[Jammy2211]

Summary

The gradient-search admission measurement the point-source source-plane campaign (PyAutoPulse point-source-source) is parked on. af.MultiStartAdam runs in the analysis-declared forward gradient_mode. It uses the same L5 target as the Nautilus leaf, on the released 2026.10.17.1 wheels (private RAL venv, jax 0.11.2), RAL CPU (ral) + A100 (--exclusive), seeds 0–4, two step ceilings. Closes #25 on merge.

Evidence, not a verdict. At a 3000-step ceiling every seed converges in 259–1432 steps to the Nautilus log-L range (16.43–16.49) with all parameters ≤ 0.80σ (Nautilus σ) from truth. At the library-default 300, 4/5 seeds stop on the budget, at up to 1.44σ. The likelihood with its gradient is 0.8–3.7% of search wall on RAL CPU, 0.3–2.3% on A100. The rest is a fixed per-fit cost (~7 s CPU / ~19 s A100: compile, start draws, output) and per-step optimiser overhead. A100 is slower end to end than an 8-CPU slice here, because the step is launch-bound. Whether this answers the admission question, or a BlackJAX NUTS/SMC leaf is still wanted, is the human's call.

Scripts Changed

• scripts/point_source/simple/multi_start_adam.py (new): MultiStartAdam leaf, library-default 300-step ceiling.
• scripts/point_source/simple/multi_start_adam_n3000.py (new): same, ceiling 3000 (auto-convergence ends the run).
• scripts/misc/searches/_point_runner.py: gradient admission bar (batched value_and_grad per eval, value-only cost, compile, steps, evals = steps × starts, share), MAP vs truth in committed-Nautilus σ, package_versions provenance for installed-release runs.
• scripts/misc/test/test_point_runner.py: tests for the new pure helpers.
• scripts/misc/wall/check_submits.py: cell-id mapping for the two setup-layout leaves (keeps the leaf-is-instrument check).
• hpc/batch_cpu/submit_search_multi_start_adam_simple_ral_cpu, hpc/batch_gpu/submit_search_multi_start_adam_simple_a100 (new; private venv, ral / exclusive A100).
• .github/workflows/lint.yml: smoke the two new leaves.
• Results: 20 rows under results/searches/point_source/multi_start_adam{,_n3000}/, SLURM logs results/logs/point_source/*.41249{5,8}_*, regenerated README / dashboard/summary.json / dashboard/catalogue.json; journal in wiki/project/state.md; setup + searches README notes.

Runs

RAL 412494 (venv build), 412495 (CPU array, 10 tasks), 412498 (A100 array, 10 tasks): all COMPLETED 0:0. Cortex: PyAutoCortex a87fd8c (runs logged). Companion timing PR: PyAutoLabs/autolens_profiling#432.

Validation

ruff check / format, pytest scripts/misc/test (131 passed), build_readme.py --check, export_inference_summary.py --check, check_submits.py --check, smoke short-circuit of both new leaves: all pass locally. Heart: organism STALE; scoped (autolens_inference, autolens_profiling) GREEN, no in-scope blockers.

Notes for review

• autolens_inference is also claimed by hst-cleanslate-light-mode (feat: lens-light mode variant, row provenance and pinned clean-slate submits #24); paths are disjoint, but README/summary/catalogue are regenerated files and may need a re-render on whichever merges second.
• No Nautilus control was re-run on this stack: same-config rows would overwrite the 2026-09-28 rows. The reference σ is the committed Nautilus rows.
• Tier: judge — human /prm.

🤖 Generated with Claude Code

https://claude.ai/code/session_01QBeyBLN7ntDChx9v1fmt6M