A PG book could support atomic, as Atomic positions are indistingishable from Normal positions. (At least if you ignore legality of the position, which Polyglot format does. You could not have Nalimov tablebases for Atomic, though, as Nalimov format does not allow the Kings to stand next to each other, while in Atomic this is not only perfectly legal, but often the only way to salvage a draw.)
But for this reason it would be quite confusing to have such a book, as it could also be probed in Normal games (or Suicide, Losers, 3Check), and suggest moves that might be disastrously wrong.
I guess I should change the book-probing code to XOR in a key that depends on the current variant, to prevent one variant from using a GUI book written for another variant that shares positions, but has different rules. For Chess960 this variant key could be 0, as identical Normal and Chess960 positions are indeed fully equivalent. (In other words, Chess960 is not really a variant, it is an extension.)
I generalized PG format such that it supports books for all variants supported by WinBoard. For the extra piece types and board squares it uses rotated versions of the standard keys (in steps of 8 bits), so that there is no problem on boards of upto 128 squares with upto 24 piece types. And I indeed do have a Polyglot book for Xiangqi. (Did not make it with Polyglot, though!
) This did require a different coding scheme for the moves, though. I reqisitioned the extra code space occupied by the promotion piece for this.