Add more FITS comparison support for regression tests

[stscijgbot-jp]

Issue JP-3789 was created on JIRA by Melanie Clarke:

The jwst regression tests primarily use astropy FITSDiff for comparing test products to truth files. With FITSDiff, it isn't possible to set a different tolerance for header value comparisons and data comparisons, and sometimes the data need a broader tolerance than the header does. It is possible to call FITSDiff multiple times, once for the primary header extension with a strict tolerance and once for the data extensions with a looser tolerance, but it's inconvenient, and it would generate two separate reports that are not easy to report in a single test failure.

It would be useful to have a convenience function to wrap FITSDiff calls, to be used in regression tests, that allows a different tolerance by extension and aggregates the reports appropriately. We could also consider adding some more readable summary reports on the differences found per extension: the default list of pixel differences, starting at the edges in, can be hard to parse.

[stscijgbot-jp]

Comment by Ned Molter on JIRA:

Adding another thing to the wishlist for FITSDiff calls: at present, when a new header keyword is added or removed, the FITSDiff thinks there's a problem in every header keyword below the one that was added or removed; see e.g. https://github.com/spacetelescope/RegressionTests/actions/runs/11803937276/attempts/1  This leads to very long outputs from what is really a small change.  Also, the line that says "extra keyword in a: KEYWORD" is confusing because it just takes the last keyword in the list even if in reality the new keyword was added somewhere in the middle.

It would be nice to find a way around this, e.g., by ignoring order / position in the list when checking for extra keywords.

[stscijgbot-jp]

Comment by Melanie Clarke on JIRA:

Adding another item to consider, stemming from an update to pathloss files from the NIRSpec team in CRDS context jwst_1300.pmap.

Regression tests for data that picked up the new pathloss file (a point source in S200A2 in regression test data jw01309022001_04102_00004_nrs2_cal.fits), showed differences in the pathloss extension in the cal file, but not in the SCI/ERR/VAR arrays. This turns out to be because of the default tolerance settings for fitsdiff: with rtol=1e-5, atol=1e-7, the differences don't show up because the data is on the order of 1e-12. With rtol=1e-5 and atol at default, the differences do show up. This means that essentially no differences will show up in regression tests for NIRSpec science data post flat-fielding, with default tolerances.

We may need to revise the tolerances for all comparisons on NIRSpec data post flat-fielding, or reconsider whether the defaults are doing what we expect them to.