Actually, I've been thinking about this a bit more.
The reasoning that 16 bit = 96 dB dynamic range and so using more than 16 bits for anything with a SNR less than 96 bits is maybe a bit misleading or oversimplifying.
One, the noise floor of an actual recording is a complex function of frequency, and different shapes of noise (that may still integrate to something around-96 dB) will be audible to different degrees. Digital quantization noise may be be more objectionable than analog noise that is summarized by the same single number.
Two, I think it's pretty well established that dithered audio sounds better than undithered. Dithered 16 bit audio has a signal to noise ratio of about 91 dB (depending on the exact form of dither). So maybe the noise of the R09's A>D is just a bit (hah!) below the 16 bit noise floor after all.
Also, going back to my earlier post, I do think all noise is additive, be it analog or quantization noise. So it seems like 24 bit could be quieter than 16 bit even with an input that is technically noisier than the theoretical limit for 16 bit.