[Jammy2211]

Summary

mp.Chameleon.potential_2d_from did not match its own deflections. For Chameleon(ell_comps=(0.2,-0.1), intensity=1.0), ∇ψ came out ≈1.57 × α (≈ 1/q), an error of about 52% of max|α|. The deflections were already correct (div α = 2κ everywhere). The cause was the MGE potential: it passed ellipticity_convention="circularised" (copied from Sersic), but Chameleon's convergence uses the major-axis elliptical radius. Switching the convention alone still left ~0.35% of max|α| (8–11% at r≈3″), so this PR replaces the MGE route with the exact closed-form potential of the two cored isothermals (Keeton 2001). It reuses the same psi_from, normalisation and rescaled core radii as deflections_2d_via_analytic_from.

Closes #682.

API Changes

None to signatures. Behaviour change: Chameleon.potential_2d_from / ChameleonSph.potential_2d_from now return the exact analytic potential, normalised so ψ(centre) = 0, instead of the MGE approximation. Values change, most for elliptical profiles. It is also faster, since there is no MGE decomposition or quadrature.
See full details below.

Test Plan

• New tests in test_chameleon.py: the central-finite-difference ∇ψ matches α to < 1e-4 of max|α| (elliptical, spherical, offset-centre large-core with M/L=2); the 5-point ∇²ψ matches 2κ to 1e-3; spot values (ψ = 0 at the centre, plus a regression value).
• Control: on main, 7 of the 14 Chameleon tests fail; with the fix all 14 pass. The measured residual is ≈2e-8, which is finite-difference noise.
• Full test_autogalaxy: 1461 passed. PyAutoLens test_autolens: 1015 passed, 1 xfailed (run on the first commit).
• JAX eager (xp=jnp) matches numpy to 2e-16; new jit-over-ell_comps tests (mass deflections + potential, light image) pass, and fail on the pre-follow-up code.
• CI.

Also fixed in this PR (follow-ups from review)

• jax.jit: Chameleon.axis_ratio (mass and light) capped q with a Python if, which raised TracerBoolConversionError once parameters were traced. It now uses xp.where, as Gaussian does. lp.Chameleon.image_2d_from also dropped xp when calling image_2d_via_radii_from, which raised TracerArrayConversionError. Deflections, potential, convergence and the light image now all jit over ell_comps and match numpy to ~3e-16.
• Docstring vs code (decision taken — flagged for review): the mass Chameleon / ChameleonSph docstrings claimed the second core is s0 + s1 ("increment"). The code uses core_radius_1 directly as the second core, in convergence, deflections and potential alike, with 2s/(1+q) rescaling and a 1/(1+q) prefactor. I changed the docstrings to state the implemented convergence and left the parametrisation alone. Changing the code would shift every existing Chameleon fit, and the light Chameleon docstring already describes core_radius_1 as the second core itself. The corrected formula reproduces the existing pinned convergence test value (0.018605).

Full API Changes (for automation & release notes)

Changed Behaviour

• autogalaxy.profiles.mass.stellar.chameleon.Chameleon.potential_2d_from — now analytic (difference of two cored-isothermal potentials, Keeton 2001); gradient equals deflections_yx_2d_from; ψ(centre) = 0. Previously an MGE approximation with the wrong ellipticity convention.
• ChameleonSph.potential_2d_from — inherits the above.

Changed Behaviour (follow-ups)

• Chameleon.axis_ratio (mass mp.Chameleon, light lp.Chameleon) — now xp.where-capped; same values, now traceable under jax.jit.
• lp.Chameleon.image_2d_from — passes xp through (JAX path no longer falls back to numpy).

Migration

• None required; any stored Chameleon potential values (e.g. time delays / Fermat potentials) will change.

Generated by the PyAutoLabs agent workflow.

🤖 Generated with Claude Code