[Jammy2211]

Summary

Timing side of PyAutoLabs/autolens_inference#29 (gradient-search admission evidence for the parked source-plane campaign). Closes #428 on merge.

• fix: steady single-JIT median beside the A100 first-block statistic #371 fresh-median evidence. The source-plane runtime cell re-run on the released 2026.10.17.1 wheels (private venv, jax 0.11.2).

  • A100 (hpc_a100_fp64, job 412517, euclid-ral-gpu-2, --exclusive): full_pipeline_single_jit_median 0.233 ms (p10–p90 0.229–0.262); single_jit, the first block after compile, 0.347 ms.
  • RAL CPU (hpc_ral_cpu_fp64, job 412518, ral 8490H, shared node at load ~10): median 0.147 ms (0.126–0.168); single_jit 0.177 ms.
  • Pinned value passed on both. These rows are the first in the cell stamped with the release version (installed wheels) rather than 2026.8.17.1.

• Campaign ledger + wiki. A 2026-10-17 section records the MultiStartAdam admission evidence:

  • converges to the Nautilus log-L range in 259–1432 steps;
  • likelihood+gradient ≤ 3.7% of search wall on CPU, ≤ 2.3% on A100.

  The parked BlackJAX candidate's admission call stays with the human.

Scripts Changed

• hpc/batch_gpu/submit_runtime_point_source_source_source_plane_solved_a100_fp64, hpc/batch_cpu/submit_runtime_point_source_source_source_plane_solved_ral_cpu_fp64: optional PYAUTO_VENV override (activate a venv of installed wheels, clear PYTHONPATH). Default behaviour unchanged.
• Results: results/runtime/point_source_source/source_plane_solved/source_plane_solved_{hpc_a100,hpc_ral_cpu}_fp64.{json,png} + comparison.*, logs results/logs/point_source_source/point_source_source_plane_2026_10_17_ral_job_41251{7,8}.out; results/notes/point_source_source_plane_campaign.md, wiki/campaigns/point_source_source_plane.md; re-rendered README, dashboard/ (incl. catalogue shards: generator timestamps/revision) and setup wiki pages.

Validation

scripts/misc/test/test_headline_completion.py: the snapshot test "no committed runtime row carries a median" became the estimator-rule invariant, because these are the first committed median rows (CI caught it; full suite 1457 passed locally).

build_readme.py --check, build_catalogue.py --check --validate-with ../PyAutoPulse, check_wiki.py --check, build_dashboard.py --check, check_results_layout.py --check, check_submits.py --check, ruff: all pass locally (rendered from the task worktree beside the Brain/Pulse symlinks). Heart: organism STALE; scoped GREEN.

Notes for review

• The two rows replace the committed 2026-09-28 rows of the same configs. The CPU row moves from the EPYC gpu node (old exemption) to the ral 8490H, which is the partition rule now. This is not a full release sweep.
• autolens_profiling is also claimed by hst-mge-pilot-qualification (docs: design independent HST MGE pilot qualification #418) and touched by a sibling point-source-image arm. Dashboard/catalogue are regenerated files: whichever PR merges second re-renders.
• Tier: judge — human /prm.

🤖 Generated with Claude Code

https://claude.ai/code/session_01QBeyBLN7ntDChx9v1fmt6M