lean_closure.py: score the working tree's repo, merges and faster proofs

[alex]

Why

Other sessions used ci/lean_closure.py score after #1641 and found four ways it misled.

1. Wrong repository. It read the repository it was installed in, not the working directory's, so running main's copy from another worktree scored main.
2. Branches behind REF. It compared with REF itself, so a branch behind origin/main counted main's newer changes as its own. In one example, 670 modules looked changed instead of 5.
3. Merges. A module that merges others took only its own old time, and the other members' time disappeared. A run that rebuilt a merged-away member wasn't counted as rebuilding the merged module either. For one RSA merge plan the tool said −108 s on replayed runs; corrected, it's −31 s.
4. Faster proofs. Modules the branch changed kept REF's times, so a PR that makes proofs faster scored only as the cost of what it added. One such PR showed +7 s while cutting about 140 s of CPU.

The change

All in ci/lean_closure.py:

• Every command reads the working directory's repository by default, with --repo for another. The tool sets lean_shards.LEAN accordingly.

• score compares with git merge-base REF HEAD.

• score --merge NEW=A,B,… gives NEW the members' times together, less IMPORT_TIME for each member but one. A run's rebuilt members map to NEW. If NEW has a time from the branch (below), that time wins.

• Branch times for the modules whose sources differ from the fork point:

  • score --times FILE reads them from a JSON object of module names and seconds;
  • score --profile-branch reads them from the profile of the local build under lean/.lake, after building the branch.

  The changed modules left without a branch time are listed on stderr, since they keep REF's time (or the median, if new).

CLAUDE.md's paragraph on the tool mentions --merge and --profile-branch.

Tests

ci/test_check_lean_closure.py gains 7 tests:

• merged times;
• a run that rebuilt a merged module: before the fix it rebuilt nothing on the branch, after it rebuilds NEW;
• a branch time beats the merge sum;
• a faster proof shortens the replayed run;
• the changed modules of a git working tree, untracked ones included;
• the merge base of a branch behind REF;
• use_repo reading another repository's modules.

python3 -m unittest discover -s ci -p 'test_check_*.py': OK.

Checked by hand: run from the worktree of the closed #1640, the tool now scores that worktree's 5 changed modules, where before it scored 670.

🤖 Generated with Claude Code

https://claude.ai/code/session_01TyNFiywMtzf7eAdKcGMA5z

---

Generated by Claude Code