GTA2 native USER readiness contract

Status: native USER repair published and qualified; one new original launch reached its menu, then refused before input on a renderer identity change. Ordered GL and Sims final releases, personally checked surviving evidence, current process closure and free locks admitted this worker at 07:08 UTC. Foreign host/GL/D3D/Sims work remains unadopted.

Evidence destination: scratch/runs/20261116-gta2-native-user-readiness-repair/. Predecessor implementation: 6d8b95c2cb4403b71d8e6ed2a21222d97fd3d12c; report: a0f642dd34b39095585a49ef4bb87fcdef828f2d.

The retained original attempt fails in workerBridge's initial snapshot during install with unhealthy native USER context, before capture and input. It does not retain the actual public process/thread/window snapshot or failed predicate list. Its exact unhealthy predicate and original cause remain UNKNOWN. The endpoint remains permanently closed; no requests are sent to it.

Retained console evidence includes Worker instantiation, native mss32/comctl32 DLL loading and both DllMain returns. The stdout file is empty. Console also records an api_table.json response parse failure and a missing tweetnacl error. These observations establish neither a native USER failure cause nor readiness.

Static install audit: the controller's startup test observes running, Worker and first slice, then installs the native bridge. Those host predicates do not check native USER health. workerBridge installs its onmessage handler and immutable global before invoking its initial healthy snapshot. A failed installation can therefore have mutated the endpoint already; no repeated installation is safe.

Authored controls keep every existing health predicate, report individual failures and preserve real public values including eip, quit and yieldReason. Their separate unpublished observer models bounded, positively UNSENT startup: delayed USER can remain unready; native exit, faults, retirement, identity change or deadline terminate without publication. A timeout never means readiness. This model is not proof of actual main/Worker scheduling or native behavior. Native authored before/after tests, all guards, full builds and publication are still required before considering the single conditional new original attempt.

The fixed 500 ms COMPLETE, 1750 ms independent UP and 3000 ms aggregate gesture limits remain requirements. No original capture, input, walking, car entry, driving, native Quit or WebGL success is claimed. No box or worker is created.

The new native harness is gated before runtime imports and PE generation. Its startup observations request real public exports on the actual owning CPU instance, recording the first host-ready state and native predicate details. It uses an observable prearmed independent watchdog and the supervisor's actual heavy flock. Its six fixture variants are authored but not generated or run: delayed USER, missing USER, early native ExitProcess, fault, capture and modal. The gate-refusal control and six standalone authored tests pass. These are host authored controls, not native qualification.

Further retained HTTP evidence has status 500 for /src/api_table.json?v=dev, /lib/vendor/tweetnacl.js?v=dev and /lib/vendor/webaudio-tinysynth.js?v=dev, all twelve-byte responses with SHA-256 8c7f644a14467e06ee798a89532dbff95329ac8e639e2b3b250f7d3fbaf8911e. The predecessor's runtime source-view selector excludes src/api_table.json and nested lib/vendor paths, while its tracked-source guard refuses omitted tracked source. The new authored selector includes these actual startup dependencies. An independent sixth authored control covers them and traversal refusal. This is a source-serving defect consistent with the retained HTTP/console errors, not proof of the exact old native USER failure cause. No closed endpoint is queried to investigate it.

Execution diagnostics must remain separate from the binding's scene identity. In particular EIP changes during ordinary execution; adding it to the snapshot compared by validate() would wrongly invalidate a stable native scene. The existing stable native binding fields and every existing health predicate must remain intact. A current readiness sample may report EIP/quit/yieldReason to explain refusal, without using those moving values as a reviewed scene token.

At 06:58 UTC, the new gated scratch source-view recipe, the predecessor recipe, and earlier control stdout/stderr files were found absent after successful reads or writes. The cause is UNKNOWN; this worker did not remove them. Wait JSONs, patches, the authored test log and this document remain. The separate unexpected-file-loss.json records current presence checks. Earlier artifact verification is a dated observation, not proof that those artifacts remain present. Native readiness execution remains unrun pending the ordered release and resolution of run evidence preservation.

Actual contract repair

The actual authored delayed-USER executable sleeps before RegisterClass, CreateWindow and SetFocus. At the first host-ready observation both CPU modes have public native PID 1000, TID 1 and nonzero EIP, quit 0 and yieldReason 0, but main HWND/generation 0 and focus 0. The exact false predicates are mainLive, mainVisible and focusLive. The old Worker install throws before capture/input; the old main install incorrectly returns INSTALLED with that same unhealthy snapshot. This reproduces a generic startup contract defect. It does not prove that the old original failed for the same reason.

The repair separates a read-only, positively UNSENT stage from endpoint publication. A CDP RemoteObject retains exact current process, renderer, module, CPU Worker/slot/endpoint and original handlers. Each sample reads actual public native exports on the owning CPU, with real current PID/TID, window generation, thread, visibility, focus, capture, modal, EIP, quit and yieldReason. It pins observed owners and rejects retirement, replacement, wrong thread, native quit, exit, faults and physical held input. No observer global or bridge handler is published. An actual healthy sample can advance to an atomic same-turn preflight and publication. Health never overrides a terminal generation-change sample.

The entire install stage has one fixed 3000 ms bound. Expiry is a refusal, not readiness. A failure seals the transport; pre-publication failures are UNSENT, and a failure after a publication request is UNKNOWN with no later requests. There is no retry or deadline extension. Both main and Worker bridge factories now complete their initial healthy snapshot before mutating handlers or globals. The public native ABI and the stable scene identity remain unchanged. Moving EIP/quit/yield diagnostics stay outside the scene identity. All existing native health predicates remain mandatory.

The source view now serves and pins actual src/api_table.json and nested lib JS/JSON dependencies. This fixes the separately proved source-serving omission. Build and runtime evidence names exact Git SHA and a deduplicated owned binary patch. Live shared-tree checks also record the relevant foreign patch as unadopted; HEAD qualification excludes it. No source tree or WASM dump is saved.

Intermediate actual main/Worker cases cover delayed USER, real MessageBox modal focus, real SetCapture, missing USER, actual ExitProcess, UD2 fault, window retirement/replacement, forced owner stop and a real TID 2 window with current owning TID 1. Only delayed healthy USER and valid modal install. Every startup case has zero capture and zero input; all negatives remain UNSENT. Forced host stop is explicitly distinct from native Quit. The final runs supersede these intermediate qualification states.

A candidate actual Worker input fixture also completed both modal dismissals, one ordinary ArrowUp and native ExitProcess through Escape. COMPLETE took 366/338/331/373 ms and independent UP 300/208/233/256 ms, within unchanged 500/1750 ms bounds. This is authored fixture evidence, not original gameplay. The published full fault matrix and main/Worker fresh normal runs are required before the single conditional fresh original attempt.

Final HEAD-plus-owned qualification: full flock build passed and all fourteen focused checks passed, including the sealed immutable source-view guards. The live shared checkout focused suite passed thirteen checks; the immutable-view check requires and passed its separate sealed HEAD run. Twenty actual final startup cases passed: four valid installs and sixteen UNSENT refusals. Every case has zero capture/input. Every negative has zero commit requests and its actual protocol wire sequence remains unchanged after seal. A real replacement window (HWND 65538, generation 2, TID 1) is healthy but terminal because the observed owner generation changed; it correctly remains unpublished. Final case source-state references tie each run to the exact owned source patch.

Published qualification and one actual original launch

Implementation 1305d45eb47203f32bfa690117265c4967c72a9d is on main and strict pull/push passed. The published source passed all fifty actual native fault cases and eight normal main/Worker gestures. Forty-eight faults have zero physical DOWN; the two intentional stalls after DOWN have independent physical UP at 1469 ms (main) and 1388 ms (Worker), ACK at 1470/1389 ms. All fifty retain permanent closure and zero-wire local replay/rebind/after-UNKNOWN refusals. Normal COMPLETE was 117/100/134/150 ms on main and 398/296/281/380 ms on Worker; independent UP was 148/139/158/149 and 314/232/204/216 ms. Both authored native Quit routes exited the real native process. These are fixture successes.

The audit distinguishes actual module instances (one main, two Worker) from freshness passes. It verifies the independent before-ARM and confirmation freshness phases. The original controller retains both final parent freshness calls, independent confirmation, COMPLETE 500 ms, physical UP 1750 ms and aggregate 3000 ms. No guard or deadline was weakened. One expired Worker review, one expired context-fault review, one host-deadline case and the prelaunch recipe cancellation are retained as unqualified attempts, not passing cases.

Original qualification uses source-equivalent current main 035e10cf849b6d14086814b10c51df3f821246e1: the only committed difference from the implementation qualification is TODOS.md. Source/module pins remain exact; foreign shared-tree WIP is excluded. A controller-entry preflight first refused ENOENT before Chrome/game/endpoint creation because its child output directory lacked launch metadata. It is recorded separately from the single actual game launch. The recovery recipe supplies a link from the child launch.json to the actual newly prearmed supervisor launch metadata; it neither fabricates nor extends the 900 second budget or independent 920 second watchdog.

The single actual original software launch now proves its current false predicates: public PID 1000/TID 1/nonzero EIP/quit 0/yield 0, HWND 0/generation 0, mainLive=false, mainVisible=false, focusLive=false. Fifteen UNSENT samples span 541 ms until actual HWND 65537/generation 1/TID 1, visible 648x508, matching focus, no capture/modal, quit 0/yield 5 and every health predicate true. Only then are the bridge and capture installed. This is proof for the new launch; the exact cause of the old closed original remains UNKNOWN.

The selected complete host image shows the original GTA2 demo menu with PLAY highlighted above OPTIONS and QUIT. Before the requested finite Enter could reach release-child creation or native input, the permanent-owner guard refused: identity changed: renderEndpoint, actual 1 versus startup none. The renderer published after the startup identity was bound. This is the first new blocker; it is not evidence that the renderer's generation may be ignored or rebound. The endpoint became permanently UNKNOWN and closed. No later command, review, rebind, replay, retry or WebGL run occurred. Zero release children and zero native input were sent. No walking, car entry, driving or original native Quit is claimed. OS browser shutdown is recorded as host-only closure.

Next work is to prove renderer publication/ownership readiness before permanent scene binding, preserving every current guard and qualifying it with authored controls. This run does not authorize another original launch or contacting either closed endpoint. Final result/release evidence in the run directory names selected full screenshots, exact Git/patch states, actual child/watchdog waits with both EOF, current PID-group absence and free locks. No box was created.

Final shared restore full flock build passed, followed by shared thirteen and sealed-HEAD fourteen focused checks. Current closure personally checks 895 retained actual child-wait receipts with stdout/stderr EOF and all 114 recorded native PID groups absent. No process with this run's profile remains. Both wine-heavy-runtime and wine-build locks are free; /home has over 7.8 GB free. The admitted foreign source pins remain byte-identical and unadopted. Seven selected full screenshots are named by result.json; redundant captures were removed with their hashes recorded. Box ID/state: null / NOT_CREATED; expiry and stop receipt are not applicable. SOURCE / BUILD / RUNTIME RELEASED at final publication; no further original work is authorized by this report.