Restructure gallery: separate dataset issues and update all metrics

Major improvements to gallery organization and accuracy:

1. Separated Dataset Issues into Dedicated Section
   • Created new expandable "Dataset Issues (18)" section
   • Cyan theme (#06b6d4) to distinguish from errors
   • Removed from errors grid filter buttons
   • Independent expand/collapse functionality
   • Clear description about exclusion from accuracy

2. Updated Errors Section
   • Now shows only 45 real errors (down from 63)
   • 5 filter buttons: All, Incorrect Parse, Missed Parse,
     Not ADE Focus, Prompt/LLM Misses
   • Updated "All" button: 63 → 45
   • Simplified descriptions

3. Updated All Hardcoded Metrics (99.082% → 99.156%)
   • Title: "99.156% Accuracy"
   • Header subtitle updated
   • Error count: 49 → 45
   • Dataset issues: 14 → 18
   • Correct answers: 5,286/5,335 → 5,286/5,331
   • Improvement: +3.72pp → +3.80pp
   • Footer accuracy updated
   • All JavaScript console logs updated

4. Technical Improvements
   • Used BeautifulSoup for proper HTML parsing
   • Filter JavaScript now targets only #errors-container
   • Prevents filter interference with dataset section
   • Proper grid layout maintained (3 columns)

Result: Clean separation between system errors and annotation
quality issues. All metrics now accurately reflect 4 cards
reclassified as dataset issues.

Category breakdown:
• Real errors: 45 (Incorrect Parse: 13, Missed Parse: 5,
  Not ADE Focus: 9, Prompt/LLM Misses: 18)
• Dataset issues: 18 (shown separately)
• Final accuracy: 99.156% (5,286/5,331)

Author: Ankit Khare