Thanks for the update. I disagree it’s the same problem, though.
The problem described in the article you link is “if you model a perfect sphere with no friction, it will roll forever.”
The problem we’re reporting is “even when adding friction, spheres roll forever – the integrator ADDS ENERGY.”
These two problems, on the surface, seem very similar, but the root cause is very different. If PhysX uses an integrator that adds energy to the system, there are many kinds of simulations that will never work right. (And this may explain some things about Kerbal Space Program, now that I think about it … 
I’m happy that you’re actually working with NVIDIA on this, and if you can put pressure on them to fix their stuff, that’s great! We just need the fix to be for the right problem, not the surface-similar wrong problem.