Test adequacy review

A green suite proves nothing about the assertions that cannot fire. This lens judges the tests; it does not write them — but every finding names the missing or corrected test concretely.

Method

1. Name vs power. For each test: state the reason it exists (its name and comments claim one), then describe the concrete code change that should make it fail. If no such change exists — because the guarded path is not reachable from the test's setup, or the assertion holds vacuously — that is the top finding class: a test that cannot fail for its named reason. (Mutation thought-experiment; run a real mutation when cheap.)
2. Trigger surface. Enumerate every path that can reach the guarded behavior (all call sites, all gateway branches, all event orders). Which does the suite exercise? An untested trigger on a money or safety path is a finding even when the tested trigger is identical-looking code.
3. Barriers and time. Every wait must be tied to an observable signal. sleep(N) as a synchronization barrier, assertions racing a timer, and "must finish before X fires" enforced by nothing are findings — they pass on fast machines and lie on slow ones.
4. Fixture honesty. A test helper that hand-duplicates a production payload, contract, or constant will drift green: the test keeps passing while production moves. Shared structs/constants between test and prod are the fix; duplication is the finding.
5. Isolation and ordering. Global resets (shared stub state, scenario resets) that couple suites, and tests correct only under sequential execution, are findings with a deferred fuse.

Evidence rule: quote the test lines that make each finding true; name the concrete input, timing, or mutation that demonstrates it.

Rules

• Run in a fresh context with only the diff and the test files — not the author's reasoning, not other lenses' output.
• Never rerank against other lenses; adequacy of tests and correctness of code are separate verdicts.
• Findings go to the findings-ledger.

Provenance

Every finding class above shipped past multi-model review at least once in an internal retrospective (2026-08): a feature-off test that could not fail for its named reason, a second trigger branch exercised by zero tests, a 3-second sleep as a barrier, tests racing a short fallback timer, and a helper duplicating the server's start payload. Meta's mutation-guided test generation (FSE 2025) grounds the mutation discipline.