Nox custody scan after durable service

Evidence: scratch/runs/20261109-nox-kernel-thread-lineage-shot-2225/.

The Nox adapter now checks real pending creation stops after its durable journal flush and before a queued userspace CONT. A creator can enter clone while that flush blocks, after the preceding event-loop scan. Releasing another thread into exit_group can then replace the creator's unconsumed message. The late scan defers that queued CONT; the ordinary consumed wait and GETEVENTMSG still provide the only creation proof. It does not register from a peek, infer a creator from TGID/PPID, or resolve missing lineage from terminal receipts.

The shared custody provider remains byte-identical and SHA-authenticated. First-observed unlinked EXIT retirement, every-CONT full-prefix durability outside ESRCH handling, sticky service errors, bounded batches, deadlines, terminal transport behavior, and original-media identities are retained.

Retained ordering and its limits

The independently decoded previous journal matches SHA-256 3e9a974509e6854471bd5ffd4fd7bd3c99111cb3fc93690d31a79bc01f61b8ad. Leader 6760 was registered by consumed FORK/GETEVENTMSG at 1807–1809. Its last retained consumed CLONE identified 6769 at 1912–1914. Leader 6760 resumed at 1917; 6769's initial stop was consumed at 1919 and resumed at 1920. 6771 then appeared unlinked at 1921, parked, consumed EXIT/GETEVENTMSG and was reaped at 1932. 6770 first appeared at EXIT at 1933 and was reaped at 1941. Leader 6760's EXIT was consumed at 1999, released at 2001, and reaped at 2008. Neither missing child has a retained creator message. There is no nonleader EXEC transfer in the journal, nor an EXEC stop for leader 6760.

The journal records event observations, consumed waits, messages, and CONT returns; it does not trace syscalls. It cannot establish the actual creator of either missing thread or which thread initiated group termination. The durable-service race below is a real reproduced mechanism, not a recovered historical syscall sequence. The old run remains refused.

Causal controls

tools/nox-kernel-lineage.test.py compiles a small pthread fixture and controls only scheduling around actual fsync. While the killer thread's resume barrier is servicing the journal, an already admitted creator reaches a real CLONE stop. No wait status or event message is synthesized. The previous adapter releases the killer, loses lineage, reaps all five tasks, and refuses. The repair consumes the CLONE and message and records custody before the killer CONT; all five tasks reap with complete lineage. Every actual CONT is checked against the full prefix witnessed at successful fsync.

A real SIGKILL that destroys the creator evidence still refuses despite all tasks reaping; the unlinked child never executes userspace. The generated adapter also passes the real leader EXIT/nonleader EXEC/successor-wait control; missing former-TID binding and missing successor wait remain refused. Failed fixture descendants are independently adopted and actually waited.

Validation and native outcome

Code commit ca5880f0114978186b13eda5c0ee40ad457d7be0 was published to main before allocation. lineage-tested/ retains all six real-kernel controls and actual child waits. The locked build passed, as did the generated actual-consumer 20-control suite, custody durability/first-unlinked-EXIT/sticky-service suite, seven terminal cleanup cases, and three native-input controls. The supported LEFT-click/150ms command also passed generated command validation; this is not a native staff-shot claim.

Both native phase manifests were sealed and independently reread before the single allocation (833 pins each). The sole changed native fixture passed: Chrome 0, consumer 0, actual parent wait 0 with separate stdout/stderr EOF, 146 owned/146 reaped, complete kernel lineage, no failure or cancellation. Its final boundary recorded zero original launches. Only then did the single conditional original launch start. It reached the Bunker and ordinary movement worked; every returned image was separately viewed before dependent input. The last reviewed image shows full health and the fist HUD in a stone corridor. The staff was not acquired and no staff shot was attempted.

The original terminated at request 29 with current reviewed capture required at 22:46:43.983Z, before dispatch. I omitted label from request 28; movement succeeded and the driver saved the capture under the undefined filename. I retrieved and viewed that existing capture, then incorrectly supplied the string undefined as the next review reference, which did not match the driver's undefined label. This is a task-local orchestration error, not a kernel lineage failure. The guard was retained. There was no subsequent game protocol, input, capture, retry, or quit. The original's Chrome exited 0; the consumer and actual parent wait exited 1, both output EOFs were observed, and all 343 tasks reaped with complete kernel lineage. Both runs had no remaining or unreaped tasks, no cancellation, closed ports, absent processes and profiles, and closed gzip journals with command-log EOF.

retrieval-verified.json independently rereads all 390 remote artifacts by SHA-256 and size, validates gzip and PNG CRCs, and reports no content errors. Only after that receipt did the independent watchdog issue exactly one delegated normal unforced stop. Box bx_6cre9386 is independently confirmed archived; receipt stop_4fc816999dde403a completed at 22:47:36.380Z. Its immutable UUID is db6fee49-ed88-4a5b-abdf-8fa81a0cf02a; deadline was 2026-11-09T23:01:41.507860Z and initial API archive expiry was 23:01:42.929Z. The watchdog's actual wait returned 0 with separate stdout/stderr EOF; a separate API state and stop-receipt read confirmed completion and unchanged config SHA.

The tested custody repair and native fixture succeeded. Native staff charge and target-change validation remain untested. Any future shot requires a separately authorized fresh run; this terminal run must not be resumed.