The streaming video encoder dropped 104 frames (100 wrist, 4 top) in 101 episodes, so after each drop the images ran one frame ahead of the data rows and the last row read the next episode's first frame (or past the end of the file, for episode 6180).
Each missing frame is restored by inserting a copy of the nearest keyframe packet at the located drop position (no re-encoding; all other packets are byte-identical), and the per-episode video timestamps and image-stat counts are updated. Every episode now has exactly one image per data row.
Previous revision: 00f832183b401be6b101874f83ef59a54a2ffaea
