Selfplay outputs chess960 "capture rook" moves

[Mardak]

lczero-client converts the gameready moves to pgn:
https://github.com/LeelaChessZero/lczero-client/blob/33e15f9ec6ce147de76802a0d173804845de219c/lc0_main.go#L403

A quick test:

convertMovesToPGN(strings.Split("d2d4 g8f6 g1f3 b7b6 g2g3 c8b7 f1g2 c7c5 e1h1", " "), "")

errors with:

movstr: chess: invalid move e1h1

Perhaps either https://github.com/Tilps/chess needs to get updated or selfplay should output e1c1/e1g1:

lc0/src/selfplay/loop.cc

Line 120 in 58428e8

for (const auto& move : info.moves) res += " " + move.as_string();

[mooskagh]

I didn't think about that, good catch. For now I think we should just convert moves in selfplay, or maybe introduce a chess960 flag in selfplay too.

[Tilps]

selfplay always starts from standard startpos so for now it makes sense to convert moves.
When we add support for starting from x-fen, we should make the output conditional on the x-fen being non standard.

[borg323]

Surely there must be a simpler way to do it than this:

diff --git a/src/selfplay/game.cc b/src/selfplay/game.cc
index f865a47..f0d725b 100644
--- a/src/selfplay/game.cc
+++ b/src/selfplay/game.cc
@@ -213,6 +213,18 @@ std::vector<Move> SelfPlayGame::GetMoves() const {
     flip = !flip;
   }
   std::reverse(moves.begin(), moves.end());
+  PositionHistory history = tree_[0]->GetPositionHistory();
+  for (size_t i = 0; i < moves.size(); i++) {
+    Position position = history.GetPositionAt(i);
+    ChessBoard board = position.GetBoard();
+    if (position.IsBlackToMove()) {
+      moves[i].Mirror();
+      moves[i] = board.GetLegacyMove(moves[i]);
+      moves[i].Mirror();
+    } else {
+      moves[i] = board.GetLegacyMove(moves[i]);
+    }
+  }
   return moves;
 }