Q3 generated-parent signal cancellation shadow repair — 2026-11-09

The production drain_remote in commit 36d5529d2e614e4ee25056ea2919c0a683e57a6c still rebound its output-directory parameter to the closed caller-cancel-sent.json file handle. A parent signal leaves requested nonempty; the next pump iteration republishes cancellation through Path(output) and raises TypeError before the generated parent's child join. The repair0737 note's assertion that this shadow was corrected is false for those committed production bytes. Its accepted transport-publication cases did not exercise this nonempty signal-request path.

The production change only renames that marker file handle. No cancellation, command, CDP, ACK, runtime or cleanup deadline changes. Publication-only custody remains UNKNOWN. Native custody is unchanged at SHA256 5ec810c4d4704ff86afc367b38f7335edf1214e3450ebb98a62ce3df6386b5fe; marker remains 1858eea7efb8875e44088ffe9f5ea434c3b22915444113204fb51673e4f5ac0d; build CPU remains 2185596 bytes, SHA256 a320d4c85156644fd12b6134d66b11784f3100b9b135756206c087cdacd43295.

Evidence is scratch/runs/20261109-q3-signal-cancellation-shadow-repair-0805/. The extended tracked remote-cancellation test executes the actual generated parent, strict prelaunch, composed Node consumer and current native supervisor with inert children. Signal cases hold the runtime's pipe open in a fixture finally after native drain, ensuring multiple .02-second pump iterations. The parent receives HUP, TERM during pending CDP, or INT, then repeated copies and all three signal kinds. Cancellation marker, first-failure content and the single signal-sent receipt remain byte-identical. The parent stays alive during the held pipe, and cannot write its child-wait receipt until that pipe is released. No production source is replaced by a hand-written parent or drain mock.

old-final-HUP, old-final-TERM and old-final-INT each fail on the unchanged production source with the exact Path(TextIOWrapper) traceback and missing generated-parent prelaunch-child actual-wait receipt. Those missing joins remain UNKNOWN; they are failure demonstrations, excluded from accepted custody. The test's independently dispatched parent/request processes and watchdogs have actual outer wait receipts. Rescue releases only the fixture pipe and signals its actual supervisor; no original execution or capture retry is attempted.

focused-final passes all eight cases: three signal cases, exit/EOF during a command wait, exit/EOF during pending CDP, and normal completion. All accepted runtime/prelaunch, remote parent/sender, local client transport and independent watchdog waits are recorded, with EOF separately observed. Native drains are complete with command EOF true; ownership/reap counts are 19/19 for each signal case, 26/26 for exit and normal, and 25/25 for the other three transport cases. Cancellation yields no later capture or successful numbered ACK. Normal quit keeps its successful ACK and actual runtime/prelaunch status0; cancellation retains status1. First-failure stickiness also retains the existing pre-cancel control's earlier failure and refuses bootstrap with zero children.

The focused launch-client suite passes (SIGPIPE/exit0/exit7, pending and before next, normal numbered requests, early EOF and held pipe). The ten prelaunch controls pass. flock /tmp/wine-build.lock bash tools/build.sh passes. The prior accepted driver, launch-client, prelaunch and remote-cancellation proof is reused by source match against commit36d5529d2e: 792 retained proof files independently reread for SHA256/size/gzip CRC. Another 595 historical reference entries were reread against their accepted manifest, without copying them or rerunning the unrelated driver suite. Historical0052's cause, 33 missing waits and incomplete command EOF remain UNKNOWN; no historical wait or cause is inferred here.

Development focused passed the five transport controls but its first signal assertion failed because the test initially put its pipe hold after a supervisor call that raises on cancellation. The corrected fixture puts the hold in finally; the final old and repaired signal runs use that same corrected test. The initial prelaunch invocation omitted its output directory and failed before its controls; prelaunch-final supplies it and passes. Neither development run is acceptance evidence.

Recovery records Git base plus small binary diffs for owned source/docs and small run-specific scripts; Git holds source history. Logs are gzip from first byte; final inventory and an independent reread check stored/original SHA256, size and gzip CRC. There are no archives, media or module copies. All accepted checks and their prearmed observable independent deadline watchdogs are joined. No Chrome, guest, original game, box, worker, agent or deployment was started. No Q2, Redline, shared host/custody/WAT, HeroesII or CLAUDE edits. Free /home remains above 1.5GB. Box ID/state/expiry/stop receipt: NONE / NOT_CREATED / N/A / N/A.

Next is separately authorized fresh-manifest remote qualification; offline signal delivery does not prove Boat delivery. This single task stops after publishing its own tested paths.