fix: synthetic positions fallback in test mode and PYAUTO_SMALL_DATASETS

[Jammy2211]

Overview

When PYAUTO_TEST_MODE=2 is active the random/garbage mass model can produce empty or NaN image-plane multiple-image positions, causing positions_likelihood_from to crash inside np.nanmax with ValueError: zero-size array to reduction operation fmax. Separately, simulators currently work around PointSolver by popping PYAUTO_SMALL_DATASETS from the environment, calling solve, and restoring it — awkward and easy to forget. This task fixes both at the library layer with synthetic position fallbacks, then deletes the workspace workarounds.

Plan

• Add a test-mode guard in Result.positions_likelihood_from that swaps empty/NaN positions for [(1.0, 0.0), (-1.0, 0.0)] so threshold + PositionsLH still build cleanly.
• Add a PYAUTO_SMALL_DATASETS=1 short-circuit at the top of PointSolver.solve that returns [(1.0, 0.0), (0.0, 1.0)] directly, skipping the triangle-tiling solve.
• Add unit tests for both fallback paths in test_autolens/.
• Delete the os.environ.pop("PYAUTO_SMALL_DATASETS", None) / restore dance from scripts/group/simulator.py, scripts/group/features/multi_gaussian_expansion/simulator.py, and scripts/group/features/no_lens_light/simulator.py.
• Library PR ships first; workspace cleanup follows after merge.
• Smoke verify: PYAUTO_TEST_MODE=2 python scripts/imaging/features/pixelization/delaunay.py must run cleanly.

Detailed implementation plan

Affected Repositories

• Jammy2211/PyAutoLens (primary)
• Jammy2211/autolens_workspace (follow-up after library merges)

PyAutoArray is intentionally not changed — the fix lives one layer up so the underlying Grid2DIrregular.furthest_distances_to_other_coordinates primitive stays pure.

Work Classification

Both — library first, workspace after.

Branch Survey

Repository	Current Branch	Dirty?
./PyAutoLens	main	clean
./autolens_workspace	main	clean
./PyAutoArray	main	clean
./PyAutoConf	main	clean

Suggested branch: feature/positions-test-mode-fallback
Worktree root: ~/Code/PyAutoLabs-wt/positions-test-mode-fallback/ (created later by /start_library)

Implementation Steps

1. PyAutoLens/autolens/analysis/result.py — positions_likelihood_from

   • Import is_test_mode alongside the existing skip_checks import from autoconf.test_mode.
   • After the existing positions = aa.Grid2DIrregular(np.asarray(...)) normalization (~line 329) and before positions_threshold_from(...), add a test-mode guard:
     • Compute the raw array.
     • If arr.shape[0] < 2 OR np.isnan(arr).any() OR np.isinf(arr).any(), replace positions with aa.Grid2DIrregular(values=[(1.0, 0.0), (-1.0, 0.0)]) and emit a single logger.warning(...).
   • Outside test mode the behaviour is unchanged — bad positions still surface as the original error so production fits aren't silently masked.

2. PyAutoLens/autolens/point/solver/point_solver.py — PointSolver.solve

   • Add import os at module level.
   • At the top of solve, before the super().solve_triangles(...) call:

     if os.environ.get(\"PYAUTO_SMALL_DATASETS\") == \"1\":
         return aa.Grid2DIrregular(values=[(1.0, 0.0), (0.0, 1.0)])

   • Plain numpy Grid2DIrregular is fine — PYAUTO_SMALL_DATASETS is smoke-test land, never inside jax.jit.

3. PyAutoLens unit tests

   • test_autolens/analysis/test_result.py: monkeypatch os.environ[\"PYAUTO_TEST_MODE\"] = \"2\", build a Result whose image_plane_multiple_image_positions returns an empty / NaN grid, call positions_likelihood_from, assert returned PositionsLH.positions equals [(1.0, 0.0), (-1.0, 0.0)].
   • test_autolens/point/solver/test_point_solver.py: with PYAUTO_SMALL_DATASETS=1, call PointSolver(...).solve(tracer=<any>, source_plane_coordinate=(0.0, 0.0)), assert result is exactly [(1.0, 0.0), (0.0, 1.0)] and the underlying triangle solver was never invoked.

4. autolens_workspace cleanup (after library PR merges)

   • scripts/group/simulator.py lines 294–309: remove import os, small_datasets = os.environ.pop(...), and the if small_datasets is not None: restore block.
   • scripts/group/features/multi_gaussian_expansion/simulator.py lines ~228–245: same cleanup.
   • scripts/group/features/no_lens_light/simulator.py lines ~257–273: same cleanup.
   • PointSolver then returns the fallback positions automatically when PYAUTO_SMALL_DATASETS=1.

5. Smoke verification before pushing

   • cd autolens_workspace && PYAUTO_TEST_MODE=2 python scripts/imaging/features/pixelization/delaunay.py — must not raise the zero-size array error.
   • Standard pytest test_autolens/analysis/ and pytest test_autolens/point/ via /ship_library.

Key Files

• PyAutoLens/autolens/analysis/result.py — empty/NaN guard in positions_likelihood_from
• PyAutoLens/autolens/point/solver/point_solver.py — PYAUTO_SMALL_DATASETS short-circuit in solve
• test_autolens/analysis/test_result.py — fallback unit test
• test_autolens/point/solver/test_point_solver.py — solver short-circuit unit test
• autolens_workspace/scripts/group/simulator.py
• autolens_workspace/scripts/group/features/multi_gaussian_expansion/simulator.py
• autolens_workspace/scripts/group/features/no_lens_light/simulator.py

Decisions / Trade-offs

• Fix at the positions_likelihood_from layer (not in Grid2DIrregular.furthest_distances_to_other_coordinates) so the autoarray primitive stays pure and production fits still surface bad positions loudly.
• Hard-coded fallback positions match the values from the prompt; they're geometrically separated enough for downstream threshold/likelihood code to behave normally.
• Short-circuit PointSolver.solve (the common entry point), not solve_triangles, to keep the fast path obvious.

Original Prompt

Click to expand starting prompt

The following error is caused when in PYAUTO_TEST_MODE=2 the mass model is some random awful garbage that does
not creatr positions that are good or physical or reliable for the lens model:

((PyAuto) ) jammy@DESKTOP-H143S82:~/Code/PyAutoLabs/autolens_workspace$ PYAUTO_TEST_MODE=2 python scripts/imaging/features/pixelization/delaunay.py

Traceback (most recent call last):
File "/home/jammy/Code/PyAutoLabs/autolens_workspace/scripts/imaging/features/pixelization/delaunay.py", line 983, in
source_pix_result_1 = source_pix_1(
^^^^^^^^^^^^^
File "/home/jammy/Code/PyAutoLabs/autolens_workspace/scripts/imaging/features/pixelization/delaunay.py", line 655, in source_pix_1
source_lp_result.positions_likelihood_from(
File "/home/jammy/Code/PyAutoLabs/PyAutoLens/autolens/analysis/result.py", line 333, in positions_likelihood_from
threshold = self.positions_threshold_from(
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/jammy/Code/PyAutoLabs/PyAutoLens/autolens/analysis/result.py", line 254, in positions_threshold_from
threshold = factor * np.nanmax(positions_fits.max_separation_of_plane_positions)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/jammy/Code/PyAutoLabs/PyAutoLens/autolens/point/max_separation.py", line 87, in max_separation_of_plane_positions
return max(self.furthest_separations_of_plane_positions)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/jammy/Code/PyAutoLabs/PyAutoLens/autolens/point/max_separation.py", line 83, in furthest_separations_of_plane_positions
return self.plane_positions.furthest_distances_to_other_coordinates
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/jammy/Code/PyAutoLabs/PyAutoArray/autoarray/structures/grids/irregular_2d.py", line 250, in furthest_distances_to_other_coordinates
furthest = self._xp.sqrt(self._xp.nanmax(sq_dists, axis=1))
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/jammy/venv/PyAuto/lib/python3.12/site-packages/numpy/lib/_nanfunctions_impl.py", line 486, in nanmax
res = np.fmax.reduce(a, axis=axis, out=out, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
ValueError: zero-size array to reduction operation fmax which has no identity

Can you make it so that a check is performed to see if the positions computed in positions_likelihood_from are
sensible or exist or not nans (check what they are when this script runs) and put in a PYAUTO_TEST_MODE fix
which simply converts the positions to (1.0, 0.0) and (-1.0, 0.0), the rest of the code should then run functionally
as expected.

can you combine this solution with another relatrd issue I drafted:

This is in group/simulator.py and annoying for users:

small_datasets = os.environ.pop("PYAUTO_SMALL_DATASETS", None)

solver = al.PointSolver.for_grid(
grid=al.Grid2D.uniform(shape_native=(500, 500), pixel_scales=0.1),
pixel_scale_precision=0.001,
magnification_threshold=0.01,
)

positions = solver.solve(
tracer=tracer, source_plane_coordinate=source_galaxy.bulge.centre
)

if small_datasets is not None:
os.environ["PYAUTO_SMALL_DATASETS"] = small_datasets

Can you make it so if PYAUTO_SMALL_DATASETS=1, the PointSOlver knows to just return two positions at some
small but convenient values (e.g. (1.0, 0.0), (0.0, 1.0)), so that testing doesnt break.

[Jammy2211]

Library PR Created — 2026-04-28

PR

• fix: synthetic positions fallback in test mode and PYAUTO_SMALL_DATASETS #479

Status

Workspace changes needed. Next: /start_workspace

API Changes

Two test-mode-only behavior additions, no public signature changes:

• Result.positions_likelihood_from substitutes [(1.0, 0.0), (-1.0, 0.0)] when PYAUTO_TEST_MODE is set and resolved positions are empty/NaN/inf.
• PointSolver.solve short-circuits to [(1.0, 0.0), (0.0, 1.0)] when PYAUTO_SMALL_DATASETS=1.

Workspace follow-up: delete os.environ.pop("PYAUTO_SMALL_DATASETS") workaround in scripts/group/simulator.py, scripts/group/features/multi_gaussian_expansion/simulator.py, scripts/group/features/no_lens_light/simulator.py.

[Jammy2211]

Smoke Test Results — 2026-04-29

All 7 smoke tests passed across autolens_workspace.

Workspace	Passed	Failed	Total
autolens_workspace	7	0	7

[Jammy2211]

Shipped — 2026-04-29

PRs

• fix: synthetic positions fallback in test mode and PYAUTO_SMALL_DATASETS #479 (library)
• fix: drop PYAUTO_SMALL_DATASETS workaround in group simulators autolens_workspace#102 (workspace)

Summary

Result.positions_likelihood_from now substitutes [(1.0, 0.0), (-1.0, 0.0)] when PYAUTO_TEST_MODE is active and resolved positions are empty/NaN/inf, so test-mode fits with random mass models no longer crash inside np.nanmax. PointSolver.solve short-circuits to [(1.0, 0.0), (0.0, 1.0)] when PYAUTO_SMALL_DATASETS=1, letting the three group simulator scripts drop their env-var pop/restore workaround.

Session Notes

• Both fallbacks were intentionally placed at the higher (positions_likelihood_from and solve) layer rather than in autoarray's Grid2DIrregular.furthest_distances_to_other_coordinates primitive, so the autoarray API stays pure and production fits still surface bad positions loudly.
• During smoke verification a separate downstream bug surfaced in PyAutoArray/autoarray/inversion/mesh/border_relocator.py:450 ('NoneType' object has no attribute 'array' during inversion mesh build under PYAUTO_TEST_MODE=2). It's pre-existing, was masked by the original positions crash, and is worth a separate ticket — out of scope for this PR.
• The committed dataset/group/simple/positions.json was regenerated with the synthetic pair as a side effect of the smoke run; discarded that artifact during cleanup so the observational dataset stays correct on main.