If you decode the same files twice with the MAD plugin they are not the same!

Any explanation?
Very weird thing.

Is Auto Clipping Attenuation active?

At first it was...then I also thought that was the cause...but it wasn't.

It probably uses random dither data.

Maybe you can check this random dither issue, just decoding some digital silence encoded files.

This is a long-known issue...

MAD does not reset the data for its dithering with each track, instead it accumulates data throughout the whole Winamp session. (although I have no idea what that data is )

AFAIK if you open Winamp, decode to one file, close and open winamp again, and decode to another file, the files should be identical. B)

Indeed this is the case.

First-order noise shaping algorithms work by keeping an error term that is the sum of all the previous errors caused by bit depth truncation. When each new sample is truncated, it is done to minimize this running error sum rather than simply rounding to nearest (which only minimizes the error for that sample). The net effect of this is that the absolute amplitude of the noise resulting from truncation is slightly higher (compared to pure rounding) but falls off at 6 dB per octave toward lower frequencies. The final result is that there is less audible noise at important parts of the audio bandwidth.

In theory, if you decoded two contiguous tracks using this method and you reset the error term at the track break and played them gapless, there could be an audible click at the point that the error term was zeroed because there would be a DC jump. However, in the real world with real signals I believe that this is highly unlikely. And, in fact, I think that there is a greater chance for the case where a track that ends very loudly might make a small click in the silent beginning of the next track.

Best to stop losing any more sleep and just leave everything 24-bit.