CDs accurately reproduce sounds up to 22KHz which is good enough for mere mortals. I suspect an audiophile’s hearing goes a bit beyond that frequency.
I think that CDs have a 50,000 sample per second rate, which give a Nyquist frequency of 25,000 Hz. The trouble is that any signal content at frequencies above the Nyquist rate are “aliased” - after sampling, a 26,000 Hz frequency component is indistinguishable from a 24,000 Hz frequency. And if you had significant signal level at 49,000 Hz, after sampling it would sound like a 1,000 Hz signal after sampling. I don’t know the technology being used to address that issue; in the old days an analog filter was the best that was practical, but today it should be easy to sample initially at 100,000 per second - or even double that, I suspect - and use digital processing to kill the frequency content above 25,000 Hz - finally reducing the sample rate down to the target 50,000 per second with essentially no aliasing, and no attenuation below the 22 kHz frequency you cite. If, indeed, not flat all the way up to 25 kHz.The other issue is the reconstruction of the waveform in you CD player. The trivial approach to the D-A process is to always output a voltage proportional to the last sample, thus producing a step waveform - and assuming that the ear can’t tell the difference. If you wanted to go full-on purist, you would create a digital “filter” with a sin(x)/x impulse response function, where x is the time after - or before the nominal time you assign to each sample as you are playing it back. When x is zero, sin(x)/x is unity, and when x is an integer multiple of pi, sin(x)/x is zero. Thus, each “dog has his day” when its value is the only thing that matters to the output - but at other values of x (which imply a higher sample rate than the 50,000 per second on the CD), theoretically all past and future samples have a more or less meaningful, positive or negative, effect on the output of the “filter.”
There is of course a practical limit to how many samples you could give influence on the filter’s output, so you couldn’t implement sin(x)/x perfectly, and there is a limit to how high an output sample rate it would be considered “worthwhile" to generate. But subject to those limitations you can theoretically reproduce the original waveform that you digitized. It would seem improbable that anyone could hear any difference between 100,000 samples per second and any higher rate, so it would seem that such a filter if implemented would be used to only double the sample rate. But in a world of gigahertz graphics chips . . .