[codex] Harden public AI summary invariants

[jeffreyzeng10]

Summary

• add unsafe AI output fixtures for public forbidden language from test(ai): add AI sanitizer and fallback invariants #152
• sanitize public AI prompt context before model upload while preserving private authenticated context
• extend public unsafe-output detection to private reviewability/private scoreability/public score-estimate language

Verification

• npx vitest run test/unit/ai-summaries.test.ts --testNamePattern "unsafe public AI output|public AI prompt"
• npx vitest run test/unit/ai-summaries.test.ts
• npm run typecheck
• npm run test:coverage
• git diff --check
• npm run test:ci

Closes #152

[JSONbored]

@jeffreyzeng10 This is ready from my side. Public AI summary generation now uses stripped public context, rejects unsafe public model output, and includes regression coverage for forbidden output terms plus private-context separation.

Current checks are green and the merge tree is clean.