TL;DR: better bits made a big difference.
Okay, I finally got some time to do a couple more experiments.
Answer #1 (not an experiment): I have the HD conduit (Wheatstone or something is the brand). I haven’t yet measured the flex; I’ll put that on the list.
Answer #2: My drillman mills came, and they are way sharper than the bits I got first (amazon link above). The chinese amazon bits were burning the wood right away [although I note I still haven’t done anything about the high RPM); these bits cut much more easily. (Noticeably sharper; I sliced a slit through my finger without realizing it.)
The calibration test: with z=-20 (the surface of the wood 20mm below the highest point I can lift the tool without dropping the nut off the rod), I got rectangles that were 10.2 and 9.6mm wide. Not bad! Worst case was about 0.2mm error per path. Not too shabby. With z=-80 (the work surface was all the way down on the bed), I got 10.1mm and 9.45mm. Not sure why the wider direction actually got better (other than just error; this in an N=1 experiment). The narrower direction is now off by about 0.3mm per path. Not a huge difference with the change in Z depth.
In these cuts, the working wood was all clamped down pretty well. I’m not far off your reported 0.2mm. Should I be able to reach higher accuracies?
I still have more things to play with, such as pulling off the extra camar0 contraption on the middle z, measuring the conduit flex on the moving rails, and trying to quantify the belt stretch. (There definitely is some, but I’m not sure how big a factor it is. The pull test above on the rollers should bound it.) I’ll report back.