Missing input validation across Array2D, Grid2D, Mask2D, Imaging, regularization

[rhayes777]

Title: Missing input validation across Array2D, Grid2D, Mask2D, Imaging, regularization

Repo: Jammy2211/PyAutoArray

Tag: @Jammy2211

---

A handful of inputs that should be rejected at construction are silently
accepted and propagate downstream as confusing NumPy or numba errors. None
of these is performance-critical, but each one converts a clear
ValueError("…") into a head-scratch debugging session for new users.
Found while auditing 2026.5.21.1.

B6. Array2D accepts pixel_scales=0, negative, NaN

import numpy as np, autoarray as aa

aa.Array2D.no_mask(np.ones((4, 4)), pixel_scales=0.0)            # accepted
# first .derive_grid call: ZeroDivisionError: float division by zero

aa.Array2D.no_mask(np.ones((4, 4)), pixel_scales=-0.1)           # accepted; geometry flipped
aa.Array2D.no_mask(np.ones((4, 4)), pixel_scales=float("nan"))   # accepted; NaN propagates

B7. Mask2D.circular_annular(inner > outer) silently returns zero pixels

aa.Mask2D.circular_annular(
    shape_native=(20, 20), pixel_scales=0.1,
    inner_radius=0.8, outer_radius=0.3,
).pixels_in_mask
# 0

Almost always a typo / swapped arguments.

B8. Grid2D.uniform(shape_native=(0, 0)) builds an empty grid

aa.Grid2D.uniform(shape_native=(0, 0), pixel_scales=0.1).shape_slim   # 0
aa.Grid2D.uniform(shape_native=(0, 5), pixel_scales=0.1).shape_slim   # 0

B5 (from autolens). al.Imaging accepts mismatched data / noise_map shapes

import autolens as al
data  = al.Array2D.no_mask(np.ones((10, 10)), pixel_scales=0.1)
noise = al.Array2D.no_mask(np.ones(( 5,  5)) * 0.1, pixel_scales=0.1)
psf   = al.Convolver.from_gaussian(shape_native=(3, 3), sigma=0.1,
                                   pixel_scales=0.1, normalize=True)
ds    = al.Imaging(data=data, noise_map=noise, psf=psf)
ds.shape_native    # (10, 10) — mismatch swallowed

(Filing here rather than in PyAutoLens since Imaging ultimately lives in
autoarray/dataset/imaging/dataset.py.)

B13. Negative regularization coefficient silently accepted

import autolens as al
al.reg.Constant(coefficient=-1.0)   # accepted; produces same log_evidence as +1.0

The coefficient appears to be squared internally, so positive and negative
give identical answers — but the value is unphysical and the user gets no
heads-up.

Suggested fix shape

A small _validate_* helper used in each constructor:

if pixel_scales <= 0 or not np.isfinite(pixel_scales):
    raise ValueError(f"pixel_scales must be a finite positive number; got {pixel_scales!r}")

…and analogous one-liners for the others. Happy to PR if useful.

Environment

• autoarray == 2026.5.21.1, Python 3.12.8.

Full repro at work/edge_case_tests/01_core_data_structures.py and
work/edge_case_tests/09_more_edges.py in the audit workspace.