I did some more thinking about my variation this morning, and I think I'm already doing really well on variation. I'd say I'm doing pretty good if I get each reading within +/- 0.2 arc seconds. But because the height measurements those readings translate to are cumulative, that means I could be anywhere between +/- 0.000068" after 70" of measurements. If I then reverse and measure back towards zero, that would potentially double the variation to 0.00014". So the fact that I'm already getting less variation than that is really pretty good. This also helps explain (I think) why on my last set of measurements where I went from near to far twice instead of down and back I saw more variation towards the far end of the straight edge.
I think the conclusion from this is that improving the resolution of setting the autocollimator crosshairs to the target could ultimately be necessary to get to my goal, and that may or may not be possible. Reducing the variation at each position towards this level is still my primary goal, but I'm already there part of the time.