AoWII audio wait completion and gameplay, 2026-11-17

Evidence: scratch/runs/20261117-aow2-audio-wait-continuation-gameplay/.

The two Mss32 audio EIP0 exits are caused by duplicate completion of an ordinary parked WaitForSingleObject frame. This is not a ThreadProc return. No audio disable, guessed return fix or crash suppression is used. The fix is published on main as 81ad6c8022229f25aa35b2baf0f0ee8bbeac9df7. Two ordinary army moves and menu Quit/native exit remain unproved.

Offline examination of the previous first-fault stacks found valid DLL wait return addresses 24 bytes below terminal ESP, with the original event and 15/5 millisecond timeout arguments intact. One completion owns 12 bytes. Running the actual shipped Worker completion handler against those stack fixtures reproduced terminal EIP0/ESP/EAX after two completions. That fixture alone did not establish what occurred in the original.

One newly authorized registered original retained a passive, bounded history of secondary Worker wait requests and replies. The actual native path proves the missing transfer for both slots:

Slot First completion Duplicate completion Following slice
2 880 resumes 0x13439f5 881 sets EIP0 882: zero blocks, ESP 0x7eabfff8
1 2123 resumes 0x13245c5 2124 sets EIP0 2125: zero blocks, ESP 0x7ebbfff0

The history contains two slice replies for the same parked frame before the two completion requests. No guest slice separates the first completion reply from the duplicate request. The duplicate reads the saved-register/local word above the completed frame, which is zero. Both terminal replies have EAX 0x102, yield0, threadReturned false and no reported WASM trap. The evidence therefore identifies an emulator continuation defect, rather than an ordinary native worker return or a WASM trap.

ThreadManager._runWorkerThread previously marked only the awaited slice as in flight. Direct nested pumps could enter the method concurrently, and the flag was cleared before asynchronous result/wait-completion handling ended. The fix synchronously admits one invocation and holds that admission through the entire operation, releasing it in the outer finally. It changes neither the Win32 wait result nor its stack cleanup. A duplicate invocation does not issue a slice or completion. Existing owner, callback, lifetime and release checks remain in force.

The focused regression drives the production Worker completion handler while interleaving nested pump admission at slice and completion-acknowledgement boundaries. It checks immediate and already-parked waits, exact stack/EIP/EAX, unsatisfied waits and completion transport failure. Original HEAD fails with two slices instead of one; the corrected scheduler passes all 59 checks. Three older VM fixtures now provide the browser-input, RenderWorker and RPC dependencies their production imports require; their assertions are unchanged. The canonical flock /tmp/wine-build.lock bash tools/build.sh passed. Wait reservation, bootstrap, wave-callback and wait-identity tests passed. The broader cross-thread test has the same two failures (24/26) with the original and corrected manager; the separate scoped render-yield test also fails identically on both. These existing limitations are retained, not reported as passes. Initial shared-build failures during concurrent Sims WAT editing are retained; the final locked build succeeded after those corrections.

The first original reached the personally reviewed default Julia opening dialog. The independent disk watchdog stopped it before a world afterimage, army movement or Quit proof. The Click-to-begin input was attempted; no native consumption claim is inferred from the interrupted screenshot. Actual Chrome0, driver1, watchdog0 and supervisor143 waits with both EOFs are retained; owned scopes and children are empty. Custody was retrieved and hash-verified before removing its disposable /dev/shm profile. No terminal was reopened. Its final offline response assessment passes all 793 required bodies.

The published 44115957 supervisor and existing direct188/package191, source/module/media proof are reused. The original standard module is 87ef39648a61ed6f54ca4ab8de1926de02c651eb2d99655015dc7e5de5407009. Git plus relevant patches is the source archive. There are no source archives, media copies, repeated lifecycle suites, boxes or deployments.

After publishing the distinct proved fix, the one conditionally authorized original served the same known source/module/media, replacing only the thread-manager body with the published hash. The required vclx50.bpl request was canceled with net::ERR_ABORTED during DLL loading. The sticky collector fault is retained, including subsequent failures after collector closure. Its cause is UNKNOWN; this does not prove a new audio/emulator defect. Only registered Launch was delivered, with no game input or replay. The unchanged driver captured two loading images before the fault was recognized; no menu, Mss32 thread progress or post-fix native audio validation is claimed.

Local custody cancellation ended that run promptly after recognition. Actual Chrome0, driver124, watchdog0 and supervisor1 waits with both EOFs and empty owned scopes/children are retained. The watchdog's independent deadline and actual disk floor were armed and did not fire. Its unique custody was retrieved and hash-verified before removing the disposable profile. The keeper's actual wait on each supervisor is recorded; an external wait on the detached keeper itself was not captured and remains UNKNOWN. Process absence is not substituted for that missing wait. No box was created: ID null, state NOT_CREATED, expiry null, stop receipt NOT_APPLICABLE.

No further original is authorized by this task. Next work must first explain the required DLL-body cancellation offline, then obtain a newly authorized native run for live audio progress, two distinct Party destinations with remaining-point changes, and ordinary menu Quit/native exit. Both terminals remain closed. The final result indexes the exact pins, relevant patches, selected result-named whole images, tests and missing proof.