Just found time to run a full 1mV sweep of Ch3.
And the answer is... yes, Ch3 has better granularity. Eyeballing the data, it seems to have about 0.22mV granularity, which is a bit more than twice as small as Ch1. See histogram attached. I should have added axis titles, but the histogram is of the measured voltage difference between two consecutive steps (separated by 1mV, the programming granularity). The peaks show that the distance is usually around 1.045mV, and if not, most likely around 0.83mV (and extremely infrequently around 1.25mV), thus the derivation of the DAC granularity.
Note that actual absolute accuracy and linearity are much worse than this, so I don't think that this has practical applications. On my home-calibrated DP832, Ch3 stays within about +/- 1.5mV of programmed value throughout the whole sweep, which is quite good.