[Jammy2211]

Summary

Profiling evidence for point-source CPU campaign phase 2 (PyAutoArray#568). An interleaved in-process A/B cell compares the JAX PointSolver with the jnp.unique vertex dedup (control) against the library change that drops it (no-dedup). This PR adds the RAL CPU and RAL A100 rows, the phase-2 section of results/notes/point_source_cpu_campaign.md, and a README line.

Median ms/call, control → no-dedup:

Cell	RAL CPU (AMD EPYC, 5 calls/round, job 350582)	RAL A100 80GB fp64 (20×20 calls, job 350587)
simple, solved	24.00 → 5.38 (4.47×)	1.695 → 0.843 (2.01×)
simple, plain	23.76 → 5.39 (4.40×)	1.684 → 0.822 (2.05×)
vmap-4	31.31 → 11.91 (2.63×)	1.865 → 0.958 (1.95×)
cluster, solved	155.2 → 78.3 (1.98×)	4.206 → 2.457 (1.71×)
cluster, plain	153.9 → 77.7 (1.98×)	3.860 → 2.105 (1.83×)

All correctness gates are bit-identical on CPU and GPU: logL, solved positions, non-zero gradients and vmap. Compile time falls about 20% on CPU and about 40% for the simple cell on GPU. The GPU job imported PyAutoArray 25894d10, the head of the linked library PR.

Caveats:

• The CPU A/B ran on EPYC, while phase 1 ran on Xeon, so only within-host ratios are quoted.
• The CPU A/B used 5 calls per round, not 20.
• The GPU protocol resolves about 4%: hash-identical library and no-dedup programs differ by 3–4%.

Scripts Changed

• scripts/point_source/likelihood_breakdown/vertex_dedup_ab.py — new interleaved vertex-dedup A/B cell (control vs no-dedup body, correctness gates, HLO sort counts, body provenance captured at import time)
• hpc/batch_cpu/submit_breakdown_point_source_vertex_dedup_ab_ral_cpu_fp64 — RAL ral-partition CPU submit
• hpc/batch_gpu/submit_breakdown_point_source_vertex_dedup_ab_a100_fp64 — RAL A100 fp64 submit
• results/breakdown/point_source/vertex_dedup_ab_{local_cpu,hpc_ral_cpu,hpc_ral_a100}_fp64.{json,png} — laptop witness, RAL CPU (350582) and A100 (350587) rows
• results/notes/point_source_cpu_2026_09_23_ral_job_350582_vertex_dedup_ab.out, results/notes/point_source_cpu_2026_09_24_ral_job_350587_vertex_dedup_ab_a100.out — job logs
• results/notes/point_source_cpu_campaign.md — phase-2 section (A/B, GPU regression check, decision)
• README.md — dashboard line

Upstream PR

PyAutoLabs/PyAutoArray#569 — library-first: merge that PR before this one.

Test Plan

• scripts/misc/tooling/build_readme.py --check passes
• scripts/misc/wall/check_submits.py --check passes (9 wall-basis submits, 0 failing)
• RAL CPU job 350582 and A100 job 350587 COMPLETED, every correctness gate Δ = 0
• image_plane_solved profiling pin unchanged
• CI green (profiling lint checks out the PyAutoArray main, so the no-dedup library arm is only exercised once the library PR merges)

Heart RED development override

Shipped under the AUTONOMY.md "Human override for Heart RED (development only)".

• Authorization (live human, 2026-09-24, this session), faithfully quoted: asked to authorize the development-only override for PyAutoArray#568 after a GPU check, the user replied "GPU check and then I authorize". The GPU check passed (RAL A100 job 350587), so the authorization is in effect.
• Scope: commit, push and open the pending-release PRs (PyAutoArray, then the linked autolens_profiling PR). No merge, no release. Merge needs its own human /prm with every required check green.
• Exact current Heart RED reasons (pyauto-heart readiness, verdict RED, score 45):
  • release validation FAILED (stage integrate)
  • workspace validation not passing (4 failed, cloud#35579888156: autolens notebooks/cluster/modeling.ipynb, autolens notebooks/weak/a2744.ipynb, autolens scripts/cluster/modeling.py, +1 more)
  • manifest drift: hub organism blurb (organs present) — 7 mismatch(es) vs PyAutoMind/repos.yaml
• Passed branch gates: pytest test_autoarray 1626 passed at 25894d10 (re-run at ship); +10 new JAX tests, including an HLO sort-count guard that fails on main (39 sort ops) and passes on the branch. Downstream autolens_workspace_test point-source jax_likelihood ×4 and jax_grad/gradient.py print likelihoods and gradients identical to main. The autolens_profiling image_plane_solved pin is unchanged. A/B correctness gates are bit-identical on CPU (RAL job 350582) and GPU (RAL job 350587). Profiling build_readme.py --check and check_submits.py --check pass.
• This branch does not fix Heart. Heart stays RED for release purposes.

Generated by the PyAutoLabs agent workflow.

🤖 Generated with Claude Code