fix: make the PYAUTO_SMALL_DATASETS fast-mode cap even (15×15 → 16×16)

[Jammy2211]

Problem

The fast-mode cap (PYAUTO_SMALL_DATASETS=1) shrank masks/grids/arrays to an odd 15×15. Any feature requiring even mask dimensions then failed under fast-mode — notably PyAutoLens' potential-correction dpsi mesh (RegularDpsiMesh(factor=2) → PairRegularDpsiMesh), which enforces:

raise ValueError("both mask dimensions must be divisible by dpsi_factor")

This failed the interferometer/features/potential_correction/{start_here,likelihood_function}.py scripts in PyAutoHeart's release-fidelity validation (and the nightly it feeds), even though they run fine at full resolution (their real-space mask is 72×72, divisible by 2). Surfaced while auditing the nightly reds; it is not related to the autoconf→autonerves rename.

Root cause

(72,72) real-space mask → capped to (15,15) → 15 % 2 != 0 → mesh rejects it. cleaned_mask_from preserves shape (it only toggles pixels to masked), so the odd shape reaches the divisibility check unchanged.

Fix

Cap to an even 16×16 across all three cap sites and the shared constant:

• Mask2D.circular (mask/mask_2d.py)
• Grid2D.uniform (structures/grids/uniform_2d.py)
• cap_array_2d_for_small_datasets + SMALL_DATASETS_SHAPE_NATIVE (util/dataset_util.py)

Fast-mode is a CI-speed convenience only (16×16 vs 15×15 is negligible) and real runs are unaffected. The potential-correction mesh keeps its strict even-divisibility invariant — the fast-mode cap simply no longer violates it.

Validation

python -m pytest test_autoarray/ → all small-datasets tests pass; full suite green apart from 3 pre-existing nufft_pynufft transformer tests that also fail on clean main in this env (unpinned pynufft), unrelated to this change. Cap-value assertions updated to 16×16; test_dataset_util.py now derives the crop from the constant rather than hardcoding 15.

Downstream

Workspace smoke suites run with PYAUTO_SMALL_DATASETS=1, so they'll use 16×16 after this merges — the change that makes the potential-correction scripts pass. No API change.

🤖 Generated with Claude Code

https://claude.ai/code/session_013ciVftxvYpefh59wSkR7jN

---

Generated by Claude Code