I like the "no kings" idea for castleable kings - though I think with smaller pawn sizes, that gives another opportunity for compression:
- king that can't castle: king
- king that can castle queenside: any other piece in king's position, no king of that color on board
- king that can castle kingside: black pawn on 1st or 8th rank
- king that can castle: white pawn on 1st or 8th rank
Because you are often going to have positions where the king can castle kingside and positions where the king can castle either way, this should maximize how often you manage to save space w/ the pawns.
Another thought I had (which might contain other problems, not sure yet) is to use pawns on the 1st or 8th rank to denote pieces which are in their starting position - the decompression algorithm can then derive what piece it is based on the known starting position. Once we start having data saving because of pawns taking less bits, we want to be able to use them as much as possible to save space.