Once audio has undergone lossy encoding, it no longer has an inherent bit depth. The encoded bits can be thought of as a mathematical description of how to reconstruct an approximation of the waveform that went into the encoder. The decoder usually calculates the reconstruction using floating point or integer math with at least 24 bits of precision. We chose to output 24-bit PCM in order to avoid introducing unnecessary dither to the decoded audio. Dither is basically a necessity when dropping audio down to 16 bits per sample from a higher bit depth, but it’s problematic when talking about lossy decoders:
If a lossy file was created from a 16/44 source like a CD, the audio already had dither applied to it as part of the mastering process. Decoding to 16 bits per sample would add additional dither.
If a lossy file is the product of a more modern mastering process like Mastered for iTunes, it’s likely that the original source for the encoding was undithered 24+ bit audio. In that case, decoding that file to 16 bits per sample would introduce dither where there was none before.
In addition to improving sound quality for lossy files, this change also simplifies some of the associated code, making everything a little more efficient (and less work to maintain).