Expanded registered startup crash sweep, 7 November 2026

The actual registered Hexen II, Heretic II and Forsaken originals were each run once under cooperative scheduling and once with the real-worker backend. All six reached the harness deadline without an escaped emulator trap, unimplemented API, guest exit, setup error or external timeout. This establishes bounded startup outcomes only: Hexen II remained at its opening notice; Heretic II and Forsaken reached their original menus. Sweep exit zero does not establish that all cases pass gameplay or graphics requirements.

Evidence: scratch/runs/20261107-graphics-expanded-crash/. Predecessors: native GL failure and repair and repaired native originals. No production source, guest setting, installation registry, clock override, media or configuration was changed. No nested worker or Claude pane was used.

Inputs and preflight

The pinned actual-main source was df08c786ca331adeb21b76c122b8b99cefcc8888, including the native shader and presentation repairs and the subsequently published owning-Worker MIDI repair abf1b8c5b31c501e81871392256740ffafdcd138. The relevant tracked binary diff is empty. The existing untracked lib/wininit-boot.js is preserved as its own new-file patch; run-specific inputs have an owned-input patch. Git supplies the source archive. No source tree or media archive was made.

The actual registry/media closure was checked before allocation and again remotely before any guest: Hexen II 16 paths, Heretic II 1,826, Forsaken 873. The manifest entries, registered executable and static-import DLLs are included. registry-media-closure.json records registrations, original paths, sizes and SHA-256 references without copying media into evidence. All manifest size/hash constraints matched. No known title prerequisite was unmet. Heretic II retained its registered 1998 clock and OpenGL arguments; Hexen II retained its root executable layout avoiding the bundled MiniGL. The opening Hexen notice still requires an ordinary OK action to proceed; the requested no-input sweep does not dismiss it. The earlier browser MIDI failure is repaired, but this run does not reach MIDI initialization or requalify that repair.

All 3,166 staged runtime/input pins matched before guests and after completion. The 856 observed JS/JSON/WASM reads match their runtime/native pins, with no unmatched read. Each of the six guests actually read the standard module: 2,182,493 bytes, SHA-256 e41511fe861363fbe1639d80f20cbcdd7240f46937c93a801eaa256ca094c467. Both standard and compatibility modules were recovered; the latter was not executed. Read provenance is in retrieved/loaded-audit.json.

Native dependencies were installed separately: @node-3d/glfw 7.4.1 and @node-3d/webgl 6.0.3. Node v24.22.1, display :0, Mesa llvmpipe LLVM 20.1.2, WebGL 1.0 and GLSL 4.50 were observed. A real native context returned red RGBA [255,0,0,255], GL error zero, and zero live contexts after destruction. The actual headless-GL test and real WGL fixed-function creation/MakeCurrent/ red-readback regression passed before any original. The 162 native dependency pins and Node binary hash are recorded; native binaries were recovered individually. No held-lane retry or RAM gate was used.

Serial paired outcomes

node tools/crash-sweep.js \
  --apps=hexen2_demo,heretic2_demo,forsaken_demo --modes=coop,threads \
  --headless-gl --jobs=1 --seconds=20 --stuck-after=0 --no-build \
  --jsonl=<run>/actual3/sweep.jsonl

The JSONL was fresh. The six guests ran serially from 15:35:31 to 15:37:34 UTC, with one heavy runtime at a time. An external preload retained each individual raw log, PNG and PCM before sweep cleanup, recorded actual module reads and observed the existing GL bridge and scheduler without changing guest behavior.

Registered title Cooperative Real-worker backend Personally inspected final display
hexen2_demo ok/content, 20.3 s, 440,009 batches ok/content, 20.3 s, 428,888 batches Both show the original GLHexen2 accelerated-version notice with OK focused. No menu or GL context reached.
heretic2_demo ok/content, 21.5 s, 31,623 batches ok/content, 20.6 s, 35,055 batches Both show the original book menu: Game, Multiplayer, Options, Video/Sound Settings, Quit.
forsaken_demo ok/content, 20.4 s, 40,333 batches ok/content, 20.4 s, 41,764 batches Both show SINGLE PLAYER / DEATH MATCH / SETUP BIKER / PLAY DEMO / OPTIONS / EXIT GAME over the 3D backdrop.

All six PNGs were personally inspected and decoded. Hexen II's pair is byte-identical, as is Heretic II's pair; Forsaken's attract/menu backdrop differs slightly. No blank capture occurred. All six audio signals are none, zero bytes; no audio-quality conclusion follows.

Heretic II created and made current a genuine native GL context in each mode, with nonzero WGL results and the observed llvmpipe provider. There was no recorded GL bridge exception or software fallback. Hexen II never passed its notice, so native capability preflight is not title-level GL execution proof. Forsaken used the registered inline software Direct3D route (dx slot 6), not native GL; the availability of the factory did not change that route.

Actual scheduler observations confirm cooperative and worker backends respectively. Hexen II and Heretic II created no auxiliary guest threads on these observed paths: their threaded arms establish backend selection, not concurrent auxiliary guest execution. Forsaken's cooperative arm ended with two active auxiliary instances, advanced EIPs and a sleeping thread's 7,056 sleep observations. Its worker arm actually executed three auxiliary guest threads in slots 2, 5 and 6, with 20,060, 8,691 and 18,488 completed slices. One had exited by the final snapshot. The two still-live Node worker IDs were 4 and 8 before cleanup; the exited worker's final ID was unavailable. Slice counts include parking/empty work and are not throughput measurements.

No unexpected real emulator crash was demonstrated, so there was no justified cause-based production repair or fail-before/pass-after regression to add. Unequal batch counts were not treated as a scheduler or rendering defect. The retained complete raw logs contain no escaped crash, unimplemented API or error line. These observations do not qualify graphics parity, audio quality, FPS, gameplay, historical27 membership or broad crash freedom.

Recovery, checks and lifecycle

Recovery verified 54 artifact references / 46 unique hashes, including all six individual raw logs, six PNGs, empty PCM artifacts, both WASM modules and seven native binaries. Large logs are retained losslessly as gzip, with local restoration hashes/lengths. Runtime and native postflight mismatches are zero. The crash-sweep regression script passed 20/20 tests. One initially attempted nonexistent focused-test filename is retained separately; it ran no test. The corrected test invocation and JavaScript syntax checks passed. No production code changed, so this documentation-only publication required no new build; the requested sweep reused the pinned module with --no-build. All three title result records pass schema, artifact existence and containment checks. The global evidence checker retains unrelated preexisting Alice and Tunnel Twist failures; no global pass is claimed.

A separate local lifecycle watchdog was armed before the one fresh allocation, with a 25-minute cap and delegated normal-stop authority. A persisted UUID Idempotency-Key created bx_kku6pzkp at 15:33:38 UTC, original expiry 15:58:38.545 UTC. No old box was resumed. After artifact retrieval and hash verification, normal stop stop_6365d3a6fe0723ac completed at 15:38:07.334 UTC, forced:false, error null. The box is archived, stoppedAt 15:38:07.318 UTC, active expiry/archiveAfter null. No running box ownership remains to hand back.

Terminal checks found no owned runtime process and removed the owned temporary directory. No browser, profile, input writer or control port was created. System listener/socket censuses are retained. Local lifecycle/messageboard watchers were closed. Sampled shared /home free space stayed above 2 GB; no peer data was deleted. The task ends with no original replay.

The remaining startup-coverage prerequisite is a separately scoped ordinary OK dismissal for Hexen II before assessing its post-notice native GL/MIDI startup under both schedulers. Heretic II and Forsaken menu startup is covered; further gameplay or rendering qualification requires its own evidence.