[samlange04]

Discussion: https://github.com/orgs/PyAutoLabs/discussions/38.

corner_anesthetic produced no plot for any model with two components of one class (most lens models) and logged
"posterior estimate not yet sufficient" instead: anesthetic keys its columns by label, the LaTeX labels repeat
(every Gaussian of a basis is $\epsilon_1^{\rm g}$), plot_2d raised KeyError and @log_plot_exception
swallowed it. This PR fixes that and makes the plots of a many-component model readable.

Fix. Columns are keyed by position (p0, p1, ...) and the LaTeX labels passed as anesthetic display labels
(NestedSamples(labels=...), make_2d_axes(labels=...)); labels joins the reserved kwargs.

Labels. AbstractPriorModel.superscripts is the dot-joined path of the names the owning components were given
in the model (lens.bulge, source.bulge, lens.shear), via new component_paths: Collection containers,
tuple names and the list index of a basis sharing one prior are dropped, an index distinguishing independent priors
is kept, a prior shared between differently named components takes the path its samples column is named after
(unique_prior_paths). The notation config's class abbreviation applies only to a prior on a bare unnamed Model.
Paths three or more levels deep keep the first name and cut the rest to three characters
(superscript_from_component_path: source.pix.reg). Tests in test_model_mapper.py cover each rule.

Per-component plots. New parameter_indices_by_component (top-level component -> parameter indices; root
priors as "model") and a parameter_indices kwarg on corner_anesthetic; WeightedDefaults.plot_results writes
corner_anesthetic_<component>.png per component when there is more than one. Config
visualize/plots_search.yaml nest.corner_anesthetic_components: true; PosteriorDefaults.should_plot(name, default=) falls back for a key an older workspace config lacks.

Latent parameters. autofit/mapper/prior_model/latent_parameters.py: a component class declares
latent_parameters = {("gamma_1", "gamma_2"): fn}; fn takes every parameter of the component by name (free as
arrays over the samples, fixed as floats, tuples flattened name_0/name_1) and returns {name: array};
declarations merge down the MRO. AbstractPriorModel.latent_parameters_from(parameter_array, weights=None) returns a
LatentParameters (names, superscripts, component paths, full model paths, array, mask); has_latent_parameters.
Periodic(values, period).recentred(weights) wraps an angle into one period centred on its weighted circular mean,
so a posterior at the branch cut is not drawn as two peaks; exported as af.Periodic.

• corner_anesthetic(latent=True) plots the latent set; WeightedDefaults.plot_latent_results writes
  corner_anesthetic_latent.png and corner_anesthetic_latent_<component>.png for the components with a latent
  parameter (none for a single-component model), under config nest.corner_anesthetic_latent (default false),
  independent of the other two switches.
• latent_samples_from appends the model's latent parameters to the latent samples under their model path
  (galaxies.lens.shear.magnitude), with or without analysis latents, so they reach latent/samples.csv,
  latent.results and latent_summary.json.
• The latent plot is latent output: a new plot_latent_results on NonLinearSearch / PosteriorDefaults (no-op
  for chain and point kinds), which SearchUpdater.visualize calls once the fit completes and during it only if
  output.yaml's latent_during_fit is on. plot_results and the iterations_per_full_update cadence are
  unchanged; searches overriding plot_results(samples) are unaffected.

Compatibility. Superscripts in model.results and plot labels change for named components (from the class
abbreviation to the component path). Two new config keys, both with in-code fallbacks. No search, samples or
output-format change otherwise.

Before / after (a SLACS lens's 19-parameter source-light stage; more in the discussion):

Before: old labels

After: corner_anesthetic_lens.png

After: corner_anesthetic_latent_lens.png

Tests. test_model_mapper.py (component paths, names only, shared priors, abbreviation),
test_latent_parameters.py (new), test_nest_plotters.py (repeated labels, subset, latent), test_family_bases.py
(per-component / latent plot files, after-fit hook, single-component), test_updater.py (latent plot schedule),
test_latent.py (model latents in the latent output). Full suite: 3720 passed locally (macOS, Python 3.13).

Commits: 9db2921aa fix, 941bd9d1b labels + per-component + latent plots, 43dcf0f65 latent plot off by default,
587a75b32 names only, 7db4b9cf8 deep-path abbreviation, 5d9ba5167 latent naming / latent.csv / after-fit,
13bd430cf shared priors.

🤖 Generated with Claude Code

---

Maintainer changes (f9377f3, tracked in #1710)

The review findings are resolved on this branch rather than in another review round. Maintainer decisions (2026-10-21): (a) superscripts are the component's dot-joined path everywhere, including model.results; (b) the latent corner plot stays a separate plot, behind plots_search.yaml nest.corner_anesthetic_latent (default off).

API Changes

Corner plots label each parameter with its component path, can be split per top-level component, and can show the model's declared latent parameters (e.g. a shear's magnitude and angle). Models can declare latents with a latent_parameters class attribute; af.Periodic marks angles. On top of the contributor's work, this commit adds two config keys and one keyword, and changes three behaviours: per-component plots follow corner_anesthetic when not configured, deep labels are guaranteed unique, and shared priors are converted by every holder. latent.csv now keeps an angle's fixed range.
See full details below.

Test Plan

• pytest test_autofit: 3881 passed, 1 skipped, 1 xfailed (local, f9377f3 content)
• New Witness tests: (i) test_weighted_plot_results_components_follow_corner_anesthetic_when_key_absent, (ii) test__visualize__raising_latent_conversion_is_logged_not_raised, (iii) test__latent_parameters__shared_prior_is_converted_by_every_holder, (iv) test__superscripts__deep_paths_stay_unique / ..._in_a_model_stay_unique. All 13 new or changed tests fail on bf43aa6.
• Workspace smoke pyauto-heart smoke against this branch: 165/165 passed (autofit 10, autogalaxy 16, autolens 43, autolens_workspace_test 37, euclid 9, HowToLens 50). The autocti environments did not build locally (missing GSL headers); not run.
• Independent review: no code defects. The docstring finding is fixed; workspace config keys follow in phase 3.
• CI green on f9377f3

Full API Changes (for automation & release notes)

Added

• AbstractPriorModel.latent_parameters_from(parameter_array, weights=None, recentre=False) — model latents (component latent_parameters declarations) over samples; recentre=True recentres Periodic latents on their circular mean (plots only)
• af.Periodic(values, period) — marks a periodic latent (angle)
• AbstractPriorModel.component_paths, parameter_indices_by_component
• autofit.mapper.prior_model.abstract.superscripts_from_component_paths(paths), indices_by_top_level_component(paths)
• Config output.yaml: latent_model_parameters: true — model latents join latent.csv / latent.results / latent_summary.json
• Config visualize/plots_search.yaml nest: corner_anesthetic_components, corner_anesthetic_latent (default false), corner_anesthetic_latent_components
• Plots corner_anesthetic_<component>.png, corner_anesthetic_latent[_<component>].png

Changed Behaviour

• Superscripts (corner plots, model.results) are the dot-joined named-component path (lens.bulge, source.pix.reg), not the class's notation.yaml superscript: entry, which now applies only to priors directly on an unnamed Model
• Deep-path names are shortened consistently and only as far as they stay unique among siblings (lens.bulge_i.she / lens.bulge_o.she)
• When corner_anesthetic_components is absent from a config it follows corner_anesthetic; corner_anesthetic_latent_components follows corner_anesthetic_latent
• A latent conversion that raises is logged; the fit's output continues without the latent plots and columns
• Parameters with no named component group under "model", or "model_root" when a component is named model

Migration

• Workspaces: add the new plots_search.yaml / output.yaml keys (library defaults apply meanwhile); update notation.yaml superscript: comments (phase 3, after release)

Heart RED development-only override

Live maintainer authorization, 2026-10-21: "do 2, Override and proceed" (start_dev for #1710). This covers commit, push and PR work only. Merge needs a separate /prm with every required check green. No release. This branch makes no claim to fix Heart. Heart readiness: STALE ("release validation incomplete: no rehearsal for current source"). Monitoring feed at ship: Heart: RED — Release RED; monitoring RED · 14/100 · 464 unresolved (updated 23h ago)

Exact RED reasons

• PyAutoFit: CI failure — https://api.github.com/repos/PyAutoLabs/PyAutoFit/actions/runs/39111874604
• PyAutoFit: PyAutoFit: CI failure
• PyAutoFit/Tests: failure
• PyAutoFit/workflows/Tests: failure
• PyAutoHands/workflows/Python Version Matrix: failure
• PyAutoHeart/workflows/Workspace Smoke: failure
• PyAutoMind/workflows/Spawn Drift: failure
• PyAutoArray/Tests: red
• PyAutoArray/Tests: red
• PyAutoFit/Tests: red
• PyAutoFit/Tests: red
• PyAutoFit/Tests: red
• PyAutoFit/Tests: red
• autofit_workspace_test/Smoke Tests: red
• autogalaxy_workspace_test/Smoke Tests: red
• autolens_workspace_test/Smoke Tests: red
• CI wall-clock: 27 gates · slowest PyAutoBrain Nightly Release 49m · 9 hang events
• PyAutoFit: PyAutoFit: CI failure
• Workspace validation: NOT ready — 1414p / 1f / 133s @ cloud#39224774724
• euclid: red
• test_run: red
• PyAutoArray: red
• autolens_inference: red
• autolens_profiling: red
• Worktree drift: 5 orphan / 0 missing / 4 dirty
• euclid_strong_lens_modeling_pipeline: red

Generated by the PyAutoLabs agent workflow.

🤖 Generated with Claude Code

[Jammy2211]

Hi @samlange04, thank you, this is a really careful piece of work. The silent KeyError hidden behind "posterior estimate not yet sufficient" was nasty, and the before/after plots make the case on their own.

On your two questions. (a) Yes, change the superscripts everywhere, not plot-only; \gamma_1^{\rm lens.shear} is clearer than \gamma^{\rm ext} twice. (b) We want the latent corner plot as a separate plot, switched by config and written next to corner_anesthetic.png as part of PyAutoFit, which is what you have built. Keeping corner_anesthetic_latent in visualize/plots_search.yaml next to corner_anesthetic, default off, is right.

Rather than ask for another round, I will push the remaining changes onto your branch myself, as we did for PyAutoGalaxy#635, and merge. For the record they are: (1) per-component plots fall back to on when an older config lacks corner_anesthetic_components, even with corner_anesthetic: false (posterior_kind.py:134-146); (2) latent_parameters_from in plot_latent_results runs outside the plot-exception guard, so a failing conversion would raise at the end of a fit; (3) model latents reach latent/samples.csv with no switch, so I will add one to the output.yaml latent block; (4) three-character path truncation can collide (lens.bulge_inner.shear and lens.bulge_outer.shear both become lens.bul.she); (5) a power law's ell_comps linked into a multipole can lose the mass's axis ratio and angle under the last-holder rule, which needs a test; (6) recentred angles will go into plots only, not latent.csv; (7) notation.yaml comments noting the superscript: block now applies only to a bare unnamed Model. Your workspace and PyAutoGalaxy branches will follow once this merges and releases. Thanks again!

[Jammy2211]

@samlange04 I've pushed the review changes onto your branch as f9377f3, so there's no extra round for you. Summary:

• Per-component plots follow corner_anesthetic when a config doesn't mention corner_anesthetic_components. The latent per-component plots get their own key, corner_anesthetic_latent_components.
• A raising latent conversion is now logged rather than failing the final visualize, and the latent plots are skipped in test mode.
• output.yaml latent_model_parameters (default on) switches the declared latents in or out of latent.csv / latent.results / latent_summary.json.
• Deep labels are shortened consistently and only as far as they stay unique: lens.bulge_i.she / lens.bulge_o.she, and a basis and its profiles read lens.bul / lens.bul.0.
• Shared priors are seen by every holder's declarations, so a power law's ell_comps linked into its multipoles keeps axis_ratio/angle whichever component the column is named after.
• Periodic latents are recentred for the corner plots only. latent.csv keeps the conversion's own range, so the column means the same thing in every fit.
• Minor: the root group becomes model_root when a component is named model, and component paths are cached while the model is frozen.

The full suite passes locally (3881), and the workspace smoke tests pass against the branch (165/165). Once CI is green I'll merge. Your PyAutoGalaxy declarations and workspace config PRs can then follow the release, as you planned. Thanks for this!