TDD

Generated tests that arrive with generated code prove only that the code matches the generator's understanding — which is exactly what is in question. Writing the test first, and watching it fail, is what turns a test into evidence.

Where it earns its cost

• Date/time and timezone math (DST transitions, "next business day")
• State machines and status transitions (who may move what, when)
• Policy functions: pricing, quotas, rate limits, permissions, scoring
• Parsers, serializers, and anything with an inverse to round-trip
• Concurrency decisions expressed as pure functions (poll intervals, staleness cutoffs, retry budgets)

Where it doesn't

Glue, wiring, UI layout, one-line delegations. Testing those is ceremony; /verify-live covers them.

The loop

1. Red — write one test for one behavior. Run it. Watch it fail, and read the failure: a test that fails for the wrong reason (import error, wrong fixture) is not yet a test. State what the failure proves.
2. Green — the minimal change that passes. Resist bundling the next behavior in.
3. Refactor — with everything green, remove the duplication you just created. Tests must not change meaning during refactor.
4. Repeat, one behavior per cycle. Commit at green boundaries (test: + feat: pairs, or one commit per cycle).

Design rules the loop enforces

• Injectable clock and randomness — any function that reads "now" takes it as a parameter (or a clock interface). Untestable time is where the worst silent bugs live; pin fixtures to hostile dates (DST boundaries, leap days, year ends, month-length edges).
• No network, no database, no filesystem in unit tests. If the logic can't be exercised without infrastructure, extract the decision into a pure function and test that; the wiring gets covered live.
• Boundary-first cases: empty, one, maximum, one-past-maximum, and the hostile middle (the 2:30 AM that doesn't exist on spring-forward day).

Honesty rules

• Never weaken an assertion to make it pass. If the test was wrong, say so and rewrite it; if the code is wrong, fix the code.
• Report results as they are: "12 pass, 1 fail" with the failure, not "mostly passing".
• A behavior discovered during /diagnose gets its regression test through this loop — red first, on the real reproduction.