3D line of sight ended up being not-trivial computationally, so we have had to make some allowances for what’s feasible to compute and have deviated from the rules as written in the CRB. The difficulty is that for both the AI and user interface, I need to know “can this 3D voxel see this 3D voxel” for the entire 3D map ahead of time instead of doing an ad-hoc raycast from single point to point like Baldur’s Gate 3 does (which can actually be exploited by clicking a target really fast before their single calculation completes).
I’ve tried to make our faster but less-realistic method reasonable to the eye, but there’s definitely still some rough edges. This one seems fine-ish to me because the snipers on the very edge of vision here, but I agree it’s borderline.