MCM current GPS readiness and original startup boundary

The production MCM capture/admission adapter now identifies the actual d3dim-gps bridge endpoint instead of guessing that a sole render port owns the scene. Startup readiness waits for endpoint readiness and a current first PRESENT. Capture and admission independently recheck the actual bridge, port, window, layer, guest and renderer instances, recovered module inventories, renderer adapter, executor and context. A requested WebGL backend cannot stand in for an actual software fallback. Context loss, endpoint/ring errors, retirement, stale publication and replacement refuse admission.

The small shared change in lib/gpu-stream-render-worker.js exposes a passive identity accessor because actual executor/context objects otherwise live only inside its closure. It reads existing objects, counters and ring error state; it performs no drain, fence, export call, GL synchronization or scheduling change. The owning endpoint and instance scope its object tokens. Only immutable recovered CDP module bytes are cached. The 12,000-ms age gate, repeated live checks, independent release and stale pre-effect refusal remain in place. Identity acquisition failures now close the chain even for direct reads or captures; a stale pre-effect admission refusal still permits a new capture.

Evidence is scratch/runs/20261120-mcm-current-gps-readiness-quarry/. The supplied parent RECOVERY.md and the requested note named mcm-worker-service-latency-quarry-control-20261120.md are absent. The existing mcm-worker-service-latency-20261120.md, parent result, retained latency spans and November 7 notes were read. No latency fix, promise-flag A/B, artificial busy work or fixture qualification campaign was attempted.

Production-path result

The first finite Boids check caught the manager's port publication before the bridge's asynchronous open completed. It closed with no input; it was never resumed. The startup predicate was then made explicit. The fresh production check used unmodified HTTP-served runtime sources and actual guest/owning renderer Workers, each with the single executing module fab7c60e39c079c2972458d30c3a36143c6afd68c80f76d97e9e4a0c2f789d92. The legacy d3dStats.renderEndpointId remained absent (old readiness fails). The new full chain captured and admitted a personally reviewed whole frame at 6,985 ms (new readiness/admission passes), identifying an actual WebGL executor and live context. That callback sent no native input. The initially reviewed frame had aged out and correctly refused with zero effects before the fresh capture. This is an adapter repair, not a performance result.

production/boids-fresh.png shows the whole page and Boids2 window: dark background, blue wire boundaries, colored triangles and a partially visible textured globe. Publication/slice/doorbell counters advance naturally. No static reachability or requested backend is credited as exercised gameplay. The retained parent's 184-ms renderer / 182.6-ms drains / 166-ms guest sync / 32-ms live validation observations are reused; historical multi-second service latency was not reproduced or repaired here.

One fresh original

After full current admission and the final locked build passed, one untouched registered MCM original launched with all 311 existing input hashes verified and its normal working directory. The selected whole original/startup.png shows the video-memory test dialog. Full live admission accepted ordinary Enter at age 9,990 ms. Its independent UP began 102 ms after DOWN started; the completed hold receipt records acknowledged dispatches and passive samples.

The retained console then logs ExitProcess(1), followed by program exit. The next capture refused because no registered app remained. The runtime closed permanently; no replay or subsequent scene/guest/renderer request was made. There is no claim about the internal cause of that exit. This startup exit is the precise dependency before name, Single Player Event, Stunt Quarry, Practice, T-rific, yellow bike, Start, throttle/turn consequence or ordinary Quit can be established. No original renderer was created, so its actual renderer/executor identities remain absent. Browser termination is cleanup, not ordinary game Quit.

Validation, resources and recovery

The 12 focused tracked tests cover absent legacy readiness, software fallback, wrong/unbound/replaced owners, replaced executor, context loss, ring errors, missing modules, detached renderer and unchanged scene age. The actual doorbell retirement regression passes. The final required flock /tmp/wine-build.lock bash tools/build.sh passes.

The first build exited 0 but is excluded as a resource-valid build: six 100-ms samples breached the disk floor (minimum 1,446,711,296 bytes). Its run-local monitor recorded closure without enforcing cancellation; that monitor defect is explicitly retained. The corrected final monitor signals only its exact owned supervisor on a breach, and the final build uses compiler TMPDIR on tmpfs. Its samples and all three runtime scopes stayed above 1.5 GB. Sampling does not establish continuous free space. No foreign cleanup or RAM gate was used.

Finite native supervision, direct parent/child joins and honest stdout/stderr EOF receipts are retained. All tracked processes are nonlive, supervisors report no leaks, and owned temporary profiles are removed. Exact Git identity, relevant binary diff, owned new-file patches, source/media hash references and unique module recovery references are retained; no source tree snapshot, media archive or bytecode dump was made. Publication receipts record the actual commit and strict pull/push outcome. No agents, workers, Python, Wine, public deployment or box lifecycle action occurred: localhost box ID null, state NOT_CREATED, expiry null, stop receipt null. Coordinator shutdown ownership is unchanged.

The initial tool-owned board watcher returned exit 143; its signal sender and enclosing tool-pipe EOF are unknown. This is separate from the directly joined runtime/watchdog/board children whose pipe EOF is recorded in the run receipts.