PVS re-search at PV nodes

[hgm]

Well, the point is how unstable the search is. Which percentage of null-window fail-highs turn into open-window fail-lows. If it is only 10%, (which I would consider pretty unstable), you would do the open-window reduced search in vain 90% of the time, while in only 10% of the case you would save yourself doing an unreduced search. So if the reduced search needs to be more than 10 times faster as the unreduced search before you break even.

[lucasart]

Reductions (Nullmove, LMR etc) and pruning (futility, LMP) create an unstable search.

Yes, as well as *anything* that treats PV and non PV nodes diffently. Example, even without reductions:
(1) you are at a PV node and searching the second move (zero window)
(2) if doesnt' fail low. you research wih a full window. now it fails low !?

the sub tree that was researched with full window has its root a PV node in (2) and non PV node in (1). So the search can behave differently and give inconsistent results between (1) and (2) for that reason only...