What I am saying is use raytracing for the primary bounce, and only use cube-mapping in the secondary bounces. So this works fine for accurate rear and side view mirrors, as the mirror would reflect objects correctly.
But if you were to look through the raytraced mirror, at the paint of the car, the car paint would be cube-mapped. This prevents the issue of the car looking flat in reflections because the fallback is cube-mapped instead of flat.
This is done by forcing the cubemap to be sampled only when in the raytraced view.