Is the number of legal chess positions odd or even? Two positions are not the same if they differ in castling rights (i.e. whether K or R have actually moved) or en passant capability (i.e. whether the move can actually be made) or who has the move.
The number of times position has been repeated is not part of the definition of position for obvious reasons; nor is the number of moves since last capture or pawn move.
I don’t have the answer myself. This is hard but doable with a bit of programming I think.
EDIT: The answer so far and the comments are clearly heading along the right lines, although not always correct yet: terrific work. I want to give multiple +1s! I will summarize the key points to help focus the work. Pairing mirrored positions is essential i.e. those which allow triangulation. I term positions which have no mirror image “vampires” :-) Any vampire (except for the starting position) must be immediate offspring of a vampire, so it’s maybe easiest to just focus on them.
Someone touched on castling, which is very important. We can partition the vampires into 16 different clans. The head of each clan is an "initial game array" position (i.e. with 32 pieces on apparently original squares) but with specific castling rights disabled (by some initial dance of rooks and knight taking 2n.0 moves).
In a clan, if a castling right remains, then that implies that the rook's pawn can safely capture off the file, the f-pawn can move forward a single square or capture or (on the kingside) the bishop can be captured, without allowing triangulation. If the right is lost then the rook's pawn can only move forward a single square, and the bishops cannot be captured.
En passant is less important but amazingly there are vampire en passant positions. For example:
r1bqkb1r/2pppppp/8/1pP5/8/8/PPPPP1PP/R1BQKB1R w KQkq b6 0 12
where White retains at least one castling right, Black has at least queenside castling rights and White can capture e.p.