perf: reuse Convolver state, cache operated-matrix dict, hoist linear-func pair loop (#496 phase 1)

[Jammy2211]

Summary

Companion PR: PyAutoLabs/PyAutoGalaxy#588

Phase 1 of the numba CPU sparse-operator likelihood speed restoration (#496, epic numba-cpu-likelihood). Three exact, likelihood-preserving changes on the CPU (use_jax=False) inversion path, plus one companion fix:

• Convolver.state_from reuses the precomputed state. The old shape test compared the mask to the kernel, so Imaging's psf_setup_state was never returned for convolve_over_sample_size == 1 and every convolution rebuilt ConvolverState (mask resize + rfft2). Now returns it iff ConvolverState.is_for_mask(mask).
• linear_func_operated_mapping_matrix_dict is a cached_property on AbstractInversionImaging (one build per inversion; the numba subclass already did this via its memo). Companion fix: the numba override reached the parent through .fget, which autonerves.CachedProperty does not have — now .func.
• Linear-func × linear-func curvature blocks (imaging/sparse.py, imaging_numba/sparse.py): noise-weight each operated matrix once instead of inside both loops, compute only the upper triangle of blocks and mirror the transpose. Exact for any per-pixel (non-constant, non-symmetric) noise map since block(i,j) = M_iᵀ diag(1/σ²) M_j.

Pairs with PyAutoGalaxy PR (batched MGE convolution) — both are independently mergeable; the measured gain needs both.

Measured (memo disabled, fresh FitImaging per call, OMP=1): MGE-60 operated matrix hst 1.11 s → 0.30 s, euclid 0.68 s → 0.125 s; matrices bitwise identical. Breakdown steps "Mapped reconstruction" and "Model data + chi²" ~3.9× faster, "Blurred image" ~1.9× (state reuse). hst pinned log-likelihood 27661.910133664103 unchanged before/after; euclid log-likelihood bit-identical before/after.

API Changes

None — internal changes only. ConvolverState gains source_mask and is_for_mask(); state_from returns the stored state in more cases (identical output).
See full details below.

Test Plan

• pytest test_autoarray — 1237 passed
• New: test__state_from__precomputed_state_reused_for_matching_mask, test__precomputed_state__convolution_bit_identical_to_no_state, test__convolved_mapping_matrix_via_real_space_np__matches_image_convolution_per_column (test_convolver.py); test_curvature_matrix_func_list_blocks.py (both sparse inversion classes, random non-symmetric noise map, 1e-12 vs brute force)
• autolens_profiling likelihood_breakdown/pixelization_numba.py hst pin PASSED (rtol 1e-6), euclid before/after bit-identical
• CI green on both PRs

Full API Changes (for automation & release notes)

Added

• autoarray.operators.convolver.ConvolverState.source_mask — the pre-resize mask the state was built from
• autoarray.operators.convolver.ConvolverState.is_for_mask(mask) — whether the state can be reused for mask

Changed Behaviour

• Convolver.state_from(mask) — returns the precomputed _state whenever is_for_mask(mask); previously only when the mask shape equalled the kernel shape. Output identical.
• AbstractInversionImaging.linear_func_operated_mapping_matrix_dict — cached_property (was property); value identical, built once per inversion.

Notes

• Two formatting-only hunks in convolver.py come from black on pre-existing drift.

Generated by the PyAutoLabs agent workflow.

🤖 Generated with Claude Code

https://claude.ai/code/session_01N6xyNMYmffpHodBrkc2d91