GL/D3D startup crash sweep after waveOut correction, 2026-11-07

Twelve current-runtime startup windows across six registered originals reach their 20-second harness deadlines with no classified escaped crash, unimplemented API, guest exit, setup failure or external timeout. The full raw logs also contain no crash/RuntimeError/Error marker. Reviewed endpoint output matches the preceding sweep's checkpoint pattern. No new reproducible emulator failure was established, so no production patch or speculative clock change was made. Cooperative GLTron remains an unresolved startup blocker; threaded Forsaken's black endpoint does not establish a persistent regression.

Evidence: scratch/runs/20261107-graphics-post-audio-crash-sweep/ (ignored). This uses the named six-app union and passed controls from the post-allocator sweep. That predecessor predates the audio correction and supplies controls, not current-runtime acceptance. No historical27 census, new accounting framework, guest configuration change or extra gameplay execution was performed.

Tested composition and command

The staged source is main 5c54c4bb895ee3e3c23a4b7501811deae5bb2fa2, including published waveOut idle-gap fix ac1bcecee. The relevant tracked runtime diff is empty. new-lib_wininit-boot.js.patch preserves the existing untracked dependency. source-commit.txt, relevant-source.patch, owned-helpers.patch, runtime-pins.json and media-refs.json provide recovery from Git plus exact helper patches and existing media path/size/SHA-256 references. No source tree or media archive was created. Later concurrent main commits are not substituted for the tested composition.

All twelve actual standard-module reads match: 2,182,500 bytes, SHA-256 f9650607181301b5d2a9990420f437bd9f1ca10d40ac73067339f07499d53121. The actual loaded lib/host-audio.js reads match 211,499 bytes, SHA-256 e2ae7f782e7fbeb246471acfb3a93d7ae80eec9b195e44f02072310e81373d15. analysis.json retains each loaded audio-host observation, not just a local file hash. All 1,590 observed JS/JSON/WASM reads match staged/native pins. Remote preflight and postflight verify all 3,629 staged pins and 162 native dependency pins. Both WASM modules were retrieved; compatibility was not run.

Registered media closures are complete and hash-matched: Quake II 67 paths, Quake III 307, GLTron 92, Hexen II 16, Heretic II 1,826 and Forsaken 873. Renderer, CLI, scheduler and standard-module controls match their predecessor hashes; the corrected audio host differs as intended. Fresh native headless-GL and real WGL create/MakeCurrent/red-readback controls pass remotely.

Node v24.22.1 and native @node-3d/glfw 7.4.1 / @node-3d/webgl 6.0.3 use Mesa llvmpipe LLVM 20.1.2, WebGL 1.0, GLSL 4.50 on display :0. GL context observations, bridge results, loaded native binaries and their hashes are retained. Quake II/III, GLTron and Heretic II create native GL in both modes. Hexen II remains before GL. Forsaken uses its registered software Direct3D route. This is not browser WebGL or physical-GPU evidence.

The existing tool ran once, one heavy runtime at a time:

node tools/crash-sweep.js \
  --apps=quake2_demo,quake3_demo,gltron,hexen2_demo,heretic2_demo,forsaken_demo \
  --modes=coop,threads --headless-gl --jobs=1 --seconds=20 \
  --stuck-after=0 --no-build --jsonl=<run>/actual/sweep.jsonl

The retained preload selects the existing --png-canvas CLI option before argument parsing. PNGs are actual composed display output, not selected raw textures. Other passive observers retain module reads, native context and existing scheduler/bridge results. No input, private handler or alternate clock was used. The original CLI clock/budget defaults remain unchanged.

Outcomes and inspected output

All twelve full logs explicitly report [max-seconds] stopping after 20s. ok means that bounded harness outcome, not guest completion or playability. analysis.json classifies raw crash, unimplemented, guest-exit and setup markers independently of image content; the tool separately records external timeouts. All those failure categories are empty.

Original Cooperative Threaded Personally inspected composed endpoint
Quake II ok/content, 20.6 s ok/content, 20.6 s Both original Game/Multiplayer/Options/Video/Quit menus with console text behind.
Quake III ok/content, 20.6 s ok/content, 20.5 s Both original Arena Single Player/Multiplayer/Setup/Demos/Cinematics/Exit menus.
GLTron ok/blank, 20.5 s ok/content, 20.6 s Cooperative uniform black; threaded colored Game/Video/Audio/Quit menu.
Hexen II ok/content, 20.3 s ok/content, 20.4 s Both accelerated-version notices with OK focused; no dismissal, world or GL startup.
Heretic II ok/content, 20.7 s ok/content, 20.7 s Both textured original Main book menus.
Forsaken ok/content, 20.6 s ok/blank, 20.4 s Cooperative Single Player menu over textured 3D backdrop; threaded uniform black.

All twelve PNGs were inspected in two ordered contact sheets; original PNGs and individual SHA-256 reviews are retained. The reviewed menus show no newly observed visual regression at these checkpoints. This is not an equal-progress pixel-parity assertion. GLTron PCM is silent in both modes (cooperative reaches the 4 MiB cap); the other cases have no PCM bytes. No audio-quality conclusion follows. Threaded selection alone does not prove auxiliary guest work for titles that create no guest threads; scheduler receipts preserve that distinction.

Black endpoints and retained progress evidence

The retained Forsaken threaded investigation was read first, and its 10-second black, 20-second menu, corridor movement and fire/ammunition captures were inspected again without rerunning gameplay. That run measures black primary/back targets during continuing allocation and conversion, then new geometry and nonblack target output in the same process. The fresh sweep ends at main EIP 0x004cd2cb, with its preceding batches progressing through 0x004cd2f2..0x004cd340. This is inside the same retained 0x4cd2c2..0x4cd40e conversion loop, not a newly established stuck EIP or lost compositor output. It supports the known progress-phase explanation. Sparse sweep observations cannot prove every operation correct or the precise clear that began this interval. No persistent regression or scheduler fix is justified.

Cooperative GLTron's fresh main samples still progress through the known decoder work (0x7a71b7, 0x7916b7, 0x7ac752; final 0x7917cf). Its black endpoint is consistent with the retained audio-service and clock-budget evidence. The latter already adopts ac1bcecee, observes 50 actual pumps and 40 correctly eligible completions, and pairs all 40 with callback/semaphore grant and header resubmission before the next publication. Every full publication gives runnable main 4,000 retired blocks. The retained 80-second black capture and boundary analysis were inspected; this task does not repeat that census or claim its observations are new sweep measurements.

The remaining causal boundary is the CLI's unconditional outer-batch time charge before the audio pump, and the resulting decoder-work/playback capacity ratio. A correction requires an explicit emulated CPU-time/idle-deadline model and a work-preserving slice-partition regression. Neither a smaller tick nor another broad original run would establish that model. The idle-gap correction is tested, but it has not fixed cooperative GLTron menu startup. No new incorrect completed primitive was demonstrated here.

Exact retained run IDs are 20261107-graphics-post-allocator-crash-sweep, 20261107-forsaken-threaded-black, 20261107-gltron-audio-service and 20261107-gltron-clock-budget. retained-progress-refs.json records hashes of the consulted unique evidence; black-progress-alignment.json retains fresh deadline-adjacent registers. No broad gameplay repetition was performed.

Verification, recovery and shutdown

Existing crash-sweep regressions pass 20/20; comparison tests pass 2/2. waveout-audio, waveout-callback-function-wasm, wave-callback-scheduler and batch-clock tests pass. Runtime pin/load audits, full raw-error inspection, retrieval verification and own-run validation pass. This is documentation-only work against the pinned module with --no-build; no new production build or fail-before/pass-after repair is claimed. The global evidence checker reports seven unrelated containment errors and older missing/invalid records; this run has no checker findings. Unregistered apps remain investigation evidence, not unrelated candidate-corpus completions.

All 72 remote references / 58 unique contents / 21,132,348 referenced bytes, including twelve raw logs, twelve compositor PNGs, PCM, both modules and unique native binaries, were retrieved and size/SHA-256 verified before shutdown. Large logs are losslessly gzip-compressed. Local and remote launch free disk exceed 5 GB; local free space stays above 8 GB. Recovery uses Git SHA, relevant diff/new-file patches, exact runtime/module pins and existing media references.

One UUID-idempotent fresh box bx_38zwh4es was created at 23:27:29Z with original expiry 2026-11-07T23:52:29.188Z. A detached independent 25-minute watchdog and separate normal-stop delegate were armed before allocation and observed alive during the sweep; retained maximum heartbeat gap is 0.784 s. After verified retrieval, explicitly delegated normal stop stop_d97e44a67fca3ae8 completed 2026-11-07T23:34:12.362Z, forced=false, error=null. The box is archived, current expiry null; archive retention is unknown. normal-stop-confirmed.json is the receipt. Both lifecycle delegates exit zero and their supervisor has no remaining children. Remote cleanup finds no owned live runtime and removes its temporary directory. The exact local board watcher is closed. No browser, profile, control port or held input was created; no box was resumed/deleted and no public deployment occurred.

This bounded changed-runtime sweep is complete. No newly reproducible crash needs repair. Cooperative GLTron's documented clock-model boundary remains; Hexen II post-notice startup is untested. Startup absence of crash and retained menus do not establish full playability, browser parity, FPS, audio quality or full27 completion.