the mesh groups are relevant only. it seems, that whenever two faces have one edge on the same coordinates, these faces are being welded ingame unless they are different mesh groups in milkshape. however i experienced unwelding a geometry entirely, converting it to cmp and seing the welding on some face pairs while not on others.

Just thought I'd throw my rusty 2 cents into this one, I hope I can help you, w0dk4.

I remember reading a thread once, a long time ago - there were even pictures attached to prove its relevancy. It was about getting smoothing in FL right. Well, based on long-term memory, here it goes:

Smoothing groups are not exported by the CMP exporter in any way, in fact, the FL engine does smoothing on its own. The way smoothing is implemented in FL is not based on smoothing groups of any kind, but, surprisingly enough, texture coordinates.
To put things short, if a texture is continued/repeated over an edge, that edge is going to be smoothed. If there is a texture boundary, like a border with a different texture (this seems to explain your partial control over smoothing), or just a "break" in the continuity of the UV map - that is, the UV map continues on the same texture, only at different coordinates, FL considers that edge to be, well, an "edge" and so it won't be smoothed.

I know that this implies that, contrary to what you just said, it has to do with the engine and not the exporter, I am almost confident that I have fixed a smoothing problem using this method a long-long time ago.
See if this helps you out.

Ok, this is definately an exporter problem and not being done by the engine.

To reproduce, just do a simple test. Create a box in milkshape. It will automatically have the proper smoothing groups and will have its edges unsmoothed.
Now, export that to a .cmp file.
Open up the .cmp file and look at the VMeshData. It will have normals that are interpolated (-> smoothed). To verify that the exporter screwed up, re-import into Milkshape and render it with smoothing enabled. You will see that two corners of the box are smoothed.

Same happens when you export a sphere for example.
The smoothing you do in any modeling program wont matter, the exporter will apply its own smoothing, based on mesh groups it seems. That would be "okay" if the exporter would smooth reliably based on mesh groups, but what's really bad about all this is that the exporter doesnt even smooth correctly.
Try the sphere example, there will be hard-edges on it.

How on earth did nobody ever notice this? Am I doing something wrong?

(btw. it doesnt matter if you use cmp exporter 0.2 or 0.3, both have the same issue)

The problem over there, if my memory serves me correctly, was a simple duplicate poly that was invisible in Milkshape, but caused the above mentioned thing.

The model was also a bit troubled, because of all the manually created faces. Never ever got a SUR file to work properly on that one, though I remember Lancer's Sur builder produced very good results. Too bad it always forgot about a wing.

Back to topic: in this case, that thread about smoothing was completely wrong. I do remember seeing one, can't find it, though. If that's the case, sorry for messing things up. Really, I wonder what's up with this one, though. If the normals exist but are bad in the produced CMP file, it's definately the exporter, that's for sure.

+1 for above post...lol