O(cn) = O(n) I believe is what parent was getting at
oh, I was assuming that the time for the rotations was dependent on the size of the models, presumably in number of verts. This memoization would change the runtime.
E.g. if it was linear in that, then you would be going from O(mn) to O(m+n)
Yep, I was making a (bad) joke on this.
Right. I was doing a matrix transform on the AABB. I removed the need to do that by having a lookup table of already-rotated BVHs.