I almost feared I failed to convince you…
Therefore I prepared a few images closer to the setup you wanted to achieve (Erm, this is supposed to be a stylized dead end street with a patch of water on the ground and some random spheres…):
So depth of field in Cycles will blur reflections independently of the surface reflecting them. And this is IMHO in line with the behaviour in real world: If depth of field blur through reflections (and refractions, for that matter) worked on the surface creating the effect (rather than the object being reflected or refracted), then how would camera lenses or mirror telescopes work? You would have to focus on the lens element/the mirror itself and then everything would be in focus - that isn’t right.
I think this is a “sign of quality” for Cycles as a renderer, because some scanline renderers calculate DOF only from the z-buffer and will fail miserably in scenarios like this, because to them the reflection on the mirror is just like color information painted on. Same goes for post-pro DOF solutions, which will also not produce realistic results in these scenarios.
All this can be different, though, if the mirror is not flat, but has a curved surface. Curved mirrors are basically lenses (“mirror lenses”: http://en.wikipedia.org/wiki/Curved_mirror) and can show distant objects in focus, while outside of the curved surface the distant objects are blurred.