Alice native fault latch: bounded repair, 2026-11-03

STOP: explicit fault termination still fails. No runtime change is published. The candidate retains the earliest real main-thread UD2 fault and leaves it inspectable, but its input cleanup assumes the CLI input queue is non-null. The first explicit termination throws Cannot set properties of null (setting 'length'). The two-build bound is exhausted. The candidate and complete native evidence remain in ignored scratch/runs/20261103-alice-native-fault-latch-repair/; all owned production files were restored byte-for-byte to the input revision before publication.

Input was 9a6516bbacd02860fdb5e1991b679904a99544b3, committed blobs only. Jedi's concurrent decoder/ALU/string WIP was excluded. This was one task, one fresh boat, no subworkers and no successor. Alice was never launched. productionReady=false, usableForFooting=false.

Failure and retained candidate

The preceding held-phase investigation preserved actual PID14844's UD2 at 01022a74: RuntimeError: unreachable bypassed the success-only observer and reached process.exit(1). This task read and sealed that recovered runtime log. Its retained cli-adapter.cjs, held-gate.cjs, main breakpoint fixture and cooperative breakpoint fixture were reused byte-identically, without relabeling receipt tests as native results. The accepted original-PE relocation fixture/result was also sealed unchanged; it was not rerun or claimed as new qualification.

The candidate adds a host native-fault-latch.js, CLI control/error boundaries, and a four-line cooperative scheduler change. Explicit arming requires the discovered process UUID/PID/generation and frozen cooperative control; Worker mode is refused. The native error is installed before diagnostic reads. Main calls park before post-run work; a cooperative catch returns before crash reporting, retirement or another sibling. Original error/stack, owning instance, registers, guest stack, thread census, batch and pending credits are retained. There is no fault-resume operation and no conversion to halt5 success.

The candidate protocol is fault-status, fault-arm, fault-inspect and fault-terminate. Only status, inspection, ping and snapshot remain available after a fault. Eval, input, step, unfreeze and ordinary quit are rejected. Termination is intended to release host input, persist the original failure, reject the pending step with that error, and exit1 without another native run. That last path is not qualified: releaseInput() clears renderer maps, then executes inputQueue.length = 0 when inputQueue === null, before file persistence or scheduler release. The complete candidate is retained under candidate/, with candidate.patch and candidate-pins.json.

Actual native results and first remaining prerequisite

Fixture Result
Unchanged main breakpoint fixture PASS: actual halt5, four-credit hold, repeated reads, explicit once-only breakpoint release and original capture-error retention.
Unchanged cooperative breakpoint fixture PASS: actual target-thread halt5, sibling freeze and four-credit release.
Main UD2 candidate FAIL at explicit termination, after the retained-fault checks below passed.
Cooperative UD2; unarmed main/thread failure Not run: the suite stopped at the first remaining failure.

At 17:08:49.157Z, actual CLI PID10821, process UUID aaef5701-dd90-4e01-a211-cd532348a591, generation1, retained the native main thread1 fault at 01022a74, path ordinary. Original error remained RuntimeError: unreachable, with WASM frames536 and12483. Registers included ESP01024a6c, EAX0/EBX0, last_run_halt=0; a128-byte guest stack was copied. Batch0, all four pending credits and the original previous-batch flag stayed held. This was a failure, not a breakpoint.

Three delayed inspection passes preserved status, registers, thread census and credits. Actual instruction bytes remained 0f0b. A disposable sibling's native absolute-memory increment remained unexecuted: its four-byte sentinel stayed zero. Additional step/eval/input/unfreeze/quit and breakpoint-release requests were rejected. Inspection and termination rejected wrong PID, process UUID or generation; inspection rejected a wrong fault token.

A real unreadable span produced a secondary inspection error. A deliberately closed HTTP connection did not replace the original native error. Both attempted termination paths instead produced the same null-input-queue TypeError; there is no tested disk-persistence-error or successful key-release claim. Valid read-only recovery after those failures still returned the original native error, 0f0b, unchanged registers and four credits. These commands and their actual responses are in main-ud2-commands.json and remaining-prerequisite.json. No later cooperative/unarmed case was admitted.

The failed fixture's SIGTERM did not release the parked gate. After recovering the live read-only fault evidence, only its own PID10821 was OS-killed to retire the failed helper, without guest advance. This is recorded separately from the failed explicit termination in own-failed-helper-retirement.json. It does not qualify production termination. No invalid instruction was resumed.

Build, input closure and recovery

Before allocation, all15,792 committed blobs were read and size/full-SHA sealed, all19 ignored PE/input pins passed, and191 exact relative import edges were checked, including launch-plan.json. The FINAL polling bundle full SHA b7aefa122818d5c353e3df59ef03b21081620b342f149530f8abd0cb1ae29410 was verified:16 protocol checks, four wrapper checks and the delayed1.5-second single-submission zero-wait case passed. No pending command was resubmitted. The remote source closure was streamed directly from committed Git blobs, with only the explicitly owned candidate overlay. All source/input pins and exact imports passed again before build.

Build attempt1 stopped before compilation at the required whole-file callback ownership seal. The four-line scheduler delta was reviewed and its seal updated, retaining ownership as pending. Attempt2 passed the locked full flock /tmp/wine-build.lock bash tools/build.sh, keyboard-message-lparam, frozen-stdin-control, CLI DI-mouse and thread-manager suites without skips. The build retains its explicit absent-Git-HEAD API comparison and offline Big5 check. Native fixture budget: three processes used, maximum30; suite deadlines300seconds. Two build attempts total; no further build or native retry.

The2089591-byte WASM SHA256 is 0a58fd7b9ac82c5303048776f442b15e72528b55ddf85564e0df92fcb1aa384d. Both modules, candidate/source hashes, all original inputs, commands, errors, build/focused logs and actual fixture results were recovered locally. All 5939 payloads passed size/full-SHA checks before normal STOP. Archive 117332948bytes, SHA256 f14771531955590bfb752548c970343a2f5aabdd0b57cc4f059a5ea28904685a. See RECOVERY.md, archive-verification.json, recovered payloads.json and SHA256SUMS. Evidence remains ignored, not committed.

Own boat bx_74kksjy9, finiteTTL3600, original expiry 2026-11-03T18:02:12.159Z. Identity, expiry and original-task delegated normal STOP were recorded before use. STOP stop_df57295b0bd2ad32 completed at 17:11:40.727Z, error:null, forced:false. Final state STOPPED, completed available snapshot verified17:11:40.653Z; own helpers absent and port8123 closed. Custody returns to the coordinator. No foreign/predecessor boat or process, Claude, HeroesII, heldCACA003B or public deployment was touched.

Next prerequisite is concrete: handle absent CLI input queues during host-only fault cleanup, then qualify successful persisted termination, cooperative native fault ownership and unchanged unarmed failures. This task does not authorize that successor. Callback/Player lifetime, actual retirement/reuse and authentic supported/unsupported ground comparisons remain later work. There were no Alice guest writes, injected engine calls, opening/route execution or gameplay claims. AUTHOR RELEASE; single-task stop, no successor or reuse.