Current six-original startup crash sweep, 14 November 2026

Twelve completed 20-second windows produced no classified trap, unimplemented API, guest exit, exception or runtime timeout. Four final compositor images are blank. No concrete emulator defect was established, so no production repair or speculative test was added. This is bounded startup evidence for six named originals, not historical27, browser, physical-GPU, parity or gameplay acceptance.

Evidence is in ignored scratch/runs/20261114-graphics-six-app-crash/. result.json names exactly twelve selected final PNGs; all were personally inspected. observations.json contains the individual signatures, full-log audit, window events, loaded-module hashes and actual process waits.

Tested composition

The locked full build passed on main a9448409727cc11d065591831103f03ae1b71c22, including published D3D9 repair 2a567ae263f968d0304b70e0376c67ef48c302ba, plus the existing foreign WIP. build/source.patch records the relevant tracked WIP; build/new-wininit.patch records the existing untracked lib/wininit-boot.js. Per-launch source pins match the recorded built-source pins across all twelve windows. This does not claim a pristine committed source tree or an audit of every runtime JS read.

All twelve observed standard-WASM reads were 2,199,488 bytes, SHA-256 c7f5bada4da4b4ed95b235add0aa1c9843176805d29d5b1ee79359752a19d348. The actual fresh build matches the D3D9 predecessor's module; no module was selected by inference. Compatibility WASM was built but not exercised.

Environment: local Node v24.18.1, display :0, @node-3d/glfw 7.4.1, @node-3d/webgl 6.0.3, llvmpipe LLVM 20.1.2, WebGL 1.0 / GLSL 4.50. The actual native shader/readback control passed. Registered game-path manifests were hashed before and after: Quake II 67, Quake III 307, GLTron 92, Hexen II 18, Heretic II 1,826 and Forsaken 873 references, with no missing paths or changed bytes. media-paths.json records existing paths and SHA-256 values; no media was downloaded or copied. No guest reverse engineering, private-payload decoding, guest patch, game input or configuration edit was performed.

Executed windows and reviewed images

Each app/mode used the existing tool, serially, with the heavy-runtime lock released between launches:

node tools/crash-sweep.js --apps=APP --modes=MODE \
  --headless-gl --jobs=1 --seconds=20 --stuck-after=0 \
  --no-build --jsonl=RUN/APP.MODE/sweep.jsonl

The retained host-only observe.cjs preload appends supported --real-ticks and --png-canvas before test/run.js parses arguments. It records actual WASM reads and retains the final PNG and compressed log before the sweep removes its temporary files. No back-canvas, surface dump or frame series is enabled. PCM is classified by the existing sweep and discarded; no audio claim follows.

Original Cooperative final compositor Threaded final compositor Signature
Quake II Original Game/Multiplayer/Options/Video/Quit menu over console/backdrop Same menu Both ok/content
Quake III Uniform gray Uniform gray Both ok/blank
GLTron Uniform black Uniform black Both ok/blank
Hexen II Original Single Player/Multiplayer/Options/Help/Quit menu Same menu Both ok/content
Heretic II Textured original Main book menu Same menu Both ok/content
Forsaken Original heart/sword logo splash Same splash Both ok/content

Hexen II's existing registry startup flag already bypasses its first-launch notice; this is beyond the notice-only November 7 observation without a task change to guest settings. Forsaken retains its registered software D3D route; passing --headless-gl does not convert it into GL coverage. The two scheduler flags establish requested modes, not auxiliary guest-Worker execution for every title. All final captures are 640×480; sweep ok means a completed harness window, not successful rendering, guest completion or gameplay.

Quake III creates and shows Quake 3: Arena in both logs, but completed source draws/swaps and their compositor ownership were not instrumented. GLTron's logs show an SDL_app creation at 0×0 and no subsequent creation/show event. Those public observations do not distinguish slow startup, context/draw failure and presentation failure. No modal or intentional guest exit was observed in the qualified twelve windows. The blank cases remain visually unqualified.

Checks, exclusions and cleanup

Passed: flock /tmp/wine-heavy-runtime.lock flock /tmp/wine-build.lock bash tools/build.sh; actual native test/test-headless-gl.js; 20/20 test/registry-crash-sweep.test.js cases; and test/test-crash-sweep-against.js. These controls do not upgrade the originals' bounded outcomes into broad acceptance.

An initial setup attempt found pruned old native-package folders and a skipped preflight. Its Quake II cooperative menu and cancelled threaded attempt are excluded; the reviewed setup PNG was deselected and removed. Fresh pinned native dependencies were installed only inside this run, then positively qualified before the twelve windows. Their 164 file hashes and npm recovery lock remain; the regenerable installed distribution was removed after checking its hashes and joining all runtimes.

One later 170-second lock-queue wait expired before any observed module load. Only the seven remaining app/mode cases were resumed, without repeating the five completed cases. The resumed independent watchdog preserved the original qualified-sweep absolute deadline, approximately 09:49:14.750 UTC. All twelve actual app runs ended normally at their 20-second deadlines, and both qualified supervisors and the excluded setup have terminal watchdog waits with both EOFs. Recorded process/group checks found no survivors or residual-group cleanup in the twelve completed cases. The final app ended at approximately 09:44:25.508 UTC. Locks were released; final free space exceeded 14.19 GB decimal.

No box was created: ID null, state NOT_CREATED, expiry null, stop receipt N/A. No source-tree archive, WASM dump, browser, deployment or other worker was created by this task. Foreign WIP remains untouched.

The next causal blocker is the missing public context-bind/draw/swap versus compositor discriminator for gray Quake III and black GLTron. A bounded passive trace at those boundaries is needed before proposing a generic fix. Repeating the same unattended windows alone would not resolve it.