Crazyhouse analysis strange behavior

Not analysed games for a while, and bumped into the next:
lichess.org/analysis/crazyhouse/r1bqkbnr/pppp1ppp/2n5/4p3/4P3/5N2/PPPP1PPP/RNBQKB1R/_w_KQkq_-_2_3
local analysis (Stockfish 11+HCE, depth 26) indicates that white has +2.7 advantage and suggests Bc4, but right after the move the advantage drops to +0.4.
Okay. Let's say this is a depth issue.
lichess.org/analysis/crazyhouse/r1bqkb1r/pppp1ppp/2n2n2/4p3/2B1P3/5N2/PPPP1PPP/RNBQK2R/_w_KQkq_-_4_4
The depth is 66 (!), the advantage is #6 and suggestion is Bxf7+, but right after that it is not #5, but -2.8. Also, followup moves are also unobvious, but I'm not that good to be sure that there are better. If someone could point how it is #6 in the first place, I would be grateful.

I understand that the SF is intended for classical chess, but I remember that earlier there was no such behavior. Now there is almost no sense in using the engine to analyse crazyhouse openings.

For me, the first position suggests Nc3... and then right after I move the evaluation changes to 1.1.
The second position suggests Ng5 for me and with no sudden changes in evaluation after I move.

I think there is a confusion in the Lichess server evaluation cache. They give sometimes evaluations from their cache that belong to another variant. Some other players and myself have already suggested that this cache should be deleted and all the wrong evaluations redone, but probably this would require much CPU time.

@superflash2022 said in #2:
> For me, the first position suggests Nc3... and then right after I move the evaluation changes to 1.1.
> The second position suggests Ng5 for me and with no sudden changes in evaluation after I move.

Just checked in the mobile app, the second position is +3.1, still suggests Bxf7+, then drops to -2.7.

@zen_queen said in #3:
> I think there is a confusion in the Lichess server evaluation cache. They give sometimes evaluations from their cache that belong to another variant. Some other players and myself have already suggested that this cache should be deleted and all the wrong evaluations redone, but probably this would require much CPU time.

Looks like you're right. Sad, the SF analysis is very helpful to improve. Or was.