LMarbles Start confirmation cause investigation
The original Start failure remains UNKNOWN. No runtime or controller cause fix is justified by the retained evidence, and no new original-game scope was opened. Puzzle8 entry, manual clear, onward and ordinary Quit remain unverified. Published 4a20c54c9 pre-body NOT_SENT recovery and its 18 controller cases, two native cases and passing locked build are reused; this task changes no production code. Existing deadlines and sticky closure are preserved.
Evidence: scratch/runs/20261118-lmarbles-start-confirmation-cause-manual-puzzle8/. The predecessor note, sealed result and Start timing record were reviewed; 18 selected retained records have fresh hashes in retained-cause-audit.json. Start dispatched at 1795007204341 ms. Its native stdout contains DOWN followed by UP, and the matching fallback release ACK arrived at 1795007205789 ms. These observations cannot certify NOT_SENT. The terminal record at 1795007205939 ms follows the independent release-child wait; it is not the raw reply timestamp.
The exact error originates after the awaited HTTP call returns a parsed successful transport result and before the controller saves its receipt. Therefore the old Start call returned after the 500 ms elapsed bound, but its raw receipt, native consumption timestamp, socket/body timings and server command phases were discarded or never recorded. Native totals (6940 batches over 49.175 seconds) cannot identify this particular input's queue or consumption latency. Current offline listener scans measured roughly 0.5–6.5 ms; they do not establish historical cost. Queue, body delivery, driver/native processing and client scheduling remain unresolved. Increasing deadlines, accepting a late ACK or changing scheduling without attribution would not be a supported repair.
Two finite authored SDL native sessions exercise the actual production controller, HTTP transport, guard, public native-input-return boundary and independent release process. They use the verified 0d383140f6d4207d18860e3cb22fc8fcc47a5d260b8d5f2443222396b245fdd5 module and supervisor-sealed source inventories. The diagnostic records client socket/body/raw-reply phases and native driver/HTTP phases without reading guest memory. The baseline consumed normally and its complete green surface was personally viewed. The delayed case injects 650 ms solely between the actual native consumed receipt and the HTTP reply, preserving cancellation forwarding.
In the delayed case, dispatch was 1795008136534 ms, body attempt 6539, native handler application 6567, native consumption 6604, and driver receipt 6615. The client timed out at 7051 (517 ms after dispatch); the delayed ACK was released at 7267. The independent fallback release reached native at 8067 and received matching UP ACK. There was exactly one DOWN, no NOT_SENT reclassification, no accepted DOWN receipt, no replay and no later controller endpoint request after closure. Local opened() refusal was also checked without any endpoint access. This is a meaningful delayed native ACK rejection test, not a reproduction or attribution of the old 500ms confirmation exceeded error: the injected case reports timed out.
Both authored diagnostic assertions pass. Native, watchdog, outer child, outer parent, release child and enclosing tool waits are exit 0 with both EOFs. The two independent finite native supervisors report no leak. These are local cancellations, not ordinary guest Quit. The baseline client observer emitted one MaxListenersExceededWarning because it registered connect observers on reused sockets; this does not change the recorded protocol outcome. No production repair or fresh browser qualification is claimed. The old DirectX pixel-oracle FAIL remains separate and unfixed.
The predecessor original's 251 source hashes still match. The new authored sessions have their own coherent source inventories at 72cd365561b20e2f182bb37fa46146106f88b179; they are recorded separately from original provenance. Recovery uses Git, relevant binary diffs, the small authored diagnostic script, existing fixture generation and retained module paths/hashes; there are no source snapshots or game-media copies.
Initial offline /home observations were below 1.5 GB. Native tests began only after the separately owned capacity recovery restored space; their sampled minima were 1626304512 and 1624899584 bytes. No task-wide floor compliance is claimed. Common helpers, controller, SDK/ReVolt work, foreign WIP and holds were untouched. No original gameplay, solver, autoplay, OCR, board-memory read, save/config/media/clock/polling/focus modification, worker, box, Python, Wine or deployment was used. Box ID, expiry and stop receipt are null; state is NOT_CREATED. Ownership is released after the report and evidence are sealed.