It looks like Stockfish still thinks black is winning. I'd be curious to see if Leela Chess Zero also thinks black is winning.
Here's what's confusing me: if the chess engine has an initial numeric value for the current position, and the positions going out to the end of its search all have the same value or worse, shouldn't it be pretty trivial to conclude it is not making progress?
https://lichess.org/analysis/8/1p3K2/kPp5/2P5/2P5/1p6/1P1Q4/...
It's not visible on the board itself, but this position's 50-move rule counter is set to 98 ply -- two ply short of a draw. (That's the 2nd-to-last-field in the FEN representation -- the editable text field just below the board, and also in the URL). The only way to avoid that draw, as Stockfish does in fact suggest, is to a sacrifice a piece in a specific and ridiculous way, to compel black to re-capture on the 100th ply (captures reset the 50-move rule). If you edit the 50-move counter down, Stockfish instead shows you the simple mate-in-3.
Nakamura famously tricked a different chess engine, relying on this 50-move draw avoidance (as one component of the hack) -- a sibling thread is discussing that example: