lame-3.96.1 here.

And the Xing headers it produces do NOT include the Xing header frame in either the framecount or bytecount fields of the Xing header.

I also added a +/- 1 fuzz factor the the VBR seek table validation code in mp3tool, as their math seems to produce ever so slightly different offsets from what mp3tool calculates, likely due to rounding somewhere.

Cheers