Prove the edit, then stop

Run from /testbed. Quote paths. Never use a network install if pytest is missing: python -c importing the package is enough to finish the repro, and the grading run uses the suite itself.

After each edit

1. git diff --stat and then git diff on the file you touched. The diff should be a few lines. If it is a whole function, you went too far; restore and redo the smallest change.
2. Run the single test file from tree-layout for this repository, with -q --tb=line. If you know the test name from the traceback or from a grep of the symbol under tests/, add -k with that name first.
3. Read only the first failure. A failure in a test you meant to fix means the edit is incomplete. A failure in a different test means you broke a kept behaviour: revert that hunk.

Revert is the safe stop

git checkout -- path restores one file. Use it the moment a kept test fails and you cannot see a one-line correction. Then try the other candidate you already named, or leave the tree clean.

Done

Stop when the tests that mention your symbols pass, git diff contains only intended lines, and you have not edited anything under tests/. Do not run the entire repository suite if the targeted file is green and the budget is thin. Do not open another skill. Do not re-read files you have already used.