closed-board-row-reopened-by-cron-tasks-jsonl-collapse

Answer

Solution written to /workspace/solution.md (273 lines), with runnable artifacts in /workspace/sol/.

Summary

Root cause (two compounding defects): 1. Destructive bookkeeping — qa-cron does read → filter → open(path,"w"), keeping only the current tick's pending rows. This is a truncate, not a merge, so terminal statuses (closed/passed/failed/wontfix) from earlier ticks are wiped (and a crash mid-write truncates the whole board). 2. No pre-dispatch dedup — the foreman trusts the board and re-dispatches a resurrected row even when the fix is already at HEAD and already graded.

Fix (3 parts + backfill): - 2a Idempotent merge keyed by row id, terminal status always wins; atomic tmp+os.replace under an fcntl lock. - 2b already_landed() guard: git log --all --grep=<row-id> and jq verdict .passed (live curl probe as corroboration only). - 2c On a hit: zero new commits, re-close with a duplicate-premise note, enqueue only a fresh tier2 re-grade. - 2d One-shot backfill that reconstructs terminal rows from verdicts, with the caveat to ship 2a first.

Verified by an executable harness (python3 verify.py) — all 8 checks pass: - bug reproduced (closed row lost/reverted) - merge is idempotent and terminal-safe - guard skips dispatch, makes zero new commits, re-closes with the note, queues tier2 - no false positives (unknown rows still dispatch) - 20 concurrent ticks lose no rows

The markdown also includes a copy-paste single-file self-test (SELFTEST OK) and manual git rev-parse HEAD / jq acceptance commands matching the required bar.