Shared kernel SIGTERM exit retirement, 2026-11-08

Repaired the first new Recoil custody failure under published 551b7bdc8: unlinked thread6812, GETEVENTMSG15, failure sequence1075. A real local kernel regression reproduces that refusal before the repair and completes normal terminal retirement afterward. This satisfies the substantive-repair stopping condition; no new original, browser, gameplay, or box was started.

Evidence R: scratch/runs/20261108-darkstone-exit-terminal-1610/. The exact starting Git SHA is in R/source-head.txt. Git plus R/owned-code.patch and the small fixture inputs are the recovery source; no source archive or media copy was made. Predecessor: parked exit repair.

The peer's actual retained journal remains at scratch/runs/20261108-recoil-generation-ready/retrieved-firstfailure/custody.ndjson, SHA256 74df1b218c47f97deae7b653a773711ffa90599be109559f2b7c125cc8499927. R/retained-cause.json independently hashes it and extracts the relevant rows. Thread6812/start15111/TGID6781 registers at1068 and parks at1070. The creator's clone stop peek at1072 is superseded by its consumed EXIT stop at1073. Thread6812 then presents EXIT at1074; GETEVENTMSG15 triggers the old SIGKILL-only assertion at1075. No clone lineage receipt for6812 exists in this journal. The independent clock census child6781 had a 2000ms execFile timeout. The timeout's default SIGTERM is the inferred sender; the kernel journal records the termination status, not who sent it. The historical failed drain remains a failure, with unresolved lineage; it is not reclassified as successful drain.

The real regression creates a process with a running sibling and a second thread parked at its initial autoattach stop. Only wait selection and signal timing are controlled: TRACEME, WNOWAIT, waitpid statuses and GETEVENTMSG are actual kernel results. SIGTERM passes the running sibling's signal-delivery stop. Its fatal group exit supersedes the creator's pending clone event, and the parked thread reaches PTRACE_EVENT_EXIT with message15. The predecessor fails with the exact assertion in R/before-confirmed-2-test.log and its journal. The earlier attempts remain retained: direct process SIGTERM first allowed lineage to arrive normally; another fixture attempt selected the clone after the child stop had already been consumed. Neither is labeled a reproduction.

tools/native_kernel_custody.py now permits retirement for a consumed PTRACE_EVENT_EXIT/SIGTRAP with a valid terminal kernel wait status, independent of the termination signal. EXIT is irreversible: continuing this stop finishes death without running the thread's userspace. A signal-delivery stop alone does not qualify. GETEVENTMSG remains required, and the retirement receipt is fsynced before CONT. WNOWAIT identity receipts remain durable before consuming or reaping; initial and fork-parent barriers are unchanged. Other unlinked stops remain parked. No TGID inference creates a parent receipt.

R/after-verified/ records successful normal drain of four tasks, terminal message15, no unlinked userspace marker, and a checked fsync-before-CONT receipt. The missing lineage stays in unresolvedLineage with kernelForkLineageComplete:false. Late real fork lineage after retirement/reap still binds the retained identity. The PID-reuse boundary uses actual fork, termination and reap events with only the later /proc identity observation injected; it refuses before binding the event to the changed identity. It does not claim to force real PID reuse. Its intentionally failed supervisor is contained by a fixture subreaper and EXITKILL, then reaped. A first test-suite attempt exposed adopted fixture zombies between tests; the added isolation fixes the harness and the initial failure log remains retained.

Eight focused custody tests pass in R/focused-tests-verified.log, including the real SIGTERM case, existing SIGKILL case, late lineage, observed reuse refusal, autoattach parking, WNOHANG-zero retry and both immutable-deadline refusals. Two actual startup integration tests pass in R/startup-integration.log: exact five-input bootstrap, actual custodian/fixture lock, kernel barriers, nested WNOWAIT reaper and complete drain; changed bootstrap refuses before uploads or starts. Four endpoint tests pass. The locked full build is recorded in R/build.log. Accepted browser controls/hover/content-barrier/kernel evidence are independently hash-checked by reference, with no unchanged Chrome rerun.

R/runtime-pins-selected.json preserves the predecessor's immutable selected generation and changes only the shared kernel pin. All548 exact providers and 42 physical helpers resolve. The current bounded endpoint wait stays included, SHA256 11683651fa033d56f861adeb64a3f46cb8422553b9651ba051f5200111614ecc. Historical plans and peer integration files remain untouched. Existing source and staged checkpoint media are referenced and hash-verified, never copied. The selected closure check causes no allocation or remote effects.

All52 recorded local fixture task/supervisor PID-start identities are independently absent in R/local-identity-drain.json. Startup uses a fixture lock. No heavy runtime ran and the local runtime lock was never acquired. There are no task browser endpoints, ports, profiles, writers or held inputs. No new box exists: ID/state/expiry/stop receipt are N/A. The peer's bx_a83xxze6 and files were never altered; its own cleanup and stop evidence remain peer-owned.

Publish the exact repair SHA to the board for generation integration, RELEASE, then stop for a fresh gameplay task. Native helmet restoration and onward progress remain unperformed. Custody does not claim exhaustive historical lineage: unresolved receipts, CLONE_UNTRACED audit limitations and the existing nonleader-exec refusal remain explicit. Full27/parity/audio/FPS remain open.