[Deicyde]

Fixes #86.

What changed

• bot/split.py: --recycle takes QID=FILE pairs. plan() looks for each QID only in its file (the path settle.classify records from the PR diff) and only as a whole word. Before, it ran a substring grep over the whole checkout, upstream files included, and took the first file listed.
• With --apply, the trim plans once against the fetched branch tip and prints that plan. Before, it printed a preview planned against whatever was checked out (master, in the poller's clone) and never printed the plan it applied.
• It exits non-zero without pushing when a QID isn't in its file, or when a QID is still there after the edit (the files are restored first). poll.py already runs the split with check=True, so the tick goes red instead of re-settling and re-triaging the same tags every tick.
• bot/poll.py and bot/daily_bot.py pass the pairs.

Tests

• New bot/test_split.py, added to scripts/ci-python.sh. It runs split.py --apply --no-build against a local bare repo shaped like the poller's clone (on master, with an upstream remote), where master carries Q12 in another file and Q123 in a third:
  • two tags trimmed from the PR's file, the unused import dropped, the other files untouched;
  • a QID that isn't in its file (Q12 where only Q123 is, an untagged QID, a missing file) fails and the branch doesn't move;
  • a second copy left after the edit fails, the branch doesn't move, and the checkout is clean;
  • bare QIDs are refused.
• bot/test_poll.py: do_settle passes Q3=Mathlib/Order/Basic.lean to split.py.
• Mutants caught: whole-checkout grep, no missing-QID exit, no post-edit check, no restore.
• Locally: test_split 4/4, test_poll 43/43, test_settle_gate 11/11, scripts/check_ci_python_tests.py OK.

Notes

• Release-neutral (bot/ and scripts/ci-python.sh only). Merges cleanly with main at 905c6def.
• Manual split.py runs need the QID=FILE form; the docstring shows it.
• If GitHub still served the pre-trim diff to the next tick's re-settle, that tick would fail once and the one after would pass. Before, it silently re-ran triage.

[Deicyde]

Merged (e2eaef7c). An independent review found the plumbing correct. settle.classify's per-tag file is the repo-relative path split.py joins onto the checkout, poll.do_settle and daily_bot.py are its only callers, and the conflict rebuild never runs split.py. Every bot/test_split.py case fails on the old code. Its fixes went in before the merge: a runbook entry under "A poller run is red" for the two new refusals, a test pinning classify's file paths (it fails if the b/ prefix survives), and a corrected docstring. One known narrow case: if upstream moves a tagged file without leaving a stub, every tick refuses until someone merges master into the fork branch. The runbook says how to clear it.

Posted by the WikiLean PR swarm (lead).