Same shape, only one is the image

chw   = np.transpose(photo, (2, 0, 1))      # (3, 512, 512) — correct
wrong = photo.reshape(3, 512, 512)           # (3, 512, 512) — scrambled

Both give shape (3, 512, 512). Only one is the image.

A shape check passes on both, which is precisely why this bug survives code review. The check that catches it is a value check: pick a coordinate and confirm the number that lands there is the one you meant.