Learn and Guard

Procedure

1. Add the recipe: class, root cause, correct fix, wrong fixes seen, tests, guard.
2. Build the strongest guard that fits: a shared helper or middleware (the right way becomes the easy way), a lint/semgrep rule, or a type that makes the bad state impossible.
3. Prove the guard: add a fresh instance of the class on a scratch branch; the guard must flag it. In our lab one guard could never fire until this test exposed it.
4. Give the tester the rule for its static scan and a seeded bug for its benchmark.
5. Track recurrence per class across runs; a recurring class means the guard is too weak.
6. Calibrate the fixer itself: run it on the seeded false-done benchmark after changing its instructions; the target is zero fake "done" claims.

Exit criteria

Library updated, the guard catches a fresh instance (evidence saved), and the tester has the new rule and seeded bug.

Counterpart of tester skill: qa-test-the-tester.