The level's around 1e-5 in arbitrary pfsMerged units (where the brightest sky lines are c. 10);

I think that this is due to the use of the sample, not population, variance to weight the sky spectra. For high flux levels (> 100 counts) this gives a bias of 1, but things are more complicated for low fluxes and when there's also read noise present. Investigate and fix.

We could just offset the sky spectrum to give mean zero residuals, but that's a hack.