3.3 KiB
Same-review repair: completed-epoch observation coverage
MeltAssist's actual review SHAfe873785b734467decd69ac1947c05b64a610361dd6f7f2226ff14964eccfa45 blocked original v2c76719ff/95ff82b9 solely on V2-EPOCH-OBSERVATION-COVERAGE. Budget, total projection, actual40000 safeguard, old-cost accounting, active-unmeasured reserve and scientific recipe passed. Review evidence SHA219bc5bb74575ae880803a89bf58bb1ae65dd55dc618c3710680d9d8af99371c and reproduction SHAace732bf9beaece291a56b342739a6c8e9e528b27c889132baacf0a8a38819cb were independently checked. This is a code-review finding, not automatic-approval rejection or scientific failure.
The original BLOCKED, exact56-file implementation, contract/proof/binding and reproduction are preserved byte-for-byte under artifacts/phase3-seed1-v2-observation-fix-2026-11-18. Original local Git commit and remote publicationbf7363bf remain. No v2 training/started/launch existed before repair.
Only three code files change: v2 observation helper, queue integration and tests. Common budget/project/dispatch helpers, worker, guard, source dataset pins, scientific comparison and costs are unchanged. Observation coverage uses a separate per-fold set, not the last projected epoch. Every poll accounts for all unseen completed-epoch receipts; collect drains before deleting a terminal fold. Global-stop and exception cleanup drain before termination/removal, with explicit coverage-error evidence if malformed input prevents coverage. The latest-epoch cost projection cursor and formula stay unchanged.
Records are keyed by completed epoch and hash-bind its receipt. Terminal, stopping, multi-receipt backlog and receipts older than5s produce OBSERVATION_INCOMPLETE with a specific cause, actual record timestamp, receipt mtime/hash, no GPU query and null snapshot times. This5s distinction is observation classification only, not a compute budget or stopping threshold. Fresh live singleton receipts get a current bounded GPU snapshot with actual start/end timestamps. If the assigned worker disappears between poll and snapshot, the snapshot is explicitly incomplete; current observations never claim historical occupancy. Existing observation records are checked and retained, not overwritten; repeated scans do not duplicate them. Foreign processes are only observed and never modified.
The two exact reported failure modes are exercised using the coordinator's actual observe/collect/poll AST fragments: terminal-before-poll preserves epoch200 before active deletion; a backlog199/200 produces both records. Added cases cover global-stop, delayed single receipt, fresh snapshot timing/hash and worker disappearance during observation. Original12 budget/science checks plus6 coverage cases =18CPU checks PASS. Initial17-check pass is also retained; the final extra race case and classification correction motivated the18-check rerun. No actual GPU query, labels, model arrays or scientific workload was used.
The new binding is submitted for the same narrow review. The old BLOCKED is not a usable launch authorization. No new main review, scientific computation, Track A changes or modification of either old seed1 STOP is implied. Actual new exact PASS is required before any v2 launch. Slow epoch34's cause remains unproven; previous stopped-coordinator CPU gaps remain unfilled.