TuxRacer retained cooperative mutex handoff acceptance — 2026-10-30
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 exact retained candidate passes independent mutex-boundary tests. One actual original-game execution consumed native Enter and rendered the main menu. Practice and race were not reached; no gameplay screenshot is claimed. This qualifies the historical retained library only, not current main.
Exact inputs and independent tests
Parent 65d3f211cb2394fc7511e8057fe09fb1d0877618 and its five requested candidate
artifacts were independently read through exact Git paths and checked against
the retained files. The previous task's remote ref was independently verified
using the coordinator's exact trusted SSH command. The report in that parent
was also read. The actual retained library is 227988 bytes, SHA256
92a5671ad712d9ac943559a009496000e28a7fe98f34e66c2f453ba8fc9bac3d. The original recovered from the exact closed archive is pinned by the tests to
f4ba190b2dbf5e2cf2dac82398577bedfb13aa26d332fb7480b4d9d7e4ea2f78. Neither is silently identified as the historical a998bf36 Git library.
acceptance.cjs loads these actual complete libraries and exercises their real
releaseMutex, waitSingle, checkMainYield and _completeWait methods. The fixture supplies main register accessors and a return-address reader;
ordinary wait completion itself is not stubbed. The tests cover final release,
recursive acquisition/release, wrong owner, invalid handle, worker backend,
unrelated sync records, different main wait, multiwait pointer and stack shape. Main's EAX/EIP/ESP/yield and completion count are checked. No interpreter is run
recursively. The original control reproduces same-slice reacquisition; removing
only the candidate's handoff call rejects at owner0 versus expected owner1,
after the exact source pin, without any title prerequisite.
One original rendering/input execution
App tuxracer_original_061a, Node v24.22.1, software GL and cooperative guest
threads used the unchanged 692809-byte runner SHA256
f96e9bb924700ee571e2904e097f80742ff55638399898ae1dc32454af05b02d
and 1881806-byte module SHA256
5a39225e1a8381ce1c4530d2ce7cd6311bd75576d10968c20e963baf033d41eb. Canonical lib/app-files.validateLocalManifestSync validation enumerated all157
actual files before execution, including a check for comma-bearing names. The exact 512MiB-heap Node/preload startup succeeded with unlimited DATA/AS,
CPU60 seconds and generated-file limit12MiB. The independent tests used256MiB
heap. These heap values are not total-process memory limits.
The revised controller accepts a personally inspected rendered title instead
of demanding the obsolete frozen-title hash. Its optional JPEG helper then
failed because skia-canvas was not installed. That original error is retained. The original PNG remained valid. During this same live execution, the worker
retrieved and inspected it: Tux Racer and PRESS ANY KEY TO START were visible. Through supported API execution, the worker reopened the existing runner's
stdin pipe at /proc/9796/fd/0 and sent one ordinary keydown:13, then
keyup:13, then an ordinary PNG capture. No guest state was edited.
At batches14420/14451 the trace records both renderer dequeues, focus routing
to hwnd0x10001, and both DispatchMessageA calls. The second inspected frame at
batch14468 displays Enter an event, Practice, Credits and Quit. This is actual
main-menu progression, not a claim inferred from command acknowledgements. The original PNG hashes and JPEG derivative recipes are in media.json.
The runtime ended with SIGKILL at11:37:33.071Z, after60.430 seconds wall time. It inherited CPU60 hard, and the supervisor recorded no wall/RSS/capture reason; the exact kill source was not separately observed. Peak group RSS337858560 bytes was below the2GiB enforcement threshold. The later inspected-menu Practice-click attempt failed opening the vanished stdin fd, before writing any click. It was not retried. Thus precisely two native key commands were sent, consumed and dispatched; zero Practice clicks were sent. No race/gameplay or final post-input yield/thread snapshot is available. The retained title snapshot contains real thread/yield/counter state; it is not relabeled as terminal state. No second execution, resume, forced unlock, audio suppression or synthetic callback was used. The controller dependency consumed time inside the one-run budget, so this is a bounded partial runtime result rather than race acceptance.