fix: restore profile validation search compatibility

[Jammy2211]

Overview

PyAutoGalaxy's new profile-constructor guards correctly reject unphysical parameters, but raw ValueError now escapes model fitting where invalid candidates must be resampled. The same validation also exposes results tutorials and regression fixtures that construct profile instances from marginal error vectors or deliberately invalid all-ones samples. This task restores release/workspace compatibility without changing prior support.

Heart's exact RED reason is release validation FAILED (stage integrate) in PyAutoHeart Actions run 31354307923. The related workspace validation reports 45 failures.

Plan

• Preserve direct constructor validation while making invalid fit candidates rejectable by PyAutoFit.
• Keep existing priors and scientific parameter support unchanged.
• Avoid constructing physical profiles from marginal/error vectors that are not guaranteed jointly physical.
• Repair deliberately invalid fixtures and move intentional NaN injection downstream of constructor validation.
• Re-run the failed script surface, workspace validation, release integration, and Heart readiness.

Detailed implementation plan

Affected Repositories

• PyAutoGalaxy (primary)
• autogalaxy_workspace
• autolens_workspace
• autogalaxy_workspace_test
• autolens_workspace_test

Branch Survey

Repository	Current Branch	Dirty?
PyAutoGalaxy	main	clean
autogalaxy_workspace	main	clean
autolens_workspace	main	clean
autogalaxy_workspace_test	main	clean
autolens_workspace_test	main	clean

No active task claims these repositories. A clean unregistered autolens workspace worktree exists on feature/point-source-campaign-evidence-tail; it is an acknowledged non-blocking overlap warning.

Suggested branch: feature/profile-validation-resample-recovery

Work Classification: Both — library first, then workspaces.

Worktree root: ~/Code/PyAutoLabs-wt/profile-validation-resample-recovery/

Implementation Steps

1. In PyAutoGalaxy/autogalaxy/exc.py, introduce a narrowly scoped profile-parameter exception inheriting from both ValueError and autofit.exc.FitException.
2. In PyAutoGalaxy/autogalaxy/profiles/validate.py, raise that exception directly for ell_comps and pass it as exc_type to the shared AutoArray scalar validators. Do not change broad ProfileException behavior or catch arbitrary ValueError in PyAutoFit.
3. Extend PyAutoGalaxy/test_autogalaxy/profiles/test_validate.py to prove direct calls remain ValueError, fitting recognizes FitException, valid values/JAX tracers still pass, and error messages remain actionable.
4. Update the results guides in autogalaxy_workspace and autolens_workspace so component-wise marginal/error vectors use list/dictionary output where they are not guaranteed to form a valid physical profile. Regenerate affected notebooks.
5. Replace invalid all-ones ellipticity components in the affected aggregator regression fixtures in autogalaxy_workspace_test and autolens_workspace_test, preserving unrelated sentinel values and assertions.
6. Redesign autolens_workspace_test/scripts/imaging/jax_likelihood/delaunay.py so its deliberate NaN reaches the downstream likelihood without asking profile constructors to accept NaN parameters.
7. Run focused PyAutoGalaxy tests and the full library suite; then run all 22 scripts that failed in release integration under their declared environments, verify generated notebook parity, dispatch workspace/release validation, and refresh Heart.

Key Files

• PyAutoGalaxy/autogalaxy/exc.py — fit-aware profile parameter exception.
• PyAutoGalaxy/autogalaxy/profiles/validate.py — constructor validation exception contract.
• PyAutoGalaxy/test_autogalaxy/profiles/test_validate.py — regression tests.
• autogalaxy_workspace/scripts/guides/results/ — marginal/error summary examples.
• autolens_workspace/scripts/guides/results/ — matching lensing examples.
• autogalaxy_workspace_test/scripts/misc/aggregator/ — physical fixture vectors.
• autolens_workspace_test/scripts/misc/aggregator/ — physical fixture vectors.
• autolens_workspace_test/scripts/imaging/jax_likelihood/delaunay.py — downstream NaN injection.

Out of Scope

• Joint unit-disk priors or axis-ratio/position-angle reparameterization.
• Global catching of ValueError in PyAutoFit.
• Narrowing component priors to an orientation-dependent square.
• Tenant-firewall drift unrelated to this regression.

Original Prompt

Click to expand starting prompt

Can we target making heart red

Should it not be the priors prevent these unphysical values throughout? That might be a big task but worth assessing

Just do 1 i agree 2 is too much

[Jammy2211]

Library PR Created

• PR: fix: resample invalid profile parameters #568
• Gate: 1,091 passed, 4 skipped
• API impact: invalid concrete profile parameters now raise a dual ValueError / FitException, allowing stochastic non-linear-search candidates to be rejected through PyAutoFit resampling.
• Workspace impact: migration is still needed for deterministic component-wise result summaries, aggregator sentinel vectors, and the intentional Delaunay NaN assertion identified in release run 31354307923.

Next: /start_workspace for the approved narrow compatibility edits. The joint ellipticity-prior redesign remains explicitly out of scope.

[Jammy2211]

Smoke Test Results — 2026-08-10

Workspace	Passed	Failed	Skipped	Total
autofit_workspace	8	2	0	10
autogalaxy_workspace	14	2	0	16
autolens_workspace	34	3	0	37
autolens_workspace_test	20	4	0	24
euclid_strong_lens_modeling_pipeline	6	0	0	6
HowToLens	50	0	0	50
Total	132	11	0	143

Smoke gate failed. Per the ship-workspace contract, no workspace PRs were opened and no code was changed in response to these failures.

Failures

autofit_workspace/notebooks/overview/overview_1_the_basics.ipynb

ModuleNotFoundError: No module named 'dill'

autofit_workspace/notebooks/searches/mcmc.ipynb

ModuleNotFoundError: No module named 'dill'

autogalaxy_workspace/notebooks/imaging/modeling.ipynb

ModuleNotFoundError: No module named 'dill'

autogalaxy_workspace/notebooks/interferometer/simulator.ipynb

ModuleNotFoundError: No module named 'dill'

autolens_workspace/scripts/multi_galaxy/features/pixelization/cpu_fast_modeling.py

Traceback (most recent call last):
  File "/home/jammy/Code/PyAutoLabs-wt/profile-validation-resample-recovery/autolens_workspace/scripts/multi_galaxy/features/pixelization/cpu_fast_modeling.py", line 43, in <module>
    import numba
ModuleNotFoundError: No module named 'numba'

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
  File "/home/jammy/Code/PyAutoLabs-wt/profile-validation-resample-recovery/autolens_workspace/scripts/multi_galaxy/features/pixelization/cpu_fast_modeling.py", line 45, in <module>
    input(
EOFError: EOF when reading a line

autolens_workspace/notebooks/imaging/modeling.ipynb

ModuleNotFoundError: No module named 'dill'

autolens_workspace/notebooks/interferometer/simulator.ipynb

ModuleNotFoundError: No module named 'dill'

autolens_workspace_test/scripts/interferometer/jax_likelihood/rectangular.py

ModuleNotFoundError:
--------------------
You are attempting to perform interferometer analysis with the default JAX-native `TransformerNUFFT`.

However, the optional library nufftax (https://github.com/GragasLab/nufftax) is not installed.

Install it via the command `pip install nufftax`.

If you want to use the legacy pynufft backend instead, pass `transformer_class=TransformerNUFFTPyNUFFT` and install pynufft.

----------------------

autolens_workspace_test/scripts/interferometer/jax_likelihood/mge.py

ModuleNotFoundError:
--------------------
You are attempting to perform interferometer analysis with the default JAX-native `TransformerNUFFT`.

However, the optional library nufftax (https://github.com/GragasLab/nufftax) is not installed.

Install it via the command `pip install nufftax`.

If you want to use the legacy pynufft backend instead, pass `transformer_class=TransformerNUFFTPyNUFFT` and install pynufft.

----------------------

autolens_workspace_test/scripts/interferometer/jax_likelihood/potential_correction.py

ModuleNotFoundError:
--------------------
You are attempting to perform interferometer analysis with the default JAX-native `TransformerNUFFT`.

However, the optional library nufftax (https://github.com/GragasLab/nufftax) is not installed.

Install it via the command `pip install nufftax`.

If you want to use the legacy pynufft backend instead, pass `transformer_class=TransformerNUFFTPyNUFFT` and install pynufft.

----------------------

autolens_workspace_test/scripts/interferometer/subhalo_recovery_interferometer.py

ModuleNotFoundError:
--------------------
You are attempting to perform interferometer analysis with the default JAX-native `TransformerNUFFT`.

However, the optional library nufftax (https://github.com/GragasLab/nufftax) is not installed.

Install it via the command `pip install nufftax`.

If you want to use the legacy pynufft backend instead, pass `transformer_class=TransformerNUFFTPyNUFFT` and install pynufft.

----------------------

[Jammy2211]

Smoke Test Results — 2026-08-10 (environment-repaired rerun)

Workspace	Passed	Failed	Skipped	Total
autofit_workspace	10	0	0	10
autogalaxy_workspace	16	0	0	16
autolens_workspace	37	0	0	37
autolens_workspace_test	24	0	0	24
euclid_strong_lens_modeling_pipeline	6	0	0	6
HowToLens	50	0	0	50
Total	143	0	0	143

All 143 smoke targets passed across six workspaces after installing the required validation dependencies into the task-local virtual environment.

[Jammy2211]

Shipped — 2026-08-10

PRs

• fix: resample invalid profile parameters #568
• docs: keep marginal summaries as parameter dictionaries autogalaxy_workspace#209
• docs: keep marginal summaries as parameter dictionaries autolens_workspace#483
• test: keep aggregator fixtures physically valid autogalaxy_workspace_test#104
• test: keep profile validation fixtures physical autolens_workspace_test#256

Summary

The approved narrow compatibility fix shipped: invalid profile construction is now both a direct ValueError and a PyAutoFit FitException, and the affected deterministic summaries and invalid fixtures were corrected. The targeted smoke suite passed 143/143, and the library-first merge gate was observed. The explicitly out-of-scope joint-prior redesign was not implemented.

Post-merge validation

Fresh release integration later reported 660 passed / 1 failed / 101 skipped / 0 timeouts, with install checks A–F all passing. Issue #567 was reopened for the remaining extra_galaxies/slam.py escape path; see the subsequent validation comment.

[Jammy2211]

Post-merge release validation — RED (2026-08-10)

Fresh validation run: https://github.com/PyAutoLabs/PyAutoHeart/actions/runs/31441556729

• Release-fidelity scripts: 660 passed, 1 failed, 101 skipped, 0 timeouts
• TestPyPI install verification: A–F all passed (including Python 3.12/3.13 and nufftax 0.6.1)
• Remaining failure: autolens_workspace/scripts/imaging/features/extra_galaxies/slam.py
• Exact rejection: ell_comps=(-0.9257683911051445, -0.5502774378326348), magnitude 1.0769644249264025

The new library guard is behaving correctly: it rejects an unphysical ellipse. The remaining defect is that this workspace script still has a prior path capable of drawing |ell_comps| >= 1. Issue reopened so the organism state does not falsely claim GREEN. No live release was performed.

[Jammy2211]

The release-validation follow-up is now implemented in PyAutoFit: PyAutoLabs/PyAutoFit#1463

It fixes the PYAUTO_TEST_MODE=1 lifecycle leak that reconstructed the PowerLaw point after its constructor guard had correctly rejected it with FitException. Full PyAutoFit validation passed (1707 passed, 2 skipped). The PR is a draft with pending-release; nothing has been merged or released.

[Jammy2211]

Release validation GREEN — 2026-08-11

Green run: https://github.com/PyAutoLabs/PyAutoHeart/actions/runs/31534325304 (dispatched 20:42:55Z on Heart main b7634e2c, completed 21:50:07Z)

53 jobs, 52 green, integrate / run_notebooks skipped by design. Artifacts written: release-stage-report, workspace-validation-report, verify-install-release.

What closed the escape path

The reopen (00:22Z) named the open question: is the ell_comps escape in PyAutoFit's post-likelihood/quick-update reconstruction, or does it need a narrower workspace prior? It was the reconstruction, in three places, plus one workspace guide. No prior was narrowed.

PR	Merged	Fix
PyAutoFit #1466	16:46Z	preserve guarded sample lifecycle
PyAutoFit #1468	17:46Z	invalidate stale factor-graph subsample instances on model rebind
PyAutoFit #1470	19:52Z	expose valid stored sample/instance pairs
autolens_workspace #484	19:59Z	stop the aggregator results guide rebuilding profiles from invalid historical samples

Evidence trail

The failing surface shrinks monotonically across four Release Integrate dispatches as the fixes land:

Run	Time	Result	Failing script packages
31456732688	03:52Z	fail	autofit/cookbooks, autolens/multi_dataset, autogalaxy/multi_dataset, autolens/guides
31517075410	17:20Z	fail	5 packages
31525583134	18:59Z	fail	1 — autolens/guides
31534325304	20:42Z	success	none

Attribution across runs is from job outcomes and merge timestamps, not job logs; the last leg is the firmest (the 18:59Z run's only failure was autolens/guides, #484 touches exactly that package, and the next dispatch was green).

Heart's verdict

Heart has not yet consumed this evidence, so its published board still shows the older tier — that is release_run.py's cached-artifact pattern working as designed, not a regression. The next dev-box tick reads the latest completed run of this channel, finds no sidecar for run id 31534325304, and self-ingests the release-stage-report; nothing needs to be run by hand. Tracked in PyAutoMind active.md as heart-green-validation-ingest, with the full record at complete/2026/08/heart-red-guarded-sample-escape.md.

Closing: the regression this issue tracks is fixed and proven green at release fidelity. No live release was performed.

---

Generated by Claude Code

[Jammy2211]

Closing: the evidence gap this tracked is resolved. The four fixes were already merged (see PyAutoMind complete/2026/08/heart-red-guarded-sample-escape.md); the outstanding validation ingest completed on the dev box on 2026-08-19 with fresh evidence covering current source — PyAutoHands rehearsal run 32210450811 (TestPyPI 2026.8.19.1.dev72301) + PyAutoHeart Release Integrate run 32264548345 (663 passed / 0 failed), ingested together: rehearse:pass, integrate:pass, release_ready: true, commit_shas matching all five library mains. Full record: PyAutoMind complete/2026/08/heart-green-validation-ingest.md.

🤖 Generated with Claude Code