Sorry if I wasn't clear. Using a time-dependent BVH (where instead of two vec3 for the corners you store four, one pair for t=0 and one for t=1) is an "optimization". Given the later sentence:

After that, he applied the same optimization to triangles (for actual character geometry)

it suggested that the bug was just using a single bounding box that ignored the motion (which is correct, but slow).