A pairwise git merge-tree sweep of the 23 open PRs on 2026-09-10 (253 pairs examined, legacy three-argument form with a self-merge negative control) found 231 conflicting pairs: every pair of content PRs collides in references.bib, 33 pairs also collide in inst/WORDLIST, and one pair (#219 x #231) collides in chapters/coding-agents.qmd.
The cause is structural, not any one PR: every fragment appends its bibliography entries at the end of references.bib and inserts wordlist entries into the same sorted file, so any two PRs edit the same hunk. After each merge, every remaining PR needs a sync from main and a hand resolution whose correct answer is always "keep both".
Options, cheapest first:
- Add a deterministic resolver script (
scripts/resolve-shared-files.py) that, on a conflicted references.bib, takes the union of @entry{key,...} blocks (failing on a duplicate key with differing bodies) and, on a conflicted inst/WORDLIST, takes the sorted union of lines, so a sync-with-main round is one command.
- Declare
inst/WORDLIST merge=union in .gitattributes (duplicate words are harmless there). Do not do this for references.bib: two PRs adding the same key would merge silently into a duplicate-key bibliography.
- Longer term, split
references.bib per chapter or per fragment so unrelated PRs stop sharing a file.
The sweep script that found this is small enough to add as scripts/pr-overlap-sweep.py; it must use the three-argument merge-tree form on this machine (git 2.37 has no --write-tree), and it reports pairs examined alongside pairs conflicting so a detector that never ran cannot read as a clean matrix.
A pairwise
git merge-treesweep of the 23 open PRs on 2026-09-10 (253 pairs examined, legacy three-argument form with a self-merge negative control) found 231 conflicting pairs: every pair of content PRs collides inreferences.bib, 33 pairs also collide ininst/WORDLIST, and one pair (#219 x #231) collides inchapters/coding-agents.qmd.The cause is structural, not any one PR: every fragment appends its bibliography entries at the end of
references.biband inserts wordlist entries into the same sorted file, so any two PRs edit the same hunk. After each merge, every remaining PR needs a sync frommainand a hand resolution whose correct answer is always "keep both".Options, cheapest first:
scripts/resolve-shared-files.py) that, on a conflictedreferences.bib, takes the union of@entry{key,...}blocks (failing on a duplicate key with differing bodies) and, on a conflictedinst/WORDLIST, takes the sorted union of lines, so a sync-with-main round is one command.inst/WORDLIST merge=unionin.gitattributes(duplicate words are harmless there). Do not do this forreferences.bib: two PRs adding the same key would merge silently into a duplicate-key bibliography.references.bibper chapter or per fragment so unrelated PRs stop sharing a file.The sweep script that found this is small enough to add as
scripts/pr-overlap-sweep.py; it must use the three-argumentmerge-treeform on this machine (git 2.37 has no--write-tree), and it reports pairs examined alongside pairs conflicting so a detector that never ran cannot read as a clean matrix.