Quake III original WebGL firing and same-camera torch — 2026-11-07

The original demo reaches Q3DM1 through its ordinary menus, fires, moves across the courtyard and supplies the requested same-camera torch before/after firing pair. Both full screenshots retain the weapon, ammo, health, crosshair and owning renderer evidence. No emulator defect was demonstrated and no emulator source changed. Evidence is in scratch/runs/20261107-q3-fire-torch/ (ignored).

Reviewed gameplay

retrieved/webgl/world-settled.png shows the readable textured courtyard with 100 rounds and health 123. The first balanced MouseLeft hold produces first-fire.png: 94 rounds, health 121, unchanged courtyard camera. The next forward capture shows 92 rounds; the first fire capture therefore does not represent the final settled ammunition count of that burst.

Two bounded W holds visibly advance toward the lit arch. At the torch, the reviewed pair is:

Capture Ammo Health Scene
retrieved/webgl/torch-before.png 92 117 Wall-mounted torch and flame, arch, weapon and crosshair
retrieved/webgl/torch-after-fire.png 82 115 Same camera, flame still visible, bullet impact near crosshair

Only a balanced MouseLeft down/up occurs between these two frames; there is no movement, mouse move or turn command. I personally inspected both originals. The arch edges, trim and crosshair remain aligned. Four fixed architecture rectangles, totaling 59,150 pixels, are also byte-identical at their original coordinates. same-camera-review.json records the rectangles and review. The changed flame/weapon pixels and HUD values are retained, not normalized. The health changes alone are not classified as bot damage or a renderer defect. These captures establish firing and the requested torch view pair; they do not establish muzzle-flash timing, FPS, audio, native/software parity or a full27 sweep.

The actual render Worker owns OpenGLHostBridge, WebGLBackend v1 and FixedFunctionGL. Endpoint 1 is GL/WebGL, HWND 65543, with a non-software context. The provider is WebGL 1.0 Chromium through ANGLE (Intel, Mesa Intel(R) UHD Graphics 620 (KBL GT2), OpenGL 4.6 (Core Profile) Mesa 23.2.1). At the final fire frame: 320,796 batches, zero software batches, write sequence 845, no endpoint failure/fallback and zero completed software draws. CPU and render Workers both instantiated module SHA-256 e41511fe861363fbe1639d80f20cbcdd7240f46937c93a801eaa256ca094c467. This is owning-backend evidence, not the toolbar selection.

Closure and controls

The existing authentic PK3 at test/binaries/candidates/quake-3-arena-demo-installer/installed/demoq3/pak0.pk3 was read in 90 small chunks, transferred, individually verified, assembled and independently reread against the full 46,853,694-byte SHA-256 e77abad2466f45a0a7ea018445528f9b95a0fe7789fa1abc1a7718bbf0754b08. Temporary chunks were removed. All 914 runtime pins pass before and after the run. No media/source archive or local duplicate media was created.

The exact input checkout is 11109c52fb51ed94f8bfe016d583b65139cb5499. runtime-pins.json, relevant-source.patch, new-source.patch and the small run-specific-inputs.patch record inputs; Git remains the source archive.

Five assembly controls, local serialized delayed-timeout/UNKNOWN release tests, explicit-success transport controls and the actual-browser trusted-release control pass. A generated browser page also proves that the absolute deadline closes the serialized queue before input or original launch. All controls finish before the original. No software/native original was rerun.

The initial browser preparation rejected index.html before app selection: I had generated browser pins with size instead of the collector's required bytes. The body SHA matched the runtime pin. The rejected harness is preserved under retrieved/webgl-prelaunch-rejected/; its selected-app and launch-input records are absent, and its normal exit, processes, ports and profile were verified closed. A secondary Puppeteer connection timed out during this stalled navigation; a later raw close found the browser already closed by the driver. No guest input was involved. The pins were corrected on the same box, with a local old-schema-rejecting check. Exclusive local log-name collisions and one local helper syntax error also occurred before effects and are preserved. There was exactly one actual original launch after complete closure.

The ordinary menu route uses Tab to Single Player, retains Q3DM1, tabs to Fight, then uses the difficulty menu's Fight button. It omits the prior exploratory Enter/Escape/click probes. One acknowledged Enter needed additional host runtime to draw the level menu; a host-only resume and capture resolved it without another guest key. No guest state, configuration, map or memory was forced.

Every one of the 23 balanced original inputs has trusted down/up events, a successful receipt and a personally reviewed reference image. The maximum reference age at input was 26.855 seconds, within the enforced 90-second limit. All 27 capture/input commands succeed; no original command is UNKNOWN or failed, and no input is replayed. The serialized finally-release gate is reused.

Deadline and terminal recovery

The 35-minute assignment runs from 07:37:35 to 08:12:35 UTC. Command admission closes at 08:03:30; browser cleanup is scheduled for 08:05, reserving over seven minutes for recovery and shutdown. The successful work finished early: ordinary quit closes the browser at 07:50:08.185Z with exit code zero and no held input.

All 204 inventoried remote artifacts (13,593,081 bytes) are hash-verified locally, plus the separately retrieved final terminal audit. This includes unique logs, captures, pre/postflight and the rejected prelaunch evidence. The unchanged module already exists locally and is identified by its full hash. Terminal checks find no runtime processes, all four HTTP/CDP ports closed, both explicit profiles absent, no temporary Puppeteer profiles or profile locks, physical/async keys false, queue zero and driver-held set empty.

The sole box bx_enztnm6f used one persisted UUID Idempotency-Key. After evidence recovery and terminal verification, explicitly delegated normal, nonforced shutdown completed at 07:50:56.954Z. State is archived; expiry is null (initial expiry was 08:13:31.450Z). No stopped box was resumed. Local watcher cleanup, free /home space above 2 GB, evidence validation and publication are recorded before result.json is written last.

The firing/torch task is complete. Further compatibility or parity work needs a separate scoped request; this run supports no broader release claim.