fix(audit): exempt test code from layer-2 CRITICAL-pattern gate

[XuebinMa]

What

Exempt test code from the Layer-2 pre-commit audit gate's CRITICAL-pattern
scan (.expect() / panic!() / .unwrap() / unsafe / shell-out).

Why — observation-period unfreeze

The three-layer audit workflow ran frozen (C-layer) from 2026-05-08 to
2026-06-01. Triaging the accumulated Layer-1 findings (.audit-state/findings.log,
7 CRITICAL hits):

Disposition	Count	Detail
False positive — test code	3	.expect() / panic!() in tests.rs, an inline #[cfg(test)] mod helper, and a test assertion
Remediated prod .expect()	3	secrets.rs ×2 → Regex::new(pat).ok(); provenance.rs → removed
True positive, justified	1	bash.rs from_utf8(...).expect(...) with a documented SAFETY invariant

Signal: the gate caught 3 real prod .expect()s (all since remediated) —
it works. Its only false-positive class is test code, which is idiomatic and
never executes on the host.

How

The gate diffs with -U0, so the enclosing #[cfg(test)] attribute isn't in
the diff. The scanner now reads each flagged file and brace-matches to
compute test-scope line ranges:

• Whole test files exempted: tests.rs, **/tests/**, *_test.rs.
• Inline exempted: lines inside a #[cfg(test)] module or a #[test] /
  #[<runtime>::test] function.
• Comments and double-quoted strings are stripped before counting braces, so
  "} {" in a string literal doesn't skew depth.
• Unbalanced blocks are not exempted — bias toward flagging, never toward
  silently letting a real prod .expect() through.

Verification

Ran the actual scanner block (extracted from the script, no copy) against a
synthetic fixture mixing prod and test .expect()s plus a brace-bearing string:

• Prod .expect() flagged (lines 2, 16) — including one placed immediately
  after a test module, proving the range doesn't over-extend.
• Test-scope .expect() exempted (lines 8 inside #[cfg(test)] mod, 13 inside
  #[test] fn). PASS.
• bash -n clean.

Decisions

• LLM-dispatch kept stubbed. Deterministic gate + test exemption gives a
  high enough signal-to-noise ratio; the single justified-but-flagged prod case
  doesn't justify synchronous per-commit LLM cost. Revisit if such cases grow.
• Full triage recorded locally in docs/audits/observed-fp.md (gitignored by
  design — that dir is local Layer-3 observation state).

Scope

Only .claude/workflows/pre-commit-review.sh. No Rust source, no audit-rules.yaml
or hook-registration changes.