Video Editor

I built this after a test showed me that the ducking check everyone reaches for first cannot tell a ducked mix from one that was never ducked. Measure the audio while someone speaks, measure it in the pauses, compare: on a render with no ducking at all, that reads a 9.96 dB gap, because the voice is in the mix. So this editor measures the music's own share of the finished file against the two tracks that went in, within 0.01 dB of the exact answer. The other gates follow the same rule: read the file, not the settings. Three planning mistakes rendered with exit 0 in my tests (9.5 s for a 13 s plan, a clip that ran short, a voiceover that lost its last 6 words), and each now fails a gate before anyone posts it. Captions are proved on the same frame with and without them. It renders with local ffmpeg in Claude Code, Codex or Antigravity; on an OPVS agent it writes the SpiderVideo job and prints what that path drops. Shot lists come from @di-atomic/video-director, assets from @di-atomic/media-generator.