Likelihood lifecycle phase 2a: profiling campaign — imaging/{mge, rectangular, delaunay, delaunay_nn} (106 planned cells)

Type: research
Target: autolens_profiling
Repos:

• autolens_profiling
  Themes:
• profiling
  Difficulty: large
  Autonomy: supervised
  Priority: high
  Status: formalised
  Consequence: judge
  Witness: the coverage matrix shows zero missing and zero stale cells for class ∈ {optimal, secondary} × measurement ∈ {runtime, breakdown} for this campaign's likelihoods; every new current cell carries fit.png, provenance and parity.json; build_catalogue --check --validate-with ../PyAutoPulse valid at @3; the campaign wiki page records the pre-registered rule and verdict
  Review-minutes: 45
  Unattended: never
  Epic: likelihood-lifecycle
  Phase: 2
  Parent: draft/research/autolens_profiling/likelihood_lifecycle_epic.md
  Filed: 2026-10-24

Phase 2 sub-issue 2a of the likelihood-lifecycle epic (report §4 "Phase 2"; ledger "Phase 2" card). Depends on phase 0 (complete) and phase 1B (merged, autolens_profiling#453); start after autolens_profiling#455 (follow-ups) merges. Scope: imaging/{mge, rectangular, delaunay, delaunay_nn} (106 planned cells).

Size: the 2026-10-17 imaging results are at Delaunay 1500 / rectangular 39², not the standard size, so every cell is re-measured at 1250 / 35². That reuse is no longer possible, and the ledger card's "reuse" line is superseded by the size ruling. First, add the missing parity.json writer (#451 defined only the format): each cell computes its log-likelihood against a named reference and writes parity.json beside the row. Adopts Pulse hst_mge_cpu_pilot.

Rules (every phase-2 sub-issue; epic ledger "Phase 2" card)

• The coverage matrix (python scripts/misc/tooling/coverage_matrix.py → dashboard/coverage.md) names this campaign's missing and stale cells; those cells are the work list.
• Every cell is measured at the standard size (Delaunay 1250, rectangular 35² = 1225; human ruling 2026-10-23) on a pinned released tag, from a private RAL checkout (never the shared mirror). CPU arrays go on --partition=ral only; A100 runs use --gres. The human gives a go for each RAL batch. Never touch another campaign's jobs.
• Every cell lands as current, carrying runtime and breakdown, provenance, fit.png (phase 1B renders it automatically) and a parity.json against a reference log-likelihood. Prior rows on the key become superseded through the catalogue builder.
• A campaign page at the mirrored wiki path (wiki/campaigns/<dataset_type>/<variant>/) records the rule and the expected verdict before the runs, then the result.
• Pulse receipt refetched after merge. Phase 2 ends only on the human's "campaigns complete" confirmation (the gate for phase 3).

Plan approved 2026-10-24 (human), with rulings

• Pinned tag: 2026.10.23.1 (the latest PyPI release), shared by every phase-2 sub-issue. Rows on other likelihoods go stale and are re-measured in 2b–2g.
• Standard Delaunay = 1250 Hilbert + 30 edge (1280 recorded). Rows record source_pixels_requested: 1250 beside the actual count, and the size rule matches on the requested count. Rectangular = 35².
• PR A (code, laptop):
  • --source-pixels in the 3 JAX runtime scripts (hard-coded 1500 / 39² today) and the numba scripts (presets; jwst/ao have none); a --standard-size flag.
  • A parity.json writer: an in-process eager NumPy reference on the same inputs and size, absolute tolerance 1e-4·|ref| (1e-3 mixed).
  • New submits: a RAL CPU array on --partition=ral for the 44 CPU cells, plus one --gres=gpu:1 A100 job per instrument. The 18 jax_cpu_sparse@hpc_a100 cells run inside those GPU jobs. Each submit gets a WALL-BASIS block.
• Batch 0 (human go): one short timing arm per new submit kind, to set --time from measurement.
• Batch 1 (human go): all 106 cells from a private RAL checkout at the pinned tag.
• PR B: commit the rows, regenerate the catalogue and coverage, write the campaign wiki page (pre-registered rule: every cell current with parity match; any divergence blocks that setup), refetch Pulse, and adopt and close Pulse hst_mge_cpu_pilot.
• Tier: judge, merge by human /prm. Starts after autolens_profiling#455 merges.