What is your experience with LMP? Is it common knowledge that LMP in PV nodes is a bad idea?
I'm not sure LMP in PV nodes is a 12 elo regression by itself. Theoretically, when you don't do something at PV nodes, it should make zero elo difference, especially at high depth when PV nodes are fewer.
But it causes technical problems (eg. truncated PVs, fixed depth search can prune all moves and not return a best move etc.). That's why I would disable it. Perhaps LMP at PV nodes was causing some problem, like the best move not being updated in some cases, because you prune at Root.
Why don't you try to disable at Root only, and leave it at PV nodes, just to see if your 12 elo regression isn't just there?
I should have mentioned that in texel I have a separate root search function that handles iterative deepening, aspiration windows, time management, multi pv, and other things that are only relevant in the root node. This root search function does not have any reductions/extensions except LMR, so LMP is already disabled at the root.
I don't do LMP at a node unless at least one legal move has already been searched at that node, and the returned score was not a negative mate score. Also, I don't do LMP for the hash move, killers, SEE>=0 captures, promotions, passed pawn pushes and moves that give check. Additionally, if alpha or beta is a positive or negative mate score, I don't do LMP.
It is possible that there is some unknown problem in my LMP implementation and that skipping LMP in PV nodes somehow hides the problem. I don't know what that problem would be though.
It is also worth noting that in an attempt to get the most elo improvement per invested CPU cycle, I run almost all my tests at very short time control (80ms/move) and against the previous development version of texel. While this has worked well on average for texel so far, it is of course possible that for some changes it gives wrong results, so maybe this change would be far from +12 elo if measured at longer time control against different opponents.