perf: cross-eval memo for MGE operated mapping matrices (numba CPU inversion, fixed-MGE campaigns)

[Jammy2211]

Overview

With fixed lens-light MGE parameters, every likelihood evaluation still rebuilds the identical ~60-Gaussian PSF-convolved operated mapping matrix (~0.56 s of a ~2.4 s euclid numba CPU eval) because the sampler constructs fresh linear-func objects per call. Per user direction the fix stays inside the numba inversion bit, with minimal source interference, and must engage only when the MGE is actually fixed.

Plan

• Override linear_func_operated_mapping_matrix_dict on InversionImagingSparseNumba only: per-inversion cached_property + a module-level cross-evaluation memo keyed by sha256 of the linear func's full pickled state (profiles + grids + PSF).
• Fixed profiles fingerprint identically → matrix reused; any free parameter changes the key → recompute exactly as before. Unpicklable objects fall back to the uncached parent path. Failure modes are misses, never stale hits. Entries are read-only copies, bounded (8); AUTOARRAY_NUMBA_OPERATED_MEMO=0 disables.
• The batched-convolution half of the original PyAutoMind prompt (PyAutoGalaxy linear/abstract.py) is deferred — out of the numba bit, and largely mooted for fixed-MGE campaigns by the memo.

Results (validated 2026-08-20, 4-core cloud container, Delaunay Hilbert-1250 fiducial)

• Euclid "MGE operated mapping matrix" step: 0.56 s → 0.004 s on memo hits.
• Euclid steady-state eval: 2.34 s → 1.34 s (on top of the fnnls solver speed-up perf: in-place Cholesky buffer + copy-free numba solves for fnnls_cholesky #452/perf: in-place Cholesky buffer + copy-free numba solves for fnnls_cholesky #453 — 4.92 s → 1.34 s, 3.7× in total today).
• Both autolens_profiling delaunay_numba pins pass (rtol 1e-6); test_autoarray/ 1034 passed (+8 new memo tests); same 3 pre-existing pynufft environment failures.

Original Prompt

Click to expand starting prompt

PyAutoMind active/numba_cpu_likelihood_mge_convolution_and_caching.md — "Numba CPU likelihood phase 1: MGE operated-matrix cross-eval memo (fixed-MGE campaigns)" (rescoped 2026-08-20 per user direction from the original batched-convolution + caching plan; full profiling context in PyAutoLabs/autolens_profiling#151).