feat: benchmark calibration, semantic validation & improved prompts

[jmlweb]

Summary

This PR adds several improvements to make analysis more reliable, especially for small models:

🎯 Gold Standard Benchmark

• 50 curated prompts with human-rated quality scores
• Covers all tiers: excellent (10), good (15), fair (15), poor (10)
• Includes correlation calculation for model accuracy measurement
• Used for calibrating scores across different providers

✅ Semantic Validation

• Validates that scores correlate with issue counts
• Detects when examples are not found in original prompts
• Auto-corrects results when validation fails
• Prevents logically inconsistent outputs

🔄 Temperature Fallback Retry

• When JSON parsing fails, retries with lower temperatures
• Sequence: 0.3 → 0.1 → 0.0
• More deterministic outputs reduce parse failures

📝 Enhanced SYSTEM_PROMPT_MINIMAL

• More contrastive examples showing score progression
• Clear examples for each tier (POOR, FAIR, GOOD, EXCELLENT)
• Better calibrated scoring guidelines

Testing

• All existing tests pass
• New benchmark tests with 24 test cases
• Typecheck passes