MCM native replacement writer after Escape

Evidence: scratch/runs/20261120-mcm-native-replacement-publication-quit/. The preceding same-read Escape evidence and appnotes were read. One new diagnostic followed ordinary startup, name, Single Player Event, Stunt Quarry, Practice and track selection into the quarry, using the retained exact f5ae1ac84e3d5e13bb093aca53b8aa2fe18c000cc891587a5d75b46eb559b2ba module (2,420,767 bytes), reconstructed coherent source graph and 311 verified media references. Instrumentation changed only the served host-imports source in memory; its transform and source hashes are retained. No production source or shared renderer/region files changed.

Immediately before ordinary Escape, the driver armed the existing bounded scalar observer. Its 128-record limit was reached. Within that bounded trace, the page receives dx_trace(5,37,16,475021312,0) and an upload of presentation 0x200025, rectangle (0,0)-(640,480). The actual WineAssembly._presentDxIfDirty policy then selects surface 37. The real DIB decoder writes an unclipped 640x480 RGB565 replacement layer, pitch 1280, without a D3D9 window or gamma transform. The same-read failure reports the old GPS write sequence 187778 and replacement DX sequence 187780. This is the same first decoder-copy sequence in the scalar trace. The GPS layer is stale, still equals _gpuFrameLayer, and no longer equals _dxFrameLayer; window, process and instance checks match. GPS reports 188 presents and stamp 31594.

This establishes the actual page-side replacement writer, beyond the prior source-reachability hypothesis. It does not establish a complete native source publication. At that exact copy, ctx.d3dimPublicationJoin, read stamp, owed descriptor/serial/Worker slot, source version and canvas version are all absent. The observer's completeDecode:true means only full decoder extents; it does not certify source completion, writer exclusion or canonical display. Page ctx.instance identity is not an independent per-call guest actor proof. Later bounded copies alternate the two DIB addresses and also lack authority. No GDI-only, fake or unsupported replacement has been admitted.

The retained native contract explains the missing authority. The synchronous Flip path calls the default-off native begin hook 0x20008; only an armed producer causes the real completed hook 0x20009 and its descriptor/serial bracket around dx_present. SynchronousFlipProducer requires a sole owning command and seals source pages plus a retained copy. The host's armD3dimPublicationObservation({synchronousSource:true}) installs that observation at a real scheduler boundary. No such producer was armed in this diagnostic. Static existence of that implementation is not evidence that the observed native source completed under its contract. The bounded trace cannot prove the absence of events outside its 128 rows.

The precise remaining dependency is an actual owned begin/completed-source witness for this CPU transition, joined through its own decode and current canonical display. That must precede any adapter binding change and its real native old-fail/new-pass and ownership/stale/replacement/retirement/incomplete controls. Attaching an old GPS/canvas tag or accepting raw full-sized pixels would bypass that dependency. No production repair, new-pass, guarded build, code publication or affected-original Quit is claimed.

The first UNKNOWN permanently closed the diagnostic, with no later explicit endpoint request or replay. The independent Escape UP receipt records its actual outcome. Console events delivered during cleanup are passive delivery, not additional endpoint requests. Chrome, board watcher, driver and supervisors have direct wait/EOF evidence; native supervision reports no leaks. Local runtime watchdog samples stayed above 1.5 GB. A separate metadata monitor later ended with exit 143 and unknown signal sender; its final metadata interval is unobserved and is not credited as continuous coverage. No later runtime opened.

The labels/review text bike-selection and starting anticipated transition completion while their retained frames still showed track selection. The images are authoritative; those two text descriptions are not proof of a reviewed bike screen. The later whole quarry frame was inspected before Escape. The exact same-read screen remains terrain/rider and proves no Escape menu. Prior throttle/turn, input, causal and process-group proofs are reused unchanged. Owned storage attribution remains UNKNOWN; no storage attribution campaign ran.

Localhost box ID is null, state NOT_CREATED, expiry null, stop receipt null; root retains shutdown ownership. No workers, Python, Wine or deployment ran.