I have two questions:
Assuming I know that my ligand binds by coordinating with a divalent cation, could I use this knowledge to remove inaccurate docked models that do not coordinate as expected?
Does HADDOCK take into account the most probable coordination number of a known metal ion? Some of my docked models had no good coordination.
as far as I know HADDOCK make a covalent bond between ion and the first residue coordinating atom in your protein. This is necessary to avoid that during it1 your ion will fly away crashing your docking calculations. Said so, in my opinion your assumption is correct but I guess most of the solution will be wrong. You could force the coordination using unambiguous restraints in order to improve the solutions.
There is indeed no directionality in ion interactions that would ensure that the coordination geometry is conserved. So if you want to ensure it, define additional distance restraints between both the coordinating atoms and your ion, but also between the coordinating atoms.