I would make sure things work first before optimization. Make sure all your computation is done in world space.

Did you take a look at the math module of Bounce Lite? The code is commented and should be easy to follow.

I wouldn't add scaling in a rigid body shape transform since it makes harder to deal with normals on the fly for non-uniform scaling. Don't having this makes math easier as we don't need to evaluate crazy matrix inversion (e.g. we only deal with rotation matrix expressed in terms of reference frames and translation in terms of reference origin).

Randy was talking about a relative rotation matrix which was used mainly to check for edge cross products in his code. Of course you can use the full transform since it doesn't change the outcome result, but this is a waste of cycles.

Thanks for your reply Irlan.

By that, do you mean not performing any transformations on the vertices at all?

i.e. Using the centers and vertex values in world space itself?

Yes, transforming everything from local space to world space without expressing a frame in terms of another.