I experienced this from time to time to. I don't think that it is a fault of the code.
I worked around this in different ways.
Did you try to move the vertices of the bump to different positions, higher, lower more to a side, things like that? Moving the whole bunch of blocks a short value from that position can help. Nevertheless, when you add other blocks, the whole collision will be calculated different - at least from what I experienced.

What happens if you add an arbitrary block at an arbitrary position in the level?

Or, the other way round - although from your screens I guess this was your initial approach - did you set the vertex/world snap to a high value?

My consequence is: two different meshes: one for the collisions, one for the look.
The collision very simple, only few tested slopes, or no slopes at all, or only terrain for grounds with bumps and slopes. Or a lot of testing and modifying of the areas.
Depends on the current state of my project and the estimated available time for the work.

What happens, if you import it into MED, close the mesh, and use it in WED as a model with the polygon flag set?