Video pipeline on ALREADY-DECODED frames (uniform size): T-aware
smart_resize per frame → normalize → pad to even T by repeating the last
frame → patchify with REAL temporal pairs (element order (c, tp, py, px)
reads frame t·2+tp). File decoding is the AVFoundation sidecar's job.
Video pipeline on ALREADY-DECODED frames (uniform size): T-aware smart_resize per frame → normalize → pad to even T by repeating the last frame → patchify with REAL temporal pairs (element order (c, tp, py, px) reads frame t·2+tp). File decoding is the AVFoundation sidecar's job.