I suppose with floats you could dither the least significant bit, but the variance of level of this dither would significantly complicate quantifying the benefit of that dither. Also, with thirty two bit bounces where every available bit is used at the loudest moments, humans could never detect the precision on any system.
I believe that what Noel is saying is that dither is never applied to floats. That would mean that internal bounces never get dither, regardless of your settings, since the files created are floats. Noel?
Corollary: changing the setting for dither between internal bounce and export is not necessary.