feat: MGE/CSE fallback for zero-returning mass profile potentials

[Jammy2211]

Summary

Replace zero-returning potential_2d_from and convergence_2d_from methods across 10 mass profiles with proper implementations. Adds a new potential_2d_via_mge_from method to MGEDecomposer using the E1 (exponential integral) formula from Shajib (2019), cross-validated against Jax-Lensing-Profiles (Herculens). Phase 3 of the mass profiles refactoring epic (#445).

Three tiers of implementation:

• Analytic: PointMass (ψ = R_E² ln r) and MassSheet (ψ = ½κr²)
• MGE convergence: cNFW/cNFWSph now compute convergence via MGE (same pattern as their existing MGE-based deflections)
• MGE potential: All remaining profiles (Sersic, Gaussian, Chameleon, PowerLawBroken, dPIE families, NFWTruncated, cNFW) compute potential via the new potential_2d_via_mge_from

Self-consistency validated: laplacian(ψ_MGE) = 2κ_MGE to 0.3% median error. The ~5-9% error vs profiles with known analytic potentials is inherent to the MGE decomposition (convergence itself has similar error), not the potential formula.

API Changes

Added MGEDecomposer.potential_2d_via_mge_from() — new public method computing lensing potential from MGE-decomposed convergence using the E1 exponential integral. Added helper methods E1() (Abramowitz & Stegun approximation, JAX-compatible) and potential_func_gaussian(). All existing zero-returning methods now return physically meaningful values instead of zeros. See full details below.

Test Plan

• pytest test_autogalaxy/profiles/mass/ — 406 passed, 0 failed
• MGE potential self-consistency: laplacian(ψ) = 2κ to 0.3% median
• MGE potential vs Isothermal/PowerLaw analytic: ~5% median (decomposition-limited)
• Updated test_mass_sheet.py to match new correct potential values
• Phase 1 self-consistency tests (autolens_workspace_test scripts/mass/) — follow-up

Full API Changes (for automation & release notes)

Added

• MGEDecomposer.E1(x, xp) — static method, exponential integral of first order (Abramowitz & Stegun 1970)
• MGEDecomposer.potential_func_gaussian(grid_radii, sigma, intensity, xp) — potential of a single Gaussian convergence component
• MGEDecomposer.potential_2d_via_mge_from(grid, *, sigma_log_list, three_D, ellipticity_convention, ...) — lensing potential via MGE decomposition
• MGEDecomposer._potential_2d_via_mge_from(grid_radii, *, sigma_log_list, ...) — internal implementation

Changed Behaviour

• PointMass.potential_2d_from — was zeros, now returns ψ = R_E² ln(r)
• MassSheet.potential_2d_from — was zeros, now returns ψ = ½κr²
• cNFW.convergence_2d_from / cNFWSph.convergence_2d_from — was zeros, now MGE-computed
• cNFW.potential_2d_from / cNFWSph.potential_2d_from — was zeros, now MGE-computed
• NFWTruncatedSph.potential_2d_from — was zeros, now MGE-computed
• AbstractSersic.potential_2d_from (Sersic, SersicSph) — was zeros, now MGE-computed
• Gaussian.potential_2d_from — was zeros, now MGE-computed
• Chameleon.potential_2d_from — was zeros, now MGE-computed
• PowerLawBroken.potential_2d_from — was zeros, now MGE-computed
• dPIEMass.potential_2d_from — was zeros, now MGE-computed
• dPIEPotential.potential_2d_from / dPIEPotentialSph.potential_2d_from — was zeros, now MGE-computed

Migration

No migration needed — methods that returned zeros now return physically correct values. Any code that relied on these methods returning zeros was already broken (using a placeholder).

🤖 Generated with Claude Code