Q3 bounded delivery repair, 0950

The genuine Q3 composition now loads tools/q3-input-delivery.cjs through compileBoundInput. The historical scratch helper and adapter are unchanged. The new provider has a literal SHA pin and participates in the existing strict prelaunch dependency gate. Shared launch clients, custody, Redline and Q2 are unchanged by this task.

The 0924 first 100ms Enter had DOWN reply at 3546ms, UP attempt at 101ms, UP reply at 3547ms and zero held samples. Tracing the retained send and evaluate finds no caller queue that serializes those two submissions. Reply delay and delayed application reach the same historical assertion in the inert full-controller controls. Neither supports a native application-time, actual held-interval or guest-poll inference; the native latency cause remains UNKNOWN.

The supported correction is failure handling and release scheduling. At the unchanged hold deadline, an unresolved DOWN now fails terminal immediately, after submitting one UP for each attempted DOWN. UP submissions run independently in reverse order: an unanswered first UP cannot prevent the second submission. The old implementation's sequential UP await demonstrably prevented that second submission. No retry or later DOWN is allowed. First failure closes the actual pending CDP map and starts the existing joined browser cleanup. Existing hold, work, CDP, ACK and cleanup caps and the before-first-DOWN timer origin remain.

The actual command saves input-<id>-delivery.json on every transition, including failure, alongside gzip physical telemetry. Per-key submission and promise settlement times, sample bounds, first failure and cleanup settlement survive. A rejected promise explicitly leaves protocol reply/application timing UNKNOWN; a successful CDP reply still makes no application-time claim. Terminal failure cannot publish a positive ACK or continue into settle/capture.

Evidence: scratch/runs/20261109-q3-bounded-delivery-repair-0950/. Full composed controller tests use the real bound module, send/pending map, evaluate and numbered transport with inert browser/page seams. Four actual old-code negative controls and seven repaired cases cover reply delay, application delay, stalled DOWN, independent UP, sticky sample failure, normal key/combo completion and ordinary quit. Node children and pipe writers actually wait with separate EOF checks under independent watchdogs. These are offline controls, not native proof. Focused telemetry, dependency and prelaunch checks and the locked build pass. Accepted unchanged startup/identity/fixture source proof is hash-verified and reused; only the changed delivery/controller behavior is newly exercised.

No box, Chrome or guest was created. Box ID/state/expiry/stop receipt are null / NOT_CREATED. Native lifetime REVOKED; allwriters and edictABA UNKNOWN; actual Boat cancellation NOT_PROVEN. Next: separately authorized native fixture using the new pinned helper before any original game run. SingleTaskSTOP.