Neverhood local image expiry refusal

Evidence: scratch/runs/20261117-neverhood-known-no-dispatch-review-room/.

The frontend now distinguishes REFUSED_BEFORE_DISPATCH (CLI exit 3) from permanent UNKNOWN. Only an expired local whole image can produce this outcome, after identity, image hash, target, review and action checks. The operation must have sent no request and published no bind/action intent or arm. Outstanding intents or arms prevent eligibility. The refusal invalidates the capture's review and the operation nonce. Its immutable image and timestamps are retained.

Recovery requires an explicit new native whole-image capture, followed by a new review. Older captures cannot be reviewed for recovery. Binding remains once per target; a recovered action uses a new nonce. Independent release after a refusal requires a subsequent fresh reviewed intent. No automatic input retry, image restamp, clock reset, server guard change or UNKNOWN recovery is allowed. The original 15/30-second freshness, 500 ms confirmation, 1,750 ms physical UP, 3,000 ms observation and 1,800/1,820-second installed runtime/watchdog remain.

Authored real native tests wait for actual expiry before binding and before input on an existing binding, verify zero transport/intent/arm publication and unchanged original clocks/images/reviews, then capture and review a fresh scene and complete native input, independent UP and Quit. Separate real native semantic/owner/focus failures and an observation timeout remain permanent UNKNOWN with zero subsequent transport. Unit tests also reject changed request ledgers, pending intents/arms, capture restamping and release without a new reviewed intent. The canonical locked build and lifecycle/shortcut checks pass.

The final run outcome, published hashes, selected whole images, actual waits, EOFs and cleanup disposition are recorded in result.json. Native Quit and host cleanup are reported separately. No workers, boxes or public deployment.

The one fresh installed-shortcut preparation with published d97dc35ad refused before original execution. Its real native child Quit test caught partial JSON while reading an evolving native-child-*-wait.json: the producer truncated the public file for each exit/EOF/close update. The controller correctly latched UNKNOWN; that endpoint was never queried or retried afterward. Original game invocations, room progress and original native Quit are all zero/unproved.

The subsequent demonstrated fault is fixed at the producer. A private file is written, fsynced and closed before atomic rename replaces the public receipt. Readers see a complete old or new version. A failed write preserves the last complete receipt and removes its private partial file. This changes evidence publication only, without changing consumption, identity, exit or EOF predicates.

The regression executes the exact old and fixed production callbacks and starts an independent OS reader halfway through each write. The old callback exposes partial JSON; the fixed callback never does, including a controlled failed replacement. Fresh sealed native Play/parent exit/child capture/input/UP/Quit and negative handoff cases validate the repair with unchanged deadlines. The thirteen-case frontend suite passed before the wait-publication delta; fresh game expiry/bound-owner recovery cases also passed with the repaired producer. The native core and module are compared exactly against the prior proof; only the reviewed wait-publication delta is excluded from that equality claim. There is no full historical suite rerun claim.

No second original preparation or invocation was attempted. The repaired subsequent fault is the stopping condition for this assignment; a fresh original room/object attempt remains future work.

Two later authored positive rehearsals failed closed under the original guards: one ordinary fixture exited before its independent UP connected; another Play receipt was consumed at 461 ms and retained at 463 ms, but subsequent durable receipt publication crossed the 500 ms confirmation boundary. Both became true terminal UNKNOWN with no recovery or later transport. Their evidence is retained; neither is counted as a passing complete suite or repaired by the atomic JSON change. A concurrent auxiliary build was interrupted by SIGTERM before its own wait receipt; the final locked build runs after all native rehearsals close.