Use when results differ between runs or machines, tests are flaky, generated files churn in diffs, simulated data must be reproducible, or ground-truth files risk being hand-edited and drifting from their generator.
