All in _processMathOne(). Two major issues I see:

• If a pair is not available in one of the sets, it assumes 0, whereas it should instead use the effective kerning value for that pair, which might come from a glyph-class, class-glyph, or class-class. This one is not too hard to fix.
• If the two sides have different groups, it arbitrarily merges them. Instead, we should classify glyphs into new groups and translate kerns over. I started https://github.com/behdad/fonttools/blob/master/Lib/fontTools/misc/classifyTools.py for that purpose, though I want to extend it to support named sets.