It seems to me something that is not being appreciated here is that the accuracy of the sampling (that is to say, how closely the series of digital points matches the natural wave) is a function of how much granularity is allowed in the mapping space. And a key point is that an increase in EITHER the X or Y axes will result in greater granularity. In other words, if you increase the samples on the X axis to 96,000 per second, that means there are more Y values taken per second, so they can more closely match the natural wave. It isn’t about “hearing” pitches at 48K. Nobody can hear those frequencies.
44.1K x 16 bits = about 3 billion potential data points per second
48K x 24 bits = about 800 billion potential data points per second
96K x 24 bits = about 1.6 trillion potential data points per second
The difference from 3 billion to 800 billion is big. I’m not sure doubling that would make much of a difference.