The ellipsoid collision model just plain out sucks. It has always been a pain to deal with and always will be. In two dimensions you can find some workarounds - although this gets a lot more tricky with rotated models - but in 3D this gets a lot harder.

What you can do for the gravity trace is make min_z and max_z (in absolute terms) really small, then perform the USE_BOX vertical trace and undo those bbox changes. This will make the downward trace rather box-like or cylindrical, and will give you less stupid target and normal results.