Personally reviewed native context binding, November 16

Evidence: scratch/runs/20261116-native-reviewed-context-binding-gameplay/. The new, explicit contract is native-reviewed-context-v1. Existing native-scene-v1 full-PNG binding and historical closed endpoints are unchanged. No historical refusal or UNKNOWN is reopened or reclassified.

native-reviewed-capture retains the exact complete PNG encoder bytes in the native process, its original start/completion times, an opaque capture ID, and the surrounding public USER snapshot. The saved PNG is those same bytes. Observation is bounded by 3000 ms; different native contexts before/after encoding refuse capture. A caller must personally inspect that complete image, attest its hash and review time, and explicitly select menu/game scope and a native target HWND. This protocol records an external personal review; it cannot perform or manufacture the review itself.

Binding and queued input consumption check the retained capture, original timestamps, immutable review attestation, current native process/window/USER generation/owner/thread, focus, modal, capture and physical-UP state. All published native window/control identities, titles, geometry, styles, control IDs/classes and button flags are checked. Pointer coordinates must lie inside the explicitly reviewed target; keys require that exact focused target. The permanent input owner still pins renderer/queue/device/function identities. One capture may authorize one finite nonce, with prepare/consume revalidation; changing its scope, target or review timestamp cannot refresh it.

The supported public semantic classes are custom windows (0), buttons (1), STATIC (3), and native MessageBox (16). Other control classes refuse the new capture contract until their semantic state can be checked. For custom-drawn content, this is a native context contract, not a pixel-semantic recognizer: no OCR or inference about changing game pixels occurs. Personal targeting and the original 5-second gameplay / 30-second menu freshness bounds remain essential. A harmless repaint alone does not replace a native target, and the new contract never hashes a fresh unreviewed repaint as though it were the personally reviewed image.

The controller opts in only through a launch record explicitly naming the contract. Its captures, binding, intent and completion records label the route. It preserves 500 ms input confirmation, independent 1750 ms physical UP, 3000 ms observation/operation bounds and the irreversible Quit reserve. A receipt must also match owner/epoch/nonce/deadline and show consumption before both the immutable operation deadline and capture freshness expiry. Failure closes the caller permanently, with physical release still independent.

The authored PE fixture uses an ordinary USER window and a real guest Worker. Repeated guest GDI SetPixel calls repaint a decorative pixel while the target, menu text, focus and native control metadata stay unchanged. The final reviewed capture and later full native frame differ at exactly one pixel, (544,264). The complete frames were personally inspected. Binding accepted that unchanged native target and exactly one finite Down-arrow intent; confirmation was 36.754 ms and independently acknowledged physical UP was 255.424 ms. These receipts prove handler consumption/release, not a game consequence.

Native negatives passed for actual SetWindowText control-text changes, focus changes, renderer-handler replacement, actual DestroyWindow/CreateWindowEx replacement generations, an actual owned MessageBox opened by the fixture's explicit WM_APP hook, an unreviewed capture, and real observation timeout. Refusal/UNKNOWN callers made zero later endpoint requests. The WM_APP hooks are authored fixture fault stimuli, not physical input or original-game hooks. Unit checks additionally cover process/window/owner/geometry/control changes, unsupported classes, altered/unretained images, exact 5/30-second boundaries, changed review attestations, one-nonce reservation, queued context changes, and late/mismatched consumption receipts. The standard actual CLI guarded input test also passed. An initial invocation used the wrong .cjs suffix; that MODULE_NOT_FOUND failure is retained, followed by the actual .js pass.

The old stale fixture's desktop ETIMEDOUT remains a failed run: its stale assertion did not execute. New diagnostics use the same immutable Git loader bytes, real guest Worker mode, unchanged process-desktop source and 3000 ms broker bound, with its existing passive phase tracing enabled. One new run recorded poll submission-to-client receipt of 67.581 seconds and a further 36.375 seconds before socket write; its driver/watchdog also serviced late. This locates a real failure before timely broker servicing, not PNG equality or stale-review policy. Its exact OS scheduling cause and the old untraced failure's precise phase cannot be reconstructed. No broker timeout was relaxed. Those runs required host termination and remain failures. A later fresh run, stale-v7, completed the actual elapsed-time stale assertion, rejected with binding review expired, and proved zero later requests. Neither eventual process absence nor an unexecuted assertion is counted as a PASS.

An early timed replacement fixture had already replaced its child before capture and therefore failed its intended subsequent-change assertion. The final fault fixture uses an explicit public message after personal review to make the actual API transition observable. Earlier review-window expirations, startup stalls, missing wait receipts and forced waits are retained. An initial visual disappearance report was corrected after viewing the exact complete PNG again: only the decorative pixel changed; the window remained present.

The independently compiled canonical module is 83703369cc49c3e16077a4abd43def27262487f83dd52583d1b433adf7095f5f. Runtime sources use published Git plus explicitly owned patches, sealed in memory, excluding foreign WIP. Git and binary patches are the source archive; no source tree, media or Wasm archive is stored. Canonical locked full builds, focused tests and the unchanged 128 original media SHA-256 pins / 127 manifest rows / real-DLL closure are recorded in the run. Original-game outcome and final process/box closure are recorded in its final result.json; native fixture success alone does not claim NewQuest, a quest consequence or Quit.

The one authorized original was launched only after publication of 2593c43eaaaafb34e668cb43f6739306fc195220 and fresh unchanged-media preflight. The fixture and preflight immutable runtime used canonical Git 85f7dd61ace7b22a289d9d5b37975e5faa9797c0 plus that publication's owned source paths. The original was incorrectly launched without the recipe flag that activates sealing. Its exact loaded source/module manifest is missing; shared WIP exclusion and its actual module SHA cannot be proved. That provenance failure disqualifies the original as immutable-runtime proof, regardless of its visible scenes. No manifest was invented after execution. Its native limit was 1200 seconds, independent watchdog 1220, and irreversible Quit reserve 180, all recorded before launch. No second original, prior save, replay, OCR, solver, private board read or configuration change was used.

Personally viewed complete scenes reached Single Player, New Quest, the ordinary “The Search for Timmoria” quest list, available Undead Hordes race, and Warrior Lord / Average setup. Nine finite reviewed actions, each with its own capture and nonce, have consumption and independent-UP receipts inside 500/1750 ms. All these are setup/exit-seeking actions, not a quest consequence. The race selector's delayed changes were observed again rather than inferred from receipts. The final Enter reached “Please Wait” with an hourglass. One reviewed Escape sought the ordinary native exit/options route and was acknowledged, but no exit menu appeared. Subsequent complete frames became black. Later, after the irreversible Quit phase had begun, fresh full scenes visibly reached the quest map, an undead capital and turn indicator 1, then “Beginning day 1” with income 150/25/0/0/0. This visually records ordinary NewQuest startup and its initial day consequence in the unqualified runtime; no town interaction, manual turn advancement or battle is claimed. A fresh exit intent was NOT_SENT because its game capture exceeded the unchanged five-second limit. No DOWN was delivered for it. The day dialog was not dismissed through gameplay after the Quit latch. Native Quit remains unproved. The caller was permanently closed and the owned supervisor received a local host-cancellation file, with no further HTTP. Native exit 0 and independent watchdog exit 0 have actual waits and both EOFs; the host-triggered shutdown is not native Quit. Runtime lasted about 1117 seconds within its fixed 1200/1220 limits. Final receipts are in result.json.

The tool-managed sealed source holder and original outer launcher returned actual status 143. Their normal holder-close / outer-child-wait receipts are missing; process absence is not substituted for those receipts. The original supervisor and independently armed watchdog survived and retain their own native-child waits. Fixture/preflight source/module hashes were verified during their immutable launches; original sealing never activated. A final holder recheck failed because its /proc descriptors were already gone. The cause of that external tool-parent termination is unknown. These lifecycle limits remain explicit alongside the native fixture passes.

Final provenance audit: completionSession does not activate sealing; the supervisor gates it on moduleHolder (or its environment counterpart), both missing from the original recipe. This is a launch error, not an established foreign-WIP effect or an identified game bug. Future authorized execution must require and inspect the prelaunch immutable manifest before native creation. The single-original authorization is spent; no corrective second original was launched. The native-context contract proof remains the sealed authored fixture result.