Affected Repositories

• PyAutoArray (primary)

Branch Survey

Suggested branch: feature/sparse-operator-oversampling-cache

Detailed changes

autoarray/dataset/imaging/dataset.py

• apply_over_sampling: pass sparse_operator=self.sparse_operator to the new Imaging.
  The existing __init__ guard still rejects an operator with
  convolve_over_sample_size_pixelization > 1.
• apply_sparse_operator_cpu:
  • add the same convolve_over_sample_size > 1 check as apply_sparse_operator
    (reuse its error text);
  • pass convolve_over_sample_size_lp/pixelization through.
• apply_sparse_operator: pass the convolve sizes through (currently defaulted to 1).
• apply_mask, apply_noise_scaling: if self.sparse_operator is not None, emit
  logger.warning(...) saying the operator was discarded because the mask or noise changed,
  and must be re-applied last. Do not carry it: it is stale.
• apply_noise_scaling: also pass convolve_over_sample_size_* through.

autoarray/inversion/inversion/imaging/abstract.py

• operated_mapping_matrix_list: @property → @cached_property (already imported, l.4).

autoarray/inversion/inversion/interferometer/abstract.py

• Same change for its operated_mapping_matrix_list, after confirming it is reached more than
  once per likelihood evaluation.
• If the property relies on anything mutable after construction, leave it and note why.

Tests (test_autoarray/dataset/imaging/test_dataset.py,
test_autoarray/inversion/inversion/imaging/test_imaging.py, interferometer analogue)

1. apply_sparse_operator_cpu().apply_over_sampling(...) keeps a non-None sparse_operator.
   The inversion built from it is the sparse Numba class, and its log likelihood matches the
   operator-applied-last order.
2. apply_mask / apply_noise_scaling on an operator-bearing dataset log the warning (caplog).
3. apply_sparse_operator_cpu on a dataset with convolve_over_sample_size_pixelization=2
   raises the same exception as the JAX variant.
4. One dense log_likelihood / inversion.reconstruction + mapped_reconstructed_data pass
   calls psf.convolved_mapping_matrix_from exactly once per convolved linear object
   (monkeypatch counter). Same for the interferometer transformer.

• Each witness is run red on unfixed source first (the memory rule: a witness can already pass
  on main).

Workflow

start_dev → /create_issue on PyAutoArray (title: "fix: apply_over_sampling drops sparse
operator; cache operated_mapping_matrix_list") → register library-dev → /start_library
worktree ~/Code/PyAutoLabs-wt/sparse-operator-oversampling-cache/, branch
feature/sparse-operator-oversampling-cache → implementation delegated to an Opus subagent
(progress file + Monitor) → /ship_library (PR only; merge stays human).

• Heart is STALE (no report / install / release validation), which is informational at start.

Verification

• Full pytest test_autoarray in the worktree. JAX tests stay green: the cache is per-instance,
  and an inversion is built per evaluation or trace.
• Re-run the DR1 A/B harness (scratchpad sparse_ab.py) against the worktree:
  • mode A with the old script order must now select InversionImagingSparseNumba;
  • dense-route per-eval time should drop by ~40% (987 → ~600 ms) with logL bit-identical.
• Workspace impact: grep workspaces and *_workspace_test for apply_sparse_operator /
  apply_over_sampling ordering and for assertions on the inversion class. No change expected,
  because correct-order scripts are unaffected.

Key Files

• autoarray/dataset/imaging/dataset.py — apply_over_sampling, apply_sparse_operator(_cpu), apply_mask, apply_noise_scaling
• autoarray/inversion/inversion/imaging/abstract.py — operated_mapping_matrix_list
• autoarray/inversion/inversion/interferometer/abstract.py — interferometer analogue

apply_over_sampling drops the sparse operator; dense route convolves the mapping matrix twice

Type: bug
Target: @PyAutoArray
Autonomy: human-required

Request (verbatim)

"file both and then do them now" — the two PyAutoArray fixes found during the Euclid DR1 campaign (2026-09-27).

1. Imaging.apply_over_sampling silently drops sparse_operator

autoarray/dataset/imaging/dataset.py apply_over_sampling (~l.539-577) rebuilds Imaging without
sparse_operator=. Calling apply_sparse_operator_cpu() before apply_over_sampling() therefore
falls back to dense InversionImagingMapping with no warning. euclid_dr1 vis_pix ran dense the whole
campaign: 987 -> 180 ms and 799 -> 152 ms per eval once fixed (science-script reorder, euclid_dr1 32fbf62),
logL identical (<=2.7e-10 nats), identifier unchanged. The operator depends only on noise_map, PSF kernel and
mask, so carrying it through over-sampling is valid. Adjacent: apply_sparse_operator_cpu lacks the
convolve_over_sample_size > 1 guard apply_sparse_operator has, and neither sparse method nor
apply_noise_scaling carries convolve_over_sample_size_*; apply_mask / apply_noise_scaling change
inputs the operator depends on, so they must not carry it but should not drop it silently either.

2. operated_mapping_matrix_list is an uncached @property

autoarray/inversion/inversion/imaging/abstract.py:119. On the dense route it is evaluated twice per
likelihood call (curvature/data-vector path and mapped_reconstructed_operated_data_dict), each a full PSF
fftconvolve of the mapping matrix: ~425 ms each, ~800 of 1005 ms per eval on a DR1 tile. Caching per
inversion instance saves ~40% on the dense route. Check the interferometer analogue for the same pattern.

Witness

• A dataset with apply_sparse_operator_cpu() then apply_over_sampling(...) keeps a non-None
  sparse_operator and the inversion factory selects the sparse Numba class (red on current main).
• One likelihood evaluation on the dense route calls psf.convolved_mapping_matrix_from once per linear
  object, not twice (red on current main).

Evidence: session scratchpad lever_probe.log, sparse_ab.py, ab_*.json.