test-quality

Deterministic unit tests inject time, randomness, and IDs rather than depending on wall clock or seed luck, and avoid shared mutable state between tests. Each test sets up and tears down its own data so execution order and parallelism never affect results.

Preserve the first intermittent failure including inputs, seed, execution order, logs, and environment before changing anything. Diagnose whether the cause is a product race, test nondeterminism, or infrastructure, fix the root cause, and never retry the suite into green.

Use integration tests when the risk lives at a real boundary such as database queries, serialization, filesystem behavior, queues, or network protocols. Exercise the real boundary implementation, since replacing it with a mock turns the test into a unit test and can hide schema and lifecycle defects.

Property-based tests verify invariants over a broad generated input space, while fuzzing targets parsers, protocols, and trust boundaries where generated inputs expose failures humans would not enumerate. Both require bounded campaigns, reproducible seeds, and minimized failing inputs promoted into deterministic regression tests.

Delete a test when evidence shows it is invalid or redundant and meaningful regression coverage remains, or when it is unreadable, low-signal, or duplicates coverage from boundary analysis. Tests of trivial getters, framework behavior, and restated implementation logic are also deletion candidates.