ci: tighten attribution-guard regex and use content-based exclusions

[millerjp]

Summary

attribution-guard regex [^a-z]gpt[^a-z]? and \bllm\b have false-negative edges (gpt-4 passes if preceded by a letter; llm inside a word slips). The exclusion list is path-anchored to exact filenames; a rename silently re-enables matches.

Scope

• Tighten token list to catch common bypass patterns.
• Move exclusions from path-anchors to a content-based annotation (e.g., a // attribution-guard: allow tokens directive in the file header).

Acceptance criteria

1. Renaming an allowlisted file no longer re-enables false positives.
2. Existing allowlisted files carry the in-file directive.
3. Guard still catches a deliberate "Claude" insertion.

Source: security-reviewer.