autocrop-verticallisted

# AutoCrop-Vertical — Subject-Aware 16:9 -> 9:16 Reframer YOLOv8 detects people in each scene's middle frame, then per-scene picks one of two strategies: - **TRACK** — crop a vertical window centered on the subject(s). - **LETTERBOX** — scale the full frame down and pad with black bars (used when subjects are too spread out to crop without losing one). Frame-accurate (no scene-boundary drift), VFR-aware, audio-sync compensated. ## When to use - Multi-person interview / podcast / two-camera setup → tracking beats center-crop. - Action scenes, walking-and-talking, B-roll where the subject moves across the frame. - Cooking / sports / panel discussions where center-crop would chop people in half. - Inside the **viral-clipper** pipeline: insert this STAGE between clip-cutting and final render. Replace the simple ffmpeg `crop=ih*9/16:ih` step with a call to this script for any clip that has >1 face detected, or whenever the source isn't a static talking head. - Inside **book-video** / **channel-breakdown**: same — use whenever you cut a horizontal source into vertical Shorts. ## When NOT to use - Single static talking head dead-center → plain ffmpeg center-crop is faster and identical-looking. - Already-vertical source → no-op. - Animation / screen-recording with no people → YOLO has nothing to lock onto; falls back to letterbox anyway, plain center-crop is simpler. - Sub-3-second clips → scene detection overhead (~50% of runtime) isn't worth it for tiny clips. ## Install lo