LMarbles Puzzle8 prelaunch investigation, 2026-11-18 12:27

No fresh original was launched. Puzzle8 clear, onward tally/Puzzle9 and ordinary Quit remain unproved in this task. The earlier 31-slide attempt and its ordinary exit remain separate historical evidence; no prior endpoint, saved board or move sequence was reused.

Evidence is scratch/runs/20261118-lmarbles-current-manual-puzzle8-progress-1227/. The bounded actual-disk sampler recorded 597 samples over 60 seconds, from 1,522,176,000 bytes available to 1,511,329,792 bytes, its minimum. This interval did not itself breach the 1.5 GB floor. It leaves only 11,329,792 bytes at minimum, following the unresolved transient floor breaches in the prior Quake II/GLTron tasks. This is a decision to avoid an unsafe original under the existing disk rule, not a new resource gate, a RAM test or proof of a game defect. No foreign cleanup or writer attribution was attempted.

Instead of another resource-closed original campaign, two small existing JavaScript regressions ran under tools/native-supervise.cjs. The migrated LMarbles supervisor/source view passed three actual cold ProcessDesktop authority/client Worker handshakes and closure, and missing/tampered index/source/module controls. Sealed resolution passed normalized imports, absent-live dependencies, JSON/package main resolution, live-file replacement and outside-inventory refusal. Both test processes exited 0; the supervision receipts report no leaks. These tests establish that the earlier indexed-source startup repair still works for these fixtures. They do not establish original gameplay, module compatibility or a visible Puzzle8 consequence.

A focused authored HTTP regression also reproduced a transport limitation relevant to the required send-boundary freshness. lmarbles-controller.cjs checks freshness before calling native-supervisor-common.cjs's postJSON. That transport immediately queues req.end(data) without a connected-socket check. With an authored 120 ms socket-assignment delay and a guard initially 40 ms from expiry, the only request arrived 103 ms after its guard deadline. The authored endpoint returned UNKNOWN with consumed:false; the fixture made no further request. This proves that the pre-call check cannot guarantee that an expired request body never reaches the endpoint. It does not prove that the real native guard accepts expired input, or that this caused any historical failure. No original or native-game input was involved. send-boundary.json retains the actual timestamps and reply.

No production code changed, so no new production build or compilation was run. The current build hash is recorded as an inventory item only, not a verified matching runtime module. No module, media, source archive, frame series or screenshots were copied. Exact Git, relevant binary diff, helper patches, hashes, recovery recipe, separate stream EOF waits and process audit are retained; result.json is written last. All three focused regression supervisors exited 0 and recorded leaked:false. The complete invoking tool's separate stream EOF is not exposed; subordinate command waits record both streams.

No controller repair or safe original is claimed. Further work requires repairing and testing the connected-send freshness path, plus sufficient actual disk headroom for a finite original and its Quit reserve. Renderer/admission code owned by the Quake II/ReVolt lanes and held Blood/Sims/foreign WIP were preserved. No workers, boxes, Python, Wine or deployment were used. Box ID null, state NOT_CREATED, expiry null, stop receipt null.