LIGHT self-review

Tier remains LIGHT: one repository audit harness fix with no runtime, security, session, persistence, concurrency, cache, integration, or domain-boundary change.

• Scope is limited to script/gpt-mini-reference-audit.test.ts plus evidence.
• The content audit is not suppressed, skipped, narrowed, or allowlisted.
• git ls-files, retired-model match rules, and the empty-match assertion are unchanged.
• fileURLToPath fixes both percent decoding and Windows drive-letter paths.
• The 20-second timeout is a circuit-breaker for a 900-file scan observed at 4.28 seconds, not a sleep, poll, retry, or timing assertion.

Programming post-write review:

1. Single responsibility: the file owns the retired GPT Mini reference audit.
2. Boundary purity: the file URL becomes a native path at the boundary.
3. Variant discrimination: no tagged variants were added.
4. Escape hatches: no any, assertions, ignores, or warning suppression.
5. Defensive layer: no redundant checks or catches.
6. One-off helpers: repositoryRootFromMetaUrl serves production setup and the regression test.
7. Tests: reverting to .pathname makes the regression fail.
8. Parameter bloat: the helper accepts one parameter.

Validation:

• Pure LOC: 48.
• Focused Bun 1.3.12 test: 2 passed, 0 failed.
• Script typecheck: exit 0.
• No-excuse audit: no violations.
• Full suite: 12,626 passed, 3 skipped, 0 failed.
• Registered LSP diagnostics cannot address the sibling worktree; script tsgo is the replacement type diagnostic.