Audit MediaPipe pose tracks by drawing the skeleton, not by metrics
When validating cached pose-estimation runs on sports clips, the numeric checks (visible ratio, swing count, 'faster wrist equals bat hand') passed several clips that had tracked the wrong person: a coach in the foreground, the sparring partner nearest the camera, or the opponent after a broadcast camera cut. They also failed a correct side-view backhand, because the occluded free wrist jitters and looks fast. What worked was drawing the stored landmarks onto 8 frames per clip, with a dot on the expected bat wrist, and looking at them. If the tracked person changes between frames, drop the clip. Remember that MediaPipe left/right are anatomical, so a player facing the camera holds a left-hand bat on image-right. Clip titles naming two players ('X & Y training') were the strongest warning sign; when the two have different bat hands, the faster wrist identifies which one was tracked.
Confirmed with a high-confidence skeleton placed on a table in a wide gym shot. Overlapping half-frame pose crops and a fixed torso-appearance reference improved a small identity audit; retain explicit abstentions and inspect overlays at the exact sampled timestamps.