NFS3 finite paused-review controller, 2026-11-12

A new run-local finite controller and its emitted Chrome adapter are implemented. Nineteen ordering/refusal cases, four initial-Pause cases and five actual child process/IPC cleanup cases pass. Actual production Chrome/CPU/SoftwareD3D refused a stale Pause review before input and shut down normally. The live positive resume/drive/Pause transaction is not qualified. Fresh native NFS Pause, two settled paused frames, reviewed successive bends, lap and result remain UNPROVED. The remaining admission blocker is actual capacity, not a demonstrated game bug.

Evidence and reconstructable helper patches are in scratch/runs/20261112-nfs3-finite-pause-drive-controller-0040/. This implements controller work missing from 0020, parent publication 260adb999; it does not promote that historical Pause image or the 1815 82 MPH into completion.

The new finite-review.cjs consumes one explicitly reviewed pair of full released frames. Both must bind all22 identities, show Pause/Continue and the same observed world timer, and span at least one physical second after the first capture completes. Original-domain reviews additionally require explicit visible race-world stability review. A frame older than30 seconds, mismatched PNG, changed identity, unreleased key, absent deadline owner, unknown intent or unreviewed Continue location is refused before resume. A consumed pair cannot authorize another transaction. The controller chooses no route, creates no reviews and performs no blind repeats.

The authorized transaction uses one ordinary native Continue click, then one bounded1500–3000ms Up/optional Left-or-Right interval, followed by independent release, one balanced100ms native Escape, and a whole released capture. It then waits for a new review; Escape dispatch alone never becomes Pause proof. A separate one-use initial-Pause request can establish the first candidate image without first resuming. emitted-adapter.cjs uses the existing full22 admission chain and unchanged native per-key release subprocess. Every key's UP is armed before DOWN. Its new independently connected15-second guard is observable before resume and releases keys/mouse and normally closes Chrome on controller loss or transaction deadline. Each new native action also checks actual /home capacity.

Failure/UNKNOWN is terminal. Tests cover partial DOWN, held-capture/release overlap, controller loss, release and cleanup failures, identity replacement before and after resume, stale Pause review, changed timer/world review, consumed review and Escape uncertainty. Cleanup failure remains UNKNOWN rather than a successful release. Five subprocess tests exercise the actual emitted release and deadline helpers, IPC disconnect and both EOFs; their CDP transport is mocked. Those and the model tests prove controller ordering/cleanup behavior only, not NFS native Pause semantics. An initial subprocess harness lost its outer close notification; its receipt/wire evidence is retained as UNKNOWN. The corrected harness waits for actual exit plus each pipe EOF separately.

The actual Chrome case launched current normal04 CPU and SoftwareD3D Workers, with recovered module SHA 04c534192ee3fe39e388536c2ce710c12c47a5cd2ca4d95c19f9d00b1e8f0a17. It displayed a visibly labeled DOM fixture over a running Boids guest. The stale fixture review was refused before any Continue/key effect. Its raw report still has passed:false because the intended full success path did not run; the verification records only the observed stale-refusal behavior. This fixture does not pause the guest and cannot prove NFS Pause. Chrome and its controller have actual code0 waits and both EOFs, with PIDs absent and owned profile/temp removed. The outer tool invocation join is UNKNOWN. Later physical-release witnesses, initial-Pause path and cleanup additions are tested by the focused model/child checks, not promoted to a completed live positive test. Exact stale-candidate bytes have a separate SHA-verified reconstructable patch.

The final candidate adds before/after physical-release witnesses to whole captures. It is frozen with Git SHA, relevant tracked binary diff, untracked input patch and owned helper patch. Patch roundtrips and612 input pins passed. No source tree/media archive or production source edit was made; no locked build was required for this run-local/documentation change. Published provider/custody 5c2c410a5 exact bytes were checked read-only, not invoked. Explicit range staging, compact provisioning consumers and a complete provision join are not qualified for this new candidate.

Thirty-five retained proof rows and all133 existing media plus EXE/manifest were rehash-verified without matrix reruns or media copying. Original EXE remains 0defab3eeb22ee4b6e0007a4d5b26a99d868008ba77e2b9bd3ef770e924548ad. The exact0010 1500ms held control is reused by SHA. Its200ms overlap remains terminal UNKNOWN; no duration extension or original replay occurred. Historical timer/lifecycle UNKNOWNs remain unchanged.

Actual free space remained below1,500,000,000 bytes in the recorded samples, including1,483,190,272 at the final qualification work. No floor waiting loop, cache/reserve/vendor/DB/session changes, private-box allocation or original input was attempted. The closed-cache worker's files were untouched. This is more than a retained-hash audit, but it does not complete the requested original driving outcome. Box ID NONE, state NOT_CREATED, expiry null, stop receipt N/A; no box watchdog/manager was started and no inherited box was touched.

Next is positive qualification of the final emitted controller when actual capacity permits, followed by a freshly watched private box, complete current provision join/all SHA checks, and actual native Escape→Pause/Continue with two settled full frames and stable observed world/timer. Only that fresh review may authorize one resume/drive/Pause transaction at a time. Continue to separately reviewed bend consequences and naturally earned lap/result while actual capacity permits; the task remains scoped to those outcomes. No physical-GPU/FPS/audio, parity or full27 claim is made.