Affected Repositories

• PyAutoArray (primary)

Branch Survey

Suggested branch: feature/imaging-sparse-triplets-masking

Implementation Steps

1. autoarray/inversion/mappers/mapper_util.py sparse_triplets_from (~270): add pix_sizes_for_sub=None; when given, build slot = tile(arange(P), M_sub), valid = (slot < repeat(pix_sizes, P)) & (cols >= 0), vals = where(valid, vals, 0), cols = where(valid, cols, 0) on both numpy (~328) and JAX (~355) branches.
2. autoarray/inversion/mappers/abstract.py sparse_triplets_data / sparse_triplets_curvature (~440-470): pass pix_sizes_for_sub=self.pix_sizes_for_sub_slim_index.
3. autoarray/inversion/inversion/interferometer/sparse.py _sparse_triplets_slim_from (~131-172): pass pix_sizes_for_sub, remove the local masking block; test_interferometer.py parity tests stay green.
4. autoarray/inversion/inversion/imaging/inversion_imaging_util.py data_vector_via_psf_weighted_data_from (~70-104) and mapped_reconstructed_operated_data_via_sparse_operator_from (~280-303): add xp=np; numpy branch uses np.bincount(..., minlength=), JAX branch keeps segment_sum; callers in imaging/sparse.py (~90, ~144, ~168, ~526) pass xp=self._xp.
5. Tests in test_autoarray/inversion/mappers/ (mapper_util) and test_autoarray/inversion/inversion/imaging/test_imaging.py; emulate the nojax leg (import-hook block) locally before ship.

Key Files

• autoarray/inversion/mappers/mapper_util.py — shared triplet builder
• autoarray/inversion/mappers/abstract.py — Mapper triplet properties
• autoarray/inversion/inversion/imaging/inversion_imaging_util.py — scatter-add helpers
• autoarray/inversion/inversion/imaging/sparse.py — consumers
• autoarray/inversion/inversion/interferometer/sparse.py — local masking to retire

Imaging sparse inversion: unmasked mapper triplets in the reconstructed-image path and an unconditional JAX import

Type: bug
Target: PyAutoArray
Repos:

• PyAutoArray
  Difficulty: small
  Autonomy: safe
  Priority: medium
  Status: formalised
  Consequence: glance
  Witness: a test asserting the imaging sparse mapped_reconstructed_data_dict equals mapping_matrix @ reconstruction for a mapper whose pix_indexes_for_sub_slim_index contains unused stencil slots (slot >= pix_sizes) and a negative index, on numpy without JAX importable (nojax leg) and under JAX
  Review-minutes: 3
  Unattended: ready
  Filed: 2026-10-21

Why

Side finding from PyAutoArray#639 (merged as #641, 2026-10-21). The interferometer sparse path zeroes invalid triplet slots in _sparse_triplets_slim_from (slot >= pix_sizes_for_sub_slim_index, or cols < 0 → weight 0, column 0). The imaging sparse path does not.

Where

• autoarray/inversion/inversion/imaging/sparse.py around line 526, _mapped_reconstructed_data_dict_from, uses linear_obj.sparse_triplets_curvature, built in inversion/mappers/abstract.py:460-470 via mapper_util.sparse_triplets_from(..., return_rows_slim=False) with no masking. A -1 source index silently wraps to the last source pixel and unused slots contribute if their weights are non-zero.
• The same helper, inversion_imaging_util.mapped_reconstructed_operated_data_via_sparse_operator_from (~line 280), imports JAX unconditionally even on the numpy path (nojax leg risk).
• The imaging data-vector and curvature paths that consume the same unmasked Mapper.sparse_triplets_* likely share the gap; audit them in the same PR.

Fix

Apply the slot/negative-index masking once at the Mapper.sparse_triplets_* builders (or in sparse_triplets_from behind a flag) so every consumer is covered; make the JAX import lazy (xp-dispatched) in the imaging util. First establish whether any current mesh (Rectangular, Delaunay, DelaunayNN, Voronoi) actually emits negative indices or non-zero weights in unused slots; if none does, the fix is a guard plus the regression test, and the prompt notes the latent nature.