LMarbles late-reply diagnostics and Puzzle8 attempt

Published d7a471f0a fixes the discarded-reply evidence gap. It does not establish a cause or repair for the earlier original Start delay. In the new diagnostic original, Start returned in 106 ms. Puzzle8 was entered through the ordinary level selector, nine personally chosen slides reduced Moves 96 to 87, and ordinary Quit completed with native exit 0. Clear, tally and onward remain unproved.

Evidence: scratch/runs/20261118-lmarbles-late-reply-diagnostics-puzzle8/, with result.json published last. The predecessor note, result and retained-cause audit were read. Its UNKNOWN scope remains permanently closed; missing historical raw data cannot be reconstructed.

Diagnostic records and acceptance

The controller writes a separate, exclusively created diagnostic file before checking transport success, strict confirmation time, native consumption or targeting acceptance. It contains the exact response bytes as base64, parsed raw transport reply, delivery and return times, operation/release bounds, and optional client socket/body/response/parse timestamps. Transport exceptions also get a diagnostic record. The independent release child preserves its own raw reply and timing before its confirmation check. Existing release-fired and actual-wait records supply the later release outcome.

Diagnostics explicitly say diagnostic-only and acceptedReceipt: false; they never authorize continuation. Accepted receipt/down publication now follows consumption/deadline/late-ACK validation. The return timestamp is sampled before diagnostic disk output, matching the previous dispatch-to-return measurement boundary. Diagnostic output time still counts against subsequent existing operation bounds. No guest call, warmup, clock, polling, hold, freshness or deadline extension was added. Published 4a20c54c9 narrow pre-body recovery and 1590 freshness behavior remain intact.

Nineteen real-controller/TCP cases pass. An authored client event-loop stall after a real CONSUMED response produces the exact 500ms confirmation exceeded branch: the raw reply remains on disk, no accepted receipt/down record appears, fallback UP completes, and later input is locally refused as sticky UNKNOWN without endpoint access. Fresh strict acceptance, expired socket/connection NOT_SENT followed by new capture/review/nonce, identity/review replacement, partial writes, exceptions after actual delivery, remote error forgery, lost ACK, UNKNOWN, partial effects, release errors/mismatch, held physical input and snapshot replacement remain covered.

Two finite actual-native SDL fixtures pass using the retained verified 0d383140f6d4207d18860e3cb22fc8fcc47a5d260b8d5f2443222396b245fdd5 module. The baseline's complete green surface was personally viewed. The late-return fixture consumes normally, then stalls only the authored client callback for 520 ms. Production diagnostics retain the CONSUMED reply and refuse it; matching fallback UP and sticky local closure remain mandatory. This proves diagnostics and rejection, not the historical Start cause. Both independent native supervisors report no leaks and the retained nested waits record both EOFs. Startup, sealed-resolution and guarded-input focused checks and the required locked build pass. Earlier focused invocations with nonexistent filenames and then an omitted fixture environment failed before the corrected final checks; their logs are retained separately.

Diagnostic original and limits

One preparation was refused before native/watchdog creation because this worker supplied a shorter completion recipe. It remains closed. A separate diagnostic original restored the existing required 3600-second/300-second-reserve completion contract; no supervisor rule was changed. Its 251 retained runtime source hashes and existing compressed module were verified, and the published controller diagnostics plus existing server phase trace were exercised. No old runtime was queried.

New Game, Start, ordinary Tier1/Puzzle8 selection and confirmation succeeded. Start's body attempt was at 1795008864583 ms, native consumption at 1795008864678, response end at 1795008864682 and controller return at 1795008864683, from dispatch at 1795008864577. These successful timings do not explain the earlier failure or justify a speculative runtime change.

All gameplay decisions came from personally viewed whole boards. Nine slides reached Moves 87. Numerous subsequent intents were locally NOT_SENT because the manual capture-to-command round trip expired the five-second game review or left less than the required 500 ms targeting budget. Each new attempt used a fresh capture/review/nonce; no late ACK or UNKNOWN recovery was accepted. Whole-desktop reductions were tried for viewing; retained native PNGs were unmodified. There was no solver, autoplay, OCR, board-memory read, saved-state, configuration, game-media or clock shortcut. The unsolved attempt ended through Escape, the visible Y prompt, then setup Escape. Matching UP ACKs preceded native exit 0; native, watchdog, outer child and outer parent waits are exit 0 with both EOFs. No cancellation was requested in this original, and no endpoint was called after native exit.

The enclosing tool returned 143 despite the separate supervisor's exit-0/no-leak receipt. The enclosing runner's final disk sampler and holder-close output were not flushed; no successful enclosing tool wait or continuous original disk-floor proof is claimed. Available-space observations remained above 1.5 GB at recorded original/task milestones and the native-fixture samples; missing memory-only samples remain unknown. No foreign cleanup or RAM gate was used. Redundant review/transition PNGs were pruned with a hash inventory; selected whole screenshots, raw replies, timing records, exact Git/patch recovery and unique fixture inputs remain.

SuperTux's supervisor/sealed-resolution ownership and SDK/ReVolt helpers/holds were preserved. Controller/common/test ownership was released after the original; any later shared edits are outside this tested change. No workers or boxes were created. Box ID, expiry and stop receipt are null; state is NOT_CREATED. The old DirectX pixel-oracle failure remains separate and unfixed.