Over-provisioning with spares needs a write-back, not just a log line
A pipeline selected 50 target items plus 25 spares from a committed manifest, then warmed each one and substituted a same-bucket spare whenever a target failed. The substitution logic was correct and the run printed every swap it made. The manifest was never updated.
Everything downstream read the manifest. So after a run with 15 substitutions, the packaging step built its archive from the original 50 target ids: it shipped the 15 items that had failed and omitted the 15 that actually worked. The health check reported the failed ids as missing forever. The demo UI would have offered a reviewer an item proven not to work. A whole-branch review caught it before the first real run; the per-task reviews had not, because within any single task the behavior looked right.
The general rule: if a process can change which items are the real set, the record of that set has to be updated by the process, not by the operator reading scrollback. Printing the substitution makes it visible; it does not make it true. Three checks worth running on any retry/fallback/replacement design: does any consumer read the pre-run list? can the post-run state be reconstructed if the terminal output is lost? and is there tooling to apply the correction, or does regeneration silently produce a different answer?
In our case regeneration was NOT idempotent (it re-selected from scratch against live state), so the printed list was the only record and nothing could consume it. The fix reused the existing manifest writer to swap whole rows, keeping each id paired with its own metadata.