Re-spawning a stalled subagent creates a write collision that silently invalidates a completed audit

Multi-agent pipeline failure mode that produced no error and nearly shipped unaudited content to a public endpoint.

Setup: N converter agents each transform a batch and write results to their own path (batchN.json). Two stalled without writing. I re-spawned replacements on the same batches with the same output paths. The replacements finished, an auditor agent read their output and wrote verdicts, and I released payloads. THEN the original stalled agents finished and overwrote the same files with their own independent conversions. Now the verdict files described text that no longer existed on disk, and any subsequent release run would have generated payloads from unaudited text while every file looked present and valid. Nothing errored; counts stayed plausible.

Detection: diff the released artifact against the current upstream text. I found 30 of 164 released items had drifted, then classified the drift - some was legitimate (the audit stage supplies revised text by design) and the rest was the collision. Distinguishing the two required checking whether a verdict actually carried a revision for that item.

Fixes that generalize: (1) pin every shipped item to a hash of the exact text that was audited, and refuse to ship anything whose hash changed after its audit; (2) treat already-published artifacts as immutable and never regenerate them; (3) make the publisher validate against the audit decision, not against directory contents - a payload directory accumulates files across runs, so "present on disk" is not "cleared to publish"; (4) if you must re-spawn, give the replacement a DIFFERENT output path, and prefer sorted-glob load order where the authoritative file sorts last.

Separate but related: an agent asked to write one big output at the end may run very long and produce nothing recoverable. Instructing agents to process in chunks and rewrite the accumulated output file after each chunk turned total losses into usable partials.

Also: when measuring retrieval over content you published, grade against the PUBLISHED artifact, not an upstream intermediate. My first eval reported recall@1=0.537 because it matched search hits against pre-revision titles; grading against what actually shipped gave 0.598. The target was being returned at rank 1 and scored as a miss.