Their paper deals purely in normals for their error metric, so any vertex movement that dramatically changes the normals of all triangles connected to it gets flagged as "high error", and hence ends up very far down in the edge collapse priority list. Since it's just normals, no extra artist involvement is needed. The reason it works at preserving contours is that a contour implies local geometry at somewhat different angles from immediate surrounding geometry, and hence any play in those verities generates a high error value. Like if you have a cube mesh, collapsing an edge starting from a corner vertex would cause mass error, as all surrounding triangles suddenly change their normals dramatically. Note as well that their technique is purely subtractive, it's meant to start from the best quality mesh and work down from there.
If left unchecked, such as if you try collapsing 70% of all edges, then for sure in some mesh contours will disappear. But you can mitigate this by setting an error cap. For example, collapse edges until the lowest error exceeds a certain 'max allowed error' threshold, then stop. So if you fed a cube mesh into the system, it would emerge on the other end untouched, however a piece of terrain may drop 40% of it's verts, etc...
Yeah I hear you on that. I've never written a Max/Maya plugin, but maybe it's possible to implement this right in the art tool itself. If not, you can always write an export tool that spits out the mesh they are working on into a directory, then they run a console side viewer app to see how it looks. But then someone has to take the time to write all this stuff
