Native thread exit delivery and cancellation causality, 2026-11-08

The custody tracer delayed real thread terminal waits behind unrelated fork events. A native pthread fixture reproduces the resulting /proc disappearance failure using Chromium's unchanged 30-iteration exponential sleep check. The repair prioritizes final delivery after PTRACE_EVENT_EXIT, and consumes already pending kernel fork messages before cancellation signals can replace them. No Chrome, gameplay, remote runtime, or box was started for this investigation. Chrome startup with the repaired tracer remains unvalidated; this is a native causal repair, not a claim that the original Chrome deadline now passes.

Evidence: scratch/runs/20261108-native-thread-exit-causality-2038/. The exact predecessor is Git 3021d134c, kernel SHA-256 e36d34990d738e616611ce8a8f4309548abcb751bd5ef5ec119d20e561c49360. Recovery uses the recorded Git SHA and small relevant binary diff, with a patch for this new note. Git remains the source archive. Fixture C files are small run-specific inputs generated by the focused test; no media or module is copied.

Observed mechanism

Read-only examination of Q2's recovered custody and Chrome stderr confirms: thread 7378/25787, TGID 7377, reached pre-exit-stop at 20:28:52.671620 UTC (sequence 559). Its actual terminal reap was not recorded until 20:28:53.618146 (568), about 947 ms later. Between those records the tracer serviced 7471's autoattach and 7430's clone attribution. Chrome 7377 reported the 30-iteration thread-disappearance fatal at 20:28:54.116935; Crashpad's ptrace permission failure followed. DevTools appeared after the unchanged startup cap. The fatal does not identify its checked TID, so this timing establishes a compatible delayed-delivery mechanism, not an exclusive historical root cause.

pthread_join can finish after the kernel clears the child TID futex while a traced thread still exists in /proc/self/task. Its final tracer wait remains necessary. waitid(P_ALL, ..., WNOWAIT|__WALL) does not promise to prioritize that thread over competing tasks. Synchronous journal storage makes the delay between selected statuses material. The repaired tracer remembers exit-stop identities and probes their actual P_PID waits before unrelated work and command output pumping. A nonblocking probe with no status falls through so other threads can finish; ECHILD never counts as retirement. All normal identity, durable before-wait logging, actual waitpid, and PID/start checks still apply. The code retains the same ptrace options, per-record journal fsync, deadlines, signals and escalation interval. No browser flags or backend are changed.

Native proof

The new focused fixture compiles a small C program with pthreads and real fork children. A separate PID/start watchdog is observable and armed before the supervisor's fork barrier; both roles have actual terminal waits. The pthread program joins its worker and checks /proc/self/task/TID with exactly 30 exponential nanosleeps, without increasing that cap. An explicit scheduling shim selects real competing kernel statuses for P_ALL, and injects 15 ms storage latency after actual fsync. It fabricates no wait status, fork message, identity, or terminal receipt. This is a controlled reproduction of a legal ordering, not a claim to reproduce every detail of the remote scheduler or storage.

In the retained full-suite pair (exit-before-kifdbtng and exit-fixed-kbe_q76m), the published predecessor delays the worker's final receipt for 3.746 seconds: all 30 checks observe it and the program returns 42. The repair records final delivery after 0.033 seconds, the check succeeds at iteration 24, and the program returns zero. Both runs retain all 27 identities' actual terminal receipts, EOF and finalized custody/command gzip streams. complete: true in the failing predecessor means custody drained; its actual nonzero root exit remains a failure. A delayed thread zombie can receive the later group exit status; the test preserves the kernel's actual result instead of substituting its earlier exit-stop message or expected worker return.

An additional C control releases eight pthreads together and joins them. All nine identities have actual terminal reaps, no remaining tasks, and complete custody. This checks that one pending final delivery cannot block other threads whose exit stops still need servicing.

Cancellation lineage

The second paired fixture exposes a real child's initial autoattach stop before its creator's real fork event and requests supervisor cancellation at that barrier. With the predecessor's signal-before-wait ordering, SIGKILL can replace the pending creator event. The child is actually reaped but lineage stays unresolved and drain correctly remains false. With the repair, the tracer probes only already owned tasks for actual pending fork/vfork/clone stops, consumes their real GETEVENTMSG attribution, and then sends cancellation signals. The fixed pair has two actual reaps, complete lineage and custody; supervisor cancellation still exits nonzero. No PPID inference or descendant scan is used.

This preserves available attribution, not information already destroyed by an external fatal signal or a race after the last probe. Existing real SIGTERM and SIGKILL fixtures that supersede clone attribution before observation still report complete: false with explicit unresolved lineage. The PID-reuse, late lineage, actual tracer/adoption double-wait, wrong-status, third-receipt, WNOWAIT-zero, ESRCH, no-wait and no-pipe-EOF refusal controls remain passing.

Historical Q2 child 8107/30705 remains unresolved: sequence 951 registered its autoattach stop with PPID 7201; 952 consumed it and 953 parked it. Sequence 1272 sent SIGKILL. Its real EXIT stop was consumed at 1307, irreversible retirement at 1308 explicitly recorded lineageResolved: false, and 1311/1312 recorded the actual adopted zombie and reap with status 9. There is no matching observed parent fork event. Its later PPID 7188 and eventual absence establish neither the missing attribution nor any additional wait. Q2's original 132 real reaps and complete: false are unchanged. TR2's owned artifacts were not edited.

Validation and limits

python3 tools/native_kernel_custody_test.py: all 33 tests pass, including both causal before/after pairs and the eight-thread control. The first development run had two assertion failures: it incorrectly assumed a delayed worker's final status must be zero and disallowed one competing event before its final status became waitable. Those fixture artifacts are retained. Assertions now accept the actual legal kernel status and permit that single nonblocking interleaving; the failing predecessor still exhausts all 30 checks and the repair succeeds.

The locked flock /tmp/wine-build.lock bash tools/build.sh passed. Its result, original log hashes/sizes, source identities and exhaustive regular-file inventory are in the run receipts. Custody and command streams are gzip from their first stored byte; the audit verifies CRC, original SHA-256/size and EOF against each drain receipt. Symlinks are recorded without following them. Historical raw compatibility tests use temporary directories; no raw command/custody stream is retained in this task's evidence. Sampled /home free space exceeds 1,500,000,000 bytes.

No new box: ID none, state NOT_CREATED, expiry NULL, stop receipt not applicable. No old-box resume/delete, original retry, local Chrome, guest write, public deployment or subworker occurred. The next separately authorized startup validation must select the newly published kernel by hash, retain the same Chrome flags and caps, and continue to reject incomplete kernel lineage. Full Chrome startup and original gameplay remain open.