The precision of floating point arithmetic and the size of the quanta used for digitization are separate design decisions.
If 24-bit audio is clipping, it is not due limitations in dynamic range, but due to setting the gain too high. At 24 bits can set it much lower without concern for running out of dynamic range at the low end, and then set it at say -3dB from peak during the mastering process.
Human hearing has up to about 20 bits of dynamic range, from a pin drop to a 747 taking off at close range. 24 bits provides 16x that amount.
What is true is that it is worth recording at a larger quantum size than the size for the final distribution. This provides extra padding to reduce digital noise by reducing floating point round-off errors. Dither down to the desired quantization as the final mastering step.