LMarbles local pre-body refusal recovery
The controller now distinguishes a locally proved freshness/admission refusal before any body attempt from uncertain delivery. The HTTP helper keeps private WeakMap provenance only for the synchronous connected callback, before end, with no headers sent, pending body or ended request. Exceptions from end, partial writes, remote error JSON and HTTP error responses cannot obtain this provenance. The controller additionally requires its private budget-refusal type, zero earlier receipts/effects, the actual successful independent release wait, a matching owner release ACK, physical UP/capture zero, and unchanged native/owner/review/window identity before and after the confirming snapshot. Existing targeting, confirmation, release, operation and runtime deadlines remain strict.
The rejected review remains reserved and the rejected nonce is never reused. Recovery requires a new capture, personal review and distinct intent nonce. Old UNKNOWN runtimes remain permanently closed. Identity/review replacement, possible sent bytes, partial effects, lost ACKs, failed/mismatched releases and uncertain physical state retain sticky UNKNOWN. Completed release-child waits now clear their unused timeout timer.
Evidence is in scratch/runs/20261118-lmarbles-local-not-sent-recovery-puzzle8/. Eighteen final real-controller/HTTP/TCP tests pass, using the production native input guard and a separate real release process. Delayed assignment and TCP connection past review expiry produce zero DOWN bodies/effects, followed by successful new capture/review/nonce and refusal of the old review. A strict freshness-budget case also recovers. Negative cases include owner/review replacement, replacement during the confirming snapshot, UNKNOWN, partial effect, lost ACK, partial write, end throwing after a real request attempt, forged remote refusal, failed/mismatched release and physical input held. The end test actually observes a DOWN despite the thrown exception and remains closed.
Threaded startup, sealed-source resolution and guarded-input regressions pass. Actual native windib positive/refusal tests pass using verified module 0d383140f6d4207d18860e3cb22fc8fcc47a5d260b8d5f2443222396b245fdd5. The native fixture seals its own source inventory. The original reuses the predecessor's separately verified 251-file source graph; all hashes matched before sealing. The earlier DirectX pixel-oracle FAIL remains separate and unfixed. No browser behavior changed; no browser acceptance is claimed. The first 17-case wrapper hit its finite watchdog after all cases passed because release-wait timers kept the process alive; the final run includes the timer cleanup and exits normally. The first locked-build wrapper lost its outer wait with tool exit 143; its supervisor records child exit 141 and no leak. That incomplete invocation is not a passing build.
The affected original ran once in a fresh bounded scope. A personally viewed complete main menu led to one consumed New Game action with matching independent UP ACK and a complete setup image. Start then exceeded the strict 500 ms confirmation bound. It became permanent UNKNOWN at 1795007205.939; its matching fallback release ACK does not prove that DOWN was unsent. Local cancellation followed with no further controller endpoint calls or replay. The actual native wait is exit 0 with both EOFs after cancellation, not ordinary guest Quit. Puzzle8 entry, clear, onward and ordinary Quit remain unverified. Final identity revalidation hardening was added after this original closed and is covered by the final controller regression; it does not repair that confirmation failure or justify another original retry.
Final qualification: all 18 controller cases, startup, sealed-source and guarded-input checks passed. Both native windib cases passed with the verified module. The required locked build passed on retry with actual wait 0 and both EOFs; the first wrapper exited 143 and its build supervisor recorded 141, so that invocation is not a passing build. No browser rerun applies to this native controller/transport change. The prior DirectX pixel oracle FAIL is separate and unfixed.
Original disk sampling stayed above 1.5 GB (minimum 1508655104 bytes). Separate qualification/task observations crossed the floor; no task-wide compliance or external-cause fix is claimed. No RAM gate, foreign cleanup, worker, box, Python, Wine, public deployment, solver, autoplay, OCR, board-memory read, configuration/save/media/clock/polling/focus change was used. Box ID, expiry and stop receipt are null; state is NOT_CREATED. Recovery records retain Git identities, relevant patches, source/module hashes, selected consequences and actual waits. The author qualifies and publishes the change in this same task, then releases ownership.