fix: SLaM advanced-pipeline inversion + adapt-image cascade

[Jammy2211]

Overview

A parked cluster of advanced-pipeline SLaM fit failures (2026-07-20 sweep), reproduced and scoped on clean main 2026-07-21 under PYAUTO_TEST_MODE=2 PYAUTO_SMALL_DATASETS=1. This task covers the pixelized-SLaM inversion / adapt-image cascade only. The multi-wavelength SersicCore alpha=0 failure proved unrelated and was split into a separate issue/prompt (bug/autogalaxy/sersic_core_alpha_zero_division.md).

Diagnosis

Two distinct signatures within the pixelized-SLaM cluster:

1. double-Einstein-ring (imaging + group) — AttributeError: 'NoneType' object has no attribute 'array' at PyAutoArray/.../inversion/mappers/abstract.py:476 (self.adapt_data.array), inside pixel_signals_from → adaptive-regularization weights. Occurs in phase source_pix[1]_source_1 (imaging slam.py:367, group slam.py:476). Confirmed same root cause in both scripts.
2. imaging/features/pixelization/slam — TypeError: add got incompatible shapes for broadcasting: (786, 786), (210, 210) at PyAutoArray/.../inversion/inversion/abstract.py:366 (curvature_matrix + regularization_matrix). A curvature/regularization size mismatch — needs phase-1 diagnosis (real pipeline bug vs synthetic-samples/test-mode artifact).

Root cause (double-ring — confirmed real, not a test-mode artifact)

The SLaM setup uses regularization_init = al.reg.Adapt (adaptive regularization → needs a per-galaxy adapt image). source_pix_1_source_1 pixelizes source_1, but its stitched galaxy_name_image_dict carries only lens + source_0. source_0 obtained its adapt image from an earlier light-profile phase; source_1 was only ever a bare redshift galaxy, so no adapt image can exist for it yet → AdaptImages yields no entry for source_1 → the source_1 mapper's adapt_data is None → crash.

Plan

• Phase 1 — diagnose: confirm the double-ring root cause end-to-end; diagnose the pixelization/slam (786 vs 210) mismatch (same adapt/synthetic-samples gap, a real mesh-size mismatch, or test-mode-only?); determine whether any fix is library (PyAutoArray/PyAutoGalaxy) or workspace-only; reconcile the HowToLens target (the parent note's HowToLens features/ paths do not exist — HowToLens uses a chapter_N_* layout).
• Phase 2 — fix double-ring: seed source_1 an adapt image (from the current tracer's lensed source-plane model image) for its first pixelized fit, so adaptive regularization has data to weight on. Re-run both double-ring scripts green in test mode.
• Phase 3 — fix pixelization/slam per the phase-1 verdict; update the HowToLens equivalent.

Detailed implementation plan

Affected Repositories

• autolens_workspace (primary)
• HowToLens (pixelization tutorial equivalent — real path TBD in phase 1)
• PyAutoLens / PyAutoArray / PyAutoGalaxy (only if phase-1 finds a library cause)

Work Classification

Workspace (primary) — with a possible library follow-up gated on phase-1 diagnosis of the pixelization/slam shape mismatch.

Branch Survey

Repository	Current Branch	Dirty?
./autolens_workspace	main	clean
./HowToLens	main	clean
./PyAutoLens	main	1 (unrelated: paper_jax/paper.md)
./PyAutoGalaxy	main	clean
./PyAutoArray	main	clean

Suggested branch: feature/slam-adapt-inversion-cascade
Worktree root: ~/Code/PyAutoLabs-wt/slam-adapt-inversion-cascade/ (created later by /start_workspace)

Chosen fix direction (approved at scoping)

Seed source_1 an adapt image from the current tracer's lensed source model before its first pixelized fit. (Alternative rejected: bootstrap with Constant regularization then switch to Adapt.)

Reproduction

cd autolens_workspace
NUMBA_CACHE_DIR=/tmp/numba_cache MPLCONFIGDIR=/tmp/matplotlib \
PYAUTO_TEST_MODE=2 PYAUTO_SMALL_DATASETS=1 PYAUTO_SKIP_FIT_OUTPUT=1 PYAUTO_SKIP_VISUALIZATION=1 PYAUTO_SKIP_CHECKS=1 PYAUTO_FAST_PLOTS=1 \
python scripts/imaging/features/advanced/double_einstein_ring/slam.py

Key Files

• scripts/imaging/features/advanced/double_einstein_ring/slam.py — source_pix_1_source_1 (~line 286-367); stitched galaxy_name_image_dict lacks source_1.
• scripts/group/features/advanced/double_einstein_ring/slam.py — same phase (~line 476).
• scripts/imaging/features/pixelization/slam.py — curvature/regularization shape mismatch.
• PyAutoGalaxy/autogalaxy/analysis/adapt_images/adapt_images.py — AdaptImages galaxy→image mapping (line ~327-335), galaxy_name_image_dict_via_result_from.
• PyAutoArray/autoarray/inversion/mappers/abstract.py:476 — pixel_signals_from (adapt_data.array crash site).

Testing

Both double-ring scripts + pixelization/slam (+ HowToLens equivalent) must run green under the smoke env vars above. Regenerate notebooks after the workspace scripts pass.

Original Prompt

Click to expand starting prompt

Parked cluster of advanced-pipeline SLaM fit failures (2026-07-20 sweep). First step: reproduce imaging/features/advanced/double_einstein_ring/slam on clean main; confirm whether the failure is the synthetic-samples adapt_images gap (fixable in the SLaM pipeline / samples_summary construction) or downstream of the pixelization-inversion-not-PD cluster. Split multi-wavelength into a separate prompt if it proves unrelated (it did — SersicCore alpha=0 ZeroDivisionError, now bug/autogalaxy/sersic_core_alpha_zero_division.md).

Affected: autolens_workspace imaging/features/pixelization/slam, imaging/features/advanced/double_einstein_ring/slam, group/features/advanced/double_einstein_ring/slam; HowToLens pixelization tutorial equivalent.

[Jammy2211]

Phase-1 diagnosis: imaging/features/pixelization/slam mismatch (786 vs 210) — root cause found

The curvature/regularization shape mismatch is not the pix-inversion-not-PD cluster and not a test-mode artifact. It's a mesh + regularization incompatibility:

scripts/imaging/features/pixelization/slam.py:482-483 pairs:

mesh=af.Model(al.mesh.RectangularAdaptImage, shape=mesh_shape),
regularization=al.reg.AdaptSplit,

But AdaptSplit (and all *Split regularizations) are only valid on irregular Delaunay/Voronoi meshes — the sibling group script documents this explicitly:

• scripts/group/features/pixelization/slam.py:23 — "AdaptSplit, which is only appropriate for irregular meshes such as Delaunay or Voronoi"
• scripts/group/features/pixelization/slam.py:280 — "Rectangular meshes use Adapt (not AdaptSplit, which is reserved for irregular meshes)"

Mechanism

InterpolatorRectangular._mappings_sizes_weights_split (PyAutoArray/.../mesh/interpolator/rectangular.py:310) is a stub that returns the non-split mappings:

def _mappings_sizes_weights_split(self):
    # ... split regularization reuses the same mappings.
    return self._mappings_sizes_weights

AdaptSplit.regularization_matrix_from then computes pixels = int(len(splitted_mappings) / 4), assuming a genuine 4-point split-cross (which Delaunay/Voronoi produce). Fed the rectangular non-split mappings, it derives the wrong pixel count -> a 210x210 regularization matrix vs the 786x786 curvature matrix -> TypeError: add got incompatible shapes at inversion/inversion/abstract.py:366.

Fix options (phase 3)

1. Pair AdaptSplit with a Delaunay mesh — preserves the example's teaching intent ("demonstrate AdaptSplit"); mirrors imaging/features/pixelization/delaunay.py.
2. Swap regularization to Adapt on the existing rectangular mesh — matches slam_start_here.py; loses the "AdaptSplit" demonstration.

Recommend option 1 (keep the example's purpose).

Optional library follow-up (candidate bug/autoarray)

The incompatible combo currently fails with a cryptic broadcasting TypeError deep in the inversion. Per fail-loud / no-silent-guards, *Split regularizations should raise a clear error at composition/first-use when paired with a mesh whose interpolator provides no genuine split (or the rectangular stub should raise rather than return non-split mappings).

[Jammy2211]

Folded in: Delaunay NEEDS_FIX cleanup (from #301)

The parked 2026-04-10 Delaunay pixelization failures are the same cluster as this task and touch the same config/build/no_run.yaml, so #301 is being folded here (closed as a duplicate scope).

Verified on clean main (2026-07-21, build/smoke env TEST_MODE=2 SMALL_DATASETS=1 DISABLE_JAX=1):

• imaging/features/pixelization/delaunay.py -> exit 0 (~19s) — the FitException is gone
• interferometer/features/pixelization/delaunay.py -> exit 0 (~34s) — the (2,2) vs (1032,1032) broadcast is gone
• HowToLens chapter_4_pixelizations/tutorial_7_adaptive_pixelization.py -> exit 0

Both were fixed by the intervening interferometer inversion work (non-PD numpy path PyAutoLens#607, zero-fill extrapolate, potential-correction port). No PyAutoArray change needed.

Add to this task's checklist (do alongside the SLaM no_run edits, in the same PR set):

• autolens_workspace/config/build/no_run.yaml — remove imaging/features/pixelization/delaunay # NEEDS_FIX 2026-04-10 …
• autolens_workspace/config/build/no_run.yaml — remove interferometer/features/pixelization/delaunay # NEEDS_FIX 2026-04-10 …
• HowToLens/config/build/no_run.yaml — remove the two .../features/pixelization/delaunay entries (dead paths — HowToLens has no features/ layout; teaches delaunay via chapter_4/tutorial_7, already green)
• autolens_workspace/smoke_tests.txt — add imaging/features/pixelization/delaunay.py and interferometer/features/pixelization/delaunay.py for per-PR regression protection (interferometer ~34s, heaviest entry — acceptable)

[Jammy2211]

Fold-in request: mask_irregular no_run un-park (from draft/bug/autogalaxy/mask_irregular_silent_failure.md)

The imaging/data_preparation/manual/mask_irregular NEEDS_FIX marker (2026-04-10 "silent failure") is stale — the bug is already fixed. On clean main the script runs PASSED (14.0s) through the build's own execute_script runner (has_failures=False). Root cause was Convolver.from_gaussian API drift, since fixed by the PyAutoArray convolver consolidation (#360/#361); headless, no GUI dependency.

This task owns autolens_workspace + HowToLens, whose no_run.yaml files carry this marker. Please remove it as part of this task's no_run cleanup (its worktree isn't on disk right now, so it couldn't be folded in directly like the autogalaxy/HowToGalaxy half was under #140):

• autolens_workspace/config/build/no_run.yaml — remove - imaging/data_preparation/manual/mask_irregular # NEEDS_FIX ...
• HowToLens/config/build/no_run.yaml — orphaned marker (no such script in HowToLens); remove it too

Verified green, one-line removal per file, zero behavioural change beyond letting the (passing) script run.

[Jammy2211]

Shipped — 2026-07-21

PRs

• fix: SLaM advanced-pipeline inversion + adapt-image cascade (#300) #302
• chore: drop stale/dead NEEDS_FIX markers from no_run.yaml (#300) HowToLens#40

Summary

Both pixelized-SLaM root causes fixed and verified green under the CI smoke env: (1) the double-Einstein-ring adapt-image gap — seed source_1's adapt image from its SOURCE-LP light-profile model image (imaging + group); (2) pixelization/slam mesh/reg incompatibility — AdaptSplit → Adapt on the rectangular mesh. Stale NEEDS_FIX markers cleared across both repos (run_smoke.py: 9/9 + 6/6).

Session Notes

• Split out (separate prompts): the multi-wavelength SersicCore alpha=0 ZeroDivisionError (unrelated) → bug/autogalaxy; and the interferometer/features/pixelization/delaunay non-PD FitException at analysis.py:182 (folded-in note's "already green" claim was wrong — kept its marker, updated the reason) → bug/autolens.
• Rule confirmed: AdaptSplit/ConstantSplit are only valid on irregular Delaunay/Voronoi meshes; rectangular meshes use Adapt. The rectangular interpolator's _mappings_sizes_weights_split is a non-split stub — a candidate PyAutoArray hardening follow-up (fail loudly on incompatible mesh+reg instead of a cryptic broadcast error).
• Did not add these heavy SLaM scripts to smoke_tests.txt (curated-subset rule + CI cost); markers removed is sufficient.
• Heart was RED at ship on two pre-existing, unrelated library reasons (PyAutoFit behind; PyAutoLens paper_jax/paper.md dirty); human-authorized PR-open.

[Jammy2211]

Both PRs merged (autolens_workspace#302, HowToLens#40). Closing.

[Jammy2211]

Follow-up: the mask_irregular fold-in noted above wasn't actioned before this task shipped, so the two remaining markers are now handled standalone (both repos are free):

• autolens_workspace → PR chore: un-park mask_irregular no_run marker (bug already fixed) #305
• HowToLens → PR feat: update group scripts to two main lens galaxies #42

Both are one-line no_run removals of the verified-green mask_irregular marker. autogalaxy_workspace + HowToGalaxy were already done via #140. This completes the 4-imaging-repo un-park.

[Jammy2211]

Correction: PRs #305/#42 above were duplicates and are now closed — the autolens_workspace + HowToLens markers had already been removed by a concurrent session via #304 and HowToLens#41 (both merged). All 4 imaging repos are clean; nothing outstanding.

[Jammy2211]

Delaunay portion of the folded checklist completed in #307:

• autolens_workspace/config/build/no_run.yaml — imaging Delaunay entry removed (done in this PR/fix: SLaM advanced-pipeline inversion + adapt-image cascade #300)
• autolens_workspace/config/build/no_run.yaml — interferometer Delaunay entry removed (test(smoke): add imaging Delaunay to smoke gate; interferometer stays parked (flaky NaN) #307)
• HowToLens/config/build/no_run.yaml — two dead Delaunay entries removed (done in fix: SLaM advanced-pipeline inversion + adapt-image cascade #300)
• autolens_workspace/smoke_tests.txt — both Delaunay scripts added (test(smoke): add imaging Delaunay to smoke gate; interferometer stays parked (flaky NaN) #307)

On the interferometer entry #300 re-parked: the claimed residual FitException (analysis.py:182) is not reproducible on merged main — prior-median eval clean, 20/20 random prior draws clean, guide-section real inversion computes log_evidence. run_smoke.py = 11/11 passed (imaging delaunay 15.3s, interferometer 20.5s). Details in #307. If you have a config where it does fail (full non-test-mode / CI py3.13), flag on #307 before merge.