You’re spot on about the performance reason I didn’t want to do this originally, but I did some testing and turns out move generation in the library I use is Blazing Fast and wouldn’t be a bottleneck
<title>Posts on A blog</title>
<link>https://mbuffett.com/posts/</link>
<description>Recent content in Posts on A blog</description>[1] Fabian Giesen's sample implementation is already as good as is, and also contains a good alias table implementation if you want multi-symbol inputs: https://github.com/rygorous/ryg_rans/blob/master/rans_byte.h
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.