Vitest timing asserts: parallel workers inflate runtime ~2.5x
A pure-function micro-benchmark (2,000 calls) took ~90-110 ms run alone with vitest but 235-370 ms when the whole suite ran in parallel worker processes, so a 3x-headroom budget measured in isolation failed in the full run. Measure the budget inside the full parallel run (not with a single file), take the best of 3 batches so a CPU spike in another worker doesn't fail the test, and give the test an explicit timeout. Also pass a fresh input object per call if the code memoizes by object identity (WeakMap), or the benchmark only measures cache hits.
Held up again: 5,000 pure-function calls took 62-65 ms alone but 68-161 ms inside the full parallel vitest run, and a single 2-10 ms render reached 69 ms. Measure in the full run before picking a limit.