Affected Repositories

• PyAutoLens (primary — the fix and its tests)
• PyAutoGalaxy (read-only — Galaxy.image_2d_from semantics are not changed)
• euclid_strong_lens_modeling_pipeline (workspace follow-up behind the library-first gate)

Work Classification

Both — library PyAutoLens first, workspace follow-up second.

Branch Survey

Surveyed via the GitHub API from a web session (no local worktree).

Remote branches on PyAutoLens: main, dev_Q, two archive/condemned/*; no existing feature branch for this task. No active.md worktree claim conflicts.

Suggested branch: feature/pixelized-source-magnification-latent

Worktree root: ~/Code/PyAutoLabs-wt/pixelized-source-magnification-latent/ (created later by /start_library; a web session works on the clone directly)

Implementation Steps

1. autolens/analysis/latent.py — add _pixelized_source_flux(fit, xp=np):
   • inv = fit.inversion; if it is None, or the source galaxy (fit.tracer.galaxies[-1]) carries no Pixelization, return 0.0 (the light-profile path already covers that case).
   • Iterate fit.tracer_to_inversion.mapper_galaxy_dict.items() (autolens/lens/to_inversion.py:394; fit.tracer_to_inversion at autolens/imaging/fit_imaging.py:153) and keep the mappers whose galaxy is the source galaxy. Caveat to verify: the converted tracer's galaxies are new objects — key the lookup on whichever galaxy object actually keys mapper_galaxy_dict (check against galaxy_image_dict's keys, fit_imaging.py:204); do not assume identity with fit.tracer_linear_light_profiles_to_light_profiles.galaxies[-1].
   • For each kept mapper, xp.sum(inv.reconstruction_dict[mapper] * mapper.mesh_geometry.areas_for_magnification) (reconstruction_dict at autoarray/inversion/inversion/abstract.py:667). A mesh geometry with no areas_for_magnification → xp.nan plus a once-per-process warning, following the _maybe_magzero_warn pattern already in the module.
   • Trace safety: reconstruction_dict values are xp arrays; Delaunay dual areas are traced on the JAX path after PyAutoArray#524; rectangular areas_transformed already is. No Python-level branching on traced values.
2. total_source_flux (latent.py:121-139) and total_source_flux_mujy (latent.py:189-215) — add the pixelized flux to the light-profile sum before the µJy conversion. magnification (latent.py:218-229) is unchanged in form. Update the three docstrings to state the new definition.
3. test_autolens/analysis/test_latent.py — extend the existing file (NumPy path only, per the repo's JAX rule):
   • FitImaging with a Delaunay pixelized source → magnification finite and equal to Σ galaxy_image_dict[source] / Σ reconstruction·areas_for_magnification.
   • Same with RectangularBilinearAdaptDensity.
   • Linear light profile + pixelization on the source galaxy → both contributions summed.
   • Light-profile-only fit → total_source_flux, total_source_flux_mujy, magnification numerically unchanged from main.
   • A KNN mesh → NaN (and one warning).
4. ship_library → PR-open. The PR body's API Changes section: changed behaviour — total_source_flux, total_source_flux_mujy and magnification now include pixelized sources (were 0 / inf). State that end-to-end Delaunay accuracy needs PyAutoArray#524 merged (it is, 2026-09-05, still pending release); the unit tests assert internal consistency and hold on either PyAutoArray.
5. Workspace follow-up (/start_workspace → ship_workspace, behind the library-first gate) in euclid_strong_lens_modeling_pipeline:
   • scripts/tools/diagnose_latent_vis_pix.py:42 docstring — magnification is not stage-invariant; state the new definition.
   • tests/test_compute_latent_variable.py (or test_latent_run_level.py) — assert the vis_pix Delaunay stage's magnification is present and finite when LatentEuclid.variables is evaluated directly on a test-mode Delaunay fit of the committed dataset/simulated/euclid_dr1_like (test-mode CI skips latents at every level, so run-level summaries can only be structural).
   • Ledger note for Mind phase 9 / Cortex phase 4: archived vis_pix rows carry a 0.0 sentinel and must be re-derived; catalogue/scripts/magnitudes.py:305 ingests the latent unchanged.

Key Files

• autolens/analysis/latent.py — the latent registry; total_source_flux, total_source_flux_mujy, magnification, LatentLens.variables (jit batch path)
• autolens/lens/to_inversion.py — TracerToInversion.mapper_galaxy_dict (mapper → galaxy)
• autolens/imaging/fit_imaging.py — tracer_to_inversion, galaxy_image_dict, inversion
• test_autolens/analysis/test_latent.py — existing latent tests to extend
• PyAutoGalaxy/autogalaxy/galaxy/galaxy.py:245-278 — Galaxy.image_2d_from zeros for a pixelization-only galaxy (read-only)
• PyAutoArray/autoarray/inversion/pixelization/mesh_geometry/delaunay.py — areas_for_magnification (dual areas since feat: weak-lensing fit class (FitWeak + plotter) — step 3 #524)

Notes for the implementer

• The Feature/Bug agents' heuristic sizing scores this "too-large" purely on prompt length and repo count; the declared size is small-medium and the design above is the human-approved plan — do not re-split into phases.
• This defect is independent of the Delaunay area defect (PyAutoArray#524): the latent never touched areas_for_magnification before, and neither bug masks the other.

PyAutoMind/draft/bug/autolens/magnification_latent_zero_for_pixelized_source.md

magnification latent is inf/0.0 for any pixelized source: Galaxy.image_2d_from returns zeros for a Pixelization-only galaxy

Type: bug
Target: autolens
Repos:

• PyAutoLens
• PyAutoGalaxy
• euclid_strong_lens_modeling_pipeline
  Themes:
• euclid
• pixelization
• latent
  Difficulty: small-medium
  Autonomy: supervised
  Priority: high
  Status: formalised
  Consequence: judge
  Review-minutes: 20
  Unattended: ready
  Epic: euclid-dr1-prep
  Parent: draft/feature/euclid/euclid_dr1_prep_epic.md
  Filed: 2026-09-04

The finding

autolens/analysis/latent.py::magnification (PyAutoLens/autolens/analysis/latent.py:218-229)
is total_lensed_source_flux_mujy / total_source_flux_mujy. The denominator,
total_source_flux (:121-139, via total_source_flux_mujy at :189-215), calls
tracer.galaxies[-1].image_2d_from(grid=fit.dataset.grids.lp) on
fit.tracer_linear_light_profiles_to_light_profiles. Galaxy.image_2d_from
(PyAutoGalaxy/autogalaxy/galaxy/galaxy.py:245-278) returns xp.zeros((grid.shape[0],))
at :278 whenever the galaxy has no non-linear LightProfile — and a Pixelization is
not a LightProfile. So for any pixelized source the source flux is a hard 0.0, then
ab_mag_via_flux_from(0.0, magzero=25) = inf and flux_mujy_via_ab_mag_from(inf) = 0.0
exactly, so magnification = lensed / 0.0 = +inf. PyAutoFit's latent writer
(autofit/non_linear/analysis/latent.py:258-263, the np.isfinite(all_values[:, i]).any()
guard) then drops the whole magnification column.

Confirmed empirically, not only by reading. A real al.FitImaging built on the
pipeline's shipped dataset/simulated/euclid_dr1_like (magzero 24.6) with the truth lens
from truth.json and a Delaunay source with zeroed_pixels=30 — the vis_pix /
full_model configuration — gives total_source_flux = 0.0,
total_source_flux_mujy = 0.0, magnification = inf. The Sersic control (the vis_lp
configuration) on the same fit gives total_source_flux = 0.8679858287,
total_source_flux_mujy = 0.4555254383 and magnification = 15.146299, reproducing
truth.json's 15.146298766165936 bit-for-bit — so the latent's definition is
sound and only the pixelized-source route is broken.

Production evidence: 9/9 archived initial_lens_model/vis_pix latent_summary.json
files record magnification = 0.0 alongside total_source_flux = 0.0 — a meaningless
sentinel that a downstream catalogue would silently ingest — while the matching
initial_lens_model/vis_lp results record real values (6.2727, 10.0782, 18.4297, …). The
vis_pix key set is otherwise complete (9 keys, same as vis_lp), so the NaN-drop
signature does not fire on those particular files; absence of the key is the signature
under current code and 0.0 under whatever version produced the archive. Either way the
Euclid pixelized stages have no magnification.

Note this defect is independent of the Delaunay area defect, now shipped as
complete/2026/09/delaunay-dual-area-magnification.md (PyAutoArray#524, PR #525): the latent
never touches areas_for_magnification at all, and neither bug masks the other.

Impact

Every Euclid pixelized stage — vis_pix and full_model — has no magnification.
catalogue/scripts/magnitudes.py maps latent.magnification into the catalogue at
:305, so the sentinel propagates. The docstring of
scripts/tools/diagnose_latent_vis_pix.py, which claims that the library latents not
depending on the source reconstruction (magnification among them) are "identical
between the two stages", is false — magnification depends on the source model,
which is exactly what changes at vis_pix, and the production numbers (6.27/10.08/18.43
vs 0.0/0.0/0.0) disprove it directly. That docstring must be corrected.

Proposed fix (for the implementer to verify, not to take on trust)

Preferred: for a pixelized source, define the source-plane flux as the reconstruction
integrated with the mesh's exact quadrature areas — the barycentric dual areas for
Delaunay (see the sibling prompt above), the transformed cell areas for rectangular
meshes (whose own correctness is an open lead in the cluster epic's
draft/test/workspaces/mesh_magnification_correctness.md) — so that magnification =
image-plane mapped flux / source-plane integrated flux.

Alternative, if that is judged out of scope for the latent layer: make the latent return
NaN explicitly for pixelized sources, and stop the 0.0 sentinel from reaching the
catalogue. What must not survive is a 0.0 that reads as a measurement.

Tests: a unit test that a pixelization-only galaxy's magnification latent is finite (or
explicitly NaN, per whichever route is chosen), and a Euclid pipeline test-mode assertion
on the vis_pix latent keys so a silently-dropped or sentinel column fails loudly.

Provenance

Proven by the euclid-dr1-prep phase 8 audit, PyAutoArray#522 (audit posted on the issue),
scripts part3_latent_reach.py, part3b_fit_latent.py.

Gate note

This directly affects Cortex phase 4's witness ("catalogue numerics match the 20260623
reference") for the vis_pix magnification column, and Mind phase 9's magnification
layer. Neither can be scored on the vis_pix magnification until this is fixed.

Implementation design (architect, 2026-09-04 — approved plan; execute from this)

Sequence: the PyAutoArray dual-area fix is merged (2026-09-05, PyAutoArray#524 / PR #525,
record complete/2026/09/delaunay-dual-area-magnification.md), so the Delaunay accuracy test may
be written against main in-worktree. /start_dev this prompt → task
pixelized-source-magnification-latent; classification both: library PyAutoLens first
(PyAutoGalaxy read-only), workspace follow-up in euclid_strong_lens_modeling_pipeline behind the
library-first gate.

Facts (do not re-derive): latents run inside a per-sample JAX jit (LatentLens.BATCH_MODE = "jit", autolens/analysis/latent.py:304-336, xp = analysis._xp) — new code must be trace-safe.
total_lensed_source_flux (:106-115) reads fit.galaxy_image_dict[fit.tracer.galaxies[-1]], which
already includes the inversion's mapped image — the numerator is sound. total_source_flux
(:121-139) / total_source_flux_mujy (:189-215) call
fit.tracer_linear_light_profiles_to_light_profiles.galaxies[-1].image_2d_from(...), zeros for a
pixelization-only galaxy (autogalaxy/galaxy/galaxy.py:278). Per-mapper reconstructions:
fit.inversion.reconstruction_dict (autoarray/inversion/inversion/abstract.py:667); mapper →
galaxy: fit.tracer_to_inversion.mapper_galaxy_dict (autolens/lens/to_inversion.py:394;
fit.tracer_to_inversion at fit_imaging.py:184). Caveat: the converted tracer's galaxies are
new objects — key lookups must use whichever galaxy object actually keys mapper_galaxy_dict
(verify against galaxy_image_dict's keys).

Library changes (PyAutoLens):

1. autolens/analysis/latent.py — add _pixelized_source_flux(fit, xp): inv = fit.inversion;
   None or no Pixelization on the source galaxy → 0.0; else for each (mapper, gal) in
   fit.tracer_to_inversion.mapper_galaxy_dict.items() belonging to the source galaxy,
   xp.sum(inv.reconstruction_dict[mapper] * mapper.mesh_geometry.areas_for_magnification). A mesh
   geometry with no areas_for_magnification (DelaunayNN, KNN, KNNBarycentric) → xp.nan with a
   one-time-per-process warning (no silent zero). total_source_flux and total_source_flux_mujy =
   light-profile flux (existing path) + pixelized flux; magnification unchanged in form.
   Trace-safe: reconstruction_dict values are xp arrays; Delaunay dual areas are xp after the
   autoarray fix; rectangular areas_transformed already is.
2. Tests in test_autolens/analysis/ (existing latent test file, else test_latent.py): FitImaging
   with a Delaunay pixelized source → magnification finite and equal to
   Σ galaxy_image[-1] / Σ reconstruction·areas; same with RectangularBilinearAdaptDensity; linear
   light profile + pixelization sums both; light-profile-only fit numerically unchanged; a KNN mesh →
   NaN. Numpy path (JAX only if the file already has a JAX pattern).
3. ship_library → PR-open. API Changes: Changed behaviour — total_source_flux,
   total_source_flux_mujy, magnification now include pixelized sources (were 0 / inf). The PR body
   states that end-to-end Delaunay accuracy needs the autoarray fix merged; these tests assert
   internal consistency and hold on either PyAutoArray.

Workspace follow-up (euclid_strong_lens_modeling_pipeline, /start_workspace → ship_workspace):
4. scripts/tools/diagnose_latent_vis_pix.py:42 docstring — magnification is NOT stage-invariant;
state the new definition.
5. tests/test_compute_latent_variable.py (or test_latent_run_level.py) — assert the vis_pix
Delaunay stage's magnification is present and finite when LatentEuclid.variables is evaluated
directly on a test-mode Delaunay fit of the committed dataset/simulated/euclid_dr1_like (a direct
call — test-mode CI skips latents at every level, so run-level summaries can only be structural).
6. Ledger note for Mind phase 9 / Cortex phase 4: archived vis_pix rows carry a 0.0 sentinel and
must be re-derived; catalogue/scripts/magnitudes.py:305 ingests the latent unchanged.

Acceptance (human, after both merges): the audit's part3b_fit_latent.py construction (Delaunay
fit on euclid_dr1_like, magzero 24.6) reports a finite magnification close to the Sersic
control's 15.146; record it on the issue.