Work Classification

Both (library first: PyAutoArray → PyAutoLens, then autolens_visualization)

Affected Repositories

• PyAutoLens (primary)
• PyAutoArray (tick precision fix)
• autolens_visualization (re-rendered figures)

Branch Survey

Suggested branch: feature/point-fit-source-plane-panel
Worktree root: ~/Code/PyAutoLabs-wt/point-fit-source-plane-panel/

Implementation Steps

1. PyAutoArray autoarray/plot/utils.py: _round_ticks raises its significant figures until distinct inputs stay distinct (used by apply_extent and yx); test in test_autoarray/plot/test_utils.py.
2. PyAutoLens new autolens/point/plot/source_plane_panel.py: presentation table per Fit class, traced positions (same path as FitPositionsSource.model_data), σ/|μ| radii, A·σ ellipses from precision_tensor_components_from(fit, "jacobian") (covariance σ²AAᵀ), own-plane caustics via LensCalc.from_tracer(tracer, use_multi_plane=True, plane_i=0, plane_j=fit.plane_index), zoom extent, corner-ranked inset/legend/χ² box.
3. positions_panel.py: shared image-number placement (generalised from point_solver_plots.number_placements_from), 1..N numbering; drop the source-plane crosses from the image panel.
4. fit_point_plots.subplot_fit: 1×2 / 1×3 layout with the new panel.
5. Tests: per-class presentation (all 7 classes, both weightings for source classes), panel artists, fluxes layout, multi-plane caustic plane.
6. autolens_visualization: re-render the three point domains, rebuild gallery + manifest.

Key Files

• autolens/point/plot/fit_point_plots.py — subplot layout
• autolens/point/plot/positions_panel.py — shared panel helpers
• autolens/point/plot/source_plane_panel.py — new
• autolens/point/plot/point_solver_plots.py — numbering helper reuse
• autoarray/plot/utils.py — tick rounding

Point-source fit: source-plane panel showing how closely the positions trace

Consequence: judge
Autonomy: human-required
builds-on: complete/2026/10/point-dataset-positions-panel.md (shipped 2026-10-21; shares the fit_point_plots.py styling: markers, σ legend, titles, solver extent)

@PyAutoLens @autolens_visualization

Request (human, 2026-10-21)

A follow up issue would be to do a source plane plot of the fit showing how close things trace, which is relevent for image chi and source chi.

Intent

Add a source-plane panel to the point-source fit.png, possibly also as its own figure. It shows:

• every observed position ray-traced to the source plane through the fit's tracer
• the model source centre
• the caustics

How tightly the traced positions bunch around the source centre is the quantity FitPositionsSource (source chi-squared) minimises. For image chi-squared (FitPositionsImagePair and its variants) it is the intuitive check that the model is self-consistent.

Open design questions (settle in /start_dev)

• Scale bar for the scatter. Options are the source-plane noise from magnification-scaled σ (what source chi-squared uses, roughly σ/√μ), a circle per traced position, or a σ annotation.
• Extent. Zoom to the traced positions plus a margin, the caustic extent, or both as an inset. Too zoomed out and the scatter is invisible; too zoomed in and the caustic context is lost.
• Multiple planes. For multi-plane and cluster fits, either one panel per source plane or only the point's own plane.
• Figure placement. In fit.png for both image and source chi-squared, or a separate fit_source_plane.png toggled by config.
• Re-render set. point_source_image_chi_squared, point_source_source_chi_squared and cluster in autolens_visualization.

Prototype findings + recommended defaults (2026-10-21; NOT yet approved — confirm at /start_dev)

Prototypes. /mnt/c/Users/Jammy/Science/scratch/eyes/autolens/source_plane_prototype/ (19 PNGs; producer in producer/). The best views are S3_fit_mockup_image_chi_perturbed_RE.png and S3_fit_mockup_source_chi_truth.png.

Recommended design. fit.png gets a second panel, "Fit Source Plane (point_0)", for every positions fit (image- and source-χ²):

• Zoom: the traced observed positions plus the model source centre, with a 30% margin.
• Inset: the full caustic in the emptiest corner, with a box marking the zoom.
• Traced positions: numbered black dots.
• Model source centre: red ★.
• Noise ellipse: an orange A·σ ellipse on every traced point (the honest propagated noise).
• Noise circle: the σ/|μ| purple circle as well, but only when the fit uses weighting="magnification".
• χ² box: the fit's χ², plus the Jacobian χ².

Other recommendations:

• Move the source-χ² "Model (source plane)" ✕ off the image-plane panel into this one.
• Number images 1–N to match the PointSolver figure (the prototype used 0–N−1).
• No model-image markers for image-χ² fits (they all land on the ★).
• Absolute arcsec ticks, with an apply_extent precision fix: at mas zoom its 2-sig-fig rounding collapses all ticks to one value ("1.2″" ×3 on cluster point_1).
• Cluster inset frames only the tangential caustic.
• With fluxes, fit.png becomes 1×3.
• Multi-plane: the point's own plane (LensCalc.from_tracer(..., plane_j=fit.plane_index) works today for caustics). This pairs with the critical-curve redshift bug and the cluster pass.

Related. The default-weighting question is filed separately as draft/bug/autolens/fitpositionssource_default_magnification_weighti.md (check against Gravity.jl).

Side note (uninvestigated). At truth, the image-pair point_source model solves 5 images for 4 observed.

Human approval (2026-10-21)

Points 1–6 and all of the recommendations above are APPROVED.

Additional binding principle: pair each figure to its Fit class.

• Image-plane χ² fits (FitPositionsImagePair / PairRepeat / PairAll / …): the source-plane panel is context only (traced positions, model source ★, caustic, inset). Draw nothing that looks like a χ²: no noise circles or ellipses, no χ² box. Subtitle e.g. "Image-plane χ² — traced positions for context".
• Source-plane χ² fits: clearly indicate WHICH source-plane χ².
  • The subtitle names it, e.g. "Source-plane χ² (magnification weighting)" or "(jacobian weighting)".
  • The footprint the χ² actually uses is primary: the σ/|μ| circle for magnification, the A·σ ellipse for jacobian.
  • Under magnification weighting, the A·σ ellipse is also shown, faint or dashed, as "propagated noise (reference)" (approved point 1).
  • Free-centre and solved-β* (FitPositionsSourceSolved) fits use visibly different centre markers (★ vs ◆), with legend text naming them.
  • The χ² box appears only on source-plane fits.
• Enumerate every positions Fit class in autolens/point/fit/positions/ and map each to one of these presentations, covered by tests.