Registered graphics startup sweep — 2026-11-07

Actual guest execution on one fresh separate box, bx_psf77a49. At source pin time main was 0c66fe75482d3a86bacdd9b971ee377abdb4dc34; this is a bounded sample of four requested registered titles, not historical27 membership or backend parity. Evidence is in scratch/runs/20261107-graphics-registry-crash-sweep/.

All requested original registered inputs were available: Quake II 67 paths, Quake III 307, GLTron 92, Forsaken 873. Availability and SHA-256 pins are in runtime-pins.json; no replacement media was downloaded, substituted or modified. Runtime, shared DLL/type-library/font inputs and pngjs were pinned before the successful launch: 1,782 files, all unchanged in remote postflight. Git plus relevant-source.patch and the new lib/wininit-boot.js patch identify source; no source/media archive was made. Node v24.22.1 was recorded before launch; the executable SHA-256 was additionally recorded at cleanup, rather than claimed as a prelaunch binary pin.

Both shipped modules were pinned and recovered:

All eight actual sweep processes read the pinned standard module, recorded by a run-only read observer. Compatibility was retained, not executed. No fresh-build or compatibility-module behavior claim is made.

Successful command, from the isolated runtime root:

node tools/crash-sweep.js --apps=quake2_demo,quake3_demo,gltron,forsaken_demo --seconds=20 --modes=coop,threads --jobs=1 --stuck-after=0 --no-build --jsonl=/home/user/graphics-sweep-runtime/scratch/runs/20261107-graphics-registry-crash-sweep/actual2/sweep.jsonl

Only one guest process ran at a time; there was no RAM gate. The tool ran 03:58:28–04:00:37 UTC. Raw per-process logs, final PNGs and PCM outputs were retained before the stock tool removed its temporary files. Large logs were losslessly gzip-compressed for recovery. retrieval-verification.json verifies 88 recovered artifacts, 45 unique hashes, including both modules. Cleanup and scheduler comparison were recovered separately afterward.

App Cooperative Threads Personally inspected final captures
quake2_demo ok/content, 20.2 s ok/content, 20.2 s Both show GAME/MULTIPLAYER/OPTIONS/VIDEO/QUIT menu at 320×240.
quake3_demo ok/blank, 20.2 s ok/blank, 20.3 s Both uniformly gray at 640×480; no menu proved.
gltron exit:1/blank, 3.7 s exit:1/blank, 3.5 s Both uniformly teal at 640×480; no menu proved.
forsaken_demo ok/content, 20.3 s ok/content, 20.3 s Cooperative shows SINGLE PLAYER and other menu entries over the 3D backdrop; threads shows an attract-scene view, debug labels and Acclaim logo.

The stock comparison reports 0 of 4 apps disagree on its signature, frame-content and audio classifications. Forsaken's pictures differ despite matching coarse classifications; equal wall time does not mean equal guest progress. All eight audio signals are none, not evidence of working or silent audio. No unimplemented API or escaped trap was logged. Tool exit 0 also covers GLTron's guest exit 1 and is not an all-games pass.

Early-exit investigation and limits

One additional GLTron cooperative execution used existing selected API tracing, 0–40,000 batches, maximum 8 seconds/40,000 batches. It ended after 32,790 batches, again with guest exit 1. investigation/trace.log records ChoosePixelFormat returning 1, SetPixelFormat returning 1, then wglCreateContext(0x00310001) returning 0, followed by guest shutdown. The trace's final teal PNG was also inspected.

test/run.js supplies a GL canvas factory only for its explicit headless-GL route (or the separate software-Glide condition). This mandated sweep command requests neither headless GL nor software OpenGL. lib/gl-compat.js therefore has no canvas in Node and returns zero from context creation. Microsoft's wglCreateContext contract specifies NULL on failure. This is consistent with the existing GLTron notes' warning about generic sweep exit 1 on a different graphics route. No Win32/x86 semantic regression was demonstrated, so no production fix or speculative regression test was added. Quake III's gray frames remain a limitation of this actual command; they are not promoted to successful rendering or a diagnosed scheduler defect. No second graphics route was substituted.

Two earlier staging failures are preserved separately: the first omitted the shared test/binaries/tlbs/stdole2.tlb, and the second used a TMPDIR long enough to fail desktop Unix-socket startup with ECONNECTION. Both failed before guest execution. Existing shared runtime input staging and a short owned TMPDIR corrected these run-local setup errors; each subsequent invocation used a fresh JSONL path. No cybersecurity flag, access403, held-lane retry, guest patch, public deployment or nested worker occurred.

Cleanup and next work

All unique evidence was recovered and verified before delegated normal stop stop_5800d4a664d25d2d. It completed 04:02:07.270 UTC, forced=false, error=null; box state archived, active expiry null (original expiry 04:26:07.799 UTC). No lifecycle ownership remains to hand back. Remote cleanup found no live owned runtime processes or owned Unix listeners and removed all three owned temporary trees. No browser, TCP server or profile was started. Local work used no emulator, server, build or source-edit lane; Aranna/Lords2 coordination is on the append-only board. Free /home remained above 2 GB.

This documentation-only publication records the sweep, API investigation, module/hash checks, image inspection and cleanup. It does not establish controlled gameplay, audio quality, FPS, full graphics correctness, historical27 coverage or broad crash freedom. A separately requested software-GL or real browser run would be useful for Quake III/GLTron rendering; it is outside this mandated command. No emulator repair is warranted by this evidence.