The idea behind using FP for "in between" data is that the (effective)
resolution of 24 bits is so high, that no dithering is required. The amount
of noise required to make something at -150dB (which is inaudible to begin
with) audible is way too much...

The use of floats instead of integers is mainly to prevent overflow or
underflow.

If one is certain of the range (e.g. adding 4 16 bits channels: you'll need
add most 18 bits to hold the result) of the values, there is no need to use
floats. Using 32 -bit integers will be faster (and more accurate...) for
predictible operations like adding signals and volume changes.

When using convolution filters (e.g. bandpass), inbetween values tend to be
very small or very large compared to the final values, especially when
cascading several of these filters. In these cases, the use of floats
eliminates the need for range checking. The final step will usually consist
of normalizing and converting to integers, with dithering if the resolution
is to be 16 bit (or less).

ML