In this post I had reported moving to 300mS mitigated the hunting issues: http://smartdcgenerator.blogspot.com/2013/03/300ms-helped-lot.html
Well, today I tested out a PID-lite implementation of calculating the PWM from VBat, and the results are even better! Going back to 100mS cycle times, the system now has a +/- 30mV regulation range around the target VBat while in Acceptance mode.
So much better, and I get to move back to the 100mS overall sampling time (better to respond to load changes).
So, PID-lite it is!