flake: walk_registry m664 tests share SEMANTICS_LOG static without cross-test serialization

[mlieberman85]

Observed on

PR #719 CI run 32931567792 — Lint + test (macos-latest) lane:

test scan_fs::walk_registry::walker::tests::walker_skips_default_noise_dirs ... FAILED
test scan_fs::walk_registry::walker::tests::walker_survives_symlink_loop   ... FAILED

thread 'scan_fs::walk_registry::walker::tests::walker_skips_default_noise_dirs' panicked at
    waybill-cli/src/scan_fs/walk_registry/walker.rs:609:9:
    top-level file should be visited; log=[]

The subsequent re-run (SHA 1b6d87a) passed the same tests — classic parallel-test flake symptom, not a real regression.

Root cause

Three tests in waybill-cli/src/scan_fs/walk_registry/walker.rs share one static log for their assertions:

static SEMANTICS_LOG: Mutex<Vec<String>> = Mutex::new(Vec::new());

Each test calls SEMANTICS_LOG.lock().unwrap().clear() at its start, populates it via a record_visit callback registered on the walker, then asserts against the collected entries.

Sites (as of dc8018d):

• walker_survives_symlink_loop (line 488) — #[cfg(unix)]
• walker_respects_exclusion_set (line 525)
• walker_skips_default_noise_dirs (line 568)

Cargo runs test fns from the same target in parallel by default (--test-threads = nproc). When two of these three tests interleave, one calls .clear() after another has populated the log but before that other reads its assertions, leaving log=[] and tripping the "should be visited" assertion.

Local runs on macOS pass because scheduling happens to serialize the three tests; CI macOS-arm64 has different core counts and interleaves them differently. This is why the m664 merge lane passed and this PR's first run failed even though walk_registry/ had zero changes on this branch.

Fix options (any one)

1. Per-test-local log: replace the static Mutex with a Vec<String> owned by each test, threaded into record_visit via the SharedWalkerContext.state payload. Cleanest — removes the shared mutable state entirely.
2. File-scoped serial Mutex: add a SERIALIZE: Mutex<()> static and have each test acquire it BEFORE SEMANTICS_LOG.clear(). Cheap but only prevents interleaving inside this one file; other tests in the same binary that touch the log would still race.
3. #[serial] from the serial_test crate: standard-idiomatic. Adds a dev-dep the workspace doesn't currently use. Skip unless the pattern shows up in more places.

Recommend option 1 — it's the right shape for a walker that already threads a state: Option<Box<dyn Any + Send + Sync>> through SharedWalkerContext per m664 contract C4.

Not blocking anything

Neither test is a functional gate; both are unit-tests over walker semantics. The flake surfaces intermittently on macOS lanes and is worked around by re-running the failed lane. Filing so someone can fix at leisure and so the next occurrence doesn't get root-caused a second time.