Quake II and GLTron original gameplay with native GL and threads — 2026-11-07

Evidence: scratch/runs/20261107-graphics-repaired-native-sweep/. Both registered originals reached inspected ordinary gameplay using the supported --headless-gl --threads route on current main. This closes the two menu-only followup gaps in the repaired-original replay. No new emulator failure was demonstrated; no production code changed.

Identity and prerequisites

The tested checkout was main 74c0ef3aeac845d871114a804faa258524cce4cb, including shader repair aaaf6e370 and presentation repair 36a212cac. runtime-pins.json records 606 individual inputs, including complete original Quake II (67 paths) and GLTron (92 paths) closures. All registered hashes matched. The relevant tracked source patch is empty; the existing untracked lib/wininit-boot.js is retained as its own binary diff. Git plus these patches is the recovery recipe; no source/media archive or repeated media was created.

All four original processes actually read standard WASM SHA-256 e41511fe861363fbe1639d80f20cbcdd7240f46937c93a801eaa256ca094c467 (2,182,493 bytes). Both standard and compatibility modules were recovered; the compatibility module was not executed. The runtime manifest and retrieved/identity-postflight.json pin loaded modules and source independently.

One fresh UUID Idempotency-Key created box bx_jb7v82vs at 06:38:00Z, with original expiry 07:13:00Z. No earlier box was resumed. Native packages were installed separately from the pinned runtime package files: @node-3d/glfw 7.4.1 and @node-3d/webgl 6.0.3. Their normal installation scripts supplied native binaries. There was no access/403 hold or retry. Node v24.22.1 used display :0, Mesa llvmpipe LLVM 20.1.2, WebGL 1.0, GLSL 4.50. The 162 dependency pins and Node binary hash are retained in retrieved/native-pins.json; nine native binaries were individually recovered.

Before originals, native clear/readback returned [255,0,0,255], GL error 0, then zero live contexts. test/test-headless-gl.js and the real WGL regression passed, including actual fixed-function shader creation, MakeCurrent and red readback. Local clock-domain/release mutation tests and CLI fault key-release cleanup passed. Two older source-extraction test fixtures did not pass: test-cli-tick-ms-input.js --source-only expects an outdated DispatchMessage marker, and test-input-consumption-state.js omits require in its extracted VM context. Those failures are retained, not presented as emulator faults. Actual original control snapshots below independently establish input delivery and release. An attempted nonexistent test filename also remains in the raw preparation log; it did not execute a test.

An early preflight ran before the last large Quake II PAK upload completed and failed with FileNotFoundError before launching a guest. The complete 607-file transfer ledger then finished; the subsequent fresh preflight matched every pin before originals. Neither guest media nor guest configuration was edited.

Bounded serial startup

node tools/crash-sweep.js --apps=quake2_demo,gltron --seconds=20 \
  --headless-gl --modes=threads --jobs=1 --stuck-after=0 --no-build \
  --jsonl=<run>/actual3/sweep.jsonl
Original Startup result Personally inspected startup display
Quake II ok, 20.5 s Original Game/Multiplayer/Options/Video/Quit menu
GLTron ok, 20.5 s Original colored Game/Video/Audio/Quit menu

Both native WGL context creation and MakeCurrent returned nonzero. Raw startup logs, exit records, native loaded objects and captures are retained. Signature ok and similar wall times are not used to establish gameplay or performance. The sweep completed before either controlled gameplay run started.

Inspected gameplay and owning route

Followups used unchanged test/run.js, --app, --headless-gl, --threads, --png-canvas, --real-ticks, batch size 200000, bounded harness deadlines and ordinary keyboard down/up commands. Read-only control observations inspected host objects; no guest state, assets, bindings or configuration was forced.

Quake II: Return opened Game, then Return selected Easy. world01.png shows the textured first-person room, blaster and health 100. Paired W moved into the room (moved.png); paired Right changed the view (turned.png). Ctrl was paired but did not establish firing. Paired Enter did: the held firing-held.png shows weapon recoil and orange impact sparks compared with fired.png. No ammunition decrease is claimed for the blaster.

At the frozen final observation, all 12 key edges had been dequeued, routed and dispatched, known host queues were empty, and Return/Ctrl/Right/W physical and async down states were false. The owning window 0x10001 layer reported api: opengl, WebGLBackend, llvmpipe, and advancing surface write sequence 2621 to 3267. The actual displayed layer's native GL object supplied the provider strings, rather than a separate capability probe.

The scheduler backend was worker, but Quake II created zero auxiliary guest threads in this session. This proves gameplay under the real-worker backend selection, not concurrent guest CPU execution for this single-threaded observed path. No artificial guest thread was introduced.

GLTron: Return opened Game and Start Game; the arena initially displayed GAME IS PAUSED. Paired Return released the pause. gt-moving.png shows the cycle and growing trail. Paired A changed the minimap trail direction and arena view. A closer second ordinary round retained gt-pair-before.png, gt-pair-left.png and gt-pair-right.png: the player trail acquires distinct left and right corners after paired A/S, with corresponding views. The original scripts/config.lua maps player-one left/right to a/s. A natural collision/result sequence is retained; no player victory or survival claim is based on it. Old round-result text remains visible during the second round, and distant black outlines/trail shadows are not a parity judgment.

GLTron had a genuine guest Worker: scheduler slot 1, handle 0xe1000, Node worker thread ID 3, two completed execution slices, 10 synchronous and 20 asynchronous RPCs. It was chiefly parked; no sustained worker throughput or audio claim follows. The displayed window's layer itself reported OpenGL / WebGLBackend / llvmpipe, surface write sequence 838 to 1873. Final input observations counted all 24 edges dequeued/routed/dispatched, empty known host queues and released Return/Escape/A/S physical and async states. Guest post-queue and SDL-consumption snapshots remain unknown; visible turns supply the gameplay evidence.

Both controlled originals exited 0. Their terminal records are strictly serial: Quake II ended 10.735 seconds before GLTron launched. All 35 recovered PNG references were personally inspected in contact sheets; the key paired world captures were also inspected at original size. PNG CRC/IDAT decoding passed. The loaded-read audit matched 523 reads, including all four original module reads, with no unmatched reads, bridge errors or escaped trap/API failure. There was no demonstrated API/shader/CPU mismatch warranting an emulator repair or an old-behavior-rejecting production regression.

Recovery, cleanup and limits

Recovery verified 124 references / 119 unique hashes, including raw startup, crash/exit/scheduler/control logs, unique captures, both WASM modules and nine native binaries. Fresh postflight matched all 606 runtime and 162 native pins. validation.json, both retrieval-verification files, native objects, control snapshots and capture-reviews.json retain the checks. No code changed, so no new build was required for this documentation-only publication.

Remote cleanup found no owned live runtime process, closed control port 18123, no owned listening Unix socket, removed owned temporary directories, and no browser/profile created. After evidence recovery, explicitly delegated normal shutdown completed at 06:44:54.020Z, nonforced, error null. Box bx_jb7v82vs is archived, stoppedAt 06:44:54.003Z, no active expiry. No box ownership remains outstanding. Local free disk was above 2 GB at preparation and final verification; no foreign evidence was removed.

These are scoped Node native GL startup and gameplay observations. Browser WebGL/hardware GPU parity, audio, FPS, level completion, historical27 and broad crash freedom remain unqualified. Cooperative gameplay and Forsaken were not repeated. Next useful work is the already identified matched-state Quake III torch investigation or deeper original gameplay; neither menu-only gap remains.