Tux Racer post-title input observation
Historical source last updated, 2026-10-30. Runtime and implementation claims below describe that dated source; see current app notes and source for subsequent changes.
The retained original-game run dequeued and dispatched Enter at batch21787, rendered the main menu, then host-delivered keyup/click/Enter without another observed dequeue before SIGXCPU at90CPU. This does not prove a renderer, coordinate, SDL callback, cached-message or scheduler defect. No causal behavior correction is justified from those observations alone. Practice/race remain unqualified. No new guest run, compilation, Boat allocation or resume occurred.
The opt-in --trace-main-scheduler observer now records returned/thrown main
slices and poll attribution, plus at most128 cumulative summaries independent
of the existing256 detail-record cap (default cadence1000 batches). Summaries
remain window-local and count import entries, not consumed MSGs or SDL events. Compare their deltas with --trace-input-dispatch dequeues:
- Runnable/returned slices rising with mainPolls unchanged: admitted guest execution has stopped calling the host poll; inspect guest execution next.
- Suspended decisions rising with no runs: the recorded reason/yield identifies the actual scheduler gate to investigate, rather than assuming restoration.
- Main/worker polls rising with no dequeues: polling continues without an event;
inspect host/cached queues using the existing frozen
input-statecontrol. - Worker polls are counted separately; they do not establish main consumption.
threwidentifies an exception even if an outer interleaved caller catches it.
No renderer, shared input plumbing, guest WAT, clock, CPU budget or accepted vertex/color source is changed. The observer reads no guest exports or queues. It is constructed only under the existing explicit flag. It preserves run return values, thrown objects and nested worker attribution. No summaries are emitted without that flag. Counts saturate at UINT32_MAX and declare saturation. 128 summaries bound the default coverage; later silence is not evidence of a stopped guest. A main-run without a finish is not proof of a hang if detail was already capped or output was interrupted.
Verification
node --max-old-space-size=512 test/test-tuxracer-main-menu-poll-observer.js
executes byte-exact sections of production initialization, host check_input,
scheduler admission and ordinary run/worker-import call sites with controlled
exports. This is a source-boundary test, not the complete runner or guest SDL
execution. Both the actual retained packet runner(blob30b8b056) and current-main
runner(blob579bca6f) pass; all four extracted spans are byte-identical, hashbound
in boundary receipts. The unmodified observer fails the persistent-summary
assertion. Removing the actual production poll hook fails attribution. Disabled
mode, exception/return identity, restoration and independent summary caps pass.
The retained packet source and current-main runner differ outside these spans; neither is silently substituted for the other. Module5a39225e1a8381ce1c4530d2ce7cd6311bd75576d10968c20e963baf033d41eb and all141 accepted canonical compiler inputs remain unchanged. Their earlier accepted vertex/color tests and original title/menu runtime are reused, without new gameplay or individual-call claims.
Next actual observation (coordinator assignment required)
Proposed command within the retained closure (not executed here):
node --max-old-space-size=512 test/run.js --app=tuxracer_original_061a --no-build --wasm=build/wine-assembly.wasm --gl-renderer=software --no-threads --max-seconds=100 --max-batches=1000000 --quiet-api --quiet-blocks --quiet-fs --png=capture/after.png --gl-census --input=17000:png:capture/title.png,18000:keydown:13,26000:keyup:13,30000:png:capture/before.png,31000:mousemove:320:260,32000:click:320:260,38000:png:capture/practice.png,40000:keydown:13,44000:keyup:13,49000:png:capture/course.png,51000:keydown:13,55000:keyup:13 --trace-input-dispatch --trace-main-scheduler --trace-from=21700 --trace-to=56000
This adds diagnostic observations to the unchanged reproducer. Preserve the first denial, raw output and final queue/scheduler state when available. The next decision must follow measured poll/run deltas, not another unchanged replay. Callback completion, guest instruction location and coordinates/state remain unknown until an actual guest observation distinguishes them.
Recovery and retained evidence
Local evidence: input-consumption. Original receipt and capture/runtime.stdout remain in the parent unchanged;
runtime stdoutSHA256e7fa512ea59ae8a9ea98b3ce4fd03349ea41e421943e196cb52770da24e6009b. Recover the retained runner from exactly test/run.js in the hashverified
standing packet packet.tar.gz
(archiveSHA2560c1b282fcb6f720c78090dce8156b2d321f9ee5839f414b09a068d033961629b). Recover current-main runner with git show 52de1079f2ce7f89e082b033e80d01231fc53d5f:test/run.js;
verify blob579bca6fb9b6ab117a403dead93eddd6ca0e2338. Both can be passed as the
test's first argument; an optional second argument selects an observer control.