fix: Delaunay areas_for_magnification returns barycentric dual areas, not Voronoi cells (#524)

[Jammy2211]

Summary

Fixes #524, the first of the two defects proven by the euclid-dr1-prep phase 8 audit (#522, PR #523).

MeshGeometryDelaunay.areas_for_magnification returned scipy Voronoi cell areas (only unbounded cells zeroed). The Delaunay mapper is a barycentric-linear interpolant, so the exact quadrature weight for Σ s_i × area_i — the integral of the reconstruction over the mesh hull — is the barycentric dual area (Σ triangle_area/3 per vertex). Bounded boundary Voronoi cells can be orders of magnitude larger than their dual areas, which the audit measured biasing magnification by −13 % to −99 % on realistic meshes.

• The dual areas were already computed on both interpolator paths (numpy scipy_delaunay, in-graph JAX jax_delaunay) and discarded after the split-point offsets. They are now returned (sixth tuple element; the Matérn variants return them too, as a fourth element, through one shared in-graph helper _dual_areas_padded_jnp), carried on DelaunayInterface.dual_areas, exposed as InterpolatorDelaunay.dual_areas, and handed to MeshGeometryDelaunay(dual_areas=) by mesh_geometry. On the JAX path the value is traced, so areas_for_magnification is safe inside the per-sample jax.jit PyAutoLens evaluates latents in.
• areas_for_magnification returns the dual areas. A geometry built standalone (dual_areas=None, e.g. the tests, or the kNN interpolators which build no Delaunay tables and override dual_areas to None) computes the identical value host-side from a scipy triangulation, only if asked. voronoi_areas / voronoi_areas_numpy are unchanged and remain available under their own name. Docstring rewritten to say which quadrature this is and why.
• _plot_delaunay: docstring no longer claims Gouraud shading (tripcolor defaults to flat, whose integral equals the dual-area integral by linearity), and the mapper's own simplices (padding rows dropped) are passed as triangles= so the figure shows the triangulation the inversion used; mappers without Delaunay tables fall back to matplotlib's own.
• Tests: the two phase-8 tests that deliberately pinned the Voronoi semantics are flipped (bounded_boundary_cells_are_kept → equals_barycentric_dual_area; uniform_lattice now asserts the hull total (n-1)² instead of (n-2)²). New mapper-level identity test test__areas_for_magnification__integrates_reconstruction_like_mapping_matrix and JAX-parity test test__areas_for_magnification__jax_matches_numpy in mappers/test_delaunay.py.

Identity-test numbers (8×8 mesh, outer ring pinned to the data footprint, interior jittered, random positive reconstruction; hull area = footprint area = 4.0 exactly, no out-of-hull fallback):

quantity	value
F_map = Σ (mapping_matrix @ s) × pixel_area	4.730578587726577
F_dual = Σ s × areas_for_magnification (new)	4.730741058881734 — rel err 3.4e-5
F_voronoi = Σ s × voronoi_areas (old semantics)	3.453335121760202 — rel err 27 %

One base file, test_autoarray/inversion/pixelization/interpolator/test_delaunay.py, was not black-clean on main; black touched about a dozen pre-existing lines in it alongside this PR's edits.

Out of scope, as filed: the rectangular-mesh lead (cluster epic), DelaunayNN/KNN area definitions, a library-level magnification API, and the second audit defect (magnification latent 0/0 for pixelized sources — separate PyAutoLens prompt).

API Changes

Changed behaviour: MeshGeometryDelaunay.areas_for_magnification now returns the barycentric dual areas (was Voronoi cell areas with unbounded cells zeroed). Any consumer summing reconstruction × areas_for_magnification gets the exact integral of the reconstruction; the only known consumers are the four source_science.py workspace scripts, which call the same attribute name and need no code change.
Added: MeshGeometryDelaunay(dual_areas=), DelaunayInterface.dual_areas, InterpolatorDelaunay.dual_areas; scipy_delaunay / jax_delaunay return a sixth element and the Matérn variants a fourth (internal helpers).
See full details below.

Test Plan

• pytest test_autoarray/inversion -q — 476 passed, 1 skipped
• pytest test_autoarray -q -n auto — 1420 passed, 1 skipped
• black --check on every changed file
• CI green on this PR

Full API Changes (for automation & release notes)

Changed Behaviour

• autoarray.inversion.mesh.mesh_geometry.delaunay.MeshGeometryDelaunay.areas_for_magnification — returns the barycentric dual area of every mesh vertex (sums to the convex-hull area; exact quadrature of the barycentric-linear reconstruction). Previously returned voronoi_areas with the -1 unbounded-cell sentinel zeroed.
• autoarray.plot.inversion._plot_delaunay — passes the mapper's simplices as triangles= when available (rendering unchanged where matplotlib's triangulation coincided, which it did on every audited mesh).

Added

• MeshGeometryDelaunay.__init__(..., dual_areas=None, xp=np, **kwargs) — accepts the interpolator's dual areas.
• DelaunayInterface.__init__(..., xp=np, dual_areas=None) / DelaunayInterface.dual_areas — trailing keyword, so positional subclass callers (DelaunayNNInterface) are unaffected.
• InterpolatorDelaunay.dual_areas (property) — the interface's dual areas; InterpolatorKNearestNeighbor.dual_areas returns None.
• scipy_delaunay / jax_delaunay return (points, simplices_padded, mappings, split_points, splitted_mappings, areas); scipy_delaunay_matern / jax_delaunay_matern return (points, simplices_padded, mappings, areas).
• _dual_areas_padded_jnp(points, simplices_padded) — private in-graph helper shared by the JAX paths.

Migration

• Callers unpacking the scipy_delaunay / jax_delaunay / matern tuples directly (one in-repo test did) must accept the extra trailing element.
• Code that relied on areas_for_magnification zeroing hull cells should use voronoi_areas and zero the -1 entries itself.

Generated by the PyAutoLabs agent workflow.

🤖 Generated with Claude Code

https://claude.ai/code/session_01VVLuDrRfkZ81TBh8yVbFrN

---

Generated by Claude Code