research: fnnls memo after-measurement on the #332 ALMA cells (2026.10.17.1) #426

Overview

autolens_profiling#332's memo lever (levers.memo in results/breakdown/interferometer/{delaunay,pixelization}_numba_hpc_ral_cpu_fp64_levers.json, alma r3.5) measured the fnnls warm-start memo 2.17× slower (mean, Delaunay solve) on an iid evaluation stream, 1.08× on rect. PyAutoArray#613 was addressed by #615 (scattered-stream back-off, released 2026.10.7.1). This issue owes the after-measurement of that same lever on the released 2026.10.17.1 stack. It reports a ratio with its interval; it does not conclude "fixed".

Human-authorized compute 2026-10-17 ("Do all that, I authorise all"). Measurement only: no library edits. Mind prompt active/interferometer_nnls_memo_after_measurement.md; Pulse task tasks/interferometer_nnls_memo_scattered_stream_guard.md.

Plan

• Same cells (scripts/interferometer/{delaunay,rectangular}/likelihood_breakdown_numba.py), same dataset (alma, mask r3.5), same arm (numba, one thread) and same memo lever as research(interferometer): numba vs FFT CPU crossover + CPU lever list — campaign 3/3 phase 2 #332, at the five libraries' 2026.10.17.1 tags from a private RAL checkout.
• CPU --partition=ral (the research(interferometer): numba vs FFT CPU crossover + CPU lever list — campaign 3/3 phase 2 #332 job ran on euclid-ral-gpu-2 CPUs with no --gres; the host class therefore differs, and this is recorded beside the result).
• Six independent rounds (one process each) per cell, so the ratio carries the repo's paired whole-round bootstrap 90 % interval (round_bootstrap.round_median_ratio, ≥ 5 rounds per ab_verdict.MIN_AB_ROUNDS), median headline (fix: intervals for the memo-policy, scaling and logdet gates — timing-noise phase 3b (#362) #410–fix: median headline everywhere, wall basis and repeat-summary support — timing-noise phase 10 (#362) #412). The original mean ratio is reported beside it for comparison with the 2.17×.
• Two stream sizes: the exact research(interferometer): numba vs FFT CPU crossover + CPU lever list — campaign 3/3 phase 2 #332 flags (--n-instances 6, so 4 timed iid solves) and --n-instances 17 (16 iid instances, matching the walk).
• Tier: judge (or undeclared) — merge mode: human /prm

Detailed plan

1. Harness: _lever_memo (scripts/misc/likelihood_breakdown/interferometer_pixelized_numpy.py) clears the memo with nnls_memo.memo_clear() when the library has it (2026.10.17.1 does). It clears the #615 back-off state too. Clearing only _nnls_passive_set_memo, as it did for research(interferometer): numba vs FFT CPU crossover + CPU lever list — campaign 3/3 phase 2 #332, would carry the iid pass's back-off into the walk pass. The method used is recorded in the lever block.
2. Submit hpc/batch_cpu/submit_memo_after_interferometer_numba_ral_alma_fp64: one job, one node, rounds interleaved across cells and stream sizes; import guard on the private checkout; provenance header.
3. Summary script scripts/interferometer/delaunay/memo_after_summary.py reads the round JSONs and writes results/breakdown/interferometer/memo_after_hpc_ral_cpu_fp64_v2026.10.17.1.json: per cell × stream the median on/off ratio with its paired round interval, the mean ratio, the figure-of-merit Δ, the per-solve slow-seed counts and the host/load.
4. A journal entry in wiki/campaigns/interferometer_likelihood.md, then dashboard regen and ship_workspace (one PR with research: interferometer streaming vs in-memory scaling (phase 1, CPU) #368's phase 2). Judge CI once; no merge.

Branch: feature/interferometer-streaming-memo-after.
Claim note: autolens_profiling is also claimed by hst-mge-pilot-qualification (#418, shipped, awaiting merge, touches only scripts/misc/jax_compile/*.md). The files do not overlap.

Original Prompt

B. Pulse task tasks/interferometer_nnls_memo_scattered_stream_guard.md (PyAutoArray#613 fixed by #615, released 2026.10.7.1). Owed: re-run the autolens_profiling#332 ALMA Delaunay memo on/off measurement that showed 2.17× slowdown on scattered streams, now on 2026.10.17.1, same cell/dataset/host class as the original, CPU ral partition, quiet core, memo on vs off, iid vs walk streams. Report the ratio with its interval using the repo's timing-audit conventions (median headline, repeats as #410–#412 define). Never "fixed" as a conclusion — report the after-measurement.