I think I've found the problems with my C# drop-cutter algorithm. The first bug was a trivial one where in the edge-test the rotation of the segment-under-test was messed up. The second one was not the facet-test itself but incorrect surface normal data. There's something going on that makes pre-calculated surface normal data not 'stick' correctly to the triangle object for later use. So here I'm re-calculating the normal data again just before the facet test.
The next step is to speed up things with a smarter kd-tree based search for which triangles are under the cutter. I've added bounding-boxes to the Cutter and Triangle classes. Running the above example the DropCutter function is called a total of 4735000 times, and of those only less than 5%, or 236539 to be exact, are useful calls, i.e. the triangle bounding box intersects the cutter bounding box, which means there's a chance that we are contacting the cutter against the triangle. The idea with the kd-tree then is to pre-search for triangles under the cutter to make less of those (supposedly expensive) redundant calls to DropCutter.
On my T5600 processor machine the above example runs in about 5 seconds. (1894 triangles, about 4.7 Million calls to DropCutter of which ca 240k calculate something). I've made a list of useful CAM-related things to work on: CAMToDo.