With PBC, imaging is performed for all atom pairs before calculating the distance between them. So the target value for a distance restraint (these are RESTRAINTS not CONstraints!) between any pair of solute atoms must not exceed 1/2 the box length. (dap)
Not only that, but the *actual* distance between two atoms must not exceed this limit, and with weak constraints, the actual distance could be greater than the target. (dac)
In SANDER, aren't the restraints computed in DISNRG? Maybe we could just turn off imaging in that subroutine, and have people use RESTRAINTS. I would actually be in favor of throwing away CONstrained bonds, angles, torsions (including the CBOND subroutine) and aksing people to use DAPs restraint code instead, which is more powerful and (in my view, anyway) more intuitive. [I don't know if this sort of change makes sense for Gibbs, but I've never liked putting constraints into the "prmtop" file, so that when they are changed, you have to re-run PARM.]