MCM owner published before first PRESENT

Evidence: scratch/runs/20261120-mcm-owner-first-present-loading-repair/.

The earlier original scope remains permanently closed. Its post-Start capture failed in PAGE_READ, but the exact failed predicate was not retained. This task does not infer that predicate or repeat the original to diagnose it.

The adapter now reports nine boolean refusal predicates from the same page read: missing layer, missing window, GPU alias mismatch, stale layer, zero presents, null stamp, zero write sequence, WASM owner mismatch, and process mismatch. It requests no later diagnostic object after a refusal.

Before capture, a bounded readiness path permits only the pristine Presenter state: layer/window/stamp null and presents zero, with the live bound GPS endpoint and registered process. It checks actual guest and renderer identities, module inventories, executor and context on each observation. Process, main instance, renderer, Presenter, endpoint, Worker, module, executor, context and epoch remain pinned while waiting. Only the initial absent window/layer may become a completed publication. The wait lasts at most 10 seconds and admits neither capture nor intent while pending. An unknown read, changed owner, retirement, stale or incomplete publication closes the chain. Input admission and the post-screenshot read still require a complete publication.

The focused native fixture uses the real compiled D3DIM producer's PRESENT, native HWND publication, software ring executor and production Presenter. The qualified 8e7989e5c chain rejects the pristine owner before capture. The new chain waits, then captures only after the real native output arrives; green pixels, native HWND publication and both production layer aliases are checked. Worker/CDP transport in this fixture is simulated; this is not an original-game or actual browser-Worker qualification. Negative controls cover timeout, renderer/Presenter/main-instance replacement, wrong owner, lost context, executor generation, retirement, incomplete and stale frames. Closed chains make zero subsequent reads. The existing actual owning Worker CPU Flip, retained-copy, decoder, filtered compositor and native replacement controls also pass on the frozen qualified graph.

Run helpers reconstruct the qualified source graph in memory from Git plus the retained runtime diff, with the owned MCM adapter overlay. The exact retained module is 2420767 bytes, SHA-256 f5ae1ac84e3d5e13bb093aca53b8aa2fe18c000cc891587a5d75b46eb559b2ba. All 311 original media references were verified without copying media. The prepared original driver would arm synchronous source observation on the startup owner before its first reviewed capture or producing input command. This preparation was not executed against an original process.

The guarded flock /tmp/wine-build.lock bash tools/build.sh fails at tools/callback-ownership-census.js: host-file SHA/lifetime reviews disagree for tracked-clean lib/dll-loader.js, lib/guest-thread-host.js, lib/guest-worker.js, lib/ne-guest-bridge.js, lib/ne-lifecycle.js and lib/render-worker.js. Those producer files and the common Hex browser-input WIP were not edited. This required gate blocks committing, strict publication and the conditional affected original. Adapter, test and this note remain uncommitted WIP. Escape/current-display/ordinary Quit are unproved. No original was launched and no endpoint request was made to any previous closed scope.

Localhost box ID null, state NOT_CREATED, expiry null, stop receipt null. Root owns shutdown; this task joins its own supervisors and finite watchdogs.

Menu stage source arming, 2026-11-20

Evidence: scratch/runs/20261120-mcm-menu-stage-source-arm-quarry/. The fresh driver removes automatic startup arming. Ordinary hardware-change and profile dialogs, followed by quarry setup, advance through the existing reviewed capture/admission chain. armMenuSource requires an explicit fresh menu-before-start review. Its same page read records the actual legacy endpoint and thread-manager object tokens and six prerequisite predicates. An absent/closed endpoint, absent threads/native instance, wrong main owner, or existing observation refuses before the host arm call. A successful arm pins endpoint/threads through the host call and consumes the review, requiring a new capture before Start. The host boundary, readiness, input, retained source-copy and current-publication guards are unchanged.

Ten focused stage tests and 28 readiness regressions on the frozen graph pass. The unchanged prior native29/CPU source proofs are referenced by hash; no native calibration was repeated. Preflight recovered 183 source files, all 305 WASM imports, the exact f5ae module and 311 media references. The deploy-only build-info.js is explicitly optional in the frozen index and remains absent locally. The full locked build passes.

The sole affected original passed the hardware-change and profile dialogs, Single Player, Stunt Quarry, Practice, T-riffic selection and rider/bike setup. The actual guest Worker module was recovered and matched once. At the visible rider/bike Start menu, the exact same-read source prerequisites were: missingEndpoint=true; closedEndpoint, missingThreads, mainOwnerMismatch, missingNativeInstance and alreadyArmed all false. The owning legacy endpoint remained absent even at this menu. Arming refused before calling the host; Start was never dispatched. This is the first precise blocker, and the scope is permanently closed. No replay, later endpoint request, fake readiness, repaint or historical source promotion followed. Quarry loading, source copy/publication, Escape/current display and ordinary Quit/guest retirement remain unproved. Browser termination is cleanup only. After closure, an offline-only hardening check also pins the legacy endpoint and threads to the reviewed menu across the asynchronous live read. Its equal-looking endpoint replacement negative passes. This additional guard was not rerun against the original; its runtime-tested predecessor is retained as a patch with the preflight adapter hash. It follows the unchanged predicate that refused the actual absent endpoint.

All owned native/browser/watchdog/board participants physically joined with EOF, no tracked process leak, and the temporary profile was removed. The sampled /home floor stayed above 1,500,000,000 bytes. Localhost box ID null, state NOT_CREATED, expiry null, stop receipt null; root shutdown delegates this completed owned cleanup. Strict publication and the final own-path commit are recorded in this run's result; foreign held regions/maps remain untouched.

The follow-up evidence is scratch/runs/20261120-mcm-tested-first-present-affected-original/. It adopted the exact tested adapter/test patch SHA-256 1fc6dc5291c979d71813f7b673bcbc696f1a7b9792954312ea10946f936a3896, reusing the unchanged 29-test and native CPU qualification. Preflight verified the retained module, all 311 media paths, imports, and qualified arm/producer contracts before the fixed runtime budget. No archived required input was missing. The browser served the recovered frozen graph, not current producer or common Hex WIP.

One affected original was explicitly authorized before commit. The inherited driver called source arming before capture, which would bypass the new pending capture readiness. The executed run-local driver instead called the new startup capture first, then attempted source arming before any input. It recovered and matched the actual guest Worker's CDP module bytes once. The reviewed screenshot shows the real hardware-change startup dialog, not quarry loading. Its capture identity has no render endpoint; it does not exercise a pending GPS publication.

Automatic source arming at that startup checkpoint was premature. The actual host call threw live shared legacy render endpoint required, permanently closing the runtime scope before any input or successful source arm. The same-read exception proves the host guard !endpoint || endpoint.closed || !this.threadManager; it did not serialize those three individual values. No later request was made to fill that gap, and no individual missing value is asserted. This is a run-driver prerequisite blocker, not evidence that the first-PRESENT repair failed. The automatic startup-arm order is not qualified for future use. No replay occurred.

Quarry loading, owning begin/completed source copy, publication binding, Escape, current-display qualification, ordinary Quit/ExitProcess and guest retirement remain unproved. Browser shutdown and physical process joins are cleanup only. The callback census now passes. The first locked full build failed because the foreign-owned region-map mirror was stale. After the coordinator refreshed that mirror, this task reran the full guarded locked build and passed with the exact adopted adapter/test hashes. This worker changed no foreign file, hash or gate. The passing build permits the conditional own-path source commit and strict publication, recorded in the follow-up run's publication receipt. All owned native supervisors, browser, board watcher and independent watchdog joined with EOF and no recorded process leak; the temporary profile was removed. Box ID/state/expiry/stop remain null/NOT_CREATED/null/null; root retains shutdown ownership.