Quake II remote startup failure and readiness repair, 2026-11-08

The upper northern opening remains unqualified. The one fresh remote run failed before connecting to Chrome, before original game launch, native input preflight, or any gameplay. No screenshot, passage, target, or onward consequence was observed. This task ends with a tested startup repair and a concrete custody blocker, without retrying the browser or allocating another box. Evidence: scratch/runs/20261108-quake2-upper-opening-1940/.

Exact selection and execution

Selection Git was 27a902b0c75b27fcedaae72a447c4d239c7ba63e, with 703 path/size/SHA pins totaling 109,875,817 bytes. All 704 uploads, including the manifest, joined before remote validation of every selected file and launch. Existing media/module paths were read directly; no local media/module copy or source archive was made. Selected CPU SHA remains d8facd0658b7f158c6e60bef2de0936471f438f8bf41e9240fd75635f1d319dc; it was not observed executing in this failed browser run. Historical cc438/f965 results remain distinct. The unchanged 187 sightline controls are referenced by hash as preparation only, excluding hull/dynamic-entity qualification.

The owned Q2 consumer directly integrates tools/native-browser-close.cjs. Concurrent close callers share one promise; profile removal and quit ACK require an actual terminal and pipe-drain receipt. UNKNOWN retains the profile and suppresses cleanup/ACK. The exact selected custody source is imported normally with journal_compression='gzip'; no runtime source rewriting or injected sink was used. A real /bin/true control verified terminal wait and gzip original hash/size before runtime.

At 19:41:57.233 UTC, the selected driver's fixed 100 × 100 ms readiness loop ended with ENOENT for DevToolsActivePort. Chrome stderr contains a DevTools listening endpoint on port 42799, but that line has no timestamp; its exact publication time is unproved. No CDP connection or original launch followed. The missing endpoint is a startup failure, not the previous local disk failure.

Tested readiness repair

tools/native-devtools-ready.cjs waits for a complete, valid port/browser-path pair until an explicit immutable deadline. Missing or partial publication can progress; malformed publication cannot become an endpoint. A terminal child or read error fails closed. The owned Q2 candidate calls it with the earlier of 60 seconds from startup and the existing browser deadline minus 30 seconds. Neither the browser cap nor cleanup reserve is extended.

Nine focused controls pass: the exact historical Q2 readiness block reproduces ENOENT with a delayed real file under scaled host-test waits, while the repaired consumer accepts the same publication; direct consumer integration; actual terminal-before-cleanup and concurrent close; UNKNOWN retaining the profile; delayed and partial publication; malformed/missing finite timeout; a real child exit; propagated read errors and invalid deadlines. These are host fixtures, not guest clock changes or fresh browser acceptance. Final focused and locked full-build logs are gzip-written from the outset, with original stream hashes. The candidate was not launched again.

Custody blocker and retained limitations

Browser close returned UNKNOWN at its 30-second cap, with no terminal/drain receipt and no quit ACK. The selected supervisor then failed at AssertionError('duplicate terminal receipt'), with 133 registrations, 131 recorded terminal reaps and two unreaped identities, 7173/16882 and 7179/16927. Its parent recorded an actual supervisor wait with return code 1.

The journal localizes the duplicate to Chrome PID/start 7185/16988: its first before-reap records PPID 7173 at sequence 1381, followed by actual wait status 0 at 1382. Sequence 1448 observes the same identity as a zombie with PPID 7172, then the duplicate guard fails. This is evidence of a parent/adoption transition; it needs a focused tracer-versus-parent terminal-ownership regression. This worker did not change the shared kernel or simply discard the duplicate.

The gzip trailer is complete and independently validates all 225,135 original bytes, SHA-256 c4532cf5cac202072ea129be6604f55c2085102883e117ac7c1528edff28bbc3. kernel-drain.json explicitly says complete: false. The original historical supervisor 143/cause UNKNOWN and truncated-prefix proofs remain unchanged.

At cleanup census, all 133 registered PID/start identities were absent, ports 42799/42301 refused connections, and the heavy lock was free. These observations do not replace the missing terminal waits. No input had been held. The profile was retained on UNKNOWN; no accessible FD referenced it, but 51 processes had unreadable FD directories, so exhaustive writer drain is unproved. No exhaustive CLONE_UNTRACED or nonleader-exec coverage is claimed.

Retrieval, shutdown, and recovery

All 19 selected remote evidence files, totaling 88,831 bytes, were retrieved and independently rehashed before requesting normal stop. Runtime stdout and custody were compressed from the outset; other small remote logs were compressed before local retrieval, with original-byte hashes independently verified. Recovery uses the exact Git SHA, relevant binary diff, small owned input patches and existing media/module path/hash pins. The executed driver patch separately matches its selected SHA ae0c74e3ad5a364a560f45277ba2a154f3abd1754d34bdcf4c9356bc91202371; the repaired candidate is distinct. No tree archive or peer WIP was copied. The reused local watchdog and upload-ledger helpers initially wrote six JSONL logs uncompressed. After all roles exited, their 617,068 original bytes were gzip-compacted, independently rehashed, and only those owned raw duplicates removed. This exception to the requested gzip-from-outset rule is recorded in local-log-compaction.json; runtime stdout/custody and focused/build logs used gzip from their first bytes.

The initial retrieval filter mistakenly omitted kernel-command.log. A later read-only archived-file retrieval was refused with HTTP 409; the box was not resumed. That unique raw command log is not recovered, and complete log recovery is not claimed. The separately retained failure JSON, supervisor traceback and finalized custody journal remain available. This is an evidence limitation in addition to the incomplete terminal drain.

The UUID was durably recorded before create, and ID/expiry before runtime. The independent observable watchdog was armed before allocation with an immutable 25-minute normal-stop cap. Box bx_rxvzubfd is ARCHIVED, expiry NULL, stopped at 19:44:15.274 UTC. Receipt stop_e37948f9ddc48635 completed at 19:44:15.291, forced: false, error: null. Both local watchdog roles have actual supervised exit-zero waits; their supervisor records no remaining children. Its own detached exit is not promoted to an external terminal-wait receipt.

No local Chrome, second allocation, restart/replay, private guest writes, handler calls, config/binding/backend/clock/seed changes, old-box operations, public deployment, subworkers, or old-run deletion occurred. A peer recorded a transient shared-disk floor breach; this run makes no continuous disk-compliance claim. Next: reproduce and repair the newly localized custody ownership failure before another authorized original. Full27/parity/audio/FPS remain open.