Native Quit critical-section retirement sweep
No critical-section retirement defect was reproduced on current main. The retained LMarbles original remains a proved native Quit status 0 with an unexplained held-main-section diagnostic. This task adds native regression coverage; it makes no runtime repair and launches no original game.
Evidence: scratch/runs/20261118-native-quit-critical-section-retirement-crash-sweep/.
The final result.json records publication, exact source/module identities,
foreign diffs, checks and process cleanup. Git plus those diffs is the source
recovery recipe; the generated module and logs are retained separately.
Historical comparison and actual observation
Dispatch main was 9191d4afbcd7eb6d84139fef87502f9273ae8d2f. The prior run used
source a00eb35bebb621ebb3f2b42b6af18f1bd874b9f6 and module SHA-256
83703369cc49c3e16077a4abd43def27262487f83dd52583d1b433adf7095f5f.
Its original/native.stdout.log records, in order:
- Worker 2 exits with yield 2; waveOut closes; process Exit code 0.
- One registered section is still held by main, lock 0, recursion 1.
- Final thread reporting still lists worker 1 active.
- Worker 1 reports
slice failed: Callback owner retired, then an unexpectedsliceDonemessage. Native stderr is empty. The original lifecycle receipt reports native exit 0 and no known surviving owned process or listener.
src/09a-handlers3-sync.wat, including $cs_release_owned, is byte-identical
between that pin and dispatch. The ThreadManager retirement fanout and CS
release hook are also unchanged; the intervening scheduler changes concern
Worker in-flight ownership, clock boot holds and D3DIM parking. The exact
historical-to-dispatch diff and current foreign WIP are retained. This is not
evidence that a later CS repair fixed the historical warning.
The held-section report in test/run.js reads the registry before final
Worker shutdown. It reports ownership, not a Wasm trap or host exit failure.
Main process termination does not call the worker-thread retirement hook for
main's section. Shared memory and endpoint destruction subsequently discard
that process's state; releasing main's locks merely to silence this report
would not establish a repair.
The requested win16-native-quit-post-exit-crash-repair-20261115.md is absent
from this checkout and its Git path history. The existing
win16-native-quit-exit-20261115.md was read instead. That investigation proves
a different defect: a real far callback resumed after DOS3Call termination and
the host exited 1. Its real native regression remains part of this sweep's
focused checks. That diagnosis cannot be transferred to LMarbles from a
warning string.
New native coverage
Run node test/test-native-quit-critical-section-retirement.cjs. It compiles
the full canonical WAT closure and generates its own small PE and x86 code.
It uses production host imports, ThreadManager, GuestMainOwner and real Node
Workers, with no retained-run dependency or substituted retirement functions.
The tested module is SHA-256
fef4be45c6072558a6f94d12b8b163523abd2ddae3651390b797626e4033e0b7
(2,408,726 bytes). Source and host-library hashes are in tested-source.json.
The locked build emits SHA-256
7d347a6a631ab77335a14935a533feb896be71aa2508fc5aa19b1e5df8f096ce.
Removing only its 37-byte region-layout custom section produces the tested
module byte for byte (canonical-module-equivalence.json).
All 24 controls pass:
- Local and Worker main owners execute EnterCriticalSection followed by ExitProcess(0), a seeded root callback return, or a deliberate UD2 trap. Clean ExitProcess leaves its post-exit marker untouched and reports exactly one exit 0. Root return reports no process exit. Both preserve the same held main-section state as the warning. The genuine trap remains a failure with no process-exit notification. Main-return coverage is the seeded callback return contract, not an original executable's CRT shutdown sequence.
- On cooperative and native Worker schedulers, an owner recursively acquires an aligned or packed section and sleeps. A second real guest thread parks on it. ThreadProc return, ExitThread(37), UD2, and TerminateThread(23) each retire ownership once; the actual parked waiter then acquires, writes its progress marker and releases. Exact exit status and cleanup completion are asserted; a duplicate release cannot change the section.
- For aligned and packed sections, an actual pending
_runWorkerThreadslice races GuestThreadHost STOP. The scheduler recordsworker-error, releases the section once, and does not resume the stopped waiter. A borrowed callback endpoint rejects late admission. Both Workers are joined and the owner endpoint has no remaining native Worker.
The STOP controls naturally reproduce both Callback owner retired and the
late sliceDone diagnostic while the test process exits 0 and the supervisor
confirms no leaked process. This supplies a concrete shutdown-race control
for the retained wording. It does not prove the original game's exact
interleaving or explain why its main section was held.
The first three development attempts exposed fixture assumptions, not runtime repairs: Win32 ExitProcess does not set Win16's yield reason 2; cooperative scheduling clears yield 9 after recording a park; and late admission must be tested on the retained link because the host removes its link on STOP. These logs and their finite cleanup receipts are retained, separately from the passing final control.
Validation and limits
The focused checks cover critical-section struct semantics, concurrent packed and aligned ownership, exit callback cleanup, ThreadProc return/fault status, and the native Win16 Quit regression. All five focused suites and the canonical locked build pass. Exact commands, exits and logs are retained in the run.
Every new native fixture uses finally-owned Worker shutdown. An independent
parent process kills a stuck fixture after 120 seconds, including synchronous
compilation; native-supervise.cjs supplies a separate process-group deadline
and cleanup receipt. No Wine, boxes, worker agents, deployment, media copies
or held investigations were used. Disk observations stay above 1.5 GB; there
is no RAM gate. Box ID, expiry and stop receipt are null (not created).
No production source changed and no actual repair was warranted by these observations. Consequently there is no old-fail/new-pass repair claim and no unchanged original retry. The precise origin of LMarbles' held-main-section state remains unproved; the new result establishes current retirement and shutdown-race behavior within the authored native scope.