Any feedback on this PR?

this looks reasonable, but would be good to see if things are actually equivalent, it is, for a given UCI_Elo a functional change. Probably a match could be done with master and patch for a given UCI_Elo, to check this doesn't introduce a large change.

Note that it also overlaps with what is discussed in #3635

Current implementation maps ELO values into a (float) level

double floatLevel = Options["UCI_LimitStrength"] ?
                      std::clamp(std::pow((Options["UCI_Elo"] - 1346.6) / 143.4, 1 / 0.806), 0.0, 20.0) :
                        double(Options["Skill Level"]);

And then converts the level into the upper or lower integer according to a statistical formula that takes into account how much the float differs from the lower bounding integer.

int intLevel = int(floatLevel) +
                 ((floatLevel - int(floatLevel)) * 1024 > rng.rand<unsigned>() % 1024  ? 1 : 0);

If you agree, I'd propose the following.

1. Verify smoothness: run tests among new patched engines (not vs master) with different UCI_Elo settings and verify that new formula is consistent, i.e. winning ratio increases smoothly with UCI_Elo

2. Verify mapping: run test of new patch against master but without enabling the float to int statistical tweak, because it would just add noise for the target of this test, and verify that the two engines have comparable strenght. To do this we could test patch vs master setting an UCI_Elo for which floatLevel - int(floatLevel) == 0, so that second formula becomes moot.

I would suggest to run just 3 matches patch against master, both engines with the same UCI_Elo at fixed values (e.g. 1800, 2000, 2200), each for e.g. 20k games, and see that the Elo difference is 'small'. This can be done on fishtest, passing additional options.