Well... without c_updatehull I had no updated collision hull. But I needed to find out what I'm doing wrong before, as there was no hint about that.
The dngine did not update the hull for sure - and I was wondering why c_trace delivers odd results just to find out it is using the original (flat) hull of the ent_createterrain created terrain.

I'm using ent_fixnormals on (unchunked) terrain and it seems to work... good enough at least.