perf(multi-hmr): autocast opt-in

MULTIHMR_AUTOCAST=1 enables MPS mixed precision for the ViT-S
backbone. Tested 2026-05-13: slower than fp32 baseline (400ms vs
270ms) -- overhead cast within forward exceeds matmul savings on
M5. Off by default; FP16 .half() crashes MPS matmul accumulator,
left out entirely.

apple_vision_pose face parser short-circuits to return immediately
since pyobjc 11 cannot dereference VNFaceLandmarkRegion2D
pointsInImageOfSize_ result. Removes ObjCPointerWarning spam at
30 fps (9 regions per face).

data_only_viz/apple_vision_pose.py

@@ -557,6 +557,7 @@ class AppleVisionPoseWorker:
             # API d'acces simple. Face parsing depuis Apple Vision est
             # actuellement bloque ; on garde MediaPipe (CPU XNNPACK) pour
             # face/hand fin tandis que Vision sert body 2D sur ANE.
+<<<<<<< HEAD
 <<<<<<< HEAD
             # Skip pour eviter le spam ObjCPointerWarning a 30 fps.
             return
@@ -566,6 +567,10 @@ class AppleVisionPoseWorker:
                 LOG.info("face: %s wrote %d points", region_name, n_written)
                 setattr(self, "_logged_face_write_" + region_name, True)
 >>>>>>> 28d562b (feat(av-live): wire Apple Vision body pose on ANE)
+=======
+            # Skip pour eviter le spam ObjCPointerWarning a 30 fps.
+            return
+>>>>>>> 623c479 (perf(multi-hmr): autocast opt-in)
 
         # faceContour
         fill("faceContour", *FACE_OFFSETS["contour"])