Computers Get It Wrong Again

I WAS losing for much of today's round-five game at Guernsey.
But after sacrificing the exchange to slow White's attack, I reached the following position:

Black to make his 45th move in Jan Maarten Van den Boogaarten (2009) - Spanton (1923)

The obvious move is 45...a5, after which my main analysis engines, Komodo9 and Stockfish9, reckon White's advantage is worth about three-quarters of a pawn.
Instead, I played:
45...Bd5!?
After the obvious …
46.Rxa6
… I sprang a surprise with …
46...Bxg2!
The game continued:
47.Qe5+
Forced as 47.Kxg2?? loses the rook.
47...Qxe5 48.fxe5 Bd5
When I loaded the game into ChessBase, the engines reckoned White was winning. However, as in my round-three game, this is an ending computers do not understand.
The engines eventually tone their evaluation down to plus-over-minus, but most humans, I am sure, would realise this position is dead-drawn.
49.h4 Bc4 50.Rf6 Bd5 51.Kg3 Bb3 52.Kf4 Bd5 53.Ke3 Be6 54.Kd4 Kf8 55.Rf2 Kg7 56.Kc5 Kf8 57.Kd6 Kg7 58.Rb2 Bc4 59.Ke7 Be6 60.Rb6 Bc4
But not 60...Kg8?? 61.Rxe6, which wins for White.
61.Rd6 ½-½
Going back to the position after 45...Bd5!?, Black had a major alternative in 46.Rc3. I had planned 46...Qe2, when the engines give a move I had not considered: 47.Rg3.
They continue 47...Qe4 48.Qe5+ Qxe5 49.fxe5, evaluating the resulting position as winning for White.
I find that hard to believe, and they have been unable to prove a win in analysis (with me taking the Black pieces). Even after reaching a position with White having just h pawn and rook against g pawn and bishop, they still believe White is winning, but it ain't so!