feat: split rectangular adaptive meshes — Bilinear (rank-CDF) vs RTU (kernel-CDF)

[Jammy2211]

Summary

Splits the rectangular adaptive mesh family into two user-facing variants (part of #461 — library phase; workspace PRs follow once this merges):

• RectangularBilinearAdaptDensity / RectangularBilinearAdaptImage — the empirical rank-CDF lattice transform (sort + cumsum, no kernel hyperparameters), recovered from the pre-refactor: consolidate rectangular meshes — kernel takes plain names, retire linear/spline/rotated #402 implementation (22b28463^) as create_transforms_rank. O(N log N) on CPU, eliminating the O(M_sub × N_data) erf sum that dominates the numba CPU likelihood (55% of a euclid eval, 89% at hst). Becomes the workspace default in the follow-up PRs.
• RectangularRTUAdaptDensity / RectangularRTUAdaptImage — pure renames of the current kernel-CDF classes (implementation and likelihood values unchanged). The advanced option for GPU, JAX gradient samplers and the interferometer sparse path, where they remain the only adaptive rectangular meshes with usable gradients. The RTU implementation is deliberately kept forever.
• RectangularUniform unchanged (now inherits under the RTU name).

The seam is a transform="rank"|"kernel" selector threaded through the create_transforms dispatch, the adaptive_rectangular_* functions, InterpolatorRectangular and MeshGeometryRectangular, so the geometry's areas/edges always route through the same transform as the mapper. Both transforms warp the same lattice and share the existing 4-pixel bilinear interpolation.

Docstrings encode the guidance split: gradient inference on the Bilinear meshes needs over_sample_size_pixelization >= 4 (imaging — FD-certified in the July gradient audit); at over-sampling 1 the rank-CDF likelihood is exactly piecewise-constant in mass/shear (zero gradients — the reason it was deleted in #402). Interferometer gradient work needs RTU.

API Changes

Breaking rename plus two new mesh classes — no back-compat aliases:

• RectangularAdaptDensity → RectangularRTUAdaptDensity, RectangularAdaptImage → RectangularRTUAdaptImage (pure renames, kernel-CDF internals and likelihood values untouched).
• New RectangularBilinearAdaptDensity(shape) / RectangularBilinearAdaptImage(shape, weight_power, weight_floor) — rank-CDF adaptive meshes with no bandwidth/n_knots hyperparameters.
• RectangularUniform unchanged.

Downstream workspaces (autolens_workspace ~224 uses, autogalaxy_workspace, HowToLens, autolens_workspace_test pins, prior-config yamls) update in follow-up PRs.
See full details below.

Test Plan

• Full suite python -m pytest test_autoarray/ on Python 3.12: 1114 passed; the only 3 failures are the pre-existing test_transformer.py pynufft failures, reproduced identically on pristine main in the same environment.
• New unit tests: rank-CDF transform exactness at the traced points (ranks / weighted cumsum), monotonicity/bounds, exact round-trip, mappings/weights normalization, area telescoping to the bounding box, density adaptivity, invalid-transform rejection, Bilinear mesh construction/dispatch/signature, split-regularization guard extended to all five rectangular meshes.
• JAX/xp parity + gradient behaviour: unit tests are NumPy-only by project rule — validated by the parity scripts in autolens_workspace_test/autogalaxy_workspace_test in the workspace follow-up (pins regenerated there).
• CPU speed-up measured in the existing autolens_profiling cells (follow-up, recorded as a versioned result on autolens_profiling#153 / PROGRAMME Phase 14).

Full API Changes (for automation & release notes)

Renamed

• aa.mesh.RectangularAdaptDensity → aa.mesh.RectangularRTUAdaptDensity (module autoarray/inversion/mesh/mesh/rectangular_rtu_adapt_density.py)
• aa.mesh.RectangularAdaptImage → aa.mesh.RectangularRTUAdaptImage (module autoarray/inversion/mesh/mesh/rectangular_rtu_adapt_image.py)

Added

• aa.mesh.RectangularBilinearAdaptDensity(shape=(3, 3)) — rank-CDF density-adaptive rectangular mesh (no kernel hyperparameters).
• aa.mesh.RectangularBilinearAdaptImage(shape=(3, 3), weight_power=1.0, weight_floor=0.0) — rank-CDF adapt-image-weighted rectangular mesh.
• create_transforms_rank(traced_points, mesh_weight_map=None, xp=np) in autoarray/inversion/mesh/interpolator/rectangular.py — the recovered empirical rank-CDF transform pair (plus restored forward_interp / forward_interp_np).

Changed Signature

• adaptive_rectangular_transformed_grid_from, adaptive_rectangular_areas_from, adaptive_rectangular_mappings_weights_via_interpolation_from: new keyword transform="kernel" ("rank" selects the empirical rank CDF; bandwidth/n_knots are then unused).
• InterpolatorRectangular.__init__: new keyword transform="kernel".
• AbstractMeshGeometry.__init__ / MeshGeometryRectangular: new keyword transform="kernel" — areas/edges route through the same transform as the mapper.

Changed Behaviour

• None for existing classes: the RTU meshes are pure renames (likelihood values unchanged); RectangularUniform is untouched.

Migration

• Before: aa.mesh.RectangularAdaptDensity(shape=(30, 30), bandwidth=1.0)
• After (same behaviour): aa.mesh.RectangularRTUAdaptDensity(shape=(30, 30), bandwidth=1.0)
• After (fast CPU default): aa.mesh.RectangularBilinearAdaptDensity(shape=(30, 30)) — gradient users set over_sample_size_pixelization >= 4 or stay on RTU; interferometer gradient users stay on RTU.

Generated by the PyAutoLabs agent workflow.

---

Generated by Claude Code