[Jammy2211]

Closes #665 — Source & Cluster arc phase 4b.

What changed

• New MeshGeometryKNearestNeighbor, built by InterpolatorKNearestNeighbor / InterpolatorKNNBarycentric. Its areas_for_magnification is the integral of each vertex's interpolation basis function over the convex hull, computed as follows:
  • evaluate the interpolator's own weights at the centres of a 256² grid (KNN_AREA_QUADRATURE_PIXELS);
  • drop the centres outside the hull;
  • sum per vertex with bincount, and rescale the total to the hull area.
• These meshes previously borrowed the Delaunay dual areas, which are exact only for the linear Delaunay interpolant.
• The areas are computed host-side and cached on first access, like the existing Delaunay dual-area fallback, so likelihood evaluations never pay for them (about 1–4 s when magnification is requested).

Measured

mesh	unit identity μ−1 (before → after)	lensed vs truth (before → after)	identity vs truth
KNearestNeighbor	+1.6e-2 → +1.3e-5	+6.10% → −0.23%	−5.11% → −0.12%
KNNBarycentric	−4.4e-2 → +1.8e-3	−1.71% → −0.40%	−3.57% → −0.14%

KNNBarycentric's remaining 1.8e-3 is the coarse sub-pixel reference in the unit test; its clipped weights are discontinuous. Quadrature convergence: the identity μ moves by less than 3e-5 between 200, 400 and 800 cells per axis. Every mesh is now within 1% of simulator truth except RTUAdaptDensity lensed (−1.82%, limited by the reconstruction). DelaunayNN is the only mesh still on approximate areas (within 0.5%).

Not in this PR: #609

The partial-block neighbour-search fix was prototyped here and then split out by human ruling. It passes a brute-force oracle, but breaks the KNN + AdaptSplit gradient check in autolens_workspace_test jax_grad/knn.py. The finding and the patch are on #609. These areas integrate whatever neighbours the interpolator selects, so they're exact with or without that fix.

Checks

• PyAutoArray full suite: 2251 passed, 4 xfailed.
• PyAutoGalaxy full suite: 1452 passed. PyAutoLens full suite: 1015 passed, 1 xfailed.
• Emulated no-JAX job, pixelization tests: 278 passed, 27 skipped.
• autolens_workspace_test scripts passed: imaging/mesh_magnification.py (KNN tolerances tightened to 1%), jax_grad/knn.py, misc/jax_assertions/knn_interpolator.py, misc/jax_assertions/knn_barycentric.py, imaging/mesh_interpolator_source_recovery.py.

Companion: autolens_workspace_test PR (tolerances). Merge this one first.

🤖 Generated with Claude Code