GeezerVison

What this sweep found

#135 fixed the setup-matrix decoy race and documented the gotcha: #[serial] only excludes other #[serial] tests. Sweeping every test binary in both crates with a helper-aware analyzer (attributes each test's mutation status including mutation hidden inside helpers like crawler_npm_e2e::with_empty_path) found the same hazard at much larger scale:

The serial tests swap HOME, VIRTUAL_ENV, PATH, XDG dirs, and (composer) the process cwd around in-process crawler calls. Unlike #135's child-spawn snapshot race (microseconds window), this variant exposes siblings for the mutator's entire body: any concurrently-running sibling whose crawl probes HOME-based global caches mid-swap sees the wrong universe. All mutating tests were already #[serial] — the gap was solely their unmarked siblings.

The fix

#[serial_test::parallel] on all 205 unmarked siblings. Parallel-marked tests keep running concurrently with each other but take the shared side of serial_test's lock, so they're excluded while any #[serial] test runs — suite wall-times are unchanged (verified: all ten binaries' runtimes identical before/after; 280 tests green).

Blanket #[serial] (the #135 approach, right for those tiny suites) would have serialized 58-test binaries for no reason.

Mechanism proven RED/GREEN

Throwaway probe pair (not committed): a #[serial] test swapping HOME for 150ms alongside an unmarked sibling asserting HOME stays stable — fails 3/3 runs with "HOME changed mid-test"; after marking the sibling #[parallel], passes 5/5. Same primitive as #135's deterministic proof, in-process variant.

Not touched

Single-test binaries with env mutation (e2e_golang_build, setup_matrix_env_guard) — no siblings, no intra-process race.

🤖 Generated with Claude Code