LMarbles Puzzle8 attempt after headroom recovery — 2026-11-18
The single authorized fresh original reached Normal/Original Puzzle8 and accepted 31 manual slides (96 to 65 moves remaining), but did not clear. Native tally and Puzzle9 remain unproved. The worker ended the attempt through the ordinary quit prompt and setup-menu Escape; native exit code was 0. This is an incomplete gameplay result, not a successful clear or an established engine defect.
Evidence: scratch/runs/20261118-lmarbles-recovered-headroom-puzzle8-clear/result.json, written after publication. image-selection.json names nine retained whole PNGs, including level confirmation, initial Puzzle8, an intermediate gray row, the final unsolved board, quit prompt and setup. Other captures and resized derivatives were pruned with hash inventory. The previous 13-slide board was historical reference only; no saved progress, imported route, solver, OCR, board-memory inspection, autoplay, or game/config/clock/pump edits were used.
The manual route assembled a gray row, then mishandled the directional-arrow constraint while trying to use gray stoppers. Recovery did not complete the pattern. Review latency repeatedly exceeded the unchanged five-second gameplay freshness or remaining 500 ms admission budget: 42 local NOT_SENT admissions and 31 differing whole-image refresh comparisons were recorded. Fourteen exact whole-frame refreshes could reuse personally reviewed images before one chosen action. These checks did not relax targeting or execute a queued route. The worker voluntarily quit before the deadline; neither review latency nor the unsolved route proves an engine impossibility or a hard runtime blocker. A new original requires new authorization because this scope is closed.
All 52 dispatched DOWN operations were CONSUMED and all 52 independent UPs ACKed. Maximum dispatch-to-UP was 501.000166 ms; maximum complete operation was 1.507 seconds. No three-second operation timeout or actual operation UNKNOWN occurred. After ordinary native exit, main-quit capture failed the local native-terminal check before contacting the endpoint. No later endpoint request or replay followed.
The independent watchdog was prearmed. Its 35,920 samples recorded a minimum of 2,010,533,888 bytes free, above the 1.5 GB floor. It did not fire; its explicit closing reason was actual-native-wait. No RAM gate was used. Native, watchdog, inner supervisor and outer parent receipts show actual exit-0 waits and separate stdout/stderr EOF. Independent release-child waits likewise recorded both EOFs. The invoking tool exposes combined output only: separate outer EOF for the holder and every one-shot helper is not established. See tool-waits.json and lifecycle-audit.json; this prevents claiming the complete requested lifecycle proof. A holder SIGTERM was observed with unknown sender; descriptors survived and were hash-verified before explicit closure. Parent signal observations are retained, without assigning a sender. The native log also records a held-critical-section retirement warning; exit 0 does not establish general engine correctness.
The runtime reused source Git a00eb35bebb621ebb3f2b42b6af18f1bd874b9f6 with the published ad27 loader overlay and exact module SHA-256 83703369cc49c3e16077a4abd43def27262487f83dd52583d1b433adf7095f5f. Git, overlay diffs, source inventory and compile recipe replace source archives. The sealed module/source descriptors were checked before closure. All 88 referenced media entries remained hash-identical; the focused media verifier passed (86 original client files, authentic CRT and mappings). Prior native/pixel/lifecycle and controller fullqueuedroutes/build proofs are hash-referenced in reused-proofs.json, not broadly rerun.
A JavaScript-only owner policy arrived after the original and its holder had launched. No further Python invocation occurred: the published JavaScript controller was used against the existing owner/lease, without relaunching the original. Pre-policy local helper bytes were retained as patches and removed after their processes ended. Subsequent helpers were JavaScript and stayed in ignored evidence. No production code changed, so no new production tests or locked build are claimed; the retained published proofs were reused as instructed. This commit contains this report only.
All known owned runtime, watchdog, release-child and holder PIDs were absent at audit; the control listener was absent. No box was created: ID and expiry null, state NOT_CREATED, stop receipt not applicable. Other owners and investigations were preserved. Remaining work is an actual Puzzle8 clear, tally and Puzzle9 in a separately authorized fresh attempt; this attempt cannot be resumed or replayed.