Blunder? The Engines Disagee

PLAYING in the Highlands Open in the southern Czech town of Trebic.
I had a lucky win in round one today against a much lower-rated opponent.

Position after Black exchanged queens on f2 in Spanton (1927) - Jan Divis (1526)

I do not know what was going through JD's mind here, but I guess it was an example of relaxing when the danger seems over, as the game saw:
33...Kf6??
Strangely, the analysis engines Komodo9 and Stockfish9 flash yellow, but not red, when this move is played. Black had to play 33...b6, or 33...Bf6 (threatening b2).
34.Bxb7 Ke5
The engines give 34...Bc1 with a big, but not winning, advantage for White after 35.b3 or 35.Be4 (threatening h7).
35.Be4 h6 36.a4
This is why I am positive the position is winning for White. By creating a passed pawn on the queenside, and having the threat of a passed pawn on the kingside, White makes it impossible for Black to defend on both sides of the board.
36...bxa3 37.bxa3 Kd4 38.a4 Bd8 39.Kf3 Ke5 40.Kg4 Kf6 41.h4 Ba5 42.Kf3 Be1 43.h5 Kg5 44.g4 Bc3
During the game I was worried Black could prevent my king penetrating to the queenside by playing 44...Kf6 45.Bf5 Ke5. But, as Ian Heppell pointed out, White wins easily by creating a second passed pawn with g5. In fact, White can play 46.g5! immediately as the bishop is immune, ie 46...Kxf5 loses to 47.gxh6.
45.Bf5 Kf6 46.Ke4 Ke7
The game finished:
47.Kd5 Be5 48.Kc6 Kd8 49.a5 Bc3 50.a6 d5 51.a7 1-0
That was pretty easy. So why don't the engines rate Black's 33rd move a blunder? Could it be the horizon effect of the winning process being long drawn-out, even though a human can quickly assess the position as (almost) certainly winning?