Hi Engin,
This has been my experience with Na sites as well. Phenix.refine water-update would always move/delete my waters out of coordination spheres (so I would turn off solvent updates, and strictly restrain these waters). I always assumed that since Mg and Na coordinate water with very short distances (2-2.4 Angstroms) compared to water to N and O atoms, they were disliked by solvent updates. Since I have not looked at the code (and haven't used the recent versions for Mg/Na containing structures), I cannot confirm this, but does solvent-update assume waters should be within 2.3-3.5 A? If that is the case, that may kill some true water sites.
I have special handling of "water - other_isolated_non_water_atom" situations that in theory should allow shorter distances, but apparently this fails sometimes (not always though). I should review that code given that I have more practical examples now. Pavel.