MCM synchronous source command adapter

Evidence: scratch/runs/20261120-mcm-synchronous-source-adapter-implementation/.

The preceding completed-source task launched only startup and closed at its disk floor before arming. This task changes source before any ordinary launch. The existing real command lifecycle is sufficient: host boundary admission installs SynchronousFlipProducer in the owning guest Worker; native 0x20008/0x20009 seal descriptor, serial, page generations and retained RGB565 pixels before the upload broker parks. The actual decoder already consumes that request through the synchronous publication join. No WAT, shared renderer or guest Worker implementation change is needed for that path.

tools/mcm-bound-input.cjs now exposes chain.armSynchronousSource(). It calls the production armD3dimPublicationObservation({synchronousSource:true}), rechecks the actual process/Worker identities, and requires a new reviewed capture afterward. Duplicate arming or owner replacement cannot admit input.

When the compositor selects the CPU replacement, the page reader uses tools/mcm-synchronous-source.cjs to require its own completed native source, descriptor/serial/command batch/Worker slot, decoder canvas proof and qualified current publication. The former GPS endpoint and its actual executor remain separately checked. No old GPS tag, full decoder extent, timestamp, repaint or unsealed canvas is accepted as authority. Capture age remains 12000 ms. The adapter conservatively refuses a subsequent repaint without a new qualified row; it does not promote a historical row to current publication.

The focused fixture compiles the native source and runs the actual owning Node Worker, native Flip, RGB565 seal, retained copy, host decoder and compositor. The pinned pre-change page adapter at 759080889 rejects that same native replacement; the new page expression accepts it. Wrong process owner, endpoint, replacement layer, stale publication, retirement, incomplete decoder and native partial HUD controls reject. The fixture also exercises a later genuine native replacement, concurrent native source rejection and producer closure. Mock arming controls supplement this proof. The 15 GPS/arming/freshness/observer regressions pass. The guarded locked build passes.

Original-game execution and strict publication receipts are recorded in the run result after the implementation checks. Prior quarry throttle/turn and foreground evidence is reused; no storage or signal attribution investigation is part of this change. Localhost box ID null, NOT_CREATED, expiry null, stop receipt null; root owns shutdown and this task cleans up its own process groups.

Commit 753deb61e passed the native fixture, 15 regressions and guarded locked build. Strict pull failed on foreign lib/region-map.generated.js and src/00-regions.wat; no push or foreign conflict resolution occurred. The affected original used the exact retained module and 311 verified media inputs. Actual Worker module bytes matched that module. Whole frames show the video-memory dialog and Enter Name with MCM typed. The screenshot main-menu still shows Enter Name. Two expired review leases refused before dispatch.

At 2026-11-20T09:49:02.326Z the independent outer watchdog sampled 1,485,811,712 free bytes and permanently closed the scope before source arming, quarry or Escape. No later endpoint request or original retry is allowed. Native supervisors report exit 143 with no leaks. Driver and inner supervisor/ watchdog have direct wait/EOF records; graceful Chrome/board waits were interrupted. Cleanup uses local process evidence, not a later browser request. Ordinary Quit and original current-display binding remain unproved.

After closure, read-only inspection of lib/gpu-stream-host.js lines 182–183 found a missing control in the earlier GPS fixture: the real GPS layer occupies both _gpuFrameLayer and _dxFrameLayer. The committed selection condition would send that GPU layer to the synchronous CPU validator. A one-line CPU-kind condition and a corrected GPS fixture are prepared as uncommitted WIP in tools/mcm-bound-input.cjs and test/test-mcm-current-gps-readiness.js. No tests or build were run after closure, so this corrected final tree is not qualified or published. The earlier native CPU proof remains valid for its tested version; it is not evidence that the corrected GPS branch passed. The precise remaining dependency is resource admission for testing/building that correction in a new task, followed by the original source arm/Escape/current-display/Quit sequence. No lease, timestamp or repaint workaround was introduced.

Dual-slot correction follow-up

Evidence: scratch/runs/20261120-mcm-final-dual-slot-source-quit/. The page selector now excludes GPU layers before invoking the completed native CPU-source validator. The GPS regression uses the production Presenter.present and Win98Renderer._overlayFrameLayer, so the actual presenter installs one GPU layer in both window slots. No hand-assigned GPU kind or alias creates the positive case. Against 753deb61e, this same fixture fails specifically in the CPU-source validator; the corrected selector admits its current GPS owner.

The real compiled native Flip/owning Worker fixture also passes on the corrected tree, including the native replacement, decoder, filtered compositor and pinned pre-adapter old-fail/new-pass proof. Wrong owner/endpoint, replaced layer/display, moved window, stale publication, retirement, incomplete decode, native partial HUD, concurrent source writing and a repaint without a new qualified row still reject. All 15 GPS/arming/freshness/scalar observer regressions pass. No source producer, WAT, historical-row promotion or forced repaint changes were made. The guarded locked build passes. The follow-up result records strict publication and whether the affected original became eligible for Escape and ordinary Quit.