Tracer terminal notification followed by adopted-child reap

The selected Quake II kernel was e1a413983f251534ad77dbf337844ab9ca3afb74, tools/native_kernel_custody.py SHA-256 04a3598c7e23966d57173df5f42ba740fd36106ca4e8f78c84c435f49451221d. Evidence remains at scratch/runs/20261108-quake2-upper-opening-1940/remote/: custody.ndjson.gz, kernel-drain.json, supervisor-actual-wait.json. The journal's original 225135 bytes hash to c4532cf5cac202072ea129be6604f55c2085102883e117ac7c1528edff28bbc3.

PID 7185/start 16988 was registered by fork at sequence 52 with parent 7173. Sequence 1381 records CLD_EXITED/0, zombie, TGID 7185, PPID 7173; 1382 records the actual terminal wait status 0. Sequence 1448 records another CLD_EXITED/0 for the same PID/start/TGID, now PPID 7172 (the subreaper). The following wait hit the duplicate assertion; its consumed status was not logged, so that exact historical second status remains unknown. Sequence 1449 records failure. The finalized drain is incomplete with unreaped 7173/start 16882 and thread 7179/start 16927; supervisor 7172 actually exited 1. This is evidence of tracer notification followed by real-parent adoption, not evidence of exec or PID reuse. No historical nonleader exec is inferred. The original game was never launched; the omitted kernel command log cannot be reconstructed here. The separate earlier supervisor-143 cause stays unknown.

The repair retains the first actual terminal wait as the identity receipt, including its observed terminal parent. A later actual terminal wait is accepted as adopted-reaped only when the original parent was external to the supervisor, the new observed parent is the supervisor, the PID/start matches, the task is a process leader, the consumed status matches, and no adoption receipt already exists. The journal links both receipts and the durable WNOWAIT observation preceding the second consumption. All other duplicates still fail. Parent-side zombie disposal does not fabricate another receipt; the tracer's actual terminal wait already proves that identity's death.

Registration, pending lineage/ESRCH reconciliation, nonleader-exec refusal, identity checks, bounded cancellation, gzip finalization and incomplete-drain rules remain enforced. A present zombie still prevents a complete drain.

Offline evidence: scratch/runs/20261108-custody-terminal-identity-1951/. python3 -B tools/native_kernel_custody_test.py passes 25 tests, including six terminal-adoption controls. A real Python thread forks a child, which exits 7 (or receives SIGTERM). The real ptrace terminal wait releases a file barrier; the parent then exits without waiting, or consumes the child itself. All waitid, waitpid, ptrace and adoption results in the positive controls are real. The exact selected historical source fails on this adoption control, while the repaired source drains completely. This reproduces the lifecycle cause, not the historical scheduling of every Node thread: the historical two unreaped siblings are not recreated as an assertion target.

Two negative controls deliberately corrupt only the second consumed status, or inject a third receipt after the real second wait. Both remain incomplete with the duplicate assertion. They are fault injection, not kernel evidence. Each control has a separate observable seven-second watchdog armed before the supervisor fork barrier, an immutable supervisor deadline, actual terminal waits for supervisor/watchdog, and an outer subreaper draining any adopted children. Watchdogs did not fire; final controls' outer adopted-wait lists were empty. Intermediate stale-directory controls did require outer adopted waits (exit 0), retained in their isolated-waits.json files. Each invocation uses fresh fixture directories so retained barrier files cannot release a later run. A failed intermediate rerun using stale directories is retained as focused-repeat-stale-fixture.*; final results use the corrected unique-directory fixture. Watchdog targets are checked by PID/start identity. Another intermediate suite run failed the existing runtime-deadline control with an incomplete drain (focused-deadline-incomplete.*). Its temporary fixture details were not retained, so the specific cause remains unknown. The final suite retains gzip controls in gzip-final/ and passes with the original .4-second runtime/.6-second additional cleanup bounds unchanged. Custody journals and test/build logs are gzip from their first record/byte. Small fixture barrier and wait JSON files are retained directly.

The required flock /tmp/wine-build.lock bash tools/build.sh passed; the module remains SHA-256 d8facd0658b7f158c6e60bef2de0936471f438f8bf41e9240fd75635f1d319dc. focused.json, build.json, and manifest.json record results, stream hashes, source commit, relevant binary diff and existing executable/ module path hashes. Git plus the small owned patch is the source recovery recipe; no tree archive or media copy was made. Existing game media were not used. No browser, original, remote service, box, deployment or worker was launched. Box ID/state/expiry/stop receipt: N/A, NOT_CREATED.

Next: a separately assigned fresh Quake II run must explicitly select the repaired kernel and readiness helper. Full27, parity, audio and FPS stay open.