I took your advice and compared the lamp images from the evening where I had calibration problems. There was a shift of minus 3 A between the first and second lamp image. The meridian flip was between the first and second images. Thereafter the lamp image shifted plus 0.4 A and stayed there.
I thought I had checked everything was tightened, but today I’ve discovered the two Allen head bolts holding the slit in place had worked loose. No idea how!
I don’t think this explains everything though. The star spectra have a different amount of shift at different wavelengths, while the lamp spectra have the same shift all the way along. I can only think that the different light paths of the lamp and the stars are at the root of the problem as you have been saying.