FINAL TIMEBOX SNAPSHOT — 2026-10-08 20:09 UTC. Evidence checkpoint verified: https://github.com/CodexCoder21Organization/kompile-core/commit/ee062059b21987736ef52f983925cf1df7132185. PR remains open/draft; four test shards are in progress and Fat-JAR passed. All owned local workers and watcher have stopped. The strict actual-policy comparison and timeout/headroom work remain incomplete. Claim is being released; no block on a person is recorded.

Resolve the final CI failure and mark the kompile-core PR ready

Re-verified 2026-10-08 20:03 UTC. The PR is open and draft: https://github.com/CodexCoder21Organization/kompile-core/pull/385. Final source is https://github.com/CodexCoder21Organization/kompile-core/commit/3dde5d09a8ee2e22f31320cd485d68bae9701e29. Do not merge, enqueue, publish, or deploy. The upfix6 timebox ends at 20:11 UTC.

Current source and validation

The retry ordering and ordinary failure cleanup are corrected in the recovery test. It acquires the existing public runtime lease before asserting absence of abandoned state. It retains the original halted attempt, both normal attempts, every assertion, and the 30-second limit. A real halted-JVM fixture produces the deterministic failure; removing the initial reclaim call fails the old assertion. After the final cleanup correction, that deliberate failure leaves no owner runtime directory.

The exact 13 regression selectors passed 13/13 first-attempt at final source. Recovery took 25.045 seconds; the explicit-failure/XML scenario took 26.235 seconds. Both independent review passes completed and the reported failure cleanup gap was addressed. No product source, child policy, iteration count, timeout, or existing fixture consumer changed.

Measurements and limits

Pinned CLI 0.0.112 already sets the outer test JVM to C1 on both main and PR. Main run XML confirms a first-attempt recovery timeout despite final success: https://github.com/CodexCoder21Organization/kompile-core/actions/runs/37331069611. The proposed newly changed outer C1 mechanism is refuted. The PR ledger has 12 retry-only passes versus 8 main; 10 of those 12 first failures are outer timeouts, predominantly compiler frames.

Manifest revision parsing runs in the outer JVM: 228 checks, 62,328 total characters, largest 443characters with three records. Four nested JVMs recorded zero checks. Diagnostic timings show builds, preparation, child execution, and 2.354 seconds of parent-directory scans. The first timeout's full mechanism remains unresolved.

All 40 candidate-overlay measurements passed first-attempt. These changed both JVM layers and are explicitly not strict actual-main-versus-PR comparisons. No broad 20% C1 slowdown appears, but no measured candidate establishes both the 13.4-second compile budget and 2x recovery headroom. No per-test exception was introduced.

Additional nested-only comparisons preserve the outer flags. The first two parallel-cache pairs passed 53.669/52.393 and 71.985/71.577 seconds; their launch traces show only an outer C1 JVM and no nested launch, so this path does not exercise the changed child policy. Recovery, warm-classpath, and Maven actual-policy comparisons are incomplete because legitimate foreign builds own the shared slots. Do not claim the original 40 overlay rows complete those requests; do not remove foreign reservations.

Live CI and remaining work

Full suite: https://github.com/CodexCoder21Organization/kompile-core/actions/runs/37833615163. All four shards were in RunTests when inspected at 19:58 UTC. Fat-JAR build passed: https://github.com/CodexCoder21Organization/kompile-core/actions/runs/37833615214. The bounded watcher uses plain --pr385 --all-checks. Mark ready only after final-head CI is green; do not rerun code failures blindly. Inspect completed XML retry histories.

Source branch: https://github.com/CodexCoder21Organization/kompile-core/tree/fix/c1-native-sha256-child-jvms. Evidence branch: https://github.com/CodexCoder21Organization/kompile-core/tree/wip/sol-upfix6. Its handoff-artifacts/upfix6 directory contains findings, complete candidate tables, retry ledger, review, diagnostic source, and XML. Refresh its head from the final checkpoint.

Remaining items are final CI confirmation, strict actual-policy comparison rows, and resolution of the timeout cost or headroom design decision. The80-minute cutoff requires stopping local processes, checkpointing, releasing this claim, and deleting the fresh clone. No production publication or deployment was performed. Unfinished work is not a block on a person, so the blocked reason stays clear.

Prior verified snapshot

The following 18:42 UTC snapshot preserves all prior branch and deployment rows. Its PR head and CI statuses are superseded by the current snapshot above.

Resolve the final CI failure and mark the kompile-core PR ready

RE-VERIFY — 2026-10-08 18:42 UTC: The snapshot below was checked against the open PR and final-head Actions run. Recheck https://github.com/CodexCoder21Organization/kompile-core/pull/385 and its run links before acting. The PR is OPEN and draft at b3cfa5177542a5f472998fd0e5fc735e6f5f034b; the base is main at 7de0729dedab4871143e138de85db581667a71f8.

Mission and authorization

Finish the final CI gate for the child-JVM C1/native SHA-256 change and mark https://github.com/CodexCoder21Organization/kompile-core/pull/385 ready only after a green final-head suite. Do not merge, enqueue, publish, or deploy. The upfix5 brief allows a shard rerun only for a demonstrated infrastructure failure; no such error appeared in this run.

What was found and done

The requested 13-test local gate passed 13/13 on the first attempt at the PR head, with existing assertions and budgets unchanged. Four prior 30-second prepared-test timeouts were caused by stale checked-in compiled fixtures that forced compilation inside the timed tests. The generated runner source and ordered compiler classpath changed the fixture hash; regenerating the fixtures restored the no-compilation assertions. The captured-classpath fixture now serves the real Conscrypt Linux x86_64 classifier through its isolated Maven server and routes the configured Maven mirror to that same server.

The generated-runner cleanup mutation omitted only the provider/loader close. It made the public N1 lifecycle test fail before process shutdown: providerRemoved=false;nativeResourceAvailable=true;serviceFactory=created. The source was restored. The packaged measurements remain unchanged: native hashing 1.372 s and compilation 10.354 s; removing native hashing took 4.376 s, while removing C1 took 12.996 s and stayed below the 13.4 s budget. The absolute compile-control claim remains refuted.

Final-head CI ledger

Test run: https://github.com/CodexCoder21Organization/kompile-core/actions/runs/37820921933. Fat-jar run: https://github.com/CodexCoder21Organization/kompile-core/actions/runs/37820922002 (success).

Shard 1's only final failure is testExplicitDirectoryKilledAttemptRecovery. Its first attempt timed out at 30 seconds while org.json.JSONTokener.back was called from WorkspaceInputManifestKt.inputManifestHasRevisions. Attempts 2 and 3 failed at the test's preparation assertion because the prepared runtime directory from attempt 1 remained. The same test timed out on its first attempt on main run https://github.com/CodexCoder21Organization/kompile-core/actions/runs/37331069611 and passed on that run's second attempt. The other explicit-build test also had two timeout attempts on that main run before passing on its third attempt; it passed after automatic retries in the final-head run. This is the brief's main-attributed case, though the leftover-directory assertions from this run were not seen in the main-run result.

The 12 other first-attempt failures passed on automatic retries: shard 1 had eight, shard 3 had two, and shard 4 had two. No manual retry was made. The result contains no infrastructure error. The watcher exited RED, so the PR body was not patched and the PR was not marked ready.

Relevant PRs / refs

No code was deployed or published by this lane. No deployment or publication command was run.

Next steps

A new owner should decide whether to investigate the leftover prepared-runtime state after a timed-out retry as a separate defect. Do not rerun the shard to diagnose it. Once its cause is addressed or an authorized gate path is chosen, rerun the required final gate, patch the PR body only through gh api -X PATCH ... -F body=@..., reread it, and mark the PR ready only after green. Do not merge or enqueue.