Some thoughts:
I done one more quick test with 32 bit sample files from SD website
https://www.sounddevices.com/sample-32-bit-float-and-24-bit-fixed-wav-files/1. I lower item volume to -150db (lowest possible volume in Reaper app)
2. I render one file as 24 bit and another as 32 bit.
3. I open both rendered files in Reaper and Normalize them back.
After Normalize rendered 32 bit file looked 100% like original
After Normalize rendered 24 bit file was crazy distorted.
Next i repeat test with source item volume lowered to -130db and -100db, but even so 24 bit file show very visible distortion after normalize.
Next i repeat test again and again and only with source item volume lowered somewhere to -60 -70db (at these levels waveform start to became rather visible), rendered 24 bit file starts looks more less like original after normalize. But i can still notice some tiny digital noise pattern artifacts if look very close on spectre image. Where is that truly safe limit in 24 bit file? It is fixed limit or smooth undefined limit? I don't know...
I see that 32 Bit Float allow to create really huge volume adjustments not only during recording but also during any other workflow steps and allow re-save truly lossless files 100% without any digital limits at all possible volume levels. So it seems like perfect archiving and editing format.
--------
For apps like DaVinci Resolve that can read 32 bit files but don't fully support 32 bit processing yet, the most universal compromise way is to normalize 32 bit files in Reaper, Soundforge, Izotope RX, and export with preserved Timecode in same 32 bit float depth.
In this case you can work with file in Resolve and same time you keep your actual 32 bit source data formally untouched, so it may be useful for other apps and processing plugins that operate in 32 bit float.
--------
You may notice in older Curtis Judd's videos he discover that the normalized sound character is slightly different at different gain amplification levels.
My guess it is not because A/D converters, but because analogue circuits inside microphone body itself.
In 32 bit at any gain level we always capture all possible dynamic range of microphone and so we can always digitally normalize it without any loss. So when we provide more gain to microphone we actually provide more voltage to its circuit and so internal electronic components inside microphone reacts somehow to this voltage and adds some specific character to the sound.
When we provide less gain/less voltage to its circuit, the internal electronic components inside microphone reacts somehow in different way to this voltage and adds some other character to the sound (less bass, less warm in that video example)
So probably with 32 bit recorder we can shape the sound of microphone by experimenting with different gain settings and find which gain will produce better sound on specific microphone model for specific needs.