Why is MD capacity always the same (74:59 exactly)...

On a frame by frame (11.6 ms) basis, ATRAC is always trying to give as much resolution as it can to each spectral component that it deems important. However it is continually short of bits, and except for the case of zero input, or a very simple signal (e.g. a few pure tones), ATRAC is using all the bits per second it has available. If it had more bits per second available it could use them, and record the signal with better sound quality.
In Sony's AES paper on ATRAC they talk about the bit allocation algorithm, and how it in general allocates more bits than are available, so they have a method for subtracting an offset from the allocations in order fit within their bit budget for a sound frame.
In the zero or simple input case, there can be unused bits in some frames, but since there is an unvarying (i.e. fixed) relationship between 11.6 ms of audio and a 424 byte "sound group" on the disc, the system is not flexible enough to actually take up less disc space when it doesn't need it.