Original graphics titles after the native GL shader repair

2026-11-07. Run: scratch/runs/20261107-graphics-fixed-gl-originals/. Quake II, Quake III and GLTron now have personally inspected original gameplay through the supported Node native GL bridge containing aaaf6e370. Forsaken has inspected gameplay through its registered software D3D route. No new emulator defect was demonstrated and no production source was changed here. Earlier original captures from before the shader repair remain pre-repair evidence; they do not become GL passes through this result.

Tested identity and recovery

The single fresh box was bx_pfx5b88a, created 04:32:02.917Z with its UUID Idempotency-Key persisted in boat-create-key.json. No predecessor box was resumed. Preparation waited for Lords2's source/build release at 04:34:44Z. The tested main commit was ce0330c444ea43742259624db5e9cd0240b2f156, which includes the published WinExec repair 9b1da000e as well as the GL repair. runtime-pins.json records every staged path, size and SHA-256. The relevant tracked source diff is empty (relevant-source.patch); the then-untracked lib/wininit-boot.js is retained as new-lib_wininit-boot.js.patch. Git plus these patches is the source recovery recipe; no source/media archive was made.

Tested input SHA-256
Standard WASM, executed in all 14 original runs e41511fe861363fbe1639d80f20cbcdd7240f46937c93a801eaa256ca094c467
Compat WASM, recovered but not executed 04acbb89f66ad9814836b5e2c34de659b518f4e1994f50c4ee8069c991f02f72
lib/gpu-backend.js 7913166b52ad34b5a7657aefec747aa66bf4b8f4c1cd3e50fa195a7956274ad7
lib/headless-gl.js 7e79d93edd690f4c2338874bcde1961360547bfd031d4b58827484e7b03689fb
Remote Node v24.22.1 e32e2d16081cf644f88392934288086e7662529d0d43277c34f7aad2207cfafb

Registered originals were complete and matched the recorded fixture hashes: Quake II 67 paths, Quake III 307, GLTron 92, Forsaken 873. The 1,355 predecessor fixture/shared inputs matched, with fonts pinned separately. original-fixture-preparation.json and runtime-pins.json preserve the exact paths and hashes without duplicating game media. A local preparation typo (test/binaries/fonts instead of recorded fonts/) was corrected before any guest ran; see preparation-correction.json.

The native environment used @node-3d/glfw 7.4.1 and @node-3d/webgl 6.0.3, display :0, Mesa llvmpipe LLVM 20.1.2, WebGL 1.0 / GLSL 4.50. The 162 native dependency files are pinned in native-pins.json; nine native libraries were also recovered individually. The red-pixel context probe returned [255,0,0,255], error zero and zero remaining contexts after destruction. Fresh native fixed-function/shader and WGL regressions passed before originals (retrieved/fix/native-test-after2.json, native-wgl-after2.json).

Original startup sweep

The supported command was node tools/crash-sweep.js --apps=quake2_demo,quake3_demo,gltron,forsaken_demo --seconds=20 --headless-gl --modes=coop,threads --jobs=1 --stuck-after=0 --no-build --jsonl=.../actual3/sweep.jsonl. All eight serial guests ended with signature ok, without an escaped trap or unimplemented API. WGL creation and make-current returned success for the six GL guests; Forsaken created no native GL context.

Original Coop capture at about 20 seconds Threads capture at about 20 seconds
Quake II Original main menu Original main menu
Quake III 512×512 logo texture; not display evidence Same texture; not display evidence
GLTron Black capture at this budget Original colored main menu
Forsaken, software D3D Original main menu Near-white capture at this budget

The existing sweep PNG selector chose Quake III's colorful DX texture slot 35. Its logo cannot establish visible startup. Followups used the existing --png-canvas / control-server compositor capture instead. The GLTron and Forsaken equal-time differences are observations, not scheduler diagnoses.

Ordinary input and actual game progress

Controlled followups used the unchanged test/run.js, supported --headless-gl, --png-canvas, --real-ticks, batch size 200000 and bounded ordinary control inputs. All six ended with code zero. Each retrieved/progress-*/ directory contains its exact launch arguments, runtime log, input/capture evidence, loaded module hashes and terminal status.

Title / mode Personally inspected progress and evidence
Quake II / coop Enter → Game → easy → first-person room with weapon/HUD. Paired W and Right inputs changed the view: world.png, moved.png.
Quake III / coop Actual main menu → choose level → difficulty → Q3DM7 loading → first-person world with bots, weapon and health 122: world02.png. W down/up completed; final.png shows another view and health 116. The subsequent Left down had no matching up because the 180-second harness guard ended the run; do not claim a fully balanced turn sequence.
GLTron / coop Return through menu/pause → moving cycle/trail; paired A changed direction/minimap: gltron-unpause.png, gltron-left.png. Later S capture shows player 1 crashed/winner 3; Escape returned to menu. This is game progress, not a survival or level-completion claim.
Forsaken / coop Logo/menu → Start/Volcano → briefing → rendered world with lives 5, shield/hull 128, Pulsar 2000. Fire produced visible green pulses and ammo 1988; Up changed the view: world.png, forsaken-fire.png, forsaken-move.png. Software route.
Quake III / threads Main menu → Q3DM1 Arena Gate → rendered world health 124; Major killed player, mouse input respawned and paired W moved the view: world01.png, world02.png, q3t-forward.png. Black shapes near the torch remain visually questionable and need targeted rendering diagnostics.
Forsaken / threads Original menus/briefing → Volcano world; paired fire produced green pulses/ammo 1992 and Up changed view: ft-continue.png, ft-fire.png, ft-look.png. Software route.

Quake II threads and GLTron threads reached inspected menus only; gameplay in those modes was not exercised. Black GLTron trail shadows follow the original configuration; distant visual artifacts and Quake III torch shapes preclude a full visual-parity claim. No trap/API/rendering defect was causally established that justified another emulator repair. In particular, a black/white short capture alone did not justify changing scheduler semantics or guest files.

All 72 recovered PNG references (66 unique images) were inspected; hashes and contact-sheet indexes are in capture-index.json / capture-reviews.json. The coordinator additionally inspected Quake II movement, Q3DM7 gameplay, GLTron trail/turn and Forsaken view change (coordinator-visual-review.json). This establishes Node GL original startup/gameplay only: not browser WebGL, historical 27-title qualification, audio qualification, FPS, or game completion.

Audit, shutdown and next work

The remote postflight checked 1,784 runtime and 162 native pins unchanged. The recovered loaded-read audit independently matches 1,781 reads, including 14 standard WASM reads, with no observed bridge errors or escaped traps/API failures (identity-audit.json, raw loaded-modules.jsonl and bridge.jsonl). Recovery verifies 268 artifact references / 248 unique hashes, including both modules, native libraries, unique captures and losslessly compressed logs. The detailed postflight filename was reused by inventory; the raw traces and audit-remote-result.json retain the remote audit, and identity-audit.json reconstructs it locally.

The startup sweep used jobs=1. Q3 coop started 0.107 seconds before the previous Q2 supervisor wrote its terminal record, after the Q2 quit request; this small launch/teardown overlap is recorded rather than described as strict terminal serialization. Subsequent launches explicitly required previous terminal records. There was no intentionally concurrent heavy title workload.

Before stopping, retrieved/cleanup.json confirmed no owned live process, closed control port 18123, no owned Unix socket, removed temporary directories and no browser/profile created. After all unique evidence/modules were recovered, delegated normal shutdown completed at 04:50:14.095Z: box archived, stoppedAt 04:50:14.074Z, forced false, error null, no active expiry (stop-confirmation.json). No box ownership remains outstanding.

Next useful work is a matched-state Quake III torch alpha/blend investigation, and optional Quake II / GLTron threads gameplay. Neither is a proved emulator blocker from this run. No public deployment was performed.