feat: dense frame extraction + key frame detection

[Overusedhydra]

What changed

• Frame interval: 3s → 1s (max frames 20 → 60) — catches the moment a creator moves their face aside to reveal on-screen text/prompts
• Key Frame section: every note now includes a prompt for the vision model to flag the one frame where text is fully visible
• Skill sync: skills/reels-vault/scripts/extract_reel.py now identical to main script (was missing all frame extraction code)
• Gitignore: excludes local tooling (.agents/, skills-lock.json)

Why

Previously at 3s intervals, frames often landed while the creator's face was still blocking on-screen content (prompts, code, text). The reveal moment fell between samples. At 1s density + the Key Frame prompt, the vision model can identify and mark the clean frame.

Testing

• Both script copies verified identical (diff returns no differences)
• Frame extraction defaults confirmed: interval=1.0, max_frames=60
• Key Frame section present in format_output()