Possible bug in contact/joint solver in two-directional Coulomb friction model

Around April/May this year I reported that we found some strange behaviour in joints and contacts in PhysX 3.2 compared to that of PhysX 2.4 (https://devtalk.nvidia.com/default/topic/541615/?comment=3792179). The strange behaviour is that joints and contacts become very compliant/wobbly. After having made some new scenes we could not reproduce the wobbly joints and so the problem seemed a one scene bug, until now.

Last week we made a new scene that also has the wobbly joints (wobbly contacts cannot be verified). Somehow we discovered that both scenes required accurate friction modelling and therefore having the flag eENABLE_TWO_DIRECTIONAL_FRICTION. Removing this flag (using the default friction model) resolves the problem. On the following links two PVD files are stored showing the exact same scene with first the default friction model and then the two-directional Coulomb friction model.

https://dl.dropboxusercontent.com/u/2268947/test_default_friction.pvd
https://dl.dropboxusercontent.com/u/2268947/test_two_dir_coulomb.pvd

I have run through the release info files of all new PhysX 3.x releases, but found no evidence that anything has changed on the friction model. I also did not find any related topic on this forum of the internet.

We now wonder what internally happens when the two-directional Coulomb friction model is activated. Is the contact/joint solver (accidentally) changed to be more compliant? Or is the Coulomb friction model inherently more compliant? And why does it affect the joints stiffness?

We hope that the stiffens of the contact and joints can be restored so that we can use the Coulomb friction model with stiff contacts and joints.

Best regards, Jeroen

anyone any idea about the problem?