Run parallel area audits

Audit code areas with parallel sub-audits and verification

Fast multi-perspective code audits with adversarial filtering of false positives; HIGH findings must be independently verified before inclusion in report.

• · [object Object]

• · [object Object]

• · README.md + CLAUDE.md (for invariants from repo)
• · docs/ (for canonical spec)
• · Source code (read-only audits, no writes)

• · subAreas array with key (slug) and prompt (audit mandate) for each area
• · One agent per sub-area (parallel), returns structured findings
• · HIGH findings adversarially verified immediately (pipeline barrier)
• · Verification can refute (real=false) or confirm (real=true, adjusted severity)
• · Synthesize aggregates verified HIGH + all MED/LOW into REVIEW.md

• · FALSE POSITIVE HIGH finding → adversarial verify refutes it, downgrades to RESOLVED
• · Security claim not independently reproducible (guards exist caller-side) → verify refutes
• · Confidence score inflated (hallucinated line number) → verify catches when file:line doesn't exist
• · Sub-area prompt ambiguous → agent returns few findings rather than inventing
• · Missing file cited → verify reads actual source, flags as non-real if absent

• · Every HIGH finding verified adversarially before report inclusion
• · Confidence 0-100 honestly assessed (not just binary)
• · File:line cited and read for verification (not hallucinated)
• · verifiedCorrect[] lists things checked and found OK (not just problems)
• · Refuted findings documented with rationale (RESOLVED status)