Stockfish lost forced mating sequence

I just saw an announced mate in something like 22 moves in a client analysis switch back to a large negative evaluation (in the negative fifties). Is that supposed to happen, or is it a bug? I would have expected such a line to be retained, and so far that was always what happened.

(I'm not talking about a server analysis, where the evaluation in large print sometimes seems to reflect the server analysis while the variants shown change dynamically and reflect the client analysis -- this is a pure client analysis (at least I didn't request a server analysis).)

(I tried to get create a study with the variants I was analysing, but clicking on "New study" only put the game in the study, not the variants, so unfortunately I have nothing to reproduce this with.)

Cannot say much without the actual game or position.

I didn't quite reproduce the phenomenon, but a similarly counterintuitive one: The mate announcement jumped from #-19 back up to #-21. This occurred in analysing this game:



I continued after 50... c3+ with 51. Kd1 Rh3 52. Rb7+ Kc4 53. Ke1 and then let Stockfish chew on the position for a while; the mate was announced for 53... d3. Since the analysis is different every time, this isn't a sure way to reproduce the phenomenon, but perhaps having a position helps nevertheless. (This may also have been the position in which the forced mate was temporarily lost altogether.)

The engine is going to show you what it thinks best play by the opponent yields. For something insane like a mate in 22 there are an absurdly huge amount of variations way beyond what any engine can calculate (at least in remotely complex positions). But all lines that it had calculated lead to shorter mates, so that was the best option. But then it churns a bit more and finds that the opponent can somehow avoid the mate, and so the line changes to the new best which is e.g. -50 or whatever.

In this particular case it's probably best to download any engine with 6-man tablebases or consult online tablebases for 6-man endgames. I know the position has 7 pieces, but not for long.

Why does server analysis not use 6-man syzygy?
It only needs the "scores" (win/draw/loss) "on hand", and the table with best moves and depth numbers can be stored apart on the server.
That way we would always have mate scores or 0 in 6-man endgames.