Q3 launch-client failure observation — 2026-11-09

The retained0052 caller launched task.py launch and later ran separate task.py command processes. Each command called c.remote, which owned only that command's Boat child; it never inspected the already terminated launch client. The remote launch parent drained the prelaunch log independently, so continuing remote execution did not establish local transport continuity.

tools/q3-native-launch-client.py is the production replacement entry for those separate invocations. It checks the selected task's SHA256, compiles its actual launch and command branches, and installs one owned remote function at both real call sites. The original remote Python launch-parent code and Node numbered request sender are generated by those retained branches. No historical task or input file is rewritten. For a separately authorized future run, start the entry with --task TASK.py --task-sha256 SHA --evidence UNIQUE_DIR and send JSON lines {"command":{...},"review":"..."} through its stdin. Allocation, provisioning, and independent box watchdog remain outside this entry; it does not authorize another original. The former standalone command entry bypasses this owner and must not be used for the managed session.

The first observed owned launch exit, signal or stdio EOF closes admission permanently. A direct process poll also checks before each command spawn. The observer performs an actual parent wait independently of pipe draining; a pipe holder cannot conceal child termination. Pending local command transport is cancelled promptly, and all later command/review/request calls are rejected before execution. First-failure evidence cannot be overwritten by cleanup. Cleanup begins automatically, even when stdin remains open. It uses a two-second local budget, bounded reader teardown, actual waits and individual stdout/stderr EOF flags. Missing EOF stays false; localComplete never asserts remote custody. EOF on stdin requests local cleanup and cannot fabricate a command ACK.

The production Q3 source composer also closes its own admission on unexpected browser exit/spawn error, CDP close/error or send failure. It immediately rejects pending CDP entries, aborts cancellable waits, and guards subsequent sends, execution and captures. Existing expected browser-close receipt handling and its 30-second cap remain unchanged. No input, screenshot option, clock, command work, ACK, runtime or cleanup deadline was extended. Marker SHA256 remains 1858eea7efb8875e44088ffe9f5ea434c3b22915444113204fb51673e4f5ac0d; built CPU remains a320d4c85156644fd12b6134d66b11784f3100b9b135756206c087cdacd43295.

Offline evidence: scratch/runs/20261109-q3-prompt-launch-failure-0130/. The launch-client test executes the actual retained caller branches and actual numbered Node request sender over local processes/pipes. SIGPIPE, exit7 and exit0 are exercised while a request is pending and before the next request. All six cases settle promptly without another request or invented ACK. Ordinary two-command flow uses the real numbered consumer and retains both successful ACKs. Separate EOF-before-exit and SIGPIPE-before-EOF cases preserve the distinction between wait and pipe drain; the latter fails cleanup at its test cap. Every fixture child, including the independent pipe writer, has an actual parent wait. The source SHA rejection is checked before composition.

The eight-case composed-driver suite executes the real initializer and command loop with Node process/pipe fixtures. New cases terminate the owned child during pending capture CDP and during a 10-second command wait: both settle in under 1.5 seconds, with no later DOWN or numbered request. Existing delayed startup, ordinary success, pending close, escalation, missing preload and drain-cap controls remain passing. q3-native-prelaunch.test.py passes unchanged; its nine audit-hook cases start no runtime. The required flock /tmp/wine-build.lock bash tools/build.sh passes. Unchanged accepted PATHHASH controls are verified and reused rather than rerun. Source recovery is Git SHA plus small owned patches only.

These are local production-boundary tests, not a Boat/browser/gameplay acceptance run. Cancelling a local Boat child cannot prove cancellation or absence of effects already dispatched remotely. Remote cleanup/custody still requires its existing independent deadline and receipts. The reason for0052's SIGPIPE is unresolved. Its historical33 missing waits and absent normal-browser-close receipt remain UNKNOWN; this change does not repair or relabel them. Native kernel custody was not refactored. A test-reader gzip-path mistake and an initially absent prelaunch test output directory were corrected before the accepted checks; failed logs are retained.

No box, original, browser, heavy runtime, worker or agent was started. Box ID/state/ expiry/stop receipt: NONE / NOT_CREATED / N/A / N/A. No peer files were deleted; free /home stayed above1.5GB. Next work is the separate bounded-custody shutdown failure before any separately authorized original. This offline task stops after publishing its tested code.