fix: relabel at-or-below-cap data at the capped pixel scale - #431

Summary

cap_array_2d_for_small_datasets handled one case and silently dropped the other. Data larger than the 16x16 PYAUTO_SMALL_DATASETS cap was cropped and rebuilt at SMALL_DATASETS_PIXEL_SCALES (0.6); data already at-or-below the cap early-returned, keeping the caller's uncapped pixel_scales (0.1). A capped simulator writes its data at 0.6, so the loader mislabelled the frame 6x — ±0.8" instead of ±4.8" for a 16x16 field.

Off-centre galaxies then fell outside the mislabelled frame, their non-negative linear intensity solve correctly returned exactly 0.0, total_luminosity became 0, and min(5 * 0.5 * 0**0.6, 5.0) collapsed a UniformPrior to lower == upper == 0.0. The PriorException surfaced four steps downstream in autolens_workspace/scripts/group/slam.py:321, naming neither the loader nor the pixel scale.

The at-or-below-cap branch now rebuilds the Array2D at the capped scale, mirroring the crop branch. Shape is preserved — that branch must never crop. Rebuilding is required rather than returning a corrected scalar: the Array2D is constructed by the caller before the call and carries its own geometry.

Fixes #430. Fixes the group/slam.py + group/slam.ipynb failures in PyAutoHeart Workspace Smoke run 30790463134.

API Changes

No signature changes. One behaviour change, confined to runs with PYAUTO_SMALL_DATASETS=1:
autoarray.util.dataset_util.cap_array_2d_for_small_datasets now returns SMALL_DATASETS_PIXEL_SCALES (0.6) and a rebuilt Array2D for input already at-or-below the cap, where it previously returned the caller's pixel_scales and the input array untouched. Shape is unchanged in that branch. Reached in practice through Imaging.from_fits, which routes data and noise_map (not psf) through it.

With the cap unset — every normal run — behaviour is byte-identical, and the crop branch is untouched.

See full details below.

Test Plan

• pytest test_autoarray/ — 929 passed
• scripts/group/slam.py (autolens_workspace) under the capped smoke profile, cleared dataset + output → exit 0, all six SLaM stages genuinely run (0 cached resumes; note output is namespaced under output/test_mode/)
• Crop path unchanged — 209x209 cosmos_web_ring gives (16,16) @ 0.6 identically before and after
• Normal uncapped operation unaffected — scripts/imaging/start_here.py (ENV: full_datasets) exit 0
• Committed datasets audited: of 31 committed dataset FITS, exactly one is at-or-below the cap (slacs1430+4105/psf.fits, 11x11) and PSFs never route through the capper — so no committed dataset changes scale

Downstream effect on two parked scripts

• imaging/features/scaling_relation/slam → now exit 0 (6 real searches). Its NEEDS_FIX park in autolens_workspace/config/build/no_run.yaml can be removed in a separate workspace PR.
• multi_galaxy/features/scaling_relation/slam → gets past the 0.0-luminosity cause (real luminosities now measured) but then hits a separate latent script bug: slam.py:863 computes image_half_width from the script's own hardcoded pixel_scale (0.1) while the mask is built from dataset_full.pixel_scales (now correctly 0.6), producing an empty mask. Not a regression — on unpatched main it fails earlier with the documented 0.0-luminosity error. Stays parked; filed separately.

Generated by the PyAutoLabs agent workflow.