I think it was you who did some research into what was causing the problem and resolved it by hacking lame to produce VBR nogap files, then running them through a VBR header fixer. Then I looked into the lame source and found that the problem was that it was always writing the VBR header to show one fewer frame than there actually was. The reason it did that was because it failed to count the VBR header itself, which is a valid mp3 frame. It turns out that that bug existed in all VBR files, but it was only really noticeable on nogapped files, as chopping the last frame off a track that has basically already faded out was unnoticeable.

It's been fixed in CVS for a while, but I don't think that there's been a release since then. Actually, it looks like there's just been a new release. Checking to see if the fix is in there and if VBR nogaps are reenabled...

It looks like it, but I'm too lazy to compile it and make sure.