OK, I have now manually calibrated using:-
The standard set of lines used by ISIS for the automatic fit (lamp lines only)
The same set with the 3946 line removed
The set of lines used for Hugh’s calibration in post #36, less the H alpha line which I do not see in my spectrum
The calibrated star spectra were identical every case with errors the same as those in my post #37, within the fit uncertainty of +-0.5A, even down to 3750A
Lamp spectra calibrated using these fits showed no significant wavelength errors so the problem, in my case at least, appears to be caused by some systematic difference in the way the spectrograph sees the internal calibration lamp compared with the star. The error is wavelength dependent rather than a simple offset so is going to be difficult to correct for.
I think my next step will be to check what my “filly dot” lamp, which is mounted over the telescope aperture, gives.