[Jammy2211]

Summary

Fixes #480. PointSolver.solve returned no images for a source on an intermediate plane of a multi-plane tracer, making any point-source modeling beyond "all sources at the same redshift" impossible.

The triangle search honoured plane_redshift and traced to the requested plane, but _filter_low_magnification did not: it built LensCalc.from_mass_obj(tracer), whose deflection callable is the whole tracer, i.e. the last plane. Candidate images of an intermediate-plane source are heavily de-magnified measured that way (~1e-3–1e-5 in the issue's configuration), so magnification_threshold discarded every one and the solve returned an empty grid. point_solver.py had plane_redshift in scope and did not forward it.

This is the pattern the fit side already uses. AbstractFitPoint.magnifications_at_positions and solved.py's _lens_calc_for both measure with LensCalc.from_tracer(..., plane_j=...) at the source's own plane — and positions/image/abstract.py already passed plane_redshift into solver.solve. So the two halves of one likelihood evaluation disagreed: model_data measured at the last plane and returned nothing while magnifications_at_positions measured at the source's plane and was right. The argument was plumbed end-to-end through the callers; only the last consumer ignored it.

No-op for the default case, verified numerically rather than argued. Tracer.deflections_yx_2d_from dispatches to deflections_between_planes_from with plane_i=0, plane_j=-1 whenever the tracer has more than one plane, so from_mass_obj(tracer) and from_tracer(tracer, plane_j=-1) wrap the same callable. Magnifications compare bit-identical (max|diff| = 0.0) for multi-plane and single-plane tracers alike.

API Changes

No public API removed or renamed. Behaviour changes:

• PointSolver.solve(..., plane_redshift=z) now returns the correct images for an intermediate plane, where it previously returned an empty grid. Results for plane_redshift=None and for the last plane are unchanged.
• A plane_redshift matching no plane now raises ValueError naming the tracer's actual redshifts. It previously reached traced_grids_list[None] and raised a bare TypeError.

The plane-index resolution is now shared by the search and the filter (ShapeSolver._plane_index) — the bug was those two disagreeing about the plane, so a second copy of the resolution invites the same drift back.

Test Plan

• New test_autolens/point/triangles/test_solver_multi_plane.py (9 tests). Nothing passed plane_redshift to the solver in any test before this — the multi-plane solve path had no coverage, which is why the bug shipped.
• 5 of the 9 fail against the unfixed solver; the other 4 pin behaviour that must not change.
• Full test_autolens: 544 passed, 1 skipped (coolest not installed).
• JAX solve path exercised directly — same 4 images as NumPy.

Magnification accuracy was cross-checked three ways at the solved plane, agreeing to 1.7e-08:

Method	µ at the four solved images
NumPy Richardson FD (shipped path)	-27.89795126 -9.05806212 23.12205173 15.99681364
JAX exact autodiff (float64)	-27.89795173 -9.05806221 23.12205188 15.99681364
Ray-traced Jacobian (no LensCalc)	-27.89795168 -9.05806221 23.12205188 15.99681364

Full API Changes (for automation & release notes)

Added

• ShapeSolver._plane_index(tracer, plane_redshift=None) -> int — resolves the plane being solved for, shared by _plane_grid and _filter_low_magnification. Raises ValueError when no plane matches.

Changed

• ShapeSolver._filter_low_magnification(tracer, points, xp, plane_redshift=None) — gained plane_redshift; now constructs via ag.LensCalc.from_tracer(..., plane_j=plane_index) instead of from_mass_obj(tracer).
• PointSolver._solve_array — forwards plane_redshift to _filter_low_magnification.

Migration

None. Both changed methods are private, and the public solve signature is unchanged.

Follow-ups filed (not in this PR)

Cross-checking this fix surfaced three separate items, filed as PyAutoMind prompts rather than widened into here:

• LensCalc's NumPy Richardson Hessian uses a hardcoded 0.01″ step. At the last plane of this same tracer it disagrees with both JAX autodiff and the ray-traced Jacobian by 122%, with the wrong sign on three of four points. Pre-existing and independent of bug: PointSolver magnification filter uses last-plane magnification instead of plane_redshift #480, but it is the default path.
• Multi-plane time delays — the time-delay equivalent of this fix. tracer_util.time_delays_from raises on >2 planes, and LensCalc.from_tracer binds the deflection callable to the plane while leaving potential_2d_from as the tracer's all-planes sum.
• A refactor giving plane-bound lensing quantities one construction path; three call sites currently hand-roll their own and resolve the plane by different keys.

Generated by the PyAutoLabs agent workflow.

---

Generated by Claude Code