Over 1 hour 7 minutes, the images offset by a total of 32 pixels vertical, 8 pixels horizontal, but almost all the horizontal drift was on the first two images (presumably taking up the backlash). So the real drift was about 32 pixels or 1 pixel every 2 minutes.
Looking at the subs I thought I could spot a few where there was more noticeable movement between subs - then I checked and these 'jumps' were where I had dropped subs because of aeroplane trails, causing nearer to a 2-pixel jump instead of one. There didn't seem to be any of the jumps I would have expected if there was significant periodic error in the worm wheel.
What was most striking is that every single sub showed nice round stars - as would be expected if the camera had strayed less than 0.5 pixels either side of the mean position.
I won't pretend that these results are good enough for long subs, but they do show that an EQ3 mount properly balanced and aligned with a bit of care on a solid tripod is capable of long enough exposures for imaging DSOs.
It also suggests to me that it will be worth me upgrading to autoguiding before upgrading my mount - which is against conventional wisdom. It also lends support to my suspicion that the weakness in the normal EQ3 setup is the aluminium tripod not the mount.
Something I want to try is taking long, unguided, wide field exposures. With a 28mm lens the tracking errors should be under a pixel even at 10 minutes exposure. This should be also be a way to see if there is significant periodic error.