Reversal is the one edit that runs against the grain of a decoder. Decoders go forward; a reversal needs the end first.
How it works
The file is split into windows of at most 30 seconds. Those windows are visited from the last to the first, and each one is decoded into a buffer, mirrored in place, and handed to the encoder before the next is touched.
Memory is one window rather than one file. That is the whole reason a 45 minute source behaves the same as a 45 second one here.
Seams
The windows are butt-joined, not crossfaded, because they are contiguous samples of one waveform: a fade at the join would attenuate real audio rather than hide anything.
Making that safe means the boundaries have to be exact. Every window is measured in whole frames rather than in seconds, and a lossy decoder is asked for a quarter second of run-up before each window so its first frames are correct rather than the garbage a cold start produces.
What reversal is good for
A reversed cymbal or piano chord fading into a downbeat is the oldest trick in the book and still works. Reversing speech makes an unintelligible texture that keeps the rhythm of the original.
The more useful engineering trick is the reverse reverb tail: reverse a phrase, add reverb, reverse the result, and the tail arrives before the sound that caused it.
Format
The output keeps the container of the source, at a bitrate matched to it. Reversing an MP3 means decoding and re-encoding, which costs one generation of quality; reversing a WAV or FLAC does not.
If the file is going through further processing, convert it to WAV first and reverse that.