Simplify Probcut Conditions

[xu-shawn]

Passed Non-regression STC:
LLR: 2.93 (-2.94,2.94) <-1.75,0.25>
Total: 248608 W: 64453 L: 64467 D: 119688
Ptnml(0-2): 784, 29486, 63796, 29436, 802
https://tests.stockfishchess.org/tests/view/6759244386d5ee47d95420ea

Passed Non-regression LTC:
LLR: 2.94 (-2.94,2.94) <-1.75,0.25>
Total: 248988 W: 63217 L: 63230 D: 122541
Ptnml(0-2): 142, 27325, 69584, 27290, 153
https://tests.stockfishchess.org/tests/view/675a7ad886d5ee47d9542341

bench 1032143

[github-actions]

clang-format 18 needs to be run on this PR.
If you do not have clang-format installed, the maintainer will run it when merging.
For the exact version please see https://packages.ubuntu.com/noble/clang-format-18.

(execution 12331921922 / attempt 1)

[peregrineshahin]

According to this commit, SF prevented fully returning mates from TTs at PvNodes making a hash collision impossible to make us announce a wrong mate as we did in some case
7e890fd
So this re-allows that and leaks it from another place, so a guard would be needed. So I'm not sure that this would be simplifying anything if it confuses the matters and require a fix.

[xu-shawn]

hmm, @Disservin any opinions on this?

[peregrineshahin]

It's probably worth noting that we don't know the actual root bounds with such cutoffs, so if any value returned at pv-nodes it might propagate to root as an exact score, and that's because we update alpha and keep switching sides with the negamax, so a fail-high or a fail-low on a pv-node can easily resolve to an exact root score (e.g. a worst case scenario of a mate in 2 where there isn't).

[xu-shawn]

It's probably worth noting that we don't know the actual root bounds with such cutoffs, so if any value returned at pv-nodes it might propagate to root as an exact score, and that's because we update alpha and keep switching sides with the negamax, so a fail-high or a fail-low on a pv-node can easily resolve to an exact root score (e.g. a worst case scenario of a mate in 2 where there isn't).

How would that happen? I thought a fail high can only propagate back as a fail low score in the parent node?

[peregrineshahin]

How would that happen? I thought a fail high can only propagate back as a fail low score in the parent node?

Good question. while it makes sense for fail-lows and one can imagine it easily it doesn't theoretically make much sense for my concern to happen on fail-highs since bestValue is tracked and we make sure to align that with alpha.
but because of the upcoming_repetition code as we don't necessarily update bestValue to alpha, we can have a double upcoming_repetition setting the window to (-1, 0) without bestValue getting changed suddenly leading to all sort of weird things, so probably what you noted is true for fail-highs if bestValue is always aligned.

[xu-shawn]

OK, I can see how that would lead to bad exact scores in some cases. Do you think some of these guards can be removed if bestValue is also updated in upcoming 3-fold detection?