is a virtual-resolution constant mismatch, closer to D164 (a front-end layout constant wrong on the PC path). Also note: the "re-syncs whenever the target is held at a screen edge" comment at `input.c:141-143` describes behaviour the code doesn't implement — the only estimator reset is the activation re-home (`input.c:545-549`), so once pinned at the clamp the estimate never recovers. Confidence: **medium-high** (constants + clamp math unambiguous in source; static-only, no build/run this session; exact reachable block "~4×3" vs the reported "3×3" within tolerance).

**Status (from §F table):** **FIXED (M-33, `port/src/input.c`, port-only, no `#ifdef PORT`).** The `menuMode` pointer branch now derives its clamp bounds from the live virtual screen — `[screenleft+20, screenleft+screenwidth-20] × [screentop+20, screentop+screenheight-20]` via `getPlayer_c_screenwidth/height/left/top()` (`src/game/bondview.c:880-895`) — falling back to the old 320×240 constants when the front-end screen isn't set (`sw` outside 200…2000). The estimator seed on activation is now the real `cursor_h_pos`/`cursor_v_pos` (`front.c:285`, externed in `front.h`) instead of the 160/120 centre guess. `MENU_CURSOR_HI_H/_V` / `MID_H/_V` kept only as the fallback. **Verified:** builds + links clean (`getPlayer_c_screen*` and `cursor_[hv]_pos` all resolve — non-static engine symbols, as expected); `GE_STARTMENU=7` mission-select boots crash-free 600+ frames; `-level_09` unregressed (framediff 3/3 within threshold, 91.6% nonclear). The in-level path is untouched (menu-only branch). **Interactive feel-check still owed** — headless input can't drive the mouse pointer, so "every grid tile is now reachable" is inferred from the corrected clamp math, not observed. `Input.MenuPointerMode=0` (legacy velocity) and the constant fallback remain as escape hatches.

## D178 — Pre-mission briefing screen: objectives blank / missing, briefing pages blank (also the D143 "briefing text blank" side effect) — FIXED (M-36).

**Pre-mission briefing screen: objectives blank / missing, briefing pages blank (also the D143 "briefing text blank" side effect) — FIXED (M-36).** Root cause: the briefing segment (`Ubrief*Z`, source `assets/obseg/brief/*.c`) is a raw ROM image of `struct BriefStruct` = `{ u16 brief[4]; struct { u16 textid; u16 enabled_difficulty; } objective[10]; }` (48 bytes), loaded by `front.c load_briefing_text_for_stage()` via `_fileNameLoadToAddr()` — **no converter and no BE→LE fixup anywhere in the chain**. On the LE host all 24 `u16` read byte-swapped. Measured on Dam (`GE_D178=1`): raw `brief=002c,012c,022c,032c obj0=042c/0100 obj3=072c/0000` vs correct `2c00,2c01,2c02,2c03 / 2c04/0001 / 2c07/0000`. Two independent symptoms follow: (a) `textid` `0x2C04` (= `getStringID(LDAM,4)`, bank 11 slot 4) reads as `0x042C` → bank 1 slot 44, never loaded → `langGet()` NULL → **blank text** (this is exactly the D143 residual, `front.c:6732` and `brief[0..3]` at `front.c:6855-6867`); (b) `enabled_difficulty` `0x0001` (Secret Agent) reads as `0x0100 = 256`, so `selected_difficulty >= enabled_difficulty` (`front.c:6729`) is false for every difficulty-gated objective — on Agent only the one `DIFFICULTY_AGENT`(0) objective survived the filter, which is why a single bare "a." bullet printed. **Fix:** `romdataFixupBriefing()` in `port/src/romdata.c` (+ `port/include/romdata.h`), called from a `#ifdef PORT` block in `load_briefing_text_for_stage()` right after the load — the same shape as `langFixupLoadedBank()` in `language.c` (BE-serialized-struct reconciliation, semantics-preserving, no game logic touched). `GE_D178=1` prints the raw and fixed words. **Verified:** Dam briefing on Agent now shows "a. Bungee jump from platform"; on `GE_STARTMENU_DIFF=3` (00 Agent) all four (Neutralize all alarms / Install covert modem / Intercept data backup / Bungee jump from platform), correct difficulty gating. `-level_09` framediff 3/3; `GE_STARTMENU=7`/`=13` crash-free.

**Status (from §F table):** **FIXED (M-36)** — high confidence

## D175 — In-game stutter / brief hang during normal play

**In-game stutter / brief hang during normal play** (user QA report): opening a door on Runway (`-level_35`, mission 3) and also observed on Surface. Self-recovered; no backtrace captured. Likely one of the known transient-hang classes (D155 catch-up spiral / D156 anim NaN loop / D134 task-done event / D147-D152 audio-lock steal) or a benign new-room texture-import frame spike on door open. See `GRAPHICS-BACKLOG.md` D175 for the gdb pattern-match sheet.

**Status (from §F table):** OPEN — observed, not investigated

## D172 — Bullet-impact / blood / spark particles render magenta or cyan instead of dark red (M-82, ROOT CAUSE FOUND + FIXED).

**Bullet-impact / blood / spark particles render magenta or cyan instead of dark red (M-82, ROOT CAUSE FOUND + FIXED).** Particle billboards (`explosion.c` `explosionRenderPart`, `gSPVertex`+`gSP2Triangles`) take their CC/tile state from ROM state records replayed just before them (`g_ExplosionDisplayLists[]` = `assets/oddtextures.c` `globalDL_0x078..`). The dominant record sets `gsDPSetCycleType(G_CYC_2CYCLE)` + `gsDPSetTextureLOD(G_TL_TILE)` + `gsDPSetCombineMode(G_CC_INTERFERENCE, G_CC_MODULATEIA2)` and binds **two** tiles: tile 0 = IA8 "smoke" @ TMEM 0, tile 1 = RGBA16 "fire" @ TMEM 0x188. `G_CC_INTERFERENCE` cycle-0 = `TEXEL0 * TEXEL1`. **fast3d's `gfx_lod_tile_offset()` returns `0` unconditionally** on the `!gfx_detail_textures_enabled` branch — a D107 fix for GE's mip textures (whole chain loaded in one LOADBLOCK at TMEM 0; `port/src/video.c:225` sets the flag `false`). So texunit 1's sampled tile = `first_tile_index + 0` = tile 0 → **TEXEL1 samples the smoke texture, not fire** → `smoke * smoke` → wrong colour. Two earlier M-82 candidates were false trails: (1) "0xB9 unhandled → D146 abort" — 0xB9 is `G_SETOTHERMODE_L` in this non-F3DEX2 build, handled; zero D146 aborts in a sustained-fire run. (2) "records never set 2-cycle" — they do (`0xBA`), and fast3d applies it (probe-confirmed `cycletype=2CYC`). **Probe evidence (`GE_D172=1`, user Silo firefight, `d172_silo.log`):** 40/40 particle tris → `cfg1[tile=1 tmem=392 fmt=RGBA siz=16b]` but `SAMPLED1 = tile0 (tmem=0 fmt=IA)`, `tex_lod=0`.

**Status (from §F table):** **FIXED (`port/fast3d/gfx_pc.cpp` `gfx_lod_tile_offset`): `return rdp.tex_lod ? 0 : i;`** — fold to the base tile only when LOD is actually active. D107's mip case (`tex_lod=1`, blurry ceilings) unchanged; a genuine non-LOD 2-texture combine now samples tile `i`. `bunker1` verify PASS (the D107 repro level), no framediff regression. PR #34. **Owed: in-game eyeball — blood/sparks should read dark red.** Env-gated `GE_D172` probes left in tree (`#ifdef PORT`, inert).

## D174 — "No blood effect" (user QA report).

**"No blood effect" (user QA report).** Likely the unverified D120 decal fix: `d43_emit.py` now emits the opcode-0x18 `PointUsage[]` chain, but it was never interactively verified and requires a full sidecar regen to take effect (`debug.ps1` does not regen). Spray path is D172 (draws, wrong colour) — total absence would be new. See `GRAPHICS-BACKLOG.md` D174.

**Status (from §F table):** OPEN — observed; first step = BUNKER1 firefight with regenerated sidecars

## D173 — Third-person Bond model spawns too high

**Third-person Bond model spawns too high** (user QA report): the player-representing figure at level start floats well above the ground "a lot of the time" while the actual player spawn is correct; same on the Cuba end credits — JB ~6 ft in the air above Natalya. D75 animated-model family but a *position* defect, not absence (M-33: gun-barrel Bond renders fine). Suspects: intro/credits puppet chr spawn position vs model base transform; anim root-motion accumulation (`modelSetAnimFrame2WithChrStuff`, cf. D156); packed-float `coord3d` decode in intro setup. "Player spawning OK" localises it to the render-side model, not `g_CurrentPlayer`. See `GRAPHICS-BACKLOG.md` D173.

**Status (from §F table):** OPEN — observed, not investigated. **M-87 user playtest: still present** — "Bond floating in the sky / cutscene glitch thing still exists." Related, newly reported: **in some cutscenes Bond spins in circles** (named instance: end of Cradle, when Trevelyan falls) — likely the same actor-position/anim-root class rather than a new bug; logged under D148/D160 (`GRAPHICS-BACKLOG.md`) since it's cutscene-triggered. **M-106 investigation (static only — no build/runtime access in this worktree, see note below):** checked the two most D209-shaped candidates for a puppet Y-position aliasing bug and RULED OUT both. (1) `chrSpawnAtPad`/`chrSpawnAtCoord`/`chrAdjustPosForSpawn` (`src/game/chraction.c:10653-10794`) — the puppet-spawn path a cutscene `guard_play_animation`-style AI cmd would use. `PadRecord.stan` (`src/bondtypes.h:1744-1751`) is a ROM-serialized *pointer* field (`StandTile *stan` at offset 0x28) — exactly the tail-pointer shape D122/D126/D132 warn about — and `tools_pc/d88_emit.py`'s `emit_pad()` zeroes it on conversion with a comment claiming it's "dead-on-load, always set before use." **Verified this claim is TRUE, not a converter bug**: every `PadRecord` literal in the N64 source asset tables (checked `assets/obseg/setup/UsetupdamZ.c:37-51`) initialises `stan` to the literal `0` — the ROM binary never carries a live pointer there on N64 either, so zeroing it on PC is byte-identical behaviour, not a regression. `chrAdjustPosForSpawn` receiving a NULL seed `StandTile*` is therefore the original game's normal case, not a port defect. (2) The anim-root-motion Y-accumulation chain (`modelTickAnim`→`modelSetAnimFrame2WithChrStuff`, `model.c:3060+`, `pos24 = pos34 + rotate(pos)` at `model.c:3322`) was already exhaustively audited for exactly this class of bug during D193/D209 (M-80/M-81) and found self-consistent, with the `padding[]`-union sweep that found D209/D210 explicitly noting `act_anim`/`act_dead` (the union arms active during a played-animation/dead puppet) are "layout-stable" (`porting-notes.md` §A, D210 entry) — i.e. that sweep already covers the cutscene-puppet actiontype, and it came back clean. So the two most probable D209-style aliasing sites for this specific symptom are both cold. **Did not find the actual root cause.** Did not get to: the *separate* Cuba end-credits case (`front.c` `init_menu18_displaycast`/`constructor_menu18_displaycast`, `front.c:7821-8320`+ — a **different**, single-actor "cast roster" display model shared by the intro roster and the credits roster, using `getsuboffset`/`subcalcpos`/`cast_rootpos_smoothed` root-motion smoothing, not the in-level chr path at all) — spot-checked `getsuboffset` (`model.c:684-700`) and it is `#MATCH`-verified clean, but did not trace the smoothing math or confirm this is even the code path the "JB above Natalya" report hits (that report implies two actors on screen at once, which this single-`cast_model` code doesn't obviously support — may be a distinct, not-yet-located two-actor cutscene-puppet path). **Confidence: LOW-MEDIUM that this is the D209-style struct-aliasing family at all** — the two most likely aliasing sites check out clean, which argues *against* the shared-root-cause theory for D173 specifically (unlike D160, where the analogous check also came back clean but for a different reason — no pointer fields to alias in the first place). Next investigator: get a live Cradle (`-level_41`) capture of Trevelyan's chr during the death sequence with a position-logging probe (none shipped yet — would need a new `#ifdef PORT` env-gated trace, not added this session per the no-build-verification constraint), and separately locate the actual two-actor Cuba credits code path (grep for wherever `intro_char_table`/cast display is invoked with two simultaneous models, if it exists — not found this session). **M-141: no new lead this session** (see D243's M-141 entry) — checked whether D243's stale-memory front-end/menu bug could share a mechanism with this puppet-position bug; no code-level link found, kept as an unconfirmed hypothesis to watch for during the next live cutscene capture, not a new suspect to chase on its own. **M-141: Steam Deck playtest confirmation** — user reports Bond floating at level start on both Dam and Facility ("chemical plant"), "but fine when it moves to first person" (matches this row's own framing exactly: player spawn/first-person is correct, only the third-person render-side puppet floats). No new mechanism; corroborates the bug is general across levels, not Dam-specific. **User confirms D292** (Bond sunk into the floor on death) **is the same underlying bug**, a third instance alongside level-start floating and the Cradle cutscene spin — see D292 for the consolidated cross-reference.

## D179 — Packaged build crashes after the logos — the D43/D69 model + bg sidecars are ROM-derived and absent from any build that has no ROM (M-34, alpha-release QA).

**Packaged build crashes after the logos — the D43/D69 model + bg sidecars are ROM-derived and absent from any build that has no ROM (M-34, alpha-release QA).** The first `goldeneye-pc-port-*-win64.zip` from CI (bundle-win.sh) shows the Rare/Nintendo logos (compiled-in `assets/rarewarelogo.c`) then AVs at `0x6b157a88`/`0x70157a88` — symbolised: `modelPromoteNodeOffsetsToPointers` (`model.c`) ← `load_object_fill_header` (`objecthandler_2.c`) ← first prop/item model load. Root cause is **not** a code regression: the port loads PC-layout model geometry from `data/pcmodels-<region>/{pcmodels.bin,manifest.csv}` and stage bg/stan from `data/pccg-<region>/{pccg.bin,manifest.csv}`, both produced offline by `tools_pc/d43_emit.py` / `d69_emit.py` from the ROM (see `port/src/pcmodels.c` / `pccg.c`; `pcmodelsReserveSize` logs *"pcmodels.bin not found — model loads will fail"* and returns 0). CI has no ROM so it never runs the emit scripts, and the two `data/` dirs are (correctly) gitignored ROM-derived game data — **cannot be committed or shipped** (converted Nintendo/Rare geometry, DLs, collision, stan nav; = distributing assets). A local build with those dirs present runs clean (verified M-34: 1741 frames, `romdataInit ... mapped at 0x10000000`, `pcmodels: 512 sidecars`, `pccg: 73 sidecars`, no crash — so `0x10000000` is *not* the problem here). The `docs/building.md` sidecar-gen step was also missing entirely. **Fix = release-side, backlogged** (`docs/BACKLOG.md` → "Alpha release"): bundle the emit scripts + their committed inputs as a pure-stdlib-Python asset-prep tool the user runs once against their own ROM (~5 MB output, no MIPS toolchain). Secondary/latent: the fixed-address `VirtualAlloc((LPVOID)0x10000000)` in `romdata.c` has no working fallback (`"using heap copy — direct ROM reads will fail"` then limps into the same crash) — didn't bite M-34 but will on a machine where something occupies `0x10000000`; harden separately (reserve earliest in `main`, or retry low bases and derive all segment math from the base obtained).

**Status (from §F table):** Diagnosed, not fixed — release packaging gap + latent `romdata.c` fallback

## D180 — Native-PC input pass (M-XX QoL run, `qol/native-pc-input-menu`, port-only).

**Native-PC input pass (M-XX QoL run, `qol/native-pc-input-menu`, port-only).** Three parts, all config-gated with default = prior behaviour: **(WI-1) `Input.MouseCaptureMode`** (0 = legacy always-grab; 1 = Quake-style click-to-lock) in `port/src/input.c` + `port/src/video.c` — in mode 1 the OS cursor is free until a click lands in the game window (`SDL_MOUSEBUTTONDOWN` → `inputNotifyClick`), and ESC / focus-loss / entering any front-end menu (`current_menu != MENU_RUN_STAGE`) frees it again; re-entering a stage while still "armed" re-locks so unpausing needs no click. `reconcileGrab()` runs once per controller-0 poll. Mouse buttons are suppressed from the game while the cursor is free in-stage (no phantom fire). Controller paths untouched. **(WI-2) absolute-cursor menu tracking** — when capture mode is on and the cursor is free in a menu, the D165/D169 pointer P-controller takes its target from the real cursor's absolute window position mapped onto the live virtual front-end rect (`getPlayer_c_screen*`), giving true 1:1 tracking instead of the relative-delta estimator. Relative path unchanged for legacy/grabbed. **(B3) `Input.MouseAimSpeed` default 25 → 16** (aim mode still overshot at 25 per the backlog).

**Status (from §F table):** LANDED, port-only, no `#ifdef PORT`. Build 241/241, `-level_09` framediff 3/3 within threshold (unregressed), `GE_STARTMENU` menu boot crash-free. **Feel-checks owed** (headless can't drive the mouse): click-to-lock ergonomics, absolute menu tracking, the new aim-speed default.

## D181 — `Game.ScreenShakeIntensity` — first route-(b) `src/` gameplay-cosmetic hook (M-XX QoL run).

**`Game.ScreenShakeIntensity` — first route-(b) `src/` gameplay-cosmetic hook (M-XX QoL run).** `src/fr.c viShake()` gains a single `#ifdef PORT` line — `intensity *= portScreenShakeScale;` (extern `f32`, defined + `configRegisterFloat`'d 0.0–10.0 in `port/src/video.c`) — before the existing clamp. Default `1.0f` ⇒ exact no-op, every headless golden dump byte-unaffected; `0.0` disables explosion/effect screen shake, up to `10.0` exaggerates it. N64 build (no `-DPORT`) keeps the original line verbatim. Precedent per `docs/BACKLOG.md` "Fun features → screen-shake intensity slider" and AGENTS.md #2's documented-exception path. Same policy class as a future FOV hook.

**Status (from §F table):** **Menu row pulled (v0.2.1, user testing).** The hook is correct on its own narrow terms — `viShake()` is called only from `explosion.c`, so the slider does exactly what it says (scales explosion/effect shake) — but that's a much narrower scope than the "Screen shake" label implies to a player: it does not touch the always-on walking head-bob or any getting-shot reaction, so a user swinging it from 0 to max during ordinary play sees nothing and reasonably calls it broken. Removed from `port/src/optionsoverlay.c`'s row table until it covers all screen-shake/view-bob sources, not just explosions; config var (`portScreenShakeScale`) and the `fr.c` hook are untouched. Build green.

## D187 — HUD ammo-type/weapon icon vanished at `-O2` (M-46 root-cause, M-47 fix + verify).

**HUD ammo-type/weapon icon vanished at `-O2` (M-46 root-cause, M-47 fix + verify).** `set_rgba_redirect_generate_microcode()` (`src/game/gunfire.c:5929`, sole emitter of the bottom-right magazine icon; call sites 6057/6123/6217/6219, all `gdl = f(...)`) is `Gfx *` with **no `return`** — the decomp banner above it flags exactly this. Harmless under `-Og` (GCC left the tail-call result in `$v0`); commit `34885535` (Sep 2) switched release to `-O2` → undefined return reg → callers resume the DL from the icon's start offset and overwrite its `texSelect`/`display_image` commands with the following `gunDrawHudInteger` output. Ammo *numbers* use a separate correctly-returning path — hence the exact symptom split. Fix = one line, `return` under `#ifdef AVOID_UB` (already `-DAVOID_UB=1`, `CMakeLists.txt:207`); N64 byte-match path unchanged under `#else`. **Verified M-47 (interactive console):** `-O2` build, `-level_09` `GE_PCDUMP` 200–440 — orange clip glyph absent pre-fix, present post-fix; `framediff.py` frame 320 worst-cell dmean **21.8 → 4.0**, frame 440 **21.7 → 4.6** (converges to the pre-regression golden); other levels unaffected (change is one HUD function). Do **not** revert `34885535`.

**Status (from §F table):** FIXED (`src/game/gunfire.c`, `#ifdef AVOID_UB`); PR #19

## D188 — Linux `-O2` boot crash: `_Printf` passes `&args` of a by-value `va_list` param (M-48).

**Linux `-O2` boot crash: `_Printf` passes `&args` of a by-value `va_list` param (M-48).** On the x86-64 System V ABI `va_list` is an array type → a by-value `va_list` parameter decays to a pointer, so `_Printf`'s `_Putfld(&x, &args, …)` handed `_Putfld` the address of its own local pointer slot, not the arg list; `_Putfld`'s `va_arg(*args, …)` then walked garbage. Latent at `-Og`, SIGSEGV at `-O2` (PR #9) — crash path `bossInitMainthreadData` (`boss.c:247`, D121 per-level `-m` sprintf) → `sprintf("-level_%c%c %s", …)` → `_Putfld` @ `xprintf.c`. Same UB class as D187. Windows x64 `va_list` is a scalar so `&args` worked there (Windows build unaffected — `-level_09` framediff 3/3). Fix: `#ifdef PORT` — `_Printf` takes `va_list *`, threads it straight to `_Putfld`; `src/sprintf.c` (sole PC caller; `rmon.c` excluded) passes `&args` of a real object. N64 build untouched (`#ifdef PORT`; file byte-matches ROM).

**Status (from §F table):** FIXED (`src/libultrare/libc/xprintf.c`, `src/sprintf.c`, `#ifdef PORT`); PR #22

## D189 — Linux `-fstack-protector` aborts on the decomp's latent stack-buffer overruns (M-48) — FIXED.

**Linux `-fstack-protector` aborts on the decomp's latent stack-buffer overruns (M-48) — FIXED.** Two hit on `-level_09`: (1) `sub_GAME_7F0B1DDC` (`stan.c`) — local `StandTile *tileStack[39]`, bail check `if (cat >= 41)` runs only once per outer iteration so `cat` reaches `40 + pointCount` and the append writes past the end; BUNKER1 prop room-list setup legitimately visits ~40 tiles (confirmed in GDB: 30 valid distinct tile ptrs and climbing; `standTileStart=0x701ae160`, so **not** a pointer-truncation bug — the earlier "blob >4 GiB" triage premise was wrong). (2) `bondviewCalcIntroSwirlCamera` (`bondview2.c`) — `f32 pointbuf[10]` written `[-3 .. 11]` via `&pointbuf[i*3]`, i∈[-1..2]. Both harmless on N64 + the MinGW build (no stack protector); only Ubuntu gcc's `-fstack-protector-strong` default makes them fatal `__stack_chk_fail`. **Fix:** `-fno-stack-protector` globally (`CMakeLists.txt`) to match the N64/MinGW builds, + `tileStack[64]` under `#ifdef PORT` for the one cleanly-bounded case, + `stan.c:2225` link-resolve moved to the D177 `PORT_PTRADD` macro for consistency. A *hardened* Linux build would need the whole overrun class swept. **Verified:** Windows `-level_09` framediff 3/3; Linux `-O2` `-level_09` runs to frame 2100+ no crash. porting-notes.md §C.

**Status (from §F table):** FIXED (`CMakeLists.txt`, `src/game/stan.c` `#ifdef PORT`); PR #23

## D190 — Linux `-level_45` SIGSEGV at load in `add_ptr_to_objective` (`objective.c:64`) (M-48, open).

**Linux `-level_45` SIGSEGV at load in `add_ptr_to_objective` (`objective.c:64`) (M-48, open).** `objective_ptrs[objective->menu] = objective;` faults — `objective` wild or `objective->menu` a garbage index → OOB write. Almost certainly the D122/D126/D132 propDef-stride / ROM-serialized-struct-widening family (`struct objective_entry` / the type-23/type-1E setup-text record chain read at the wrong stride on PC). Found in the M-48 Linux level sweep: **5/6 spot-checked levels (`-level_09/20/27/34/37`) load + render fine; `-level_45` is the one failure.** Not a v0.1.0-alpha blocker (alpha is explicitly "expect crashes"). Windows: `-level_45` load status not re-checked this session.

**Status (from §F table):** **CLOSED — not a port bug; level 45 is MP-only and has no solo setup file in the ROM (root-caused M-105, reproduced on Windows).** Full chain: `setup_text_pointers[45]` = `"UsetupimpZ"` (`chraidata.c:824`), but that name is **not in the ROM file table** (`assets/obseg/file_resource_table.inc.c`) — only the MP variant `"Ump_setupimpZ"` is. BASEMENT is a 4-player MP level (`front.c:557` `multi_stage_setups[]`, alongside STACK/TEMPLE/COMPLEX/CAVES/LIBRARY — exactly the other missing solo names); MP play prepends `mp_` to the name (`prop.c:1261-1269`), so the solo name was never shipped. A solo `-level_45` boot (debug flag only; no solo mission references stage 45) requests a nonexistent file: on PC the inert rmon stubs (`osReadHost`/`osWriteHost`, `n64stubs.c`) leave the CheckFileExists response in `fileGetIndex`'s dynamic path (`ob.c:421`) **uninitialized** → fresh table entry with size 0 → `fileIndexLoadToBank` allocates bank memory, loads 0 bytes → the setup rebase + propDef walk in `proplvreset2` runs over **stale pool memory** → garbage records (M-105 Windows crash: `modelLoad` via `domakedefaultobj`, `obj=0x784f`; M-48 Linux hit `add_ptr_to_objective` — same garbage walk, different stale contents). N64 parity: the file is genuinely absent there too (indy host answers not-found → index-0 load) — this boot is broken in the original as well. No fix taken: a port-layer guard cannot synthesize a valid setup (the walks need `-1`/`PROPDEF_END` terminators no zeroed buffer provides), and fabricating ROM content for an out-of-scope MP level has no ground truth. The M-48 "Linux-specific" framing was wrong — the failure is platform-independent; it was merely found on Linux first.

## D191 — Windows `-O2` release-bundle crash: SIGSEGV in `modelGetNodeRwData` (`src/game/model.c:478`) — the `root->Parent` node-tree walk (M-49, user playtests of the…

**Windows `-O2` release-bundle crash: SIGSEGV in `modelGetNodeRwData` (`src/game/model.c:478`) — the `root->Parent` node-tree walk (M-49, user playtests of the v0.1.0 win64 bundle).** **Two instances, identical fault PC `0x1400801b1`:** (1) fresh save, killed the first guard right after escaping the cell in Bunker ii; (2) Statue (`-level_49`?), right after the Trevelyan meeting cutscene. `ge007.crash.log` (auto-dump; symbols intact in the shipped exe, `addr2line` resolves): PC = `modelGetNodeRwData` at `model.c:478` — the `switch (root->Opcode & 0xff)`, i.e. `root` (a `ModelNode *`) is bad; `addr2line` puts it in the inlined self-recursion at `:532`, so the `while (root->Parent) { root = root->Parent; … }` walk (`model.c:527-540`) handed a garbage node into the recursive call. Instance 1: `Rdx = 0x70267bc000000000` — a `~0x70266xxx` heap address in the **high 32 bits, zero low half** = 32-bit pointer field read as 64-bit (or a byte-swapped `u32` pair). Instance 2: `Rdx = 0`, FAULT ADDR `0x0` — `root` (or `root->Parent`) resolved to NULL. Both: `Rcx=Rdi` = the `~0x7008xxxx`/`~0x7007xxxx` object ptr, `R8 = 0x1d` (29). So the `Parent` chain in some model contains a **truncated or NULL link** — pointer-width / byte-swap family (D3x, D119 `weapons_held[]->chr` type-pun, D122/D126/D132 propDef-stride). NOT weapon-fire-specific — common factor is `modelGetNodeRwData` running on a character/cutscene model whose node tree has a bad `Parent`. Crash logs saved: `scratchpad/crashlogs/`. **Repro aid:** shipped exe has full symbols; `debug-crash.ps1` in the bundle dir runs it under gdb, `GE_D51=1` → `d52rw.log` rwdata-pool trace. The crash.c EBP backtrace breaks immediately (`#01: 0x1d`) — need a real `bt` + `p *root` + `p *Objinst` at the fault. Not yet reproduced under a debugger; no fix.

**Status (from §F table):** FIXED + PLAYTEST-VERIFIED (M-79). User replayed Bunker ii (first-guard kill) and Statue (post-Trevelyan cutscene) on the fix build — **both no longer crash** → 21/21 solo levels crash-free (v0.2.0 gate met; Cradle's remaining issue is D193, not a crash). Build clean; bunker1 golden framediff 3/3. **Real gdb backtrace (M-79):** the fault is NOT in the `root->Parent` walk — `root` arrives *already* garbage (`0x70267a9000000000` = a half-word-swapped `u32` `0x70267a90`) from frame #1 `bondviewSelectCuff(model=0x70076808, header=0x700763a0, switchindex=29)`, called from `gunUpdateAndFire(GUNRIGHT)` on the frame the player fires. `bondviewSelectCuff` byte-indexes `header->Switches` (a `ModelNode *` array — **8-byte stride on PC**, widened per D43/D45) with `offset = switchindex << 2` (N64 4-byte stride), so every `base[N]` deref lands mid-slot and yields a misaligned half-pointer that is non-NULL by luck → `modelGetNodeRwData(root=garbage)`. **Exactly the D140 bug class** — D140 fixed the identical pattern in `gunfire.c:1740`; this call site was missed. Fix: PORT-only `offset = switchindex * (s32) sizeof(ModelNode *)`; the `base = (u8*)switches + offset` math and the `#else` N64 line are unchanged. ABI/pointer-width only (D3x class), no game-logic change. The Statue instance (`Rdx=0`, NULL fault addr) is the same site landing on a slot whose bytes read as 0.

## D193 — NPC AI locomotion is too slow — characters don't move to their destinations at the original game's rate (M-49, user playtest; clarified after the FULL 21-level…

**NPC AI locomotion is too slow — characters don't move to their destinations at the original game's rate (M-49, user playtest; clarified after the FULL 21-level playthrough — "the biggest overall flaw").** Not (primarily) an animation-playback bug — the *travel speed* of scripted/AI movement is wrong: guards, named NPCs (Trevelyan, Ourumov) and friendlies (Natalya) all take noticeably longer to run/walk to their goal positions than on N64. General, not chr-specific. **Leading hypothesis — the D117/D134/D155/D156 wall-clock-timing family:** `chr.c` / `chrai.c` integrate movement as `pos += speed * g_GlobalTimerDelta` and step sim loops `while (i < g_ClockTimer)` / `-= g_ClockTimer` (`chr.c:1437,1446,1491-1502,1862,2359,2569`; `chrai.c:76`). On the port `osGetCount()` is free-running wall-clock (D117), `g_GlobalTimerDelta`/`g_ClockTimer` are derived from it, and D155's `FRAMETIMING_PORT_MAX_CATCHUP = 6` clamps `nextFrameTime`. If the effective per-second sum of `g_GlobalTimerDelta` (or the `g_ClockTimer` tick count) runs low — quantization, the catch-up clamp biting during normal play, or a units mismatch (`g_JP_GlobalTimerDelta` vs `g_GlobalTimerDelta`, `chr.c:1435/1437`) — every delta-scaled AI motion integrates slow while **player** movement (input-driven per frame) still feels right, which is exactly the reported split. Related: **D170** (Ourumov/Trevelyan "flee" looks wrong — flagged there as possibly the same locomotion-speed issue rather than an AI divergence). Also possible: `chrGoToBond/Chr(SPEED_RUN/WALK/SPRINT)` speed enum → path-follow step, or pathfinding advancing ≤1 waypoint per frame-quantized tick. **Instances seen (M-49 full playthrough):** (a) Facility — the post-scripted-scene beat with Ourumov + the soldier squad: the soldiers **walk in super slowly and take a long time to path into firing positions** before they start shooting (this is regular squad combat AI, not a named-NPC script — good evidence the slowdown is general delta-scaled `chr` locomotion / path-following, not one ailist); (b) the named-NPC flee behaviour on Silo (Ourumov) and Cradle (Trevelyan) — see D170; (c) **Natalya (the escort NPC) always walks really slow** — on every escort level (Bunker ii, Statue, Control, Archives) she trails far behind, forcing the player to wait. Escort follow-AI is the same `chrGoToChr`/path-follow path as (a). **WORST CASE — Cradle (`-level_51`) is "bugged out completely" due to Trevelyan's behaviour (M-49).** The final level is Trevelyan-fleeing-and-scripted end to end (he runs the antenna cradle, climbs, drops to the cradle floor, the helicopter beat); if his locomotion timing / path-following is off he never hits the trigger points and the level's scripted progression stalls or goes haywire. This makes D193 effectively a **can't-properly-finish-the-game** bug, not just cosmetic sluggishness. **Next:** log `g_GlobalTimerDelta` + `g_ClockTimer` per frame during normal play and check the 1-second sum vs the N64's ~60; time a guard's run across a known distance vs reference footage; then watch Trevelyan's chr state / current AI command on Cradle.

**Status (from §F table):** **FIXED + user-playtest-verified (M-81) — root cause is D209**: `act_ubytes.padding[45]`, a raw-byte union alias for `act_gopos.unk59` (the SPEED tier), lands on offset 45 which is `unk59` on N64 but is byte 5 of `waypoints[1]` at 64-bit — always `0x00`. So the locomotion-animation selector `get_sound_at_range()` always received tier 0 and bound `ANIM_DATA_walking`/`walking_unarmed` for **every AI chr in the game**, and since travel is anim-root-motion driven they all moved at walk pace no matter what the ailist commanded. User verified on Cradle: Trevelyan and the guards run. Full elimination trail below — **steady-state timing keystone MEASURED and the "sim sums low" hypothesis is CONTRADICTED (M-80).** Added a `GE_D193=1` per-wall-second probe in `waitForNextFrame()` (`src/game/frametiming.c`, `#ifdef PORT`, env-gated): logs render frames, sim ticks delivered (= sum of the `deltaFrames` fed to `updateFrameCounters` = sum of `g_ClockTimer`), catch-up clamp hits, and max raw pre-clamp delta. Built ntsc-final, ran `-level_09 / _20 / _27 / _34` idle-after-load, ~20 s each. **Every level, rock-steady: render = 30.0 fps, sim = 60.0 ticks/s, clamped = 0, maxraw = 2.** i.e. each rendered frame carries `g_ClockTimer = 2` and the per-second sum is *exactly* the N64's ~60 — the wall-clock timer is NOT running low in steady state, and `FRAMETIMING_PORT_MAX_CATCHUP` never bites in normal play. VSync on/off made no difference (still 30/60). So the B1 premise ("`g_GlobalTimerDelta`/`g_ClockTimer` sum low over a real second") is wrong for the common case; D193 must come from elsewhere. **Residual quantization risk (not yet observed):** `nextFrameTime = (dt_counts + 387937) / 775875` is integer-rounded per frame with no remainder carry, so a scene that renders at a *steady* rate between 30 and 60 that is not an integer divisor of 60 (e.g. a solid 50 fps → dt≈20 ms → rounds to 1 → sim 50/s, ~17 % slow) would under-advance; the fix there is a fractional accumulator in the `#ifdef PORT` path. But the port renders a hard 30 (software RSP ≈ 60 µs/frame — not GPU-bound; the cap is the `pendingGfx < 2` double-buffer + sched round-trip), so this only matters if a real AI-heavy scripted scene drops into the 20–30 band. **Next (needs input, interactive session):** drive Bond into the Facility Ourumov/squad beat (D193a) and Cradle Trevelyan run with `GE_D193=1` — confirm render/sim hold at 30/60 under real AI load; if they do, pivot off timing entirely and instrument a single guard's world position vs. distance/time (anim root-motion rate, `chrGoToChr` SPEED_* enum → step size, path-node spacing) — the travel-speed bug is likely in the locomotion/anim path, not the frame clock. `GE_D193` catalogued in `docs/dev/GE-ENV-PROBES.md`. **M-80 cont. — LIVE-LOAD CONFIRMED, pivot is final.** User ran Cradle (`GE_D193=1`, `-level_41` — see the level-id note below) and drove Bond through the Trevelyan sequence while Trevelyan was visibly "walking really slow". Probe throughout: **`render=30 sim=60 clamped=0 maxraw=2`, dead steady** — no dips, no clamp hits, no `[D156]` guard fires. So under the *exact* failing scene the sim clock delivers a perfect 60 ticks/s. **D193 is NOT a timing/scheduler bug — it is in the character locomotion/animation path.** Best current hypothesis: GE character *travel* is animation-root-motion driven — `modelTickAnim` advances `frame += playspeed*speed` per tick (numticks = g_ClockTimer, correct), then `modelSetAnimFrame2WithChrStuff` (`model.c:3060+`) extracts per-anim-frame translation via `sub_GAME_7F06D3F4(...&pos)` and applies `scale = model->scale * model->anim_translation_scale`. The **player does NOT use anim root motion** (explicit speedgo/speedstrafe physics in bondview2.c), which explains the "player fine / all AI slow" split precisely. Suspects, in order: (1) the anim translation keyframe data is decoded with wrong units/scale by its sidecar converter (D43/model family); (2) `model->anim_translation_scale` or `model->scale` misread (ABI/struct-stride — D3x/D99 class); (3) `model->animrate`/`playspeed`/`speed` wrong. **M-80 cont. — port side instrumented (`GE_D193A` in `chrTick`, `GE_D193B` in `sub_GAME_7F06D3F4`; both `#ifdef PORT`, catalogued in GE-ENV-PROBES.md). Findings:** (i) `g_GlobalTimerDelta = 2.0`, `model->playspeed = 1.0`, `model->scale = 0.1`, `model->anim_translation_scale = 1.0` for every ticked chr on Silo + Facility — the scalars that feed the root-motion accumulation are all nominal, matching the identical N64 C (no `#ifdef PORT` in this hot path bar the non-firing D156 NaN guards). `model->animrate = 0.0` universally but that is the default — only written by `modelSetAnimPlaySpeed` on a blended transition (`startframe > 0`), so not the bug. (ii) Patrolling guards (`act=14 ACT_PATROL`, `model->speed=0.5`) travel **~60–99 world-u/s** on the port — no N64 number to compare against yet. (iii) `GE_D193B` (Facility, walk anim `706af4d8`): the decoded root-motion triple `tmp` ≈ (±0-3, ~1080-1094 hip height, ±0-50) with **`base = 0` and `angle = 0` on every call** (could be legit for a straight-line walk — joint 0 channel A at index 0, no per-frame facing change — or could be a mis-selected/zeroed channel; can't tell without the N64 side). **Bottom line: every port-side input to the locomotion math reads nominal**, so the divergence is either (a) subtle wrong magnitudes in the decoded anim bitstream values themselves (`tmp` triple / the descriptor decode), or (b) something in the `pos34` accumulation / `header->unk*` state — neither visible without a reference. `ModelSkeleton` **ruled out** — the skeletons are statically defined in C (`New_ModelSkeleton`/`MODELSKELETON` macros, `chrobjdata.c` etc.), not ROM-serialized, so `Joints` is a native x86-64 pointer and `ModelJoint` is 6 B on both platforms; `base` is correct. That leaves the ROM-serialized `ModelAnimation` bitstream (`bitDescriptors`/`bitStream`, D32-widened) and the anim-table load path as where a magnitude error could enter. **M-80 cont. — anim-data endianness (D33) audited, looks complete for this path.** `ptr_animation_table` is `romCopy`'d then byte-swapped by `romdataFixupAnimationData` (`port/src/romdata.c`, D33): per 20-byte record it bswaps `+0x00/+0x04/+0x08/+0x0C/+0x0E/+0x10` and per 6-byte `ModelAnimBitField` descriptor bswaps `+0 bitOffset` / `+4 valueOffset` (bitCount/pad identity); the bit*stream* is left raw (correct — `modelAnimReadRootMotionValue` walks it byte-by-byte, MSB-first, endianness-neutral, same as PD's `anim_read_bits`). `PTR_ANIM_walking/running/sprinting` **are** in `animation_table_ptrs1` so guard-locomotion records do get fixed up; `_ptrs2` is only 3 helicopter/plane anims (no overlap → no double-swap). No `bad descriptor range` / `bad record offset` errors in any run. Every `anim->unkNN` the locomotion path reads (`unk04`, `unk06`, `unk07`, `unk0C`) is within the fixed-up 0x14 bytes. `sizeof(ModelAnimation)` = 64 B both platforms (D32 kept the embedded ptrs as `u32`); `sizeof(ModelAnimBitField)` = 6 B both. So the anim decode is structurally sound as far as static review goes — **a magnitude error, if any, needs the N64 reference to see.** **Next — two viable paths (needs a fresh session; toolchain not present here — no `mips-linux-gnu-*`, no `qemu-irix`, no emulator):** (A) set up the MIPS toolchain + IDO-recomp + an N64 emulator with `osSyncPrintf` capture, build the ROM with `#ifndef PORT` `GE_D193B`-equivalent logging, navigate to the same scene, diff `tmp`/`base`/`angle`/travel-speed against the port capture; (B) user records N64 footage (real HW or their emulator) of one guard traversing a measured distance, frame-count it → gives the port-vs-N64 speed ratio (a clean 2× would point at a per-frame vs per-tick doubling; a fractional ratio at a scale/data error). Recommend (B) first — cheap, and the ratio narrows (A) massively. **Highest-impact playtest finding — breaks Cradle.** **Level-id note:** the earlier "Cradle = `-level_51`" in this row and HANDOFF was WRONG — `LEVELID_CRADLE = 41` (`src/bondconstants.h:1662`; STATUE=22, so CRADLE = 22+19). `-level_51` = `LEVELID_EAR` (unused slot) and SIGSEGVs at load in `stanBuildRoomData` (`stan.c:281`, NULL deref) — a non-issue (bad level id), noted here only so it isn't re-investigated as a real crash. **M-81 — SOLVED, and none of the M-80 suspects were the cause.** Re-verified the whole timing chain independently (`waitForNextFrame` -> `speedgraphframes` -> `g_ClockTimer` -> `g_GlobalTimerDelta`; `g_GlobalTimer += g_ClockTimer` at `lv.c:1022` is **tick**-based, so the ~75 AI timeout comparisons against it are frame-rate independent) and re-checked the D33 fixup *widths* against the real structs (they match exactly, and `(bs-bd) % 6 == 0` passing is independent evidence the two embedded pointers decode correctly). Root-motion accumulation is self-consistent (`pos24 = pos34 + rotate(pos)` at `model.c:3322`, so `pos` is a per-frame delta in x/z and an absolute hip height in y — matching the observed `tmp` shape). Extending `GE_D193A` with the visibility-gate counters (`tick`/`move`, i.e. chrTick entries vs `chrUpdateAnim` calls) killed the culling hypothesis too — chr 0 is `31/31` every second. The tier was then shown correct (`tier=1 SPEED_RUN`) and the *animation* wrong (`aidx=40/107`, never 42/41), which is **D209**. Note for future readers: `chrlvApplySpeed`'s `speedPtr` out-param is a **turn** rate, not a travel rate (`chraction.c:8070` passes `&act_runpos.turnspeed` into it), so `act_gopos.speed == 0` just means "running in a straight line" — it is not evidence of a stalled character.

## D194 — Mouse input needs another tuning pass — RMB aim over-sensitive + sensitivity couples to frame/sim rate (M-49, user playtest).

**Mouse input needs another tuning pass — RMB aim over-sensitive + sensitivity couples to frame/sim rate (M-49, user playtest).** Two parts. **(a) Aim mode (RMB) is near bang-bang, not proportional** (`port/src/input.c:697-707`): any per-poll delta past `AIM_MOVE_THRESH` emits stick `m = clamp(61 + |Δ|·(MouseAimSpeed/100)·AIM_GAIN, 61, 60+AimBand)` = `[61,80]` — with the M-29 defaults (`MouseAimSpeed=16`, `AIM_GAIN=4.0`, `AimBand=20`) `m` is already 61 (fast) at ~1.5 px/poll and saturates at 80 by ~30 px/poll, so the usable proportional band is tiny and sits near the top of GE's aim range → "way too sensitive," turn rate barely tracks how fast you actually move the mouse. Needs a real response curve (low floor, wider proportional band, maybe a gamma), not a threshold+clamp. **(b) Overall sensitivity correlates with movement/frame rate** (user's words: "correlated with movement speed, does not feel good"): `input.c` emits raw stick counts **per controller poll** with no `g_GlobalTimerDelta` / polls-per-frame normalization (hipfire yaw `sx += edx·(MouseTurnSpeed/100)·6.0` at `:709`; aim as above). GE integrates the stick scaled by its sim delta, and the port's poll cadence vs render/sim cadence drifts with scene load — so effective sensitivity shifts when the framerate does (which reads as "when I'm moving"). Fix direction: accumulate mouse delta and convert to stick using a fixed reference dt (or feed the game a dt-consistent value), so a given hand motion = a given view rotation regardless of fps. Lineage: D118a (hipfire pitch digital vs analog), D165 (menu pointer), D166 (hipfire pitch pulses), D180-B3 (`MouseAimSpeed` 25→16). New adjacent QoL ask, root-caused → **D223** (mouse-wheel direction should reverse weapon cycle direction, currently both directions cycle forward). Config knobs exist (`Input.MouseAimSpeed/AimBand/MouseTurnSpeed/HipfirePitchSpeed`) but the *shape* of the mapping is the problem. **M-87 design-goal note (user):** overall mouse feel should move toward **Quake Remaster / the Perfect Dark PC port** — raw, direct, responsive 1:1 mouse-look — while explicitly **not losing GE's original aiming nuance** (the analog-stick aim-mode vs digital-hipfire split this whole D118/D194 lineage exists to preserve). Reframes the fix direction from "just widen the proportional band" to: normalize by a fixed reference dt first (part (b), the frame-rate coupling — this alone should fix most of "feels unresponsive"), then re-derive the aim-mode response curve, and treat raw hipfire mouse-look (PD/Quake-style, bypassing the N64 digital-stick emulation entirely when the player is *not* in aim mode) as a real design option to evaluate against "keep GE's aim-mode feel," not assume the N64 stick-emulation path is mandatory everywhere. **M-87 — found it: GE already has exactly this control scheme, unused by the port.** User asked why vertical mouse feels less responsive than horizontal; investigation found the port forces `controldef = CONTROLLER_CONFIG_HONEY` (1.1), under which hipfire pitch is genuinely digital-only (C-up/C-down) while yaw is analog — an intentional N64-era asymmetry (D166), not a port bug, but avoidable. **`CONTROLLER_CONFIG_SOLITARE` (1.2) / `GOODNIGHT` (1.4)** — user's own recollection, verified in code — redirect forward/back/strafe-L/R onto digital C-buttons/D-pad (`bondview2.c:5195-5253`, `moveData.digitalStep{Forward,Back,Left,Right}`), which frees the **entire primary analog stick** for continuous, unclamped pitch **and** yaw: `moveData.canNaturalPitch`/`canNaturalTurn` (set `!insightaimmode`, i.e. active outside aim mode, not gated behind holding aim) drive `analogPitch/70.0f` / `analogTurn/70.0f` directly (`bondview2.c:5950`/`6008`) — no threshold, no digital step, no `[61,80]`-style clamp band (that clamp is specific to the *aim-mode* zoom-turn path, a separate codepath, `insightaimmode && stick_y>60`). `analogPitch`/`analogTurn` default to the primary stick's raw X/Y for every 1.x scheme (`bondview2.c:4881-4882`) unless a 2.x dual-controller scheme overrides them from a second controller. **This is architecturally identical to a modern FPS mouselook**: continuous analog look on the stick, discrete step movement on buttons — a legitimate, unmodified original GE control scheme, not a port invention. Recommended fix: force the port's `controldef` to SOLITAIRE or GOODNIGHT instead of HONEY, map WASD to the digital step-movement buttons (already close to how keyboard hipfire movement works today) and mouse X/Y directly to the primary stick's continuous X/Y — this should resolve the vertical-vs-horizontal asymmetry completely (no more digital-pulse emulation needed at all, D166 becomes largely moot) and likely fixes much of "feels unresponsive" too, since it's direct continuous stick input rather than any thresholded emulation. The frame-rate-coupling issue (part b) is orthogonal and still needs fixing regardless of scheme. Cross-ref D166, D223 (control-preset forcing — this gives a concrete "which preset" answer). **Aim-mode note:** the port's aim-mode code (`input.c:777-784`) already applies the same gain formula to both mouse axes symmetrically — if vertical still feels worse *while aiming* (RMB held), that's a separate, unconfirmed issue, not explained by any of the above. **M-87: user independently confirms part (a)** ("right mouse aim is way too sensitive") and suggests the design fix — aim sensitivity should scale off the **same base per-pixel sensitivity as normal (hipfire-turn) mouse movement**, not its own disconnected curve. Concretely: today hipfire turn is a direct accumulation (`sx += edx·(MouseTurnSpeed/100)·6.0`, `input.c:709`) while aim is the separate floored+clamped `[61,80]` stick-value synthesis (`AIM_GAIN`/`AimBand`, unrelated constants) — two independently-tuned systems that don't agree on what "1 inch of mouse" should mean. Aim mode is a distinct codepath from hipfire (gated by `insightaimmode`, **untouched by the SOLITAIRE scheme change above** — natural-pitch/turn is explicitly `!insightaimmode`), so this needs its own fix regardless of the control-scheme switch. Fix direction: derive the aim-mode mapping from the same base sensitivity constant as hipfire (`MouseTurnSpeed`-equivalent), scaled down only by the aim-zoom's FOV narrowing if any adjustment is wanted at all, rather than the separate `AIM_GAIN`/`AimBand` floor+clamp system — so moving the mouse the same physical distance feels the same whether or not aim is held, without losing GE's own aim-precision character.

**Status (from §F table):** **PARTIAL — M-120 human playtest: hipfire/primary look is GOOD now, aim-mode (RMB/Shift) is STILL too sensitive, unusable for precision aiming.** Part (b) (dt-decouple) + the SOLITAIRE natural-pitch scheme switch (part of this PR's combined pass) are user-confirmed working well on PR #62 — normal look feels right. **Part (a) (aim-mode response curve) does not deliver the intended fix despite being "implemented."** User's words: "right mouse / shift aiming with crosshair is still way too sensitive and does not feel right / is not usable in game." Code-level re-read of `port/src/input.c`'s `aimHeld` branch (~line 939) finds a likely concrete cause the original D194(a) design didn't eliminate: `m = AIM_STICK_MIN + pow(normX, gamma) * (AIM_STICK_GAME_MAX - AIM_STICK_MIN)` — the response curve (`aimCurveGamma`, gamma=1.6 default) only shapes the range **above** the `AIM_STICK_MIN=61` floor; the instant `|aimEdx| >= AIM_MOVE_THRESH` (a tiny 0.3 px/poll), stick jumps straight to 61, i.e. **10% of the game's max aim-turn speed (`(stick-60)/10` clamped 1.0), with no ramp from zero.** For hipfire comparison, the non-aim path (`hipEdx * hipSens * MOUSE_TURN_GAIN`, no floor, no threshold-snap) is fully proportional to zero — structurally different, and the one the user says feels good. Aim mode's floor+threshold-snap behavior reads exactly like the original "near bang-bang" M-49 complaint the curve was meant to fix, just now confined to the bottom of the range instead of the top. **Not yet build/runtime-verified** — this is a static code-read hypothesis, not a confirmed root cause; needs an env-gated probe or a tuning pass (e.g., drop `AIM_MOVE_THRESH` toward 0 and make the curve span `[0, AIM_STICK_GAME_MAX-AIM_STICK_MIN]` from a true zero floor, or lower `AIM_STICK_MIN` itself for a genuine near-zero low end) before another playtest. **PR #62 (`fix/d194-mouse-dt-decouple`) stays held, unmerged** — part (a) needs another iteration; parts (b)/SOLITAIRE are ready to land whenever the aim-mode fix is folded in and re-tested together. **M-122: all four stick-synthesis models above were abandoned.** GEPD's own source (`MouseInjectorPlugin/games/goldeneye.c`) was found and read: it never synthesizes a stick for aim at all — it directly overwrites `crosshair_x/y_pos`, `gun_azimuth_angle/turning` (its formulas, RATIOFACTOR=1, basefov) and edge-scrolls the camera past 72% of the way to the screen edge. The port now mirrors that exactly in `port/src/input.c aimGepdCompute()` (knobs `Input.AimAbsolute`, `Input.GepdSens`); no look stick is emitted while aiming, and the game's own damped crosshair update (`caclulate_gun_crosshair_position_rotation`, gunfire.c:4819) is left running — GEPD patches nothing there either, its per-frame overwrite simply wins. **User verdict: "mechanically it works... the aim mode works how we'd want."** Committed `ef49d9e6`. Residuals: ① crosshair/arm spazz glitch, ② aim-mode FPS drop, ③ sensitivity tuning (see M-122 in HANDOFF). **M-123: playtest log (`d194_gepd.err`, 1964 gepdaim lines) mined for ①.** Result: the port's own writes are rock-stable — zero crosshair/camera jumps while aim is held; every large camera delta in the log turned out to be user hipfire stick swings between polls. Exactly ONE true anomaly: a +359° one-tick spike in raw `vv_theta` (cam 0.5 → 359.7, log line 534) — the port's own `[0,360)` wrap of the edge-scroll accumulator firing when the game held a small positive angle and a leftward scroll crossed zero. GEPD does a bare `camx += ...` with no wrap, and the game itself never wraps `vv_theta` in on-foot play (`bondviewApplyVertaTheta` only sin/cos's it; tank code is the sole wraper), so the wrap was removed (port-only). Note: the handoff's theory that this feeds the crosshair via `speedtheta` does not hold — `speedtheta` comes from stick input (`bondview2.c:6056`), never a `vv_theta` delta; all solo-play consumers of `vv_theta` are sin/cos or `(360−θ)` rotations (raw-delta consumers at bondview2.c:1361/9397 are MP-only). The spazz must therefore live in the game-side terms added on top of our overwrite each tick (`turn_x = autoaimx | speedtheta*0.3`, `turn_y = autoaimy | −speedverta*0.1`, gunfire.c:4819 loop) — prime suspect: auto-aim pull oscillating as lock-on jitters, worst near enemies (matches the user's report). The gepdaim log line now also prints the pre-overwrite residue (`game=` = last tick's `pos·damp+turn`), `aa=(autoaimx,autoaimy)`, `st/sv` (speedtheta/speedverta) so a live capture can confirm or kill that. Tick ordering checked: our write runs BEFORE `bondviewProcessInput` each tick (osContStartReadData → inputComputePad, libultra.c:1094), vs GEPD writing after game code — but the displayed fixed point (`X·damp+turn`) is identical either way in steady state, so ordering is not a spazz source by itself. Suspect (c) 2px deadband does not apply to the GEPD path (deadband lives only in the legacy fallback); suspect (d) centre-spring nudge is ~0.1–0.3° over its 2 ticks — negligible, kept (it clears `docentreupdown`, which must not fight aim). **M-123 user playtest (post wrap-fix + extended logging, `GepdSens` 25): "looks good... other than sensitivity we're good, 95%"** — residual arm/gun/crosshair glitching accepted as playable; ① closed as acceptable, remaining D194 work is ③ live sensitivity co-tune (`Input.GepdSens`, ini at 30 at session end) and the still-uncontrolled ② aim-mode FPS retest (logging off). **D194 CLOSED (post-merge):** ③ done (hipfire/aim aligned, defaults `AimModeSens=38`/`MouseTurnSpeed=50`, merged `f239adf5`) and ② resolved by user report -- "no performance [issue] is good now" during the calibrated playtest sessions; the formal logging-off control was never run but the original ≤10fps complaint did not recur, so it is closed on user evidence (revisit only if an FPS regression appears later).

## D195 — Transparency / alpha surfaces broken on Control (M-49, user playtest)

**Transparency / alpha surfaces broken on Control (M-49, user playtest)** — Natalya's console room, the glass partitions + translucent projected wall displays. Originally reported as a blend-state defect. Reinvestigated this session (M-116) after the user reported the glass rendering **solid black** — traced to a real, separate, much bigger bug: `G_TEXTURE_GEN` (RSP environment-mapped texture coordinates, used by shiny/reflective materials game-wide — gold/silver weapon skins, chrome trim, and this glass) was never implemented in fast3d. D72.1 (an earlier session) removed the PD-inherited envmap UV-generation block wholesale to fix the Rareware logo, reasoning "GE never generates texture coordinates from vertex normals" — true for the logo (lit, but does not set `G_TEXTURE_GEN`) but false in general: every envmap draw was left sampling whatever raw `tc[]` happened to be authored (near 0), landing on a fixed corner texel of the reflection map every frame regardless of view angle — for a typically dark-edged reflection texture, a solid near-black surface.

**Status (from §F table):** **FIXED (`port/fast3d/gfx_pc.cpp`, `gfx_sp_vertex`) — restored envmap UV generation gated on the real `G_TEXTURE_GEN` bit** (ported from the reference Perfect Dark PC port's already-correct implementation), so D72.1's actual fix is untouched (the logo never sets this bit) while every other envmap draw gets real reflection coordinates. **User-verified live:** console glass and shiny gold/silver guns are now visibly reflective instead of flat black. Confirmed crash-free on a full 21-level sweep. **The remaining "not translucent" complaint is NOT a bug**: user confirmed after seeing the fixed reflective glass that it was never see-through on the original N64 either — it's a reflective (mirror-like) surface by design, not transparent glass. Traced (before the user's confirmation settled it) to the material's own combine formula, `alpha = (TEXEL0/1 LOD-crossfade alpha) * SHADE_alpha + PRIMITIVE_alpha` — an additive constant term that saturates alpha near 1.0 by design, matching "reflective, mostly opaque" intent, not a blend-state defect. **Closed.** Cross-ref: this same envmap fix likely also explains part of D228 (white NPC-weapon patches) if any of those are chrome/shiny weapon variants — worth checking before further D228 investigation.

## D197 — Character face/head textures wrap around the head, seen on Silo (M-49, user playtest).

**Character face/head textures wrap around the head, seen on Silo (M-49, user playtest).** Reads as `G_TX_CLAMP` not honoured (or a non-power-of-two wrap period) on a head/face texture tile — the texture repeats around the mesh instead of clamping at the UV edge. Not yet isolated to a specific character or texture; no `GE_PCDUMP`/`GE_TEXDUMP` capture taken. Likely fast3d texture-tile setup (clamp/wrap mode bits from the tile descriptor not round-tripped) rather than a data bug, given it's cosmetic and geometry-independent. **Next:** `GE_TEXDUMP` the offending head texture's tile descriptor on Silo, diff clamp/mirror bits against `gmain.s`'s expectations for that material. **v0.2.0-pre playtest (user):** no longer observed on Silo — closed, dropped from release known-issues.

**Status (from §F table):** CLOSED — user playtest v0.2.0-pre

## D198 — `alAdpcmPull`/`_decodeChunk` used `K0_TO_PHYS` (an unconditional `& 0x1FFFFFFF` mask) on 64-bit heap pointers (`src/libultra/audio/load.c`).

**`alAdpcmPull`/`_decodeChunk` used `K0_TO_PHYS` (an unconditional `& 0x1FFFFFFF` mask) on 64-bit heap pointers (`src/libultra/audio/load.c`).** `K0_TO_PHYS` is the N64 kseg0→physical mask, valid only because N64 pointers fit in 32 bits; on PC's 64-bit heap it truncates any pointer whose value exceeds 512MB, corrupting the `aLoadADPCM`/`aSetLoop`/`aADPCMdec` DRAM addresses at 3 call sites (`load.c:71`, `:468`, `:474`). Every sibling call site in `env.c`/`resample.c`/`reverb.c` already resolves state pointers via `osVirtualToPhysical` (the port shim, `port/src/libultra.c`), which is bit-identical to `K0_TO_PHYS` for real N64 kseg0 addresses but correctly PC-shimmed. **Fix (ABI-layer, `#ifdef`-free — same call resolves correctly on both platforms):** replace the 3 `K0_TO_PHYS(...)` calls with `osVirtualToPhysical(...)`. No behavior change on N64 (identical result for kseg0 pointers); fixes PC. Surfaced while building the Phase-3 software mixer (D199) — the acmd list these opcodes feed now actually executes on PC instead of being ignored.

**Status (from §F table):** FIXED (`src/libultra/audio/load.c`) — part of the Phase-3 mixer landing, not yet build/runtime-verified in isolation

## D199 — Phase-3 software audio mixer: macro-swap `aXxx` opcode execution + reverse-engineered `aSetBuffer` persistent-context semantics (`port/src/mixer.c`…

**Phase-3 software audio mixer: macro-swap `aXxx` opcode execution + reverse-engineered `aSetBuffer` persistent-context semantics (`port/src/mixer.c`, `port/include/mixer.h`, `include/PR/abi.h`).** The RSP audio ucode (`aspMain`) that would normally execute GE's acmd list never runs on PC (`port/src/ucode.c`); adapting the Perfect Dark PC port's macro-swap trick (`docs/dev/AUDIO-PLAN.md`), `include/PR/abi.h` now `#include`s `port/include/mixer.h` under `#ifdef PORT`, which redefines every `aXxx` macro to call an `Impl` function in `mixer.c` immediately against a small software "DMEM" scratch buffer, instead of packing RSP command words for later execution. GE uses the classic IDO libaudio ABI (verified against every call site in `src/libultra/audio/*.c` + `src/libultrare/audio/{env,reverb}.c`), not PD's differently-shaped "naudio New" ABI — DSP math (ADPCM decode, linear resample, linear envelope mix) is the same Nintendo/SGI ucode PD's `mixer.c` already ported; only the ABI/addressing layer is GE-specific and new here. Several opcodes (`aADPCMdec`, `aResample`, `aEnvMixer`, `aLoadBuffer`, `aSaveBuffer`) take only a state/DRAM pointer — their DMEM source/dest addresses and byte counts come from the most recent `aSetBuffer` call(s), which there is no RSP disassembly to check against; the persistent-context shape (`sCtx.{in,out,count,dryR,wetL,wetR}`) was reconstructed purely from reading every `aSetBuffer`+opcode call-site pair across `resample.c`/`load.c`/`reverb.c`/`mainbus.c`/`save.c`/`env.c` (see the field comments on `sCtx` in `mixer.c:52-82`). `aSegmentImpl` is a no-op — GE's audio DRAM addresses are already resolved via `osVirtualToPhysical` before reaching the mixer (see D198), so segment/base never factor into DMEM addressing here. **Risk:** the `aSetBuffer` context reconstruction is an inference, not a verified spec — if a call site this session didn't audit sets the context differently, audio for that path will mix from the wrong DMEM offsets. **Verified:** `-level_09` now runs full music + SFX 120s+ crash-free (post-D200 fix); no cross-level or opcode-level audio-correctness verification done yet (no golden-audio comparison tooling exists).

**Status (from §F table):** LANDED (`port/src/mixer.c`, `port/include/mixer.h`, `include/PR/abi.h`) — functional on `-level_09`, context-inference risk noted, needs a wider level sweep

## D196 — OS mouse cursor stays visible after closing the F10 options overlay in click-to-lock mode (M-49, user playtest).

**OS mouse cursor stays visible after closing the F10 options overlay in click-to-lock mode (M-49, user playtest).** `Input.MouseCaptureMode=1`. Open F10 overlay → `optionsOverlayToggle()` (`optionsoverlay.c:391`) calls `inputSuspendForOverlay()` which forces `SDL_ShowCursor(SDL_ENABLE)` + drops the grab. While open, `inputComputePad(0)` early-returns at `input.c:512` (overlay owns controller 0) **before** `reconcileGrab()` / any cursor-visibility call. On close: the next poll runs `reconcileGrab(menuMode)` (`input.c:522`) which re-grabs + hides the cursor **only if `captureArmed && !menuMode`** — i.e. only if you had already clicked-to-lock in a stage before opening F10. If you opened the overlay from a menu, or in a stage you hadn't clicked-locked yet, `want` computes 0, `mouseGrabbed` stays 0, and **nothing calls `applyCursorVisibility()`** — so the cursor `inputSuspendForOverlay()` made visible never gets re-hidden (correct free-but-focused state per `applyCursorVisibility()` `input.c:842-846` is *hidden*). **Fix (port-only, ~1 line):** call `applyCursorVisibility()` right after the `reconcileGrab(menuMode)` at `input.c:522` (self-heals every poll once the overlay-open early-return stops firing), or add an `inputResumeFromOverlay()` that runs `reconcileGrab` + `applyCursorVisibility` and call it from `optionsOverlayToggle()` on close. D180/D184/D165 input lineage. Minor.

**Status (from §F table):** **FIXED (`6e758f1d`, merged to `main`) — the first option above was applied:** `applyCursorVisibility()` now runs right after `reconcileGrab(menuMode)` in the controller-0 poll, so once the overlay-open early-return stops firing the cursor visibility self-heals on the next poll. Status cell was left stale at OPEN; corrected M-105.

## D192 — Front-end mission/level-select grid pointer still can't reach the outer cells — but only under `Input.MouseCaptureMode=0` (legacy always-grab) (M-49, user…

**Front-end mission/level-select grid pointer still can't reach the outer cells — but only under `Input.MouseCaptureMode=0` (legacy always-grab) (M-49, user playtest).** D169 was marked FIXED (M-33) by replacing the hard-coded 320×240 pointer clamp with the live `getPlayer_c_screenwidth/height/left/top()` rect; the feel-check was flagged "owed" on both D169 and D180 and never done. It now reproduces: with `MouseCaptureMode=0` the menu pointer runs the relative-delta P-controller path (`port/src/input.c:659-690`), which clamps target+estimate to `getPlayer_c_screen*` — but that accessor returns `g_CurrentPlayer->c_screenwidth` (`src/game/bondview.c:882`), which in the front end is the ~320×240 stage viewport, **not** the front end's actual 440×330 (`front.c:8570` `viSetViewSize(440,330)`). So the effective clamp is still ~300/220 while the mission-select grid's outer split points sit at 317 / 235.5 (`front.c:516/519/3180/3194`) — outer columns + bottom row unreachable, P-controller winds up and never settles. `MouseCaptureMode=1` (the shipped default, click-to-lock) writes `cursor_h_pos`/`cursor_v_pos` directly from the absolute OS cursor (`input.c:629-657`) and *feels* fine, though it shares the bad `hiH` so the last few px at the right/bottom edge may still be short. **Not a v0.1.0 blocker** — default mode is acceptable, always-grab is opt-in via INI, front-end roughness is already in the release notes. **Fix (port-only, small):** in the `menuMode` branch, when `current_menu` is a front-end menu use the real front-end rect (hard-code `front.c`'s cursor clamp `[20,420]×[20,310]`, or read the front-end viewport global) instead of routing through `g_CurrentPlayer`. Confirm with one `GE_INPUTLOG=1` capture in mission-select — sweep the mouse to each corner, watch where `cursor=(x,y)` plateaus. D165 · D169 · D180 lineage.

**Status (from §F table):** CLOSED — the buggy mode 0 was removed entirely in the v0.2.0 prep pass (see D239); click-to-lock is now the only capture mode, so the unreachable-cells path no longer exists

## D184 — F10 port-layer options overlay — the approach-(C) surface from `OPTIONS-MENU-PLAN.md` (M-37).

**F10 port-layer options overlay — the approach-(C) surface from `OPTIONS-MENU-PLAN.md` (M-37).** New `port/src/optionsoverlay.c` + `port/include/optionsoverlay.h`: an immediate-mode overlay drawn as its own fast3d 2D DL, appended after the game DL in `gfx_pc.cpp gfx_run()` (`optionsOverlayEmit()` → NULL when closed ⇒ **zero bytes appended, golden dumps byte-identical**). F10 toggles (`video.c videoPumpEvents`, next to F12); while open, `input.c inputComputePad(0)` returns a neutral pad and routes arrows/enter/wheel/mouse-buttons to `optionsOverlayHandleInput()` (mirrors the WI-1/D180 swallow). ESC closes; `configSave()` on close. `config.c` gains `configForEachOption()` + a `configSetOptionMeta()` label/step/enum side table (no per-key knowledge in config.c). Live knobs (VSync/FpsCap/TextureFilter via `videoRequestLiveConfig()`; mouse + screen-shake via the registered pointer) apply immediately; MSAA tagged "(restart)". v1 rows per plan §3: VSync, FpsCap, MSAA, TextureFilter, MouseAimSpeed, MouseTurnSpeed, MouseInvertY, MouseCaptureMode, ScreenShakeIntensity.

**Status (from §F table):** **LANDED, port-only, no `#ifdef PORT` in `src/`.** Build links 241/241. **Runtime verification NOT run this session** — the headless env kills the GUI process after ~8 s and `gfx_opengl_dump_bound_fbo` writes 0-byte PPMs here, so framediff / 60 s soak / overlay-frame eyeball are all owed on an interactive machine. Overlay-closed no-op is code-evident (emit returns NULL before touching the DL buffer).

## D203 — Steam Deck (SteamOS, x86_64 Linux) v0.1.0 release bundle: Facility (`-level_34`) crashes "after loading in as James Bond" (user bug report).

**Steam Deck (SteamOS, x86_64 Linux) v0.1.0 release bundle: Facility (`-level_34`) crashes "after loading in as James Bond" (user bug report).** Idle repro attempts on WSL Ubuntu with the actual `dist/goldeneye-pc-port-0.1.0-linux-x86_64.tar.gz` are **negative**: 120 s idle → frame 3300+ clean exit (the intro auto-advances, so "sitting as Bond" is covered), and a ~4-min stick-forward-only run → frame 6900+ also survived. So the trigger needs real gameplay input (fire/action) or is Deck-environment-specific. **Leading suspect: the D191 family** — both known v0.1.0 field crashes are SIGSEGVs in `modelGetNodeRwData` (`model.c:478`) over a truncated/NULL `root->Parent` node link, and both were triggered by *gameplay state* (first guard kill on Bunker ii; post-cutscene on Statue), never by idling; Facility opens into the Ourumov + soldier-squad scripted beat (D193a instance), which fits the trigger profile. Alternatives: Deck GL driver path (RADV/zink vs llvmpipe) or 40/50/60 Hz timing (D193 family). **Next:** (1) WSL run under gdb with a walk+fire `GE_INPUTSCRIPT` — generate the script into a file first (earlier attempts were blocked by a `bash -lc` quoting quirk that emptied the loop var, so no fire buttons ever went out); on SIGSEGV compare the fault PC against `modelGetNodeRwData`; (2) if not reproducible on WSL, get `ge007.crash.log` from the Deck (the Linux build writes it — proven by a stale one in the smoke tree) + ask what Bond was doing at crash time; **RESOLVED (D253):** reproduced on WSL Ubuntu by walking Bond into the crouch-forcing linked tile near the Facility opening; gdb capture showed SIGSEGV in `stanCheckLinkedSpecialTile` (`stan.c:2441`) storing to a stale `%r9` (0x3) while the caller's `record` local was valid — an x86-64 SysV int/float argument-class mismatch, not the D191 family. Fixed in D253; verified on WSL (same route no longer crashes).

**Status (from §F table):** CLOSED — root cause + fix in D253.

## D201 — `src/libultra/audio/bnkf.c` bank/instrument/sound/wavetable relocation offsets were `s32`, truncating the real 64-bit base pointer.

**`src/libultra/audio/bnkf.c` bank/instrument/sound/wavetable relocation offsets were `s32`, truncating the real 64-bit base pointer.** `alSeqFileNew`/`alBnkfNew` compute `offset = (s32) base` from a heap pointer and thread it through `_bnkfPatchBank/Inst/Sound/WaveTable` as `s32 offset, table` parameters, used to rebase every embedded pointer field in a loaded bank/inst/sound/wavetable/book/loop record. On N64 (32-bit pointers) the truncation is a no-op; on PC, sign-extension of the truncated `s32` during the rebase arithmetic reconstructs the wrong 64-bit address whenever the base allocation's low 32 bits have the top bit set (~50% of allocations, allocator-dependent) — corrupting every relocated pointer in the bank file, a plausible contributor to bad model/audio pointers surfacing downstream (D191-family symptoms are a node-tree walk, not this, but the class is the same). **Fix (ABI-layer, no logic change):** widen `offset`/`table`/`woffset` and all four `_bnkfPatch*` parameters from `s32` to `uintptr_t` (`<stdint.h>`) — already this codebase's idiom for pointer-safe integers in `src/game/*.c`. No behavior change on the 32-bit N64 build.

**Status (from §F table):** FIXED (`src/libultra/audio/bnkf.c`) — part of the Phase-3 mixer landing, not yet build/runtime-verified in isolation

## D208 — Player weapon-fire sound goes permanently silent mid-firefight (all other SFX unaffected) — `struct` field `bondview.h:200 s32 field_A48` is used as a second…

**Player weapon-fire sound goes permanently silent mid-firefight (all other SFX unaffected) — `struct` field `bondview.h:200 s32 field_A48` is used as a second `ALSoundState *` sound-handle slot, but is only 32 bits wide on PC (M-78c, ROOT CAUSE FOUND, D3x pointer-width class).** This is the "doesn't recover after the alarm stops" half of the user's D207 report — a distinct bug from D207's voice-pool question. The player's per-hand fire sound is double-buffered across two handle slots: `hands[h].audioHandle` (`ALSoundState *`, correctly pointer-width) and `hands[h].field_A48` (`s32`). `gunfire.c:3186-3200` (`bondfirefunc`) each trigger interval calls `sndPlaySfx(bank, idx, (ALSoundState *)&handptr->field_A48)`, and `sndPlaySfx` writes the new state through it as `pendingState->link.next = nextState` — an **8-byte pointer store** into a 4-byte field, overrunning into the adjacent `field_A4C` (which is the auto-fire sound-cadence timer, written back every frame at `gunfire.c:3168`). `sndUnlinkClearSound` likewise does an 8-byte `state->state->link.next = NULL` there. The two writers stomp each other; after enough double-buffered fire cycles `field_A48` holds a value that is neither `0` nor a valid pointer, so the `else if ((ALSoundState *)handptr->field_A48 == 0)` re-arm gate in the double buffer is **never true again → no further player gunshot is ever issued.** `audioHandle` alone can't cover because the code only replays into it `if (audioHandle == NULL)` and the async `sndDeactivate(audioHandle)` leaves it briefly non-NULL, deferring to the jammed `field_A48` branch. **Evidence (user Bunker capture, `audiotrace.log`, silenced PP7 = idx 46):** idx-46 `sndPlaySfx` requests stop dead at t=776609203515 (log line 18930 of 25086) while guard fire (109), pickups (232), impacts, doors all continue to the end; at line 19159 the player-gun slot `0x70076630` is caught holding `was 0x000001db00000000` — a torn 64-bit value = `field_A48=0x00000000` + `field_A4C=0x1db` (475, a `g_GlobalTimer`-relative cadence stamp) read as one word. The alarm is not required — it just raises `sndPlaySfx` churn so the tear is hit sooner. On N64 both fields are 4 bytes, `audioHandle` is a 4-byte pointer, and `&field_A48` punned as `ALLink*` is a 4-byte `link.next` write — all consistent, no overrun. **`bondview` / `hands[]` is runtime-only state (`g_CurrentPlayer`), never ROM-serialized**, so widening `field_A48` to pointer width is layout-safe (no by-offset or ROM-cast access — grep-verified; only `field_A50`/`field_A4C` sit after it and are plain `s32` timers).

**Status (from §F table):** FIXED + user playtest-verified (M-78d) — Bunker firefight: player gunshots persist through the whole engagement and survive the klaxon stopping. Committed 7601b9db (PR #29, merged).

## D209 — Every AI character in the game is locked to a *walk* animation regardless of the speed the ailist commands — `act_ubytes.padding[45]`, a raw-byte alias into…

**Every AI character in the game is locked to a *walk* animation regardless of the speed the ailist commands — `act_ubytes.padding[45]`, a raw-byte alias into the action union for `act_gopos.unk59` (the SPEED tier), shifts from offset 45 to 81 at 64-bit (M-81, ROOT CAUSE of D193, D3x pointer-width class).** GE character *travel* is animation-root-motion driven, so the bound locomotion animation sets the speed. `get_sound_at_range()` (`chraction.c:3573`, the locomotion-anim selector — misleading name) picks `ANIM_DATA_sprinting`/`running`/`walking` (and the one-handed/female variants) from its `arg1` tier. Its **only** caller, `play_hit_soundeffect_and_proper_volume()` (`chraction.c:3665`), passes the tier as `self->act_ubytes.padding[45]` — a raw byte read out of the `ChrRecord` action union rather than the named field. On N64 that alias is exact: `struct act_gopos` = `{coord3d targetpos@0, StandTile *target@12, waypoint *target_path@16, waypoint *waypoints[6]@20, u8 curindex@44, u8 unk59@45}`, so `padding[45]` **is** `unk59`. On x86-64 the three pointer members widen 4→8 B (`target@16, target_path@24, waypoints@32`), pushing `curindex`→80 and `unk59`→**81**, while the literal `45` now lands inside `waypoints[1]` (spanning 40..47) — specifically **byte 5 of a heap pointer**. The port's arena is low-4GB (`0x00000000_70xxxxxx`), so byte 5 is **always `0x00`** → `arg1 = 0` → `ANIM_DATA_walking` / `ANIM_DATA_walking_unarmed`, unconditionally, for every AI chr on every level since the 64-bit transition. The named-field reads (`chrlvApplySpeed(self, ..., (s32) self->act_gopos.unk59, ...)` at `:9104`, and the `unk59 == 2` / `== 1` `modelSetAnimSpeed` tiers at `:9106-9130`) resolve correctly, which is why the tier *looked* right everywhere it was inspected — only the alias was broken. **Runtime evidence (Cradle `-level_41`, `GE_D193A` extended with `tier`/`aidx`):** all six chrs report `act=15 ACT_GOPOS tier=1 SPEED_RUN` on 420/440 samples, yet `aidx` (live `model->anim` reverse-resolved against `animation_table_ptrs1`) is **40 = `PTR_ANIM_walking`** for Trevelyan (chr 0) and **107 = `PTR_ANIM_walking_unarmed`** for the guards on every single sample — `aidx 42 (running)` / `41 (sprinting)` never appear. Travel speed sat in a flat 140–155 u/s band for all of them, independent of animation, which is the walk-cycle stride. `act_ubytes` has exactly one use in the whole tree, so this is a single site. Both callers of `play_hit_soundeffect_and_proper_volume` reach it with `actiontype == ACT_GOPOS` (`:3723`, immediately after `actiontype = ACT_GOPOS; act_gopos.unk59 = speed;`; and `:9087`, inside the explicit `else` of `actiontype == ACT_PATROL`, mirroring the `chrlvApplySpeed` split directly below it), so reading the named field is exactly the byte N64 reads — no behavioural change.

**Status (from §F table):** **FIXED + user-playtest-verified (M-81).** `chraction.c:3665`, `#ifdef PORT` → `get_sound_at_range(self, self->act_gopos.unk59, ...)`; the N64 line kept verbatim under `#else`. ABI/pointer-width only, no game-logic change (rule-2 narrow exception, D3x class). User ran Cradle on the fix build: **Trevelyan and the guards run.** Also corrected the `// guess: room` comment on `unk59` (`src/bondtypes.h:2191`) — it is the SPEED tier.

## D210 — `chrToPatrol()` leaves `act_patrol.lastvisible60` uninitialised on PC — a raw-byte union alias (`act_init.padding[0x13]`) that shifts when `act_patrol` widens…

**`chrToPatrol()` leaves `act_patrol.lastvisible60` uninitialised on PC — a raw-byte union alias (`act_init.padding[0x13]`) that shifts when `act_patrol` widens (M-82, A1 class, sibling of D209).** `chrToPatrol` (`chraction.c:3872`, sets `ACT_PATROL`) does `self->act_init.padding[0x13] = -1;` — union-relative byte `0x4c`. On N64 `struct act_patrol` = `{patrol_path *path@0, s32 nextstep@4, bool forward@8, waydata@0xc (0x40 B), s32 lastvisible60@0x4c, f32 speed@0x50}`, so byte `0x4c` **is** `lastvisible60`, initialised to `-1`. On x86-64 the leading `path` pointer widens 4→8 B, so everything after shifts +4: `lastvisible60` is now at `0x50`, and the literal `0x13`/byte-`0x4c` write lands in the last 4 bytes of `waydata` (`segdisttotal`). Net PC effect: `lastvisible60` keeps whatever garbage the previous union arm left, and the "chr hasn't seen the player recently → break off / resume patrol" gate at `chraction.c:~9347` (`(act_patrol.lastvisible60 + CHRLV_DEFAULT_TIMER) < g_GlobalTimer`) reads that garbage on the first tick(s) after entering PATROL — until `chr.c:2569` / `chraction.c:3482` / `:9359` overwrite it (only when the chr actually sees the player). Same root class as D209 (raw index into a union arm whose pointer members widened); found by the post-D209 A1 sweep. Only new instance in the sweep — `act_init.padding[0..3]` in `chrlvTickAnim`/`chrlvTickDead` alias `act_anim`/`act_dead`, both pointer-free, so those are layout-stable.

**Status (from §F table):** **FIXED (`chraction.c:3872`, `#ifdef PORT` → `self->act_patrol.lastvisible60 = -1;`, N64 line under `#else`). ABI/pointer-width only (D3x/A1 class), no game-logic change.** Build clean; bunker1/facility/archives/dam verify PASS, no crash. **Owed:** a playtest eyeball of patrol-guard behaviour on entry to patrol (do they over/under-react to a player they haven't seen). Low individual impact but exactly the class the D170/D193 "AI behaviour looks off" reports keep landing in.

## D212 — Anisotropic-filtering config wiring — `Video.Anisotropy` (M-83 QoL wave, port-only).

**Anisotropic-filtering config wiring — `Video.Anisotropy` (M-83 QoL wave, port-only).** The fast3d GL backend already implemented `set_anisotropy_level` / `get_max_anisotropy_level` and applied `GL_TEXTURE_MAX_ANISOTROPY` on mipmapped tiles with a hardcoded `current_anisotropy_level = 4`; there was no way to configure it. Added the `gfx_set_anisotropy_level()` extern-C wrapper in `gfx_pc.cpp` (clamps to `[1, GL max]`, `reset_texture_state()` so live changes take), declared in `gfx_api.h`, and `Video.Anisotropy` (1–16) in `port/src/video.c` wired through `videoApplyImageOptions()` on init + live-config. **Default 4 = the value fast3d already applied ⇒ no visual delta at defaults.** F10 overlay row "Anisotropic".

**Status (from §F table):** **LANDED, port-only.** Build clean. **Owed:** eyeball at 16× on a grazing surface (Depot roof / Dam walkway) vs 1×.

## D213 — On-screen FPS readout — `Video.DisplayFPS` (M-83 QoL wave, port-only, PD parity).

**On-screen FPS readout — `Video.DisplayFPS` (M-83 QoL wave, port-only, PD parity).** `port/src/optionsoverlay.c` samples frame time once per emitted frame (`fpsTick()` via `sysGetMicroseconds()`, ~0.5 s window) and, when enabled, emits a small top-right green "N FPS" DL even while the F10 panel is closed. Config-only knob (kept off the 13-row panel, which is at its documented layout limit — matches PD, whose `DisplayFPS` is also file-only). **Default 0 ⇒ `optionsOverlayEmit()` still returns NULL when the panel is closed ⇒ golden dumps byte-identical.**

**Status (from §F table):** **LANDED, port-only.** Build clean; `bunker1` no-op confirmed. **Owed:** eyeball that the counter reads a sane ~30 and doesn't overlap the HUD.

## D214 — Keyboard rebinding — `[Input.Bind]` config section (M-83, port-only, PD parity, Phase 4 debt).

**Keyboard rebinding — `[Input.Bind]` config section (M-83, port-only, PD parity, Phase 4 debt).** `port/src/input.c` gains 12 gameplay actions (Forward/Back/StrafeLeft/StrafeRight/TurnLeft/TurnRight/Fire/Aim/Action/Cancel/LeanLeft/Start), each a comma-separated list of up to 4 SDL scancode names (`Input.Bind.Forward = W,Up` …). `inputRebuildBinds()` (called from `inputInit()`, after `configLoad()`) parses them via `SDL_GetScancodeFromName`; unknown names log a warning and are skipped. The old hardcoded `keyDown(ks, SDL_SCANCODE_*)` block in `inputComputePad` is replaced by `actHeld(ks, IA_*)`. Mouse buttons (fire = LMB, aim = RMB) stay hardwired; the mouse-wheel weapon-cycle A-pulse is untouched. **Defaults reproduce the previous FPS layout exactly — verified: all 20 default key names round-trip to the identical `SDL_Scancode` the old code used (`SDL_GetScancodeFromName` test, 0 mismatches).** Config-file only (no in-overlay text entry — matches PD, whose rebinding is also file-driven); a fresh `ge007.ini` gets the full `[Input.Bind]` block with the defaults visible.

**Status (from §F table):** **PARTIAL — M-84: code PASS (alt/ctrl); one doc-only fix.** LANDED, port-only, no `#ifdef PORT`; build clean. One doc bug fixed (no code change): `config.c` groups keys by the substring *before the last `.`*, so the section header is **`[Input.Bind]`**, not `[Bind]` — earlier finding/commit/QOL text was wrong. Working syntax: put `Fire = Left Alt` under `[Input.Bind]`, or write a fully-qualified `Input.Bind.Fire = Left Alt` line anywhere (`config.c:307`). Separately: the pause→**watch menu** does not wire Esc→back (still Enter to open/close/back) — pre-existing, `options.c`, minor.

## D215 — 1P weapon viewmodel never drew in-level — `gunRenderFirstPersonGunModels` builds its `ModelRenderData` from a reinterpret across separately-declared adjacent…

**1P weapon viewmodel never drew in-level — `gunRenderFirstPersonGunModels` builds its `ModelRenderData` from a reinterpret across separately-declared adjacent globals that only works on N64 (M-83, D3x/ABI class; retires VIEWMODEL-RESEARCH.md).** The N64 line `renderdata = *(ModelRenderData *)&D_80035CC0;` (`gunfire.c`) casts the lone `u32 D_80035CC0 = 0;` to a whole `ModelRenderData` and reads on into the next declared global, `u32 D_80035CC4[] = {1, 3, 0, 0, …}` (`gun.c`) — yielding `{ basemtx=0, zbufferenabled=1, flags=3, rest=0 }` **only because the N64 linker lays the two globals out contiguously in declaration order and pointers are 4 B**. On x86-64 it breaks twice: `ModelRenderData`'s three pointer members (`basemtx`/`gdl`/`mtxlist`) widen 4→8 B so `flags` is no longer at struct offset 8, and C guarantees nothing about the two globals' relative placement. Measured (`GE_DVM=1`, BUNKER1): `renderdata.flags == 0`, so **every** weapon-model DL node was gated off at `modelRenderNodeGundl` (`(renderdata->flags & 1)`), and `subdraw()` for the weapon emitted just its `gSPSegment` — 1 gfx command instead of ~150. All three VIEWMODEL-RESEARCH.md ranked suspects were ruled out by the same probe first: `field_87F==1` (gate open), `RootNode` valid, `numRecords 22 < 192` (pool fine), node walk visits 24+ valid `DL`/`SWITCH`/`BSP` nodes. **Fix:** `#ifdef PORT` → set the three fields explicitly (`zbufferenabled=TRUE; flags=3;`), N64 line verbatim under `#else`. ABI/layout only, no logic change.

**Status (from §F table):** **FIXED (`src/game/gunfire.c`, `#ifdef PORT`).** Build clean. `GE_DVM=1` post-fix: `flags=0x3`, weapon `subdraw()` now emits **153** gfx commands (was 1). Regression sweep crash-free. `GE_DVM` gate+render probes left in, env-gated (cf. `GE_D193*`). **Owed:** one BUNKER1 eyeball that the gun model is actually visible + correctly posed (headless can't see pixels; the DL is now provably generated). The committed `bunker1` golden will need re-baselining — the gun now renders in-frame.

## D216 — `Game.SkipIntro` — boot straight to the file-select menu (M-83, port + one `#ifdef PORT` game-code line, PD parity, Phase 4 quick win).

**`Game.SkipIntro` — boot straight to the file-select menu (M-83, port + one `#ifdef PORT` game-code line, PD parity, Phase 4 quick win).** New `s32 portSkipIntro` (`port/src/video.c`, `configRegisterInt("Game.SkipIntro", 0, 1)`), read once in `src/game/lv.c` at `LEVELID_TITLE` stage load — in the `else` of the existing `GE_STARTMENU` `#ifdef PORT` block. When set it does what every intro stage's own `!is_first_time_on_main_menu` branch does (`front.c`): `is_first_time_on_main_menu = FALSE; prev_keypresses = TRUE; maybe_is_in_menu = TRUE; menu_update = MENU_FILE_SELECT;` — i.e. it reuses the game's own post-intro route rather than inventing one, so the legal screen + Nintendo/Rare/GoldenEye logo attract loop is skipped. `GE_STARTMENU` still wins when both are set. Default 0 ⇒ untouched.

**Status (from §F table):** **LANDED + VERIFIED this session.** Build clean. Headless: default boot frame 90 = the TWYCROSS classification screen (10% non-black), `SkipIntro=1` frame 90 = the **SELECT FILE menu** (folders/photos/cursor all render, 89% non-black), no crash across repeated boots. Screenshots captured + eyeballed. **Owed:** nothing critical — a human boot to confirm the menu is fully interactive (cursor moves, a save slot opens). **REGRESSED (v0.2.1 user testing, 2026-09-15): user reports enabling `Skip intro` breaks audio.** Not root-caused this session — leading suspect is that the intro/attract stages perform some audio-subsystem init (title/attract music kickoff, or a first-touch of the mixer) that `portSkipIntro`'s jump to `MENU_FILE_SELECT` bypasses. **Mitigation applied:** the `Game.SkipIntro` row was pulled from the F10 options overlay (`port/src/optionsoverlay.c`) so players can no longer enable it from the menu; the config var, default (0/off), and `lv.c` hook are untouched (so an existing `ge007.ini` with it already set would still trigger the bug — worth a follow-up hard-disable if that turns out to matter). Do not re-add the menu row without root-causing the audio interaction first.