Pixel scale from the WCS: ngmix prior and PSF star preselection

[cailmdaley]

Closes #134.

The pixel scale now comes from the image astrometry everywhere in the committed pipeline. No 0.186 or 0.187 is left in workflow/config/cfis/ or in the code paths it runs, and ngmix has no PIXEL_SCALE option.

• ngmix. The centroid prior is one pixel wide, and each object uses its own pixel: sqrt|det J| of the per-epoch WCS Jacobians already built for the fit, averaged over the object's epochs. The prior is still built from the object's position-seeded RNG before the fit, so only its width changes. pixel_scale_from_wcs and its scan of the merged headers are removed. That scan raised AttributeError on real headers, and the PIXEL_SCALE = 0.186 pin had been hiding it.
• Star selection. The 0.3–1.5″ preselection bounds were FWHM_IMAGE > 0.3 / 0.187. They are now FWHM_WORLD bounds in degrees, which SExtractor computes from the local WCS at each object. The ±0.2 px mode window is unchanged, since it is already in pixels. Diagnostics report arcsec via FWHM_WORLD*3600, and the stars/deg^2 stat, which was built from 0.187, is removed.
• Record. fit_priors describes the per-object width and pins NGMIX_RUNNER.PIXEL_SCALE = absent. star_selection_box pins the angular bounds. CENTROID_SOURCE, METACAL_PSF and BKG_SUB are also pinned absent, so a template that sets one fails the decision check (seen to fail).

On real data (exposure 2114045p, 40 CCDs):

• The mean scale is 0.1855″/px. Per position, sqrt|det J| runs from 0.1827″ to 0.1871″ (−1.5% to +0.8%, rms 0.5%). Within one CCD it varies by up to 1.3%, as much as between CCDs, so per position is the right granularity.
• In 200 paired 2-epoch sims, changing the prior width by ±0.75% moves the median |Δg| by 7e-6 at S/N 20 and moves ⟨R11⟩ by less than 0.1%. At S/N ≥ 50 the changes are 1e-6 or smaller.
• Star selection on four CCDs, same catalogue before and after: preselection grows by 0–4 objects per CCD. Selected stars are identical on CCDs 0 and 20, 67 → 68 on 39, and 74 → 66 on 10. The CCD-10 change comes from the mode estimator's sensitivity to edge objects, which Centre the PSF-star box on a KDE peak of the FWHM preselection #934 addresses.

Tests.

• test_process_centroid_prior_is_each_objects_own_pixel_scale runs the process loop on objects with rotated and flipped epochs. It checks each prior width, and that the prior draws from the object's untouched position-seeded RNG.
• test_stamp_pixel_scale_reads_a_rotated_flipped_wcs checks the scale from a rotated, flipped TAN WCS.
• test_star_selection_preselects_on_angular_fwhm runs the committed .setools on synthetic stars and galaxies with per-object scales between 0.185 and 0.188.
• Each test was seen to fail under injected bugs.
• tests/unit and tests/module in the dev container: 804 passed, 1 skipped. astra validate passes. A cross-family review (GPT) found no correctness defects in the change.

Claude Opus 5.5 on behalf of Cail

🤖 Generated with Claude Code

[cailmdaley]

can we use the correct per-CCD pixel scale?

[cailmdaley]

Yes, and it now goes finer than per CCD. Each object's centroid-prior width is sqrt|det J| of its own epochs' WCS Jacobians, the same ones the fit already uses, averaged over those epochs. The global mean, the merged-header scan and the PIXEL_SCALE option are gone. On 2114045p the scale varies by up to 1.3% within a CCD, as much as between CCDs, so per position is the right granularity. Star preselection was already per position, through SExtractor's FWHM_WORLD. The effect on shear is negligible: a ±0.75% width change moves g by about 1e-5 at S/N 20, and the RNG stream is untouched.

Claude Opus 5.5 on behalf of Cail