[Jammy2211]

Summary

Phase 1 of #623. DelaunayNN was ~40x slower than Delaunay on the NumPy (use_jax=False) path because sibson_mappings_weights_from_tables ran _sibson_single_from_tables once per query in a Python list comprehension (~1.5 ms × ~11k queries per likelihood). This adds a parallel numba kernel (prange over queries, module-level in interpolator/_sibson_numba.py, lazily imported so numba stays optional and the on-disk cache hits across processes) that ports the per-query code with the same arithmetic order: cavity walk + tolerance, cavity overflow, Watson contributions, outside-mesh / on-vertex / on-edge rows, degenerate and neighbour-overflow → NaN, and the six outputs with identical dtypes and padding. The NumPy branch dispatches to it when numba is importable and general.yaml numba.use_numba is true (float64 inputs); otherwise the unchanged reference loop runs. JAX path untouched. Phase 2 (routing audit of the numba imaging/interferometer likelihood paths) follows on the same issue.

Aris SMA interferometer (5.5M visibilities, threads=1, steady median per likelihood; same-run comparison):

mesh	vertices	before	after	Sibson step
Delaunay	1405	0.43 s	0.41 s	–
DelaunayNN	1405	18.1 s	0.57 s (1.39× Delaunay)	8.46 s → 12.9 ms (11,116 queries)
Delaunay	5526	12.7 s	8.2 s	–
DelaunayNN	5526	56 s	12.1 s	63 ms (27,600 queries)

logL identical to the reference loop (14092443.357190056). All four cases route to InversionInterferometerSparseNumba (DelaunayNN nnz/source column 17.8 @1405, 1.0 @5526, gate 60) — measured only; routing is phase 2. Numba cache: first process 18.8 s compile, second process cache hit, first call 0.53 s.

API Changes

None — internal changes only. (New sibson.sibson_numba_kernel() builder and private helpers; no public signature or behaviour change beyond speed. Weights agree with the reference to ≤1e-14, not bit-for-bit, because the reference groups vertex ids with an unstable argsort.)
See full details below.

Test Plan

• New test_sibson_numba.py (12): kernel vs reference on random meshes (outside-hull, on-vertex, on-edge-midpoint queries), cocircular grid, genuine degenerate triangle, five small cavity/neighbour cap combos (overflow asserted), dispatch both ways — mappings/sizes/diagnostics identical, weights 1e-12
• Full test_autoarray/: 2026 passed, 4 xfailed (local, worktree)
• Emulated unittest-nojax leg (path finder blocks jax/jaxlib/…): interpolator dir 95 passed, 8 skipped, 3 xfailed
• autolens_workspace_test: misc/jax_assertions/delaunay_nn.py, delaunay_nn_caps.py, delaunay_walk.py, imaging/mesh_interpolator_source_recovery.py — all exit 0 (DelaunayNN recovery pearson 0.99935)
• Independent review (fresh-context Opus): no correctness bugs; two findings fixed (cross-process numba cache miss; availability probe now catches ImportError)

Full API Changes (for automation & release notes)

Added

• autoarray.inversion.mesh.interpolator.sibson.sibson_numba_kernel() — returns the compiled parallel Sibson kernel (imports _sibson_numba on first use)
• autoarray/inversion/mesh/interpolator/_sibson_numba.py — private module: _sibson_from_tables_parallel + pairwise-sum helpers

Changed Behaviour

• sibson_mappings_weights_from_tables(..., xp=np) — dispatches to the numba kernel when available/enabled; outputs unchanged up to ≤1e-14 weight rounding

Heart RED override (development only)

Authorised by the live human in this session on 2026-10-15 for #623: "I override, tests will be handled once the epic refactor elsewhere is sorted". Permits commit, push and this PR only — merge needs a separate human /prm with required checks green; this PR does not claim to fix Heart.

• Current pyauto-heart readiness --json: verdict stale (ts 2026-10-14T14:57:28Z) — "release validation incomplete: no rehearsal for current source".
• Current Heart feed RED reasons (verbatim): "Heart: RED — Release RED; monitoring RED · 10/100 · 415 unresolved (updated 4h ago)"; "[red] release validation FAILED (stage integrate)"; "[red] PyAutoHands/workflows/Python Version Matrix: failure"; "[red] PyAutoHeart/workflows/Release Integrate: failure"; "[red] PyAutoMind/workflows/Lifecycle Drift: failure"; "[red] PyAutoArray/Tests: red"; "[red] PyAutoArray/Tests: red"; "[red] autofit_workspace_test/Smoke Tests: red"; "[red] autogalaxy_workspace_test/Smoke Tests: red"; "[red] autolens_workspace_test/Smoke Tests: red"; "[red] CI wall-clock: 26 gates · slowest PyAutoBrain Nightly Release 63m · 5 hang events"; "[red] Release readiness: release validation FAILED (stage integrate)"; "[red] autofit_test: red"; "[red] Release validation: NOT release_ready — v2026.10.15.1.dev81701 profile=release (2026-10-15T07:04:34+00:00)"; "[red] validation_report/stages/integrate: fail"; "[red] PyAutoArray: red"; "[red] autolens_inference: red"; "[red] autolens_profiling: red"; "[red] Worktree drift: 5 orphan / 0 missing / 4 dirty"; "[red] euclid_strong_lens_modeling_pipeline: red". (Note: PyAutoArray main Tests at ccddfba is green on all three legs.)
• Branch gates passed: tests, emulated nojax, smoke scripts, independent review (above).

Generated by the PyAutoLabs agent workflow.

🤖 Generated with Claude Code