You're not wrong Juls, you want to pre-empt physics issues before they occur for best results, although in practice having the fall-back of the physics saying "hang on, I need to react to this" is not as bad you might think.
But if you think about it you actually have all the event triggers you need anyway, you know when the car is near a bump or kerb or gravel trap because you know where the car is. You know when forces are about to effect the physics because you know the controller inputs, so pre-empting when the physics need to be running at a higher frequency is actually extremely easy.
I hadnt got as far as car collisions (which created some interesting scenarios in our multiplayer testing!), I think it was the thing I was working on implementing when the project was abandoned, certainly I was getting nagged for it by the test team, but I wonder if any practical issues would have arrisen from that. Although I can't see anything that would cause major concern. It's not like the physics are not running, they're just optimised is all.
I too was suprised by the physics rate of some of those games, and wonder just what they encompass within the word "physics".