rebuild_f5.sh cannot converge on a delete/modify conflict — rerere never records it nf-nggo ← Beads

Twinned from https://github.com/cfsh/ninefold/issues/465 by tools/beads/import_github.py's reconcile pass.

Hit while rebuilding f5/integration after pushing to #455. The script stops with:

``` ✗ UNRECORDED CONFLICT merging fix/stale-dispatch-corroborate: tools/pr/poll.sh Resolve, then: git add -A && git commit --no-edit (rerere records it) Then re-run this script. ```

Resolving and re-running does not help — it recurs identically every time. I resolved it three times.

Cause

The conflict is DU (deleted by us, modified by them): design 001's tooling migration deletes tools/pr/poll.sh, while fix/stale-dispatch-corroborate modifies it. Both are open PRs, so rebuild_f5.sh merges both.

git rerere does not record delete/modify conflicts — it only replays *content* conflicts within a file. So the script's core assumption ("resolve once, rerere replays it thereafter") does not hold for this conflict class, and the loop it prints is unreachable: every fresh rebuild re-hits it, and the resolution can never be pre-recorded.

The result is that f5/integration cannot be rebuilt at all until one of the two PRs merges — it is not a matter of resolving more carefully. origin/f5/integration currently sits at the last successful rebuild and is stale.

Worth considering

• Detect the class and say so. git diff --diff-filter=U plus a check for delete/modify would let the script print "rerere cannot record this class — resolve and force-push by hand, or skip this branch" instead of a loop that cannot terminate.
• rerere.autoUpdate / merge.conflictStyle do not help here; the only git-side lever is git config rerere.enabled + a *content* conflict.
• Skipping the branch (with a loud line naming what was dropped) may be the honest fallback — the branch is a throwaway test-drive aggregate, and losing one branch from it is much better than the whole thing going stale. CLAUDE.md's rule that a silent skip is unacceptable still holds; this would be a *loud* one.
• Cross-checking: the same shape will recur for every file design 001 deletes while another PR edits it, so this is likely to bite again during that migration rather than being a one-off.

Filed unclaimed. Neither colliding PR is mine — I only ran into it rebuilding the aggregate.