Now I can't help but wonder how far one would get with reverse Catmull-Clark on more general meshes. You know, if a mesh has mostly quads and few poles, then most of its topology looks like the result of a subD step anyway. The few bits where it isn't shouldn't be too hard to deal with. The trickier part of course is then to determine the vertex positions of the reduced mesh so that after applying subD, the result comes as close as possible to the original.

-- I'm not mad at you, just Westphalian.