Engine gives -3 but it looks drawn

Am I missing anything? It says black has a substantial advantage but I don't see how they break through.

thanks!

I agree, it's totally drawn, the computer just has trouble to see that it cannot make progress (as said due to the horizon effect). Black's king has no squares it can use to enter the opponents position (e.g. with a black king on any of h1,h2,h3,h4,g4,g3,g2,g1,d4,c5,c3,c2,c1,b3,b2,b1,a3,a2 or a1 this should be a win). The only way to make progress would be to win the pawn on f4 with the bishop, when the pawn on g5 would also fall. But white can always keep his king on f3 or g3/e3, which prevents that. One could think of sacrifices of the bishop either on f4 or g5, but that does not make progress (on Bxg5 white plays fxg5 and no progress can be made, on ...Bxf4 Kxf4 nothing happens, either) and would perhaps even run a very serious risk of losing to some c4-c5 idea (did not check this, but that would be a serious threat).

UglyDuckling is correct. Sadly, even if you trade your Bishop for the f-pawn by recapturing with your e-pawn, you just can't win because your king can't assist or white will play c5 since you let your king wander too far from the potential passers on the queenside. Additionally, if trade your bishop for your opponents g-pawn, recapturing with your king, you get to queen your passed pawn but white gets to queen back with check, gaining a perpetual. If only you had a way to gain a tempi in that line you could stage a brilliant endgame win! :(

Zerazork if black plays Bf6 at some point, white just won't take it. There's no way for black to create any winning chances, and no such concrete calculations are needed.

Nothing to do with the horizon effect. It's the contempt configuration.

Contempt causes the engine to over-evaluate the pieces of the side it is evaluating. In this case, because black is a bishop up, contempt = 20 (Stockfish 9's default contempt setting) causes the evaluation to heavily dip in black's favour.

27 Bxh4 and it's easy. ;)

A high evaluation which has no trend is basically a draw. Or as our engine-free predecessors stated: a good position which cannot be improved is not a good position.