Current original Tux Racer browser races, 2026-11-08

Original tuxracer_original_061a completes Practice → Frozen River on both current browser backends. Personally inspected native Race Over and subsequent Best result screens agree within each run:

Actual backend Race Over / matching Best result Reviewed race progression
WebGL 1 02:02.96, Herring 19, Score 0 / 000000 00:12.17 / 0 fish; 00:34.87 / 2; 00:51.26 / 7; 01:07.77 / 7; 01:27.12 / 12; 01:48.08 / 18; finish.
WAT software GL 01:50.50, Herring 15, Score 0 / 000000 00:19.11 / 1 fish; 00:34.02 / 2; 00:48.86 / 3; 01:07.12 / 6; 01:30.36 / 11; finish.

Evidence belongs to ignored scratch/runs/20261108-tuxracer-current-browser-backends/. This documentation-only change requires no emulator repair. The accepted historical browser finish and GLU power-of-two/alpha repair in tuxracer.md remain accepted. Earlier October 30 CLI title/loading investigations do not override the later browser success or these current completed races.

Original route and controls

Each fresh browser follows ordinary Enter → Main Menu → P → Practice / Frozen River → Enter → native race. Native Up, Left and Right are finite physical key holds with independently armed release timers. Up uses 2000 ms; Left and Right use 750 ms; menu keys use 100 ms. Captures include the held interval and released state. Up frames show paddling/downhill progression; steering changes Tux's bank/orientation and course position. After each hold, the bound capture reports all four host arrow states zero. These observations establish ordinary race input and release, without claiming isolated physics conformance or equating passive counters with guest keyboard polls.

Both routes show textured snow/ice and mountains, Tux, flags/herring, recognizable trees with transparent silhouettes, and readable changing HUDs. The reviewed finish frames include a black rectangular backing behind the speed readout. This bounded review is not a claim of flawless rendering. Different viewpoints, race paths and displayed times do not establish pixel parity, competitive clock accuracy or FPS. Audio quality and full27 remain open.

Before allocation, the actual request-safe.py → numbered publication → consumer → exact ACK path passed generated positive requests with real 0, 1000, 2000 and 10000 ms waits, the exact upcoming Enter/P/Escape/arrow keys, 100/750/2000 ms holds, and qualified optional Up+steering combinations. Only the single-key routes were used in the original. No mouse action was needed. Oversized waits, post-UNKNOWN input and insufficient immutable-deadline budget publish no command. All eight existing MW3 deadline regressions pass; the independent-UP control passes slow observation and injected UNKNOWN.

Preallocation qualification first exposed that the reused run-local helper lacked p. That generated consumer stopped; no original existed yet. Adding the ordinary P mapping to the owned helpers and using a new generated stream passed. A second generated attempt rejected reused local IDs before publication; the final qualification uses distinct IDs. These failures are retained. The runtime source and guest bindings were unchanged. Original commands have no UNKNOWN, replay or refused input. Capture/admission binds the actual process, guest Worker/module, owning render endpoint/Worker/module, epoch and canvas. Every dependent scene was personally inspected before further native input.

Actual backend and build identity

WebGL's unique live WebGLBackend belongs to the owning render Worker and GL endpoint 1. It owns a non-lost WebGLRenderingContext, WebGL 1, with ANGLE reporting Intel UHD Graphics 620 / Mesa 23.2.1, OpenGL 4.6. The main guest and renderer CPU exports both return gl_sw_enabled() == 0.

Software GL's unique live OpenGLHostBridge reports backend software, software-consumer true, context 1 for HWND 65537, current owner 0/context 1, and a software backend. Its export memory is the renderer instance's memory. The main guest (endpoint 1 / slot 0) and renderer CPU exports both return gl_sw_enabled() == 1; a race observation records 709 guest software presents. The audio thread's separate CPU reports zero, and legacy endpoint 2 uses WebGL for presentation. Neither is substituted for the owning software GL context.

Both browsers run Chrome 151.0.7922.108, V8 15.1.206.16, with native viewport and browser guest threads enabled. Live debugger-retrieved guest and renderer modules match the staged 2,182,500-byte standard module: f9650607181301b5d2a9990420f437bd9f1ca10d40ac73067339f07499d53121. The compatibility module is also retained by hash.

Runtime source Git is 7ce76eae9ed98a9cac0b23f609c2fbc5857e4f3f, with empty relevant tracked binary diff and the retained untracked new-lib_wininit-boot.js.patch dependency. The exact pinned closure is in runtime-pins.json. fixture-identity.json records all 158 canonical original media paths/sizes/SHA-256 values and the 157 browser mounts; the original inventory hash remains 2c2d29a9dc7e5f9bc9e972b408fd83dbb17e8ca48a6aed09ffd375830284d8d8. No media or source-tree archive was created. Owned helpers and this note are preserved as new-file patches; Git plus the recorded patches is source recovery.

Two extra historical static test invocations did not provide new acceptance: the registration test stops at its stale test/run.js SHA pin, before asset checks; the staging test was invoked without its required argument. It was not rerun because its recipe moves the shared fixture and writes outside this run. The complete canonical hash/mount preflight and actual served browser closure checks provide this task's original-media validation.

Resource closeout

One fresh UUID-created box, bx_dsjw226m, was allocated at 01:08:27.557 UTC, initial expiry 01:33:27.397 UTC. The detached supervisor, 25-minute watchdog and normal-stop delegate were observable before allocation; the immutable local deadline is 01:33:27.200 UTC.

WebGL Chrome exited normally at 01:13:16.681 UTC. Its controller/browser PIDs, two ports and profile were proved absent before software GL launched at 01:13:30.768 UTC. Software Chrome exited normally at 01:17:22.623 UTC. Both controllers and browsers are gone, all four ports are closed, both profiles and writer locks are absent, and all 801 staged pins match after shutdown. All 24 original commands (13 WebGL / 11 software, including quits) have exact successful ACKs; all 16 key holds have independent release receipts. All 40 retrieved PNGs pass CRC/IDAT validation, bound capture hashes match, and the 697 successful served reads match pinned bytes. The final 324 artifact references / 290 unique contents, 21,289,617 reference bytes, were retrieved and independently size/SHA-256 verified before shutdown.

Delegated normal stop stop_b87ce352ed3b906a completed at 01:18:21.315 UTC, forced=false, error=null. Final box state is archived, current expiry null. Both lifecycle delegates exit zero; the supervisor reports no remaining children. The independent verification helper also exited. No box remains running for this task.

No guest code, state, config, save or clock was written by the harness; no private game handler or fabricated result was used. No RAM gate, stopped-box resume/delete, extra box, archive or public deployment occurred. Sampled local and remote free disk remained above 5 GB.

Next work is separately assigned broader course/browser/visual coverage. Full27, full parity, competitive timing, sustained FPS and audio remain open. This single-task worker releases its resources and stops.

The corpus manifest currently has no Tux Racer candidate entry. Per the dashboard workflow, the reviewed outcome remains in this note and run-local completion.json; it is not attached to another game's dashboard card.