MCM recovered-space attempt — 2026-11-07

The first blocker was a run-specific cleanup/retrieval harness error, not an emulator defect. Race movement and steering remain untested. Evidence is in scratch/runs/20261107-mcm-recovered-space-race/.

Before allocation, the actual production direct-initializer IIFE passed seven generated controls, including filtered-out mcm, delayed restoration, cancellation and explicit shell.launchApp arguments. The real uppercase keyboard helper typed RIDER, balanced down/up per character, followed by ordinary Enter. Invalid mixed-case, empty, numeric and oversized names emitted zero events. Generated production renderer controls and the extracted browser executor exercised the intended literal menu points, click/doubleclick, capture, Enter, throttle, separate steering, durations and quit arguments. Atomic transport, UNKNOWN/no-replay and invalid-duration controls passed.

Fresh boat bx_957bs67f was created at 02:44:17 UTC after persisting and fsyncing a UUID Idempotency-Key. On the boat, actual Chrome DOM events from the same typing helper were trusted: five keydown/keypress/keyup triples for RIDER and a final Enter down/up pair. Invalid names left that event list unchanged. Native Surface4/7 COM/vtable-tail checks passed against a remotely prepared, then consumed fixture module. No local guest runtime or build ran.

The generated dx_boids direct route proved actual guest Worker and WebGL renderer ownership. Renderer target B98A37E59DBEF996C6AB99058DA0435F, session 798DCC70EEA3B74B42EB8ADDBC752B5C, owned legacy/WebGL endpoint 1. Its completed fences advanced 42 → 50, control error 0, adapter present, endpoint failure null, guest fallback count 0 and completed software draws 0. The renderer reported ANGLE / Mesa Intel UHD Graphics 620, WebGL 2.0. I personally reviewed the fresh sample scene. I also reviewed the generated native COMCTL property sheet and its disappearance after an ordinary balanced OK click at (173,418). Both generated browsers closed normally before the original.

At approximately 02:46:09 UTC, generated-finish.py successfully completed its ordinary quit, then tried to retrieve native-ok.png using the already-used local evidence label generated-ok-pin. The exclusive evidence-file create raised FileExistsError before another remote request. That image had already been retrieved and personally reviewed. This is the first terminal blocker. The shell sequence incorrectly lacked fail-fast handling and continued through pin verification to original launch. This continuation was a worker workflow error; no retry or repair was attempted after noticing it.

Exactly one original browser was created, at 02:46:09.435 UTC; production direct launch ?debug&threads&d3d-renderer=webgl&app=mcm was recorded at 02:46:13.464. No original mouse, name, menu or driving input was sent. Its only queued command was ordinary quit, acknowledged successfully. The retrieved startup screenshot was personally reviewed and shows the video-memory prompt. Original guest Worker execution is observed, but no original renderer endpoint exists at that checkpoint. Enter Name, profile creation, Single Player, track/rider, Start, settled throttle and steering were not reached. No emulator defect, audio, FPS, full-corpus or parity claim follows.

Staging records main 615a5bed188ce4151f22e492350de38e45221ac4, the relevant and all-tracked dirty patches, and the untracked wininit-boot source patch. All 938 staged pins matched remotely before the original and at cleanup. The actual guest and generated renderer Worker bytes were captured through CDP: standard module SHA-256 07f1cc970038f8ab31313b8b200edf26caa63473f293f5fc937f5cfe5c5c7f6e, 2,181,179 bytes, equal to the staged module. This is executed-byte identity, not an assertion that this module was freshly built from that staging HEAD. All 311 registered original inputs, 45,460,823 bytes, match before/after; MCM.EXE remains f0370d1d526be46eaf6ea5f941f185a393a8432baec695e1529f8672a314a71b. Guest media/config/save, installation-root VFS mapping, working directory and public catalog were unchanged. No profile was created. All 1,491 complete successful HTTP responses match staged pins, with no unmapped success. Non-200 requests identify build-info, auth-user, favicon, sample/MCM icons and the optional font-bundle endpoint; individual fonts loaded successfully.

All 102 remote artifacts were retrieved or deduplicated against identical owned local bytes; retrieval-verification.json maps every remote path. Chrome exits were 0. All six tracked browser/controller PIDs, six ports, three private profiles and writer locks were independently verified absent. The literal pending: UNKNOWN; never replay in generic cleanup records is a static helper label, not an outstanding command: both actual quit commands have successful numbered ACKs. No gameplay command was issued.

Delegated normal stop stop_550f13be50ff0283 completed 02:47:05.032 UTC, forced false, error null. Boat state is archived with no active expiry. Allocation through retrieval and completed stop took about 168 seconds, within 35 minutes. Observed local disk measurements remained above the owner's 1.5 GB floor. No cybersecurity flag, held-risk retry, old-boat resume, nested worker, source/media archive or deployment occurred. The prerequisite for a future fresh assignment is removing the duplicate retrieval from the generated session closeout and making the sequence stop on the first local error.