diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml index 5d88cc03..03981458 100644 --- a/.github/workflows/ci.yml +++ b/.github/workflows/ci.yml @@ -5,7 +5,10 @@ on: branches: [ main, master ] tags: [ 'v*' ] # No paths-ignore here: tag pushes must always build so a release can be - # cut even when the tagged commit only touched docs. PRs keep the filter. + # cut even when the tagged commit only touched docs, and a workflow that + # never triggers would leave a required status check pending forever. + # Docs-only branch pushes are instead skipped by the `changes` job below, + # which reports the build jobs as "skipped" (green for branch protection). pull_request: paths-ignore: [ '**.md', 'docs/**', 'LICENSE', 'NOTICE' ] workflow_dispatch: @@ -42,7 +45,42 @@ concurrency: # only job granted contents: write. jobs: + # Did this push touch anything outside docs / licenses? Branch pushes to main + # get no workflow-level paths-ignore (see the `on:` comment), so this is where + # a docs-only merge is filtered out. Tags and pull_request events always build. + # Mirrors the pull_request paths-ignore globs: **.md, docs/**, LICENSE, NOTICE. + changes: + runs-on: ubuntu-22.04 + outputs: + code: ${{ steps.filter.outputs.code }} + steps: + - uses: actions/checkout@v7 + with: + persist-credentials: false + fetch-depth: 0 + + - id: filter + run: | + if [ "${{ github.event_name }}" != "push" ] || [ "${GITHUB_REF_TYPE}" = "tag" ]; then + echo "code=true" >> "$GITHUB_OUTPUT"; exit 0 + fi + base='${{ github.event.before }}' + if ! git cat-file -e "${base}^{commit}" 2>/dev/null; then + echo "no usable base ${base} — building to be safe" + echo "code=true" >> "$GITHUB_OUTPUT"; exit 0 + fi + changed="$(git diff --name-only "$base" '${{ github.sha }}')" + echo "Changed files:"; echo "$changed" + if echo "$changed" | grep -qvE '(\.md$|^docs/|^LICENSE$|^NOTICE$)'; then + echo "code=true" >> "$GITHUB_OUTPUT" + else + echo "docs-only push — skipping build" + echo "code=false" >> "$GITHUB_OUTPUT" + fi + validate: + needs: changes + if: needs.changes.outputs.code == 'true' runs-on: ubuntu-22.04 steps: - uses: actions/checkout@v7 @@ -78,6 +116,8 @@ jobs: # address-constant romassets_.s. The ROM (.z64) and the ROM-derived # blobs are runtime-only dependencies — CI cannot run the game, only build it. windows-build: + needs: changes + if: needs.changes.outputs.code == 'true' runs-on: windows-latest defaults: run: @@ -139,7 +179,7 @@ jobs: # gets contents: write. release: if: github.ref_type == 'tag' - needs: [ validate, windows-build ] + needs: [ changes, validate, windows-build ] runs-on: ubuntu-22.04 permissions: contents: write diff --git a/CLAUDE.md b/CLAUDE.md index 6888f20a..b2816cee 100644 --- a/CLAUDE.md +++ b/CLAUDE.md @@ -66,3 +66,28 @@ Full template: `docs/dev-process.md` -> "Investigation-brief template". - **Per batch / integration checkpoint:** one 60s `-level_09` crash-free run. - Subagents on small disjoint subsystems (input, audio, saves, converters) don't run the game — the overseer verifies. + +## Local Qwen dispatch (cost-free, data-local worker) + +A local Qwen3.8-27B is wired in via the `delegate-local` MCP server (see +`C:\Users\james\Source\Repos\20260902-qwen38claudepair`). Use it as a +*subagent whose compute runs on-machine* for scoped, checkable work — the lead +still owns judgment, planning, and the verification gate. + +- **Tools:** `local_backend_status()` (run once before first dispatch), + `list_local_agents()`, `delegate_to_local_agent(agent_name, task, workdir, max_turns?, max_tokens?)`. +- **Workers** (`~/.claude/agents/`): `repo-mapper` (read-only ingest/map), + `mechanical-coder`, `test-writer`, `triage` (read-only root-cause + patch), + `refactor-bot`. +- **Dispatch when ALL hold:** narrow blast radius, clear spec, output checkable + by a build/test/diff. Keep on Claude: ABI/handedness/format reasoning, novel + debugging, anything you can't cheaply verify. +- **Backend reality:** ~1 inference slot, slow per token — dispatch + **sequentially**, `max_tokens` ~4–8K, treat as batch work. Needs the LiteLLM + proxy (`litellm/run.ps1`) + Unsloth Studio both up. +- **Windows quirk:** the worker's shell is `cmd.exe`, not bash — its agent + prompt tells it to use `read_file`/`findstr`, not `cat`/`grep`. Always + spot-check the returned artifact against the real files before integrating; + the 27B will confidently fabricate if a read fails. +- Fits the **dispatch preflight** rules above (FILES / BUDGET / ON EXPIRY / + CONSTRAINTS / REPORT) — write the `task` string to that shape. diff --git a/README.md b/README.md index 3ba471a1..19b15482 100644 --- a/README.md +++ b/README.md @@ -187,13 +187,13 @@ the front end into the early game. It is not finished and it is not stable. ## Download -Pre-built **Windows x86-64** bundles are published under -[Releases](../../releases) as marked pre-releases — the engine executable plus -its runtime DLLs, no ROM and no game assets. You still supply your own ROM (see -[Requirements](#requirements)) and drop it in `data/` as described in the -bundled README. Every download is a research alpha with the caveats in -[Status](#status) above; there is no stable build yet. To build from source -instead, see [Building](#building). +There are no pre-built binaries yet. The project is Phase 2 WIP with the +caveats in [Status](#status) above — nothing runs end to end — so the only way +to try it is to build it yourself; see [Building](#building). You supply your +own ROM (see [Requirements](#requirements)) and drop it in `data/` before +building. Pre-release **Windows x86-64** bundles (engine executable plus +runtime DLLs, no ROM or game assets) may be published under +[Releases](../../releases) once the port is playable. ## Requirements @@ -388,19 +388,19 @@ GitHub. It follows the same conventions they do: - **No ROM and no game assets are distributed** — not in this repository and not in any release. Textures, audio, models, level data and in-game text are - extracted from a ROM *you already own*, on *your* machine (at build time for - a source build; at first run for a pre-built binary). + extracted from a ROM *you already own*, on *your* machine, at build time. - The repository is a fork of the public [GoldenEye 007 decompilation](https://github.com/n64decomp/007) and inherits its contents unmodified (see [`NOTICE`](NOTICE) for what that includes). - No official logos, box art, or marketing assets are used. "GoldenEye 007", "007", "James Bond" and related marks belong to their respective owners (Nintendo, Microsoft/Rare, MGM, Danjaq, EON Productions). -- **Pre-built binaries** offered under [Releases](../../releases) contain only - the engine — the `port/` layer plus the compiled decompilation, with no game - data of any kind — bundled with permissively-licensed runtime libraries - (SDL2, zlib, the MinGW runtime; their licenses travel in the download). Any - build, yours or ours, is useless without a ROM you supply. +- **If** pre-built binaries are published under [Releases](../../releases) in + the future, they will contain only the engine — the `port/` layer plus the + compiled decompilation, with no game data of any kind — bundled with + permissively-licensed runtime libraries (SDL2, zlib, the MinGW runtime; their + licenses travel in the download). No such release exists yet, and any build — + yours or ours — is useless without a ROM you supply. This project is **not affiliated with, endorsed by, or sponsored by** Nintendo, Rare, Microsoft, MGM, Danjaq, EON Productions, or any rights holder in diff --git a/docs/_includes/head-custom.html b/docs/_includes/head-custom.html index c8965e02..253dcf8d 100644 --- a/docs/_includes/head-custom.html +++ b/docs/_includes/head-custom.html @@ -250,4 +250,39 @@ .page-header { padding: 3rem 1rem 2.75rem; } .main-content { padding-top: 2rem; } } + + /* ---- Mermaid diagrams ------------------------------------------------ */ + /* Kramdown renders ```mermaid fences as 
.
+     The script below hands their text to mermaid and swaps in the SVG. Keep
+     the pre unstyled (no dark code background) while it waits to render. */
+  .main-content pre.mermaid,
+  .main-content pre > code.language-mermaid {
+    background: transparent;
+    border: 0;
+    color: var(--ink);
+    padding: 0;
+  }
+  .main-content pre.mermaid { text-align: center; overflow-x: auto; }
+  .main-content pre.mermaid:not([data-processed]) { visibility: hidden; }
 
+
+
diff --git a/docs/dev/agentic-development.md b/docs/dev/agentic-development.md
index 34ecbe19..67bc71a5 100644
--- a/docs/dev/agentic-development.md
+++ b/docs/dev/agentic-development.md
@@ -3,7 +3,7 @@
 *A case study in AI-agent collaboration on a large, low-level codebase. A
 GoldenEye 007 Nintendo 64 → desktop port, taken from "builds a ROM" to
 "playable front end" in about two weeks — driven mostly by two coding agents (a
-local open-weight model, Qwen 3, on a single RTX 5090, and a hosted frontier
+local open-weight model, Qwen 3.8, on a single RTX 5090, and a hosted frontier
 model, Claude / Claude Code), handing work back and forth through shared
 written notes under one person's part-time direction. This is the goal, the
 method, the timeline, and an honest read on what did and didn't work.*
diff --git a/docs/dev/findings.md b/docs/dev/findings.md
index 7850df94..1daabd6b 100644
--- a/docs/dev/findings.md
+++ b/docs/dev/findings.md
@@ -392,7 +392,7 @@ covers D24–D69; the log continues in §H (D32 procedure, D70–D121).
 | 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).** 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). | **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).** 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. | **FIXED (M-36)** — high confidence |
 | D175 | **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. | 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 nothing → see "D176(a)" below, partial, not root-caused. **(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. See `GRAPHICS-BACKLOG.md` D176. | (a) partial / (b) open — both M-36 |
+| 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) root-caused, fix owed / (b) open |
 | 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 |
 | D174 | **"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. | 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 |
@@ -4556,6 +4556,24 @@ Remaining candidates, in order:
    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
@@ -4636,6 +4654,51 @@ sunset cloud gradient. Headless `-level_36` + `GE_D176=1` probe
 `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).
+
 ---
 
 ## D177 — Ladders non-functional: `count`/`rooms` land in the high half of a widened pointer (M-36)
diff --git a/docs/dev/release-dropin-rom-plan.md b/docs/dev/release-dropin-rom-plan.md
new file mode 100644
index 00000000..c374f266
--- /dev/null
+++ b/docs/dev/release-dropin-rom-plan.md
@@ -0,0 +1,244 @@
+# Release plan — "drop in your ROM and play"
+
+Status: proposal / not yet scheduled. Owner: TBD.
+
+## Goal
+
+An end user downloads the GitHub release zip, unzips it, drops their
+legally-owned GoldenEye 007 N64 ROM (`.z64`, big-endian) into a folder,
+double-clicks the exe, and plays. No Python, no MSYS2, no manual
+asset-extraction, no command line.
+
+## Where we are today
+
+The Windows bundle produced by `tools_pc/bundle-win.sh` (and uploaded by the
+`windows-build` CI job) already contains only: the engine exe, its MinGW /
+SDL2 / zlib DLLs, a README, and licenses. It deliberately ships **no** ROM and
+**no** game assets. That part is fine.
+
+The gap is everything the runtime needs *alongside* the ROM before a level
+will render:
+
+| Dependency | Produced today by | Consumed at runtime by |
+|---|---|---|
+| `port/src/romassets_.s` — absolute cart-address symbols for every obseg/ramrom/music asset + `ge007.ld` segment markers | `python3 scripts/gen_romassets.py `, compiled **into** the exe | linker; `&symbol` + `romCopy()` in game code |
+| `data/pcmodels-/{pcmodels.bin,manifest.csv}` (~1.3 MB) — every model file re-laid-out to PC struct layout, RZ-compressed | `python3 tools_pc/d43_emit.py ` | `port/src/pcmodels.c` (`pcmodelsReserveSize`/`LoadSidecars`/`PatchTable`) |
+| `data/pccg-/{pccg.bin,manifest.csv}` (~3.6 MB) — every stage `bg/*.seg` + `Tbg_*_stanZ` re-laid-out to PC layout | `python3 tools_pc/d69_emit.py ` | `port/src/pccg.c` |
+| ROM present at `data/ge007..z64` (or `baserom..z64`) | user copies it | `port/src/romdata.c` maps it at `0x10000000` |
+
+Without the two sidecar trees the game shows the intro logos and then crashes
+in `modelPromoteNodeOffsetsToPointers` (finding D179).
+
+The full decomp asset-extraction toolchain (`scripts/extract_baserom.u.sh`,
+`tools/extractor`, MIPS binutils) is **not** on the runtime path and never
+needs to reach an end user — it only regenerates the committed `assets/**`
+`.inc.c` / `.bin` metadata that a plain `git clone` already has.
+
+### Important facts established while writing this plan
+
+1. **`gen_romassets.py` does not need the ROM.** It resolves every symbol for
+   `u`, `e`, and `j` from committed files (`scripts/filelist..csv`,
+   `assets/obseg/file_resource_table.inc.c`, `assets/ramrom/ramrom.s`,
+   `assets/music/*.s`, `ge007.ld`). The ROM is only opened to size the
+   trailing `images` segment, and there is already a CSV-derived fallback when
+   it is absent. So all three `romassets_.s` files can be generated in CI
+   with no ROM. (`docs/building.md` claims PAL/JP need "region-specific
+   prop/font offsets from that region's ROM" — the current generator code
+   shows no such ROM read; treat that sentence as stale, but verify before
+   relying on it — see Open questions.)
+
+2. **`d43_emit.py` / `d69_emit.py` are pure-stdlib Python 3.** They read only
+   the ROM plus committed repo files, and their output is a deterministic
+   function of the ROM. Both carry extensive built-in round-trip validation
+   (re-parse the emitted image, check every field against the N64 source) and
+   exit non-zero on any mismatch. `d43_emit.py` also runs a battery of
+   buffer-fit cross-checks.
+
+3. **The engine is region-locked at compile time.** `REGION_DEFS` in
+   `CMakeLists.txt` differ per region (`VERSION_US/EU/JP`, `BUGFIX_*`,
+   `LEFTOVERDEBUG`, …) and are used in 60+ game files. One exe cannot serve
+   all three ROMs; the release must ship three exes (or one launcher that
+   dispatches to three).
+
+4. **Data/config location is already sane.** `sysResolvePath()` maps `$S/` to
+   `./data/` (next to CWD if present, else next to the exe). `ge007.ini`,
+   `ge007.eep`, and the sidecar trees all live under `$S/`. First launch
+   already writes `ge007.ini` there. The sidecar loaders already search
+   `$S/`, `$E/`, and `./`.
+
+5. **ROM identity checking already half-exists.** `port/src/romdata.c`
+   validates magic / title / cart-ID / country byte. `tools_pc/romverify.c`
+   goes further (recomputes the `n64cksum` 6102 checksum over
+   `[0x1000,0x100000)` using `tools/n64cksum.c`). Canonical SHA-1s for all
+   three regions are in the README requirements table.
+
+## Recommended approach
+
+**Two-part: (A) fold the ROM-independent piece into CI now, and (B) add a
+native first-run converter to the engine so the sidecars are generated on the
+user's machine from their own ROM.**
+
+### Part A — ship all three regions, pre-generate all `romassets`
+
+Small, low-risk, unblocks a "download three exes, pick the one for your ROM"
+release even before Part B lands.
+
+- CI `windows-build` becomes a 3-way matrix (`ntsc-final`, `pal-final`,
+  `jpn-final`). Each leg runs `gen_romassets.py ` then `build-pc.sh`.
+- `bundle-win.sh` collects all three exes (`ge007.ntsc-final.exe`,
+  `ge007.pal-final.exe`, `ge007.jpn-final.exe`) into one bundle, with shared
+  DLLs.
+- Add a tiny **launcher** `ge007.exe` (≈150 lines C, its own CMake target):
+  reads `data/*.z64` (and `baserom.*.z64`), checks the header country byte +
+  SHA-1 against the three known-good hashes, then `CreateProcess`es the
+  matching region exe. On no ROM / unknown ROM it shows a message box naming
+  the accepted versions and hashes and the folder to drop the ROM in.
+  - Alternative if a separate launcher is unwanted: make each region exe, on
+    detecting a ROM whose country byte is for a different region, re-exec its
+    sibling. Slightly hackier; avoids a fourth binary.
+
+### Part B — native first-run sidecar converter
+
+Port `d43_emit.py` and `d69_emit.py` to C as a new engine module
+`port/src/pcconvert/` (files: `pcconvert.h`, `d43.c`, `d69.c`,
+`pcconvert.c` orchestrator). On startup, `romdataInit()` (after it has the
+mapped ROM and region) calls `pcconvertEnsureSidecars(region)`:
+
+1. If `$S/pcmodels-/pcmodels.bin` **and** `$S/pccg-/pccg.bin`
+   exist and their embedded format-version tag matches the build → return
+   immediately (the common case after first run).
+2. Otherwise show a "Preparing game data from your ROM (one-time, ~5 s)…"
+   splash / log line, run the two converters, write the four output files
+   into `$S/`, then continue boot. Cache a `pcconvert.stamp` with the ROM
+   SHA-1 + converter version so a ROM swap or engine upgrade re-runs it.
+
+The converters need, besides the ROM:
+
+- `scripts/filelist.u.csv`
+- `assets/obseg/file_resource_table.inc.c`
+- `assets/**/ModelFileHeader.inc.c` (NS/NT per model)
+- the stage `bg` / `stan` `.inc.c` metadata `d69` reads
+
+These are committed text/CSV, a few hundred KB total. **Embed them into the
+exe at build time** via a CMake codegen step (`file(READ ...)` →
+`.byte`/`const char[]`), so the release stays "exe + DLLs" with nothing loose.
+A build-file change is required here (new `add_custom_command` + generated
+source) — out of scope for this doc, but noted as the one unavoidable
+`CMakeLists.txt` edit.
+
+Reuse: `zlib` is already linked (RZ = `0x11 0x72` + raw deflate — the exact
+codec `pcmodels`/`pccg` already decompress). The engine already has robust
+byte-swap helpers in `romdata.c`. The Python round-trip validators should be
+ported too and run in debug builds / behind an env flag.
+
+**Effort estimate:** `d43_emit.py` is ~1000 lines of intricate model-graph
+simulation (LOD/SWITCH rewire, BSP splice, marker-expansion bounds, DFS
+placement) + validation; `d69_emit.py` ~500 lines. Call it a 2–4 week
+focused task with careful diffing of C output against the Python output
+byte-for-byte on all three ROMs before trust. This is the single biggest
+work item.
+
+### Interim option (faster Part B, if timeline pressure)
+
+Freeze the two emit scripts with **PyInstaller / Nuitka in the Windows CI
+job** (the runner has Python) into one `ge007-convert.exe`, ship it in the
+bundle, and have the engine `CreateProcess` it on first run when sidecars are
+missing. Pure-stdlib scripts freeze into a single clean exe with no user
+Python. Downsides: ~8–15 MB added to the bundle, Windows-only, and it is a
+second binary to sign / trust. Recommend this **only** as a stopgap that buys
+time for the real C port; it satisfies "no user-visible Python/MSYS2" but not
+"single native binary" or non-Windows.
+
+### Rejected options
+
+- **Precompute and ship the sidecars.** They are ROM-derived copyrighted game
+  data (`data/pcmodels-*`, `data/pccg-*` are gitignored precisely for this).
+  Redistribution is not acceptable. Rejected.
+- **Region-independent sidecars.** `d43`/`d69` output is a function of the
+  ROM; there is no region-independent form to precompute, and even if there
+  were it would still be ROM-derived data. Rejected.
+- **Bundle CPython + the .py scripts loose.** Meets the letter of "no MSYS2"
+  but not "no Python", ages badly, and is bigger than a frozen exe for no
+  gain. Rejected in favour of the interim freeze if a stopgap is needed.
+
+## CI / release changes
+
+1. `windows-build` → matrix over the three ROMIDs; each leg generates its
+   `romassets_.s` and builds.
+2. Commit all three `port/src/romassets_{u,e,j}.s` (or generate in each CI
+   leg — committing is simpler and lets non-CI builds skip the generator).
+   Add an `e`/`j` regen + `git diff --stat` check to the `validate` job
+   mirroring the existing US one.
+3. `bundle-win.sh`: bundle all three region exes + the launcher; keep the
+   "no ROM / size ≤ 60 MB" guards (raise the limit only if the interim frozen
+   converter is included).
+4. `release` job: unchanged shape (draft pre-release); update
+   `.github/release-notes.md` with the drop-in-ROM instructions + the
+   requirements/hash table.
+5. Add a CI check that runs `d43_emit.py --check-only` / a `d69` equivalent
+   against a **CI-only** ROM if one can be provided via encrypted secret;
+   otherwise this stays a manual pre-publish smoke test (the release job is
+   already a human-gated draft).
+6. Once Part B lands: a CI step that builds the C converter, runs it and the
+   Python converter on the same ROM (secret), and `cmp`s the outputs.
+
+## Code to be written
+
+| File(s) | Rough scope | Notes |
+|---|---|---|
+| `port/src/launcher/ge007_launch.c` + CMake target | ~150 lines | ROM detect, SHA-1 match vs 3 known hashes, `CreateProcess` region exe, message-box on failure |
+| `port/src/pcconvert/d43.c` | ~1000–1300 lines | C port of `tools_pc/d43_emit.py` incl. graph sim + validation |
+| `port/src/pcconvert/d69.c` | ~500–700 lines | C port of `tools_pc/d69_emit.py` |
+| `port/src/pcconvert/pcconvert.c` + `.h` | ~200 lines | orchestrator, stamp file, splash text, wired into `romdataInit()` |
+| CMake codegen for embedded metadata | ~40 lines CMake + generated `.c` | embeds `filelist.u.csv`, `file_resource_table.inc.c`, `ModelFileHeader.inc.c`, bg/stan `.inc.c` |
+| `tools_pc/bundle-win.sh`, `.github/workflows/ci.yml` | edits | matrix, 3 exes, launcher, notes |
+| SHA-1 helper (or reuse a vendored small SHA-1) | ~100 lines | for launcher + `pcconvert.stamp`; `romverify.c`'s n64cksum path is an alternative identity check |
+
+## ROM SHA-1 / region handling
+
+- Accept a ROM at `data/ge007..z64`, `data/*.z64` (any name — sniff
+  the header), or `baserom.{u,e,j}.z64`.
+- Verify: N64 magic + "GOLDENEYE" + "GE" + country byte (already done), then
+  SHA-1 against the README table. On a header-valid but hash-mismatched ROM
+  (bad dump, byte-swapped `.n64`/`.v64`, ROM hack): warn but allow a
+  `--skip-rom-check` / ini opt-out; byte-swapped images should be detected and
+  the user told to supply a true big-endian `.z64`.
+- Region is taken from the country byte (`E`→ntsc-final, `P`→pal-final,
+  `J`→jpn-final), which already drives `pcmodelsRegionForCountry()`.
+
+## Config / data location
+
+No change needed. `$S/` = `./data/` (or `/data/`). First run creates
+`ge007.ini`; Part B additionally creates `pcmodels-/`,
+`pccg-/`, and `pcconvert.stamp` there. Document that the user makes a
+`data/` folder next to the exe and drops the ROM in it (the launcher can
+offer to create it).
+
+## Open questions
+
+1. **Does PAL/JP `romassets` generation truly need nothing from the ROM?**
+   `docs/building.md` says it needs region-specific offsets; the code says
+   otherwise. Generate `e` and `j` in CI and diff against a
+   locally-ROM-generated copy to confirm before committing them.
+2. **Does `d43_emit.py` really use `scripts/filelist.u.csv` for all three
+   regions?** It hard-codes `FILELIST = "scripts/filelist.u.csv"` even when
+   `REGION` is `pal-final`/`jpn-final`. Either model file offsets are
+   US-identical across regions (plausible — models live in obseg which may not
+   shift) or PAL/JP model conversion is currently untested. Confirm before
+   promising PAL/JP support in the release.
+3. **Launcher vs self-re-exec vs three shortcuts** — pick one UX.
+4. **Splash UI for the first-run convert** — reuse the SDL window / a simple
+   progress line, or a pre-SDL Win32 dialog? ~5 s is short enough that a log
+   line + busy cursor may suffice.
+5. **Non-Windows.** Part B's C converter is portable; the launcher and
+   bundling are Windows-only today. Out of scope but the C port keeps the
+   door open.
+6. **Signing.** Unsigned exes + a "reads your ROM" first-run step will trip
+   SmartScreen. Not blocking, but worth a release-notes note.
+
+## Suggested milestone ordering
+
+1. Part A (3-region matrix + pre-generated `romassets` + launcher) — days.
+2. Interim frozen `ge007-convert.exe` wired to first-run — days, optional.
+3. Part B C port of `d43`/`d69` + embedded metadata + output diffing — weeks.
+4. Drop the interim binary; single-binary drop-in experience complete.
diff --git a/docs/index.md b/docs/index.md
index 5019d7ec..19697a13 100644
--- a/docs/index.md
+++ b/docs/index.md
@@ -13,7 +13,7 @@ graphics coprocessor in software. It follows the architecture of the
 
 It is also a **research project on AI-agent collaboration in a large,
 unfamiliar, low-level codebase**: two coding agents — a local open-weight model
-(Qwen 3) on a single RTX 5090, and Claude / Claude Code — driven by one person
+(Qwen 3.8) on a single RTX 5090, and Claude / Claude Code — driven by one person
 part-time through ~230 translation units of unmodified big-endian MIPS game
 code, handing work back and forth through shared written notes.