feat(models): implement structured JSON output for OpenAI

[runtimebug]

Summary

Implements Phase 1 Task 3: Structured JSON output from LLM to replace fragile regex parsing.

Changes

• New schema module (src/schemas/):

  • review-response.ts - TypeScript types for structured output (severity, category, confidence, title, explanation, suggestion)
  • review-response.json - JSON Schema for OpenAI's response_format
  • validate.ts - Validation utilities with fallback to legacy parsing
  • validate.test.ts - 33 comprehensive tests

• Updated OpenAI adapter (src/models/openai.ts):

  • Uses json_schema response format for reliable structured output
  • Per-comment severity filtering (critical | warning | suggestion | nitpick)
  • Confidence threshold filtering (default min: 0.6)
  • Backward-compatible legacy parsing as fallback
  • Rich comment formatting with category badges and confidence scores

Why

The previous regex-based parsing (/([\w/.-]+):(\d+)\s*[—–-]\s*(.*?)(?=...)/gs) was fragile:

• Comments were silently dropped if LLM used slightly different formatting
• No per-comment severity or confidence scores
• Colons in comment content could break parsing

Testing

All 94 tests pass:

• 33 new tests for schema validation
• 20 existing tests for OpenAI adapter (still pass with backward compatibility)
• Full test coverage for legacy fallback mode

Backward Compatibility

The implementation gracefully falls back to legacy free-text parsing if:

• The response doesn't look like JSON
• JSON parsing fails
• Validation fails

This ensures existing behavior is preserved while enabling the new structured format.