Thanks Phil, I'm using my own post-processor that I wrote for Fusion 360 so it would be easy to enhance that to optionally swap axes when it generates the G-code. I thought about doing that up front, but I wanted to make my video work for folks that might be using different post-processors without the ability to modify them. It looks like GCode Ripper would be a good option for them.
The Harmonic Drive gearboxes are advertised as "zero backlash" but of course it's not quite zero. The type that I have is sold in two grades, specified to have backlash under 3 arc-min or under 1 arc-min. I can't figure out which mine is from the part number on it, but I presume it's the cheaper 3 arc-min grade which means backlash up to 0.05 degree. I plan to measure the actual backlash of my system, including that of the 3-jaw chuck since its jaws may permit some movement that is backlash-like. Measuring that and the holding torque might make an interesting video.
I was worried that the 45:1 ratio of my gearbox would be too slow but it actually seems pretty snappy at least when rotating with no load. I probably can't rotate it as fast under load (when cutting) without losing steps. I was lucky to find one with this ratio, which seems about optimal for my needs. With a 200-step motor and 1/8 microstepping being roughly the practical limit for positional accuracy of a stepper, the 45:1 ratio gives a resolution of 0.005 degree. Assuming that the physical backlash of the system is roughly 10x that as noted above, the positional accuracy is not significantly limited by stepping resolution and mainly limited by the backlash. I will try to measure it but I'm guessing that with all things combined, and some opposing torque applied, I should get a positional accuracy of about 0.1 degree total or +/- 0.05 degree. That would give me an error up to about +/- 0.001" / 0.02mm on the circumference of a 2" / 50mm diameter part.