This improves detection accuracy:
1. Add name normalization for numbered design frame names (e.g., '12. Withdraw' → 'withdraw')
2. Add pages.json cross-validation to ensure no pages are missed
3. Add structure check using .pen node tree before visual comparison - detects missing blocks early
4. Add enforced checklist for correct difference counting - icon mismatch must be counted as moderate
5. Add examples for correct grading
