First of all, I love the new puzzle filters.
But I think I have found the first endgame one that ran across the engine horizon effect (cannot be avoided...).
lichess.org/training/rqa2l
The accepted solution is 1...f5, after gxf5 Nxh5 Black's path to victory is pretty obvious (White will soon have to sacrifice their bishop for a running h-pawn).
However, 1...g6 also leads to a Black victory, though Stockfish evaluates it as 0.0 at first (and it loses the puzzle).
Here, the plan (after hxg6 Nxg6) is a little more complicated.
First, you need to manoeuvre your knight to attack the pawn on b2 so that it is forced to advance to b3.
Next, the knight is manoeuvred to d4 (dark square! No way to premanently control it for the bishop), from where it at the same time attacks b3, covers f5 to prepare an exchange of pawns there and covers both e6 and c2, the only two squares from which the bishop can reach both g6 and f5.
White is then forced to either allow the exchange (freeing the Ph6) or lose the Pb3 (freeing the Pb4), after which the normal plan from the accepted solution comes into play (trade one pawn for the bishop, promote the other).
All in all, this seems to take too many moves for the engine to see within the horizon of puzzle generation (lots of different paths to execute the first two steps due to many possible bishop move options). However, it looks like a forced win nonetheless (black evaluation advantage after hxg6 Nxg6 is constantly increasing if you let Stockfish run, too) and the puzzle should thus be deleted.
But I think I have found the first endgame one that ran across the engine horizon effect (cannot be avoided...).
lichess.org/training/rqa2l
The accepted solution is 1...f5, after gxf5 Nxh5 Black's path to victory is pretty obvious (White will soon have to sacrifice their bishop for a running h-pawn).
However, 1...g6 also leads to a Black victory, though Stockfish evaluates it as 0.0 at first (and it loses the puzzle).
Here, the plan (after hxg6 Nxg6) is a little more complicated.
First, you need to manoeuvre your knight to attack the pawn on b2 so that it is forced to advance to b3.
Next, the knight is manoeuvred to d4 (dark square! No way to premanently control it for the bishop), from where it at the same time attacks b3, covers f5 to prepare an exchange of pawns there and covers both e6 and c2, the only two squares from which the bishop can reach both g6 and f5.
White is then forced to either allow the exchange (freeing the Ph6) or lose the Pb3 (freeing the Pb4), after which the normal plan from the accepted solution comes into play (trade one pawn for the bishop, promote the other).
All in all, this seems to take too many moves for the engine to see within the horizon of puzzle generation (lots of different paths to execute the first two steps due to many possible bishop move options). However, it looks like a forced win nonetheless (black evaluation advantage after hxg6 Nxg6 is constantly increasing if you let Stockfish run, too) and the puzzle should thus be deleted.