Conservation is easy. But elastic collisions are generally incorrect for vehicles, they'd look goofy at least, and mid-corner punts would be catastrophic. The losses are typically in sound (cyclic deformations) and, uh, more permanent deformations - the tyres' slipping plays a part, too. You can fudge some of that, e.g. the larger the impact, the more inelastic it becomes, up to the limit of deformation, which varies by axis; and by car; and by prior deformation etc.
Conservation gets a little hard when bits decide to detach and fly off, spinning and such. But only really for the big bits.
Redistribution is not so easy, and is the aspect with the more substantial gameplay consequences for car-to-car impacts. Relative stiffness plays a huge role there, and that implies deformation, right down to the tyres. You can fudge that, too, partly by assuming suspension (and tyre) deformation dominates at low energies.
You still can't do any of it on the raw geometry (there's too much of it), you'd have to use proxies. One way is to build a complex damage model to collect data for a stiffness-loss lookup on, say, location/angle-energy values based on a proxy (e.g. a bounding box) per car, and return a different force for each car, or something. But then you don't get visual deformation without running a separate process that recreates some of the stuff in the original simulation, and you don't get any variation as the car deforms or as bits fall off, unless you include all of that in your lookup...
Sure, some of that is overkill, but all of it, even the basic stuff, pre-supposes that the cars react appropriately to those general forces applied to the body or chassis in the first place. I'm still waiting for a proper multi-body approach and the end to break-dancing cars in GT games, because currently (despite the progress) they only respond convincingly to forces applied at the tyres.
