[Jammy2211]

Summary

With serial numba, the interferometer curvature matrix F = Aᵀ W̃ A can now be built by a new symmetric, row-restricted kernel, curvature_direct_conv_symmetric.

• Ranking. Source columns are ranked by their top extent row (ymin, stable sort, empty columns last).
• Row restriction. Each column convolves only the extent rows that later-ranked columns span, and only each row's mask x-range.
• Gather and mirror. A branch-free ranked-CSC sweep gathers the ranked upper triangle, then a mirror pass fills the rest.
• Exactness. Every computed entry is bit-identical to curvature_direct_conv.

Routing.

• The kernel runs when its geometry work fraction f is at or below interferometer_direct_conv_symmetric_max_work_fraction. The packaged default is 0.7.
• Mapping operators with non-finite weights fall back to the full kernel, so NaN evidence propagates as before.
• The parallel path is unchanged.
• The route cost model multiplies t_dc by min(1, κf) with κ = 1.32, so the factory routes on the kernel that will actually run. real_space_mask_from passes the mask, and only on NumPy.

Human decision (2026-10-16, on #635): rank by ascending ymin (not centroid) and raise the default gate from 0.5 to 0.7. The model says the kernel wins while κf < 1 (f < 0.76). Real Aris Delaunay columns are tall because of multiple imaging: the median span is 90 of 341 rows, against 15 on the synthetic geometry. That puts f at about 0.50–0.51, which the old 0.5 gate excluded.

Campaign: PyAutoLabs/PyAutoPulse#62 (task interferometer_large_mask_aris, lever L2).

Closes #635

API Changes

These are additive. The default behaviour changes only in speed: serial numba interferometer inversions with f ≤ 0.7 use the new kernel, with F identical to round-off.

• New Settings(interferometer_direct_conv_symmetric_max_work_fraction=...) and the general.yaml key of the same name (default 0.7; 0 disables; null means the default).
• New numba util functions and constants: curvature_direct_conv_symmetric, symmetric_restriction_inputs_from, symmetric_work_fraction_from, SYMMETRIC_KERNEL_KEYS, DIRECT_CONV_SYMMETRIC_WORK_FACTOR.
• curvature_route_seconds_from gains an optional work_fraction argument.

See full details below.

Witness (RAL, euclid-ral-compute-1, ral partition, --exclusive, 1 thread, load ≤ 1.0 during timing)

Software: PyAutoArray e59d5fa9 plus this diff (v3 patch), PyAutoNerves 7f3055b, PyAutoFit 9b394e73, PyAutoGalaxy 9f9ac947, PyAutoLens 4723668a; numba 0.65.1, numpy 2.2.6. Compile time is excluded: each figure is the minimum of two repeats with a warm numba cache.

Real Aris SPT2319-55 data (5,516,930 visibilities, 512×512 mask, extent 341², regular overlay Delaunay mesh, ConstantSplit). Jobs 412490, 412491 (FFT arm), 412492 and 412493 (control).

vertices	f (ymin)	full direct_conv	symmetric	speed-up	κf model	FFT route	log L / FoM full vs symmetric
1,459	0.511	22.43 s	15.08 s	1.49×	1.48×	15.32 s	bit-identical
2,577	0.504	22.83 s	14.95 s	1.53×	1.50×	not run	rel 3e-16 / 5e-16
5,746	0.495	24.22 s	15.35 s	1.58×	1.53×	not run	rel 2e-12 / 2e-10 (see below)

• Gate (≥ 1.3× on the curvature step on real data, matching log L): passed.
• 5,746 vertices. The log det moves by 0.0071 nats (rel 8e-8). This is round-off amplified by cond(F+H) = 7.2e13:
  • Both F matrices agree to 4.4e-16 relative.
  • The full kernel's own F, factorised from its lower triangle versus its upper triangle, already differs by 0.0077 nats in the log det (control job 412493).
• At 1,459 vertices the gain over the previous default route is about 1.02×. The old default sent this cell to FFT (15.32 s on this node), so most of the 1.49× there is against the full kernel. Two things follow:
  • The cost model still chooses direct_conv (correct, but marginally).
  • On RAL, FFT/full direct_conv is 0.68, against 0.89 predicted from laptop calibration. Refitting the FFT constant belongs to the next phase (FFT size levers) and is not part of this PR.
• At 2,577 and 5,746 vertices, direct_conv was already the route, so the speed-up applies in full.

Synthetic Aris-512 geometry (341² extent, 91,361 px, 3-NN, nnz 274,083), ymin order, job 412490:

S	f	full	symmetric	speed-up	κf model	A-A / C-C spread
1,428	0.414	22.08 s	12.08 s	1.83×	1.83×	1.008 / 1.007
2,540	0.408	22.53 s	12.11 s	1.86×	1.86×	1.001 / 1.002
5,712	0.403	23.94 s	12.59 s	1.90×	1.88×	1.001 / 1.004

The ranked triangle is bit-identical in all three, and the full-matrix max relative difference is ≤ 1.3e-15, which equals the full kernel's own asymmetry.

Heart

pyauto-heart readiness --json (canonical, 2026-10-17T11:48:54Z): STALE, score 85. red_reasons: []. The stale reason is "release validation incomplete: no rehearsal for current source". STALE passes for development shipping, so no RED override was needed.

Test Plan

• PyAutoArray full suite: 2086 passed, 4 xfailed. The new test_curvature_direct_conv_symmetric.py has 21 tests.
• PyAutoGalaxy 1357 passed; PyAutoLens 832 passed, 1 xfailed (against this branch).
• RAL real-data and synthetic witnesses (tables above).
• Independent review of v2: CLEAN (600-trial adversarial fuzz). v3 changes only the ordering, the default and the docstrings.
• Workspace smoke (pyauto-heart smoke, local library roots): 165/165 passed across autofit, autogalaxy, autolens, autolens_test, euclid and howtolens. The autocti and autocti_test environments failed to build before execution because GSL headers are missing on the laptop; that is unrelated to this change, which does not touch CTI.
• CI green

Full API Changes (for automation & release notes)

Added

• Settings(interferometer_direct_conv_symmetric_max_work_fraction: Optional[float] = None) and the property of the same name. Default 0.7 from general.yaml inversion.interferometer_direct_conv_symmetric_max_work_fraction; 0 always uses the full kernel; negative raises ValueError.
• inversion_interferometer_numba_util.curvature_direct_conv_symmetric(..., order, rank, ylo_suffix, xlo_row, xhi_row): the serial symmetric kernel.
• inversion_interferometer_numba_util.symmetric_restriction_inputs_from(iy, ix, cscptr, csc_row, ny, nx, pix_pixels): ordering, row restriction, x-ranges and work_fraction.
• inversion_interferometer_numba_util.symmetric_work_fraction_from(...): f from the mapper triplets, used by the router.
• SYMMETRIC_KERNEL_KEYS and DIRECT_CONV_SYMMETRIC_WORK_FACTOR = 1.32.
• InversionInterferometerSparseNumba.symmetric_restriction_inputs and .use_symmetric_kernel.
• interferometer.abstract.real_space_mask_from(dataset).

Changed

• curvature_route_seconds_from(source_pixels, nnz, extent_shape, work_fraction=None): t_dc is scaled by min(1, κ·work_fraction) when the argument is given.

Migration

• None. To restore the pre-perf: symmetric row-restricted direct_conv curvature kernel (serial numba, large extents) #635 kernel choice, set interferometer_direct_conv_symmetric_max_work_fraction: 0.

Generated by the PyAutoLabs agent workflow.

🤖 Generated with Claude Code

https://claude.ai/code/session_01QBeyBLN7ntDChx9v1fmt6M