What I remain to be convinced of is that 32bit floating point somehow improves "resolution" of low level signals through the A>D conversion. This claim seems to contradict Nyquist–Shannon sampling theorem.
Not at all. There are two different variables affecting audio sampling.
1) Sampling frequency: Nyquist-Shannon deal with it. The sampling frequency limits the bandwidth of the audio signal we can sample and reconstruct failthfully. In the simlpe case(*), the maximum audio frequency we can represent.
Note that Nyquist-Shannon don't deal with the number of bits of the sampling. Actually what they say can be applied to an ideal, analog sampling system in which you are not quantizying the signal to a distrete number.
2) Resulution. That's the number of bits. It doesn't affect the signal bandwidth, but it's signal to noise ratio. When you have a poor resolution, your reconstructed signal will have "steps" which means a kind of noise will appear: quantization noise.
So, both analog and digital systems have a noise floor. In the case of digital systems that noise floor (ignoring of course the noise inherent to the analog stages like microphone, preamp, A/D converter, etc) is determined by the number of bits you are sampling.
The reason why you don't want to record with a level too low in a digital system is that. Low sounds will have a much better S/N ratio than the very quiet ones.
So, increasing the number of bits is beneficial because it will increase the S/N ratio. But it won´t affect the frequency response.
Using 32 bit floating point doesn't mean that you will get the full dynamic range of a 32 bit floating point number, of course, but it means that internal calculations won't overload the container (the simplest case would be making a stereo mix of the input channels while you are recording) and that in case you have a special A/D converter arrangement (like dual A/D converters or a higher precision A/D converter with a 32 bit sample format) you can effectively obtain a larger resolution and you can store it without the limitations of a 24 bit sample.
I hope it's more clear.