Leela's search depth and forced win/draw in chess arena

So, ií I understand correctly, the reason why Leela seems to calculate "slowly" is that the simulated games (if I take the naive understanding of Monte-Carlo tree search) are not counted in the total amount of nodes, as they do not consider all opponent's responses.

But the depth shown in the GUI is probably correct, as the engine is able to produce the principal variations. So if I see "depth=5/10", can I be assured that all variations in depth 5 were considered thoroughly? In comparison, Stockfish shows 12/25, which I guess are also thoroughly calculated depths (except for alpha-beta cutoffs, which is an "exact" heuristics).

In that sense, if there was a forced mate/draw in 5 moves, would the engine spot it and actually play that variation instead of using any heuristics? Because I have seen these neural engines picking up the "worse" continuation just because it is "easier to play", and the rationale being that mate in 3 is valued as good as mate in 2.