Uncovering Hidden Subspaces in Video Diffusion Models Using Re-Identification