test-value

Identify tests that protect the same failure through the same implementation path, then consolidate or remove duplicates while tracing call paths to confirm surviving tests still detect the mistake. Record the burden, existing protection, and surviving evidence for substantial cuts.

Replace arbitrary waits with real readiness signals and inspect ownership and cleanup of shared state. Distinguish a product race from a test timing assumption, and only quarantine with a reason, replacement protection, and a clear restoration condition.

Skip new tests for pure copy changes, mechanical edits, and behavior already covered effectively by existing tests, type checks, or other checks. Adding a test for every edit is not the objective; protection must justify its execution and maintenance cost.

Snapshots are valuable for public serialization, protocol bytes, and promised deterministic output, but exact matches on automatically chosen output freeze noncontractual details. Assert outcome properties and justified tolerances while preserving explicit constraints.

When AI writes both implementation and tests, they can share the same mistaken assumption. Establish expectations from user requirements, public contracts, or independent calculations instead of copying the production algorithm into expected-value generation.