fix: AnalysisImaging.log_likelihood_function CPU branch returns figure_of_merit (not log_likelihood)

[Jammy2211]

Summary

• The CPU branch of AnalysisImaging.log_likelihood_function returned fit.log_likelihood while the JAX branch returned fit.figure_of_merit. For pixelization (inversion) fits these differ by the regularization log-det terms of the Bayesian log evidence.
• As a result, any nested sampler driven by the CPU path was optimising a chi²-only target instead of the log evidence — and drifted toward outer_coefficient ≈ 0 (noise-overfit, degenerate "no regularization" mode). JAX runs of the same model converged to the physical Bayesian-evidence maximum.
• Fix: return fit.figure_of_merit from both branches. For purely parametric fits this is a no-op because figure_of_merit == log_likelihood; for inversion fits the search now uses the mathematically correct objective.

Test plan

• New regression test test__log_likelihood_function__returns_figure_of_merit_for_pixelization exercises a Rectangular Pixelization with Constant regularization, asserts analysis.log_likelihood_function(instance) == fit.figure_of_merit, and verifies the two quantities genuinely differ (so the test isn't a tautology).
• Existing test__figure_of_merit__matches_correct_fit_given_galaxy_profiles (parametric Sersic) still passes because figure_of_merit == log_likelihood when there's no inversion.
• Full test_autolens/imaging/model/test_analysis_imaging.py passes (5/5).
• End-to-end verification on a simulated lens: Nautilus + CPU sparse + 4-core converges from the degenerate mode (einstein_radius=0.997, outer_coefficient=0.000) to the physical mode (einstein_radius=1.001, outer_coefficient=95.6), matching the JAX baseline to within numerical precision.

Impact

• All previous Nautilus / DynestyStatic / nested-sampler lens-imaging fits run on CPU with a pixelization source were optimising the wrong objective. Re-runs may produce different (correct) posteriors.
• JAX runs are unaffected — they were already using figure_of_merit.
• Non-pixelization (parametric-only) runs are unaffected — figure_of_merit == log_likelihood for those.

🤖 Generated with Claude Code