In this screenshot: ibb.co/n26xpS, the evaluation #-3 for the move 22. Kh4 is correct. This screenshot: ibb.co/hJY267 shows the three different moves with which black can win in three moves after 22. Kh4. However, note that 22... g5+, the continuation given in the first screenshot, isn't among them. Indeed, 22... g5+ takes one move longer, as seen in this third screenshot: ibb.co/feUaR7, since white can respond with 23. Bxg5 (rather than 23. Kxg5 as given in the first screenshot), giving check and delaying the win by one move.
It seems that this shouldn't happen -- for a forced win in 3, the engine shouldn't just output some line that ends in a win after 3 moves, but one of the forced lines. The moves in the variation should be the best ones it found, and the information that 22... g5+ isn't the best move must have been generated in determining the evaluation #-3.
Here's the game: lichess.org/CgrPQeBq
It seems that this shouldn't happen -- for a forced win in 3, the engine shouldn't just output some line that ends in a win after 3 moves, but one of the forced lines. The moves in the variation should be the best ones it found, and the information that 22... g5+ isn't the best move must have been generated in determining the evaluation #-3.
Here's the game: lichess.org/CgrPQeBq