On my laptop, SF 17 says that 15... Qxb2 was the blunder.
Qxb2 looked fine on lower depth, but then White's eval rose to over +1 after a few minutes.
15... Rxd3, instead, seems to hold. Exchanging queens.
16. Nxe5 ...Rd4
17. fxg7 ...Bxg7
18. Nxf7 ... 0-0
I haven't looked past that, but I'm thinking Black is fine from there on out, now that he has avoided the ...Qxb2 line. Black's down a pawn, but he's got the bishop pair on an open board. He'll snag that pawn back at some point.
Wow. Well, you may be right that it's playable. I note that only one human plaer has ventured that queen trade out of 62 database games, and it does look like a sad solution. We're down a pawn, and e6 looks pretty weak, plus White's pieces are more active for the time being. I'm sort of astonishedt he engine calls it as close to equal as it does.
The thing is - do you really want to sign up for learning all this theory just to suffer through a +0.7 position? I mean, you could learn the same amount of theory in the Poisoned Pawn and get a +0.1 position. Still, I admire the folks who take up the challenge.
I'd say the opposite trend seems to be (surprisingly) true: a lot of openings and defenses that were thought to be dubious have been proven to be playable by modern engines.
Yes, very often true. There are some lines though, that are pretty much refuted, though I'm struggling to think of one that was ever particularly popular. One I'm rather fond of that seems, alas, to be busted, is the Polugayevsky Variation of the Najdorf. Last time I played it in a daily game, I lost, and, even looking later with an engine, I couldn't figure out what I did wrong other than choosing to play the Polugayevsky.
Then there are the middle cases, of which the King's Indian is perhaps emblematic. It's not refuted, and I think a lot of people think that engines' high opinions of white's chances are inflated. But it's clear that White has a number of ways to get a position, that, without being fully winning, can make Black's life harder than it needs to be. On the other hand, at the amateur level, it remains popular as ever, and since sub-2000 players are typically out of theory by move 12-14 anyway, those engine problems doen't seem to have any serious negative effect on Black's practical chances.
And this is the crux of the whole situation here. Even top level players are out of theory after 30 to 40 moves, usually much sooner. Engines, of course, are never out of theory; they're incapable of 'forgetting' anything. So an engine that finds a flaw in a line that's 40+ moves long has little relevance for a human player. I remember one time Fabi won some game in a tournament, and afterwards the commentator told him he missed a mate in 24, or some such ridiculous number, that Stockfish had found. He just laughed. He wasn't too upset about missing the move. It was a rather 'human' mistake to make.