feat: oversampled PSF blurred images — operate/image consumer (phase 2c)

[Jammy2211]

Overview

Phase 2c of oversampled PSF convolution: the PyAutoGalaxy consumer switch, per §5 of the approved design (PyAutoMind/feature/autoarray/oversampling_design.md, approved in PyAutoLabs/PyAutoArray#353). When the dataset's PSF is oversampled, operate/image.py must evaluate images at the fine resolution before convolving — today the over_sample decorator bins immediately, so the Convolver only ever sees image-resolution arrays.

Depends on PyAutoArray phase 2a (#354, merged as #355). Independent of phase 2b (#356 / PR #357, inversion path).

Design refinement at implementation: instead of threading a binned= kwarg through every image_2d_from signature (Galaxy → Galaxies → light profiles, plus Tracer in PyAutoLens), the oversampled branch evaluates on aa.Grid2DIrregular(grid.over_sampled) — the decorator already passes irregular grids through unbinned, and grid.over_sampled is in per-pixel sub-block order, exactly the oversampled Convolver's input format. No signature changes anywhere; Tracer inherits the behaviour through OperateImage unchanged.

Plan

• blurred_image_2d_from: when psf.convolve_over_sample_size > 1, evaluate the image and blurring image on the over-sampled irregular grids and pass them to psf.convolved_image_from(..., mask=grid.mask); the LightProfileOperated branch is unchanged (operated images are added at image resolution, unblurred).
• blurred_image_2d_list_from and galaxy_blurred_image_2d_dict_from: same switch per galaxy image.
• padded_image_2d_from / unmasked_blurred_image_2d_from stay at image resolution (visualization-only; documented limitation per the design, revisited in the docs phase).
• Tests (numpy-only): s=1 strict parity; s=2 blurred image through operate/image.py equals the same computation done directly through the phase-2a Convolver API; list/dict variants consistent with the scalar path.

Original Prompt

Click to expand starting prompt

See PyAutoMind/issued/oversampling_phase_2c_autogalaxy_consumer.md (phase 2c split of oversampling_phase_2_core_api.md; design approved in PyAutoArray#353).

[Jammy2211]

Phase 2c implemented — at the ship sign-off checkpoint

Committed locally on feature/psf-oversample-galaxy (a4e6561c, not pushed). Diff: 2 files, +154/−21.

Verification

Leg	Result
Tests	PyAutoGalaxy 943 passed (2 new); PyAutoLens 335 against the branch (Tracer inherits OperateImage, so its suite exercises the changed base class)
Equivalence	s=2 through operate/image.py equals the direct phase-2a Convolver call to 1e-14; list/dict variants match the scalar path exactly
s=1 parity	Untouched code paths; existing 12 operate-image tests unchanged and green
Review	Verdict: CLEAN
Design note	Implemented via the Grid2DIrregular(grid.over_sampled) pass-through instead of threading a binned= kwarg — no signature changes anywhere, Tracer gets the behaviour for free (recorded as a design refinement on this issue's opening post)

Awaiting ship approval (supervised --auto): on approval the branch is pushed and the pending-release PR opens. This is the last code phase — after it, phase 3 (workspace tests + simulator example, including the autolens_workspace_test numerical tests deferred earlier today) and phase 4 (docs) remain, plus the filed refactor follow-up.

🤖 Generated with Claude Code

[Jammy2211]

➕ Scope addition shipped to PR #481 (user-directed): linear light profile support under an oversampled PSF. Their operated_mapping_matrix_override was convolving binned image-resolution evaluations — at s>1 it hit the phase-2a binned-input guard. It now evaluates each profile on the over-sampled coordinates and convolves at the fine resolution (same pattern as the consumer switch). Column-level test vs the direct Convolver to 1e-14; suites re-run green: PyAutoGalaxy 944, PyAutoLens 335.

Support matrix at convolve_over_sample_size > 1 after this PR: standard light profiles ✅ · linear light profiles ✅ · operated light profiles ✅ (unblurred by definition) · pixelized via mapping formalism ✅ (2b, merged) · sparse formalism / data_linear_func_matrix preload / adaptive over sampling ❌ raise loudly (deferred by design). Phase 3 workspace tests will cover all four supported surfaces plus the guard assertions (prompt updated).

🤖 Generated with Claude Code

[Jammy2211]

✅ Merged as PR #481 (squash). Phase 2c complete — all four model surfaces (light / linear-light / operated / pixelized-mapping) now support oversampled PSF convolution on main. Phase 3 (workspace tests + simulator) is next.

🤖 Generated with Claude Code