No testTimeout is configured, so Windows inherits a POSIX-calibrated 5s default that measures contention

[nokkies]

No testTimeout is configured anywhere, so every lane inherits vitest's 5s default. That number is POSIX-calibrated, and on Windows it measures contention rather than correctness.

Measured on a Windows runner: a full run failed 30 tests across 13 files, and every one of those files passed when run alone. Windows pays hundreds of milliseconds to start each process — Windows PowerShell 5.1 alone is expensive — and this suite runs many spawn-heavy files in parallel.

Raising the deadline to 30s left 2 failures; 60s left 3. Past roughly 30s there is no threshold to find, only residual flake that shuffles between runs — so 30s is the knee, and a larger value would only make a genuine hang cost more.

Fix

A per-platform deadline: 30s on Windows, vitest's 5s default retained on POSIX so CI keeps a fast failure signal, with an env override for host tuning. Applied in a fork at 411dfd772a.

One thing that bit us afterwards, worth knowing before adopting this

A raised default only reaches tests that do not override it. One describe-level { timeout: 15_000 } sat below the new Windows figure, so it did not raise that lane — it halved it, in a file whose tests drive real git and node processes and take ~4s each with the machine idle. It surfaced as a single timeout in a full run that CI had passed on the same commit.

If you take the per-platform deadline, grep for { timeout: at the same time. Every such literal is part of the change.