actually, you *must* dither at every truncation step -
if you don't, you
will lose information _and_ introduce signal-dependent requantisation
noise, which can never be removed again.
so the best approach is to only reduce the wordlength once, at the very
end of the chain, before going to CD.
This is a correct recommendation.
some mastering people (bob katz among them) even go as
far as demanding
dither at every level control in the chain.
Strictly speaking, this does not save you (and by you, I mean Bob
Katz). Further operations after dithering can 'break' the dither and
reintroduce distortion. Then you have the worst of both worlds; the
distortion _and_ the added noise.
Working at 24 bit / 32 float, this is academic and won't hurt you in
any audible way. It's too far below the audible floor to care about.
Monty