So you're looking at why the 2-D data isn't significant, but the 1-D result is?

Is this related to the problem that we use the background as an estimate for its variance? (population v. sample)?

That's the first thing I'm investigating. There may be other causes that I will chase down as well.

I don't think this is related to how the variance is measured. At the moment, the question is how a single-pixel value of -0.14 with a variance of 21 gets extracted to a spectrum value of -1.0 with a variance of 0.11.

Ah, I had mismatched the image and spectrum. The actual image value is -10.8 and variance of 12. That's a S/N of about 3, which matches the S/N in the spectrum. So that question is answered.

Now, how is that single 3-sigma low value leaking into everything else?

The low value is in a sky fiber, and so the low value is an ingredient for the 1D sky subtraction model, and the subtraction of a low value makes a high value.

The 1D sky subtraction is using FitFocalPlaneTask.fit, the docstring for which says:

        This implementation is a placeholder, as it simply averages the input
        vectors instead of doing any real fitting or rejection of outliers, and
        no attention is paid to the position of the fibers on the focal plane.

I should add that there's no treatment of errors either.

So I think I now understand what's causing the problem, and the solution is to replace FitFocalPlaneTask with a proper implementation. We'll always be limited by the mere two sky fibers in the integration test, but we should be able to make that single bad pixel go away.

If we need to add more sky fibres, let's do it.  I'd be OK with a 50:50 ratio for a test suite (as opposed to a guessed 20:80 sky:science on the sky)

Debugging is done. Will be fixed on PIPE2D-635.