fix: repair ConstantZeroth dead code (eye(P) shape bug + missing neighbors_sizes)

[Jammy2211]

Closes #448. Phase 8C of the inference-methods programme (autolens_profiling#134, results/notes/inference/PROGRAMME.md on its PR #135).

What was broken

al.reg.ConstantZeroth was dead code presenting as a feature — two independent defects, both reproduced on clean main (7d1906e) before any fix (tracebacks pasted on #448):

1. eye(P) shape bug — constant_zeroth_regularization_matrix_from built the zeroth-order term as xp.eye(P) where P = neighbors.shape[1] is the neighbor-column count (e.g. 4), not S the mesh pixel count, so const + zeroth raised ValueError: operands could not be broadcast together with shapes (9,9) (4,4) whenever S != P.
2. Missing neighbors_sizes at the class API — ConstantZeroth.regularization_matrix_from omitted the required argument, raising TypeError before the shape bug was even reached.

The fix

• Zeroth term is now the full S x S scaled identity (xp.eye(S) * coefficient_zeroth**2), matching zeroth.py's semantics (eye(pixels)).
• The class API threads neighbors_sizes=linear_obj.neighbors.sizes, mirroring Constant.regularization_matrix_from.
• Sibling check: adapt_split_zeroth.py and brightness_zeroth.py do not carry the copy-paste pattern — their zeroth terms are xp.diag over per-pixel weights; no change made there.
• Untouched by design: Constant/Adapt/AdaptSplit numerics, autoarray/__init__.py (version-stamp branch owns it). Nothing else bundled.

Null-mode-lift verification (spectrum before/after)

On the 9-pixel rectangular MockMapper fixture with lambda_neighbor = lambda_zeroth = 1.0, through the class API:

Scheme	min eigenvalue	max eigenvalue	condition number
Constant	1.000e-08 (the jitter floor — graph-Laplacian null mode)	6.000e+00	6.0e+08
ConstantZeroth (fixed)	1.000e+00 = lambda_z^2	7.000e+00	7.0e+00

The zeroth-order term lifts the null mode exactly as intended: max |(CZ − C) − lambda_z^2·I| = 0.0 element-wise, and the smallest eigenvalue rises from the 1e-8 conditioning floor to lambda_z^2. This fix makes the scheme WORK — it does not recommend it over AdaptSplit (ConstantZeroth is an alternative regularization model; target-changing relative to AdaptSplit).

Tests

New test_autoarray/inversion/regularizations/test_constant_zeroth.py (numpy-only per repo policy):

• matrix shape (S, S) on a fixture with S != P (the shape that previously raised),
• no null mode: min(eigvalsh) >= lambda_z^2,
• reduction to Constant's matrix + lambda_z^2 * I where the schemes overlap,
• class-API path returns without raising (diagonal = Constant's 8.0000001 + lambda_z^2 = 9).

Full suite: pytest test_autoarray/ → 988 passed, 59 skipped, 3 failed — the 3 failures are test_transformer.py::test__nufft_pynufft__* (scipy.linalg.pinv2 removed in modern scipy, inside pynufft), identical on clean main in this environment and unrelated to this change.

Merge stays with @Jammy2211.

---

Generated by Claude Code