Rule-based pose gestures: gate on visibility, make rules exclusive
Simple gesture recognizers over MediaPipe pose landmarks (wrists 15/16, elbows 13/14, shoulders 11/12, nose 0) work well for a game if each returns a confidence and you pick the max. Return 'none' when key landmarks have visibility below ~0.5, remember normalized y grows downward (raised arm = wrist.y < shoulder.y minus a margin), and add disambiguating conditions so gestures don't overlap: a 'both arms high' gesture needed 'wrists spread apart' to stop it firing for 'hands together'. Keep the returned move ids typed as the server's move union; a mismatch like 'powerstance' vs 'power_stance' silently drops moves.
Left/right caveat: MediaPipe Pose names landmarks from the person's point of view in the frame it receives (15 = the subject's left wrist). If you mirror a selfie camera frame before inference, left and right swap, so a 'raise left arm' rule fires on the other arm. Mirror only the preview, or swap the index pairs when you flip the input.
Porting these rules to Apple Vision (VNDetectHumanBodyPoseRequest) flips one sign: Vision's normalized points have origin at bottom-left, so y grows upward and a raised arm is wrist.y > shoulder.y. Use VNRecognizedPoint.confidence as the visibility gate (it is 0 for joints not found), and convert with VNImagePointForNormalizedPoint plus a y flip before drawing in UIKit/SwiftUI.