Redline closure validation: measured custody stall repaired

The 1634 generated launch's 120-second timeout was caused by durable diagnostic work in syscall custody. It was not a CPU compilation or provider hash failure. The preserved caller spent its final 101.448 seconds between closure-validation-begin and cancellation. Its observer filled after another 0.322 seconds, at dashboard.html, but custody still made 7,572 syscall resumes after the observation cap, with no capacity to observe those calls. The capped trace correctly remained incomplete.

Evidence: scratch/runs/20261109-redline-validation-delay-repair-1648/. preserved-analysis.json derives these counts from the original archived journal. diagnose.py uses the exact published collector and observer against 128 real provider paths, hashes and sizes, without copying providers. The original observed path hit its fixed 45-second bound; 44.183 seconds were measured inside fsync (735 observer and 733 custody calls). Ordinary custody verified all 128 in 0.267 seconds. A labelled cap-only control also timed out: returning to CONT at exhaustion alone did not address pre-cap durability cost. The repaired control verified all 128 in 8.399 seconds, with 35 observer and 37 custody fsyncs. These bounded controls used independent armed watchdogs, actual waits and separate stdout/stderr EOF. They did not run Chrome.

tools/redline-bounded-custody.py is a narrow explicit adapter over the exact published launcher, observer and collector. It buffers syscall diagnostics for already admitted tasks in bounded batches and returns to ordinary fork/exit custody when observation fills. Every ownership admission still persists its complete custody prefix before release. Observer batches are limited to 64 KiB and 256 records, flushed at ownership barriers and close. The original custody journal's bounded backpressure remains. Durability failures propagate; capped observation remains incomplete, and retrieval still uses the published snapshot validator. No new task is discovered or retired by inference.

The current publication route authenticates this additional provider in its closure and selects the adapter explicitly. Existing published provider pins, closure hashing, HTTP checks, admission algorithm, physical input/capture custody, deadlines and CPU bytes remain unchanged. Primary CPU is 3e491c0840bb7771328738a0d160b4bbaa3d7ab72074cbd38a7b41e5d7e8ec18; compatibility CPU is ecb45716b1ee7fc5218b1670dd270b19c7664ee30d000b3360a12786a1c21eaf.

The fresh actual generated provision passed offline, including byte-checked uploads, media assembly and the real startup barrier. The fresh actual generated launch passed at its original 120-second deadline: closure validation of all 904 providers / 149,109,355 bytes took 234.674 ms. DRIVER_PREFLIGHT then intentionally exited zero before server/Chrome; the controller correctly reported missing ACK and exited one. All seven owned tasks were actually reaped; generated parent, outer parent and independent watchdog waits retained separate EOF. This proves the repaired offline gate, not native readiness.

Focused checks passed: six bounded-custody tests (record cap, minimum cap, byte cap, late fork/exec, cancellation, digest/sequence validation, bounded buffer and failed fsync); generated caller corruption checks for host, both CPU modules and the new adapter; actual staged caller HTTP failure preserving root/custody/controller exit one with 16 missing/changed-provider controls and forged-repin refusal; complete current startup/admission suite; direct generated caller import audit with no outside dependency; and flock /tmp/wine-build.lock bash tools/build.sh. Initial regression mistakes (expecting a separate lineage record when vfork was registered directly, and reusing a PID's write-once phase files in a test) are retained with corrected passing runs. An initial startup invocation omitted its required environment variable and refused before execution.

Runtime followed Q2-reviewed1638's final published RELEASE STOP at main 1beaf26d20b056c306057994e956b340d696326c. Independent rereads verified its unique census SHA/size/gzip CRC, physical drain, immutable watchdog join and exact API ARCHIVED normal stop. Its runtime kernel/parent waits were NOT_CREATED and browser EOF remained UNKNOWN.

One observable independent immutable 25-minute UUID watchdog was prearmed before the sole box allocation: bx_y2mt6zju, UUID 5f1b9750-d477-4aa6-b921-d8f213e13034, initial archiveAfter=2026-11-09T17:21:47.659Z. The actual generated provision uploaded all providers and completed its remote barrier (ordinary transport parent actual wait zero, separate EOF). It then failed locally with EEXIST while persisting remote-barrier.json.gz: the offline provision had already used that write-once evidence label. This worker introduced that harness collision. The remaining seal/plan uploads were not reached, so this was not a completed provision gate. The first failure ended runtime. No fixture, original game, later input/capture, retry or deadline extension followed.

Read-only recovery independently verified a unique compressed census before delegated normal stop: 904/904 exact providers, no missing/changed provider, no runtime argv/FD holder or profile, and physical drain. Runtime kernel and parent waits were NOT_CREATED; ports were NOT_CREATED with no connect proof; browser EOF remains UNKNOWN. Recovery initially refused an out-of-scope /tmp upload locally; the corrected read-only census used the owned run path. This was evidence recovery, not another runtime attempt.

Final exact API state: ARCHIVED; normal unforced stop stop_35784f4489ea758a completed 2026-11-09T16:58:10.532Z, without error. Final archiveAfter=null; expiresAt was NOT_RETURNED. The actual watchdog joined zero with separate EOF and its immutable configuration hash unchanged. The box was not resumed or deleted. No box remains billing idle.

Git base plus small relevant patches identify sources; no source/media archive was made. Unique evidence is SHA/size checked, including gzip CRC and decoded identities. Empty/invalid artifacts remain invalid. Free /home stayed above 1.5 GB. No local Chrome, agent, worker or public deployment was started.

Outcome: causal offline repair tested; gameplay screenshot UNVERIFIED. Future authorized runtime must isolate offline provision's evidence directory from native provision labels before a new fixture/original attempt. This task makes no further runtime attempt and ends with final published RELEASE STOP.