v2026.5.29.4

PyAutoGalaxy v2026.5.29.4

What's New

Breaking Changes

• fix(mass): convergence_func on PowerLawBroken, PowerLawMultipole, cNFW family (#467)
  • Added convergence_func overrides on PowerLawBroken, PowerLawBrokenSph (inherited), PowerLawMultipole, and the cNFW family (cNFWSph + MCR variants inherit). PowerLawBroken._convergence is a new private radial helper. PowerLawBroken.potential_2d_from now raises NotImplementedError explicitly (it was already non-functional). No signatures of existing public methods changed. See full details below.
• fix: soft-fail jax_zero_contour callers in lens_calc to NaN/[] (#465)
  • Behaviour change (no signature change): when jax_zero_contour is not installed, einstein_radius_jit_from now returns float('nan') (was: ModuleNotFoundError), and the public tangential_critical_curve_list_via_zero_contour_from / radial_critical_curve_list_via_zero_contour_from (via the private _critical_curve_list_via_zero_contour) now return [] (was: ModuleNotFoundError). One logger.warning per process per feature. New private helper _maybe_optional_dep_warn and module-level set _OPTIONAL_DEP_WARNED. No imports removed.
• fix: raw-flux latent + soft-fail magzero-required µJy (#463)
• feat: MGE/CSE fallback for zero-returning mass profile potentials (#449)
  • 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.
• refactor: port CSE module to support JAX via xp parameter (#447)
  • Added xp=np keyword argument to four MassProfileCSE methods: convergence_cse_1d_from, deflections_via_cse_from, _convergence_2d_via_cse_from, _deflections_2d_via_cse_from. Existing callers passing no xp argument are unaffected (default is np). The decomposition solver (_decompose_convergence_via_cse_from) stays NumPy-only — it runs before JIT tracing. See full details below.
• feat: first-class latent variable API in PyAutoGalaxy (#441)
  • New public module autogalaxy.imaging.model.latent with helpers (ab_mag_via_flux_from, flux_mujy_via_ab_mag_from), the LATENT_FUNCTIONS registry, the latent_keys_enabled() reader, and one registered latent (total_galaxy_0_flux_mujy, disabled by default). AnalysisImaging gains a LATENT_KEYS @property and a compute_latent_variables(parameters, model) method — both new public surface. New config file autogalaxy/config/latent.yaml. Note: autoconf lowercases yaml keys, so the latent name uses lowercase mujy (not muJy) — this leaks through to the latent.csv column header.
• feat: JIT-friendly Einstein radius helper + AnalysisDataset jit latent mode (#435)
  • Two additions to PyAutoGalaxy: a new public method LensCalc.einstein_radius_jit_from, and a class attribute override AnalysisDataset.LATENT_BATCH_MODE = "jit". No existing API surface is modified or removed.
• perf(lens_calc): cache (f, ZeroSolver) for zero_contour critical curves (#434)
  • LensCalc instances now carry a private _zero_contour_cache dict. No public API surface changes — same method names, same signatures, same return values. Behavioural change: _critical_curve_list_via_zero_contour reuses its closure and ZeroSolver across repeat calls with the same parameters.

New Features

• perf: cache expensive @Property on Fit classes (#462)
• feat: vmapped_deflections_from for batched subhalo deflections (#455)
• docs: LaTeX docstrings for all mass profile classes (#453)
• feat: SimulatorInterferometer.via_galaxies_from auto-default xp from parent use_jax (#443)
• feat: SimulatorImaging.via_galaxies_from auto-default xp from parent use_jax (#442)

Bug Fixes

• fix(jax): defensive pytree dedup in imaging/interferometer analyses (#468)
• fix(mass): wire convergence_func on dPIE family for MGE decomposition (#466)
• fix: elliptical MGE potential via deflection line integral (#460)
• fix: use xp.sqrt in NFWSph.potential_func_sph (#458)
• fix: add xp=np to convergence_func across all mass profiles (#457)

Internal

• perf: vectorize MGE potential over components (#461)
• fix: cNFWSph deflection boundary bug and MCR validation (#451) (#454)
• refactor(model_util): replace simulator_start_here_model_from with direct random_galaxy_for_simulation_from (#438)
• refactor: archive quantity package to autolens_workspace_developer/legacy (#437)

Upstream Changes

PyAutoFit

• chore(deps): allow anesthetic>=2.9.0 to unblock jax>=0.7 / numpy>=2 resolution (#1306)
• fix(nss): chunked algo.init follow-up to #1303 (#1305)
• feat(nss): chunk_size kwarg for inversion-heavy A100 likelihoods (#1303)
• fix(jax): structural defense against cached_property pytree/dict leaks (#1302)
• fix(jax): keep parameterization cache off ModelInstance + auto-register pytrees (#1300)
• Cache model.parameterization; try interactive matplotlib backends (#1299)
• Prefer fit_quick.png in quick-update display candidates (#1298)
• Remove use_jax_for_visualization; add visualization warmup (#1297)
• fix: skip _compute_latent_samples in PYAUTO_TEST_MODE (#1294) (#1295)
• Add live_visual_update flag for opt-in on-the-fly visualization (#1293)
• fix: PYAUTO_TEST_MODE should write to a separate output dir (#1292)
• feat(quick_update): IPython.display.update_display for live Jupyter cells (#1290)
• feat(analysis): LATENT_BATCH_MODE attribute (vmap default, jit option) (#1288)
• Fix Sample.kwargs mixed string/tuple key bug (#1287)
• nss extras: strip git+https URLs to unblock PyPI uploads (#1286)

PyAutoArray

• fix(jax): exclude cached_property descriptors from pytree flatten paths (#343)
• fix: VectorYX2DIrregular from_dict round-trip (missing values property) (#342)
• perf: cache expensive @Property on Fit classes (#341)
• fix: make Array2D.native jit-traceable for JAX simulator path (#339)
• fix: raise ValueError on xp=np + jnp-backed-grid mismatch (#337)
• feat: SimulatorInterferometer(use_jax=True) + xp-aware preprocess Gaussian noise (#336)
• feat: SimulatorImaging(use_jax=True) + xp-aware preprocess noise (#335)
• TransformerNUFFT: add chunk_size knob to cap nufftax gather buffer (#330)
• interferometer: enable sparse_operator for nufftax TransformerNUFFT (#329)

---

Full changelog: 2026.5.21.1...2026.5.29.4