Quake II upstream writer investigation, 0301

The repair remains unproved. Unchanged production passed on the authorized small Chrome fixture, including the old observer control, so this run has no causal failing-before/passing-after pair. No production fix is published. Both rejected Q2close0244 patches remain rejected and untouched. The improved fixture helper remains uncommitted WIP.

Evidence: scratch/runs/20261109-quake2-upstream-writer-repair-0301/. Git SHA, staged provider hashes and small input patches identify the tested source; no source tree, media or module archive was created. The four0300 relay controls were referenced by SHA/size without rerunning them.

Actual Chrome observations

All three runs used the real composed spawn, CDP callbacks, numbered quit, close helper and compressed stream logging. The original game was replaced with a small page and two Workers compiling an eight-byte Wasm module. Production shutdown, phase budgets, durability and success rules were unchanged.

Case Close elapsed Result
Boundary observer baseline 237 ms normal=true, exit0, no signals, both EOFs
Inherited 250 ms observer control 207 ms normal=true, exit0, no signals, both EOFs
Real upstream writer retained for 1,500 ms 1,569 ms normal=true only after both EOFs

The baseline maps Chrome PID5737/start5287 stdout/stderr to pipes23531/23533. Cat PID5755/start5304 reads pipe23531 and writes outer socket28400; cat PID5756/start5305 reads pipe23533 and writes outer socket28402. The maps also retain exact PID/start/FD/access identities for crashpad, zygote, GPU, utility and renderer writers. They do not establish the historical writers' identities.

In the delayed case Chrome PID8853 exited at 03:14:31.104Z. Holder PID9010/start33074 exited at 03:14:32.506Z, with the two stream end events recorded independently. Terminal wait and pipe EOF are separate evidence. All twelve later capture/key/combo/quit requests were rejected without adding another command file.

The boundary scans took 9–18 ms; the legacy scans took 4–7 ms on this box. Baseline journal writes totaled 616 ms across 1,273 writes (maximum 5.05 ms). The historical baseline's 421 records during its fifteen-second close suggest a measurement question, not a measured fsync cause. Neither observer overhead nor slow journal storage was established as the historical cause here.

One initial observer control mistakenly used the baseline's manifest and refused before Chrome. Its exit, complete kernel drain and readiness timeout are retained. A fresh corrected control produced the reported 207 ms pass.

Verification and lifecycle

All eleven focused composed cases passed, including forced exit-zero rejection, terminal timeout, delayed EOF timeout, later-command rejection and the independent 30-second cleanup cap. The locked build passed. Results are recorded in offline.log and build.log. They do not convert these unchanged-production passes into repair acceptance. Historical shutdown failures remain unresolved.

Both exact predecessor releases were checked: Warcraft0244 at 03:03:57Z and Super0300 at 03:09:17Z. One box, bx_m6w3a76f, had an independent prearmed 25-minute watchdog, immutable UUID a9d9318f-2ee5-45e2-b4c3-e5da77257e6c, finite TTL1496 and initial expiry 03:34:33.727Z. The watchdog tool session returned zero; a separately persisted parent wait/pipe-EOF receipt for that local watchdog is unavailable.

Before delegated normal archive, 107 artifact rows / 88 unique contents (149,832 stored bytes) were recovered and reopened with exact SHA/size and GNU gzip CRC checks. All four kernel drains were complete, covering 415 registered identities with terminal wait receipts; command EOFs were true, remaining/unreaped sets empty, recorded identities absent and ports closed. Absence was not used to manufacture EOF or terminal receipts.

The box is ARCHIVED, stopped 03:15:46.946Z, receipt stop_c41c040c0474c08a, completed 03:15:46.963Z, forced=false, error=null. Final archiveAfter=null; expiresAt was not returned. Runtime was released promptly. No original, local Chrome, peer box, agent or public deployment was used.

The concrete acceptance blocker is the missing reproducible production failure under the permitted fixture conditions. Further work needs evidence that distinguishes the historical failing environment from this passing one; changing shutdown code on these results would be speculative. Single-task STOP.