Affected Repositories

• PyAutoArray (primary)

Branch Survey

Suggested branch: feature/adaptive-rect-mesh-node-positions

Implementation Steps

1. autoarray/inversion/mesh/interpolator/rectangular.py: helper adaptive_rectangular_node_centres_from(...) — node unit coordinates U_y = (n_y - r - 1)/(n_y - 3), U_x = (c - 1)/(n_x - 3) in mapper order, through _transforms_from inverse (as adaptive_rectangular_transformed_grid_from); clamp guard nodes to the data span; xp-compatible.
2. autoarray/inversion/mesh/mesh_geometry/rectangular.py: node_centres property (adaptive → helper; uniform → mesh_grid).
3. autoarray/inversion/mappers/abstract.py (or the rectangular interpolator): route source_plane_mesh_grid through mesh_geometry.node_centres for rectangular meshes, returned as Grid2DIrregular.
4. Check consumers: inversion/plot/inversion_plots.py CSV writer, inversion/inversion/abstract.py max_pixel_centre, inversion/mappings/mapping.py (Delaunay/kNN only), preloads (Delaunay mesh grids only).
5. Tests in test_autoarray/inversion/... per the plan; nojax leg: guard JAX-only tests with importorskip.
6. Downstream: PyAutoGalaxy/PyAutoLens test suites; workspace_test mesh_magnification.

Key Files

• autoarray/inversion/mesh/mesh/rectangular_rtu_adapt_density.py (~L273, overlay_grid_from)
• autoarray/inversion/mesh/interpolator/rectangular.py
• autoarray/inversion/mesh/mesh_geometry/rectangular.py
• autoarray/inversion/mappers/abstract.py
• PyAutoLens autolens/analysis/result.py ~L566 (position updating consumer)

Adaptive rectangular meshes report a uniform overlay as their source-plane node positions

Type: bug
Target: autoarray
Repos:

• PyAutoArray
  Epic: cluster-strong-lensing
  Phase: 4e (blocker)
  Arc: draft/feature/autolens/source_cluster_arc.md
  Difficulty: medium
  Autonomy: supervised
  Priority: high
  Consequence: judge
  Filed: 2026-10-30
  Witness: for every adaptive rectangular mesh, mapper.source_plane_mesh_grid equals the source-plane positions at which the mapper's interpolant takes each reconstruction value (node i's basis function peaks there); griddata of the reconstruction on those positions integrates to sum(reconstruction * areas_for_magnification) within a few percent; the reconstruction CSV writes those positions; RectangularUniform and Delaunay-family output is unchanged.

Original request (human, 2026-10-30)

Library fix first (Recommended) — Park this PR. File and fix a PyAutoArray bug so source_plane_mesh_grid (or a
new attribute) gives the adapted node centres, then come back and finish this workspace PR on top of it,
library-first. The whole script then gives right answers.

Finding (autolens_workspace#605 comment, 2026-10-30)

Every adaptive rectangular mesh (RectangularBilinearAdaptDensity, RectangularBilinearAdaptImage,
RectangularRTUAdaptDensity, RectangularRTUAdaptImage) builds its interpolator with
mesh_grid=overlay_grid_from(shape, relocated_grid.over_sampled)
(autoarray/inversion/mesh/mesh/rectangular_rtu_adapt_density.py ~L273): a uniform grid over the traced
data's bounding box. The mapper's interpolant instead places reconstruction value (row r, col c) at unit-CDF
node U_y = (n_y - r - 1)/(n_y - 3), U_x = (c - 1)/(n_x - 3) (guard ring at the border), mapped back to the
source plane through the inverse adaptive CDF (adaptive_rectangular_transformed_grid_from). So
mapper.source_plane_mesh_grid (= interpolator.mesh_grid) is not where the values live.

Measured on autolens_workspace imaging/simple__no_lens_light (true tracer, 28×28 BilinearAdaptDensity,
coefficient 1.0, extra galaxy noise-scaled): reported node spacing is uniformly 0.11″ (data span / 28); the
library's own areas_for_magnification at the brightest nodes is 0.0013 arcsec² against 0.0097 implied by the
reported positions; correlation between the areas and the Delaunay dual of the reported positions is 0.13.
griddata of the reconstruction on the reported positions integrates ~4.6× too high, so the workspace
interpolation-route magnification is 6.9 against a mesh-route 31.6 and a truth of 34.9, at any interpolation
resolution.

Blast radius

• autoarray/inversion/plot/inversion_plots.py ~L510: source_plane_reconstruction_{i}.csv (a documented science
  product) writes y, x from mapper.source_plane_mesh_grid, so its coordinates are wrong for adaptive meshes.
• Every workspace source_science.py / CSV-loading example that interpolates with griddata on those positions
  (interpolated source, zoom, errors, S/N masking, interpolation-route magnification), plus anything downstream
  that reads the CSV.
• autoarray/inversion/inversion/abstract.py ~L1640 max_pixel_centre: the brightest source pixel's position is
  taken from mapper.source_plane_mesh_grid, so it is wrong for adaptive meshes. PyAutoLens
  autolens/analysis/result.py ~L566 uses it for automatic position updating (brightest source pixel →
  PointSolver → image-plane multiple-image positions → SLaM positions likelihood). Quantify how far the reported
  brightest-pixel centre moves on a realistic fit, and say in the PR whether past SLaM runs with adaptive
  rectangular meshes were materially affected.
• Plots are fine: autoarray/plot/inversion.py uses mesh_geometry.edges_transformed.
• Check before changing interpolator.mesh_grid itself: the rectangular interpolator passes it to its mesh geometry
  (interpolator/rectangular.py ~L759), and preloads / to_inversion code paths touch source_plane_mesh_grid.
  A separate attribute that the mapper and the CSV writer use may be the safer shape — decide in the plan.

Work

1. Decide the API: make mapper.source_plane_mesh_grid return the adapted node centres for adaptive rectangular
   meshes, or add a node-centre attribute on the rectangular mesh geometry and route the mapper property and the
   CSV writer through it. Guard-ring nodes: state where they sit (clamped to the data span) and say so in the
   docstring.
2. Unit tests: node centres equal the positions where each node's basis weight is 1 (map a grid at the centres
   through the mapper and check the weight matrix is ~identity on interior nodes); the griddata integral check in
   the Witness; RectangularUniform unchanged.
3. Fix the CSV writer if it does not pick the change up automatically; note the coordinate change in the PR body
   (CSV contents change for adaptive meshes).
4. Re-run autolens_workspace_test scripts/imaging/mesh_magnification.py (areas untouched; it must stay green).
5. Then resume pixelization-mesh-mag-docs (autolens_workspace#605, branch feature/pixelization-mesh-mag-docs,
   WIP commit 7c30f74f): scale out the extra galaxy, interferometer coefficient, print the truth beside both routes,
   honest accuracy wording.