Wavelength calibration can be bad, both relatively between fibers and absolutely to standards: an RMS of 0.2 nm (a couple of pixels) is common. Such bad wavelength calibration will severely compromise sky subtraction, so we need to improve this before people get serious about playing with sky subtraction algorithms (either 1D or 2D).
Some ideas:

1. Fix the representation of wavelength in the DetectorMap. Currently, wavelength is represented as a spline, but the spline is derived from a wavelength vector by simply taking every Nth point as a knot — no fit is done, so the fidelity of the solution (e.g. derived from a polynomial fit) is reduced.

2. Line identification needs to be robust against inaccuracies in the DetectorMap wavelength solution. While it's true the DetectorMap wavelength solution should be solid during observing, during development and commissioning and while dealing with both simulated data and LAM data, it's not unknown for the true wavelength solution to change relative to what's in the DetectorMap. Here's a possible algorithm:

• find and centroid all peaks in spectrum
• apply rough wavelength calibration to measured peaks
• for both observed and model line lists:
  • find pairs of lines with separations between 5A and 500A (separation is not too small to allow confusion, not too large to be very non-linear)
  • for each pair, calculate mean and difference
• match observed pairs to model pairs when the mean is within 50A and the difference is within ~ 0.5A
• Instead of pairs, could use triplets, and match the ratio of spacings.

3. Identify and flag arc lines that are regularly discrepant so they won't be used in the fitting.