Dithering 24 bit files or exporting 32 bit floating point files is an "it couldn't hurt" practice firmly founded in superstition, not science.
i'm sorry for making you sad, but i can't help it:
we have a track with a bunch of plugins on it. or just one, whatever. a 32fp render of this track will be exactly the same as the source. this is easily verified with a null test. a 24 bit render will be almost exactly the same, except there'll be either truncation distortion or dither noise added. also easily verified with a null test.
how is this anything other than straight up rudimentary bulletproof science? there's nothing superstitious about it. saying "it's happening at a level that's far too low to hear" is fine and i agree with you, but that's like telling a scientist "oh please, if you need a microscope to see it, it can't possibly be important."
First of all, let's be perfectly clear about what you're claiming.
You state that a 32 bit float file represents the source better than a 24 bit file.
It also appears that you've modified the source with DSP operations and you've stated previously that you believe these operations expand the word width to 32 bits.
Are you comparing the 24 bit source to the DSP modified file in this case?
If so it appears that you're also asserting that the truncation to 24 bits has compromised the data integrity of the now 32 bit file.
Do I understand you correctly?
Now lets add some missing details:
A 32 bit floating point file contains 24 bits of audio data and an 8 bit multiplier, which allows the level to be scaled. If the 24 bit audio data is extracted from the file, it is identical to a normalized 24 bit file.
When a 32 bit float file is reproduced it is only ever a scaled 24 bit file.
Dither and truncation should only happen when the output file is being generated.
For these reasons, comparing the dithered 24 bit file to an internal DAW file format is irrelevant.
There is no output mechanism on earth that can reproduce the quantization error or noise of a truncated 24 bit signal.
And lastly, as demonstrated in the posted video, and I'm talking about the part where he's normalizing the gain for the truncation and performing a null test using the plugin he created, the self noise of a file is so far above the quantization error that the error becomes immeasurable.
It's important to remember that each additional bit only adds another least significant bit to the word width, and has no affect on the upper bits.