diff --git a/.github/release-notes.md b/.github/release-notes.md
index d75d907a..671179ad 100644
--- a/.github/release-notes.md
+++ b/.github/release-notes.md
@@ -15,8 +15,9 @@ This release is v0.5.0 plus frame rates above 60, an experimental macOS port
 and a batch of fixes.
 
 - **Frame rates above 60.** The *Frame rate cap* (`Video.FpsCap`) now offers
-  **90, 120 and 144**, and an **Auto** setting that matches your display's
-  refresh rate (up to 144). Auto is the default on fresh installs; an existing
+  **90, 120, 144, 165 and 240**, and an **Auto** setting that matches your
+  display's refresh rate (up to 240; 165 and 240 have had less testing than
+  90 to 144). Auto is the default on fresh installs; an existing
   `ge007.ini` keeps its value. Extra frames are blended between game ticks, so
   the game still simulates at its original rate. It costs one game tick of
   extra latency, does nothing on a 60 Hz display, and the *Original N64*
@@ -32,7 +33,14 @@ and a batch of fixes.
   native-widescreen picture on front-end screens, and the first launch opening
   on the wrong monitor (#151).
 - **Crash fixes:** the Linux end-credits crash (#152) and a crash when a guard
-  spawned wearing a hat (#153).
+  spawned wearing a hat (#153, #156).
+- **Original controls are now original (#145, #149).** With *Control style*
+  set to Original, a gamepad behaves as an N64 controller: South is A, West is
+  B, and the game decides what each button does, so B opens a door or reloads
+  (not both), A changes weapon, and B goes back in menus. *Invert look* works in
+  every original style and stays in sync with the watch's *Look up/down*.
+- **Mission timer (#144)**, the file-select crest drawing on a dark square
+  (#146), and a fix for a Windows exit crash.
 
 Checked on PC at 120 Hz with VSync on a G-Sync display, and on the Steam Deck
 at 90 fps.
@@ -43,12 +51,13 @@ at 90 fps.
 
 ### Frame rate
 
-- **Auto, 90, 120 and 144** join 30 and 60 in the F10 *Frame rate cap* row.
+- **Auto, 90, 120, 144, 165 and 240** join 30 and 60 in the F10 *Frame rate cap* row.
   Above 60 the port draws extra frames between game ticks by blending the
   previous and current frame. It only applies when the display is faster than
   60 Hz; on a 60 Hz display nothing changes.
-- A fresh install uses **Auto**, so a 90 Hz screen runs at 90 and a 144 Hz
-  monitor at 144. The *Original N64* display mode sets the cap to 60.
+- A fresh install uses **Auto**, so a 90 Hz screen runs at 90, a 144 Hz
+  monitor at 144 and a 240 Hz monitor at 240 (faster displays hold at 240).
+  The *Original N64* display mode sets the cap to 60.
 - A fast camera turn, or a large change in what is on screen, falls back to a
   full frame instead of a blend.
 - The Steam Deck at 90 is good on most levels (see the known issues).
@@ -59,12 +68,32 @@ at 90 fps.
   automatically by CMake). The 64-bit address fixes this needed (textures,
   input, animation and multiplayer crashes) are in the shared code and have no
   effect at base 0 on Windows and Linux.
-- Build from source with Homebrew's GNU `gcc`; there is no download. The
-  x86_64 macOS code (Julio C. Rocha's #88) is in the tree but was never
+- There is a `macos-arm64` download, or build from source with Homebrew's
+  GNU `gcc`. The x86_64 macOS code (Julio C. Rocha's #88) is in the tree but was never
   verified on Intel hardware.
 
+### Controls
+
+- **Original control style = an N64 controller** (D643). Every press sends the
+  real N64 button in levels, menus, the watch and the tank; no port-only
+  actions are added. Default layout: South = A, West = B, right trigger = Z,
+  left trigger / RB = R, LB = L, right stick = C-buttons, D-pad and Start as
+  is. Each N64 button can be moved per player in `ge007.ini`
+  (`Input.N64Pad.A = a`, `Input.N64Pad2.B = x`, ...). `Input.N64StickFeel = 1`
+  (default in Original) gives the N64's linear stick range; 0 keeps the
+  previous feel. The Modern style and keyboard/mouse are unchanged.
+- **Invert look** reaches every original style (D605) and, in Original, is the
+  same setting as the watch's *Look up/down* (D645).
+
 ### Settings and diagnostics
 
+- **Mission timer** (D606, #144): an optional always-visible timer, off by
+  default (F10).
+- F10 dropdowns stay inside the settings box instead of covering the help
+  text (D646).
+- Windows: per-monitor DPI awareness, so fullscreen on a scaled 4K display is
+  native resolution (D626); exclusive fullscreen uses the desktop mode (D639).
+
 - The old front-end **Velocity** pointer mode is removed; the retired ini keys
   are ignored quietly (D574).
 - The Linux/macOS **update check** no longer fails on a response over 64 KB
@@ -120,6 +149,17 @@ the folder:
 xattr -dr com.apple.quarantine goldeneye-pc-port-<version>-macos-arm64
 ```
 
+Or download it in Terminal, which skips the block entirely (`curl` does not
+mark the file for Gatekeeper):
+
+```
+curl -L https://github.com/jkdansereau/goldeneye-pc-port/releases/download/v<version>/goldeneye-pc-port-<version>-macos-arm64.tar.gz | tar xz
+```
+
+Without Terminal: after the first launch is blocked, open System Settings →
+Privacy & Security and click **Open Anyway** next to the `ge007.aarch64`
+message.
+
 I have only run it on one M3 Mac, so please report what Mac and macOS you
 tried and what happened.
 
@@ -139,6 +179,7 @@ Full steps, controls and troubleshooting are in the bundled `README.md`.
 | Jungle can feel slightly less smooth at 90 fps than other levels | Feel | Frame cap 60 |
 | Cradle's turret explosions may draw wrongly | Cosmetic; reported, not yet investigated | None |
 | After alt-tabbing away, the picture can look hazy until it settles; fast repeated alt-tab is wonkier and takes longer | Cosmetic, Windows | It recovers on its own; alt-tab again or give it a couple of seconds |
+| macOS: the end credits may crash | Ending only, macOS only | Your progress is already saved by then |
 | macOS (Apple Silicon) is experimental, tested on one Mac, and not notarized | Platform | Run `xattr -dr com.apple.quarantine <folder>` before the first launch; reports welcome |
 | PAL and JP ROMs aren't supported in release packages | NTSC-U only | Use an NTSC-U ROM. Both regions convert, build and boot from source; packaging is still open |
 | Changing aspect ratio inside a level can briefly glitch the gun/hand model, rarely | Cosmetic, one-off | Change the ratio from the front-end PC Options, or accept it |
@@ -164,9 +205,19 @@ workarounds, is the
   cleared (D579 follow-up).
 - **#151** the first launch opens on the primary monitor (D581).
 - **#152** Linux end-credits crash: a garbage text pointer is rejected (D582).
-- **#153** hat-spawn crash: the object record type was read from the wrong byte
-  on a little-endian machine (D580).
+- **#153, #156** hat-spawn crash: the object record type and the end-of-list
+  marker were read from the wrong byte on a little-endian machine, so the search
+  could run past the end of the list (D580).
 - **#108** SFX / text-bank pointer guards (D573).
+- **#145 / #149** Original control style as an N64 controller (D643); Invert
+  look in the original styles (D605, D645).
+- **#144** mission timer (D606).
+- **#146** file-select crest on a dark square: 4-bit IA textures now decode as
+  the N64 does (D644).
+- **Exit crash (BSOD 0x119) on some Windows GPUs:** the GL context and window
+  are released before the process exits (D632).
+- Windows DPI awareness (D626), exclusive fullscreen mode (D639), F10 dropdown
+  placement (D646).
 - Velocity removal, update-check fix, `ge007.log`, HUD-150 sliders, preset
   scope (D574-D577, D440).
 - **macOS arm64:** one address-model header, the 64-bit pointer fixes and a
@@ -174,10 +225,12 @@ workarounds, is the
 
 ## Verification
 
-I did not re-run the three-platform reference-frame gate for this release; the
-last full run was for v0.5.0. For v0.6.0 I checked the frame-rate path and the
-fixes above by playing on Windows and the Steam Deck, and the macOS build with
-a 21-level sweep and live sessions.
+Windows: the 21-level reference-frame sweep passes (Streets sits at the edge of
+its tolerance and flips run to run with or without this release's changes).
+Steam Deck and macOS: all 21 levels load, play and quit without a crash (their
+reference frames are not yet comparable, so no pixel check). The fixes above
+were also checked by playing on Windows, with the original controls compared
+against the issue reports rather than an N64 pad.
 
 ## Thanks
 
diff --git a/.gitignore b/.gitignore
index 8869a7e7..c1978f1d 100644
--- a/.gitignore
+++ b/.gitignore
@@ -205,3 +205,11 @@ tools_pc/sweep-captures/**/*.ppm
 # referenced by no project tooling. Never commit.
 /Python/
 docs/media/*.mp4
+
+# mod-loader tool build artifacts
+/tools_pc/modpatch/modpatch.exe
+
+# demo harness data (tools_pc/demobatch.sh): personal recordings + run output
+demos/
+demobatch/
+levelbench/
diff --git a/NOTICE b/NOTICE
index ef7dcf36..c93f978a 100644
--- a/NOTICE
+++ b/NOTICE
@@ -10,6 +10,9 @@ repository:
   * the port layer            port/
   * the PC build system        CMakeLists.txt, cmake/, build-pc.sh
   * PC-port tooling            tools_pc/
+  * macOS packaging            scripts/package-macos.sh, scripts/gen_macho_syms.py,
+                               scripts/strip_weak_pragmas.py
+  * agent-development tooling  pi/ (not part of the game build)
   * documentation              docs/, README.md
 
 Third-party code vendored into the port layer keeps its own license:
@@ -18,6 +21,17 @@ Third-party code vendored into the port layer keeps its own license:
     libultraship / sm64-port lineage; see port/fast3d/LICENSE.txt.
   * port/fast3d/glad/ - generated OpenGL loader (glad), public domain / MIT.
 
+Runtime libraries bundled in the downloads (not in this repository) keep
+their own licenses; each license text ships in the download's licenses/ folder:
+
+  * SDL2 (all platforms) and SDL3 (macOS, loaded by sdl2-compat) - zlib license.
+  * zlib - zlib license.
+  * Windows: the MinGW-w64 runtime (libwinpthread, libgcc, libstdc++).
+  * macOS: the GCC runtime (libgcc_s, libstdc++) - GPL-3.0 with the GCC
+    Runtime Library Exception 3.1. The GPL-3.0 text the macOS package ships
+    is tools_pc/dist/GPL-3.0.txt (the Homebrew gcc keg carries only the
+    exception text and GPL-2.0).
+
 Vendored build tooling keeps its own license (none of it is compiled or linked
 into the PC port; it belongs to the N64 decompilation build):
 
diff --git a/README.md b/README.md
index 98eea0f3..77d460ba 100644
--- a/README.md
+++ b/README.md
@@ -6,7 +6,7 @@
 ![license](https://img.shields.io/badge/license-MIT-ffb454)
 
 **GoldenEye 007 (Nintendo 64, 1997) as a native PC game for Windows, Linux
-and Steam Deck.** Native widescreen, 60 fps or higher (90/120/144 on fast
+and Steam Deck.** Native widescreen, 60 fps or higher (up to 240 on fast
 displays), 2–4 player split-screen, mouse and keyboard or a modern controller, rebindable controls
 and an in-game options menu. No
 emulator, and not a binary recompilation: the game's own
@@ -17,7 +17,7 @@ for your PC and running from your own ROM.
 **[Download the latest release](https://github.com/jkdansereau/goldeneye-pc-port/releases)**,
 then follow the [Quick start](#quick-start).
 
-[Quick start](#quick-start) · [Controls](#controls) · [Features](#features) · [Settings and files](#settings-and-files) · [Troubleshooting](#troubleshooting) · [Roadmap](#roadmap) · [Building](#building-from-source) · [Legal](#legal)
+[Quick start](#quick-start) · [Why this port](#why-play-this-port) · [Controls](#controls) · [Features](#features) · [Settings and files](#settings-and-files) · [Troubleshooting](#troubleshooting) · [Roadmap](#roadmap) · [Building](#building-from-source) · [Legal](#legal)
 
 > [!IMPORTANT]
 > **You must supply your own GoldenEye 007 ROM.** This repository and its
@@ -49,10 +49,10 @@ project is as much a study of that process as it is a port. See
 
 ## News
 
-- **RELEASE-DATE** — **v0.6.0** ([release notes](https://github.com/jkdansereau/goldeneye-pc-port/releases/tag/v0.6.0)):
-  - **Frame rates above 60.** `Video.FpsCap` now offers 90, 120 and 144, and
-    an **Auto** setting (the new default on fresh installs) that matches your
-    display's refresh rate. Frames between game ticks are blended from the
+- **2026-10-13** — **v0.6.0** ([release notes](https://github.com/jkdansereau/goldeneye-pc-port/releases/tag/v0.6.0)):
+  - **Frame rates above 60.** `Video.FpsCap` now offers 90, 120, 144, 165 and
+    240, and an **Auto** setting (the new default on fresh installs) that
+    matches your display's refresh rate, up to 240. Frames between game ticks are blended from the
     previous and current frame; the game still simulates at its original
     rate. It costs one game tick of extra latency, does nothing on a 60 Hz
     display, and the *Original N64* preset caps at 60. This is a rework of
@@ -66,9 +66,14 @@ project is as much a study of that process as it is a port. See
     for the one-line unblock. Test reports are very welcome.
   - **Fixes:** geometry vanishing when you press against a wall (#150),
     the first launch opening on the wrong monitor (#151), a Linux crash in the
-    end credits (#152), a crash when a guard spawns with a hat (#153), a
+    end credits (#152), a crash when a guard spawns with a hat (#153, #156), a
     sound/text pointer check that rejected valid pointers (#108), a Cradle
     catwalk shadow flicker, and stale watch pixels in native widescreen.
+  - **Original controls (#145, #149):** with *Control style* set to Original,
+    a gamepad now behaves as an N64 controller (South = A, West = B, the game
+    decides what each does), and *Invert look* works in every original style.
+    Also an optional mission timer (#144), the file-select crest no longer
+    drawing on a dark square (#146), and a Windows exit-crash fix.
   - **Smaller changes:** the legacy front-end Velocity pointer mode is gone,
     the Linux/macOS update check no longer fails on a large response, every
     run is also logged to `ge007.log` next to the ini, and F10 slider labels
@@ -201,6 +206,22 @@ one pad pick a 1.x style. Saves made with earlier builds may hold 1.2 (the port
 forced style back into the save), so when you switch to the Original control style, pick
 your style on the watch once.
 
+In the Original control style the pad is a strict virtual N64 controller in
+every context (stage, menus, watch, tank): each press sends the native N64
+button and the game decides what it means (so *B* uses a door or reloads, as
+on the console). Default layout, remappable per player in the ini
+(`Input.N64Pad.<Button>`, `Input.N64Pad2..4` for players 2-4; up to two pad
+buttons each) and with `Input.N64StickFeel` (1 = linear +/-80 stick with a
+small dead centre, 0 = the Ext deadzone):
+
+| Pad | N64 | Pad | N64 |
+|---|---|---|---|
+| South (A) | A | Right stick | C buttons |
+| West (X) | B | Left stick | Analog stick |
+| Right trigger | Z | D-pad | D-pad |
+| Left trigger / RB | R | Start | Start |
+| LB | L | Back | PC options overlay |
+
 **Rebinding:** *Key Bindings* in the options overlay rebinds the
 keyboard and mouse (one list: movement, then actions). Controller buttons are
 rebindable per controller under *Controller* (pick a player first): set *Layout preset* to
@@ -225,6 +246,47 @@ left/right adjusts the selected value, **A** opens a page or toggles a row,
 again) closes. With a keyboard, use `F10`, the arrow keys, Enter and
 Backspace.
 
+## Why play this port
+
+The questions people ask most about playing GoldenEye today, and how this
+port answers them:
+
+- **Real 60 fps, not a speed hack.** The game's logic counts time in 60 Hz
+  ticks. On the N64 it only drew a frame every 2–3 ticks (20–30 fps); here it
+  draws every tick, so you get 60 fps with the game, physics and AI running
+  at the N64's own speed. There is no CPU overclock or frame-skip patch involved,
+  and the few spots in the game that behaved differently at one tick per frame
+  are fixed to match the N64. Above 60 (90 to 240, or Auto), extra
+  frames are blended between ticks for smoothness, at the cost of one tick of
+  latency.
+- **Aiming that doesn't feel floaty.** *Aim style* picks the behavior.
+  **Centered** pins the crosshair in the middle, and the mouse or the right
+  stick turns the camera directly, like a modern shooter. **Original** keeps
+  the N64's crosshair that moves a little before the camera follows. The stick
+  has separate horizontal and vertical sensitivity, separate deadzones for each
+  stick, look smoothing and a trigger threshold. The mouse has raw input, an aim
+  curve, smoothing, and separate turn and aim-mode sensitivities.
+- **Dual analog and mouse and keyboard, built in.** No emulator remapping. The
+  right stick looks, the triggers fire and aim, and every button is rebindable,
+  for the keyboard, the mouse and each controller. Rare's Xbox control styles are
+  presets. If you want the N64 exactly, the **Original** control style turns the
+  pad into a strict virtual N64 controller.
+- **The whole game.** All 20 solo missions and the credits, with every
+  difficulty, 007 mode and the cheats, plus 2–4 player split-screen on every
+  multiplayer map. Your emulator saves (1964, Project64) load directly.
+- **Faithful by construction.** It is the game's own reconstructed source
+  compiled natively, not an emulator and not a remake, so levels, guards and
+  weapons behave the way the N64 cartridge does. The **Original N64** display
+  preset switches every visual improvement off at once.
+- **Modern display options:** native widescreen at any aspect ratio, high
+  resolutions, MSAA, adjustable FOV and draw distance, and HUD scaling.
+- **Windows, Linux, Steam Deck and (experimental) macOS** from one download
+  page, with an in-game options menu on F10 or the pad's Select button.
+
+What it doesn't do (yet): bots, online play, new or enlarged multiplayer maps,
+mod support, and PAL/JP ROMs are not part of this release. See the
+[Roadmap](#roadmap).
+
 ## Features
 
 - **2–4 player split-screen multiplayer** on every MP map, with the N64's own
@@ -238,8 +300,8 @@ Backspace.
   and *Aspect ratio → Original* shows the exact 4:3 frame.
 - **A steady 60 fps**, with VSync and a frame cap (30 or 60), including on
   low-end GPUs.
-- **Frame rates above 60:** 90, 120 or 144 fps, or **Auto** to match your
-  display. Extra frames are blended between game ticks (the game still runs at
+- **Frame rates above 60:** 90, 120, 144, 165 or 240 fps, or **Auto** to match
+  your display (up to 240). Extra frames are blended between game ticks (the game still runs at
   its original rate), at the cost of one tick of latency. It has no effect on
   a 60 Hz display, and the *Original N64* preset keeps 60. See
   [framerate and pacing](docs/framerate-and-pacing.md).
@@ -355,7 +417,7 @@ MasterVolume = 100
 ```
 
 `Fullscreen` 1 is borderless fullscreen; `FpsCap` -1 is Auto (match the display,
-up to 144), 0 is uncapped, otherwise a fps cap (30, 60, or 90/120/144 for frame
+up to 240), 0 is uncapped, otherwise a fps cap (30, 60, or 90 to 240 for frame
 interpolation on a faster display; an existing ini keeps its value); `TextureFilter` is 0 nearest,
 1 bilinear, 2 the N64 3-point filter, 3 trilinear; `FovScale` and `DrawDistance` are
 percentages of the original (the options overlay shows the field of view in
@@ -555,6 +617,17 @@ it comes from a download. Unpack it, then run `xattr -dr com.apple.quarantine <f
 xattr -dr com.apple.quarantine goldeneye-pc-port-<version>-macos-arm64
 ```
 
+**Or skip the quarantine step** by downloading in Terminal instead of a browser
+(`curl` does not mark the file, so Gatekeeper never checks it):
+
+```sh
+curl -L https://github.com/jkdansereau/goldeneye-pc-port/releases/download/v<version>/goldeneye-pc-port-<version>-macos-arm64.tar.gz | tar xz
+```
+
+**Without Terminal:** after macOS blocks the first launch, open System Settings →
+Privacy & Security, scroll down to the message about `ge007.aarch64` and click
+**Open Anyway** (macOS 15 removed the old right-click → Open bypass).
+
 Then set up `data/` with your ROM as in the package README and run
 `./ge007.aarch64` from that folder in Terminal. If you launch it first, the process
 stalls before the window appears (it never reached `main` in my test; that is
@@ -726,7 +799,7 @@ models used, and the harnesses that drove them:
 This is a non-commercial fan preservation/research project, in the same
 category as the many other N64 decompilation and native-port repositories on
 GitHub. It is **not affiliated with, endorsed by, or sponsored by** Nintendo,
-Rare, Microsoft, Valve, MGM, Danjaq, EON Productions, or any rights holder in
+Rare, Microsoft, Valve, Apple, MGM, Danjaq, EON Productions, or any rights holder in
 GoldenEye or James Bond. "GoldenEye 007", "007", "James Bond" and related
 marks belong to their respective owners; they are used here only to say which
 game this port runs. No official logos, box art, or marketing assets are used.
@@ -734,8 +807,10 @@ game this port runs. No official logos, box art, or marketing assets are used.
 ### What the download contains
 
 - **The engine:** the `port/` layer plus the compiled decompilation.
-- **Runtime libraries** under permissive licenses (SDL2, zlib, the MinGW
-  runtime); their licenses travel in the download.
+- **Runtime libraries:** SDL2 (and SDL3 on macOS) and zlib under the zlib
+  license, the MinGW runtime on Windows, and the GCC C++ runtime on macOS
+  (GPL-3.0 with the GCC Runtime Library Exception); their licenses travel in
+  the download.
 - **The one-time asset tool**, which reads your ROM on your machine.
 
 It contains **no ROM and no game assets of any kind**. Any build, yours or a
@@ -776,4 +851,4 @@ is covered by [`NOTICE`](NOTICE), not by that license.
 
 ---
 
-*Last updated 2026-10-10 — v0.6.0 · <a href="https://github.com/jkdansereau/goldeneye-pc-port">GitHub</a> · <a href="https://github.com/jkdansereau/goldeneye-pc-port/releases">Releases</a>*
+*Last updated 2026-10-13 — v0.6.0 · <a href="https://github.com/jkdansereau/goldeneye-pc-port">GitHub</a> · <a href="https://github.com/jkdansereau/goldeneye-pc-port/releases">Releases</a>*
diff --git a/docs/ROADMAP.md b/docs/ROADMAP.md
index d49b4da7..225f4d7a 100644
--- a/docs/ROADMAP.md
+++ b/docs/ROADMAP.md
@@ -155,7 +155,7 @@ below).
 | Dam: grate decals vanish behind a sliding edge, wall lighting shifts with view; guard tower window/doorway view of the outside glitches | closed 2026-10-09: faithful | D583 | Present with every present exact (`allexact`) and in v0.5.0; maintainer checked 1964 (GEPD): the exact same behaviour. Original-game portal/culling artefacts, not port bugs. Candidate for the opt-in fixes row below. |
 | Opt-in fixes for original-game rendering bugs (QoL) | idea, post-1.0 (maintainer 2026-10-09) | D583 | Maintainer: a nice QoL feature would be optionally fixing artefacts the original game has, e.g. the Dam wall grates cut by a sliding portal edge and the guard-tower window/doorway outside view (D583 row above), and overlapping 2D transparent textures on Cradle sorting wrongly (also in the original). Must stay off by default (fidelity is the default) and live in `port/` or an explicit option; game-logic changes would need Rule-2 sign-off. |
 | Exclusive fullscreen picks the wrong resolution with NVIDIA "OpenGL present method = Prefer native" | open (v0.7.0, maintainer 2026-10-10) | D583 | Seen 2026-10-09 while bisecting D583; borderless is fine. Probably exclusive mode switches to the ini Width/Height instead of keeping the desktop mode. Not investigated. |
-| Deck golden captures are 1280x800 when interpolation is on | open (v0.7.0, maintainer 2026-10-10) | D578, D283 | With Auto FpsCap on the Deck's 90 Hz screen, GE_PCDUMP writes 1280x800 frames against the 640x480 Deck goldens, so every frame fails to compare (and the sweep prints PASS with "stems don't match", which is not a pass). `verify.sh` now pins `Video.DeckPresetApplied = 1` (D283 preset). Fix: dump at the window size, or re-capture the Deck goldens once. |
+| Deck golden captures are 1280x800 when interpolation is on | open (v0.7.0, maintainer 2026-10-10) | D578, D283 | With Auto FpsCap on the Deck's 90 Hz screen, GE_PCDUMP writes 1280x800 frames against the 640x480 Deck goldens, so every frame fails to compare (and the sweep prints PASS with "stems don't match", which is not a pass). **2026-10-11:** the false PASS is fixed: `framediff.py` exits 3 on a size mismatch and `verify.sh` reports it as REGRESSION ("not compared"); the capture-size fix itself is still owed. `verify.sh` now pins `Video.DeckPresetApplied = 1` (D283 preset). Fix: dump at the window size, or re-capture the Deck goldens once. |
 | Crash on save slots 1/3/4, plus a host BSOD (fltmgr.sys) | decided | #94 | Needs a repro. Two maintainer retest requests (v0.3.0, v0.4.0) got no reply. Likely D299/D301 (save-slot crash, fixed) and D344 (BSOD on exit, fixed). **Decided 2026-10-01: close as stale when the release ships**, pointing at D299/D301/D344; no GitHub activity before then. |
 | Authored mip chains not sampled for explicit-LOD textures | post-1.0 | D323 | Distant textures use driver mips instead of Rare's hand-authored chain (heaviest on Statue/Aztec). Parked as a fidelity enhancement. **Decided 2026-10-01:** after 1.0, together with the LOD-fraction approximation as one "real RDP LOD" project. |
 | Approximations in the software RSP | post-1.0 | `gfx_pc.cpp` `G_CCMUX_LOD_FRACTION`; D323 | The LOD-fraction CC input is an eyeballed approximation, bundled with D323 as one "real RDP LOD" project after 1.0 (2026-09-30 triage). **WrapFix settled 2026-10-02:** the maintainer checked the Depot building ceilings against 1964 with the default `Video.WrapFix = 0` and they look right, so the default stays off (the opt-in knob remains; RC3/D167). |
@@ -195,7 +195,7 @@ A single campaign pass with these on the checklist would clear most of them.
 | Steam Deck preset skipped when first launched from Desktop Mode | verify | D283 | Fixed on `release/v0.4.1`: the Deck is detected by DMI hardware id (Valve Jupiter/Galileo; `STEAMOS` as a second signal) and the preset applies once per ini, only while Fullscreen/Maximized are untouched. Only a Windows build and a Linux `-fsyntax-only` check ran. **Owed: one run on a real Deck** (Desktop-Mode-created ini, then Game Mode; expect fullscreen 1280x800 and the log line `video: Steam Deck (DMI Valve ...)`). The Known-issues row goes when this ships. |
 | Windows 7 | open | #98 | **Re-checked 2026-10-02: likely feasible around 1.0.** The exe imports no Windows 8+ functions and the bundled SDL2 2.32.10 still supports Windows 7. The real blocker is the first-run converter `ge007-convert.exe` (PyInstaller, Python 3.9+): the native C converter (#6 Part B) removes it, or a separate Python 3.8 build. Work: one Windows 7 VM test, then manual per-release checks (CI cannot test it). Supersedes the earlier 'won't-fix' call. |
 | Alt-tab haze + fast alt-tab recovery | open (v0.7.0, maintainer 2026-10-10) | D578 | Maintainer 2026-10-10 (v0.6.0 build, Windows 120 Hz): alt-tabbing leaves a hazy picture; fast repeated alt-tab is wonky and recovers only after a while or more alt-tabbing. Not new in v0.6.0 (was an owed check). Known issue in the v0.6.0 notes. QoL fix: clean focus-loss/regain path (context/FBO state, present queue reset). |
-| Windows display scaling: the game is not DPI-aware on scaled displays | open (v0.7.0, maintainer 2026-10-10) | — | Seen 2026-10-07 during the v0.5.0 4K captures: on a 3840x2160 display at 125% Windows gives the game a 3072x1728 surface and stretches it, so the picture is softer than native. `gfx_sdl2.cpp` sets `SDL_HINT_WINDOWS_DPI_AWARENESS` only when `allow_hidpi` is off; setting it (`permonitorv2`) on the HiDPI path too renders at native resolution. Workaround today: environment variable `SDL_WINDOWS_DPI_AWARENESS=permonitorv2`. |
+| Windows display scaling: the game is not DPI-aware on scaled displays | fixed 2026-10-10 (D626); maintainer by-eye on a scaled display owed | D626 | Seen 2026-10-07 during the v0.5.0 4K captures: on a 3840x2160 display at 125% Windows gives the game a 3072x1728 surface and stretches it, so the picture is softer than native. `gfx_sdl2.cpp` sets `SDL_HINT_WINDOWS_DPI_AWARENESS` only when `allow_hidpi` is off; setting it (`permonitorv2`) on the HiDPI path too renders at native resolution. Workaround today: environment variable `SDL_WINDOWS_DPI_AWARENESS=permonitorv2`. |
 
 ## 5. Modern-remaster layer (port features beyond the N64)
 
diff --git a/docs/_layouts/default.html b/docs/_layouts/default.html
index b018e711..09c960cc 100644
--- a/docs/_layouts/default.html
+++ b/docs/_layouts/default.html
@@ -55,7 +55,7 @@
     <footer class="site-footer">
       <div class="footer-stamp">
         <a href="https://github.com/jkdansereau">maintained by jkdansereau</a>
-        &nbsp;&middot;&nbsp; last updated: 2026-10-10 &middot; v0.6.0
+        &nbsp;&middot;&nbsp; last updated: 2026-10-13 &middot; v0.6.0
         &nbsp;&middot;&nbsp; <a href="https://github.com/jkdansereau/goldeneye-pc-port">GitHub</a>
         &nbsp;&middot;&nbsp; <a href="https://github.com/jkdansereau/goldeneye-pc-port/releases">releases</a>
       </div>
diff --git a/docs/dev/RELEASE-GATE.md b/docs/dev/RELEASE-GATE.md
new file mode 100644
index 00000000..c126ab7c
--- /dev/null
+++ b/docs/dev/RELEASE-GATE.md
@@ -0,0 +1,21 @@
+# Release gate driver
+
+`tools_pc/release_gate.sh` runs the automatable pre-release checks and prints one
+PASS / FAIL / SKIPPED table. It wraps `verify.sh` (pixel gate) and `levelbench.sh`
+(scripted run, clean-quit and crash-log detection); every run ends through
+`GE_QUITFRAME`, nothing is hard-killed.
+
+```sh
+tools_pc/release_gate.sh --dry-run all       # print what would run
+tools_pc/release_gate.sh win                 # sweep vs win goldens + fire smoke + exit path
+tools_pc/release_gate.sh deck <host>         # 21 levels via deckrunner (shortcut must be up)
+GE_MAC_DIR=<remote game dir> tools_pc/release_gate.sh mac <host>
+```
+
+Options: `--secs N` (per-level run length, default 30), `--out DIR`. Hosts and
+paths come only from arguments or env (`GE_WIN_DIR`, `GE_DECK_HOST`,
+`GE_MAC_HOST`, `GE_MAC_DIR`, `GE_MAC_EXE`, `GE_MAC_BUILD_CMD`,
+`GE_GATE_SSH_OPTS`). The Deck pixel compare is reported SKIPPED (640x480
+goldens vs 1280x800 dumps); Mac has no goldens, so it is crash-free only. The
+fire smoke is `GE_INPUTSCRIPT` `MLB1`/`MLB0` (D597, the D595 regression).
+Remote job, result and script files are deleted after the summaries are pulled.
diff --git a/docs/dev/findings-index.csv b/docs/dev/findings-index.csv
index ee23750f..84b8923c 100644
--- a/docs/dev/findings-index.csv
+++ b/docs/dev/findings-index.csv
@@ -325,6 +325,8 @@ D579,FIXED,Issue #150: hugging a wall makes geometry vanish (N64 native widescre
 D583,FIXED,120 fps + VSync on G-Sync/VRR:,"FIXED on PC 2026-10-09 (G-Sync OLED, maintainer by eye); Deck 90 good on most levels (per-level pass 2026-10-10); open:"
 D584,FIXED,Cradle catwalk shadow flicker at every frame rate (regression from D579).,FIXED 2026-10-10 (maintainer Deck:
 D614,OPEN,Deck Bunker fight texture-cache thrash:,"OPEN: measured, fix not shipped (2026-10-10)"
+D605,FIXED,"Issue #145 (v0.6.0): F10 ""Invert look"" never reached the look stick in the HONEY role (game styles 1.1/1.3 -- left stick looks, right stick is the digital move…",FIXED 2026-10-09 (v0.6.0 line; port-only:
+D606,FIXED,Issue #144 (v0.6.0 triage):,"FIXED 2026-10-09 (v0.6.0 line; port-only in port/src/optionsoverlay.c, src/game zero-diff; off by default so goldens stay byte-identical; smoke-verified:"
 D407,CLOSED,Tanks cannot be boarded/exited on PC —,CLOSED (2026-09-28 bookkeeping:
 D408,FIXED,Game.SkipIntro skips the post-mission failure dossier:,FIXED 2026-09-30 (SkipIntro hook gated to first boot; live dossier check owed).
 D415,FIXED,"Shooting a light fixture does not ""kill"" its flickering light (Bunker, Caverns):",FIXED (2026-09-29 bookkeeping:
@@ -439,6 +441,27 @@ D534,FIXED,Pre-release review of the 2026-10-05 local batch —,FIXED 2026-10-05
 D540,FIXED,Fog rework (Deck playtest):,FIXED 2026-10-05 (port:
 D450,FIXED,"Unplugging and replugging a controller was not picked up (user, GameSir-G7 Pro; 2026-09-29).","FIXED — port-only: the render thread's gfx_sdl_handle_events drains the same SDL queue as video.c's host pump but had no case for SDL_CONTROLLERDEVICEADDED/REMOVED, so it usually swallowed them and i…"
 D436,FIXED,"Persistent (hipfire) crosshair, opt-in:",FIXED (shipped 2026-09-29; rule-2 sign-off:
+D617,OPEN,"The HD pack's ge007.tdb is the emulator's full per-draw key list, not the pack's PNG list —",OPEN — plan written (docs/dev/TEXPACK-RICE-DRAW-PLAN.md); whole-texture recipe confirmed correct by 9 tdb hits with exact ;WxH agreement; PNG-filename and brute-force param searches are dead ends
+D618,?,"The community HD pack's non-CI keys are per-draw rects partly unreachable from N64 tile state (440x1 strips), and a rect key cannot land in the whole-texture p…",STOPPED 2026-10-06 — tdb ;WxH are per-draw rects (440x1 ×301 over tmem's 128-byte pitch = recompile-side); ~100 byte-domain variants score 0 hits; content-join does not discriminate; rect-keyed repla…
+D619,?,"The community HD pack's remaining 1790 PNGs are blocked by four independent causes (recompile-side rect keys, wrong hash domain, rect-vs-whole-texture format,…",WRITTEN UP 2026-10-06 —
+D620,OPEN,The HD pack's non-CI keys hash the pack's own re-encoded content —,MEASURED 2026-10-07 —
+D621,PARTIAL,The PD dabs-mod texture-pack notes mine:,PLAN LANDED 2026-10-07 —
+D622,?,B0 sprite-bank confirmation —,"PASS 2026-10-07 — two human-driven GE_TEXNUMDUMP+GE_TEXDUMP runs (front end 377 unique texnums / 521 dumps, -level_09 combat 321 / 316; evidence scratch/b0/):"
+D623,OPEN,Phase B1–B3 measured:,MEASURED 2026-10-07 —
+D624,OPEN,B4 content scorer built + tuned:,"BUILT 2026-10-07 — 5-stage pipeline (geometric prefilter → content rank → GT-tuned auto-accept → vision judge → human queue) per the 2026-10-07 decision (commits 512e6d57, 95d38637)."
+D625,?,"texpack A5: downscale-to-GL-max instead of the D615 vanilla fallback, plus first-use decode-stall metering —","BUILT 2026-10-10, not yet A/B-verified —"
+D626,FIXED,Windows DPI awareness was never declared:,FIXED 2026-10-10 (gfx_sdl2.cpp:
+D627,FIXED,DexDrive / GameFAQs .n64 saves import directly:,"FIXED 2026-10-10 (eepimport.c + eep_convert.py dexdrive-to-pc); byte-identical to the reference extraction, malformed inputs rejected untouched; by-eye mission select owed"
+D628,RESOLVED,"The exact-CRC CI half of the GoldenEye-007-HD join is sound (256 texnums, independently re-verified), but its structural half was scored against GE_TEXDUMP cap…","CI HALF ACCEPTED 2026-10-10 (256/256 (crc,fmt,siz) agreement with the rice census, 256 distinct texnums, palette-variant collision resolved by ROM palcrc, builder resolves all 256 sources); structura…"
+D632,OPEN,Exit BSOD 0x119 recurred (2026-10-10) despite D344; the first minidump shows it fires in the kernel's process-exit GPU cleanup (DXGPROCESS::Destroy -> DXGCONTE…,FIX BUILT 2026-10-10 (fix/d632-exit-teardown):
+D636,FIXED,Non-CI pack textures ARE fingerprint-keyed:,FIXED 2026-10-10 (feat/texpack-fp):
+D637,?,Fingerprint fallback in the loader:,BUILT 2026-10-10 (feat/texpack-fp); C CRC == Python on 80 ROM glyphs; not run-verified (game runs on hold)
+D638,FIXED,HD-pack ammo icon cut off:,FIXED 2026-10-10 (chore/v070-agentic-batch):
+D639,OPEN,Exclusive fullscreen picked the wrong resolution:,FIX BUILT 2026-10-10 (chore/v070-agentic-batch):
+D643,FIXED,Issues #145 (b)(c) / #149:,FIXED 2026-10-12 (port-only; GE_PADMAPTEST PASS; maintainer pad pass vs the #145/#149 reports; 1:1 N64-pad stick check owed when an adapter is available)
+D644,FIXED,File-select royal crest drew on an opaque dark square (#146):,"FIXED 2026-10-12 (import_texture_ia4; 1964-confirmed symptom, maintainer-verified; Win sweep unchanged)"
+D645,FIXED,"Original 1.1: F10 ""Invert look"" did not invert vertical look (#145 a):",FIXED 2026-10-12 (F10 row and watch row synced in Original; maintainer-verified)
+D646,FIXED,"F10 dropdown popup drew over the hint/tooltip panel (popup bounded by the screen edge, not the settings box) —",FIXED 2026-10-12 (ddGeom bounded to the box; maintainer-verified)
 D607,OPEN,Address model unified into one header:,DONE 2026-10-07 (macos branch; Windows identity gate:
 D608,FIXED,D34 ANIM_DATA_x re-basing guard in assets/animationtable_data.h was __x86_64__-only:,"FIXED 2026-10-07 (macos branch; mechanical guard widening, arm64 link clean)"
 D609,FIXED,macOS arm64 SIGKILL at exec:,"FIXED 2026-10-07 (macos branch; probe-confirmed; arm64 keeps the 4 GiB pagezero, floor 11.0)"
diff --git a/docs/dev/findings.md b/docs/dev/findings.md
index 35e25476..1688c31a 100644
--- a/docs/dev/findings.md
+++ b/docs/dev/findings.md
@@ -733,6 +733,8 @@ covers D24–D69; the log continues in §H (D32 procedure, D70–D121).
 | D583 | **120 fps + VSync on G-Sync/VRR: fullscreen black frames + OLED VRR flicker from the interpolation present scheduler.** Black frames fixed (VSync presents wait for their slot time; slot grid uses the reported refresh, not the measured swap gap). The OLED flicker was the present cadence (8% of ticks got 1 slot, early/late gap pairs, dropped slots); fixed by one continuous present grid snapped to tick/N, an 8-slot present ring (no drops), a capped draw-time peak for the ready time, and showing a waiting present before the next frame's draw. 120.0 presents/s, 99%+ of gaps in 7-9.5 ms. | FIXED on PC 2026-10-09 (G-Sync OLED, maintainer by eye); Deck 90 good on most levels (per-level pass 2026-10-10); open: Bunker fight draw cost, Jungle feel (see 2026-10-10 addendum) |
 | D584 | **Cradle catwalk shadow flicker at every frame rate (regression from D579).** D579 stopped cull-testing triangles with a vertex behind the camera; Cradle's translucent two-faced catwalk walls then drew both faces. Fix: mixed-w triangles get the homogeneous facing test det[x y w] (no division by w), identical in sign to the pre-D579 XOR flip but well-conditioned. #150 stays fixed (its real fix was the D543 camera-plane clip). | FIXED 2026-10-10 (maintainer Deck: Cradle clean, Dam wall-hug clean) |
 | D614 | **Deck Bunker fight texture-cache thrash: CI8 keys hashed leftover palette entries past a partial TLUT load, so the same image re-imported whenever draw order changed (1,000-4,000 evictions per 10 s).** A loaded-extent CI8 key cut re-imports 60-70% but changed neither import time (-7%) nor pacing, and 84-315 texels per run index past the extent (fidelity risk): reverted. Not the cause of the fight drops; D583 present scheduling remains the lead. | OPEN: measured, fix not shipped (2026-10-10) |
+| D605 | **Issue #145 (v0.6.0): F10 "Invert look" never reached the look stick in the HONEY role (game styles 1.1/1.3 -- left stick looks, right stick is the digital movement C-button source), so it was inert under the game's own original styles; and the D564 pin of the game's own watch "Look up/down" stayed on in Original mode, locking the player's native in-game control** — full `## D605` entry at file tail | FIXED 2026-10-09 (v0.6.0 line; port-only: the HONEY branch inverts the left (look) channel instead of the right, the D564 pin is lifted in the Original scheme so the watch row is live again; build green, the #145 maintainer/user live pad pass owed) |
+| D606 | **Issue #144 (v0.6.0 triage): no on-screen mission timer; added the optional PD-style always-visible timer reading the game's own mission clock** — full `## D606` entry at file tail | FIXED 2026-10-09 (v0.6.0 line; port-only in `port/src/optionsoverlay.c`, `src/game` zero-diff; off by default so goldens stay byte-identical; smoke-verified: a 1200-frame @60 run shows "TIME 0:20" — the timer tracks the game clock exactly; live multi-minute level pass owed) |
 | D407 | **Tanks cannot be boarded/exited on PC — the v0.4.0 use/reload split (D378/D393) removed the B-button tap the engine's tank handlers in `bondview2.c` consume (user report 2026-09-28, Runway/Streets)** — full `## D407` entry at file tail | **CLOSED (2026-09-28 bookkeeping: tank board/exit accepted by the user; shipped in v0.4.0).** Earlier: PARTIAL (fix landed in v0.4.0, port-only): present `GE_CONT_B` on the use (E / pad A) rising edge only while `g_PlayerIsInTank == 1` or `g_BondCanEnterTank != 0`; E keeps its D378 no-reload-fallback semantics elsewhere, N64 layout unaffected. D407(b) same cycle: front PC Options page-edge highlight clamp + mouse-wheel / W-S paging (wheel queue consumed on the menu, D223); its wheel mapping shipped inverted and was fixed to match W/S (wheel up = step up). D407(c): board-animation lockout, menu-accept B gate in tank states, in-tank aim routed through the legacy velocity stick + `Input.TankAimScale` knob. RESOLVED 2026-09-28: all port-only (src/game zero-diff), 10 TANKDBG probes stripped, release binary verified clean, user live tank drive signed off (board + exit OK, aim feel good). |
 | D408 | **`Game.SkipIntro` skips the post-mission failure dossier: with it on, entering a level then aborting (watch Z+A) or dying (KIA) returns straight to the menus — no REPORT / "Mission status: KILLED IN ACTION / ABORTED" screen** (user report 2026-09-28; save/AllUnlocked ruled out) — full `## D408` entry at file tail | FIXED 2026-09-30 (SkipIntro hook gated to first boot; live dossier check owed). |
 | D415 | **Shooting a light fixture does not "kill" its flickering light (Bunker, Caverns): the fixture takes the hit but the flicker keeps going (community report, issue #87 comment by the reporter, 2026-09-28, 00 Agent playthrough of v0.4.0).** — full `## D415` entry at file tail | FIXED (2026-09-29 bookkeeping: root-caused and fixed as D430 + D431, user-verified vs GEPD/1964 on Caverns 2026-09-29; #119). Earlier: ROOT-CAUSED as D430 (see there); fix in tree, live check owed. |
@@ -850,6 +852,27 @@ covers D24–D69; the log continues in §H (D32 procedure, D70–D121).
 | D450 | **Unplugging and replugging a controller was not picked up (user, GameSir-G7 Pro; 2026-09-29).** — full `## D450` entry at file tail | FIXED — port-only: the render thread's `gfx_sdl_handle_events` drains the same SDL queue as video.c's host pump but had no case for `SDL_CONTROLLERDEVICEADDED/REMOVED`, so it usually swallowed them and `inputRescanPads` never ran. Both pumps now post a request; the rescan runs at the top of `inputUpdate`, on the thread that reads the pads. |
 | D436 | **Persistent (hipfire) crosshair, opt-in: `Video.CrosshairPersistent` / F10 > Gameplay > HUD > "Crosshair when not aiming" draws the native sight at `crosshair_angle` outside aim mode (2026-09-29).** — full `## D436` entry at file tail | FIXED (shipped 2026-09-29; rule-2 sign-off: user direction; presentation-only masked condition in `gunDrawSight` + `port/src/video.c`). Default off byte-identical; aim mode unchanged; netplay SH identical mixed off/on. |
 
+| D617 | **The HD pack's `ge007.tdb` is the emulator's full per-draw key list, not the pack's PNG list — non-CI rice keys are per draw tile, so the join needs display-list tile parameters** — ROADMAP #100 rice-pack stage (tooling); full `## D617` entry at file tail | OPEN — plan written (`docs/dev/TEXPACK-RICE-DRAW-PLAN.md`); whole-texture recipe confirmed correct by 9 tdb hits with exact `;WxH` agreement; PNG-filename and brute-force param searches are dead ends |
+| D618 | **The community HD pack's non-CI keys are per-draw rects partly unreachable from N64 tile state (`440x1` strips), and a rect key cannot land in the whole-texture pack format — the rice-pack route stops here** — ROADMAP #100 rice-pack stage; full `## D618` entry at file tail | STOPPED 2026-10-06 — tdb `;WxH` are per-draw rects (440x1 ×301 over tmem's 128-byte pitch = recompile-side); ~100 byte-domain variants score 0 hits; content-join does not discriminate; rect-keyed replacement would be a format extension. HD/custom texture support ships on the texnum path |
+| D619 | **The community HD pack's remaining 1790 PNGs are blocked by four independent causes (recompile-side rect keys, wrong hash domain, rect-vs-whole-texture format, content outside the identity space) — #100 status written up** — ROADMAP #100; full `## D619` entry at file tail, current status in `docs/dev/TEXPACK-STATUS.md` | WRITTEN UP 2026-10-06 — brute-force rect join 14.6M positions / 0 hits (three stride models); pack taxonomy puts 822 PNGs in UI/Hacks classes the engine never registers; ceiling ~1200 of 2046 via an unmeasured level-scoped content join; 256 import today |
+| D620 | **The HD pack's non-CI keys hash the pack's own re-encoded content — no vanilla/recompile byte domain produces them; the join must come from external signals (level folder + tile geometry)** — ROADMAP #100 rice-pack stage; full `## D620` entry at file tail | MEASURED 2026-10-07 — xdelta "plus" ROMs unusable (32 MB, 98 % diff, unparseable at imagelist offsets; prior recomp probes void); vanilla 953 non-CI keys → 0 tdb hits, pack 1491 `_all` keys → 0 vanilla hits, 541 pack palcrcs → 3 vanilla matches (HD palettes regenerated); joined-pair scale factors 4×–128× (dims join dead) but folder→level mapping holds (257 pairs / 21 folders); B2 rewritten, B4 narrowed (sprite-bank UI/Effects art is registered via texLoad) |
+| D621 | **The PD dabs-mod texture-pack notes mine: the HD route stays texnum-keyed — the PD port's key layer transfers nothing (its keys hash ROM bytes we hold; this pack's hash 1964tech's build, D620), but its format layer is a ready fix list; end-to-end plan lands** — ROADMAP #100; full `## D621` entry at file tail, plan in `docs/dev/TEXPACK-PLAN.md` | PLAN LANDED 2026-10-07 — Phase A (pad contract / alpha repair / F10 row / stall / manifest / merge), Phase B (B0 sprite-bank GE_TEXNUMDUMP confirmation → B1 folder→level → B2 per-level GDL census → B3 ambiguity gate → B4 content scorer → B5 review viewer), Phase C (join output as NNNN.png + per-level A/B); next task B0 |
+| D622 | **B0 sprite-bank confirmation — the `Effects`/`UI` sprite-bank texnums register and draw (explosion/smoke banks 2081–2196, crosshair 2236, front-end art 2194–2628), so those PNGs stay in the join scope; the unregistered residue is exactly the known in-exe art (font glyphs + fire particles); Phase B continues to B1–B3** — ROADMAP #100 Phase B0 (`docs/dev/TEXPACK-PLAN.md`); full `## D622` entry at file tail | PASS 2026-10-07 — two human-driven `GE_TEXNUMDUMP`+`GE_TEXDUMP` runs (front end 377 unique texnums / 521 dumps, `-level_09` combat 321 / 316; evidence `scratch/b0/`): Effects/Explosion+Smoke banks and UI/Crosshair registered at `texLoad` and drawn; `UI/Gunbarrel` PNGs are gun model textures (not a sprite bank), `UI/Letters`+`UI/Ammo` = in-exe fonts (out of plan); `GE_TEXNUMDUMP` log now flushes per line (was lost on force-kill); next: B1–B3 |
+| D623 | **Phase B1–B3 measured: the folder→level mapping holds (241/244 Areas containment with the model pool) and model DLs carry the G_NOOP texnum convention, but the B3 ambiguity gate fails — median 104 candidates/PNG, so the geometric tier is B5 human-in-loop** — ROADMAP #100 Phase B1–B3 (`docs/dev/TEXPACK-PLAN.md`); tooling `tools_pc/texpack/{level,model}_census.py` + `ambiguity_report.py`; full `## D623` entry at file tail | MEASURED 2026-10-07 — B1: 15 pack Areas folders mapped via kSoloLevels/bg.c (+2 mission offset from Statue; Bunker→09+27, Surface→36+43); bg-only containment 173/257, +model pool (new: model segs carry G_NOOP texnums, 510 models / 1462 texnums) → 241/244 Areas. B3 (pow2 scale set learned from the 251 joined pairs; 32×/64× dominate): ex-Hacks 1781 PNGs → 515 zero-cand (29 %, mostly `UI` 454 = the known in-exe fonts), 44 in 1–3, 60 in 4–20, 1162 >20 (65 %); median 104 → gate = B5 for the geometric tier; decision owed: B4 scorer budget (1162 × ~100 cands) vs B5 tail vs both |
+| D624 | **B4 content scorer built + tuned: the structural RGB-NCC scorer (dither-blurred, palette-invariant) beats CLIP on this pack — rank-1 56 % vs 33 % on the 257 ground-truth pairs; auto-accept tier at margin ≥ 0.10 is 100 % precise on GT; pipeline produces the human queue (189 auto / 1341 vision-pending / 515 out-of-scope)** — ROADMAP #100 Phase B4 (`docs/dev/TEXPACK-PLAN.md`); tool `tools_pc/texpack/content_scorer.py` (rank/tune/verdicts/apply); full `## D624` entry at file tail | BUILT 2026-10-07 — 5-stage pipeline (geometric prefilter → content rank → GT-tuned auto-accept → vision judge → human queue) per the 2026-10-07 decision (commits 512e6d57, 95d38637). Scorers measured on 255 in-scope GT pairs: CLIP ViT-B/32 33 % rank-1 (ViT-L/14 no better); structural 3-channel NCC + 1/2× LANCZOS dither-blur 56 % rank-1 / 67 % rank-≤5 → adopted (CLIP kept as `--scorer clip` A/B). Margin-vs-precision on GT: 100 % down to t≈0.05 (97/255 @ 0.10, 61/255 @ 0.20); recommended t=0.10. Verdicts @ 0.10: 189 auto (47 geometric n≤3 + 142 margin), 1341 vision-pending (contact sheets + answers.csv template under `scratch/join_review/`), 515 out-of-scope; `apply` merges vision answers → `queue.csv` + `queue/` sheets (verified). Stop point honoured: queue produced, Phase C not started. Evidence: `scratch/join_ranked.csv`, `scratch/join_verdicts.csv`, `scratch/join_review/`, caches `scratch/{clip_tex,clip_pack}_cache.npz` + `scratch/ncc_van_cache.pkl` |
+| D625 | **texpack A5: downscale-to-GL-max instead of the D615 vanilla fallback, plus first-use decode-stall metering** — a pack image over the GL max (12 of 256 community files are 2048/3072/4096) used to drop to the vanilla texture, silently killing the HD look on low-end GPUs; now one separable LANCZOS3 resize (radius 3, weight-renormalised at edges) into a lazily-grown dw×sh work buffer yields the largest size the GPU holds, one-time per file through the existing cache; the only fallback left is OOM/degenerate GL max < 2 (logged ≤8 lines). Stall meters (render-thread only): total/worst-file decode ms, per-frame accumulation committed at a new `gfx_start_frame` hook (`texpackFrameStart`, inert without a pack), evictions; summary every 16th decode and every decode > 5 ms, `fflush`'d (D622-class). `GE_FAKEMAXTEX=<n>` test hook registered in `docs/dev/GE-ENV-PROBES.md`. Port-only: `port/fast3d/gfx_pc.cpp`, `port/src/texpack.c`, `port/include/texpack.h`; full `## D625` entry at file tail | BUILT 2026-10-10, not yet A/B-verified — downscale path needs a `GE_FAKEMAXTEX` pack run + low-end pass; stall meters need one pack run for the first-use spike shape. **Label note: D621–D625 on `feature/texpack` collide with the mainline allocation; renumbered to the next free numbers at land time** |
+| D626 | **Windows DPI awareness was never declared: the hint sat behind PD's `PLATFORM_WIN32`, never defined here, so scaled displays got a stretched logical-size surface** (3072x1728 on a 125 % 4K panel) | FIXED 2026-10-10 (`gfx_sdl2.cpp`: `_WIN32` guard, `permonitorv2` always); fullscreen now native 3840x2160, Facility golden PASS; maintainer by-eye owed |
+| D627 | **DexDrive / GameFAQs `.n64` saves import directly: the GE note (game code `3BADD1E5`) is pulled from the Controller Pak dump, zero-padded to 2048 B and handed to the D514 emulator-save importer** | FIXED 2026-10-10 (`eepimport.c` + `eep_convert.py dexdrive-to-pc`); byte-identical to the reference extraction, malformed inputs rejected untouched; by-eye mission select owed |
+| D628 | **The exact-CRC CI half of the GoldenEye-007-HD join is sound (256 texnums, independently re-verified), but its structural half was scored against `GE_TEXDUMP` captures taken *with a pack selected* — the dump is the pack image, so the hold-out was circular and the "same-size replacement" premise it produced is false (packs are 4x-128x upscales, D620)** — ROADMAP #100; join table `tools_pc/texpack/gehd_join.csv`, tooling `tools_pc/texpack/gehd_{join,build}.py`; full `## D628` entry at file tail | CI HALF ACCEPTED 2026-10-10 (256/256 `(crc,fmt,siz)` agreement with the rice census, 256 distinct texnums, palette-variant collision resolved by ROM palcrc, builder resolves all 256 sources); structural half VOID (same-size gate admitted only the 52 contaminated dumps, so "0 matches" was unmatchable rather than rejected; gate rebuilt aspect-ratio/integer-scale aware); no per-level dump runs needed - the non-CI route is D624 python-decoded vanilla renders + its existing human/vision queue |
+| D632 | **Exit BSOD 0x119 recurred (2026-10-10) despite D344; the first minidump shows it fires in the kernel's process-exit GPU cleanup (`DXGPROCESS::Destroy` -> `DXGCONTEXT::DestroyContext` -> `VidSchTerminateHwQueue` -> `VidSchiDecrementHwQueueReference`, HAGS on), i.e. on the GL context the port leaves for the kernel to destroy** - `port/fast3d/gfx_sdl2.cpp`, `port/src/video.c`, `port/src/main.c`; full `## D632` entry at file tail | FIX BUILT 2026-10-10 (`fix/d632-exit-teardown`): the atexit handler deletes the GL context and window while the process is alive (only when the render thread parked). Not run-verified (game runs on hold); maintainer exit checks owed |
+| D636 | **Non-CI pack textures ARE fingerprint-keyed: the D617/D620 census only alt-row-swapped explicit-LOD textures, but the game swaps every texture (D149); swapping all gives 203 exact non-CI joins (was 17), and those overrule 30 picture-join picks (near-identical variants on the wrong texnum, e.g. 948/949, 35/36, 284/307)** - `tools_pc/texpack/rice_nonci.py`, `gehd_join.py nonci`, `gehd_join.csv` (+203 rows); full `## D636` entry at file tail | FIXED 2026-10-10 (`feat/texpack-fp`): exact join 256 CI + 203 non-CI = 459 texnums; local pack rebuilt exact-first (511 files) |
+| D637 | **Fingerprint fallback in the loader: a texture with no texture number is keyed by RiceCRC32 over its loaded texels and replaced from `<pack>/fp/GOLDENEYE#crc#fmt#siz...png` - brings in the menu/HUD font (188 glyphs, D635)** - `port/src/texpack.c`, `port/fast3d/gfx_pc.cpp`, `tools_pc/texpack/gehd_build.py --fp`; full `## D637` entry at file tail | BUILT 2026-10-10 (`feat/texpack-fp`); C CRC == Python on 80 ROM glyphs; not run-verified (game runs on hold) |
+| D638 | **HD-pack ammo icon cut off: the draw normalises texture coordinates by the padded TMEM line (a 5-texel RGBA32 icon is sampled as 8 wide), but the pack image holds only the visible texels, so it was stretched and cropped** - `port/fast3d/gfx_pc.cpp` (import_texture pack upload); full `## D638` entry at file tail | FIXED 2026-10-10 (`chore/v070-agentic-batch`): the pack image is placed on a canvas padded in the same proportion (edge texels replicated); headless Dam run: 5x12 -> 8x12, 5x28 -> 8x28, 54x54 -> 56x54, HUD cartridge renders whole |
+| D639 | **Exclusive fullscreen picked the wrong resolution: `SDL_WINDOW_FULLSCREEN` switches the monitor to the window's display mode, which SDL derives from the current window size when none is set** - `port/fast3d/gfx_sdl2.cpp` `set_fullscreen`; full `## D639` entry at file tail | FIX BUILT 2026-10-10 (`chore/v070-agentic-batch`): exclusive fullscreen now sets the desktop mode of the window's display first; maintainer check owed (NVIDIA "Prefer native") |
+| D643 | **Issues #145 (b)(c) / #149: in the Original control style the pad was not a virtual N64 controller -- port actions (dedicated use/reload, crouch, gadget cycle, prev-weapon) and a fixed menu lockout (West = B-equivalent accept) were layered on, so one button could open a door AND reload, and West accepted instead of going back** — full `## D643` entry at file tail | FIXED 2026-10-12 (port-only; GE_PADMAPTEST PASS; maintainer pad pass vs the #145/#149 reports; 1:1 N64-pad stick check owed when an adapter is available) |
+| D644 | **File-select royal crest drew on an opaque dark square (#146): IA4 decoded as I2:A2 instead of the RDP's I3:A1** — full `## D644` entry at file tail | FIXED 2026-10-12 (`import_texture_ia4`; 1964-confirmed symptom, maintainer-verified; Win sweep unchanged) |
+| D645 | **Original 1.1: F10 "Invert look" did not invert vertical look (#145 a): D605 flipped the move stick; the game's own Look up/down drives C-button/aim pitch** — full `## D645` entry at file tail | FIXED 2026-10-12 (F10 row and watch row synced in Original; maintainer-verified) |
+| D646 | **F10 dropdown popup drew over the hint/tooltip panel (popup bounded by the screen edge, not the settings box)** — full `## D646` entry at file tail | FIXED 2026-10-12 (`ddGeom` bounded to the box; maintainer-verified) |
 
 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
@@ -18855,6 +18878,21 @@ Build-verified only, not run (confidence medium). `Video.FpsCap = -1` is the new
 
 **Release strip (2026-10-10, v0.5.1):** the `d578_live` live-word file and every word it read, `holecap`, `burst` capture, `tristat`, `GE_INTERP_SELFCHECK`, `GE_INTERPPORTAL_CHECK`, PROJDIAG, the `GE_INTERPSCIS_*` / `GE_INTERPSCISSOR` / `GE_INTERPGATE_ABFILE` knobs and the matrix-counting set-change check (`GE_INTERPSET_*`, which only ran with `mvblend`) were removed; each site collapsed to its default branch (clip widening off, replay box unioned with the game box, in-place restart on, glFinish skipped under gamescope). Kept: the 10 s stats lines, `D583 SLOW`, IPLOG/`GE_INTERPLOG`, `GE_INTERPCLIP_ON`, `GE_INTERPFALLBACK_*`, `GE_INTERP_FAST_TURN`, `GE_INTERPROOM/REACH_OFF`, `GE_INTERPPORTAL_OFF/_UNION`. The removed switches are in `feat/d583-deck90` history for a future debug build.
 
+### D578 addendum (2026-10-13): Auto follows the display up to 240; 165 and 240 in the menu
+
+Maintainer ask (v0.6.0 release day): allow 240 and let Auto follow the user's display. The old Auto clamp
+of 144 had no recorded reason; it was the top menu entry when Auto was added. The real ceiling is the
+present ring: `IP_SLOTS` 8 with `maxN = IP_SLOTS / 2` (libultra.c) gives at most 4 presents per sim tick,
+i.e. 240 NTSC / 200 PAL. `videoAutoFpsCap()` now returns the refresh clamped to `4 * vi` (was 144), and
+`kFpsCapSeq` gains 165 and 240. A display above 240 Hz holds at 240 under Auto (raising it needs a bigger
+ring and the matching FBO slots, `GFX_INTERP_SLOTS`). A separate "uncapped" entry was not added: presents
+are paced to the refresh (VSync or the compositor), so uncapped interpolation = Auto, and `FpsCap 0`
+stays the old non-interpolated path (D186). Slow GPUs are already covered by the per-frame `s_ipPassEma`
+pass reduction. Verified on the dev box (120 Hz panel), Dam, 30 s, `GE_INTERP_FORCE=1 FpsCap=240`: with VSync
+on, swap-driven at 120 (2 slots/tick, as before); with VSync off, timer-paced 240.0 presents/s, 4 slots/tick
+on 600/600 frames, 0 dropped after warm-up, draw peak < 1 ms, clean GE_QUITFRAME exit. Not yet seen on a
+real 165/240 Hz panel -- confidence medium-high.
+
 ## D579: Issue #150 "geometry disappears when too close" -- FIXED: the D543 CPU near clip cut at the near plane under GL_DEPTH_CLAMP; now cuts at the camera plane (2026-10-08, fixed 2026-10-09)
 
 **Status: FIXED 2026-10-09 (maintainer-verified at the Dam gate in 16:9, floor haze intact; golden sweep owed). Root cause and fix: see "Resolution" at the end of this entry. Everything between here and there is the superseded investigation record (the cull was never the cause).**
@@ -18910,11 +18948,14 @@ Build-verified only, not run (confidence medium). `Video.FpsCap = -1` is the new
 - Fix: `#ifdef PORT` macro `REUSE_PDTYPE` reads `PropDefHeaderRecord::type`. No logic change.
 - Siblings: grepped `src/game` for `(u8)x[0]` type reads; only cleanup_objects.c (already fixed, D139) and this function.
 - Confidence: high on cause (matches RAX/fault signature); runtime verification owed (hat-spawning guard AI).
+- **Runtime check (2026-10-13, Windows, temporary probe, since removed):** at each level's first render, walked `g_CurrentSetup.propDefs` with the true stride and compared the fixed read (`PropDefHeaderRecord::type`) with the old `(u8)cmd[0]`, then called `setupFindObjForReuse(PROPDEF_HAT, ...)` with musthaveprop FALSE and TRUE (the TRUE call reaches the `obj->prop->parent` line that faulted in #153). All 21 solo levels: no crash, END reached at the true end. The fixed read finds every hat record (e.g. Bunker 1: 23 of 205 records, all with live props; Facility 51, Train 62, Jungle 40, Aztec 39). The old read matched **0 real hats on every level**: on PC, hat-slot reuse never worked before the fix, so the old search could only ever return a false match (a record whose extrascale low byte is 0x11 = PROPDEF_HAT). No record held that byte at level start; the probe's `early_end_at=-1` on every level is the missing piece, as issue #156 (same crash, same PC, reported by spektra, 2026-10-13) pointed out: `PROPDEF_END` (0x30) was also tested on the wrong byte, the END record's extrascale is 0, so the old loop **never stopped at the end of the list** and walked on into the memory after it until a stray 0x30 (stop) or 0x11 (false hat match, then `obj->prop->parent` on garbage). That explains the "after a while of play" timing (it depends on the heap contents past the list) and why the weapon-recycle caller (`setupFindObjForReuse(PROPDEF_COLLECTABLE)` from propobj.c ~11488) was exposed too. The D580 macro covers both reads (the while test and the type match). #156 = duplicate of #153, fixed by the same change. Confidence: high.
 
-## D581: First launch opened on the leftmost monitor, not the primary — FIXED 2026-10-09 (port only, #151)
+## D581: First launch opened on the leftmost monitor, not the primary — FIXED 2026-10-09, corrected + runtime-verified 2026-10-13 (port only, #151)
 
 Cause: with no saved Window.X/Y, `gfx_sdl2.cpp` centred the window using display 0 mode size as absolute coordinates (and sized it from display 0). SDL display 0 is not the OS primary on multi-monitor X11/Xwayland, and the coordinates ignored the display desktop offset, so the window landed on the left monitor. Fix: `primary_display_index()` picks the display whose bounds contain desktop origin (0,0), falling back to 0; auto size and centring use it, centring in global coordinates. Saved positions (Window.X/Y >= 0) untouched. Build verified; runtime check needs a dual-monitor desktop with non-leftmost primary. Confidence: medium-high (native Wayland may ignore requested positions).
 
+**Correction (2026-10-13, runtime-checked on the X220, Fedora 37 KDE Plasma X11, SDL 2.26.3):** the "display containing (0,0)" rule was wrong and still reproduced #151. With the reporter's layout (secondary on the left at the origin, primary on the right; here LVDS-1 at 0,0 and HDMI-1 primary at +1366+0), an SDL probe listed **display 0 = the xrandr primary** (bounds 1366,0 1920x1080) and display 1 = the laptop at 0,0, and the game opened 869x652 at X=249, sized and centred on the laptop. SDL already lists the OS primary first (X11 RandR primary, Windows MONITORINFOF_PRIMARY, macOS main display); the original bug was only the missing desktop offset. Fix: `primary_display_index()` returns 0, and every centring helper adds the display's bounds origin (`gfx_sdl_init`, `gfx_sdl_get_centered_positions` for F10 Center window, and `get_centered_positions_native` for exclusive-mode changes had the same mode-size-as-absolute bug). After the fix: 1224x918 at X=1714 (centred on HDMI-1: 1366 + 960 - 612). Regression check with the primary at the origin (HDMI-1 0,0 primary, LVDS-1 at +1920): X=348 (960 - 612), as before. Native Wayland (no positions) is untested. Confidence: high for X11/Windows.
+
 ## D582: #152 v0.5.0 Linux crash in the ending credits (Cuba), textMeasure reading a garbage string pointer — ANALYSIS ONLY 2026-10-09
 
 **Report:** issue #152 (Debian 13.7, v0.5.0 / 5c43e69, ntsc-final). SIGSEGV a few
@@ -20818,3 +20859,1126 @@ the disassembly (64-bit load) plus the maintainer's live 2P retest (die, press E
 **Conclusion.** The thrash is real waste but not the cause of the Deck fight drops; D583's present-scheduling finding (a due present stuck behind the single worker's draw) remains the lead, with the parked present thread (D585) as the structural fix. If the thrash is revisited: key CI8 on the loaded extent only when the texture's max index is known to be inside it (e.g. memoise the max index per address+size with invalidation on tex-pool reuse), and give `Hasher` the full key so variants spread across buckets.
 
 **Ops note.** Two `deckrunner.sh` instances (one from an earlier Game Mode launch) raced for the same job files: one claimed a job and the other ran it with empty arguments, clobbering results and running levels concurrently. Check `pgrep -af deckrunner.sh` before queueing.
+
+## D615: HD texture packs (first stage) — `Video.TexturePack` selects a `<texnum>.png` directory; the renderer substitutes pack images at texture import, keyed on the texture number the game knows (ROADMAP #100)
+
+**Scope:** the first stage of ROADMAP #100 (HD texture packs, post-1.0
+batch; this landed early in the `feature/texpack` worktree off
+`release/v0.4.1`). A texture pack is a sub-directory of `data/texpacks/`
+holding PNGs named `<texture-number>.png` (the 12-bit `texturenum` the
+game assigns each texture, zero-padded decimal). The new ini option
+`Video.TexturePack` (sub-directory name; empty = off) selects one; the
+port substitutes the pack's image for the ROM's decompressed texture
+wherever that number is imported.
+
+**Why the hook is in `texLoad` (the only game-code touch):** the N64
+pipeline never makes the texture number visible to the RSP — GBI carries
+pool addresses only. The game side knows the number (`tex->texturenum`,
+written at `src/game/image.c` texLoad) and the renderer sees the address
+(`import_texture`'s `orig_addr`), so one line bridges the two:
+`#ifdef PORT texpackRegister(tex->data, (s32)tex->texturenum)` after the
+pool allocation. **AGENTS rule 2:** this is a behavioral addition
+(opt-in visual replacement, not an ABI/layout edit), so it was
+implemented under the requested sign-off, behind `#ifdef PORT`, with the
+default (empty option) fully inert — the un-patched build path is
+untouched and byte-identical.
+
+**Mechanism (all in `port/`):**
+- `port/src/texpack.c` — address→texnum registry (open-addressing,
+  insert/update only: the texture pool is a ring, so recycled addresses
+  get re-registered; the D75 content hash still protects the GL cache
+  independently). K0/V1 normalised direction-independently
+  (`GE_HIGHARENA_BASE` inverts the numeric order). One mutex: the game
+  thread registers, the render thread resolves; PNG decodes under the
+  lock (each file decodes once; FIFO cache, 128 entries / 512 MB cap).
+- `port/src/pngread.c` — self-contained PNG→RGBA32 decoder (zlib, which
+  is already linked; no new dependency), adapted from the Perfect Dark
+  PC port's (MIT; NOTICE). Handles the pack-relevant subset: 8/16-bit
+  RGB(A)/palette/gray, plain + Adam7.
+- `port/fast3d/gfx_pc.cpp` `import_texture` — when the pack has a file
+  for the resolved texnum, the decoded RGBA32 takes over the upload
+  (shared `tex_upload_buffer` → `upload_texture`, mip generation per
+  `rdp.tex_lod`), bypassing the N64 pixel-format dispatch; the cache key
+  is tagged `0x5EA7xxxx | texnum` so a pack upload never aliases a
+  vanilla key. Images larger than the upload buffer / GL max fall back
+  to the vanilla import (logged once).
+- `port/src/video.c` — `Video.TexturePack` string option +
+  `videoTexPackInit()` called from `main.c` right after `configLoad()`
+  (same ordering as the Steam Deck preset).
+
+**Verification:** `-level_09` A/B with a 512-file test pack
+(`tools_pc/texpack_make_test_pack.py`, solid-colour PNGs covering
+texnums 0–511): baseline vs packed frames — the level's 10 active
+texnums (7, 257, 259, 295, 296, 315, 468, …) visibly substitute
+(walls/floor/ceiling turn solid red), every un-replaced element
+(doors, arrows, lights, the CCTV window) renders unchanged; decode log
+clean, no errors, 32-frame runs stable. Default (empty option) run:
+`texpack: no pack selected`, frame-identical scene.
+
+**Known limitations (documented for the stage):** pack files replace
+the *base image* — the N64 wrap period (the RC3/D167 mask period) is
+unchanged, so a pack image larger than the tile's authored period will
+repeat at the authored period; in-exe-image (C-array) assets are not
+registered (only the ROM-serialised `texLoad` path), so title/HUD 2D
+art is out of scope this stage; decode is synchronous on the render
+thread (first-use stall per file, ~1–3 ms); no F10 options row yet
+(string options have no options-overlay row — follow-up); rice-pack
+(rice→texnum) support is the next stage, per the dabs reference.
+
+---
+
+### Session 2 (2026-10-06): pack-authoring dumper, alpha-true dumps, and the "line artifact" fix
+
+**Texnum dumper (ROADMAP #100's dumper half).** The registry is now
+always-on (`texpackRegister`/`texpackTexNumFor` no longer gated on a
+selected pack — one hash probe + mutex per texLoad; only the pixel
+swap in `texpackReplacement` stays gated), and two things were added:
+
+- `GE_TEXNUMDUMP=1` appends every `(address, texnum)` pair to
+  `texdump/texnumd.log` at registration — the offline map that turns a
+  `GE_TEXDUMP` run into a pack-authoring job list.
+- `texpackBeginImport(addr)` / `texpackCurrentTexNum()` (import context
+  on the render thread, strictly synchronous, zero locking): 
+  `import_texture` records the address it is about to import, and the
+  GL upload layer (`gfx_opengl.cpp`) names its `GE_TEXDUMP` PPMs
+  `texdump/NNNN.ppm` (texnum) for registered pool textures instead of
+  the `tNNN_WxH_mipN` sequence name; unregistered/C-array textures and
+  the fire-shape branch keep the old names.
+
+**Alpha-true dump sidecar.** The PPM writer has no alpha channel, and
+the pack-authoring tool initially flattened every pixel to A=255 —
+which silently destroyed vanilla 32-bit cutout alpha (arrow HUD quads,
+cactus leaves). The texnum-named dump now also writes a raw RGBA32
+sidecar `texdump/NNNN.rgba` (`GE1\nW H\n` + w*h*4 bytes); the
+authoring tool prefers it and falls back to the PPM (A=255
+interleaved) only when the sidecar is absent.
+
+**The "line artifacts" (4x pack looked wrong) — root cause + fix.**
+Three separate bugs, in order of discovery:
+
+1. *UV scale bug (introduced, then reverted):* a per-tile float UV
+   multiplier (`tpack_su/sv = pack_dim / orig_dim`, applied to the
+   vertex-pass UVs) assumed the GL image size had to be folded into
+   the UV range. It did not — the vertex pass normalises by the N64
+   texel count (`tex_width` from the tile line-size), and a wrapping
+   tile's GL repeat period grows with the uploaded image size,
+   cancelling the texel growth exactly. The multiplier therefore
+   made every replaced surface Nx denser (the "line artifacts" of
+   the first 4x test). Reverted at all four sites; the pack branch
+   comment records why. (The 1x A/B: with UV scaling off and an
+   alpha-true pack, 0 texnums differ at the dump level and the
+   frame diff drops to a sub-2-level cactus-region remnant.)
+2. *Nearest-neighbour upscale grid:* N64 textures are 16–64 px and
+   most surfaces are **magnified** on screen. GL_LINEAR interpolates
+   the small originals smoothly; a 4x NN pack turns each texel into a
+   hard 4-px block, so the wall (a 32x32 original) showed a ~9-px
+   square grid at mid-range. Fix: the authoring tool's default
+   upscale is now **Lanczos-3** (separable 2x steps, pure stdlib;
+   `--nn` keeps the blocky look).
+3. *Minification aliasing:* in-level `tex_lod` is 0, so the vanilla
+   sampler is base-level-only GL_LINEAR; a 4x pack image minifies
+   4x harder and aliases into regular banding even with smooth
+   content. Fix (port-only, pack-only): pack replacements always
+   upload with a generated mip chain (`upload_texture(..., true)`) and
+   the cache node's new `sampler_mips` flag forces the MIPMAP min
+   filter (compared in the vertex pass so a pack/vanilla bind swap
+   re-sets it). Vanilla nodes and the vanilla upload path are
+   untouched (vanilla `sampler_mips` = `rdp.tex_lod` as before).
+
+**Result:** 4x Lanczos pack on `-level_09` frame 100 vs vanilla:
+mean abs diff 1.31/channel (the banded 4x NN build was ~11); no
+visible artifacts at the cactus/wall/panel regions. A
+procedurally-authored 1024x1024 demo pack (`testhd`, walls 0295/0259
+with seams/rivets/grain/scratches) shows the authored-detail
+"HD pack" look in-engine — the port needed nothing further for it.
+
+**Pack-authoring tools (`tools_pc/`, stdlib-only):**
+`texpack_from_dumps.py` (texdump dir -> pack dir; `.rgba` sidecars
+preferred, PPM fallback; Lanczos-3 default, `--nn` option; scale must
+be a power of two) and `texpack_demo_hd.py` (same pipeline, but the
+listed texnums get a procedurally *authored* 1024x1024 texture —
+seams/rivets/grain/AO/scratches composited over the smooth upscale —
+as a stand-in for hand-drawn HD art).
+
+
+**Addendum 2026-10-10 (alpha preservation + per-level overrides; port layer
+only, `port/src/texpack.c`, `port/fast3d/gfx_pc.cpp`):**
+
+*Fix 1 -- keep the original's transparency.* Glass (texture 654, an I-format
+texture whose N64 alpha is its intensity) turned solid because the pack PNG
+is fully opaque. "Fully opaque" is computed once per PNG at decode and
+cached with it. On a texture-cache miss, when the replacement is opaque:
+I (and CI drawn with the TLUT off, which the importer already treats as I)
+and IA get alpha = Rec.601 luminance of each replacement pixel (exact N64
+semantics for I; for IA a flat RGB image carries no alpha, so luminance is
+the best available). CI with any transparent entry in the used palette bank,
+RGBA16 with any alpha bit clear, or RGBA32 with any alpha < 255 cannot be
+rebuilt faithfully: the texnum falls back to the vanilla texture (logged
+once, first 32), remembered in a per-texnum mark cleared on level change.
+Caveat: the CI check looks at the palette at first use; a texnum reused with
+a different palette keeps its first decision until the next level. A PNG
+that already has alpha is used as-is.
+
+*Fix 2 -- per-level overrides.* `data/texpacks/<pack>/<SS>/<texnum>.png`
+(SS = stage number, 1-3 digits, e.g. `33` = Dam) overrides the flat
+`<texnum>.png` while that stage is loaded. The key is
+`lvlGetCurrentStageToLoad()` (what the D416 log prints as `stage=`), read
+port-side only when the pack ships at least one override directory.
+Directories are scanned at pack-select time; the decode cache and the GL
+texture-cache key include the source (stage+1 xor-folded into the 0x5EA7xxxx
+tag), so a level change never reuses another level's override. With no pack
+selected nothing changes (texpackReplacement returns NULL on the first test).
+
+## D616: Rice texture packs (GLideN64/RT64 "rice"-CRC-named PNGs) map onto the port's `NNNN.png` texpack format via a ROM-side dumper + pack importer — ROADMAP #100, rice-pack stage
+
+**Status:** TOOLING LANDED 2026-10-06 (`tools_pc/texpack/rice_dumper.py` + `tools_pc/texpack/rice_pack_import.py`; feature/texpack worktree). No port C changes. Validated against the `GoldenEye-007-HD` pack (RT64/`ge007.tdb` ecosystem, `rt64.json` + 2046 `GOLDENEYE#…#…png`).
+
+**The problem.** Rice packs ship one PNG per N64 texture named
+`GOLDENEYE#<crc>#<fmt>#<siz>[#<palcrc>]{_ciByRGBA,_all}.png` where `<crc>`
+is a "RiceCRC32" of the texture's base image **as it sits in RDRAM**. The
+pack targets the 1964tech recompile, so its tdb/offset identifiers are
+recompile-side and not portable to the vanilla texnums. The only portable
+join key is the rice CRC itself — so the port side must be able to *produce*
+the same identifiers from the vanilla ROM.
+
+**Domain solved (after a long 0-hit phase — the LE variant is the trap):**
+
+- The pack's CRCs are **NOT** the `RiceCRC32` you get by porting
+  GLideN64's `TxUtil.cpp` or RT64's `texture_hasher.cpp` as-is: those
+  read `*(uint32*)(src+x)` (little-endian on x86), and that variant
+  produces **0 hits** against the pack on every byte-domain combination
+  (swapped/raw × LE/BE × stride/bpl, 1744 zlib-CI textures).
+- The matching variant is the **GlideHQ / PD-dabs-mod convention**
+  (see `ge-port-reference/perfect-dark-dabs-mod/port/src/texpack.c`):
+  4-bit rotate-carry CRC, **big-endian** 32-bit reads, and a per-row
+  `wordHash` that carries into the row-end XOR
+  (`wordHash = pos ^ BE32(row+pos); crc = wordHash + rol4(crc); …
+  crc += wordHash ^ counter`).
+- Hashed bytes = the base image exactly as RDRAM holds it:
+  1. zlib-CI decode (`texInflateZlib`): ctrl/format/nc-1/nc×2-byte
+     big-endian palette, then `w`,`h`, then rarezip (`11 72` header +
+     **raw** deflate, `wbits=-15`).
+  2. `texAlignIndices` layout: rows of `ceil(nper)` bytes padded to an
+     **8-byte-aligned stride** (nper = w for 8-bit idx, `(w+1)/2` for
+     4-bit).
+  3. `texSwapAltRowBytes` **iff `ctrl & 0x3F` (lod) > 0** (odd rows,
+     32-bit word pairs; row advance `alignedwidth*4` bytes = the pool
+     stride). The recompile's copy has **no `#ifdef PORT` guard** — it
+     swaps on PC too — and the pack was dumped from the recompile, so
+     the swapped domain is the right one for lod>0 textures.
+- The hash's `rowStride` = the **pool stride (8-aligned)**, not the
+  RDP-tile `bpl` formula.
+- ID fields: `fmt` = 2 for all colour-index textures; `siz` = 1 for
+  8-bit indices (RGBA16_CI8/IA16_CI8, image formats 9/11), 0 for 4-bit
+  (10/12). The 4th field (`palcrc`) is the same hash over the first
+  `cimax+1` big-endian 16-bit palette entries — but it is **not needed
+  for the join**: no two pack textures share an image CRC (verified
+  against all 532 CI entries), so the 3-field key `crc#2#siz` is
+  unambiguous.
+- Odd-width CI4 (8 of 1744): `texAlignIndices` writes `ceil(w/ipb)`
+  bytes per row (the C loop `for(x=0;x<width;x+=ipb)`), i.e. a
+  partial last byte, not `w//2`.
+
+**Result on `GoldenEye-007-HD`:** dumper emits 1744 vanilla zlib-CI
+identifiers; importer joins 555 CI pack PNGs → **251 unique keys / 256
+files** written (fan-out to identical-content texnums, e.g. 299/300).
+The 304 unmatched pack CI textures have no vanilla-ROM texnum
+(recompile-side additions); the extension path is joining the pack's
+`ge007.tdb` offsets to the recompile's `imagelist.u.csv`
+(`GoldenEye-007-Plus/imagelist.u.csv`) — out of stage-1 scope.
+Spot checks: tex 4's image CRC **and** palcrc (`5f045394`) both equal
+the pack's own PNG name; written files are valid PNGs (e.g. 0004.png
+320×320).
+
+**Non-CI pack textures** (fmt 0/3/4: IA/RGBA/VP, 4597 entries) are not
+covered: their rice IDs come from a different byte domain (RDP-tile
+bpl, paletteless formats) and from recompile assets — stage 2, with the
+tdb/imagelist join. Model packs: separate stage (ROADMAP #100).
+
+**Tools** (stdlib-only, no machine paths):
+- `tools_pc/texpack/rice_dumper.py <rom> <imagelist.u.csv> <out.csv>` —
+  per-texnum `texnum,crc,fmt,siz,palcrc,w,h`.
+- `tools_pc/texpack/rice_pack_import.py <rice_index.csv> <packdir>
+  <outdir> [packname]` — lays out `<outdir>/<texnum:04d>.png` for
+  `port/src/texpack.c` (D615) to consume.
+
+**In-engine A/B (2026-10-06, feature worktree):** the `GoldenEye-007-HD`
+pack was imported through the D616 pipeline (256 files under
+`data/texpacks/GoldenEye-007-HD/`; `data/` gitignored — pack PNGs are
+game content and stay out of the tree), `Video.TexturePack` set in
+`ge007.ini`, `-level_09` runs with `GE_QUITFRAME`: the log shows the
+substitution firing for real (13 decodes: 0315/1024 at 2048×2048,
+0893/0901/0915 at 1024², …) with a clean exit. `GE_PCDUMP` A/B
+captures at the same frame stem (vanilla vs packed): identical
+camera/HUD; the packed run shows the HD grain/detail on the
+corridor's wall and floor surfaces, the vanilla run flat. (Launch
+gotchas on a fresh worktree: `data/` needs the `pcmodels-`/`pccg-`
+asset sidecars and `C:\msys64\mingw64\bin` on PATH for SDL2.dll —
+known launch requirements, not pack issues.)
+
+## D617: The HD pack's `ge007.tdb` is the emulator's full per-draw key list — non-CI rice keys are per draw tile, so the join needs display-list tile parameters (ROADMAP #100, rice-pack stage)
+
+**Status:** OPEN 2026-10-06 — investigation landed, plan written
+(`docs/dev/TEXPACK-RICE-DRAW-PLAN.md`); no code change yet. Tooling-only,
+`feature/texpack` worktree.
+
+**What this corrects in D616.** D616 assumed the pack targets a recompile
+whose texnums are unreachable, and treated the pack's PNG filenames as the
+join target. The pack page says otherwise: it is a **GLideN64 HTS pack for
+emulators** (Project64 + mupen64plus; "HTS file goes to:
+Project64/Plugin/GFX/cache"), and `ge007.tdb` is the **emulator-side
+registry of every texture key**, 35517 lines, of which the 2046 pack PNGs
+are a small subset.
+
+**Findings.**
+
+1. **Join against the tdb, not the PNGs.** Parsing the tdb's 2567
+   non-CI keys (`GOLDENEYE#<crc>#<fmt>#<siz>_all;WxH`, hex upper-case,
+   `_all` suffix must be stripped before splitting on `#`) and hashing the
+   whole stored texture with the D616 recipe (`rice_hq` BE variant over
+   pool bytes at decode dims and pool stride) yields **9 hits** where the
+   PNG-filename join yielded **0**. The 1491 non-CI PNGs are mostly added
+   HD content (Effects/Explosion ×325, Smoke ×312, Gunbarrel ×299, UI
+   letters, character hacks) — they are not vanilla replacements, which is
+   why the PNG join is empty.
+2. **The recipe is confirmed correct**, not merely close: for the 9 hits
+   the tdb's `;WxH` equals the decode dims exactly, and the fmt/siz map
+   (`TEX_GBI_FMT`/`TEX_GBI_SIZ`) and the pool-stride formula agree with
+   `texGetLineSizeInBytes(tex,0)*8` in `src/game/tex.c` (the emulator's
+   `line<<3`).
+3. **Why 9/953 and not ~953**: the tdb carries ~2.7 keys per vanilla
+   texture. GLideN64 hashes **per draw tile** —
+   `src = RDRAM + tileAddr + ul_t*bpl + ((ul_s<<siz)+1)>>1`,
+   `RiceCRC32(src, tile.width, tile.height, fmt<<8|siz, bpl)`,
+   `bpl = tile.width<<siz>>1`, `rowStride = tile.line<<3` — and most GE
+   draws set a display-list `G_SETTILE` whose width/height/`ul_s`/`ul_t`
+   differ from the stored texture (sub-rectangles, tiled draws, repeats).
+   `bpl` (the inner word-walk extent) and `rowStride` (the row advance)
+   are different values for those draws.
+4. **Brute force is not the answer.** Exhaustive `(w,h,stride,offset)` and
+   BE/LE/swizzle/nibble-swap/decoded-pixel sweeps over the strongest
+   content pairing (texnum 2214, I4 128×48, key `1478129A`) score 0 — the
+   tile parameters must be read from the display lists, not searched for.
+   The C oracle for the ROM-side decode is also not the unblocking step:
+   the pool bytes are already content-verified against the pack PNG.
+
+**Next step** is the display-list walk in the plan doc (reuse
+`tools_pc/d43_gdldump.py` / `tools_pc/bg_gdl_census.py` framing); the tdb is
+the oracle for validating the resulting key set before touching the pack.
+
+## D618: The community HD pack's non-CI keys are per-draw rects that are partly unreachable from N64 tile state (440x1 strips), and a rect key cannot land in the texnum-keyed pack format — the rice-pack route stops here (ROADMAP #100, rice-pack stage)
+
+**Status: STOPPED (design decision, 2026-10-06) — the display-list walk in
+`docs/dev/TEXPACK-RICE-DRAW-PLAN.md` is not worth running.** Measured, not
+assumed; all numbers below are from `scratch/rice_domain_probe*.py`,
+`scratch/content_join_probe.py`, `scratch/ci_faithfulness_probe.py` (local
+analysis, not committed).
+
+**Setup.** Re-tested D617's diagnosis on the GoldenEye-007-HD pack
+(`ge007.tdb`, 35517 lines) before spending the plan's work.
+
+**Findings.**
+
+1. **The tdb's `;WxH` field is a per-draw rect, and some rects cannot be
+   N64 tile state at all.** Non-CI key dims: `64x64` ×623, `16x16` ×429,
+   `440x1` ×301, `64x32` ×187, `95x32` ×148, `65x65` ×84, `54x54` ×81,
+   `32x32` ×79, `128x47` ×56, `8x8` ×49, `32x64` ×48, `96x80` ×40. None of
+   440/95/128x47/96x80 exist in `imagelist.u.csv`. A `440x1` I8 rect is
+   440 bytes/row, over tmem's 128-byte pitch — so those 301 keys are
+   **recompile-side (RT64) textures, not loadable tile state**; no display-
+   list walk over the vanilla ROM can produce them. ~1/8 of the non-CI key
+   set is unreachable by construction, and the rest is per-draw tiling.
+2. **The byte domain is not recoverable by search.** ~100 domain variants
+   (pool / pool+N64 alt-row swap / decoded RGBA32 / decoded RGBA16, BE and
+   LE word reads, aligned / packed / `w*4` strides, RGBA/ABGR/ARGB/BGRA
+   orders, alpha = source/opaque/zero) over the 604 dim-matching candidates:
+   **0 hits**. The 9 whole-texture hits from D617 are coincidences (texnums
+   2085–2100, 2 distinct CRCs).
+3. **Content-join is not a workaround.** Downsampling pack PNGs to vanilla
+   dims and scoring mean-abs-diff over RGBA gives best scores 8–20 with
+   margins <3 against 200+ same-(fmt,siz) candidates — the pack's HD art is
+   a rework, not a faithful upscale, so it does not discriminate. (The
+   near-zero diffs seen earlier were a channel-count artifact of a raw-
+   length probe, not a match.)
+4. **The 256 entries we do emit are semantically sound.** `rt64.json`:
+   `defaultOperation: "stream"`, `operationFilters: []`, two `shiftFilters`
+   (`UI\Checkmark\*`, `UI\Crosshair\*` → `half`) — entries are straight
+   replacements, not masks. Of the 256 emitted: 199 color+opaque, 35
+   color+alpha, 22 grayscale (legitimate I/IA sources). Pack PNG dims are
+   exact integer multiples of the tdb rect (64x/32x/16x/8x).
+
+**The architectural blocker (why the walk would not finish the feature).**
+Even with correct per-draw keys, a rect key names a **sub-rectangle of a
+texture, drawn tiled per draw**. Our pack format (`data/texpacks/<name>/<texnum>.png`,
+D615) replaces a **whole texture**. Dropping a `95x32` tile's HD art at the
+parent texnum changes every draw that uses that texture, not just the one
+tile. Consuming this pack needs a format extension (rect-keyed entries:
+texnum + `ul_s,ul_t,w,h`, per-draw selection) — a second feature, not a
+missing step.
+
+**Decision / what ships instead.** HD + custom texture support is the
+verified texnum path: dumper → authoring (upscale or user art) →
+`data/texpacks/<name>/NNNN.png` → D615 A/B. Community HTS packs stay
+unsupported for non-CI content; the CI half (251) already imports. Optional
+later nicety, not content: the PD dabs-mod runtime rice index
+(`texpackBuildRiceIndex`/`texpackRiceLookup`) so whole-texture rice packs
+load with no offline step — it scores the same 251+9 on this pack, so it is
+architecture, not coverage.
+
+## D619: The community HD pack's remaining 1790 PNGs are blocked by four independent causes (recompile-side keys, wrong hash domain, rect-vs-whole-texture format, content outside the identity space) — feature status written up (ROADMAP #100)
+
+**Status: WRITTEN UP 2026-10-06 — `docs/dev/TEXPACK-STATUS.md` is the
+current-status read for #100; this entry is the evidence for the blocker
+list.** Extends D618 (which stopped the rice route) with the brute-force
+negative and the pack's own content taxonomy.
+
+**Setup.** Question: can a *full* HD import be launched from the GoldenEye-007-HD
+pack (2046 PNGs)? Measured, not assumed.
+
+**Findings.**
+
+1. **Brute-force rect join closes the non-CI route without a display-list
+   walk.** The tdb gives each key's rect dims, so enumerate every
+   sub-rectangle of every vanilla non-CI texture at those dims and hash the
+   pool bytes with the GLideN64 recipe (`bpl` from the rect width, row
+   advance = the source pool pitch, word-offset term rect-relative). Result:
+   **14.6M positions, 0 hits**, and 0 for the two other plausible models
+   (row advance = `bpl` contiguous-tile copy; absolute word-offset term).
+   Combined with D618's ~100 whole-texture domain variants (0 hits), the
+   non-CI keys are not derivable from vanilla-ROM bytes by any model tried.
+2. **The pack's own folder taxonomy says the rest is not texture-replacement
+   content at all:** `Effects` 734 (Explosion 339, Smoke 326, Death 40, …),
+   `UI` 558 (Gunbarrel 299, Letters 208, Crosshair 27, Ammo 19), `Areas/<Mission
+   NN - Name>` 481, `Hacks/<Mario Characters v3.17>` 264, `Objects/Weapons` 9.
+   `UI` is in-exe C-array art the registry never sees (D615 limitation); `Hacks`
+   are character-model swaps = model packs, the other half of #100. So the
+   ceiling on this pack is ~1200 of 2046, and only through a level-scoped
+   content join (per-level `GE_TEXNUMDUMP` active-texnum sets ∩ the tdb's
+   `(fmt,siz,WxH)`, then content-match with a margin threshold) — not yet
+   measured for discriminative power.
+3. **Four independent blockers, in severity order** (full detail in
+   `docs/dev/TEXPACK-STATUS.md`): **B1** keys name draw rects, not textures
+   (`440x1` ×301 exceeds tmem's 128-byte pitch → recompile-side, unreachable
+   by construction); **B2** the non-CI hash domain is the recompile's
+   converted bytes, not the ROM's pool bytes; **B3** a rect key cannot be
+   applied by the whole-texture `NNNN.png` format (needs a rect-keyed format
+   extension + per-draw selection); **B4** most entries are outside the
+   engine's identity space (HUD in-exe art, model swaps).
+
+**Landed/verified for the record:** 256 of 2046 import (251 zlib-CI joined
+by rice CRC + 9 coincidental non-CI); the texnum path itself (D615) is
+A/B-verified. Owed on that path (shippable checklist): branch merge, F10
+overlay row for `Video.TexturePack`, in-exe image registration, wrap period,
+upload-buffer size ceiling, first-use stall, pack manifest, region-tagged
+packs, user-facing authoring walkthrough.
+
+## D620: the HD pack's non-CI keys hash the *pack's own* re-encoded content — no vanilla/recompile byte domain produces them; the pack→texnum join must come from external signals (level folder + tile geometry). The recompile-side domain B2 named is actually measurable, and B2/B4 are refined, not overturned (ROADMAP #100, rice-pack stage)
+
+**Setup.** D619 left B2 ("the non-CI hash domain is the recompile's bytes")
+on an *unmeasured* recompile side. This session had the whole recompile
+side local: the `GoldenEye-007-Plus` source tree (its `imagelist.u.csv`,
+2698 rows, offset-aligned with vanilla), its image build docs
+(`scripts/IMAGE_BUILD_SYSTEM.md`), the 1964 GEPD bundle ROMs, and the two
+xdelta patches. Measure every remaining byte domain before accepting B2.
+
+**Findings.**
+
+1. **The xdelta "plus" ROMs are not usable as a measurement domain.**
+   `xdelta3 -d` (v4 *decode* — the patches carry no xdelta3-applyable
+   stream; only the output length is honoured) of
+   `goldeneye_007_plus(.xdelta|_v2)` onto the 12 MB vanilla base yields a
+   **32 MB** ROM differing from the base in 98 % of its first 12 MB —
+   another layout/endian than any of the GEPD bundle's 12 MB ROMs
+   (`007 - GoldenEye (USA).n64`, sha1 `7dd376e9…`, 95 % different from
+   vanilla, same size). Image bytes at the (shared) `imagelist.u.csv`
+   offsets in **all three** ROMs parse as no GE bitstream under any
+   ctrl-offset × bit-order × zlib/non-zlib interpretation (brute-forced
+   header scan: 0 valid fmt/nc/w/h). Every earlier "recomp byte" probe ran
+   on the corrupt xdelta output — its "2283 undecodable" is void. The
+   *correct* recompile-side byte set is inside the recompile tree:
+   `assets/images/combined/combined.bin` / `assets/images/split/*.bin`
+   (per its `IMAGE_BUILD_SYSTEM.md`, byte-perfect against the vanilla
+   region: sha1 `044fca472bf6…`), i.e. for untouched images, identical to
+   the vanilla ROM region — a domain D619 already measured (0 hits).
+2. **The pack's keys hash pack content; no byte domain on our side can
+   produce them.** Vanilla non-CI keys (953, from the verified
+   `rice_nonci.py` pipeline — decoded pool bytes, n64-swapped when
+   lod>0, rice_HQ CRC, `scratch/van_nonci_keys.csv`) hit **0** of the
+   tdb's 2567 non-CI keys. The pack's own 1491 `_all` keys
+   (I8×573 / IA8×389 / RGBA16×343 / I4×136 / RGBA32×41 / IA4×9) hit 0 of
+   the 953. The 553 `ciByRGBA` entries carry palcrcs (541 non-`$`) that
+   match a vanilla palCrc in only **3** cases → the pack **regenerates
+   palettes**, so even the CI-side palette channel of its keys is
+   pack-side. Non-CI/CI-by-RGBA keys are hashes of 1964tech's
+   re-encoding of HD art; a join has to come from signals *outside* the
+   pack: (a) level folder ↔ the level that draws the texnum, (b) tile
+   geometry, (c) elimination among a level's texnums after the CI-joined
+   siblings are removed.
+3. **The scale factor is wild, so a dimension bijection is dead.** Over
+   the 257 confirmed CI-joined pairs (251 keys, fan-out included) the
+   pack/vanilla dim ratio is 4×–128×: 32× ×135, 64× ×83, 16× ×31, 128×
+   ×6, 8× ×1, 4× ×1. But every joined pair sits in its *correct* level
+   folder (257 across 21 folders, e.g. texnums 423/240/1193/1258 in
+   `Areas/Mission 01 - Dam`) — the folder→level mapping is a working
+   signal, and D619's "unmeasured level-scoped join" should be formulated
+   as *folder → level's drawn-texnum set (per-level GDL census) +
+   elimination + tile-geometry scoring*, not a dims match.
+4. **B4 narrowed.** Image-ID sprite art (explosion/smoke/gunbarrel-style
+   banks) loads through `texLoad`/`texLoadFromTextureNum`
+   (`front.c:4964`, `image_bank.c:327`) → **already inside the registered
+   texnum path** (D615). D619's "822 PNGs in UI/Hacks are outside the
+   identity space" therefore applies only to (i) in-exe C-array art
+   (rareware logo, font echardata) and (ii) `Hacks` (character-model
+   swaps = model packs). Owed: one `GE_TEXNUMDUMP` pass over the
+   title/menu + an effects-heavy level to confirm the sprite/UI texnums
+   appear (cheap, closes (i) vs (ii) empirically).
+
+**Decision.** B2 is rewritten: "pack keys hash pack content; an
+external-signal join is required" (not "hash-domain mismatch to fix").
+B4 narrowed per 4. Next experiments, in order: (1) `GE_TEXNUMDUMP` on
+title/menu + effects-heavy level → registered-texnum confirmation;
+(2) folder→level bijection from the 257 joined pairs + a per-level GDL
+texnum census → measure the elimination ambiguity for the 481 `Areas` +
+734 `Effects` PNGs before committing to the content-join machinery.
+
+## D621: the PD dabs-mod texture-pack notes mine — the HD-pack route stays texnum-keyed; the PD port's *key* layer transfers nothing (its keys hash ROM bytes we hold; this pack's keys hash 1964tech's build, D620), but its *format* layer is a ready fix list; plan lands
+
+**Setup.** "How did the PD PC port get HD texture support, and why can't we
+adapt that?" — answered by reading the dabs-mod tree's
+`CLAUDE-notes/texture-packs.md` in full (the PD PC port fork that ships
+the PD Plus HD packs, 3395 images).
+
+**Findings.**
+
+1. **The PD port's HD support is the *same architecture* we have.**
+   `texpack.c` indexes `texture-packs/<name>/NNNN.png` by the port's own
+   texture number, substitutes at `import_texture()`, decodes off the
+   render thread, authors from an in-game F7 dump. The PD Plus HD pack
+   was authored **natively** that way — no key join exists to solve.
+   Our D615 system is this system.
+2. **The PD port's *emulator-pack* layer is a key-domain trick, and it
+   transfers nothing to our pack.** For Rice/GLideN64/1964tech
+   artifacts it **recomputes the Rice checksums itself, over the ROM
+   bytes it holds** (glyphs: 32 rows × 16 bytes at the 8-byte stride,
+   "exactly one variant matched"), then matches pack records
+   (`texpackIndexHtc`, `.htc`/`.dat` ingest). That works because PD's
+   asset pipeline reads the stock ROM at startup — the pack's key domain
+   *is* the ROM's domain. This HD pack's keys hash 1964tech's
+   re-encoded build (D620): there is no byte stream of ours to
+   recompute from. Same texnum architecture, different key domain —
+   the plan keeps the texnum route and drops the key route.
+3. **The *format* layer is a ready-made fix list for our importer**
+   (adopted as Phase A of `docs/dev/TEXPACK-PLAN.md`): `gfx_pad_replacement()`
+   (pack image = tile, not the padded 8-byte row — 278/3395 of the PD
+   pack were "stretched" by exactly this); `bottomup.txt`/`ext_tex`
+   row-order markers (rice packs are N64-row-order, must not flip);
+   `gfx_replacement_alpha()` (opaque replacement standing in for an
+   alpha texture gets repaired alpha — the PD notes describe our exact
+   I8 failure: light beams "a solid grey sheet, every smoke puff a
+   square"); F8/F9/F10 pack keys; worker-thread decode + kept store.
+   Our decision: bake pad/alpha into the **authoring tools**
+   (`rice_pack_import.py` emits padded, right-way-up, alpha-repaired
+   `NNNN.png`) so the port stays format-agnostic; keep the
+   in-port fallback for hand-made packs.
+
+**Decision.** Plan: `docs/dev/TEXPACK-PLAN.md` — Phase A (make the
+texnum path correct/shippable: pad contract, alpha repair, F10 row,
+stall, manifest, merge), Phase B (measure the geometric join: B0
+sprite-bank `GE_TEXNUMDUMP` confirmation, B1 folder→level table, B2
+per-level GDL census, B3 ambiguity gate, B4 content scorer, B5 review
+viewer), Phase C (land the join as `NNNN.png` + per-level A/B). Next
+task: B0.
+
+## D622: B0 sprite-bank confirmation — the `Effects`/`UI` sprite-bank texnums register and draw, so those PNGs stay in the join scope; the unregistered residue is exactly the known in-exe art; Phase B continues to B1–B3
+
+**Setup.** Phase-B step B0 of `docs/dev/TEXPACK-PLAN.md` (ROADMAP
+#100, plan D621): confirm the HD pack's `Effects/`
+(Explosion/Smoke/Death) and `UI/` (Gunbarrel/Crosshair/Ammo)
+sprite-bank texnums register through the texnum hook and draw, so the
+external-signal join (D620) can reach them — otherwise those ~1290 PNGs
+move out of scope and the ceiling drops. Two `GE_TEXNUMDUMP=1
+GE_TEXDUMP=1` runs, human-driven (the `GE_INPUTSCRIPT` headless attempt
+burned most of its level time in the intro cutscene — scripted combat is
+not reliable yet, human played the runs): (a) no level flag — front-end
+menus (title → mode select → 0100 → 0200, then the Dam level loaded);
+(b) `-level_09` with combat (firing + kills). Evidence in
+`scratch/b0/` (gitignored, regenerable): `texnumd_frontend.log`
+(445 lines / 377 unique texnums, 521 texture dumps),
+`texnumd_level09.log` (335 lines / 321 unique, 316 dumps), the PPM/RGBA
+dump sets, labeled contact sheets (`fe_*.png`, `lvl09_*.png`) and zoom
+sheets.
+
+**Findings.**
+
+1. **The `Effects` sprite-bank texnums register and draw.** The
+   explosion/smoke sprite banks on `-level_09` are texnums
+   2081–2196: grey smoke puffs (2086–2098), white flash discs / smoke
+   billows (2108–2112, 2176–2196), yellow/orange spark and flash
+   sprites (2119, 2128, 2157–2160), white crescents/dots (2171,
+   2177–2180). Every one of these appears in
+   `scratch/b0/texnumd_level09.log` (registered at `texLoad`) and was
+   drawn during the combat run (texnum-named dumps exist).
+2. **The `UI` crosshair is a registered texnum.** 2236 = the red
+   targeting reticle — registered and drawn on `-level_09`.
+3. **Front-end art is registered.** Mode-select icons 2194–2196,
+   mission-select photos 2552–2553 (128x64) and 2578–2598 (68x44),
+   agent portraits 2504–2508 (32x64 CI8), the "CLASSIFIED /
+   CONFIDENTIAL / EYES ONLY" roll 2618–2628 (95x32 CI8 — the pack tdb's
+   `95x32` rect keys now read as this game image geometry, D618),
+   waveform/cube/arrow sprites 2214–2225, 2188–2193.
+4. **The unregistered residue is exactly the in-exe C-array art
+   already out-of-plan (D619):** the echardata font glyphs (unnamed
+   `tNNN_8x7`–`tNNN_8x12` dumps, a–z/digits) and the 15 fire-particle
+   DL images (`fireNN_16x14`, D219). Nothing else unregistered was
+   observed.
+5. **`UI/Gunbarrel` is not a sprite-bank class.** GE renders
+   first-person weapons as textured 3D models, so the pack's 299
+   `Gunbarrel` PNGs are gun *model* textures — ordinary registered
+   texnums scoped to the level(s) that equip the weapon, joinable by
+   the same level-scoped route as `Areas`/`Objects`. Same for the
+   in-HUD art: `UI/Letters` (208) and `UI/Ammo` (19) are the in-exe
+   font/glyph art (finding 4) and stay out of plan. The
+   `Effects/Death` sub-class (40) was not visually isolated but is the
+   same sprite-bank mechanism (chrobj effect banks load via `texLoad`).
+
+**Verdict — B0 PASS.** The `Effects`/`UI` classes stay in the join
+scope; the ~1200 ceiling stands until B3 measures it. Along the way:
+the first dump run produced a 0-byte `texnumd.log` — the log's `FILE*`
+was buffered and the force-killed dump run lost it on exit;
+`port/src/texpack.c` now flushes after every line (port-layer fix, no
+game-code touch).
+
+**Next.** Phase B1 (folder→level table), B2 (per-level GDL texnum
+census), B3 (candidates-per-PNG ambiguity report + gate), per the plan.
+
+## D623: Phase B1–B3 measured — the folder→level mapping holds (241/244 Areas with the model pool), model DLs carry G_NOOP texnums, and the B3 ambiguity gate fails: median 104 candidates/PNG → B5 human-in-loop for the geometric tier
+
+ROADMAP #100, `docs/dev/TEXPACK-PLAN.md` Phase B (B1 folder→level, B2 per-level GDL census, B3 candidates-per-PNG ambiguity report + gate). Tooling only: `tools_pc/texpack/level_census.py`, `model_census.py`, `ambiguity_report.py` (+ `scratch` CSVs). No port C, no game code.
+
+**B1 — folder→level verified.** The 21 pinned folders (D620's 257 CI-joined pairs) extend to the pack's 15 `Areas/Mission NN` folders via `kSoloLevels` (`port/src/main.c`) + the `bg.c` levelinfotable: Dam 33, Facility 34, Runway 35, Surface 36 (+Surface2 43), Bunker 09 (+Bunker2 27), Silo 20, Frigate 26, then a constant +2 mission offset (Statue 22, Archives 24, Streets 29, Depot 30, Train 25, Jungle 37, Caverns 39, Aztec 28). Re-derived the 257-pair join with folder attribution (still 21 folders / 257 pairs — D620 stable). Against the bg-only census, 173/257 texnums (67 %) appear in their mapped level's bg DLs; the misses are model/prop/weapon textures (e.g. KF7 870/887, Mines 778–942, Dam 685–688) — adding the model pool (below) lifts containment to 241/244 (98.8 %); the 3 residual misses (279, 2473, 316 — one each in Facility/Surface/Bunker) are zlib-CI images that no static G_NOOP pool draws (shared-pool or cutscene art — still joinable, just not statically verifiable).
+
+**B2 — census, extended to model segs.** `level_census.py` walks the 17 bg segments (15 missions + Bunker2 + Surface2) and extracts texnums from G_NOOP words (`w1 & 0xFFF` = LOD texnum, `(w1>>12) & 0xFFF` = DETAIL second texnum, `tex.c:919`) → `scratch/level_texnum_census.csv` (1442 rows / 930 unique texnums; 412 without vanilla dims). New fact: **model display lists carry the same G_NOOP texnum convention** — `GproximitymineZ` → 944/945, `PlandmineZ` → 2469/2470, and 944/945 are inside the pack's `Objects/Weapons/Mines` joined range. `model_census.py` scans all 510 model segs (raw-deflate behind the 0x1172 header) → `scratch/model_texnum_census.csv` (3389 rows / 1462 texnums; 3314 in the image list). Model texnums are global (a model segs into every level that instantiates it), so the candidate pool for a level-scoped PNG is *that level's bg set* ∪ *model pool* ∪ *sprite bank* (B0 dumps), minus the 256 already-joined.
+
+**B3 — the gate fails (geometric-only).** `ambiguity_report.py` scores every in-scope pack PNG (2045) against its pool with class match (`ciByRGBA` ↔ rice-CI texnum) and the D620-learned scale set — the 251 joined pairs use power-of-two upscale factors (32×/64× dominate; 4/8/16/128 occur), so a candidate must be an *exact* pow2 multiple in both axes. Result (ex-Hacks 1781 PNGs; the 264 `Hacks/Mario` are pack-original fan content — out of join scope by design, though their dims do match character model textures):
+
+| tier | ex-Hacks count | share |
+|---|---|---|
+| 0 candidates | 515 | 29 % |
+| 1–3 (auto) | 44 | 2 % |
+| 4–20 (scorer-friendly) | 60 | 3 % |
+| >20 | 1162 | 65 % |
+
+Median **104** candidates/PNG → the plan's gate (`≤~3` auto / `≤~10` scorer / else human) lands on **B5 human-in-loop** for the geometric tier. The zero tier splits as predicted by D622: `UI` 454/558 zero (Letters+Ammo are the in-exe echardata fonts, never registered — exactly out of plan), `Effects` 40, `Areas` 21 (model textures without vanilla dims in the 405-dim non-CI fill), `Objects` 9. The >20 tier is dominated by the 32×32 CI class (1007 rice texnums) matching any 1024×1024 PNG at 32×. The `joined` set is derived from pack keys (RiceCRC32 ∩ rice census = 256 texnums, D616/D620-stable); the first report build wrongly subtracted the whole 1744-row rice census (100 % zeros for the CI class) — fixed before these numbers.
+
+**Verdict.** The external-signal join is *verifiable* (folder→level + model pool = 98.8 % containment of the 257 known pairs) but **not discriminative geometric-only**: 65 % of in-scope PNGs have >20 geometric candidates and 29 % have none. Per the gate, Phase B proceeds to **B4 (content scorer)** for the >20 tier and **B5 (human review)** for the 0/low tiers — or the ceiling is reset to the 104 auto/scorer-friendly PNGs (44+60). Decision owed: scorer budget (1162 PNGs × ~100 candidates each) vs. human tail vs. both. Evidence: `scratch/join_ambiguity.csv` (per-PNG candidates), `scratch/level_texnum_census.csv`, `scratch/model_texnum_census.csv`, `scratch/b0/folder_texnums.json`.
+
+## D624: B4 content scorer built + tuned — the structural RGB-NCC scorer (dither-blurred, palette-invariant) beats CLIP on this pack: rank-1 56 % vs 33 % on the 257 ground-truth pairs; the GT-tuned auto-accept tier (margin ≥ 0.10) is 100 % precise on GT; the pipeline produces the human queue (189 auto / 1341 vision-pending / 515 out-of-scope)
+
+ROADMAP #100 Phase B4 (`docs/dev/TEXPACK-PLAN.md`), the decision committed 2026-10-07 (512e6d57 + 95d38637): replace the B3 geometric-only gate with a 5-stage AI-scorer pipeline — geometric prefilter (D623's `join_ambiguity.csv`) → content rank → GT-tuned auto-accept → vision judge (agent contact sheets) → human queue for disagreements. Tool: `tools_pc/texpack/content_scorer.py` (subcommands `rank` / `tune` / `verdicts` / `apply`). Stop point honoured: the human-queue artifacts are produced; Phase C was not started.
+
+**Scorer A/B (measured on the 255 ground-truth pairs in candidate scope — 256 known pairs, D620; 1 pair's true texnum has no geometric candidates).** The GT set is the right tuning instrument: these are *known-correct* pack-PNG → vanilla-texnum pairs, so rank-precision and the auto-accept margin are measurable without a human.
+
+| scorer | rank-1 / 255 | rank-≤5 / 255 |
+|---|---|---|
+| CLIP ViT-B/32 (native pack scale) | 70 (27 %) | 136 (53 %) |
+| CLIP ViT-B/32 (pack → vanilla-dims LANCZOS) | 85 (33 %) | 153 (60 %) |
+| CLIP ViT-L/14 (scale-matched) | 84 (33 %) | 141 (55%) |
+| grey NCC, dither-blurred | 138 (54 %) | 171 (67 %) |
+| **RGB 3-ch NCC, dither-blurred (adopted)** | **143 (56 %)** | **170 (67 %)** |
+
+The **structural scorer wins decisively** and the result is interpretable: the pack is HD *redrawn* art with regenerated palettes, while the vanilla side is 15-colour dithered N64 art — a large perceptual domain gap that hurts a learned semantic embedding (CLIP), whereas a downscale-to-vanilla-dims + mean-centred per-channel normalised cross-correlation with a 1/2× LANCZOS dither-blur on both sides is invariant to palette regeneration (the D620 hue problem) and only needs matching *structure*. CLIP is retained as `--scorer clip` for A/B, not deleted. Both sides' dither-blur is the load-bearing step (grey NCC without blur: 51 % rank-1; with blur: 54–56 %).
+
+**Auto-accept margin (tuned on the same 255 GT pairs, structural scorer).** margin = score(true) − score(runner-up):
+
+| threshold | auto tier | precision on GT |
+|---|---|---|
+| 0.00 | 163 | 89.0 % |
+| 0.02 | 132 | 100 % |
+| 0.10 | 97 | 100 % |
+| 0.20 | 61 | 100 % |
+| 0.30 | 40 | 100 % |
+| 0.40 | 23 | 100 % |
+
+(margin percentiles: p10 −0.335, p50 0.053, p90 0.369). The 100 %-precision region extends down to t≈0.02, so a 0.20 threshold (24 % of GT) is over-conservative. **Recommended t = 0.10** — 100 % GT precise, 38 % GT coverage; rows above it auto-accept, the rest route to the vision judge. Note the precision numbers are *on the 255 GT pairs* (the known-correct set); the GT pairs skew toward "easy" matches (they were CRC-joined), so the non-GT tail will be noisier — the vision-judge stage exists exactly for that, and a vision "none" on an auto row is the escape hatch (the auto tier is a *routing* decision, not a proof).
+
+**Pipeline output (`verdicts --threshold 0.10 --vision`), 2045 pack PNGs:** 189 auto (47 geometric n_cand≤3 + 142 margin≥0.10), **1341 vision-pending** (one contact sheet each — pack + top-5 structural candidates, labelled with NCC sims — under `scratch/join_review/`, plus the `answers.csv` template), 515 out-of-scope (the D623 zero-candidate tier, dominated by the in-exe `UI` fonts). 1341 splits as 17 n_cand≤15 (top-5 ≈ the whole candidate set, most vision-resolvable) vs 527 with >40 candidates (top-5 is a weak hint there; most will answer "none" and land in the human queue). `apply` merges filled answers: an on-top-5 pick = vision verdict; a "none"/off-list pick = **human queue** (`scratch/join_review/queue.csv` + `queue/` sheets for the maintainer) — verified end-to-end on a 3-row sample.
+
+**Tooling notes.** Vanilla renders: the B3/D616 decode path (zlib-CI + non-CI, N64 alt-row swap for lod>0, 5551 palette) verified against the pack's own CRC-joined PNGs (e.g. texnum 294). Caches: `scratch/clip_{tex,pack}_cache.npz` (CLIP, optional) and `scratch/ncc_van_cache.pkl` (vanilla NCC vectors, ~0.3 s to load; a full `rank` pass is ~90 s, ~20 s/variant of GT tuning with cached decodes). One bug caught in testing: the struct `rank` writer initially unpacked `(texnum, score)` pairs as `(score, texnum)` — every `margin` came out as a texnum-difference (~−425), silently zeroing the margin auto-tier (a wrong "960 auto" count); caught because the CSV columns were eyeballed against a known row before verdicts ran.
+
+**Status: B4 BUILT + TUNED, stop point reached.** The human queue exists (1341 contact sheets + template); judging the 1341 vision-pending sheets (agent or maintainer, batchable) and running `apply` to finalise `queue.csv` is the next step, then B5 review / Phase C (out of scope here, per the drop-in prompt). No port C, no game-code changes.
+
+
+**Addendum 2026-10-10 (auto-accept hygiene, `content_scorer.py verdicts`):**
+a row is never auto-accepted when (a) its folder starts with `Hacks/` or
+contains `Fan Translations` (now `out-of-scope`), (b) the pack image's
+class disagrees with the top candidate's live GBI format class (`ciByRGBA`
+-> CI only, `_all` -> non-CI only; unknown candidate -> disagree), (c) the
+top-two margin is below the threshold (the old "<= 3 candidates" shortcut
+now applies only to a single candidate), or (d) the same texnum would be
+auto-assigned to more than one pack image in the same folder (all such rows
+demote to `vision-pending`). Pure helpers (`auto_eligible`,
+`demote_duplicates`, `class_agrees`) checked offline on a synthetic list;
+the `apply` subcommand and scratch data were not touched. The 189-auto
+figure above predates these rules and will drop on the next `verdicts` run.
+
+## D625: texpack A5 — downscale-to-GL-max (separable LANCZOS3) instead of the D615 vanilla fallback, plus first-use decode-stall metering — a pack's 4K art now renders on low-end GPUs instead of silently dropping to the vanilla texture; worst-frame decode stall is attributable per frame
+
+ROADMAP #100 Phase A5 (`docs/dev/TEXPACK-PLAN.md`). Two changes, port-only (`port/fast3d/gfx_pc.cpp`, `port/src/texpack.c`, `port/include/texpack.h`):
+
+**Downscale-to-GL-max.** A5 size census: 12 of 256 community-pack files are 2048/3072/4096 on the long side — over the GL max of low-end GPUs (and over the 8K-capped upload buffer on nothing). Before D625, the D615 guard dropped such files to the vanilla texture, which on a low-end GPU silently kills the HD look for exactly the pack's hero art. Now: the PNG is already decoded RGBA, so one separable LANCZOS3 resize (radius 3, scale-normalised, weight-renormalised at edges) into a lazily-grown `dw x sh` work buffer yields the largest size the GPU can hold; the result goes through the existing texture cache (one-time per file). The downscale output is `<= max^2 x 4 = tex_upload_buffer_size` by construction, so the only vanilla fallback left is an OOM'd work buffer or a degenerate GL max < 2 (both logged, at most 8 lines each). Non-downscaled images still `memcpy` straight into the upload buffer (the D615 path, unchanged). `max_tex_size_cached` is used only by the texpack path, so vanilla uploads are unaffected. `GE_FAKEMAXTEX=<n>` (registered in `docs/dev/GE-ENV-PROBES.md`) caps the effective ceiling so the low-end path is testable on a high-GL-max machine.
+
+**First-use stall metering.** The PNG decode is synchronous on the render thread (under `g_lock`), so a first-use import's cost is exactly the time in `texpackDecode`. New counters (render-thread owned, no extra lock): total/worst-file decode ms, per-frame accumulation committed at a `gfx_start_frame` hook (`texpackFrameStart`, inert unless a pack is selected), evictions (FIFO pressure). Summary line every 16th decode and every decode over 5 ms, `fflush`'d (D622-class force-killed runs must not lose it to stdio buffering).
+
+**Status: BUILT, not yet A/B-verified** — the downscale path needs a `GE_FAKEMAXTEX` run on a real pack + a low-end (or emulated) GPU pass; the stall meters need one pack run to read the first-use spike shape. **Label note: D621–D625 on `feature/texpack` collide with the mainline allocation (main's D621–D625 are the Deck double-advance / stray-polygon / Surface-2 / cursor / rocket findings); the branch keeps its local labels and they are renumbered to the next free numbers at land time (the branch forked before mainline claimed that range).**
+
+
+## D626: Windows DPI awareness was never declared — the hint sat behind PD's `PLATFORM_WIN32`, which this port never defines, so scaled displays got a stretched logical-size surface
+
+**Status:** FIXED 2026-10-10 (`port/fast3d/gfx_sdl2.cpp`), ROADMAP §4 "Windows display scaling" row.
+
+**Symptom.** On a 3840x2160 display at 125 % Windows scaling, fullscreen rendered a 3072x1728 surface that Windows stretched (soft picture); the maintainer's v0.5.0 4K captures were all 3072x1728. Workaround that worked: `SDL_WINDOWS_DPI_AWARENESS=permonitorv2`.
+
+**Cause.** `gfx_sdl_init` set `SDL_HINT_WINDOWS_DPI_AWARENESS` only under `#if defined(PLATFORM_WIN32)` (inherited from the PD port) and only when HiDPI was disabled. `PLATFORM_WIN32` is defined nowhere here (`port/include/platform.h` defines `PLATFORM_WINDOWS`), so the hint never compiled in on any path.
+
+**Fix.** Guard on `_WIN32` and set `permonitorv2` unconditionally on Windows, before `SDL_Init(SDL_INIT_VIDEO)`. Ini Width/Height are physical pixels (a 1280x720 window is now 1280x720 device pixels, smaller on screen than before at >100 % scaling).
+
+**Verified.** Same machine (125 %): fullscreen frame dump 3072x1728 -> 3840x2160; windowed 1280x720 unchanged in pixels; Facility golden PASS. Audit: no other undefined `PLATFORM_*` guards in `port/` (the rest come from `platform.h`).
+
+## D627: DexDrive / GameFAQs `.n64` Controller Pak dumps load directly — `ge007.eep` may be the dump itself; the GoldenEye note is extracted and handed to the D514 importer (ROADMAP §5a)
+
+**Status:** verify 2026-10-10 (`port/src/eepimport.c`, `tools_pc/eep_convert.py`). Port-only; by-eye mission select with a real DexDrive save owed.
+
+**Format (confirmed from the 36,928-byte GameFAQs sample `goldeneye-007.1080.n64`).** 0x1040-byte DexDrive header (`123-456-STD`, comment text at 0x40), then the 32 KB Controller Pak (128 x 256 B pages). Pak page 1 = inode table (128 x u16 BE, low byte = next page, 1 = end of chain); pages 3-4 = note table (16 x 32 B; game code @0, publisher @4, start page u16 BE @6). The GE note is code `3BADD1E5`, start page 5, chain 5 -> 6 -> end = 2 pages = the 512-byte cartridge EEPROM. Those 512 B zero-padded to 2048 are byte-identical to the previously hand-extracted `gamefaqs_dexdrive_2001.eep`, and all 6 regions convert under D514.
+
+**Implementation.** `geEepImportDexDriveSave(path)` runs first inside `geEepImportEmulatorSave` (no libultra.c change): file must start with `123-456-STD` (else untouched, old path), be at least 0x1040+0x8000 bytes (tail ignored, 1 MB read cap), contain a note with code `3BADD1E5` and start page 5..127, whose inode chain is exactly 2 pages, in range, ending at 1 (loop/out-of-range chains rejected). Original is backed up to `ge007.eep.dexdrive.bak` (.bak2... never overwriting), then `ge007.eep` is rewritten as 2048 B and the ordinary D514 conversion runs (second backup `.emulator.bak` of the extracted bytes). Any failure: one `eep:` LOG_ERROR line, file untouched, no backup; the game then treats it like any wrong-size/invalid save (its first 2048 bytes are header junk, every region fails CRC, slots wiped on first write) — same as before this change. `eep_convert.py dexdrive-to-pc IN.n64 OUT.eep` does the same offline.
+
+**Verified (no game window).** Build clean (no new warnings in eepimport.c; `abi_ratchet` OK). gcc harness around the real `eepimport.c` with a C `fileGenerateCRC`: the sample .n64 -> extraction identical to the reference, final file byte-identical to `eep_convert.py dexdrive-to-pc` output; truncated, header-only, magic-only, no-note, looping chain, out-of-range chain, bad start page: all rejected, file untouched, no backup. Plain emulator save path unchanged (identical output).
+
+**Owed:** drop a real DexDrive dump in as `data/ge007.eep` and check mission select / unlocks by eye. Other Controller Pak dump formats (raw `.mpk`, no header) are not handled.
+
+## D605: #145 "Invert look" inert under the game's own original styles (v0.6.0) -- FIXED 2026-10-09 (port-only; live pad pass owed)
+
+**Status: FIXED 2026-10-09 (v0.6.0 line, branch `fix/v060-pad-look-invert`).
+Two port-only changes; build green (Windows, v0.5.1 base). The issue
+reporter's live pad pass (Classic controls + 1.1, both invert states,
+watch row) is owed; the face-button/menu-role questions of the issue are
+the separate maintainer decision on the v0.6.0 ROADMAP row.**
+
+**Bug (GitHub #145, reported on v0.5.0):** with F10 "Control style"
+= Classic/Original + controller style 1.1, toggling F10 "Invert look"
+(`Input.PadLookInvertY`) had no in-game effect on vertical aim, and the
+game's own in-watch "Look up/down" row looked locked (it is the game's
+own UI, but D564 pins the value to its factory default every tick, so
+any change reverts).
+
+**Part (a) -- the port's row missed the look stick.** In the HONEY role
+(game styles 1.1/1.3; the port presets 1.2 Christmas / 1.3 Frost run it
+too, the latter with the D498/D518 physical stick swap) the LEFT stick
+is the analog look stick and the right stick feeds the digital
+C-button movement (E/D/C/F thresholds) -- the reverse of the SOLITARE
+role (1.2/1.4, where the right stick looks and is inverted on its
+scaled channel already). `inputPoll` inverted the right-stick raw in the
+HONEY branch, i.e. the *movement* C-button source, while the look
+channel passed through untouched. Fix (`port/src/input.c`): invert the
+left (post-swap look) channel `py` there, and stop inverting `ry`. The
+SOLITARE branch already inverted its look channel; menus are untouched
+(menu navigation always sources the left stick before this logic).
+Semantics: `Input.PadLookInvertY` now inverts the stick that LOOKS in
+the active role, in both schemes.
+
+**Part (d) -- the D564 pin now respects the scheme.** D564 removed the
+port UI row for the game's own "Look up/down" (`Bond.Look`) and pinned
+`cur_player_look_vertical_inverted` to 0 so a stale value could not flip
+the pad stick on the game's pitch path. In the Modern scheme that stays
+correct (the port writes the view directly on the mouse/centred-pad
+paths). In the ORIGINAL scheme the game honours the setting on the N64
+stick-pitch path (`bondview2.c` builds `moveData.invertPitch` from it),
+so pinning it removes the player's native control -- exactly the
+"locked" watch row the issue reports. Fix (`port/src/watchsettings.c`):
+gate the pin on `inputControlScheme() == 0` (new accessor over the D513
+`Input.ControlScheme`); in Original mode the in-watch row is the live
+control again, as on the N64, and the F10 rows still apply on top (a
+player who sets both inverts twice -- their choice, as on any system
+with two inversion layers).
+
+**Verification owed (maintainer / issue reporter):** Classic + 1.1 with
+a pad: F10 Invert look on/off flips vertical aim; the in-game watch
+"Look up/down" row accepts its toggle and takes effect; Modern scheme
+unchanged (pin still on, F10 rows as before). Build-checked only so far
+(the invert path is a sign flip in the pad poll; no headless probe
+covers it -- `GE_PADMAPTEST` exercises the action table, not this).
+
+## D606: #144 optional PD-style always-visible mission timer (v0.6.0) -- FIXED 2026-10-09 (port-only; smoke-verified, live level pass owed)
+
+**Request (v0.5.0 issue triage, #144):** an optional, always-on-screen
+mission timer, PD-style. Port-only QoL; must not move a golden byte when
+off.
+
+**Source of the time (the faithful N64 mission clock, not a wall
+clock):** `lvlGetCurrentMultiPlayerSec()` (`src/game/lv.c:1764`,
+declared `lv.h:79`) returns `g_CurrentMultiPlayerSec`, which
+`lv.c:1306/1308` advances each tick by the region clock accumulator:
+`D_80048394 += g_ClockTimer; g_CurrentMultiPlayerSec = D_80048394 /`
+`(50.0f on VERSION_EU, else 60.0f)`. That is the same accumulator the
+in-game watch TIME page and the `AI_IFMissionTime` conditions read, so
+the on-screen number is exactly the game's own mission time -- it runs
+while a stage is active and the menu is static, matching the N64
+watch.
+
+**Implementation (`port/src/optionsoverlay.c` only, `src/game`
+zero-diff):**
+- New F10 toggle row `Game.MissionTimer` (label "Mission timer",
+  `kOnOff`), default 0, registered in `overlayConfigInit` and in
+  `kResetDefaults`; row help added to `kRowHelp`.
+- `missionTimerTick()` runs once per emitted frame (same cadence as the
+  D213 FPS readout above it in `overlayEmitLocked`): with the toggle
+  off, or outside an active stage (`current_menu` neither
+  `MENU_RUN_STAGE` nor `MENU_INVALID`), no text is produced and the row
+  simply does not exist on screen; otherwise it formats the clock as
+  `TIME M:SS` (H:MM:SS at 3600 s and up).
+- Emit path: the closed-panel mini DL (top-right, green 0x40ff60ff,
+  same style as the D213 FPS line) now draws the FPS readout at y=6 and
+  the timer under it at y=22 when both are visible; the `!s_open` early
+  return fires only when neither has text, so with the toggle off the
+  overlay appends nothing and the golden dumps stay byte-identical.
+
+**Verification (2026-10-09, worktree smoke run):** 1200-frame @60 fps
+run with `Game.MissionTimer=1` renders **"TIME 0:20"** in the top-right
+(1200 frames / 60 = 20 s -- the timer tracks the game clock exactly,
+not wall time); the crop confirms placement and the green readout
+style. Build green. **Owed:** a live multi-minute level pass (the H:MM:SS
+format at 60+ minutes, and any overlap with other top-right content on
+a small window) -- cheap for the maintainer to eyeball.
+
+## D628: the HD-pack join's exact-CRC half holds up; its structural half was scored against dumps of the pack itself (ROADMAP #100)
+
+**Status:** CI half ACCEPTED, structural half VOID, 2026-10-10. Tooling +
+join table landed on `feat/texpack-join` (`tools_pc/texpack/gehd_join.csv`,
+`gehd_join.py`, `gehd_build.py`, `geh_spotcheck.py`); no port C change. The
+work was produced by the maintainer's local model; this entry is the
+integrator's verification of it.
+
+**What holds (independently re-verified, no game run).** 256 join rows, one
+per texnum, 256 distinct texnums (4..2696), no `Hacks/` rows. Cross-checking
+every row's pack filename against `scratch/rice_ci_fresh.csv`: **256/256
+agree on `(crc, fmt, siz)`** - the join is exact texel identity (D616's
+big-endian RiceCRC32), which is what makes it safe without any art judgement.
+The one palette-variant collision (texnum 4, `UI/Checkmark`) is resolved
+towards the variant whose pack palcrc equals the ROM palcrc (`5f045394`),
+with the loser logged. `gehd_build.py --dry-run` resolves all 256 sources,
+0 missing, 0 `(stage,texnum)` duplicates. Coverage is 251/2046 pack PNGs
+(14.1 % excluding `Hacks/`): levels whose area art the pack kept in CI join
+well (Statue 12/12, Jungle 13/15, Streets 23/30), the non-CI bulk
+(Effects 734, UI 558, Aztec 27) joins nothing.
+
+**The trap.** `GE_TEXDUMP` writes the texture **the GL upload receives**, and
+`texpackReplacement` has already substituted by then - so a dump taken with
+`Video.TexturePack` set *is* the pack image. In `scratch/b0/` (the D622
+Bunker-1 + front-end runs) **52 of 346 dumps are pack-substituted**: tn 26,
+151 and 205 are byte-identical to their pack PNG (max channel diff 0), at
+1024x1024 / 512x512 where the ROM texture is 32x32. All **45** CRC-joined
+texnums that the `precision` subcommand used as its hold-out are inside that
+52, so its "45/45 rank-1, NCC = 1.000" compares each pack PNG against a dump
+of itself. It measures nothing.
+
+**What the contamination then broke.** From those dumps the report concluded
+"HD packs replace at the same pixel size (45/45 identical dims, 0 diff-dim)"
+and rewrote the structural matcher to score only candidates whose dump dims
+*exactly equal* the pack's. The premise is false - against the ROM census the
+same 256 pairs are uniform upscales: 4x(1), 8x(1), 16x(30), 32x(135),
+64x(83), 128x(6), **0 at 1x** - which is D620's own 4x-128x measurement. With
+a real (vanilla) dump the true candidate is 32x smaller than the pack image,
+so the exact-dims gate excludes it by construction: the reported
+"0 structural matches, correctly rejected by the confidence bars" was
+**unmatchable, not rejected**, and the uniform-tile top-1s it saw (texnum 668
+x407, 264 x230) are the handful of other *contaminated* dumps that happened to
+share the pack's dims.
+
+**Gate rebuilt.** `Scene.scores` now admits a candidate when its dims are
+`pw/k x ph/k` for one common factor - same aspect ratio, no larger - and
+scores at the candidate's resolution (one LANCZOS coarse per query, BILINEAR
+per target, through the pre-existing helpers). This keeps the protection the
+exact-dims gate was reaching for (the 256x96 glyph *strip* vs the 64x32 glyph
+that produced the confident wrong match on texnum 83 is an aspect mismatch,
+so it is still excluded) without excluding every true pair. Smoke on the
+contaminated dumps: candidate pool median 168 (was effectively 0), 45/45
+hold-outs rank 1 - still circular on those dumps, so it proves the code path,
+not accuracy.
+
+**Non-CI is not a CRC problem (reconfirmed).** 1488 non-CI non-`Hacks` pack
+PNGs against the 953-row non-CI census: **0** `(crc,fmt,siz)` hits and 0
+crc-only hits - i.e. D540/D620 reproduce exactly. The report's reason ("the
+pack re-encoded that content") is right; the correction is only that this is
+settled, not newly open.
+
+**Decision / next step.** The CI half ships as the `gehd` pack (exact
+identity, no art judgement). The request for per-level `GE_TEXNUMDUMP` +
+`GE_TEXDUMP` runs on the other 15 levels is **declined**: it would need the
+pack switched off to be worth anything, and the existing D623/D624 pipeline
+already scores against python-decoded vanilla renders of the ROM - no game
+run at all, scale-aware, GT-tuned, with the 189-auto / 1341-sheet queue
+already built. Picking that tail up is a cheap-worker job, post-release.
+
+**One wart for whoever resumes D624:** `scratch/join_verdicts.csv` contains
+**63 `auto` rows for pack PNGs that already have an exact-CRC answer**, scored
+against a pool from which their own texnum was removed as already-assigned -
+so all 63 name a wrong texnum. Harmless as long as exact-CRC wins the merge
+(it does in `assemble`), but they must be dropped before any `apply`, and
+they are not evidence against D624's GT precision.
+
+## D632: exit BSOD 0x119 again - the kernel's process-exit GPU cleanup, not the game's quit ordering
+
+**Status:** FIXED 2026-10-10 on `fix/d632-exit-teardown` (off
+`release/v0.7.0`); port-only. Exit paths live-verified (addendum below).
+
+**Report.** The maintainer exited the game (texpack build, manual play
+session); the PC bugchecked 0x119 VIDEO_SCHEDULER_INTERNAL_ERROR, P1 0x3000,
+and rebooted (20:30). The game log ends in NUL padding (OS went down before
+its file cache was flushed). Same bugcheck as D344.
+
+**Evidence (first readable minidump, `!analyze -v`).** Process
+`ge007.x86_64.exe`; failure bucket
+`0x119_3000_UNKNOWN_dxgmms2!VidSchiDecrementHwQueueReference`. Stack, top
+down: `watchdog!WdLogEvent5_WdCriticalError` <- `dxgmms2!VidSchiDecrementHwQueueReference`
+<- `VidSchTerminateHwQueue` <- `dxgkrnl!DXGHWQUEUE::DestroyCoreState` <-
+`DXGCONTEXT::DestroyContext` <- `DXGDEVICE::DestroyDevice` <-
+`DXGPROCESS::Destroy` <- `DxgkProcessNotify` <- `nt!PspExitProcess` <-
+`NtTerminateProcess`. No game frame on the stack: the game had already
+exited; the kernel was destroying the GPU device and the hardware queues
+(hardware-accelerated GPU scheduling) that the process left behind, and its
+queue reference count was inconsistent. dxgkrnl 10.0.19041.7725,
+nvlddmkm 2026-09-04.
+
+**Why D344 did not cover it.** D344 makes the render thread `glFinish` and
+unbind the context at a frame boundary, then `exit(0)`. The context itself
+(and the window) were never destroyed, so every exit still handed the GL
+device + hardware queues to the kernel's process-exit cleanup - the path that
+faulted.
+
+**Fix.** `portAtExit` (after the window state and config are saved, before any
+D443 relaunch) calls `videoTeardownForExit()`: if the render thread parked,
+`gfx_sdl_teardown_for_exit()` deletes the GL context (`SDL_GL_DeleteContext`)
+and destroys the window while the process is alive, so the driver retires its
+queues through the normal context-destruction path. On a park timeout the
+teardown is skipped (the render thread may still be inside the driver);
+logged either way (`video: GL context and window destroyed` /
+`video: render not parked; GL teardown skipped`).
+
+**Confidence: medium.** The dump pins the faulting path to process-exit
+destruction of leftover GPU state, and the fix removes that state first; but
+the bugcheck is intermittent and is ultimately a driver/OS fault (HAGS queue
+refcount), so only exits without a bugcheck over time will show whether it
+holds. Workaround if it recurs: Windows Settings -> Display -> Graphics ->
+hardware-accelerated GPU scheduling off.
+
+**Verification owed (maintainer, when runs resume):** a few normal exits
+(menu Quit, window close, Alt+F4); each log should end with `video: exiting
+(render parked ...)` followed by `video: GL context and window destroyed`.
+
+**Addendum (2026-10-10, first live exits).** Menu Quit: `render parked after
+3 ms` -> `GL context and window destroyed`, no bugcheck. Window close:
+`render idle after 0 ms` -> `render not parked; GL teardown skipped` - the
+host's wait loop (D344) only waited while a frame was in progress, so a quit
+landing between frames exited before the render thread parked and the
+context still fell to the kernel. Fixed: the host waits for the park itself
+(bounded by QUIT_WAIT_MS); the next frame start parks instead of drawing, so
+it costs about one frame. **Re-checked 2026-10-10 (maintainer, integration
+build):** window close `render parked after 21 ms`, Alt+F4 `render parked
+after 6 ms`, each followed by `GL context and window destroyed`; no bugcheck /
+unexpected-shutdown events. All three exit paths now tear down in-process.
+Whether this removes the bugcheck for good can only show over time (it was
+intermittent); HAGS off remains the fallback.
+
+## D636: the non-CI fingerprint join was one swap rule away - 203 exact joins, and they correct the picture join
+
+**Status:** FIXED 2026-10-10 on `feat/texpack-fp` (off `release/v0.7.0`);
+tools + join table only, no port change. Follows D635.
+
+**Cause of the old 0/17.** `rice_nonci.py` (D617) hashed the alt-row-swapped
+pool only for textures with explicit LODs; everything else was hashed
+unswapped. The game runs `texSwapAltRowBytes` on every level of every texture
+(D149), so the RDP - and an emulator's fingerprint - always sees swapped
+rows. With the swap applied to every texture the census (953 non-CI
+textures) matches **203** texnums against the pack's `crc#fmt#siz` keys (17
+before). The load-width variant tested alongside made no difference here
+(unlike the font, D635). D540/D620's "the pack hashes its own re-encoded art"
+reading was wrong for the same reason D635's first probe was: an emulator
+looks a replacement up by the original texture's fingerprint.
+
+**Cross-check.** Of the 203, 136 had an independent picture/vision/maintainer
+join: 122 agree, 14 disagree. Every disagreement is a near-identical variant
+on the wrong texnum (35<->36, 284<->307, 145/146, 948/949, 1088/1089,
+440/441, 715/716, ...): a 32-bit CRC + format key cannot produce a neat swap
+of two siblings, so the picture join was wrong. Rebuilding the local pack
+exact-first dropped 30 picture picks whose image is exactly another
+texture's; 61 matcher labels (D631) were corrected to the exact owner.
+
+**Result.** `gehd_join.csv` now carries 459 exact rows (256 CI + 203 non-CI,
+method `exact-crc-nonci`), built by `gehd_join.py nonci <packroot>
+<rice_nonci.csv> <out>`. Local pack: 511 texnums (459 exact + 52 picture
+picks that no exact join contradicts). Lesson for the join: run the
+fingerprint join first and use picture matching only for what it cannot key
+(it has no reach into textures whose pack art is keyed by a different load).
+
+## D637: fingerprint fallback for textures without a texture number (the font)
+
+**Status:** BUILT 2026-10-10 on `feat/texpack-fp`; port-only, inert unless
+the selected pack has an `fp/` folder. Not run-verified (game runs on hold,
+D632): the first launch should show the HD font in the menus and HUD.
+
+**Loader** (`port/src/texpack.c`). `texpackInit` also scans
+`<pack>/fp/` (recursively, a few levels) for files named
+`GOLDENEYE#<crc>#<fmt>#<siz>...png`, adds them to the file table under a
+stage of their own (never matched by a texnum lookup; same decode cache,
+2048 cap and LRU as every pack file) and builds a sorted key index
+(crc, G_IM_FMT, G_IM_SIZ; first file wins per key). `texpackRiceCrc` is the
+GlideHQ/RT64 RiceCRC32 (big-endian words, word hash carried into the row
+end); checked equal to the Python implementation (`font_key_probe.py`,
+`rice_hq`) on 80 glyphs read from the ROM font banks.
+
+**Renderer** (`import_texture`). After a texture-cache miss, a texture with
+no registered texture number and a non-CI format is fingerprinted: row bytes
+= the render tile's line (`(width+7)&~7` bytes for the I8 font, the D635
+match condition), height = the tile's, row stride = the full image line for
+a LoadTile window, else the row; bytes in N64 order via
+`gfx_tex_normalize_source`, clamped to the loaded extent. A hit uploads the
+pack image under the vanilla cache key (so it is computed once per cache
+entry) through the same path as texnum replacements, including the D615
+transparency rules. Textures that have a texture number are untouched.
+
+**Builder.** `gehd_build.py --fp [folders]` copies pack folders under
+`fp/` keeping their names (default `UI/Letters`, the font). The authoring
+guide documents `fp/`.
+
+**Not covered yet:** paletted (CI) textures without a texture number (their
+emulator key also carries a palette CRC).
+
+## D638: HD ammo icons cut off - pack images vs the padded TMEM line
+
+**Status:** FIXED 2026-10-10 on `chore/v070-agentic-batch` (off
+`release/v0.7.0`); port-only. Maintainer report: "the ammo icon doesn't line
+up quite right, gets cut off" (the AK's).
+
+**Cause.** The vertex path normalises S/T by `tex_width`/`tex_height`, the
+render tile's TMEM line in texels (`line_size_bytes` per `siz`), which pads
+each row to the 8-byte line: the RGBA32 HUD ammo icons are 5-14 texels wide
+and are sampled as 8 or 16. A vanilla upload is that padded width, so the
+visible texels sit where the coordinates expect them. A pack image is drawn
+at the *visible* size (e.g. 160x384 = 32 x 5x12), so stretching it over the
+padded width showed 5/8 of the art, squeezed. Power-of-two surfaces (floors,
+walls) are not padded, which is why only small odd-width art showed it.
+
+**Fix.** In the pack upload path, after the GL-max downscale: if the tile's
+line exceeds its visible size (`(lrs-uls)/4+1`, `(lrt-ult)/4+1`) and the
+pack image scales the visible size uniformly (|sx - sy| <= 1 %), the image is
+copied onto a canvas of `line x scale` with the last column/row replicated
+into the padding, so the sampled region matches the original exactly. A
+tile that windows a larger image fails the uniform-scale test and is left
+alone; the canvas must fit the GL max and the upload buffer.
+
+**Verified** (headless `-level_33`, GE_PCDUMP, GE_QUITFRAME): `D638: texpack
+image 160x384 padded to 256x384 (visible 5x12 of a 8x12 TMEM tile)`, 80x448
+-> 128x448 (5x28 of 8x28), 1728x1728 -> 1792x1728 (54x54 of 56x54); the HUD
+cartridge renders whole and in proportion. Clean exit (render parked, GL
+teardown).
+
+## D639: exclusive fullscreen took the window size as the monitor mode
+
+**Status:** FIX BUILT 2026-10-10 on `chore/v070-agentic-batch`; port-only.
+Not live-verified: needs the maintainer's NVIDIA setup ("OpenGL present
+method = Prefer native", where it was seen 2026-10-09 during D583).
+
+**Cause.** `set_fullscreen` calls `SDL_SetWindowFullscreen(wnd,
+SDL_WINDOW_FULLSCREEN)` for `Video.FullscreenMode = 1`. Exclusive mode
+switches the display to the window's fullscreen display mode; none was ever
+set (`gfx_sdl_set_closest_resolution` exists in the window-manager API but
+nothing calls it), so SDL picks the mode closest to the current window size -
+e.g. the ini's 1440x1080 window - instead of the desktop resolution.
+Borderless (`SDL_WINDOW_FULLSCREEN_DESKTOP`) never changes the mode, which is
+why it was fine.
+
+**Fix.** Before entering exclusive fullscreen, set the window's display mode
+to the desktop mode of the display it is on; an explicit
+`set_closest_resolution` (none today) still wins.
+
+**Check owed:** Video.FullscreenMode = Exclusive, a non-native window size,
+toggle fullscreen: the monitor should stay at its native resolution.
+## D643: Original control style = a strict virtual N64 pad (#145 b/c, #149)
+
+**Status:** BUILT 2026-10-12 on `feat/original-pad`; port-only (`port/src/input.c`),
+Modern/Ext path untouched. Confidence: HIGH that the mapping is exactly the
+table below (GE_PADMAPTEST proves every source, in stage/menu/tank, plus remap
+and guards); MEDIUM on the stick-feel numbers (assumption, see below); live pad
+pass owed.
+
+**Cause.** `padMapPure` mapped pad buttons to port actions: PA_USE/PA_RELOAD
+ran dedicated `bond_interact_object()` / `attempt_reload_item_in_hand()` calls,
+so one button bound to both opened a door and reloaded (#149); the game's own
+native B ("use if usable, else reload", lv.c) only ran on a native B bit.
+Outside a stage a fixed lockout sent A,X -> A and B,Y -> B, so West accepted
+instead of going back (#145 c). PA_CROUCH / PA_GADGET / PA_PREVWPN had no N64
+button at all.
+
+**Fix.** With `Input.ControlScheme = 1` the published `PadBinds` carries an
+N64-button table (`n64[NB_*]`, up to two pad sources per button, per seat =
+pads[] index, D448/D518) and all port-action masks are zero, so
+`padMapPure` (flag `PadMapCtx.orig`) just ORs the bound native bits in every
+context and returns (no crouch, gadget edge, southpaw swap, dedicated
+use/reload; the slot-0 dedicated path reads no pad). Defaults (Switch Online
+N64 convention): South=A, West=B, RT=Z, LT/RB=R, LB=L, Start, D-pad=D-pad,
+right stick=C (digital, 50% threshold, fixed), left stick=analog; North/East
+and the stick clicks unbound. Back opens F10. Keys: `Input.N64Pad.<Btn>` (seat
+1) / `Input.N64Pad2..4.<Btn>`; Btn = A B Z R L CUp CDown CLeft CRight Start
+DUp DDown DLeft DRight; values are kPadSrc names ("a", "lefttrigger,rightshoulder",
+"none"). Guards: Start unbound -> default; A and B both unbound -> both
+default. The table is part of the config signature (rebuilt lock-free into the
+spare buffer as before). UI: ini-only (the F10 rebind rows are action-shaped
+and do not generalise cheaply; not added).
+
+**Sticks.** In Original the SOLITARE role swap is no longer applied: left
+stick is always the analog stick, right stick the C-buttons (D605's invert
+on the look channel is unchanged). `Input.N64StickFeel` (1 default, 0 = the
+Modern `scaleAxis` deadzone+rescale): 1 = linear 32767 -> +/-80 (STICK_MAX)
+with a fixed 2500/32767 (~7.6%) centre deadzone, no rescale. **Assumption to
+confirm vs 1964 / the N64 manual:** the 2500 deadzone figure and per-axis
+(not radial) application. The port had no acceleration curve in the pad path;
+what it added in Original was the user deadzone (25% default) rescale -- that
+is what the switch bypasses (look sensitivity/smoothing only existed in the
+now-bypassed SOLITARE branch).
+
+**Verify.** `GE_PADMAPTEST=1 GE_QUITFRAME=3 ge007.x86_64.exe` (needs a GL
+window; SDL dummy video fails): preset0 == legacy 20971520 cases 0 mismatches,
+Original per-source table x 3 contexts, remap/seat independence,
+unbound-fallback, stick -> PASS.
+
+## D644: File-select royal crest drew with an opaque dark square (#146): IA4 decoded as I2:A2 instead of I3:A1
+
+**Symptom.** Each file-select folder's gold crest sat on an opaque dark noisy square; N64/1964 shows only the gold.
+
+**Evidence.** `GE_D526` census at frame 1150 (crest region): the crest is four 65x65 tiles, `fmt=3 (IA) siz=0 (4b)`, 2-cycle, `oml=0x0c184340` (FORCE_BL, blend CLR_IN*A_IN + CLR_MEM*(1-A_IN)), `comb=0x00f38e4f020a2d12` (alpha = tex alpha x shade alpha); the blender path was already correct (`alpha=1`). The raw texture (`GE_TEXRAW`) holds only nibbles {0, 2, 4, 15}. Real RDP IA4 is I3:A1 (nibble = I<<1 | A): 15 = opaque white, 0/2/4 = transparent (A=0). `import_texture_ia4` (D266) decoded I2:A2, so nibble 2 -> alpha 170 and nibble 4 -> alpha 85 with black intensity: the dark square. D266's "hardware-correct" claim was never verified (its own census found zero IA4 draws at its target scene).
+
+**Fix.** `port/fast3d/gfx_pc.cpp` `import_texture_ia4`: I3:A1 (intensity = nibble>>1 scaled by 255/7, alpha = nibble&1 ? 255 : 0). Removed the I2:A2 comment on `SCALE_2_8` (macro kept, unused). Before/after crops: `crest_before.png`, `crest_after.png`.
+
+**Scope.** Every IA4 texture in the game (any level or UI). Needs the lead's sweep/verify.sh re-run; the old I2:A2 tree-strip motivation (D236/D265) never reproduced as an IA4 draw.
+
+**Confidence.** High for cause and fix on the file-select crest (visually verified, gold intact). Medium on other IA4 users until the sweep.
+
+## D645: Original 1.1: F10 "Invert look" did not invert vertical look (#145 a, follow-up to D605/D643)
+
+**Symptom.** Maintainer playtest 2026-10-12 (Original control style, 1.1): toggling F10 "Invert look" did not change vertical look.
+
+**Cause.** D605 inverted the left stick's Y in the HONEY role. In 1.1 the left stick moves; hip-fire pitch is C-up/C-down (the right stick under D643's strict pad) and the game applies its own "Look up/down" option (`moveData.invertPitch`) to C-buttons and the aim-mode stick. The port's row never reached that path.
+
+**Fix (port only).** `port/src/input.c`: the HONEY-role stick flip is skipped in Original; accessors `inputPadLookInvert` / `inputSetPadLookInvert`. `port/src/watchsettings.c`: in Original, F10 "Invert look" and the watch "Look up/down" row are one setting: whichever changed since the last tick wins, and on stage entry the F10 row wins. Modern scheme unchanged (the D564 pin still applies there).
+
+**Status.** FIXED, maintainer-verified 2026-10-12 on Windows: inverts in 1.1; the F10 and watch rows stay in sync.
+
+## D646: F10 dropdown popup drew over the hint/tooltip panel
+
+**Symptom.** Maintainer 2026-10-12: the Video page "Frame rate cap" dropdown opened downward over the tooltip panel below the settings box (both translucent, text overlapped).
+
+**Cause.** `ddGeom` (`port/src/optionsoverlay.c`) only flipped the popup upward when it would pass the screen edge (`ovH() - 4`), not the settings box bottom (`o.bottom`); the hint panel sits between the two.
+
+**Fix.** The popup is bounded to the settings box (`o.titleB + 1` .. `o.bottom - 1`): it opens upward when there is no room below, and the D568 fallback now pins it flush to the box bottom instead of the screen.
+
+**Status.** FIXED, maintainer-verified 2026-10-12.
diff --git a/docs/framerate-and-pacing.md b/docs/framerate-and-pacing.md
index a323a20a..6dbc4dff 100644
--- a/docs/framerate-and-pacing.md
+++ b/docs/framerate-and-pacing.md
@@ -30,9 +30,10 @@ lighter values for draw distance, LOD distance and anti-aliasing into
 
 ### Above 60 fps
 
-`Video.FpsCap` (F10, *Frame rate cap*) accepts Auto, 30, 60, 90, 120 and 144.
+`Video.FpsCap` (F10, *Frame rate cap*) accepts Auto, 30, 60, 90, 120, 144, 165 and 240.
 **Auto** is the default on fresh installs (an existing `ge007.ini` keeps its
-value) and matches the display's refresh rate, up to 144; on a 60 Hz display it
+value) and matches the display's refresh rate, up to 240 (the renderer draws at most four
+frames per game tick); on a 60 Hz display it
 is plain 60.
 
 - **The game still simulates at its original 60 Hz (NTSC) tick.** The extra
diff --git a/docs/index.md b/docs/index.md
index bf900dfb..acb654f2 100644
--- a/docs/index.md
+++ b/docs/index.md
@@ -9,7 +9,7 @@ description: >-
 # and the WebPage JSON-LD dates). Keep these equal to the last real content
 # change of THIS page, not to the last release.
 date: '2026-10-05'
-modified_time: '2026-10-10'
+modified_time: '2026-10-13'
 locale: en_US
 # The masthead is this page's title, so its first content h2 must NOT get the
 # amber page-title look (double header). See the .flat-title rule in
@@ -32,9 +32,9 @@ Rare "Indy" engine family, one hardware generation apart.
 
 <div class="cblock" markdown="1">
 
-### v0.6.0 (RELEASE-DATE) — frame rates above 60, experimental macOS, fixes.
+### v0.6.0 (2026-10-13) — frame rates above 60, experimental macOS, fixes.
 
-- **Frame rates above 60:** the frame cap offers 90, 120 and 144, and **Auto**
+- **Frame rates above 60:** the frame cap offers 90, 120, 144, 165 and 240, and **Auto**
   (the default on fresh installs) matches your display. Frames between game
   ticks are blended; the game still simulates at its original rate. It costs
   one game tick of latency and does nothing on a 60 Hz display. A rework of
diff --git a/docs/porting-notes.md b/docs/porting-notes.md
index 0677be5f..c64364ad 100644
--- a/docs/porting-notes.md
+++ b/docs/porting-notes.md
@@ -1985,3 +1985,28 @@ On the Deck (Game Mode: stderr is dropped, so instruments must go to ge007.log v
 ## D41. A long frame time on one render thread is not draw cost until you line presents up against draws (D583)
 
 Before optimising a "slow" interpolated frame, check what its time contains: with one worker that both draws and presents, a frame's measured draw includes nothing but work, yet the presents queued behind it go out late by up to the whole draw, and the symptom (uneven 8/14 ms gap pairs at full present rate) looks exactly like an over-budget pass. Profile on-CPU time (`perf` works on SteamOS, `perf_event_paranoid` 2 is enough for your own process) and rule out the GPU clock (QAM Manual GPU Clock) before touching batching; then log each draw's start/end next to each present's due/late time (`GE_INTERPLOG=<n>`). On the Deck in Game Mode, gamescope only shows windows Steam launched: a harness started over ssh runs hidden and throttled, so launch through a Steam shortcut (`tools_pc/deckrunner.sh`).
+## D42.. Combiner alpha only reaches the framebuffer through the blender: no FORCE_BL, no blend (D414)
+
+fast3d's `use_alpha` (blender inputs CLR_MEM/1MA) is a heuristic for "this is an alpha-blended draw". On the RDP the blender only acts on interior pixels when `FORCE_BL` is set (or via coverage/`CVG_X_ALPHA`); a render mode with XLU-style inputs but no `FORCE_BL` is opaque regardless of combiner alpha. Vanilla GE never noticed because its combiner alpha there is `SHADE` alpha (255). A mod's DL with cycle-2 alpha `(0,0,0,0)` drew fully transparent = black. GL blend is now gated on `FORCE_BL | CVG_X_ALPHA`.
+
+**Tell:** rasterized with correct depth, immune to vertex-shade boosting, correct raw texels, black output. **Bisect fast:** output raw `texVal0/texVal1` in the fragment shader (bypasses combiner + blend); if that is right, the fault is combiner or blend state, not textures.
+
+## D43. A capture taken through the feature under test is not ground truth (D628)
+
+`GE_TEXDUMP` dumps what the GL upload receives, which is the *texture pack's*
+image whenever a pack supplies that texnum - so dumps captured with
+`Video.TexturePack` set cannot validate a pack join: the hold-out compares the
+pack PNG with a dump of itself and scores a perfect 1.000. 52 of 346 dumps in
+one such set were pack-substituted, and the false "pack and ROM textures are
+the same size" premise read off them then gated the matcher down to nothing
+(real packs are 4x-128x upscales). Two habits avoid the whole class: capture
+ground truth with the feature OFF (and record in the dump dir which switches
+were set), and before trusting a precision number, check that the hold-out's
+dims/format still look like the *source* data rather than the replacement.
+A 100 % score on a derived metric is a circularity smell, not a result.
+
+**"Faithful" input mode = emit native bits and add nothing (D643).** A port that routes pad buttons to PC *actions* (dedicated use/reload calls, crouch, gadget cycle) cannot also claim to be the N64 pad: the game's own handlers (e.g. B = use-else-reload) only run on the native bit, so a button bound to two actions double-fires, and a fixed menu lockout reinterprets faces. For a strict mode, give the published bind table a native-button layer, zero every port-action mask in it (so side paths that read the action bitset, like a dedicated use/reload hook, go inert without being touched), and make the pure mapper `orig`-flagged and context-free; prove it with the same boot self-test (per-source table across stage/menu/tank). Also stop role-swapping sticks to match the game's selected style: a virtual pad keeps left = analog, right = C.
+
+## D44. RDP IA4 is I3:A1; a decode tweak justified by "hardware layout" needs a texel dump (D644)
+
+IA4 = 3-bit intensity + 1-bit alpha (nibble = I<<1 | A). An I2:A2 decode (D266) turned the A=0 nibbles 2/4/6 into semi-opaque black and drew the file-select crest on a dark square. When a texture's transparent texels render opaque, histogram the raw nibbles/bytes (`GE_TEXRAW` + `GE_D526` for the draw's fmt/oml/comb) before touching blender or alpha-compare code: the importer's alpha expansion is the cheapest suspect.
diff --git a/port/fast3d/gfx_pc.cpp b/port/fast3d/gfx_pc.cpp
index c501c14d..507cdc0d 100644
--- a/port/fast3d/gfx_pc.cpp
+++ b/port/fast3d/gfx_pc.cpp
@@ -51,8 +51,6 @@ uintptr_t gfxFramebuffer;
 #define SCALE_4_8(VAL_) ((VAL_)*0x11)
 #define SCALE_8_4(VAL_) ((VAL_) / 0x11)
 #define SCALE_3_8(VAL_) ((VAL_)*0x24)
-/* D266: RDP IA4 is I2:A2 (2-bit intensity, 2-bit alpha), not the I3:A1 the
- * PD-derived importer assumed. */
 #define SCALE_2_8(VAL_) ((VAL_)*0x55)
 #define SCALE_8_3(VAL_) ((VAL_) / 0x24)
 
@@ -959,16 +957,17 @@ static void import_texture_ia4(int tile, const LoadedTexture& loaded_texture, bo
     for (uint32_t i = 0; i < size_bytes * 2; i++, dest += 4) {
         const uint8_t byte = addr[i / 2];
         const uint8_t part = (byte >> (4 - (i % 2) * 4)) & 0xf;
-        /* D266: I2:A2 -- the hardware IA4 layout. The old I3:A1 read made the
-         * A2=1 canopy pixels fully opaque and the A2=2 edge dither invisible
-         * (D236/D265 tree "wall"). */
-        const uint8_t intensity = part >> 2;
-        const uint8_t alpha = part & 3;
-        const uint8_t c = SCALE_2_8(intensity);
+        /* D644: RDP IA4 is I3:A1 (3-bit intensity, 1-bit alpha). D266's
+         * I2:A2 read was unverified; it turned the even (A=0) nibbles 2/4/6 into
+         * alpha 170/0 (and I2 as dark) and drew the file-select crest's transparent
+         * texels as a dark square (#146). */
+        const uint8_t intensity = part >> 1;
+        const uint8_t alpha = part & 1;
+        const uint8_t c = (uint8_t)((intensity * 255) / 7);
         dest[0] = c;
         dest[1] = c;
         dest[2] = c;
-        dest[3] = SCALE_2_8(alpha);
+        dest[3] = alpha ? 255 : 0;
     }
 
     const uint32_t width = rdp.texture_tile[tile].line_size_bytes * 2;
diff --git a/port/fast3d/gfx_sdl.h b/port/fast3d/gfx_sdl.h
index 4cdd9127..d4231853 100644
--- a/port/fast3d/gfx_sdl.h
+++ b/port/fast3d/gfx_sdl.h
@@ -18,6 +18,8 @@ void gfx_sdl_make_context_current(void);
 void gfx_sdl_release_context(void);
 /* D344: finish queued GL work and unbind the context (render thread, quit). */
 void gfx_sdl_park_for_exit(void);
+/* D632: delete the GL context + window (host thread, exit, render parked). */
+void gfx_sdl_teardown_for_exit(void);
 
 /* Refresh the cached drawable size after a window resize (called by the
  * host thread's event pump). */
diff --git a/port/fast3d/gfx_sdl2.cpp b/port/fast3d/gfx_sdl2.cpp
index 04119a2b..a7c42b86 100644
--- a/port/fast3d/gfx_sdl2.cpp
+++ b/port/fast3d/gfx_sdl2.cpp
@@ -61,8 +61,24 @@ static void gfx_sdl_set_fullscreen_flag(int32_t mode) {
     }
 }
 
+/* D639: set once a resolution is picked explicitly (gfx_sdl_set_closest_resolution). */
+static bool exclusive_mode_chosen = false;
+
 static void set_fullscreen(bool on, bool call_callback) {
     fullscreen_state = on;
+    /* D639: exclusive fullscreen switches the monitor to the window's display
+     * mode; with none set, SDL derives it from the current WINDOW size, so a
+     * 1440x1080 window asked the monitor for ~1440x1080 instead of its desktop
+     * resolution (seen with NVIDIA "OpenGL present method = Prefer native").
+     * Unless a resolution was picked explicitly, use the desktop mode of the
+     * display the window is on. */
+    if (on && fullscreen_flag == SDL_WINDOW_FULLSCREEN && !exclusive_mode_chosen) {
+        SDL_DisplayMode dm;
+        const int disp = SDL_GetWindowDisplayIndex(wnd);
+        if (SDL_GetDesktopDisplayMode(disp >= 0 ? disp : 0, &dm) == 0 && dm.w > 0 && dm.h > 0) {
+            SDL_SetWindowDisplayMode(wnd, &dm);
+        }
+    }
     SDL_SetWindowFullscreen(wnd, on ? fullscreen_flag : 0);
     if (call_callback && on_fullscreen_changed_callback) {
         on_fullscreen_changed_callback(on);
@@ -86,17 +102,13 @@ static void gfx_sdl_get_active_window_refresh_rate(uint32_t* refresh_rate) {
     *refresh_rate = mode.refresh_rate;
 }
 
-// D581 (#151): SDL display 0 is not the OS primary on multi-monitor setups. The
-// primary display is the one whose bounds contain the desktop origin (0,0)
-// (X11/Xwayland/Windows convention); fall back to display 0.
+// D581 (#151): SDL lists the OS primary display first (X11 RandR primary,
+// Windows MONITORINFOF_PRIMARY, macOS main display), so display 0 IS the
+// primary. The original bug was centring with display 0's size as absolute
+// coordinates, ignoring its position; callers now use its global bounds.
+// Not "the display containing (0,0)": on X11 the primary can sit right of a
+// secondary at the origin (the #151 layout; verified on KDE X11, 2026-10-13).
 static int primary_display_index() {
-    const int n = SDL_GetNumVideoDisplays();
-    for (int i = 0; i < n; ++i) {
-        SDL_Rect r;
-        if (SDL_GetDisplayBounds(i, &r) == 0 && r.x <= 0 && r.y <= 0 && r.x + r.w > 0 && r.y + r.h > 0) {
-            return i;
-        }
-    }
     return 0;
 }
 
@@ -108,12 +120,16 @@ static void gfx_sdl_init(const struct GfxWindowInitSettings *set) {
     if (!set->allow_hidpi) {
         // HiDPI control, if available
         SDL_SetHint(SDL_HINT_VIDEO_HIGHDPI_DISABLED, "1");
-#if defined(PLATFORM_WIN32) && defined(SDL_HINT_WINDOWS_DPI_AWARENESS)
-        // if HiDPI is disabled, declare ourselves DPI aware to get 1:1 window size on Windows
-        SDL_SetHint(SDL_HINT_WINDOWS_DPI_AWARENESS, "permonitor");
-#endif
     }
 #endif
+#if defined(_WIN32) && defined(SDL_HINT_WINDOWS_DPI_AWARENESS)   /* PD used PLATFORM_WIN32, never defined here */
+    // Always declare per-monitor DPI awareness on Windows so the window and
+    // fullscreen surfaces are 1:1 with physical pixels. Without it, a scaled
+    // display (e.g. 3840x2160 at 125%) hands the game a 3072x1728 surface and
+    // stretches it (soft picture) -- this used to be set only when HiDPI was
+    // disabled. Width/Height in the ini are physical pixels either way.
+    SDL_SetHint(SDL_HINT_WINDOWS_DPI_AWARENESS, "permonitorv2");
+#endif
 
     // D566: the port never takes text input, so never ask for an on-screen
     // keyboard (Steam Deck Game Mode popped one at every launch). Literal
@@ -328,16 +344,24 @@ static void get_centered_positions_native(int32_t width, int32_t height, int32_t
     const int disp_idx = SDL_GetWindowDisplayIndex(wnd);
     SDL_DisplayMode mode = {};
     SDL_GetDesktopDisplayMode(disp_idx, &mode);
-    *posX = mode.w / 2 - width / 2;
-    *posY = mode.h / 2 - height / 2;
+    SDL_Rect b = { 0, 0, 0, 0 };   /* D581: the display's desktop offset */
+    if (disp_idx < 0 || SDL_GetDisplayBounds(disp_idx, &b) != 0) { b.x = 0; b.y = 0; }
+    *posX = b.x + mode.w / 2 - width / 2;
+    *posY = b.y + mode.h / 2 - height / 2;
 }
 
 static void gfx_sdl_get_centered_positions(int32_t width, int32_t height, int32_t *posX, int32_t *posY) {
     const int disp_idx = SDL_GetWindowDisplayIndex(wnd);
-    SDL_DisplayMode mode = {};
-    SDL_GetCurrentDisplayMode(disp_idx, &mode);
-    *posX = mode.w / 2 - width / 2;
-    *posY = mode.h / 2 - height / 2;
+    /* D581: centre on the window's display in global coordinates (its bounds
+     * carry the desktop offset; the mode size alone lands on the leftmost). */
+    SDL_Rect b = { 0, 0, 0, 0 };
+    if (disp_idx < 0 || SDL_GetDisplayBounds(disp_idx, &b) != 0 || b.w <= 0 || b.h <= 0) {
+        SDL_DisplayMode mode = {};
+        SDL_GetCurrentDisplayMode(disp_idx < 0 ? 0 : disp_idx, &mode);
+        b.x = 0; b.y = 0; b.w = mode.w; b.h = mode.h;
+    }
+    *posX = b.x + b.w / 2 - width / 2;
+    *posY = b.y + b.h / 2 - height / 2;
 }
 
 static void gfx_sdl_set_closest_resolution(int32_t width, int32_t height, bool should_center) {
@@ -345,6 +369,7 @@ static void gfx_sdl_set_closest_resolution(int32_t width, int32_t height, bool s
     const int disp_idx = SDL_GetWindowDisplayIndex(wnd);
     SDL_DisplayMode closest = {};
     if (SDL_GetClosestDisplayMode(disp_idx, &mode, &closest)) {
+        exclusive_mode_chosen = true;   /* D639: an explicit pick wins over the desktop mode */
         SDL_SetWindowDisplayMode(wnd, &closest);
         SDL_SetWindowSize(wnd, closest.w, closest.h);
         if (should_center) {
@@ -457,6 +482,25 @@ extern "C" void gfx_sdl_park_for_exit(void) {
     }
 }
 
+/* D632: host thread, at exit, after the render thread has parked (context
+ * unbound everywhere, GPU drained by gfx_sdl_park_for_exit). Delete the GL
+ * context and the window while the process is still alive, so the driver
+ * retires its hardware queues through the normal path. Left alone, they are
+ * destroyed by the kernel inside process teardown (DXGPROCESS::Destroy ->
+ * DXGCONTEXT::DestroyContext -> VidSchTerminateHwQueue), which is exactly
+ * where the 0x119 bugcheck of 2026-10-10 fired (minidump, HAGS on). */
+extern "C" void gfx_sdl_teardown_for_exit(void) {
+    if (ctx) {
+        SDL_GL_MakeCurrent(NULL, NULL);   // host never holds it after init; harmless
+        SDL_GL_DeleteContext(ctx);
+        ctx = NULL;
+    }
+    if (wnd) {
+        SDL_DestroyWindow(wnd);
+        wnd = NULL;
+    }
+}
+
 /* Release the GL context from the calling (host main) thread. WGL only allows
  * a context to be current on one thread at a time, and while another thread
  * still holds it, wglMakeContextCurrent fails with "The requested resource is
diff --git a/port/include/video.h b/port/include/video.h
index 1be74bfc..9ca03a31 100644
--- a/port/include/video.h
+++ b/port/include/video.h
@@ -103,6 +103,10 @@ int  videoIsFullscreen(void);
  * on a clean exit). No-op if the window isn't up. */
 void videoSaveWindowState(void);
 
+/* D632: delete the GL context + window at exit (only if the render thread
+ * parked). Called from the atexit handler after configSave. */
+void videoTeardownForExit(void);
+
 #ifdef __cplusplus
 }
 #endif
diff --git a/port/src/input.c b/port/src/input.c
index fca36f65..5fc60b3b 100644
--- a/port/src/input.c
+++ b/port/src/input.c
@@ -465,7 +465,33 @@ static const struct { const char *key; const char *label; const char *def; } kPa
     [PA_START]   = { "Start",      "Start / pause",    "start" },
 };
 
-typedef struct { u32 mask[PA_COUNT]; } PadBinds;
+/* D643 (#145 b/c, #149): Original control scheme = a STRICT virtual N64
+ * controller. Each native N64 button is bound to up to two pad sources
+ * (Input.N64Pad[N].<Button>), and a press emits exactly that bit in every
+ * context (stage, menu, watch, tank) -- no port action (dedicated use/reload,
+ * crouch, gadget cycle, prev-weapon) is injected; the game's own code decides
+ * what A/B/Z/... mean (B = "use if usable else reload" lives in lv.c).
+ * Defaults follow the Nintendo Switch Online N64 convention. The right stick
+ * stays digital C-buttons (fixed, in addition to any C source bound here). */
+enum { NB_A, NB_B, NB_Z, NB_R, NB_L, NB_CU, NB_CD, NB_CL, NB_CR, NB_START,
+       NB_DU, NB_DD, NB_DL, NB_DR, NB_COUNT };
+static const struct { const char *key; const char *def; unsigned bit; } kN64Btn[NB_COUNT] = {
+    [NB_A]     = { "A",      "a",                        GE_CONT_A },
+    [NB_B]     = { "B",      "x",                        GE_CONT_B },
+    [NB_Z]     = { "Z",      "righttrigger",             GE_CONT_G },
+    [NB_R]     = { "R",      "lefttrigger,rightshoulder", GE_CONT_R },
+    [NB_L]     = { "L",      "leftshoulder",             GE_CONT_L },
+    [NB_CU]    = { "CUp",    "none",                     GE_CONT_E },
+    [NB_CD]    = { "CDown",  "none",                     GE_CONT_D },
+    [NB_CL]    = { "CLeft",  "none",                     GE_CONT_C },
+    [NB_CR]    = { "CRight", "none",                     GE_CONT_F },
+    [NB_START] = { "Start",  "start",                    GE_CONT_START },
+    [NB_DU]    = { "DUp",    "dpup",                     GE_CONT_UP },
+    [NB_DD]    = { "DDown",  "dpdown",                   GE_CONT_DOWN },
+    [NB_DL]    = { "DLeft",  "dpleft",                   GE_CONT_LEFT },
+    [NB_DR]    = { "DRight", "dpright",                  GE_CONT_RIGHT },
+};
+typedef struct { u32 mask[PA_COUNT]; u32 n64[NB_COUNT]; } PadBinds;
 typedef struct { PadBinds seat[MAX_PADS]; } PadTable;
 
 static int  padPreset[MAX_PADS] = {0};           /* D518: per-seat (Input.PadPreset[2/3/4]) */
@@ -475,8 +501,22 @@ static int  padPreset[MAX_PADS] = {0};           /* D518: per-seat (Input.PadPre
  * (D498: declared here so the stick-style helpers below can read it.) */
 static int naturalPitchMode = 1;                 /* Input.NaturalPitch */
 static int controlScheme = 0;                    /* D513 Input.ControlScheme: 0 Modern (port forces the style), 1 OG (the game's own selection is honoured) */
+/* D513 accessor (#145): watchsettings.c gates D564's pin of the game's own
+ * "Look up/down" value on this: in the Modern scheme the port's Invert
+ * look rows own the inversion; in Original the game's watch setting is
+ * the live control (the game honours it in bondview2's stick-pitch path),
+ * so it must stay user-settable there. */
+int inputControlScheme(void) { return controlScheme; }
+/* D645: Original scheme -- F10 "Invert look" mirrors the game's own
+ * "Look up/down" (watchsettings.c keeps the two in sync both ways). */
+static int padLookInvertY = 0;                   /* Input.PadLookInvertY: 1 = invert the look stick's Y (right stick in SOLITARE roles, left stick in HONEY roles -- #145) */
+int inputPadLookInvert(void) { return padLookInvertY; }
+void inputSetPadLookInvert(int v) { padLookInvertY = v ? 1 : 0; }
 static char padStr[MAX_PADS][PA_COUNT][32];      /* Input.Pad[N].<Action> */
 static char padStrKey[MAX_PADS][PA_COUNT][40];
+static char n64Str[MAX_PADS][NB_COUNT][32];      /* D643 Input.N64Pad[N].<Button> */
+static char n64StrKey[MAX_PADS][NB_COUNT][40];
+static int n64StickFeel = 1;                     /* D643 Input.N64StickFeel (Original only): 1 accurate, 0 the Modern deadzone/rescale */
 static PadTable padTabBuf[2];
 static void *padTabPub;                         /* published PadTable*; pointer-sized/aligned for SDL_Atomic{Get,Set}Ptr (D599) */
 static unsigned padTabSig = 0;
@@ -511,6 +551,21 @@ static void padBuildBinds(PadBinds *b, int seat, int custom, int *bad)
     for (int a = 0; a < PA_COUNT; a++)
         b->mask[a] = padParseMask(custom ? padStr[seat][a] : kPadAct[a].def, bad);
     if (!b->mask[PA_START]) b->mask[PA_START] = PSM(PS_START);   /* lockout guard: keep a pause path */
+    for (int n = 0; n < NB_COUNT; n++) b->n64[n] = 0;
+}
+
+/* D643: the Original table. Port actions are zeroed (nothing may inject
+ * use/reload/crouch/gadget/weapon-cycle); n64[] carries the buttons. Guards:
+ * Start unbound -> default; A and B both unbound -> both default (menu lockout). */
+static void padBuildN64(PadBinds *b, int seat, int *bad)
+{
+    for (int a = 0; a < PA_COUNT; a++) b->mask[a] = 0;
+    for (int n = 0; n < NB_COUNT; n++) b->n64[n] = padParseMask(n64Str[seat][n], bad);
+    if (!b->n64[NB_START]) b->n64[NB_START] = padParseMask(kN64Btn[NB_START].def, NULL);
+    if (!b->n64[NB_A] && !b->n64[NB_B]) {
+        b->n64[NB_A] = padParseMask(kN64Btn[NB_A].def, NULL);
+        b->n64[NB_B] = padParseMask(kN64Btn[NB_B].def, NULL);
+    }
 }
 
 /* D498: preset 3 is Custom; 4 = 1.4 Elektra is a fixed mapping, not custom.
@@ -559,6 +614,11 @@ static unsigned padComputeSig(void)
 {
     int anyCustom = 0;
     unsigned h = 2166136261u;
+    h = (h ^ (unsigned)controlScheme) * 16777619u;   /* D643 */
+    if (controlScheme)
+        for (int s = 0; s < MAX_PADS; s++)
+            for (int n = 0; n < NB_COUNT; n++)
+                for (const char *p = n64Str[s][n]; *p; p++) h = (h ^ (unsigned char)*p) * 16777619u;
     for (int s = 0; s < MAX_PADS; s++) {
         h ^= (unsigned)padPreset[s] * 16777619u;
         if (padPresetCustom(s)) anyCustom = 1;
@@ -584,7 +644,10 @@ static const PadTable *padTabGet(void)
         PadTable *cur = (PadTable *)SDL_AtomicGetPtr((void **)&padTabPub);
         PadTable *spare = (cur == &padTabBuf[0]) ? &padTabBuf[1] : &padTabBuf[0];
         int bad = 0;
-        for (int s = 0; s < MAX_PADS; s++) padBuildBinds(&spare->seat[s], s, padPresetCustom(s), &bad);
+        for (int s = 0; s < MAX_PADS; s++) {
+            padBuildBinds(&spare->seat[s], s, padPresetCustom(s), &bad);
+            if (controlScheme) padBuildN64(&spare->seat[s], s, &bad);   /* D643 */
+        }
         SDL_AtomicSetPtr((void **)&padTabPub, spare);
         padTabSig = sig;
         padTabValid = 1;
@@ -618,7 +681,7 @@ static u32 padActHeld(const PadBinds *b, u32 raw)
     return h;
 }
 
-typedef struct { int facePlayable; int slotNative; int southpaw; int tank; } PadMapCtx;
+typedef struct { int facePlayable; int slotNative; int southpaw; int tank; int orig; } PadMapCtx;   /* orig: D643 */
 typedef struct { unsigned button; int crouch; int gadgetEdge; u32 actHeld; } PadMapOut;
 
 /* The pure mapping: raw source bits + bindings + previous action state ->
@@ -629,6 +692,15 @@ static void padMapPure(const PadBinds *b, u32 raw, u32 actPrev, unsigned buttonI
                        const PadMapCtx *c, PadMapOut *o)
 {
     unsigned button = buttonIn;
+    if (c->orig) {   /* D643: strict virtual N64 pad; same in every context, no port actions, no southpaw swap */
+        for (int n = 0; n < NB_COUNT; n++)
+            if (raw & b->n64[n]) button |= kN64Btn[n].bit;
+        o->button = button;
+        o->crouch = 0;
+        o->gadgetEdge = 0;
+        o->actHeld = 0;
+        return;
+    }
     u32 held = padActHeld(b, raw);
     if (raw & b->mask[PA_FIRE] & PS_TRIGGERS) button |= GE_CONT_G;
     if (raw & b->mask[PA_AIM] & PS_TRIGGERS)  button |= GE_CONT_R;
@@ -1047,7 +1119,7 @@ static int padLookSensY    = 100;             /* right-stick look: vertical (pit
 static int padSouthpaw     = 0;               /* 1 = swap fire (G) / grenade (R) trigger actions */
 static int padLookSmooth   = 0;               /* right-stick look low-pass strength, 0-10 (0 = off, 10 = max) */
 static int padTriggerPct   = 25;              /* trigger press point, % of travel (~30*256) */
-static int padLookInvertY  = 0;               /* 1 = invert right-stick (look) Y */
+/* padLookInvertY: declared with the D513 accessors above (D645). */
 static int inputDefaultsRev = 0;              /* D546: 1 once the deadzone default migration ran */
 /* Per-pad look-smoothing EMA state (Wave A item 5), one X/Y pair per pad. Reset
  * on hot-unplug (inputRescanPads) so a smoothed value can't stick after the
@@ -1761,6 +1833,27 @@ void inputUpdate(void)
     mouseDY += dy;
 }
 
+/* D643 Input.N64StickFeel=1 (Original): the stick as a bare N64 pot. Linear
+ * 32767 -> STICK_MAX (80, the N64 range), a small centre deadzone and nothing
+ * else (no deadzone rescale, look sensitivity or smoothing). The ~7.6% centre
+ * (2500/32767) is an assumption of typical N64 stick dead-centre jitter --
+ * confirm against 1964 / the N64 manual. */
+#define N64_STICK_DZ 2500
+static int scaleAxis(int v, int dz);
+static int scaleAxisN64(int v)
+{
+    if (v > -N64_STICK_DZ && v < N64_STICK_DZ) return 0;
+    int out = (int)((long)v * STICK_MAX / 32767);
+    if (out >  STICK_MAX) out =  STICK_MAX;
+    if (out < -STICK_MAX) out = -STICK_MAX;
+    return out;
+}
+/* Modern/legacy deadzone scaling, or the accurate N64 response in Original. */
+static int padStickAxis(int v, int dz)
+{
+    return (controlScheme && n64StickFeel) ? scaleAxisN64(v) : scaleAxis(v, dz);
+}
+
 static int scaleAxis(int v, int dz)
 {
     if (dz < 0) dz = 0;
@@ -2481,7 +2574,7 @@ static unsigned inputComputePadSlot(int idx, signed char *stick_x, signed char *
          * combined edges prevent a held input from repeatedly using a door.
          * In menus/watch/tank the pad keeps native accept/cancel buttons. */
         SDL_GameController *slot0Pad = slotPad(0);
-        u32 slot0Held = slot0Pad ? padActHeld(&padTabGet()->seat[padSeatForSlot(0)], padReadRaw(slot0Pad)) : 0;
+        u32 slot0Held = (slot0Pad && !controlScheme) ? padActHeld(&padTabGet()->seat[padSeatForSlot(0)], padReadRaw(slot0Pad)) : 0;
         int padUse = (slot0Held & (1u << PA_USE)) != 0;       /* D469: bound actions (Jinx 1.1: A / X) */
         int padReload = (slot0Held & (1u << PA_RELOAD)) != 0;
         int useNow = scriptIsActive(0) ? scriptSlots[0].use : (actHeld(ks, IA_CANCEL) || padUse);
@@ -2880,11 +2973,11 @@ static unsigned inputComputePadSlot(int idx, signed char *stick_x, signed char *
         if (!padMenuMode && stickSwap(idx)) { int tx = lx, ty = ly; lx = rx; ly = ry; rx = tx; ry = ty; }   /* D518 */
 
         if (padMenuMode) {
-            int px = scaleAxis(lx, padDeadzoneL);
-            int py = -scaleAxis(ly, padDeadzoneL);       /* SDL up = negative -> N64 up = positive */
+            int px = padStickAxis(lx, padDeadzoneL);   /* D643: accurate N64 response in Original */
+            int py = -padStickAxis(ly, padDeadzoneL);    /* SDL up = negative -> N64 up = positive */
             if (px) sx = px;
             if (py) sy = py;
-        } else if (styleIsSolitare(idx)) {   /* D498, D513, D518 */
+        } else if (!controlScheme && styleIsSolitare(idx)) {   /* D498, D513, D518; D643: Original never swaps roles (left stick = analog stick, right = C) */
             /* D194/D238: SOLITARE swaps stick roles -- left stick becomes
              * digital-step movement (same analog-for-movement tradeoff as
              * the keyboard remap above), right stick becomes continuous
@@ -2934,13 +3027,23 @@ static unsigned inputComputePadSlot(int idx, signed char *stick_x, signed char *
                 if (rys) sy = rys;
             }
         } else {
-            int px = scaleAxis(lx, padDeadzoneL);
-            int py = -scaleAxis(ly, padDeadzoneL);       /* SDL up = negative -> N64 up = positive */
+            int px = padStickAxis(lx, padDeadzoneL);   /* D643: accurate N64 response in Original */
+            int py = -padStickAxis(ly, padDeadzoneL);    /* SDL up = negative -> N64 up = positive */
+            /* #145: in the HONEY role the LEFT stick is the look stick (1.1/
+             * 1.3 -- and, after the D498/D518 physical swap for 1.3 Frost,
+             * this channel carries the physical right stick, the 1.3 look
+             * stick). Inverting it, not the right stick (which is the digital
+             * movement C-button source in this role), is what makes F10's
+             * "Invert look" affect in-game vertical aim under the game's own
+             * original styles. The SOLITARE branch above already inverts its
+             * (look) right-stick channel. */
+            /* D645: not in Original -- there the game's own "Look up/down"
+             * (mirrored from this row by watchsettings.c) inverts natively. */
+            if (padLookInvertY && !controlScheme) py = -py;
             if (px) sx = px;
             if (py) sy = py;
             if (idx == 0 && (px || py)) s_aimDevMouse = 0;   /* D337: pad aim active */
 
-            if (padLookInvertY) ry = -ry;
             if (rx >  RSTICK_THRESHOLD) button |= GE_CONT_F;
             if (rx < -RSTICK_THRESHOLD) button |= GE_CONT_C;
             if (ry >  RSTICK_THRESHOLD) button |= GE_CONT_D;
@@ -2958,6 +3061,7 @@ static unsigned inputComputePadSlot(int idx, signed char *stick_x, signed char *
             pc.facePlayable = (idx == 0 || inputMpSplit()) && inputCanUseGameplayActions(padMenuMode);
             pc.slotNative = (idx > 0 || inputMpPadsOwnP1());
             pc.southpaw = padSouthpaw;
+            pc.orig = controlScheme;   /* D643 */
             pc.tank = !pc.facePlayable && idx == 0 && !padMenuMode && g_PlayerIsInTank == 1 &&
                       slotPlayer() && slotPlayer()->outside_watch_menu && !slotPlayer()->pause_state;
             const PadBinds *pb = &padTabGet()->seat[padSeatForSlot(idx)];
@@ -4004,6 +4108,64 @@ static void padMapSelfTest(void)
     sysLogPrintf(LOG_INFO, "GE_PADMAPTEST: binding API round trip (%d) clear keeps slot (%d) [%s]", c10, c11, padStr[1][PA_FIRE]);
     fail += !(c10 && c11);
 
+    /* D643: Original = strict virtual N64 pad. Per-source table (identical in
+     * stage / menu / tank), then remap, per-seat independence, guards. */
+    {
+        static const unsigned exp[PS_COUNT] = {
+            [PS_A] = GE_CONT_A, [PS_B] = 0, [PS_X] = GE_CONT_B, [PS_Y] = 0,
+            [PS_LB] = GE_CONT_L, [PS_RB] = GE_CONT_R, [PS_LS] = 0, [PS_RS] = 0,
+            [PS_START] = GE_CONT_START, [PS_DUP] = GE_CONT_UP, [PS_DDOWN] = GE_CONT_DOWN,
+            [PS_DLEFT] = GE_CONT_LEFT, [PS_DRIGHT] = GE_CONT_RIGHT,
+            [PS_LT] = GE_CONT_R, [PS_RT] = GE_CONT_G };
+        char savedN[MAX_PADS][NB_COUNT][32];
+        int savedCs = controlScheme, savedSouth = padSouthpaw, o1 = 1, o2 = 1, o3 = 1;
+        memcpy(savedN, n64Str, sizeof(savedN));
+        controlScheme = 1;
+        padSouthpaw = 1;   /* must be ignored in Original */
+        for (int s2 = 0; s2 < MAX_PADS; s2++) {
+            padPreset[s2] = 0;
+            for (int n = 0; n < NB_COUNT; n++) snprintf(n64Str[s2][n], sizeof(n64Str[0][0]), "%s", kN64Btn[n].def);
+        }
+        padTabValid = 0;
+        const PadTable *ot = padTabGet();
+        for (int src = 0; src < PS_COUNT; src++)
+            for (int cx = 0; cx < 3; cx++) {   /* playable / menu / tank */
+                PadMapCtx c = { cx == 0, 1, 1, cx == 2, 1 };
+                padMapPure(&ot->seat[0], PSM(src), 0, 0, &c, &o);
+                int ok = o.button == exp[src] && !o.crouch && !o.gadgetEdge &&
+                         padActHeld(&ot->seat[0], PSM(src)) == 0;
+                o1 &= ok;
+                sysLogPrintf(ok ? LOG_INFO : LOG_ERROR, "GE_PADMAPTEST: original %-13s %-8s -> %04x (want %04x)",
+                             kPadSrc[src].name, cx == 0 ? "playable" : cx == 1 ? "menu" : "tank", o.button, exp[src]);
+            }
+        { PadMapCtx c = { 1, 0, 0, 0, 1 }; padMapPure(&ot->seat[0], PSM(PS_RT) | PSM(PS_A), 0, 0, &c, &o); o1 &= o.button == (GE_CONT_G | GE_CONT_A); }
+        /* remap: seat 1 B := y,b and CUp := b; seat 0 untouched */
+        snprintf(n64Str[1][NB_B], sizeof(n64Str[0][0]), "y,b");
+        snprintf(n64Str[1][NB_CU], sizeof(n64Str[0][0]), "rightstick");
+        /* guards: seat 2 Start none; seat 3 A and B both none */
+        snprintf(n64Str[2][NB_START], sizeof(n64Str[0][0]), "none");
+        snprintf(n64Str[3][NB_A], sizeof(n64Str[0][0]), "none");
+        snprintf(n64Str[3][NB_B], sizeof(n64Str[0][0]), "none");
+        ot = padTabGet();
+        PadMapCtx oc = { 1, 0, 0, 0, 1 };
+        padMapPure(&ot->seat[1], PSM(PS_Y), 0, 0, &oc, &o);  o2 &= o.button == GE_CONT_B;
+        padMapPure(&ot->seat[1], PSM(PS_B), 0, 0, &oc, &o);  o2 &= o.button == GE_CONT_B;
+        padMapPure(&ot->seat[1], PSM(PS_X), 0, 0, &oc, &o);  o2 &= o.button == 0;
+        padMapPure(&ot->seat[1], PSM(PS_RS), 0, 0, &oc, &o); o2 &= o.button == GE_CONT_E;
+        padMapPure(&ot->seat[0], PSM(PS_X), 0, 0, &oc, &o);  o2 &= o.button == GE_CONT_B;
+        padMapPure(&ot->seat[0], PSM(PS_Y), 0, 0, &oc, &o);  o2 &= o.button == 0;
+        padMapPure(&ot->seat[2], PSM(PS_START), 0, 0, &oc, &o); o3 &= o.button == GE_CONT_START;
+        padMapPure(&ot->seat[3], PSM(PS_A), 0, 0, &oc, &o);  o3 &= o.button == GE_CONT_A;
+        padMapPure(&ot->seat[3], PSM(PS_X), 0, 0, &oc, &o);  o3 &= o.button == GE_CONT_B;
+        int st = scaleAxisN64(2499) == 0 && scaleAxisN64(32767) == STICK_MAX && scaleAxisN64(-32768) == -STICK_MAX &&
+                 scaleAxisN64(16384) == 40 && padStickAxis(5000, 7500) == scaleAxisN64(5000);
+        sysLogPrintf(LOG_INFO, "GE_PADMAPTEST: original table (%d) remap/seat-independence (%d) unbound-fallback (%d) stick (%d)", o1, o2, o3, st);
+        fail += !(o1 && o2 && o3 && st);
+        controlScheme = savedCs;
+        padSouthpaw = savedSouth;
+        memcpy(n64Str, savedN, sizeof(savedN));
+    }
+
     memcpy(padStr, saved, sizeof(saved));
     memcpy(padPreset, savedPreset, sizeof(savedPreset));
     padTabValid = 0;
@@ -4025,6 +4187,15 @@ PD_CONSTRUCTOR static void inputConfigInit(void)
             snprintf(padStr[seat][a], sizeof(padStr[seat][a]), "%s", kPadAct[a].def);
             configRegisterString(padStrKey[seat][a], padStr[seat][a], sizeof(padStr[seat][a]));
         }
+    /* D643 (#145/#149): Original-scheme N64 button map, per seat (seat 1 = Input.N64Pad.<B>, seats 2-4 = Input.N64Pad2/3/4.<B>). */
+    for (int seat = 0; seat < MAX_PADS; seat++)
+        for (int n = 0; n < NB_COUNT; n++) {
+            if (seat == 0) snprintf(n64StrKey[seat][n], sizeof(n64StrKey[seat][n]), "Input.N64Pad.%s", kN64Btn[n].key);
+            else           snprintf(n64StrKey[seat][n], sizeof(n64StrKey[seat][n]), "Input.N64Pad%d.%s", seat + 1, kN64Btn[n].key);
+            snprintf(n64Str[seat][n], sizeof(n64Str[seat][n]), "%s", kN64Btn[n].def);
+            configRegisterString(n64StrKey[seat][n], n64Str[seat][n], sizeof(n64Str[seat][n]));
+        }
+    configRegisterInt("Input.N64StickFeel", &n64StickFeel, 0, 1);
     configRegisterInt("Input.MPMode", &mpInputMode, 0, 2);   /* D416: 0 Auto, 1 PadsOnly, 2 KbmP1 */
     configRegisterInt("Input.MouseEnabled", &mouseEnabled, 0, 1);
     configRegisterInt("Input.CrosshairCursor", &cfgCrosshairCursor, 0, 1);   /* D555 */
diff --git a/port/src/main.c b/port/src/main.c
index 673b9d4a..79780411 100644
--- a/port/src/main.c
+++ b/port/src/main.c
@@ -88,6 +88,10 @@ static void portAtExit(void)
     inputRumbleStopAll();   /* D493: no pad left rumbling after exit */
     videoSaveWindowState();
     configSave();
+    /* D632: release the GPU context while the process is alive (needs the
+     * window state saved first; before any relaunch so the GPU init of the
+     * new instance never overlaps this teardown). */
+    videoTeardownForExit();
     /* D443: Restart game -- runs only on the orderly D344 exit path, after the
      * ini is written, so the new instance reads the saved settings. */
     if (videoRestartRequested()) {
diff --git a/port/src/optionsoverlay.c b/port/src/optionsoverlay.c
index 298919b1..75284d36 100644
--- a/port/src/optionsoverlay.c
+++ b/port/src/optionsoverlay.c
@@ -53,6 +53,7 @@ extern MENU current_menu;
 #include "video.h"
 #include "input.h"
 #include "front.h"   /* selected_folder_num (D356 stage probe) */
+#include "lv.h"   /* lvlGetCurrentMultiPlayerSec (D606 mission timer) */
 #include "bondtypes.h"   /* sImageTableEntry (D555 crosshair pointer: typed field reads) */
 #include "file.h"   /* save_data (D356 reset probe: second-file isolation) */
 #include "optionsoverlay.h"
@@ -144,10 +145,11 @@ static const char *const kGameplayView[] = { "Original", "Extended", NULL };   /
  * have it, but it's dropped from the menu: with VSync on (the default) it's
  * indistinguishable from 60, and with VSync off it just burns GPU time
  * re-presenting the same simulated frame -- confusing for no real benefit. */
-/* D578: 90/120/144 draw interpolated frames between the 60 Hz game frames
+/* D578: 90/120/144/165/240 draw interpolated frames between the 60 Hz game frames
  * (port/fast3d/gfx_pc.cpp); the sim rate is unchanged. */
-/* -1 = Auto: match the display refresh (60..144; the default). */
-static const int         kFpsCapSeq[] = { -1, 30, 60, 90, 120, 144 };
+/* -1 = Auto: match the display refresh (up to 240; the default). 240 is the
+ * renderer's ceiling: at most IP_SLOTS/2 = 4 presents per 60 Hz tick (libultra.c). */
+static const int         kFpsCapSeq[] = { -1, 30, 60, 90, 120, 144, 165, 240 };
 #define FPSCAP_N ((int)(sizeof(kFpsCapSeq) / sizeof(kFpsCapSeq[0])))
 
 /* Windowed-mode resolution presets. Filtered at init to those that fit the
@@ -383,10 +385,13 @@ static struct Row rows[] = {
     { .key="__HdrGame", .label="GAME", .kind=ROW_HEADER },
     /* Aim control (hold vs toggle the aim button) is a per-file watch row
      * like Auto-aim; the port honours the game's toggle. The game's own
-     * "Look up/down" (Bond.Look) row was removed (D564): the port's mouse and
-     * centred-pad look write the view directly and never consult it, so the
-     * Mouse / Controller "Invert look" rows are the only inversion controls;
-     * watchsettings.c pins the game's value to its factory default. */
+     * "Look up/down" (Bond.Look) row was removed (D564): in the Modern
+     * scheme the port's mouse and centred-pad look write the view directly
+     * and never consult it, so the Mouse / Controller "Invert look" rows
+     * are the only inversion controls there (watchsettings.c pins the
+     * game's value to its factory default). #145: in the Original scheme
+     * the pin is lifted and the in-game watch row is live again -- the
+     * rows above still apply on top (pad look stick, #145 fix). */
     { .key="Bond.AimControl", .label="Aim control", .kind=ROW_TOGGLE,
       .names=kHold, .found=1, .uiMax=1, .cfgMax=1, .saveScoped=1 },
     { .key="Bond.AutoAim", .label="Auto-aim", .kind=ROW_TOGGLE,
@@ -437,6 +442,9 @@ static struct Row rows[] = {
      * patched, read or written. Applies live (rowSetCommit). */
     { .key="Game.AllUnlocked", .label="All unlocked",
       .kind=ROW_TOGGLE, .step=1, .names=kOnOff },
+    /* D606 (#144): PD-style always-visible mission timer; drawn top-right
+     * (under the FPS readout when both are on) while a stage is active. */
+    { .key="Game.MissionTimer", .label="Mission timer", .kind=ROW_TOGGLE, .step=1, .names=kOnOff },
     SEP(G4),
     /* D551: opt-in update check; the action row appears only once a newer release is known. */
     { .key="Game.CheckUpdates", .label="Check for updates", .kind=ROW_TOGGLE, .step=1, .names=kOnOff },
@@ -791,9 +799,20 @@ int optionsBindingCaptureTick(void)
 static int      s_showFps = 0;
 static char     s_fpsText[16] = "";
 
+/* D606 (#144, v0.6.0): optional PD-style always-visible mission timer.
+ * Reads the game's own mission clock (lvlGetCurrentMultiPlayerSec: the
+ * D_80048394 accumulator scaled to region ticks -- the same source the
+ * watch TIME page and the AI_IFMissionTime conditions use), so it is the
+ * faithful N64 mission time, not a wall clock. Port-only: a text row like
+ * the D213 FPS readout; off by default, nothing appended when off, so
+ * goldens are untouched. */
+static int      s_missionTimer = 0;
+static char     s_missionTimeText[24] = "";
+
 PD_CONSTRUCTOR static void overlayConfigInit(void)
 {
     configRegisterInt("Video.DisplayFPS", &s_showFps, 0, 1);
+    configRegisterInt("Game.MissionTimer", &s_missionTimer, 0, 1);
 }
 
 /* Sampled once per emitted frame; recomputes the string every ~0.5 s. */
@@ -820,6 +839,27 @@ static void fpsTick(void)
     }
 }
 
+/* D606: refreshed once per emitted frame, like the FPS window above. The
+ * clock is stage-paced (lvl.c advances it during a run), so menu frames
+ * simply carry no text and the row disappears outside a mission. */
+static void missionTimerTick(void)
+{
+    if (!s_missionTimer ||
+        (current_menu != MENU_RUN_STAGE && current_menu != MENU_INVALID)) {
+        s_missionTimeText[0] = 0;
+        return;
+    }
+    int s = (int)lvlGetCurrentMultiPlayerSec();
+    if (s < 0) s = 0;
+    if (s >= 3600) {
+        snprintf(s_missionTimeText, sizeof(s_missionTimeText), "TIME %d:%02d:%02d",
+                 s / 3600, (s / 60) % 60, s % 60);
+    } else {
+        snprintf(s_missionTimeText, sizeof(s_missionTimeText), "TIME %d:%02d",
+                 s / 60, s % 60);
+    }
+}
+
 /* Layout (game 2D pixel space = viGetX() x viGetY(), ~320x240). Shared by the
  * emit path and the mouse hit-testing in optionsOverlayHandleInput().
  * PD-style dialog (menu PD parity R4): sized to its content and centred, one
@@ -1945,14 +1985,17 @@ static int ddGeom(s32 *x0, s32 *y0, s32 *x1, s32 *y1, int *vis)
     s32 h = v * OV_LINE + 2;
     *x1 = o.valueR + 3;
     *x0 = *x1 - w;
-    s32 H = ovH();
+    /* D646: the popup stays inside the settings box -- it used to open
+     * down over the hint/tooltip panel below the box (its limit was the
+     * screen edge, not o.bottom). */
+    s32 lo = o.titleB + 1, hi = o.bottom - 1;
     *y0 = rowY + OV_LINE;
-    if (*y0 + h > H - 4) *y0 = rowY - h;   /* no room below: open upward */
+    if (*y0 + h > hi) *y0 = rowY - h;   /* no room below: open upward */
     /* D568: no room above either (a long list on a mid-page row): keep the whole
-     * popup on screen, flush with the bottom margin. A negative top used to
+     * popup inside the box, flush with its bottom. A negative top used to
      * draw the labels above the screen while the fill wrapped down past the
      * bottom edge (an empty, misplaced box). */
-    if (*y0 < 4) *y0 = (H - 4 - h > 4) ? H - 4 - h : 4;
+    if (*y0 < lo) *y0 = (hi - h > lo) ? hi - h : lo;
     *y1 = *y0 + h;
     *vis = v;
     return 1;
@@ -2081,6 +2124,7 @@ static const struct { const char *key; double def; } kResetDefaults[] = {
     /* GAMEPLAY ini rows (port/src/video.c initializers) */
     { "Game.SkipIntro",   0 },   /* = 0 */
     { "Game.CheckUpdates", 0 },  /* = 0 (off, D551) */
+    { "Game.MissionTimer", 0 },  /* = 0 (off, D606) */
     { "Game.NoHitFlash",  0 },   /* = 0 */
     { "Game.AllUnlocked", 0 },   /* = 0 -- D257's "default ON" note was stale */
     { "Game.HudScale",   100 },  /* = 100 (D226) */
@@ -3403,6 +3447,7 @@ static Gfx *overlayEmitLocked(void)
     }
 
     fpsTick();
+    missionTimerTick();
 
     {   /* D558 test hook: GE_OPTIONSOVERLAY_ATFRAME=<n> re-applies the open request at emit n (read once). */
         static int s_atFrame = -2, s_emitN = 0;
@@ -3421,10 +3466,12 @@ static Gfx *overlayEmitLocked(void)
     }
 
     if (!s_open) {
-        if (!s_showFps || !s_fpsText[0]) {
+        if ((!s_showFps || !s_fpsText[0]) && !s_missionTimeText[0]) {
             return NULL;   /* nothing appended -> golden dumps byte-identical */
         }
-        /* D213: FPS-only mini DL (top-right), panel closed. */
+        /* D213: FPS-only mini DL (top-right), panel closed. D606: the
+         * mission timer shares the top-right corner, below the FPS readout
+         * when both are visible. */
         const s32 fw = ovW();
         const s32 fh = ovH();
         gfx_set_overlay_scale(ovScale());   /* D510 */
@@ -3434,7 +3481,14 @@ static Gfx *overlayEmitLocked(void)
         gDPSetTexturePersp(fgdl++, G_TP_NONE);
         gDPSetScissor(fgdl++, G_SC_NON_INTERLACE, 0, 0, viGetX(), viGetY());
         fgdl = microcode_constructor(fgdl);
-        fgdl = drawTextR(fgdl, fw - 6, 6, s_fpsText, 0x40ff60ff);
+        int ty = 6;
+        if (s_showFps && s_fpsText[0]) {
+            fgdl = drawTextR(fgdl, fw - 6, 6, s_fpsText, 0x40ff60ff);
+            ty = 22;
+        }
+        if (s_missionTimeText[0]) {
+            fgdl = drawTextR(fgdl, fw - 6, ty, s_missionTimeText, 0x40ff60ff);
+        }
         gDPPipeSync(fgdl++);
         gSPEndDisplayList(fgdl++);
         return s_buf;
@@ -3817,6 +3871,7 @@ static const struct { const char *key, *help; } kRowHelp[] = {
     { "Game.SkipIntro", "Skips the intro and opens the file select screen." },
     { "Game.AllUnlocked", "Unlocks all levels, 007 mode and every cheat." },
     { "Game.CheckUpdates", "Notifies you when a new version is available." },
+    { "Game.MissionTimer", "Shows the mission elapsed time on screen (top-right, under the FPS readout)." },
 };
 
 
diff --git a/port/src/video.c b/port/src/video.c
index ba878f47..caccbd0c 100644
--- a/port/src/video.c
+++ b/port/src/video.c
@@ -60,7 +60,7 @@ static int initDone = 0;
  * at modern resolutions without requiring 4x anti-aliasing.
  */
 static int cfgVSync         = 1;   /* swap interval: 0 = off, 1 = on            */
-static int cfgFpsCap        = -1;  /* frame cap in fps; -1 = Auto (match display refresh, 60..144), 0 = uncapped (vsync); menu: Auto/30/60, and 90/120/144 = D578 interpolation */
+static int cfgFpsCap        = -1;  /* frame cap in fps; -1 = Auto (match display refresh, up to 4x the VI rate), 0 = uncapped (vsync); menu: Auto/30/60, and 90/120/144/165/240 = D578 interpolation */
 static int cfgMSAA          = 2;   /* 1/2/4/8/16 samples; 2x default is lighter on low-end GPUs */
 static int cfgTexFilter     = 1;   /* 0 = nearest, 1 = bilinear (default), 2 = N64 3-point + trilinear, 3 = trilinear (Trilinear option, playtest 2026-10-03) */
 static int cfgFixMipTex     = 1;   /* RC2: clip mip-contaminated texture uploads to base height */
@@ -1033,13 +1033,15 @@ static int s_dumpSlot = -1;   /* D578: GE_PCDUMP_PRESENTS names each present */
 
 /* Auto (Video.FpsCap = -1): match the display. Refresh unknown, or no faster
  * than the sim (60 Hz panel, 59.94 reported as 59/60) -> the VI rate, i.e. the
- * plain non-interpolated path; otherwise the refresh clamped to 144. */
+ * plain non-interpolated path; otherwise the refresh itself, up to the
+ * renderer's ceiling of 4 presents per sim tick (IP_SLOTS / 2 in libultra.c:
+ * the ring holds this frame + one still waiting), i.e. 240 NTSC / 200 PAL. */
 int videoAutoFpsCap(void)
 {
     const int vi = (osTvType == 0) ? 50 : 60;
     const int refresh = gfx_sdl_get_refresh_rate();
     if (refresh <= vi + 5) return vi;
-    return refresh > 144 ? 144 : refresh;
+    return refresh > 4 * vi ? 4 * vi : refresh;
 }
 
 static void videoApplyFpsCap(void)
@@ -1277,6 +1279,24 @@ static void videoRenderPark(void)
     }
 }
 
+/* D632: host thread, from the atexit handler (after the window state and
+ * config are saved). Tears the GL context + window down only when the render
+ * thread is parked: it then holds no context and never touches GL again. On a
+ * park timeout the render thread may still be inside the driver, so deleting
+ * its context would be worse than leaving it to the kernel. */
+void videoTeardownForExit(void)
+{
+    if (!initDone) {
+        return;
+    }
+    if (!SDL_AtomicGet(&s_parked)) {
+        sysLogPrintf(LOG_WARNING, "video: render not parked; GL teardown skipped");
+        return;
+    }
+    gfx_sdl_teardown_for_exit();
+    sysLogPrintf(LOG_INFO, "video: GL context and window destroyed");
+}
+
 /* Host thread: exit once the render thread is out of the GL driver. */
 static void videoHostExitIfRequested(void)
 {
@@ -1284,8 +1304,15 @@ static void videoHostExitIfRequested(void)
         return;
     }
     Uint32 start = SDL_GetTicks();
-    while (!SDL_AtomicGet(&s_parked) && SDL_AtomicGet(&s_inFrame) &&
-           SDL_GetTicks() - start < QUIT_WAIT_MS) {
+    /* D632: wait for the park itself, not just for "outside a frame". A
+     * quit landing between frames (window close, 2026-10-10: "render idle
+     * after 0 ms") used to exit at once, before the render thread had
+     * unbound its context, so the D632 teardown was skipped and the GL
+     * context fell to the kernel's process-exit cleanup -- the bugcheck
+     * path. Once s_quitReq is set the next frame start parks instead of
+     * drawing (see the ordering note above), so this takes about a frame;
+     * QUIT_WAIT_MS stays the bound if the render thread is stuck. */
+    while (!SDL_AtomicGet(&s_parked) && SDL_GetTicks() - start < QUIT_WAIT_MS) {
         SDL_Delay(2);
     }
     sysLogPrintf(LOG_INFO, "video: exiting (render %s after %u ms)",
diff --git a/port/src/watchsettings.c b/port/src/watchsettings.c
index e879434c..7ba5469d 100644
--- a/port/src/watchsettings.c
+++ b/port/src/watchsettings.c
@@ -27,6 +27,9 @@ extern u16 musicTrack1GetVolume(void);
 extern u16 get_mTrack2Vol(void);
 extern void fileBuildWriteNewSave(u32 folder);
 extern void set_cur_player_look_vertical_inverted(u32 value);
+extern int inputControlScheme(void);   /* D513: 0 Modern, 1 Original */
+extern int inputPadLookInvert(void);   /* D645 */
+extern void inputSetPadLookInvert(int v);
 extern void cur_player_set_aim_control(u32 value);
 extern void cur_player_set_lookahead(u32 value);
 extern void cur_player_set_sight_onscreen_control(u32 value);
@@ -623,13 +626,35 @@ void watchSettingsGameTick(void)
      * during a level/menu transition. Each command's folder is re-checked
      * against selected_folder_num at apply time, so a stale cross-file edit
      * is still dropped, never misapplied. */
-    /* D564: the game's own "Look up/down" option is not exposed (the port's
-     * Mouse / Controller "Invert look" rows are the inversion controls). Pin it
-     * to its factory value (0, as a blank profile has it) so a value saved by an
-     * older build or set via the in-game watch menu cannot silently flip the
-     * pad stick on the game's pitch path. RAM only; one int compare per tick. */
-    if (stageActive() && g_CurrentPlayer && get_cur_player_look_vertical_inverted() != 0)
+    /* D564: the game's own "Look up/down" option is not exposed in the
+     * Modern scheme (the port's Mouse / Controller "Invert look" rows are
+     * the inversion controls there -- the port writes the view directly on
+     * those paths). Pin it to its factory value (0, as a blank profile has
+     * it) so a value saved by an older build or set via the in-game watch
+     * menu cannot silently flip the pad stick on the game's pitch path.
+     * #145 (v0.6.0): in the Original scheme the pin is LIFTED -- the game
+     * honours the setting in bondview2's stick-pitch path
+     * (moveData.invertPitch), so the in-game watch "Look up/down" row is
+     * the player's live control again, exactly as on the N64. RAM only;
+     * one int compare per tick. */
+    if (stageActive() && g_CurrentPlayer && inputControlScheme() == 0 &&
+        get_cur_player_look_vertical_inverted() != 0)
         set_cur_player_look_vertical_inverted(0);
+    /* D645 (#145): in Original the F10 "Invert look" row and the watch
+     * "Look up/down" row are one setting. Whichever changed since the last
+     * tick wins; the game then applies it natively (C-buttons and the
+     * aim-mode stick), as on the N64. */
+    static int lastRow = -1, lastGame = -1;
+    if (!stageActive()) lastRow = -1;   /* re-sync on each stage entry */
+    if (stageActive() && g_CurrentPlayer && inputControlScheme() == 1) {
+        int row = inputPadLookInvert(), game = get_cur_player_look_vertical_inverted() ? 1 : 0;
+        if (lastRow < 0) {   /* stage entry: the F10 row (edited in the front end) wins */
+            if (row != game) set_cur_player_look_vertical_inverted(row);
+        } else if (row != lastRow && row != game) set_cur_player_look_vertical_inverted(row);
+        else if (game != lastGame && game != row) inputSetPadLookInvert(game);
+        lastRow = inputPadLookInvert();
+        lastGame = get_cur_player_look_vertical_inverted() ? 1 : 0;
+    }
 
     if (!stageActive() || !g_CurrentPlayer) {
         wsStageWasActive = 0;
diff --git a/scripts/package-macos.sh b/scripts/package-macos.sh
index f5feb479..ef8a2af9 100755
--- a/scripts/package-macos.sh
+++ b/scripts/package-macos.sh
@@ -172,6 +172,18 @@ cp LICENSE "$OUT/licenses/LICENSE-port-MIT.txt"
 for d in /opt/homebrew/opt/sdl2/LICENSE.txt /opt/homebrew/opt/sdl2-compat/LICENSE.txt /opt/homebrew/opt/sdl2/share/licenses/*/LICENSE.txt; do
   if [ -f "$d" ]; then cp "$d" "$OUT/licenses/SDL2.txt"; break; fi
 done
+# sdl2-compat loads the bundled libSDL3 (zlib); carry its text too.
+for d in /opt/homebrew/opt/sdl3/LICENSE.txt /opt/homebrew/opt/sdl3/share/licenses/*/LICENSE.txt; do
+  if [ -f "$d" ]; then cp "$d" "$OUT/licenses/SDL3.txt"; break; fi
+done
+# The bundled libgcc_s / libstdc++ are GPL-3.0 with the GCC Runtime Library
+# Exception. The Homebrew gcc keg carries COPYING.RUNTIME but only GPL-2.0 as
+# COPYING, so the GPL-3.0 text ships from the tree.
+cp tools_pc/dist/GPL-3.0.txt "$OUT/licenses/gcc-COPYING3.txt"
+if [ -f /opt/homebrew/opt/gcc/COPYING.RUNTIME ]; then
+  cp /opt/homebrew/opt/gcc/COPYING.RUNTIME "$OUT/licenses/gcc-COPYING.RUNTIME.txt"
+fi
+ls "$OUT/licenses"
 
 # --- save converter + asset-prep tool (same assembly as bundle-linux.sh) ---
 mkdir -p "$OUT/tools"
diff --git a/tools_pc/demobatch.sh b/tools_pc/demobatch.sh
new file mode 100644
index 00000000..f848e293
--- /dev/null
+++ b/tools_pc/demobatch.sh
@@ -0,0 +1,185 @@
+#!/usr/bin/env bash
+# demobatch.sh -- batch-replay maintainer-recorded pad demos through
+# tools_pc/levelbench.sh --replay, collect frames / flicker flags / pacing
+# stats, and write one report per batch. Sequential only (parallel runs
+# invalidate frame pacing).
+#
+#   demobatch.sh [options] DEMODIR [ENV=VAL ...]
+#     --dir GAMEDIR   game dir holding data/ (default: cwd)
+#     --exe PATH      game binary (default: levelbench's autodetect)
+#     --out ROOT      results root (default: <GAMEDIR>/demobatch/<YYYYMMDD-HHMMSS>)
+#     --spot f1,f2    dump just these game frames (GE_PCDUMP lo-hi:step with
+#                     step = gcd of the spacing; extra frames are deleted)
+#     --flicker lo-hi dump every frame in the window, run flicker_scan.py
+#                     (--spot and --flicker are mutually exclusive: one dump
+#                     spec per run)
+#     --only a,b      only these demo names (file name without .demo)
+#     --dry-run       print the commands, launch nothing
+#   demobatch.sh --compare ROOT_A ROOT_B
+#                     side-by-side report rows with deltas for numeric columns
+#
+# Demo harness (maintainer workflow). Record once, per level:
+#   tools_pc/levelbench.sh --dir <game dir> --secs N --record demos/<name>.demo <level>
+# You play the level with the pad; input is saved to the .demo and a
+# <name>.demo.meta (level, secs, seed, binary md5, date) is written beside it.
+# Demos live in <game dir>/demos/ -- personal input recordings, gitignored,
+# never committed. Then `demobatch.sh --dir <game dir> <game dir>/demos`
+# replays each with the recorded level/secs/seed (GE_RSEED pinned) and writes
+# <ROOT>/report.txt. Replays drift slightly (D117): compare runs within one
+# session (--compare), never as pixel gates. Extra ENV=VAL args go to every run.
+set -u
+HERE=$(cd "$(dirname "$0")" && pwd)
+absdir() { (cd "$1" 2>/dev/null && pwd); }
+
+compare() {
+    local A=$1/report.txt B=$2/report.txt
+    if [ ! -f "$A" ] || [ ! -f "$B" ]; then echo "missing report.txt in $1 or $2" >&2; exit 2; fi
+    awk '
+    function d(a, b) {
+        if (a ~ /^-?[0-9.]+$/ && b ~ /^-?[0-9.]+$/) return sprintf("%+.2f", b - a)
+        return "-"
+    }
+    FNR == NR { if ($1 != "name" && NF >= 9) { a[$1] = $0; order[++n] = $1 } next }
+    $1 != "name" && NF >= 9 { b[$1] = $0 }
+    END {
+        f = "%-16s %-5s | %-5s %-5s | %-4s %-4s %-6s | %-8s %-8s %-7s | %-6s %-6s %-6s | %-5s %-5s %-6s\n"
+        printf f, "name", "level", "exitA", "exitB", "wdA", "wdB", "dWd", "pres/sA", "pres/sB", "dPres", "unevA", "unevB", "dUnev", "slowA", "slowB", "dSlow"
+        for (i = 1; i <= n; i++) {
+            k = order[i]
+            split(a[k], x, " ")
+            if (!(k in b)) { printf "%-16s (only in A)\n", k; continue }
+            split(b[k], y, " ")
+            printf f, k, x[2], x[3], y[3], x[5], y[5], d(x[5], y[5]), x[6], y[6], d(x[6], y[6]), x[7], y[7], d(x[7], y[7]), x[8], y[8], d(x[8], y[8])
+            if (x[9] != y[9]) printf "%-16s   flicker flags: %s -> %s (%s)\n", "", x[9], y[9], d(x[9], y[9])
+        }
+        for (k in b) if (!(k in a)) printf "%-16s (only in B)\n", k
+    }' "$A" "$B"
+}
+
+if [ "${1:-}" = --compare ]; then
+    [ $# -eq 3 ] || { echo "usage: $0 --compare ROOT_A ROOT_B" >&2; exit 2; }
+    compare "$2" "$3"; exit 0
+fi
+
+GAME=$PWD; EXE=; OUT=; SPOT=; FLK=; ONLY=; DRY=0
+while [ $# -gt 0 ]; do
+    case $1 in
+        --dir) GAME=$2; shift 2;;
+        --exe) EXE=$2; shift 2;;
+        --out) OUT=$2; shift 2;;
+        --spot) SPOT=$2; shift 2;;
+        --flicker) FLK=$2; shift 2;;
+        --only) ONLY=$2; shift 2;;
+        --dry-run) DRY=1; shift;;
+        -h|--help) sed -n 2,26p "$0"; exit 0;;
+        *) break;;
+    esac
+done
+[ $# -ge 1 ] || { echo "usage: $0 [options] DEMODIR [ENV=VAL ...]" >&2; exit 2; }
+DEMODIR=$1; shift
+if [ -n "$SPOT" ] && [ -n "$FLK" ]; then echo "--spot and --flicker are mutually exclusive" >&2; exit 2; fi
+[ -d "$DEMODIR" ] || { echo "no demo dir: $DEMODIR" >&2; exit 2; }
+GAME=$(absdir "$GAME") || { echo "bad --dir" >&2; exit 2; }
+DEMODIR=$(absdir "$DEMODIR")
+[ -n "$OUT" ] || OUT="$GAME/demobatch/$(date +%Y%m%d-%H%M%S)"
+case $OUT in /*|[A-Za-z]:*) ;; *) OUT=$PWD/$OUT;; esac
+
+# Dump spec: spot frames -> lo-hi:step (step = gcd of gaps); flicker -> lo-hi.
+DUMP=; SPOTLIST=
+if [ -n "$SPOT" ]; then
+    SPOTLIST=$(echo "$SPOT" | tr ',' '\n' | sort -n -u | tr '\n' ' ')
+    read -r -a FR <<< "$SPOTLIST"
+    lo=${FR[0]}; hi=${FR[$((${#FR[@]} - 1))]}; g=0
+    for ((i = 1; i < ${#FR[@]}; i++)); do
+        a=$((FR[i] - FR[i-1])); b=$g
+        while [ "$b" -ne 0 ]; do t=$((a % b)); a=$b; b=$t; done
+        g=$a
+    done
+    DUMP="$lo-$hi"
+    [ "$g" -gt 1 ] && DUMP="$lo-$hi:$g"
+elif [ -n "$FLK" ]; then
+    DUMP=$FLK
+fi
+
+LB=("$HERE/levelbench.sh" --dir "$GAME")
+[ -n "$EXE" ] && LB+=(--exe "$EXE")
+PY=
+for c in python3 python py; do
+    if command -v "$c" > /dev/null 2>&1 && "$c" -c 'import sys' > /dev/null 2>&1; then PY=$c; break; fi
+done
+[ -n "$PY" ] || PY=python3
+[ "$DRY" = 1 ] || mkdir -p "$OUT"
+
+metaget() { sed -n "s/^$2=//p" "$1" | head -1 | tr -d '\r'; }
+
+NAMES=()
+for demo in "$DEMODIR"/*.demo; do
+    [ -f "$demo" ] || continue
+    n=$(basename "$demo" .demo)
+    if [ -n "$ONLY" ]; then case ",$ONLY," in *",$n,"*) ;; *) continue;; esac; fi
+    if [ ! -f "$demo.meta" ]; then echo "skip $n: no $n.demo.meta" >&2; continue; fi
+    NAMES+=("$n")
+    lvl=$(metaget "$demo.meta" level); secs=$(metaget "$demo.meta" secs); seed=$(metaget "$demo.meta" seed)
+    CMD=("${LB[@]}" --secs "$secs" --replay "$demo" --out "$OUT/$n")
+    [ -n "$DUMP" ] && CMD+=(--dump "$DUMP")
+    CMD+=("$lvl" "GE_RSEED=${seed:-0123456789abcdef}" "$@")
+    if [ "$DRY" = 1 ]; then
+        echo "[$n] ${CMD[*]}"
+        [ -n "$SPOTLIST" ] && echo "[$n]   then keep only frames: $SPOTLIST"
+        [ -n "$FLK" ] && echo "[$n]   then: $PY -I $HERE/flicker_scan.py $OUT/$n/ppm --csv $OUT/$n/flicker.csv"
+        continue
+    fi
+    echo "=== $n (level $lvl, ${secs}s)"
+    "${CMD[@]}" > "$OUT/$n.console" 2>&1
+    tail -3 "$OUT/$n.console"
+    if [ -n "$SPOTLIST" ] && [ -d "$OUT/$n/ppm" ]; then
+        for f in "$OUT/$n"/ppm/frame_*.ppm; do
+            [ -e "$f" ] || continue
+            num=$((10#$(basename "$f" .ppm | sed 's/frame_//')))
+            case " $SPOTLIST " in *" $num "*) ;; *) rm -f "$f";; esac
+        done
+    fi
+    if [ -n "$FLK" ] && [ -d "$OUT/$n/ppm" ]; then
+        "$PY" -I "$HERE/flicker_scan.py" "$OUT/$n/ppm" --csv "$OUT/$n/flicker.csv" > "$OUT/$n/flicker.txt" 2>&1
+        tail -1 "$OUT/$n/flicker.txt"
+    fi
+done
+if [ "$DRY" = 1 ]; then echo "(dry run: no report written; root would be $OUT)"; exit 0; fi
+
+# ---- report ----------------------------------------------------------------
+# presents/s: mean over "D578 interp: X presents/s, ..." lines.
+# uneven%: from "D578 cadence: gaps <7:a 7-9.5:b 9.5-12:c 12-20:d >20:e | ...",
+# bucket counts summed over all lines; uneven = share outside the modal bucket.
+R="$OUT/report.txt"
+fmt="%-16s %-5s %-5s %-5s %-9s %-10s %-8s %-5s %-7s\n"
+{
+printf "$fmt" name level exit crash watchdog presents/s uneven% slow flicker
+for n in "${NAMES[@]}"; do
+    d="$OUT/$n"; L="$d/ge007.log"; [ -f "$L" ] || L="$d/stdout.txt"
+    lvl=$(metaget "$DEMODIR/$n.demo.meta" level)
+    ex=n; crash=n; wd=0; slow=0; stats="- -"
+    if [ -f "$L" ]; then
+        grep -q -E 'quit requested|video: exiting' "$L" && ex=y
+        wd=$(grep -c 'no frame rendered' "$L")
+        slow=$(grep -c 'D583 SLOW' "$L")
+        stats=$(awk '
+            /D578 interp:/ { for (i = 2; i <= NF; i++) if ($i ~ /^presents\/s/) { p += $(i-1); pn++ } }
+            /D578 cadence:/ {
+                for (i = 1; i <= NF; i++) if ($i == "gaps") {
+                    for (j = 1; j <= 5; j++) { s = $(i + j); sub(/^.*:/, "", s); g[j] += s + 0 }
+                    break
+                }
+            }
+            END {
+                tot = 0; mx = 0
+                for (j = 1; j <= 5; j++) { tot += g[j]; if (g[j] > mx) mx = g[j] }
+                printf "%s %s", (pn ? sprintf("%.1f", p / pn) : "-"), (tot ? sprintf("%.1f", 100 * (tot - mx) / tot) : "-")
+            }' "$L")
+    fi
+    [ -f "$d/ge007.crash.log" ] && crash=y
+    fl=-; [ -f "$d/flicker.txt" ] && fl=$(grep -c '^FLICKER' "$d/flicker.txt")
+    printf "$fmt" "$n" "${lvl:--}" "$ex" "$crash" "$wd" "${stats% *}" "${stats#* }" "$slow" "$fl"
+done
+} > "$R"
+cat "$R"
+echo "report: $R"
diff --git a/tools_pc/dist/GPL-3.0.txt b/tools_pc/dist/GPL-3.0.txt
new file mode 100644
index 00000000..94a9ed02
--- /dev/null
+++ b/tools_pc/dist/GPL-3.0.txt
@@ -0,0 +1,674 @@
+                    GNU GENERAL PUBLIC LICENSE
+                       Version 3, 29 June 2007
+
+ Copyright (C) 2007 Free Software Foundation, Inc. <http://fsf.org/>
+ Everyone is permitted to copy and distribute verbatim copies
+ of this license document, but changing it is not allowed.
+
+                            Preamble
+
+  The GNU General Public License is a free, copyleft license for
+software and other kinds of works.
+
+  The licenses for most software and other practical works are designed
+to take away your freedom to share and change the works.  By contrast,
+the GNU General Public License is intended to guarantee your freedom to
+share and change all versions of a program--to make sure it remains free
+software for all its users.  We, the Free Software Foundation, use the
+GNU General Public License for most of our software; it applies also to
+any other work released this way by its authors.  You can apply it to
+your programs, too.
+
+  When we speak of free software, we are referring to freedom, not
+price.  Our General Public Licenses are designed to make sure that you
+have the freedom to distribute copies of free software (and charge for
+them if you wish), that you receive source code or can get it if you
+want it, that you can change the software or use pieces of it in new
+free programs, and that you know you can do these things.
+
+  To protect your rights, we need to prevent others from denying you
+these rights or asking you to surrender the rights.  Therefore, you have
+certain responsibilities if you distribute copies of the software, or if
+you modify it: responsibilities to respect the freedom of others.
+
+  For example, if you distribute copies of such a program, whether
+gratis or for a fee, you must pass on to the recipients the same
+freedoms that you received.  You must make sure that they, too, receive
+or can get the source code.  And you must show them these terms so they
+know their rights.
+
+  Developers that use the GNU GPL protect your rights with two steps:
+(1) assert copyright on the software, and (2) offer you this License
+giving you legal permission to copy, distribute and/or modify it.
+
+  For the developers' and authors' protection, the GPL clearly explains
+that there is no warranty for this free software.  For both users' and
+authors' sake, the GPL requires that modified versions be marked as
+changed, so that their problems will not be attributed erroneously to
+authors of previous versions.
+
+  Some devices are designed to deny users access to install or run
+modified versions of the software inside them, although the manufacturer
+can do so.  This is fundamentally incompatible with the aim of
+protecting users' freedom to change the software.  The systematic
+pattern of such abuse occurs in the area of products for individuals to
+use, which is precisely where it is most unacceptable.  Therefore, we
+have designed this version of the GPL to prohibit the practice for those
+products.  If such problems arise substantially in other domains, we
+stand ready to extend this provision to those domains in future versions
+of the GPL, as needed to protect the freedom of users.
+
+  Finally, every program is threatened constantly by software patents.
+States should not allow patents to restrict development and use of
+software on general-purpose computers, but in those that do, we wish to
+avoid the special danger that patents applied to a free program could
+make it effectively proprietary.  To prevent this, the GPL assures that
+patents cannot be used to render the program non-free.
+
+  The precise terms and conditions for copying, distribution and
+modification follow.
+
+                       TERMS AND CONDITIONS
+
+  0. Definitions.
+
+  "This License" refers to version 3 of the GNU General Public License.
+
+  "Copyright" also means copyright-like laws that apply to other kinds of
+works, such as semiconductor masks.
+
+  "The Program" refers to any copyrightable work licensed under this
+License.  Each licensee is addressed as "you".  "Licensees" and
+"recipients" may be individuals or organizations.
+
+  To "modify" a work means to copy from or adapt all or part of the work
+in a fashion requiring copyright permission, other than the making of an
+exact copy.  The resulting work is called a "modified version" of the
+earlier work or a work "based on" the earlier work.
+
+  A "covered work" means either the unmodified Program or a work based
+on the Program.
+
+  To "propagate" a work means to do anything with it that, without
+permission, would make you directly or secondarily liable for
+infringement under applicable copyright law, except executing it on a
+computer or modifying a private copy.  Propagation includes copying,
+distribution (with or without modification), making available to the
+public, and in some countries other activities as well.
+
+  To "convey" a work means any kind of propagation that enables other
+parties to make or receive copies.  Mere interaction with a user through
+a computer network, with no transfer of a copy, is not conveying.
+
+  An interactive user interface displays "Appropriate Legal Notices"
+to the extent that it includes a convenient and prominently visible
+feature that (1) displays an appropriate copyright notice, and (2)
+tells the user that there is no warranty for the work (except to the
+extent that warranties are provided), that licensees may convey the
+work under this License, and how to view a copy of this License.  If
+the interface presents a list of user commands or options, such as a
+menu, a prominent item in the list meets this criterion.
+
+  1. Source Code.
+
+  The "source code" for a work means the preferred form of the work
+for making modifications to it.  "Object code" means any non-source
+form of a work.
+
+  A "Standard Interface" means an interface that either is an official
+standard defined by a recognized standards body, or, in the case of
+interfaces specified for a particular programming language, one that
+is widely used among developers working in that language.
+
+  The "System Libraries" of an executable work include anything, other
+than the work as a whole, that (a) is included in the normal form of
+packaging a Major Component, but which is not part of that Major
+Component, and (b) serves only to enable use of the work with that
+Major Component, or to implement a Standard Interface for which an
+implementation is available to the public in source code form.  A
+"Major Component", in this context, means a major essential component
+(kernel, window system, and so on) of the specific operating system
+(if any) on which the executable work runs, or a compiler used to
+produce the work, or an object code interpreter used to run it.
+
+  The "Corresponding Source" for a work in object code form means all
+the source code needed to generate, install, and (for an executable
+work) run the object code and to modify the work, including scripts to
+control those activities.  However, it does not include the work's
+System Libraries, or general-purpose tools or generally available free
+programs which are used unmodified in performing those activities but
+which are not part of the work.  For example, Corresponding Source
+includes interface definition files associated with source files for
+the work, and the source code for shared libraries and dynamically
+linked subprograms that the work is specifically designed to require,
+such as by intimate data communication or control flow between those
+subprograms and other parts of the work.
+
+  The Corresponding Source need not include anything that users
+can regenerate automatically from other parts of the Corresponding
+Source.
+
+  The Corresponding Source for a work in source code form is that
+same work.
+
+  2. Basic Permissions.
+
+  All rights granted under this License are granted for the term of
+copyright on the Program, and are irrevocable provided the stated
+conditions are met.  This License explicitly affirms your unlimited
+permission to run the unmodified Program.  The output from running a
+covered work is covered by this License only if the output, given its
+content, constitutes a covered work.  This License acknowledges your
+rights of fair use or other equivalent, as provided by copyright law.
+
+  You may make, run and propagate covered works that you do not
+convey, without conditions so long as your license otherwise remains
+in force.  You may convey covered works to others for the sole purpose
+of having them make modifications exclusively for you, or provide you
+with facilities for running those works, provided that you comply with
+the terms of this License in conveying all material for which you do
+not control copyright.  Those thus making or running the covered works
+for you must do so exclusively on your behalf, under your direction
+and control, on terms that prohibit them from making any copies of
+your copyrighted material outside their relationship with you.
+
+  Conveying under any other circumstances is permitted solely under
+the conditions stated below.  Sublicensing is not allowed; section 10
+makes it unnecessary.
+
+  3. Protecting Users' Legal Rights From Anti-Circumvention Law.
+
+  No covered work shall be deemed part of an effective technological
+measure under any applicable law fulfilling obligations under article
+11 of the WIPO copyright treaty adopted on 20 December 1996, or
+similar laws prohibiting or restricting circumvention of such
+measures.
+
+  When you convey a covered work, you waive any legal power to forbid
+circumvention of technological measures to the extent such circumvention
+is effected by exercising rights under this License with respect to
+the covered work, and you disclaim any intention to limit operation or
+modification of the work as a means of enforcing, against the work's
+users, your or third parties' legal rights to forbid circumvention of
+technological measures.
+
+  4. Conveying Verbatim Copies.
+
+  You may convey verbatim copies of the Program's source code as you
+receive it, in any medium, provided that you conspicuously and
+appropriately publish on each copy an appropriate copyright notice;
+keep intact all notices stating that this License and any
+non-permissive terms added in accord with section 7 apply to the code;
+keep intact all notices of the absence of any warranty; and give all
+recipients a copy of this License along with the Program.
+
+  You may charge any price or no price for each copy that you convey,
+and you may offer support or warranty protection for a fee.
+
+  5. Conveying Modified Source Versions.
+
+  You may convey a work based on the Program, or the modifications to
+produce it from the Program, in the form of source code under the
+terms of section 4, provided that you also meet all of these conditions:
+
+    a) The work must carry prominent notices stating that you modified
+    it, and giving a relevant date.
+
+    b) The work must carry prominent notices stating that it is
+    released under this License and any conditions added under section
+    7.  This requirement modifies the requirement in section 4 to
+    "keep intact all notices".
+
+    c) You must license the entire work, as a whole, under this
+    License to anyone who comes into possession of a copy.  This
+    License will therefore apply, along with any applicable section 7
+    additional terms, to the whole of the work, and all its parts,
+    regardless of how they are packaged.  This License gives no
+    permission to license the work in any other way, but it does not
+    invalidate such permission if you have separately received it.
+
+    d) If the work has interactive user interfaces, each must display
+    Appropriate Legal Notices; however, if the Program has interactive
+    interfaces that do not display Appropriate Legal Notices, your
+    work need not make them do so.
+
+  A compilation of a covered work with other separate and independent
+works, which are not by their nature extensions of the covered work,
+and which are not combined with it such as to form a larger program,
+in or on a volume of a storage or distribution medium, is called an
+"aggregate" if the compilation and its resulting copyright are not
+used to limit the access or legal rights of the compilation's users
+beyond what the individual works permit.  Inclusion of a covered work
+in an aggregate does not cause this License to apply to the other
+parts of the aggregate.
+
+  6. Conveying Non-Source Forms.
+
+  You may convey a covered work in object code form under the terms
+of sections 4 and 5, provided that you also convey the
+machine-readable Corresponding Source under the terms of this License,
+in one of these ways:
+
+    a) Convey the object code in, or embodied in, a physical product
+    (including a physical distribution medium), accompanied by the
+    Corresponding Source fixed on a durable physical medium
+    customarily used for software interchange.
+
+    b) Convey the object code in, or embodied in, a physical product
+    (including a physical distribution medium), accompanied by a
+    written offer, valid for at least three years and valid for as
+    long as you offer spare parts or customer support for that product
+    model, to give anyone who possesses the object code either (1) a
+    copy of the Corresponding Source for all the software in the
+    product that is covered by this License, on a durable physical
+    medium customarily used for software interchange, for a price no
+    more than your reasonable cost of physically performing this
+    conveying of source, or (2) access to copy the
+    Corresponding Source from a network server at no charge.
+
+    c) Convey individual copies of the object code with a copy of the
+    written offer to provide the Corresponding Source.  This
+    alternative is allowed only occasionally and noncommercially, and
+    only if you received the object code with such an offer, in accord
+    with subsection 6b.
+
+    d) Convey the object code by offering access from a designated
+    place (gratis or for a charge), and offer equivalent access to the
+    Corresponding Source in the same way through the same place at no
+    further charge.  You need not require recipients to copy the
+    Corresponding Source along with the object code.  If the place to
+    copy the object code is a network server, the Corresponding Source
+    may be on a different server (operated by you or a third party)
+    that supports equivalent copying facilities, provided you maintain
+    clear directions next to the object code saying where to find the
+    Corresponding Source.  Regardless of what server hosts the
+    Corresponding Source, you remain obligated to ensure that it is
+    available for as long as needed to satisfy these requirements.
+
+    e) Convey the object code using peer-to-peer transmission, provided
+    you inform other peers where the object code and Corresponding
+    Source of the work are being offered to the general public at no
+    charge under subsection 6d.
+
+  A separable portion of the object code, whose source code is excluded
+from the Corresponding Source as a System Library, need not be
+included in conveying the object code work.
+
+  A "User Product" is either (1) a "consumer product", which means any
+tangible personal property which is normally used for personal, family,
+or household purposes, or (2) anything designed or sold for incorporation
+into a dwelling.  In determining whether a product is a consumer product,
+doubtful cases shall be resolved in favor of coverage.  For a particular
+product received by a particular user, "normally used" refers to a
+typical or common use of that class of product, regardless of the status
+of the particular user or of the way in which the particular user
+actually uses, or expects or is expected to use, the product.  A product
+is a consumer product regardless of whether the product has substantial
+commercial, industrial or non-consumer uses, unless such uses represent
+the only significant mode of use of the product.
+
+  "Installation Information" for a User Product means any methods,
+procedures, authorization keys, or other information required to install
+and execute modified versions of a covered work in that User Product from
+a modified version of its Corresponding Source.  The information must
+suffice to ensure that the continued functioning of the modified object
+code is in no case prevented or interfered with solely because
+modification has been made.
+
+  If you convey an object code work under this section in, or with, or
+specifically for use in, a User Product, and the conveying occurs as
+part of a transaction in which the right of possession and use of the
+User Product is transferred to the recipient in perpetuity or for a
+fixed term (regardless of how the transaction is characterized), the
+Corresponding Source conveyed under this section must be accompanied
+by the Installation Information.  But this requirement does not apply
+if neither you nor any third party retains the ability to install
+modified object code on the User Product (for example, the work has
+been installed in ROM).
+
+  The requirement to provide Installation Information does not include a
+requirement to continue to provide support service, warranty, or updates
+for a work that has been modified or installed by the recipient, or for
+the User Product in which it has been modified or installed.  Access to a
+network may be denied when the modification itself materially and
+adversely affects the operation of the network or violates the rules and
+protocols for communication across the network.
+
+  Corresponding Source conveyed, and Installation Information provided,
+in accord with this section must be in a format that is publicly
+documented (and with an implementation available to the public in
+source code form), and must require no special password or key for
+unpacking, reading or copying.
+
+  7. Additional Terms.
+
+  "Additional permissions" are terms that supplement the terms of this
+License by making exceptions from one or more of its conditions.
+Additional permissions that are applicable to the entire Program shall
+be treated as though they were included in this License, to the extent
+that they are valid under applicable law.  If additional permissions
+apply only to part of the Program, that part may be used separately
+under those permissions, but the entire Program remains governed by
+this License without regard to the additional permissions.
+
+  When you convey a copy of a covered work, you may at your option
+remove any additional permissions from that copy, or from any part of
+it.  (Additional permissions may be written to require their own
+removal in certain cases when you modify the work.)  You may place
+additional permissions on material, added by you to a covered work,
+for which you have or can give appropriate copyright permission.
+
+  Notwithstanding any other provision of this License, for material you
+add to a covered work, you may (if authorized by the copyright holders of
+that material) supplement the terms of this License with terms:
+
+    a) Disclaiming warranty or limiting liability differently from the
+    terms of sections 15 and 16 of this License; or
+
+    b) Requiring preservation of specified reasonable legal notices or
+    author attributions in that material or in the Appropriate Legal
+    Notices displayed by works containing it; or
+
+    c) Prohibiting misrepresentation of the origin of that material, or
+    requiring that modified versions of such material be marked in
+    reasonable ways as different from the original version; or
+
+    d) Limiting the use for publicity purposes of names of licensors or
+    authors of the material; or
+
+    e) Declining to grant rights under trademark law for use of some
+    trade names, trademarks, or service marks; or
+
+    f) Requiring indemnification of licensors and authors of that
+    material by anyone who conveys the material (or modified versions of
+    it) with contractual assumptions of liability to the recipient, for
+    any liability that these contractual assumptions directly impose on
+    those licensors and authors.
+
+  All other non-permissive additional terms are considered "further
+restrictions" within the meaning of section 10.  If the Program as you
+received it, or any part of it, contains a notice stating that it is
+governed by this License along with a term that is a further
+restriction, you may remove that term.  If a license document contains
+a further restriction but permits relicensing or conveying under this
+License, you may add to a covered work material governed by the terms
+of that license document, provided that the further restriction does
+not survive such relicensing or conveying.
+
+  If you add terms to a covered work in accord with this section, you
+must place, in the relevant source files, a statement of the
+additional terms that apply to those files, or a notice indicating
+where to find the applicable terms.
+
+  Additional terms, permissive or non-permissive, may be stated in the
+form of a separately written license, or stated as exceptions;
+the above requirements apply either way.
+
+  8. Termination.
+
+  You may not propagate or modify a covered work except as expressly
+provided under this License.  Any attempt otherwise to propagate or
+modify it is void, and will automatically terminate your rights under
+this License (including any patent licenses granted under the third
+paragraph of section 11).
+
+  However, if you cease all violation of this License, then your
+license from a particular copyright holder is reinstated (a)
+provisionally, unless and until the copyright holder explicitly and
+finally terminates your license, and (b) permanently, if the copyright
+holder fails to notify you of the violation by some reasonable means
+prior to 60 days after the cessation.
+
+  Moreover, your license from a particular copyright holder is
+reinstated permanently if the copyright holder notifies you of the
+violation by some reasonable means, this is the first time you have
+received notice of violation of this License (for any work) from that
+copyright holder, and you cure the violation prior to 30 days after
+your receipt of the notice.
+
+  Termination of your rights under this section does not terminate the
+licenses of parties who have received copies or rights from you under
+this License.  If your rights have been terminated and not permanently
+reinstated, you do not qualify to receive new licenses for the same
+material under section 10.
+
+  9. Acceptance Not Required for Having Copies.
+
+  You are not required to accept this License in order to receive or
+run a copy of the Program.  Ancillary propagation of a covered work
+occurring solely as a consequence of using peer-to-peer transmission
+to receive a copy likewise does not require acceptance.  However,
+nothing other than this License grants you permission to propagate or
+modify any covered work.  These actions infringe copyright if you do
+not accept this License.  Therefore, by modifying or propagating a
+covered work, you indicate your acceptance of this License to do so.
+
+  10. Automatic Licensing of Downstream Recipients.
+
+  Each time you convey a covered work, the recipient automatically
+receives a license from the original licensors, to run, modify and
+propagate that work, subject to this License.  You are not responsible
+for enforcing compliance by third parties with this License.
+
+  An "entity transaction" is a transaction transferring control of an
+organization, or substantially all assets of one, or subdividing an
+organization, or merging organizations.  If propagation of a covered
+work results from an entity transaction, each party to that
+transaction who receives a copy of the work also receives whatever
+licenses to the work the party's predecessor in interest had or could
+give under the previous paragraph, plus a right to possession of the
+Corresponding Source of the work from the predecessor in interest, if
+the predecessor has it or can get it with reasonable efforts.
+
+  You may not impose any further restrictions on the exercise of the
+rights granted or affirmed under this License.  For example, you may
+not impose a license fee, royalty, or other charge for exercise of
+rights granted under this License, and you may not initiate litigation
+(including a cross-claim or counterclaim in a lawsuit) alleging that
+any patent claim is infringed by making, using, selling, offering for
+sale, or importing the Program or any portion of it.
+
+  11. Patents.
+
+  A "contributor" is a copyright holder who authorizes use under this
+License of the Program or a work on which the Program is based.  The
+work thus licensed is called the contributor's "contributor version".
+
+  A contributor's "essential patent claims" are all patent claims
+owned or controlled by the contributor, whether already acquired or
+hereafter acquired, that would be infringed by some manner, permitted
+by this License, of making, using, or selling its contributor version,
+but do not include claims that would be infringed only as a
+consequence of further modification of the contributor version.  For
+purposes of this definition, "control" includes the right to grant
+patent sublicenses in a manner consistent with the requirements of
+this License.
+
+  Each contributor grants you a non-exclusive, worldwide, royalty-free
+patent license under the contributor's essential patent claims, to
+make, use, sell, offer for sale, import and otherwise run, modify and
+propagate the contents of its contributor version.
+
+  In the following three paragraphs, a "patent license" is any express
+agreement or commitment, however denominated, not to enforce a patent
+(such as an express permission to practice a patent or covenant not to
+sue for patent infringement).  To "grant" such a patent license to a
+party means to make such an agreement or commitment not to enforce a
+patent against the party.
+
+  If you convey a covered work, knowingly relying on a patent license,
+and the Corresponding Source of the work is not available for anyone
+to copy, free of charge and under the terms of this License, through a
+publicly available network server or other readily accessible means,
+then you must either (1) cause the Corresponding Source to be so
+available, or (2) arrange to deprive yourself of the benefit of the
+patent license for this particular work, or (3) arrange, in a manner
+consistent with the requirements of this License, to extend the patent
+license to downstream recipients.  "Knowingly relying" means you have
+actual knowledge that, but for the patent license, your conveying the
+covered work in a country, or your recipient's use of the covered work
+in a country, would infringe one or more identifiable patents in that
+country that you have reason to believe are valid.
+
+  If, pursuant to or in connection with a single transaction or
+arrangement, you convey, or propagate by procuring conveyance of, a
+covered work, and grant a patent license to some of the parties
+receiving the covered work authorizing them to use, propagate, modify
+or convey a specific copy of the covered work, then the patent license
+you grant is automatically extended to all recipients of the covered
+work and works based on it.
+
+  A patent license is "discriminatory" if it does not include within
+the scope of its coverage, prohibits the exercise of, or is
+conditioned on the non-exercise of one or more of the rights that are
+specifically granted under this License.  You may not convey a covered
+work if you are a party to an arrangement with a third party that is
+in the business of distributing software, under which you make payment
+to the third party based on the extent of your activity of conveying
+the work, and under which the third party grants, to any of the
+parties who would receive the covered work from you, a discriminatory
+patent license (a) in connection with copies of the covered work
+conveyed by you (or copies made from those copies), or (b) primarily
+for and in connection with specific products or compilations that
+contain the covered work, unless you entered into that arrangement,
+or that patent license was granted, prior to 28 March 2007.
+
+  Nothing in this License shall be construed as excluding or limiting
+any implied license or other defenses to infringement that may
+otherwise be available to you under applicable patent law.
+
+  12. No Surrender of Others' Freedom.
+
+  If conditions are imposed on you (whether by court order, agreement or
+otherwise) that contradict the conditions of this License, they do not
+excuse you from the conditions of this License.  If you cannot convey a
+covered work so as to satisfy simultaneously your obligations under this
+License and any other pertinent obligations, then as a consequence you may
+not convey it at all.  For example, if you agree to terms that obligate you
+to collect a royalty for further conveying from those to whom you convey
+the Program, the only way you could satisfy both those terms and this
+License would be to refrain entirely from conveying the Program.
+
+  13. Use with the GNU Affero General Public License.
+
+  Notwithstanding any other provision of this License, you have
+permission to link or combine any covered work with a work licensed
+under version 3 of the GNU Affero General Public License into a single
+combined work, and to convey the resulting work.  The terms of this
+License will continue to apply to the part which is the covered work,
+but the special requirements of the GNU Affero General Public License,
+section 13, concerning interaction through a network will apply to the
+combination as such.
+
+  14. Revised Versions of this License.
+
+  The Free Software Foundation may publish revised and/or new versions of
+the GNU General Public License from time to time.  Such new versions will
+be similar in spirit to the present version, but may differ in detail to
+address new problems or concerns.
+
+  Each version is given a distinguishing version number.  If the
+Program specifies that a certain numbered version of the GNU General
+Public License "or any later version" applies to it, you have the
+option of following the terms and conditions either of that numbered
+version or of any later version published by the Free Software
+Foundation.  If the Program does not specify a version number of the
+GNU General Public License, you may choose any version ever published
+by the Free Software Foundation.
+
+  If the Program specifies that a proxy can decide which future
+versions of the GNU General Public License can be used, that proxy's
+public statement of acceptance of a version permanently authorizes you
+to choose that version for the Program.
+
+  Later license versions may give you additional or different
+permissions.  However, no additional obligations are imposed on any
+author or copyright holder as a result of your choosing to follow a
+later version.
+
+  15. Disclaimer of Warranty.
+
+  THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY
+APPLICABLE LAW.  EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT
+HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY
+OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO,
+THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+PURPOSE.  THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM
+IS WITH YOU.  SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF
+ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
+
+  16. Limitation of Liability.
+
+  IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
+WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS
+THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY
+GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE
+USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF
+DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD
+PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS),
+EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF
+SUCH DAMAGES.
+
+  17. Interpretation of Sections 15 and 16.
+
+  If the disclaimer of warranty and limitation of liability provided
+above cannot be given local legal effect according to their terms,
+reviewing courts shall apply local law that most closely approximates
+an absolute waiver of all civil liability in connection with the
+Program, unless a warranty or assumption of liability accompanies a
+copy of the Program in return for a fee.
+
+                     END OF TERMS AND CONDITIONS
+
+            How to Apply These Terms to Your New Programs
+
+  If you develop a new program, and you want it to be of the greatest
+possible use to the public, the best way to achieve this is to make it
+free software which everyone can redistribute and change under these terms.
+
+  To do so, attach the following notices to the program.  It is safest
+to attach them to the start of each source file to most effectively
+state the exclusion of warranty; and each file should have at least
+the "copyright" line and a pointer to where the full notice is found.
+
+    <one line to give the program's name and a brief idea of what it does.>
+    Copyright (C) <year>  <name of author>
+
+    This program is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation, either version 3 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with this program.  If not, see <http://www.gnu.org/licenses/>.
+
+Also add information on how to contact you by electronic and paper mail.
+
+  If the program does terminal interaction, make it output a short
+notice like this when it starts in an interactive mode:
+
+    <program>  Copyright (C) <year>  <name of author>
+    This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'.
+    This is free software, and you are welcome to redistribute it
+    under certain conditions; type `show c' for details.
+
+The hypothetical commands `show w' and `show c' should show the appropriate
+parts of the General Public License.  Of course, your program's commands
+might be different; for a GUI interface, you would use an "about box".
+
+  You should also get your employer (if you work as a programmer) or school,
+if any, to sign a "copyright disclaimer" for the program, if necessary.
+For more information on this, and how to apply and follow the GNU GPL, see
+<http://www.gnu.org/licenses/>.
+
+  The GNU General Public License does not permit incorporating your program
+into proprietary programs.  If your program is a subroutine library, you
+may consider it more useful to permit linking proprietary applications with
+the library.  If this is what you want to do, use the GNU Lesser General
+Public License instead of this License.  But first, please read
+<http://www.gnu.org/philosophy/why-not-lgpl.html>.
diff --git a/tools_pc/flicker_scan.py b/tools_pc/flicker_scan.py
new file mode 100644
index 00000000..a03fc26f
--- /dev/null
+++ b/tools_pc/flicker_scan.py
@@ -0,0 +1,114 @@
+#!/usr/bin/env python3
+"""flicker_scan.py - find A/B/A single-frame flicker in GE_PCDUMP captures.
+
+  flicker_scan.py DIR [--scale 4] [--thresh T] [--csv out.csv]
+
+Loads DIR/frame_NNNNNN.ppm in frame order (gaps skipped: only frames that
+exist are compared), block-average downscales by --scale, and flags frame n
+when diff(n, n-1) > T and diff(n, n-2) < T/4 -- i.e. n differs from its
+predecessor but the frame before that looks like n again (or n-1 was the
+odd one out). diff = mean abs channel difference (0-255). Default T = 12.
+Pure stdlib; uses numpy if present. Always exits 0.
+"""
+import argparse, os, re, sys
+
+try:
+    import numpy as np
+except ImportError:
+    np = None
+
+
+def read_ppm(path):
+    with open(path, 'rb') as f:
+        data = f.read()
+    toks, pos = [], 0
+    while len(toks) < 4:
+        while data[pos:pos + 1].isspace():
+            pos += 1
+        if data[pos:pos + 1] == b'#':
+            while data[pos:pos + 1] != b'\n':
+                pos += 1
+            continue
+        s = pos
+        while not data[pos:pos + 1].isspace():
+            pos += 1
+        toks.append(data[s:pos])
+    pos += 1
+    if toks[0] != b'P6' or int(toks[3]) != 255:
+        raise ValueError('not an 8-bit P6 ppm')
+    return int(toks[1]), int(toks[2]), data[pos:]
+
+
+def downscale(w, h, px, k):
+    """Block-average to (w//k, h//k); returns a flat sequence of ints."""
+    ow, oh = w // k, h // k
+    if np is not None:
+        a = np.frombuffer(px, dtype=np.uint8, count=w * h * 3).reshape(h, w, 3)
+        a = a[:oh * k, :ow * k].astype(np.float32)
+        return a.reshape(oh, k, ow, k, 3).mean(axis=(1, 3)).ravel()
+    out = []
+    n = k * k
+    for by in range(oh):
+        rows = [by * k + dy for dy in range(k)]
+        for bx in range(ow):
+            for c in range(3):
+                t = 0
+                for r in rows:
+                    base = (r * w + bx * k) * 3 + c
+                    for dx in range(k):
+                        t += px[base + dx * 3]
+                out.append(t / n)
+    return out
+
+
+def diff(a, b):
+    if np is not None:
+        return float(np.abs(a - b).mean())
+    return sum(abs(x - y) for x, y in zip(a, b)) / max(len(a), 1)
+
+
+def main():
+    ap = argparse.ArgumentParser(description=__doc__.split('\n')[0])
+    ap.add_argument('dir')
+    ap.add_argument('--scale', type=int, default=4)
+    ap.add_argument('--thresh', type=float, default=12.0)
+    ap.add_argument('--csv')
+    a = ap.parse_args()
+    files = []
+    for n in os.listdir(a.dir):
+        m = re.fullmatch(r'frame_(\d+)\.ppm', n)
+        if m:
+            files.append((int(m.group(1)), n))
+    files.sort()
+    t = a.thresh
+    rows, flagged = [], []
+    prev = []   # last two (frame, thumb)
+    for fn, name in files:
+        try:
+            w, h, px = read_ppm(os.path.join(a.dir, name))
+            th = downscale(w, h, px, max(a.scale, 1))
+        except Exception as e:
+            print('skip %s: %s' % (name, e), file=sys.stderr)
+            continue
+        d1 = diff(th, prev[-1][1]) if len(prev) >= 1 else None
+        d2 = diff(th, prev[-2][1]) if len(prev) >= 2 else None
+        flag = d1 is not None and d2 is not None and d1 > t and d2 < t / 4
+        rows.append((fn, d1, d2, flag))
+        if flag:
+            flagged.append(fn)
+            print('FLICKER frame %d (odd frame: %d): diff(n-1)=%.2f diff(n-2)=%.2f'
+                  % (fn, prev[-1][0], d1, d2))
+        prev = (prev + [(fn, th)])[-2:]
+    if a.csv:
+        with open(a.csv, 'w') as f:
+            f.write('frame,diff_prev,diff_prev2,flag\n')
+            for fn, d1, d2, fl in rows:
+                f.write('%d,%s,%s,%d\n' % (fn, '' if d1 is None else '%.3f' % d1,
+                                           '' if d2 is None else '%.3f' % d2, int(fl)))
+    print('scanned %d frames, %d flagged (thresh %.1f, scale %d)' %
+          (len(rows), len(flagged), t, a.scale))
+    return 0
+
+
+if __name__ == '__main__':
+    sys.exit(main())
diff --git a/tools_pc/framediff.py b/tools_pc/framediff.py
index caba350f..1c98aaa0 100644
--- a/tools_pc/framediff.py
+++ b/tools_pc/framediff.py
@@ -48,7 +48,8 @@ USAGE
 Frames are matched by filename stem (frame_000200). .ppm (P6) and .png
 are both accepted on either side. Exit status: 0 = all frames within
 threshold, 1 = at least one frame/region over threshold, 2 = usage / IO
-error (e.g. a golden frame with no candidate).
+error (e.g. a golden frame with no candidate), 3 = a candidate frame's size
+differs from its golden (never a pass: the frames were not compared).
 
 Regenerate the golden set (after a deliberate visual change):
   GE_PCDUMP="640-1120:240" ./build-pc/ge007.x86_64.exe -level_09
@@ -419,7 +420,7 @@ def main(argv):
             print("FAIL %s: size %dx%d vs golden %dx%d"
                   % (stem, cw, ch, gw, gh))
             results.append({"frame": stem, "status": "size"})
-            worst = 2
+            worst = max(worst, 3)
             continue
         mask = build_mask_bitmap(gw, gh, masks)
 
@@ -470,6 +471,7 @@ def main(argv):
     print("\n%d/%d frames within threshold%s"
           % (len(results) - n_fail, len(results),
              "" if worst == 0 else "  --  REGRESSION" if worst == 1
+             else "  --  SIZE MISMATCH" if worst == 3
              else "  --  ERROR"))
     return worst
 
diff --git a/tools_pc/levelbench.sh b/tools_pc/levelbench.sh
index 917722f2..bcb04ce6 100644
--- a/tools_pc/levelbench.sh
+++ b/tools_pc/levelbench.sh
@@ -13,6 +13,10 @@
 #     --record FILE   human plays (no script); pad input saved to FILE
 #     --replay FILE   replay a --record file instead of a script
 #                     (both pin GE_RSEED so guards make the same choices)
+#                     (also writes FILE.meta: level, secs, seed, binary_md5,
+#                     recorded, script=none -- consumed by demobatch.sh)
+#     --dump SPEC     pass GE_PCDUMP=SPEC ("lo-hi[:step]"); the ./ppm frames
+#                     are moved to <out>/ppm/ after the run
 #     --gamescope     wrap in nested `gamescope -r 90 -w 1280 -h 800 -f`
 #                     (Deck desktop mode: approximates Game Mode pacing)
 #
@@ -22,7 +26,7 @@
 # play), then 1.2 s Z holds every 4 s, so guards hear shots and start their
 # patterns. Z is fire in every control style and harmless in cutscenes.
 set -u
-SECS=45; SCRIPT=fight; REC=; REP=; EXE=; DIR=$PWD; OUT=; GSCOPE=0
+SECS=45; SCRIPT=fight; REC=; REP=; EXE=; DIR=$PWD; OUT=; GSCOPE=0; DUMP=
 while [ $# -gt 0 ]; do
     case $1 in
         --secs) SECS=$2; shift 2;;
@@ -32,8 +36,9 @@ while [ $# -gt 0 ]; do
         --out) OUT=$2; shift 2;;
         --record) REC=$2; SCRIPT=none; shift 2;;
         --replay) REP=$2; SCRIPT=none; shift 2;;
+        --dump) DUMP=$2; shift 2;;
         --gamescope) GSCOPE=1; shift;;
-        -h|--help) sed -n 2,22p "$0"; exit 0;;
+        -h|--help) sed -n 2,26p "$0"; exit 0;;
         *) break;;
     esac
 done
@@ -93,6 +98,20 @@ IENV=(GE_INPUTSCRIPT_MS=1 "GE_INPUTSCRIPT=$INPUT")
 [ "$SCRIPT" = none ] && IENV=("GE_RSEED=${GE_RSEED:-0123456789abcdef}")
 [ -n "$REC" ] && IENV+=("GE_INPUTRECORD=$REC")
 [ -n "$REP" ] && IENV+=("GE_INPUTREPLAY=$REP")
+if [ -n "$REC" ]; then
+    {
+        echo "level=$LVL"
+        echo "secs=$SECS"
+        echo "seed=${GE_RSEED:-0123456789abcdef}"
+        echo "binary_md5=$(md5sum "$EXE" | cut -c1-8)"
+        echo "recorded=$(date +%Y-%m-%dT%H:%M:%S)"
+        echo "script=none"
+    } > "$REC.meta"
+fi
+if [ -n "$DUMP" ]; then
+    IENV+=("GE_PCDUMP=$DUMP")
+    mkdir -p ppm; rm -f ppm/frame_*.ppm
+fi
 echo "input env: ${IENV[*]}" >> "$OUT/run.txt"
 env GE_QUITFRAME=$((SECS * 60)) "${IENV[@]}" "$@" \
     timeout $((SECS + 90)) "${CMD[@]}" > "$OUT/stdout.txt" 2>&1
@@ -105,6 +124,10 @@ LOG=$(find . -maxdepth 3 -name ge007.log -newer "$STAMP" 2>/dev/null | head -1)
 [ -n "$LOG" ] && cp "$LOG" "$OUT/ge007.log"
 for f in ge007.crash.log; do [ -f "$f" ] && [ "$f" -nt "$STAMP" ] && cp "$f" "$OUT/"; done
 rm -f "$STAMP"
+if [ -n "$DUMP" ]; then
+    mkdir -p "$OUT/ppm"
+    for f in ppm/frame_*.ppm; do [ -e "$f" ] && mv "$f" "$OUT/ppm/"; done
+fi
 
 L="$OUT/ge007.log"; [ -f "$L" ] || L="$OUT/stdout.txt"
 {
diff --git a/tools_pc/release_gate.sh b/tools_pc/release_gate.sh
new file mode 100644
index 00000000..30a5891f
--- /dev/null
+++ b/tools_pc/release_gate.sh
@@ -0,0 +1,166 @@
+#!/usr/bin/env bash
+# release_gate.sh -- run the automatable release checks per platform and print
+# one PASS/FAIL/SKIPPED table. Reuses verify.sh (pixel gate) and levelbench.sh
+# (scripted run + clean-quit + crash-log detection); invents no new harness.
+#
+#   release_gate.sh [--dry-run] [--secs N] [--out DIR] <win|deck HOST|mac HOST|all>
+#
+#   win          verify.sh sweep --platform win (pixel gate vs goldens/<lvl>/win)
+#                + scripted-fire smoke on -level_09 (GE_INPUTSCRIPT MLB1/MLB0,
+#                  the D595/D597 regression) + exit-path check (rc 0 through
+#                  GE_QUITFRAME, "quit requested" log line, no ge007.crash.log)
+#   deck HOST    queue 21 levelbench jobs through deckrunner (Steam shortcut must
+#                be running on the Deck), crash-free + clean-quit + fps summary.
+#                Pixel compare is SKIPPED (deck goldens 640x480 vs 1280x800 dumps).
+#   mac HOST     over ssh: optional build, 21-level crash-free sweep + fire smoke.
+#                No goldens exist -> crash-free only.
+#   all          every platform that is configured (HOSTs from env), else SKIPPED.
+#
+# Environment (no defaults are machine-specific):
+#   GE_WIN_DIR     game dir for the Windows runs (default: <repo>/build-pc, has data/)
+#   GE_DECK_HOST / GE_MAC_HOST   ssh targets for `all` (or pass as argument)
+#   GE_MAC_DIR     remote game dir on the Mac (binary + data/); required for mac
+#   GE_MAC_EXE     remote binary name inside GE_MAC_DIR (default: ge007)
+#   GE_MAC_BUILD_CMD  optional command run in GE_MAC_DIR before the sweep
+#   GE_GATE_SSH_OPTS  extra ssh/scp options
+# Every run ends through GE_QUITFRAME (levelbench sets it); nothing is hard-killed.
+# Remote temp/result files are removed after results are pulled.
+set -u
+ROOT=$(cd "$(dirname "$0")/.." && pwd)
+DRY=0; SECS=30; OUT=; MODE=; HOST=
+while [ $# -gt 0 ]; do
+    case $1 in
+        --dry-run) DRY=1; shift;;
+        --secs) SECS=$2; shift 2;;
+        --out) OUT=$2; shift 2;;
+        -h|--help) sed -n 2,26p "$0"; exit 0;;
+        win|all) MODE=$1; shift;;
+        deck|mac) MODE=$1; HOST=${2:-}; shift; [ $# -gt 0 ] && shift;;
+        *) echo "unknown argument: $1" >&2; exit 2;;
+    esac
+done
+[ -n "$MODE" ] || { sed -n 2,6p "$0" >&2; exit 2; }
+[ -n "$OUT" ] || OUT="$ROOT/scratch/release-gate-$(date +%Y%m%d-%H%M%S)"
+LEVELS="33 34 35 36 09 20 26 43 27 22 24 29 30 25 37 23 39 41 28 32 54"
+FIRE='1500:START;6000:MLB1;14000:MLB0'
+SSHO=${GE_GATE_SSH_OPTS:-}
+ROWS=(); FAILS=0
+row() { [ $DRY = 1 ] && [ "$3" = PASS ] && set -- "$1" "$2" DRY "$4"; ROWS+=("$1|$2|$3|$4"); [ "$3" = FAIL ] && FAILS=$((FAILS + 1)); }
+run() { if [ $DRY = 1 ]; then echo "+ $*"; return 0; fi; "$@"; }
+if [ $DRY = 1 ]; then OUT=$(mktemp -d); trap 'rm -rf "$OUT"' EXIT; else mkdir -p "$OUT"; fi
+
+# judge_summary FILE -> "PASS" or "FAIL:<why>" from a levelbench summary.txt
+judge_summary() {
+    local f=$1
+    [ -s "$f" ] || { echo "FAIL:no summary"; return; }
+    grep -q '^exit=0$' "$f" || { echo "FAIL:rc $(grep -m1 '^exit=' "$f" | cut -d= -f2)"; return; }
+    grep -q 'CRASH LOG PRESENT' "$f" && { echo "FAIL:crash log"; return; }
+    grep -q 'no clean quit line' "$f" && { echo "FAIL:no clean quit"; return; }
+    echo PASS
+}
+
+gate_win() {
+    local d=${GE_WIN_DIR:-$ROOT/build-pc} v rc
+    echo "== win: pixel sweep"
+    if [ $DRY = 1 ]; then
+        echo "+ $ROOT/tools_pc/verify.sh sweep --platform win"; rc=0
+    else
+        "$ROOT/tools_pc/verify.sh" sweep --platform win | tee "$OUT/win-sweep.txt"; rc=${PIPESTATUS[0]}
+    fi
+    if [ $rc = 0 ]; then row win "verify.sh sweep (21 lvls vs win goldens)" PASS ""
+    else row win "verify.sh sweep (21 lvls vs win goldens)" FAIL "rc $rc, see win-sweep.txt"; fi
+    echo "== win: scripted-fire smoke + exit path (level 09)"
+    run "$ROOT/tools_pc/levelbench.sh" --dir "$d" --secs "$SECS" --script "raw:$FIRE" \
+        --out "$OUT/win-fire" 09
+    if [ $DRY = 1 ]; then v=PASS; else v=$(judge_summary "$OUT/win-fire/summary.txt"); fi
+    if [ "$v" = PASS ]; then
+        row win "fire smoke -level_09 (MLB1/MLB0)" PASS ""
+        row win "exit path (rc 0 via GE_QUITFRAME, no crash log)" PASS ""
+    else
+        row win "fire smoke / exit path -level_09" FAIL "${v#FAIL:}"
+    fi
+}
+
+gate_mac() {
+    local h=${HOST:-${GE_MAC_HOST:-}} dir=${GE_MAC_DIR:-} exe=${GE_MAC_EXE:-ge007} l v bad=0 badl=
+    if [ -z "$h" ] || [ -z "$dir" ]; then
+        row mac "all checks" SKIPPED "no host (arg/GE_MAC_HOST) or GE_MAC_DIR"; return
+    fi
+    # shellcheck disable=SC2086  # SSHO is an option list
+    run scp $SSHO "$ROOT/tools_pc/levelbench.sh" "$h:$dir/levelbench.sh"
+    # shellcheck disable=SC2086
+    [ -n "${GE_MAC_BUILD_CMD:-}" ] && run ssh $SSHO "$h" "cd '$dir' && $GE_MAC_BUILD_CMD"
+    for l in $LEVELS; do
+        # shellcheck disable=SC2086
+        run ssh $SSHO "$h" "cd '$dir' && bash levelbench.sh --dir . --exe './$exe' --secs $SECS --script 'raw:$FIRE' --out gate-$l $l >/dev/null 2>&1; cat gate-$l/summary.txt; rm -rf gate-$l" \
+            > "$OUT/mac-$l.txt" 2>&1
+        [ $DRY = 1 ] && continue
+        v=$(judge_summary "$OUT/mac-$l.txt")
+        [ "$v" = PASS ] || { bad=$((bad + 1)); badl="$badl $l(${v#FAIL:})"; }
+    done
+    # shellcheck disable=SC2086
+    run ssh $SSHO "$h" "cd '$dir' && rm -rf levelbench.sh gate-* ppm ge007.crash.log"
+    if [ $bad = 0 ]; then row mac "21-level crash-free + fire smoke" PASS ""
+    else row mac "21-level crash-free + fire smoke" FAIL "$bad bad:$badl"; fi
+    row mac "pixel compare" SKIPPED "no mac goldens exist"
+}
+
+gate_deck() {
+    local h=${HOST:-${GE_DECK_HOST:-}} l n=0 v bad=0 badl= name t pre last max
+    if [ -z "$h" ]; then row deck "all checks" SKIPPED "no host (arg/GE_DECK_HOST)"; return; fi
+    echo "== deck: checking for duplicate deckrunner instances"
+    if [ $DRY = 1 ]; then echo "+ ssh $h pgrep -af deckrunner.sh"; else
+        # shellcheck disable=SC2086
+        ssh $SSHO "$h" 'pgrep -af deckrunner.sh' || true
+        echo "(if more than one deckrunner.sh is listed, stop the extras before continuing)"
+    fi
+    # shellcheck disable=SC2086
+    run scp $SSHO "$ROOT/tools_pc/levelbench.sh" "$h:levelbench.sh"
+    pre="gate$(date +%H%M%S)"
+    for l in $LEVELS; do
+        n=$((n + 1)); name=$(printf '%s-%03d-%s' "$pre" "$n" "$l")
+        # shellcheck disable=SC2086
+        run ssh $SSHO "$h" "mkdir -p ~/lb/queue && echo '--secs $SECS --script fight $l' > ~/lb/queue/$name.job"
+    done
+    # the runner is serial; wait for the last job's .finished marker
+    last=$(printf '%s-%03d-%s' "$pre" 21 54); t=0; max=$((21 * (SECS + 60)))
+    if [ $DRY = 1 ]; then echo "+ poll: ssh $h test -e ~/lb/$last.finished (timeout ${max}s)"; else
+        # shellcheck disable=SC2086
+        until ssh $SSHO "$h" "test -e ~/lb/$last.finished" 2>/dev/null; do
+            sleep 10; t=$((t + 10))
+            [ $t -ge $max ] && { echo "deck: timed out waiting (is the deckrunner shortcut running?)" >&2; break; }
+        done
+    fi
+    n=0
+    for l in $LEVELS; do
+        n=$((n + 1)); name=$(printf '%s-%03d-%s' "$pre" "$n" "$l")
+        # shellcheck disable=SC2086
+        run ssh $SSHO "$h" "cat ~/lb/$name/summary.txt" > "$OUT/deck-$l.txt" 2>&1
+        [ $DRY = 1 ] && continue
+        v=$(judge_summary "$OUT/deck-$l.txt")
+        [ "$v" = PASS ] || { bad=$((bad + 1)); badl="$badl $l(${v#FAIL:})"; }
+    done
+    # shellcheck disable=SC2086
+    run ssh $SSHO "$h" "rm -rf ~/lb/$pre-* ~/lb/done/$pre-* ~/levelbench.sh"
+    if [ $bad = 0 ]; then row deck "21-level crash-free + clean quit" PASS ""
+    else row deck "21-level crash-free + clean quit" FAIL "$bad bad:$badl"; fi
+    row deck "levelbench fps summary" INFO "D578 cadence / D583 SLOW lines in deck-<lvl>.txt"
+    row deck "pixel compare" SKIPPED "deck goldens 640x480 vs 1280x800 dumps (known v0.7.0 item)"
+}
+
+case $MODE in
+    win) gate_win;;
+    deck) gate_deck;;
+    mac) gate_mac;;
+    all) gate_win; gate_deck; gate_mac;;
+esac
+
+echo
+echo "=== RELEASE GATE ($([ $DRY = 1 ] && echo dry-run || echo "$OUT")) ==="
+printf '%-5s %-52s %-8s %s\n' PLAT CHECK RESULT NOTE
+for r in "${ROWS[@]}"; do
+    IFS='|' read -r a b c d <<<"$r"
+    printf '%-5s %-52s %-8s %s\n' "$a" "$b" "$c" "$d"
+done
+if [ $FAILS = 0 ]; then echo "OVERALL: no failures"; else echo "OVERALL: $FAILS FAIL"; fi
+[ $FAILS = 0 ]
diff --git a/tools_pc/verify.sh b/tools_pc/verify.sh
index da2fba57..67f9fee4 100644
--- a/tools_pc/verify.sh
+++ b/tools_pc/verify.sh
@@ -412,9 +412,13 @@ except Exception:
           # golden's frame stems don't line up with this run's --dump window
           # (goldens are captured at GOLDEN_DUMP; a
           # sweep using a different stride has nothing to compare against
-          # rather than a real regression). Only exit 1 is a verdict fail.
+          # rather than a real regression). exit 3 = a frame was captured at a
+          # different size than its golden (e.g. the Deck dumping 1280x800
+          # against 640x480 goldens): nothing was compared, so it is a fail,
+          # never a pass. Exits 1 and 3 are verdict fails.
           case "$fexit" in
             1) status=REGRESSION; note="framediff vs $gdir" ;;
+            3) status=REGRESSION; note="frame size differs from golden $gdir (not compared)" ;;
             2) note="golden frame stems don't match --dump window ($dump) for $gdir" ;;
           esac
           local tolpct; tolpct=$(golden_tolpct_for "$name")