[Jammy2211]

Summary

Lever 2 of #267, stacked on #269 (lever 1): the CPU A/B and A100 rows for PyAutoArray's sparse log det of the split-regularization H (library PR: PyAutoLabs/PyAutoArray#554), the five-pin witness, and the lever 2 section of the phase-3 verdict note.

Verdict

Control = lever 1, feature = lever 2, so the difference isolates lever 2. On the production numba CPU path (HST Delaunay N=1500, route b, memo ON, 1 thread, RAL job 343353 run alone on the node): whole call 302.7 → 267.4 ms (1.13x); inversion.log_det_regularization_matrix_term 37.5 → 6.5 ms; every other row flat; 4/4 cell gates; witness W1–W5 PASS (route fires once, log det rel 1.2e-9, log evidence rel 2.3e-10 inside the 1e-9 pin, gates admit, 6.0x same-matrix). A100 (job 343354, fixed_light.py cell): all log_evidence_terms bit-identical; the A100's dense Cholesky of H is 1.13 ms, so the lever is CPU-only by measurement. Cumulative levers 1+2 vs merge base: 413.3 → 267.4 ms, 1.55x.

Scripts Changed

• hpc/batch_cpu/submit_breakdown_imaging_fixed_light_numba_levers_l2_delaunay_ral_hst_fp64 — new: two-arm Leg D run, control lever 1 / feature lever 2 on private PYTHONPATH checkouts, gpu partition CPUs-only, thread env pinned to 1; in-job proof the arms differ by exactly lever 2 (SPARSE_LOG_DET_* present only in the feature tree); witness in the feature arm.
• hpc/batch_gpu/submit_breakdown_imaging_fixed_light_numba_levers_l2_delaunay_a100_hst_fp64 — new: two-arm fixed_light.py Delaunay cell on the A100 (the delaunay.py cell has no log-det row), fp64, fresh compile/autotune cache per arm.
• scripts/imaging/likelihood_breakdown/fixed_light_numba_levers_l2_witness.py — new: Part A is lever 1's witness verbatim (same system); W1 route fires once, W2 sparse vs dense-forced log det with the H eigenvalue range, W3 figure_of_merit ≤ 1e-9, W4 both regime gates admit, W5 same-matrix timing.
• results/breakdown/imaging/ — 4 JSON + 4 PNG + witness JSON.
• results/notes/fixed_lens_light_levers_2026_09.md — lever 2 sections added (headline, decomposition, A100 row and CPU-only verdict, verdict); Scope, Provenance, Gate roll-up, Next and Artifacts updated; lever 3 pending.

Cross-reference

Library change: PyAutoLabs/PyAutoArray#554 (feature/fixed-light-numba-levers-l2 @ 0b17c292, stacked on #553). Merge order: #269 then this; #553 then PyAutoLabs/PyAutoArray#554.

Tracked by #267 (lever 2 of 3).

Heart gate

pyauto-heart readiness was RED (score 20) at ship time with two release-chain reasons unrelated to this branch — "install verification FAILED (testpypi; checks F)" and "release validation FAILED (stage integrate)" — acknowledged by the human on 2026-09-16 for PR-open only (levers 2 and 3, same reasons as lever 1). Merge stays a human act via /prm.

🤖 Generated with Claude Code

[Jammy2211]

Closed as superseded by #272, not merged — every commit on this branch lands on main through #272, which is stacked on it and was retargeted to main on 2026-09-16.

Why not merged in stack order: the lint workflow checks out PyAutoArray main, which since PyAutoArray #555 merged (91240e43) already carries all three levers. This branch's copy of scripts/misc/test/test_fixed_light_numba.py pins the pre-lever-3 library (test_curvature_reg_matrix_is_recomputed_more_than_once, test_route_d_np_reaches_route_bs_evidence_on_a_real_fit — patched() without factor), so its lint run can no longer go green on its own (run 35156146275: those 2 failures, 534 passed). The verdict in this PR's body stands as the lever's review record; the results, note sections and submit scripts are unchanged in #272. Tracked by #267.

[Jammy2211]

Correction to the note above: GitHub marked this PR merged automatically once its commits reached main through #272's merge commit cea41bb (the PR was never merged on its own — no separate merge commit, no separate CI pass). Same content, same outcome: every commit on this branch is on main.