Affected Repositories

• PyAutoArray (primary, phase 1)
• autolens_profiling (phase 2, re-measurement record)

Branch Survey

Suggested branch: feature/interferometer-numba-route-gate

Implementation Steps

1. autoarray/inversion/inversion/interferometer_numba/inversion_interferometer_numba_util.py — beside nnz_per_source_column_from, add route_cost_from(mapper, extent_shape, batch_size) returning (t_fft, t_direct_conv) estimates: t_fft = c_fft * S * ceil-batched extent FFT work (extent_y*extent_x*log2(extent_y*extent_x)), t_dc = c_dc * nnz_total * kernel_work (derive kernel work from the direct_conv implementation: per-nonzero it touches the (2ny,2nx) operator preload rows for the column's other non-zeros; confirm by reading curvature_direct_conv). Fit c_fft, c_dc by least squares to the six 512×512 points and the 125×125 crossovers; document the fit residuals in the docstring.
2. autoarray/inversion/inversion/factory.py _use_interferometer_numba (~L242–320) — replace the nnz_per_source_column > nnz_max comparison with: explicit user value set → old behaviour (honours autolens_profiling A/B scripts); default → cost-model comparison. The extent shape comes from dataset.sparse_operator / mask shape_native_masked_pixels; pass what the factory already has (it may need the dataset, not just the mapper — adjust the signature and its two callers/tests).
3. autoarray/settings.py — make interferometer_numba_nnz_per_source_max default to None meaning "cost model" in the packaged config, keep 0 as kill switch and a positive value as a hard gate; add interferometer_route: Optional[str] ("numba"/"fft") force override for benchmarking. Update the config docstrings and the config/ default yaml.
4. Tests: test_autoarray/inversion/inversion/test_factory.py — synthetic mappers with the 125×125 geometry (5,496 px, extent ~170², nnz/col 50 → numba, 80 → FFT for Delaunay; ~77 crossover rectangular) and the 512×512 geometry (91,507 px, 682² extent, S=1459/188 nnz → FFT, 2577/106 → numba, 5746/48 → numba); assert the chosen class. Keep test__inversion_imaging__delaunay_nn__numba_* and test: pin DelaunayNN routing to the numba likelihood paths (#623 phase 2) #625's routing pins green. Settings tests for the override semantics.
5. Witness: with the scratchpad aris_evidence/step8/bench.py --route on the 512 cache at s=8/6/4 and the 125×125 inputs via step7/bench.py, confirm the default picks the faster route at every vertex count; record the table in the PR body.
6. Phase 2 (workspace, start_workspace after the library PR merges): add the re-measurement record to autolens_profiling (the campaign's results directory, same format as docs: refactor docstrings for autoarray/mask package #226) and reference this issue.

Key Files

• autoarray/inversion/inversion/factory.py — _use_interferometer_numba
• autoarray/inversion/inversion/interferometer_numba/inversion_interferometer_numba_util.py — nnz_per_source_column_from, curvature_direct_conv
• autoarray/settings.py + autoarray/config/ — gate setting and defaults
• test_autoarray/inversion/inversion/test_factory.py — routing tests

Ordering

Runs after the sparse data-vector task (same repo claim; it also lowers memory for the re-measurement). DelaunayNN routing (#624/#625) merged 2026-10-15, so no factory.py conflict remains.

Evidence

PROGRESS4 steps 8.3a/b/b'/profile; measured s/call, 1 thread. Evidence copied to this session's scratchpad aris_evidence/ (temporary).

Interferometer numba route gate should scale with mask size

Type: feature
Target: PyAutoArray
Repos:

• PyAutoArray
• autolens_profiling
  Difficulty: large
  Autonomy: supervised
  Priority: normal
  Memory: wiki/galaxies/sources/massive-ellipticals.md; wiki/lensing/sources/dark-matter-substructure.md; wiki/galaxies/sources/cosmos-survey.md
  Status: formalised
  Consequence: glance
  Witness: on both the Aris 125x125 (5,496 px) and 512x512 (91,507 px) masks, the default routing picks the faster of numba direct_conv vs NumPy FFT curvature at every tested vertex count (1.4k-5.7k), asserted by a routing test
  Review-minutes: 3
  Unattended: ready

The factory (autoarray/inversion/inversion/factory.py ~L245) routes a NumPy single-mapper
interferometer inversion to the numba direct_conv curvature route only when mean non-zeros per
source column <= interferometer_numba_nnz_per_source_max (default 60, measured on one reference
geometry). On a 512x512 mask (91,507 masked pixels) the measured crossover is ~140-150 nnz/column:
at 2,577 vertices (106 nnz/col) the default sends numba to the FFT route at 44.1 s vs 34.2 s for
direct_conv (1.29x slower). Measured (s/call, 1 thread):

FFT cost scales ~linearly with source pixels (blocks of 128 FFTs over the 682x682 extent) and with
extent size; direct_conv scales with total non-zeros x kernel work. A fixed nnz/column constant
cannot be right for both small and large masks.

Do: replace the fixed gate with a cost model using quantities known at routing time (source pixels
S, total nnz, extent / FFT grid shape), calibrated on the existing measurements (old 125x125 mask
crossover ~60 Delaunay / ~77 rectangular, autolens_profiling#226; this 512x512 data), keeping a
user override. Add tests for the routing decision. Re-measure on both masks.

Evidence: /tmp/claude-1000/-home-jammy-Code-PyAutoLabs/50c8c6cd-6ac2-47c2-9290-7ee73f9c4315/scratchpad/aris/PROGRESS3.md, PROGRESS4.md, step7/, step8/. PyAutoArray ccddfba.