Bug..? Solver iteration count vs. Wheel colliders - strange relationship

So, either PhysX is a buggy, unpredictable mess, or I am missing something.

We had a rigged vehicle with wheel colliders tuned perfectly with a global solver iteration count of 12. Strangely, increasing the global solver iteration count to 20 yielded entirely different results for the wheel colliders and to replicate the tuning we had to lower the spring force by roughly half its previous value.

Can anyone confirm what we are seeing and offer any tips for dealing with this? Ideally, the solver iteration count should be able to vary based on platform capability.

TIA!