feat: add confidence-based quality filters to seed-videos.sh

[kelleyglenn]

Summary

Closes #90

Adds confidence-based filtering to seed-videos.sh so entries with unreliable metadata are automatically skipped during loading. Uses the confidence scores already produced by the AI extraction pipeline (claude_extract.py).

Filtering rules

1. Location confidence: Skip below 0.6, unless the location has a name, street address, and coordinates (then allow down to 0.55 — enrichment-corroborated locations)
2. Non-accountability content: Skip if both amendments confidence < 0.5 AND participants confidence < 0.8
3. Video date confidence: Nullify dates with confidence < 0.5 (send video without date, don't exclude it)

Impact on current dataset (AP-after-2025-01-01)

Category	Count
Loaded	420 (93%)
Filtered by location confidence	9
Filtered as non-accountability	2
Failed (no coordinates/participants)	21
Total	452

Test plan

• Dry-run jq filter against full dataset — 28 entries filtered (matches expected)
• Apple Store entry (0.55 with street address) correctly passes
• Compilation/product videos correctly filtered
• No bc dependency (uses jq for all float comparisons — works on Windows Git Bash)

🤖 Generated with Claude Code