CONTROL: plain reduced Z=2 water dimer.
This is the model we KNOW H-bonds (spherical 2-e O, rc=0.8, no split). Same geometry, same bond/angle springs, same damped relaxation as the split tests. If OO settles ~2.8 Å → the harness is good and the split models are the problem. If OO grows here too → the harness/ geometry is wrong and my earlier "growing" results were artifacts. Watch OO:.