[Jammy2211]

Closes #643. Side finding from #639 / #641.

Summary

The imaging sparse-inversion path read Mapper.sparse_triplets_data / sparse_triplets_curvature built without the slot-size and negative-index zeroing the interferometer path applied locally, and its two scatter-add helpers imported JAX inside the function on every call.

• mapper_util.sparse_triplets_from gains pix_sizes_for_sub=None and masks invalid entries (slot ≥ pix_sizes, or cols < 0 → weight 0, column 0) on both backends, after slim_weights. Mapper.sparse_triplets_data / sparse_triplets_curvature pass it. The interferometer _sparse_triplets_slim_from now uses the shared masking (its local block removed); its triplets, data_vector, curvature_matrix and reconstruction are identical before and after (24 arrays, Delaunay + rectangular, with and without primary beam).
• data_vector_via_psf_weighted_data_from and mapped_reconstructed_operated_data_via_sparse_operator_from take xp; numpy uses np.bincount(minlength=) with no JAX import, JAX keeps segment_sum. The four imaging call sites pass xp=self._xp.

Is the defect latent? Yes. On the 7×7 fixtures (Delaunay meshes also given a grid reaching outside the hull): Rectangular (all variants), KNN and KNNBarycentric emit no unused slots and no negative indices; Delaunay (56 unused slots) and DelaunayNN (1082) hold index −1 with weight exactly 0 in every unused slot, never a negative index in a used slot nor a non-zero weight in an unused one. Those entries wrap harmlessly under segment_sum today, but the new numpy bincount branch raises on −1, so the masking is a prerequisite, not only a guard.

Narrower than planned: ImagingSparseOperator.from_noise_map_and_psf and the imaging curvature methods still import JAX unconditionally, so InversionImagingSparse as a whole cannot run without JAX even on numpy; fixing that is beyond this PR (follow-up). The interferometer _sparse_triplets_curvature_from is left as is (its util drops out-of-range columns itself).

API Changes

Additive/internal: mapper_util.sparse_triplets_from(..., pix_sizes_for_sub=None); the two imaging util helpers gain xp=np. No public behaviour changes.
See full details below.

Test Plan

• test_autoarray/inversion/pixelization/mappers/test_mapper_util.py: hand-built index table with a −1 index and an unused slot with non-zero weight → zeroed on numpy and JAX.
• test_autoarray/inversion/inversion/imaging/test_imaging.py: imaging sparse parity (data vector, curvature, mapped reconstructed image) vs the dense mapping inversion on such a mapper, numpy backend and JAX backend; subprocess no-JAX test (jax/jaxlib/nufftax blocked) covering the triplets and both helpers against the dense mapping matrix.
• Controls: masking removed → parity test fails; pre-fix util → no-JAX test fails.
• Worktree runs: pytest test_autoarray/inversion -q → 1037 passed, 4 xfailed; pytest test_autoarray/dataset/imaging -q → 36 passed; no-JAX emulation 23 passed, 10 skipped, no jax module loaded.

Heart YELLOW acknowledged by the human on 2026-10-21 (manifest drift 1 mismatch; release validation stale: source moved since rehearsal (PyAutoArray)).

Full API Changes (for automation & release notes)

Changed Signature

• autoarray.inversion.mappers.mapper_util.sparse_triplets_from(..., pix_sizes_for_sub=None) — new optional kwarg; when given, invalid stencil entries are zeroed.
• inversion_imaging_util.data_vector_via_psf_weighted_data_from(..., xp=np), mapped_reconstructed_operated_data_via_sparse_operator_from(..., xp=np) — backend dispatch; numpy branch no longer imports JAX.

Generated by the PyAutoLabs agent workflow.

🤖 Generated with Claude Code