Add star candidate and PSF-fit visualization

[dvida]

Summary

• add a --show-candidates option to RMS.ExtractStars
• display raw local-maximum candidates before PSF fitting, including over-limit sets
• after fitting, display raw candidates as red circles and accepted PSF fits as green markers
• auto-level the avepixel background from the 1st to 99.99th percentile with a flat-image fallback
• show candidate/fitted counts and a legend in the comparison plot
• process FF files sequentially while interactive candidate windows are enabled
• keep default parallel extraction unchanged
• infer integer image bit depth safely and default floating display images to 8-bit unless specified

Tests

• pytest -q Tests/TestExtractStars.py (8 passed)
• real-data run on FF_US002B_20260427_084517_340_0752128.fits (256 candidates, 253 fitted)
• python -m compileall -q RMS/ExtractStars.py Tests/TestExtractStars.py
• python -m RMS.ExtractStars --help
• git diff --check

[Cybis320]

Just a heads up that this PR collides with the yet-to-be-merged #918 (candidate subsampling), and depending on the motivation behind it, might loose some of its utility once/if #918 is merged.

The conflict

I test-merged this into improve-star-detection. One conflict in RMS/ExtractStars.py, the overflow block: this PR reorders it to plot before return False, while #918 deletes the return False outright and replaces it with the stratified top-hat subsampler.

Two of the five new tests assert the removed behavior and will fail: testDefaultOverLimitRejectionDoesNotPlot and testOverLimitCandidatesAreShownThenRejected.

In #918 max_stars stops being a threshold

Both call sites pass max_star_candidates=config.max_stars. #918 doesn't change the number, it changes what it is:

• Today: a panic threshold. Cross it, the whole FF yields zero stars. The only sane tuning is "high enough to never trip it" — hence 800 in the template.
• After Subsample instead of dropping frames with too many star candidates; improve star extraction robustness #918: a sampling budget. Cross it, you keep the best N spatially stratified and the PSF fit vets them. It binds routinely and benignly, so the right value is lower — 400 on my stations.

So "over-limit" stops being an error condition and becomes the normal operating regime on any noisy camera. The PR body's "including over-limit sets" is the half that goes away.

The part I'd keep regardless is the plotStars rework — ~30 lines, correct bit-depth handling, fig/ax instead of pyplot global state, image-or-FF input. Strictly better than the commented-out call it replaces, and reusable by the overlay above. One bug in it: bit_depth = 8*img.dtype.itemsize infers 32 for a float32 image → max_level = 2**32 - 1 → black plot. extractStars always passes it explicitly, so only future callers are exposed.

Minor: this adds Tests/TestExtractStars.py, I have Tests/test_ExtractStars.py coming on the adaptive-gate branch. Two files for one module differing only in case. Happy to fold mine into yours or rename either way.

[dvida]

Thanks for the detailed heads-up and for catching the float32 inference issue. I fixed that in 3a72121: integer images now infer their width with np.iinfo, while floating-point images default to the historical 8-bit display scale unless the caller supplies the source bit depth. I also added a float32 regression test.

For the possible future integration with #918, I suggest keeping the candidate overlay before subsampling so it continues to show the complete raw local-maximum set. We could then optionally distinguish the spatially retained sample with a second colour and show raw/selected counts in the title. The over-limit tests would become sampling-budget tests: the plot should receive all raw candidates, while PSF fitting should receive at most max_star_candidates. The ExtractStars tests can also be consolidated into one canonically named file when that branch is rebased.

For this PR I am leaving the current prerelease overflow semantics intact, since #918 is not merged.