perf(numba): speed up the curvature matrix F at HST resolution (#505)

[Jammy2211]

Summary

Phase 1 of the numba CPU-path speed-up of the curvature matrix F at HST
resolution (#505). Two changes:

1. Drop the redundant passes in F assembly. The three blocks each rebuilt
   operated_mapping_matrix / noise_map ** 2 and re-walked the mapping matrix.
   They are now per-block helpers over one shared weighted copy, filling both
   triangles directly. Verified bit-identical (np.array_equal True on euclid
   and hst).
2. FFT the mapper x linear-func block. Its dense sliding window is a
   correlation with the PSF, i.e. a convolution with the PSF reversed along
   both axes. It now runs through the existing batched Convolver (new cached
   Convolver.reversed_kernel), once per linear func instead of once per
   (mapper, linear func) pair; only the sparse scatter onto source pixels stays
   in numba.

Measured at OMP_NUM_THREADS=1, memo off, n_repeats 10, with the before column
re-measured back-to-back with the after run (this host carries 20-30 %
session-to-session variance):

cell	eval	F total	mapper x mapper	mapper x l-func
hst, rectangular bilinear	1.562 -> 0.595	1.195 -> 0.359	0.295 -> 0.277	0.858 -> 0.054
euclid, rectangular bilinear	0.349 -> 0.249	0.256 -> 0.097	0.060 -> 0.065	0.182 -> 0.023
hst, Delaunay-1250	1.367 -> 0.758	1.077 -> 0.184	0.122 -> 0.123	0.854 -> 0.052
hst, rectangular RTU	(8.501) -> 7.479	(1.250) -> 0.334	(0.304) -> 0.249	(0.929) -> 0.061

~16x on the mapper x linear-func block, 3.3x on F, 2.6x on the whole HST
evaluation. The issue's 2x goal on F is met on both HST cells. RTU is shown in
parentheses because it is GPU-only by the 2026-08-28 decision and was re-run
once for currency rather than paired.

F is no longer the dominant term at HST resolution: the largest remaining steps
on hst bilinear are the mapper x mapper sparse-operator block (0.277 s) and the
MGE operated mapping matrix (~0.224 s). That is the phase-2 candidate; no
phase-2 prompt is filed by this task since the goal is met.

API Changes

Internal to the numba imaging-inversion path. One numba-only helper
(inversion_imaging_numba_util.curvature_matrix_mirrored_from) is removed
because F now fills both triangles directly; the identically-named functions in
inversion_util and inversion_imaging_util are untouched, and the only
downstream reference in the workspaces uses that untouched non-numba namespace.
One numba util is added for the scatter half of the split block, and Convolver
gains a cached reversed_kernel property. convolver.py is purely additive
(48 lines added, 0 removed). No user-facing class, signature or default changes.
See full details below.

Test Plan

• pytest test_autoarray/ in the task worktree: 1296 passed (1290 before;
  6 new). The new tests assert the FFT path against the retained dense kernel
  using asymmetric non-square PSFs, so a missing axis reversal fails.

• F agrees with the sliding-window result to 3e-18 relative on euclid
  and hst (the FFT is not bit-identical to the direct sum, as expected).

• All 3 pinned log-likelihoods PASSED: hst bilinear 27661.91013366411
  (pin 27661.910133665442), hst RTU 27180.70471569685
  (pin 27180.704715698186), hst Delaunay 29090.527210448134
  (pin 29090.527192092646). euclid has no pin; it measured
  6213.306873885871, unchanged to every recorded digit.

• Nautilus pool run — no thread oversubscription. A single-thread win
  only counts if it survives the one-process-per-core pool, and an FFT
  spinning up its own threads would show as a pool regression. Ran
  autolens_workspace's imaging/features/pixelization/cpu_fast_modeling.py
  (first, non-SLaM fit) on an 8-core host against canonical main and
  against this branch via PYTHONPATH (autoarray.__file__ asserted both
  ways), number_of_cores 2 -> 8, with a 60-component linear MGE bulge added
  so the FFT'd block is non-empty:

  | | pool of 8 (median of 3) | serial (1 core) |
  |---|---|---|
  | main | 0.1919 s/eval | 0.4845 s/eval |
  | branch | **0.1747** s/eval | **0.4489** s/eval |
  
  The pool improves 9 % and the single process 7 % — the pool gain tracks the
  single-thread gain rather than eroding it — and the parallel speed-up ratio
  is flat across the change (**2.52x -> 2.57x** on 8 cores), which is the
  number that would drop on hidden threads. Consistent with the code:
  `Convolver`'s FFT path uses `np.fft.rfft2` (no `workers=`, single-threaded)
  and `scipy.fft` appears only as `next_fast_len`, a shape calculation.
  

Profiling artifacts and the full method: autolens_profiling,
results/notes/numba_curvature_matrix_f_split.md.

Heart readiness — human RED override

pyauto-heart readiness graded RED score 45 at ship time
(ts 2026-08-28T15:02:11Z). The diff repairs none of the named reasons, so this
is a plain human override, not the corrective-PR exception. Reasons quoted
verbatim from pyauto-heart readiness --json:

red_reasons:
  "release validation FAILED (stage integrate)"
yellow_reasons:
  "workspace validation not passing (2 failed, cloud#33179766004: autolens_test scripts/imaging/rectangular_mge.py, autolens_test scripts/imaging/rectangular_mge_rtu.py)"
  "manifest drift: session-start hooks (generated) — 32 mismatch(es) vs PyAutoMind/repos.yaml"

Human authorisation, 2026-08-28, verbatim:

I authorize things to override the heart RED.

Scope: push + PR-open for this task only. Merge and release stay human, and
this override does not extend to any other task.

The RED is pre-existing on main and unrelated to this branch.
~/.pyauto-heart/validation_report.json (ts 2026-08-28T14:59:58Z):
validation_outcome: fail, release_ready: false, stages.integrate = fail
(run 33177898708), stages.rehearse = pass, totals 692 passed / 3 failed /
1 timeout. The failing scripts are autofit scripts/plot/nautilus_plotter.py,
autolens_test scripts/imaging/jax_likelihood/rectangular_mge.py,
.../rectangular_mge_rtu.py (the known f0ef8f2 pin drift) and, as the timeout,
autolens_test scripts/multi_dataset/jax_likelihood/delaunay.py. All live in
other repos and all are on the JAX likelihood path or in autofit; this
branch touches only the numba imaging-inversion path plus an additive-only
Convolver.

Full API Changes (for automation & release notes)

Removed

• autoarray.util.inversion_imaging_numba.curvature_matrix_mirrored_from(curvature_matrix) —
  no longer needed; the numba F assembly now writes both triangles as it goes.
  Not a removal of inversion_util.curvature_matrix_mirrored_from or
  inversion_imaging_util.curvature_matrix_mirrored_from, which are unchanged
  and still used by the interferometer and non-numba imaging paths.

Added

• autoarray.util.inversion_imaging_numba.curvature_matrix_off_diags_via_mapper_and_blurred_curvature_weights_from(data_to_pix_unique, data_weights, pix_lengths, pix_pixels, blurred_curvature_weights) —
  the sparse scatter half of the mapper x linear-func block, taking curvature
  weights that have already been correlated with the PSF. The dense
  curvature_matrix_off_diags_via_mapper_and_linear_func_curvature_vector_from
  is retained and is the reference the FFT path is asserted against in the tests.
• autoarray.operators.convolver.Convolver.reversed_kernel — cached property
  returning a Convolver built on the kernel reversed along both axes,
  inheriting this convolver's use_fft policy. Turns a PSF correlation into a
  convolution the batched FFT path can run.

Changed Behaviour

• The numba curvature matrix F is assembled from one shared
  operated_mapping_matrix / noise_map ** 2 copy via per-block helpers, and the
  mapper x linear-func block is computed by FFT convolution rather than a dense
  sliding window. F is bit-identical for change 1 and agrees to 3e-18 relative
  after change 2; pinned log-likelihoods are unchanged.

Migration

• None required. No public class, signature or default changed; the one removed
  symbol has no reference outside the module it was removed from.

Generated by the PyAutoLabs agent workflow.

[Jammy2211]

Linked workspace PR: PyAutoLabs/autolens_profiling#189 (profiling harness split + re-baselined breakdown artifacts + the pool-run notes).

Library-first merge gate: this PR (PyAutoArray#506) merges first; autolens_profiling#189 must not merge until it is MERGED, since the profiling repo's "after" columns were measured against this branch.