I said it's possible (not necessary!) that dependent packages need to remove pex_logging first. From my experience, the dependency blocking sometimes happened in c++ codes; mostly it was afw and somehow I couldn't remove downstream pex_logging dependency in c++ while afw still used pex_logging. I suspected it was some SWIG thing but didn't follow through to know exactly why. If you are really curious I may go back to find an example.
On second thought, it's indeed puzzling how meas_modelfit could block ci_hsc. There seems no log passing, and after all, ci_hsc is pure python, right? I didn't try anything practical here. Maybe Perry Gee can say more what he saw?