It's also that PGN allows you to encode illegal moves, which is important if your dataset contains over-the-board games. That you only got to about 10-12 bits

It's also that PGN allows you to encode illegal moves, which is important if your dataset contains over-the-board games. That you only got to about 10-12 bits per move is actually kind of sad in a way, because it means you're not doing substantially better than the approach where you record a move as just (delete-piece-at ____, recreate-that-piece-at ____) 12-bit pairs, where castles are implicitly only recorded by moving the king further than usual and underpromotion has to be explicitly recorded with an extra 12-bit move that is otherwise semantically impossible.