TIM positively unsent review recovery, 2026-11-15

Evidence: scratch/runs/20261115-tim-unsent-review-puzzle2/. The previous original in 20261115-tim-smart-handle-puzzle2 remains closed UNKNOWN. Its five delivered menu operations reached untouched Puzzle 1; its first construction drag was refused at review age 6.304 seconds before any command write. This task neither reopens nor reclassifies that run.

The old published policy was reproduced from Git in a disposable policy-only fixture: exactly 6.304 seconds yields current scene expired, permanent UNKNOWN, zero command records and zero consumed records. This is a controller classification problem, not evidence of a native engine defect.

The implemented recovery is limited to a native-attested REVIEWING capture with no held input and no delivery, atomically refused before a command is written. A refusal burns that capture and contains no action or coordinates. Native confirmation and a new complete capture precede a new review, intent and nonce. No action is automatically retried. Five-second gameplay and thirty-second menu limits remain fixed. Any command write, hover, queue, DOWN, uncertain ACK, transport failure or conflicting decision remains terminal UNKNOWN.

The atomic frame decision either refuses before command write or permanently claims an input nonce. A refusal receipt contains no gesture or coordinates. Native code verifies its current untouched review token before acknowledging zero send and publishing a fresh complete capture. A new intent identifier and personal review are required; the refused frame cannot be used again. Expiry after command write remains UNKNOWN, even if it produces zero native callbacks. UNKNOWN permits no new cleanup input; only an already armed independent UP may release a held operation. The bound is strictly below 1500 ms for chosen holds, with the independent UP deadline capped at 1750 ms.

Twenty-six policy checks and thirteen actual 16-bit COM controls pass. The native stale construction refusal has zero adapter calls, queued events, held buttons and callbacks, including 200 subsequent native probe steps. Native confirmation, a new complete capture and a separate fixture review then admit exactly one click with native movement/press/release conditions 1/2/4. There is no game model in this fixture. Separate native controls establish right input, held drag, late/dropped ACK terminality, expiry after write, Quit reserve, capture refusal while held, cancellation, independent UP and death after an executed native press. Normal holds are 100 ms; deliberately interrupted holds release at 300 ms. Parent death is actually waited with exit 17. The native interrupted-context/caller-stack/RETF and range/ratio suite also passes.

The first late-ACK diagnostic accidentally left the separate guard unmodified. Its assertion therefore failed after a normally balanced click; it was cancelled and actually waited (native -9, watchdog 0, supervisor-child 247, both EOFs). The corrected fixture injects faults in both processes and passes. This was a test-harness error, not evidence of a game or production engine bug.

LS32 and Q3 explicitly released source/build/runtime before these controls. Personally verified LS32 evidence contains 160 named hashed wait files, 80 native/watchdog/supervisor-child wait triples and 321 absent PIDs/groups. Q3 contains 54 hashed actual wait files with both EOFs, its direct original native/watchdog/supervisor-child closure, 109 absent PIDs and 96 absent recorded groups. Historical missing outer/ancillary EOFs stay UNKNOWN. Both locks were observed free. No workers or boxes are created.

The locked full build passes. The correction was published on main as 7d27c1d28928ac2460aad7e61bf1cd6a166cb6c0 before the one newly authorized original. Native transport controls do not prove a TIM handle.

The prior smart-handle, direction/belt and construction notes were read. Actual manual printed pages 12–16 and six selected original prior gameplay images were personally viewed. The manual specifies drag, separate placement click and then a handle click; attachment belts use separate endpoints. Prior images prove floating/anchored/picked-up/invalid cage states, not a flip. Each new original action will require its own complete current scene. Native Quit, host shutdown, input acknowledgment and visible puzzle completion are distinct.

The one changed-route original is now closed, with an exact partial outcome. Its 900-second native deadline, independent 920-second watchdog and irreversible 180-second Quit reserve were prearmed. All 436 launch source, ToyVM, host, controller, test and media pins remained unchanged. The installed game used the same thirteen original files, with no imported configuration or game state.

Personal full-scene reviews passed the title and printed-page-one manual check, then selected Puzzle 1 PLAY. Three motors were dragged onto the editable board and separately clicked for placement. Later scenes show the first two persisting without their handles when another object is active. The final complete scene shows the upper motor with handles still visible after its placement click; that alone does not prove a handle-ready or direction state. Six stale reviews were refused before command write and independently confirmed by native code; fresh complete captures and newly chosen intents then admitted operations. These refusals did not reopen the earlier closed UNKNOWN original.

I then supplied a malformed belt gesture using from/to, while this CLI requires x, y, toX, toY. Policy validation raised KeyError: 'x' and permanently closed this new original UNKNOWN. There is no command for that last reviewed frame. This is an operator schema error, not an engine bug or a stale-review recovery. I sent no further input and did not reclassify, rebind, retry or open another original. Twelve operations had been consumed, including one standalone hover. Chosen held operations were 100 ms. No belt endpoint, ramp, Run, puzzle clear, Continue/Advance, Puzzle 2, direction change or native Quit was reached. Gameplay beyond these visible placements remains unproved.

The native host closed with exit 1, independent watchdog 0, supervisor child 1 and launcher child 1; all four have actual waits and both EOFs. Native Quit was not delivered. nativeExited:false with a default exit code of zero is not native-Quit evidence. The outer tool returned 143 and exposes no separate pipe EOF; that outer uncertainty remains UNKNOWN. Closure audit records 96 named actual waits with both EOFs and 97 absent PIDs and matching groups, including qualification controls. Both locks were observed free before the explicit source/build/runtime release. No boxes were created (ID and expiry null; stop receipt not applicable).

Evidence is scratch/runs/20261115-tim-unsent-review-puzzle2/result.json. It references exact Git and relevant binary patches, pin manifests, tests, actual waits and release verification. Only four selected result-named PNGs remain; discarded captures have a filename/hash receipt. No source tree, media, WASM or capture bundle is archived. The final PNG is the exact partial Puzzle 1 scene. This task stops here; another original would require a new explicit assignment.