Am I missing something or the only way to get this done is to refine
twice - once only for ligand occupancy, and then for the rest of the
model.

And what if total occupancy of the ligand is unknown? Is there a way to
allow sum(occ) to be anywhere between 0 and 1?

I may be wildly off base here, but if you give each conformation a different residue number as well as different altlocs, and set the occupancies below 1, won't this remove the sum constraint? It's a little clumsy, but you just need to change the residue numbers back before you deposit in the PDB.

-Nat