I've noticed the same thing, and no, this isn't how it's supposed to work. It's pretty much useless for my purposes, though somebody might find it worthwhile. If you can see everything well enough to manipulate with the mouse, you can avoid collisions anyway! C-D would be most useful when applying some premade motion, like a pose file, a BVH motion file, or a Python script. In those cases the motion is "blind" and needs to be stopped when it hits an unexpected obstacle. As far as I can tell, there's no Python-level representation of the C-D, so there's no way to make more complicated actions happen.... like bouncing off the target, or moving the target in response to the hit. Those would be easy in Python, if a "report" of the collision were available. But it isn't, and it doesn't work with BVH either.

My python page
My ShareCG freebies