feat(court): enhance reviewer and judge prompts to catch silent data corruption

[MrFlounder]

Summary

• Expands both Reviewer A (Claude) and Reviewer B (Codex) prompts with 4 high-risk pattern checks derived from the PR #3232 incident
• Adds a "Judge's Own Investigation" step to Phase 3 so the judge proactively checks for these patterns even if neither reviewer flags them
• The 4 patterns: called-but-not-changed code contracts, mutations on read paths, blast radius from null/default initial state, data contract completeness at merge boundaries

Context

PR #3232 (refreshMetricsIfStale()) caused mass data corruption because:

1. The bug lived in MetricsCalculationService.calculate() which was called but not changed in the diff
2. A GET endpoint silently mutated data via fire-and-forget
3. All existing evals had null metricsRefreshedAt, so every record got hit on first view
4. Object spread merged incomplete data over existing metrics, zeroing out fields

None of these were visible in the diff alone. The current reviewer prompts only say "review for bugs, security issues, and code quality" — too generic to catch this class of issue.

Test plan

• Run crab court on a test PR and verify both reviewers mention the new checks in their findings
• Verify the judge's Phase 3 output includes the "Judge's Own Investigation" section
• Confirm no regression in general review quality (reviewers still catch standard bugs)

🤖 Generated with Claude Code