I replaced some of the AD829s with other AD829s, but the offset situation didn't improve.
However, I figured that we don't really need the ~100MHZ bandwidth of the AD829, since the IMC loop limits us to a ~10kHz CARM bandwidth. Also, since we don't routinely use IN2 for anything, I felt free to try something else.
Specifically, I replaced all of the positive gain AD829s in the input 2 gain ladder with OP27s (U8B->U12B on D1500308), which should have input offset voltages ~30x lower than the AD829s.
Here is a comparison of the outputs these configurations perform, normalized to the output at the +0dB gain setting - where all of the op amps in the gain ladder are bypassed.
So, most of the transitions now result in an output offset change of less than 0.5mV, which is nice.
The exception seems to be where the +8dB stage is switched in or out. I may try replacing this one, as these transitions cause a lock loss now when trying to lock the arm with high bandwidth using POY.