Q2 retained parent-wait provider repair

Evidence: scratch/runs/20261109-quake2-manifest-cli-guard-progress-1615/.

The corrected production manifest-file provision command passed both offline and on the single allocated box. The changed native fixture then stopped at its first failure: the actual retained 1645 launch wrapper calls runpy.run_path(.../pipe-wait.py) after spawning its remote parent, but fresh production preparation omitted that helper. The remote wrapper raised FileNotFoundError. No retry, gameplay input, later capture, cancellation fixture, original guard/door attempt, or deadline extension followed. Gameplay, guard defeat and onward-door consequences remain UNVERIFIED.

tools/quake2-staged-closure-gameplay.cjs now stages the exact existing 961-byte retained parent-wait helper, SHA-256 5c67aa36ea6d6c5251129309f0cd532937bf855c5cecec46bb71d63c9804e933. Preparation checks its bytes before creating the manifest; the production --consumer-preflight /absolute/run/runtime-pins.json checks its declared pin and actual bytes before starting the consumer. This is one explicit missing runtime dependency, with no positive provider substitution or blanket repin. The canonical EXE, owning Worker slot0/native-write route, primary CPU 3e491c08…, compatibility CPU ecb45716…, browser policy, physical gates and existing deadlines remain unchanged. Retained callers/evidence are immutable.

The consumer-preflight regression uses real fresh preparation and the actual retained provision command expression with the published executable/argument contract. It passes with Git forbidden and actual fixture/health/wire consumer composition. It refuses a directory, absent manifest, omitted parent helper, missing or damaged helper bytes during preparation, omitted observer hook and damaged policy pin. It does not execute the old retained stage-closure command. The existing historical ENOTDIR proof remains in run 1607.

The caller lifecycle regression now checks declaration of the helper and audits its actual Python runpy file read against the staged manifest, preventing the shared checkout from hiding this omitted dependency. Direct and actual kernel supervisor runs pass through the real retained 1645 launch branch and composed consumer at DRIVER_PREFLIGHT, with explicit stdin start, current primary browser/policy, startup barrier, actual outer/generated/runtime parent waits and separate stdout/stderr EOF. The kernel run has complete reaps, empty unresolved sets and released lock. Local controller readiness and early-refusal checks pass with actual waits and separate EOF. Existing staged missing/damaged provider, forged-repin, authority/version and historical-default controls pass. The required flock /tmp/wine-build.lock bash tools/build.sh passes and both CPU hashes remain exact. These repaired-path controls are offline, not native qualification.

Before allocation, the actual generated run provision branch was exercised OFFLINE through the production CLI, with real pinned file reads and hashes; only upload/remote transport was local. The original run had 741 providers; repaired fresh preparation plus the same six explicit inputs has 742. The first staged-control invocation omitted its required test-output environment and refused before running tests; the configured run passes. That invocation error is retained. Initial local recovery command construction also refused an embedded NUL before spawning an executor; its corrected evidence-only recovery is recorded separately. Neither is a native retry.

Exact predecessor UUID/API, final published release, unique recovered evidence SHA/size/gzip CRC, physical cleanup and actual watchdog JOIN0/separate EOF were independently checked. Peer workers remained offline; their owned paths did not intersect the production provider set.

Our sole box was bx_y37jnhc6, UUID fd8fb7e6-b84e-4ff0-9ee0-a715bc859c6c. Its independent immutable 25-minute UUID watchdog was armed and independently observed before allocation. Initial expiry was 2026-11-09T16:42:01.752Z. Cancellation drained all six registered kernel tasks with actual reaps and empty unresolved sets. Both recorded runtime parent waits have separate EOF. The generated outer parent's actual wait is UNKNOWN: the missing helper raised after spawn, so no wait receipt exists; physical reaps do not repair that missing evidence. The local controller and both local transport children actually waited with separate EOF.

Physical census found no staged runtime argv or run FD holders; profile and writer lock were absent and the runtime lock was free. Browser endpoint/port records were NOT_CREATED, which is not a numbered port-connect proof. Browser EOF and native gameplay qualification remain UNKNOWN/false. Empty command logs are never evidence of execution. The 17 unique-artifact references deduplicated to 13 compressed objects, independently SHA/size/gzip-CRC checked before normal shutdown. No source-tree, media or module archive was made; recovery records Git base and the small patch, with media referenced by existing path/hash.

Final exact state is ARCHIVED, archiveAfter=null, final expiresAt not returned. Delegated normal unforced stop stop_2175d0a06f9c43a9 completed at 2026-11-09T16:19:07.048Z without error. The independent watchdog was actually joined with status 0 and separate EOF; lifecycle parent tool exit is 0. No box was resumed or deleted. No local Chrome, agents, workers or public deployment was launched. Observed free /home remained above 1.5 GB.

Next: Q3 may proceed only after this task's final published RELEASE STOP and its own verification of ARCHIVED/physical drain. Any future Q2 native attempt needs a separately authorized fresh lease, fresh repaired preparation and all existing fixture/cancellation gates before one original attempt. This task makes no further runtime attempt.