I'm back from IMC at Bulgaria with second prize, and ready to work!
While I was away I thought about generic convex polyhedra and actually started implementing them. I hope to finish that in a few days. This not only makes engine practically more useful but also makes some computations more cheap and robust.
Also I discovered some problem with my last collision handler. It conforms to one paper, and now I doubt that paper is correct. I'll consult local physicists in the nearest time.
Finally, I read several papers on broad-phase collision detection. The one I liked most is Broad-Phase Collision Detection Using Semi-Adjusting BSP-trees. So I'll start implementing it as soon as possible.