@bead-v I believe your problem with the vanilla model is that you're not applying an orientation transform to the normals. The torso skin mesh has a super cool 180 degree Z orientation (quat(0, 0, 1, 1.27e6)), so, actually all those torso vertex normals you drew pointing away from the cape vertex normals, all "3" of those normals go in the same outward direction (and two of them match). If not complete slop/oversight it was probably done to workaround a bug in the vanilla toolset's smoothing. I suspect when you fix that (apply orientation translation to normal vectors), the patch-based normal construction is going to work out better. I don't fully comprehend why this would be necessary, other than, interestingly, apparently we need to be doing world coordinate vector comparisons in our smoothgroup computations where we've been doing something more like object space comparisons. When I opened the binary model straight into blender, it looked ... well ... like it does in-game. For the record, I wouldn't consider the smoothing on the vanilla pfbim model something worth emulating/holding up as an example. It's reaaaaal sloppy and horrible. The weights along the torso-robeback interface don't even match, such that you can see gapping in-game when you look from the right angle. Here's a screenshot of the vanilla model hilighting the garbage smoothing between torso and robeback (basically it's OK on the 5 rear seams, but kind of sucks on the side seams)
@JCarter426 if you're tweaking this model, I'd definitely recommend applying the rotation on the torso skin mesh if that's not one of the for-some-reason-sensitive-for-skin-weights-in-3DS operations. If the model you posted in thread already has this done, sorry and nevermind! (I only looked at vanilla while debugging)