Genuine MCM GPS producer observation

Run: scratch/runs/20261120-mcm-real-gps-producer-source-observation/.

MCM D3DIM opens its real d3dim-gps endpoint through the process host. It does not create a legacy software-GL endpoint. The new explicit prepareD3dimGpsPublicationObservation installs guest writer participation and the synchronous native Flip observer at an ordinary scheduler boundary. The returned registration is PREPARED, with no endpoint or publication join.

The actual endpoint factory awaits the separate gps-producer-observation handshake before returning its real port to the bridge's first kick. That handshake requires D3DIM GPS, a shared writer epoch, a bounded source journal and retained-source storage. The legacy protocol still rejects GPS. No surrogate endpoint is created. Message dispatch, doorbell drains and native transient resets participate in installed render writer coverage.

Preparation requires an idle writer boundary. Binding occurs inside an already observed guest command and validates its captured participants and unpoisoned epoch; it does not demand that this command has already returned. This is installation, not permission to read source pixels. The existing sole-writer native Flip seal, page leases, retained copy, complete decoder, compositor and publication checks remain authoritative.

ARMED requires the real endpoint acknowledgment and owner generation. PUBLISHED requires a qualified synchronous publication row. GPS ready, commands, frames and endpoint acknowledgment alone create no CPU source authority. Replacement, cancellation, timeout, endpoint error and process or guest retirement invalidate the epoch, retire joins, clear timers, remove owned subscriptions and settle cleanup outcomes. Reinitialization awaits only an existing registration, preserving ordinary unarmed startup timing.

mcm-bound-input.cjs selects this path only with the explicit gpsSourcePreparation option. Its default legacy behavior remains covered. The final offline guard also closes a chain on a CLOSED preparation before any further Worker reads or dispatch. It was not replayed on the original.

The actual browser fixture uses the qualified graph plus the owned overlay, real CPU and render Workers, and compiled native code. The first D3DIM GPS command installs coverage and creates no legacy endpoint; the following native CPU Flip produces a retained RGB565 source. Full fallback decoding, composition and display publication reach PUBLISHED. Both actual Worker modules are recovered and hashed once. An earlier fixture configuration mistake selected the local executor; a subsequent implementation test exposed the erroneous idle-at-bind requirement. Neither failed attempt is accepted proof. The corrected fixture and exact final-source fixture pass.

Eight actual render Worker controls cover normal installation, wrong owner, replacement, cancellation, invalid journal, endpoint retirement, the preserved legacy-wire refusal followed by the new GPS protocol, and malformed legacy control. Ten lifecycle/admission controls, ten unchanged menu-stage controls and 32 existing render-manager cleanup checks pass. The owned host/render callback review remains anchored to the preceding reviewed hashes. The full flock build passes; retained older native29/CPU proofs are reused only where the exact input/source hashes match, never as qualification of the removed legacy hook.

The sole affected original used exact module f5ae1ac84e3d5e13bb093aca53b8aa2fe18c000cc891587a5d75b46eb559b2ba (2,420,767 bytes), 311 verified media references and the qualified 183-file graph plus the owned overlay. It passed ordinary hardware/name dialogs and quarry setup. Preparation returned PREPARED with endpoint null. The fresh post-prepare capture visibly reached rider/bike selection. The submitted Start review was 17.8 seconds old, exceeding the unchanged 12-second limit; admission refused it before any Start effects. The scope was permanently closed without replay or later ordinary endpoint requests. The pre-arm review text incorrectly described rider/bike selection while its capture still showed the preceding track-selection transition; only the subsequent post-prepare capture proves the rider/bike screen.

Quarry gameplay, original source publication, Escape, current display and ordinary Quit remain unproved. This is the precise first original blocker, not a successful gameplay claim. Final offline admission/reset review changes have native/focused coverage but no second original run. All owned runtime processes were joined by native-supervise and independent finite watchdogs; no remote box was created. Evidence records exact runtime and final source recovery separately, module/media references, diffs and cleanup receipts.