[Jammy2211]

Summary

Measurement leg of PyAutoArray#498 (numba CPU likelihood epic, phase 3a — the cross-evaluation NNLS warm-start memo, library PR: PyAutoLabs/PyAutoArray#501). Adds the per-solve active-set iteration diagnostic, the 11-variant / 32-cell lens-model robustness matrix, the fallback-guard re-check, the step-0 post-#453/#497 baseline re-profile, and the two notes that record the default-on decision.

Headline numbers (Delaunay-1250 + MGE-60 fiducial, AUTOARRAY_NUMBA_OPERATED_MEMO=0, OMP_NUM_THREADS=1): random-walk median active-set iterations 70→7 (euclid, 9.9×), 32→8 (hst, 4.0×); solve 0.54→0.05 s / 0.25→0.07 s; log-likelihood parity ≤ 4e-14 in all 32 matrix cells; zero seeded-solve failures; median solve time never worse than memo-off by >2%. Step-0 baseline: the solve is now 40% (euclid) / 17% (hst) of an evaluation, not the ~72% the issue was filed on.

Scripts Changed

• scripts/imaging/likelihood_breakdown/delaunay_numba_nnls_iterations.py — NEW: random-walk + i.i.d. instance sequences, memo off/on, fnnls_cholesky stats captured from inside FitImaging; --model <variant>, --n-instances; counts fallbacks; JSON + PNG per cell
• scripts/imaging/likelihood_breakdown/nnls_iterations_matrix.py — NEW: aggregates every cell JSON into results/notes/nnls_warm_start_memo_matrix.md
• scripts/misc/likelihood_breakdown/README.md — scripts table rows for both numba cells (delaunay_numba.py was missing)
• README.md, scripts/misc/likelihood_breakdown/README.md, scripts/misc/simulators/README.md — regenerated timing tables (build_readme.py, a lint.yml gate): the step-0 re-profile and the euclid simulator summary changed the numbers they read

Results

• results/breakdown/imaging/delaunay_numba_nnls_iterations_{euclid,hst}[_<model>]_v2026.8.17.1.{json,png} — 16 cells
• results/breakdown/imaging/delaunay_numba_breakdown_{euclid,hst}_v2026.8.17.1.{json,png} — step-0 baseline re-profile (overwritten in place, same version string)
• results/simulators/imaging_euclid_summary_v2026.8.17.1.* — side effect: the euclid dataset was absent locally and auto_simulate_if_missing regenerated it; timings only, pinned log-likelihood unchanged
• results/notes/nnls_warm_start_memo.md, results/notes/nnls_warm_start_memo_matrix.md; pointers in results/README.md

Test Plan

• ruff check + ruff format --check on both scripts
• AUTOLENS_PROFILING_SMOKE=1 python <script> exits 0 for both
• ruff check . / ruff format --check . / build_readme.py --check / check_submits.py --check clean repo-wide; the six lint.yml smoke scripts exit 0
• pinned log-likelihoods PASS in every run; memo on/off parity ≤ 4e-14
• CI lint smoke green

Pre-existing red, not from this branch: lint.yml's pytest scripts/misc/test/ step fails test_hazards_prior_exit.py::test__records_the_clipper_as_what_blocks_it (the ClipperPriorBox anchor is not found in finding.blocked_by). It reproduces identically on canonical main (451b85f) with this branch nowhere in the checkout, and this branch touches no file under scripts/misc/hazards/ or scripts/misc/test/. Needs its own task.

Merges after PyAutoArray#498's PR (library-first gate). Generated by the PyAutoLabs agent workflow.