Nox durable journal batching and native open-lane attempt
The Nox-only repair is published in 3c1acf80fbcc4bca0f09fa72ed18cde88217fd6b. It reduces measured journal overhead in the controlled slow-storage comparison while retaining a durable ordered prefix before every actual PTRACE_CONT. One changed native fixture passed. The conditional original reached a healthy equipped-staff open-floor aim, then ended on a discriminating command-schema refusal before the shot. No shot or target destruction is claimed.
Evidence: scratch/runs/20261109-nox-durable-journal-native-target-2110/. Exact source is Git plus code.patch, postcommit-owned.patch, tested-source.json, the generated historical small patches, and small-run-inputs.patch; no source/media archive was made. The native generation was sealed before launch and never modified afterward. A later test-only addition exercises the first-observed unlinked EXIT path.
Repair and durability
tools/nox-startup-custody.py authenticates the shared provider SHA-256 3448afd4c5ec4ea691c584a6cfde3a61b72f5c9143458189c1035f374c4c6240 before two uniquely anchored substitutions. Custody records use the existing bounded deferred batch. The common trace_task flushes pending journal bytes at entry for CONT, outside the ESRCH handler and before the real ptrace call. This includes both ordinary and first-observed unlinked irreversible EXIT releases. The existing executable admission barrier remains. Shared native_kernel_custody.py and shared providers were read-only.
The transformation changes no identity, lineage, ownership, sequence, cancellation, finalization, or bounded-pressure policy. Failed fsync remains sticky. Gzip footer finalization still requires its own successful sync. Existing bounded write/fsync/ptrace timing observations remain; metadata identifies the behavior-changing adapter as nox-durable-every-cont-prefix-v1.
The actual generated consumer test runs the whole generated Node consumer under real kernel custody using a Node browser peer, without local Chrome. Tests check source drift refusal and reversible exact substitutions; bounded/oversize ordering; sticky saturation/admission/finalization failures; failed final sync; gzip footer sync; and containment after admission failure. Real fork/exec tests compare every actual CONT's complete ordered journal prefix against bytes observed at a successful fsync, including EXIT. The additional real first-unlinked-EXIT fixture checks the successful-fsync byte prefix and complete drain.
The same 12 fork/exec children with 12ms injected at each real fsync measured 115 syncs / 1.632077251s for the shared implementation and 65 / 0.936549554s for Nox batching, with 13/13 tasks reaped in both. These paired measurements support reduced service overhead under that storage model. The generated consumer also passed with slow Journal.write, 18/18 reaped. See custody-test/pressure.json, custody-test.log, and the final rerun. Fourteen driver controls, three native-input tests, two observation-error preservation controls, and the locked build passed.
Native result and first failure
Exactly one leased box ran exactly one changed fixture, followed by one conditional original after fixture pass and independent staged-pin verification. The fixture completed normally: 148/148 owned tasks reaped, complete finalized journal, no cancellation/failure, parent return 0 with separate stdout/stderr EOF. The original reached native Nox Bunker as Conjurer Jack through ordinary UI, acquired the Force of Nature Staff of Recharge, and confirmed the equipped staff in inventory. Individually reviewed images showed full health and charge 3 while moving into an open corridor. The last reviewed image, review/barrel-aim-center.png, shows the cursor on the lower of three barrels across clear floor, well away from Jack. No enclosed-door shot was attempted.
Request 47, staff-open-lane-shot, incorrectly used action mouse instead of the consumer's supported action. This worker command error produced AssertionError: known action at 21:38:01.255Z before dispatch; mouse/hold receipts are null. The first failure was terminal: no retry, subsequent input, capture, or quit. Automatic containment closed Chrome; the original drained 340/340 tasks, no remaining/unreaped tasks or cancellation, finalized journals and command EOF. Parent return 1, stdout EOF and stderr EOF were separately recorded. No guest death or firing result is inferred from this refusal. Further target validation requires a separately assigned run using a validated supported click command.
The existing generated catch handler called close("first terminal failure") after refusal. That handler retrieves native events by evaluation and sends Browser.close. Although this worker issued no later command or screenshot, the automatic browser quit is a conformance gap against the task’s strict no-post-failure-quit requirement. This behavior was not changed or retried in the sealed run. It must be repaired and tested before another native attempt; the clean drain does not erase this gap.
Native timings do not demonstrate a speedup against the earlier unpaired box: changed fixture readiness was 8654.639ms with 8417.697ms overlapping fsync service; original readiness was 9338.295ms with 9116.682ms overlapping fsync service. These windows run from spawn return through port read and WebSocket setup. Timing samples had no drops. Service includes descheduling; box/storage conditions differ from the earlier 1314.989ms / 468.277ms fixture. The old timeout cause remains UNKNOWN.
Lifecycle closure
Box bx_7mwsptuf; immutable watchdog UUID 138d5588-0323-4b6a-898e-39717ee1e00b. It was independently observed before allocation; deadline 2026-11-09T21:41:17.112625Z, initial box expiry 2026-11-09T21:41:27.341Z. All 596 unique artifact entries were retrieved and independently reread with SHA-256/size, gzip CRC and PNG chunk CRC before the watchdog issued exactly one delegated normal unforced stop. Profiles, processes and ports were verified absent/closed.
Stop receipt stop_d423a3a7e86c0700 completed at 2026-11-09T21:39:13.800Z, forced false. Independent final API read confirms archived, stopped 21:39:13.780Z; final archiveAfter null, expiresAt not returned. Watchdog actual wait returned 0 with separate stdout/stderr EOF; its supervisor recorded the child reaped. No owned live box remains. The run retains retrieval, cleanup, stop, immutable-config, actual-wait and independent final-state receipts. Work ends here without another native attempt.