As far as I have found:

step 1: each vertex is positioned according to physics and dynamics. Done for all vertices first.

step 2: if between the vertex position and a distance of colission-depth (looking 'down' only, not two-ways) the colission surface is met, the sim does something 'clever' (which might be wrong if the vertex is already inside and the opposite side of the object is in the range) to avoid massive calculations in the future. This step becomes pretty meaningless when collision depth < offset.

step 3: if between the vertex position and the colission offset the collossion surface is met or passed, the sim starts pushing the vertex in the other way. This is an iterative process effecting surrounding vertices as well, is limited in success, and rather calculation intensive when a limited success per ietration generates a need for a lot of iterative steps. Since this is done after step 2, the routine might work with the wrong notion of 'pushing outwards'.

This is what I recall from memory from my works 4 - 6 months ago, I haven't looked into my notes or into this thread. It's somewhere above in my remarks on those two distance-parameters.

- - - - - 

Usually I'm wrong. But to be effective and efficient, I don't need to be correct or accurate.

visit www.aRtBeeWeb.nl (works) or Missing Manuals (tutorials & reviews) - both need an update though