Every new collision shape requires a lengthy implementation and test. Thus we can't implement an arbitrary number of collision shapes, but only one, if at all.
That's interesting. I would've thought you'd use the Gilbert-Johnson-Keerthi algorithm for collision detection, in which case supporting a new type of convex hull that's compatible with all other collision shapes would simply be a matter of writing a new support function for that type of hull.
As others have said, I think a cylindrical hull (as well as support by c_trace with USE_BOX) would by the most advantageous addition for us programmers if you can really only do one.
Jibb