feat: cluster likelihood per-step breakdown scripts (source-plane + image-plane chi²)

[Jammy2211]

Overview

Completes PyAutoMind feature/cluster/5_profiling.md (the last pre-LensTool item of the cluster home straight): two scripts in likelihood_breakdown/cluster/ giving the per-step JIT timing decomposition of the cluster point-source likelihoods, in the house style of likelihood_breakdown/imaging/* (per-step jit_profile lower/compile/first/steady-state, JSON + bar chart into results/breakdown/cluster/, AUTOLENS_PROFILING_SMOKE=1 short-circuit, CPU fp64 single-config).

Autonomy: supervised --auto (user directive 2026-07-09: "do the next task --auto"; heart-ack carried in-session).

Plan

• likelihood_breakdown/cluster/source_plane.py — decomposes FitPositionsSource for the standard cluster model (2 main dPIE + 10 scaling members + NFW host at z=0.5; sources at z=1.0/2.0): tracer build from the workspace dataset CSVs → multi-plane traced grids (the 149-free… 14-profile deflection stack is the expected hot spot) → per-system centroid + chi² → likelihood assembly. Known constraint designed around: the full source-plane chi² is JIT-blocked (workspace_test table), so steps are JIT-profiled individually where traceable with the eager reference asserted at rtol=1e-4, and the blocked step documented.
• likelihood_breakdown/cluster/image_plane.py — decomposes FitPositionsImagePairRepeat: tracer build → triangle-tiling PointSolver solve per source plane (dominant; JAX-jitted per the simulators/cluster.py jitted_solve precedent) → Hungarian pairing → chi². Per-plane timing table so the multi-plane fan-out cost is visible.
• Both consume autolens_workspace/dataset/cluster/simple via the standard auto-sim guard (regenerates through the workspace simulator, current reference-anchored truth), mirroring how simulators/cluster.py resolves _workspace_root.
• README auto-table hookup via the existing results/breakdown/ artifact convention.

Affected Repositories

• autolens_profiling (primary) — parallel to the active profiling-preopt-campaign claim; its diff is likelihood_runtime/sweep.py + scripts/build_baseline.py, disjoint from the new likelihood_breakdown/cluster/; re-verify at ship.

Branch: feature/cluster-likelihood-breakdown · Worktree: ~/Code/PyAutoLabs-wt/cluster-likelihood-breakdown/

Follow-up noted (not this task)

simulators/cluster.py (run-time tracking) predates the scaling-member tier and the reference-anchored truth relation — it writes a 5-component model into the shared workspace dataset path and would fight the current workspace simulator if run. Needs a small sync task.

Original Prompt

Click to expand

$(cat PyAutoMind/feature/cluster/5_profiling.md)

[Jammy2211]

Shipped to PR-open: #58 (pending-release ✓).

Headline numbers (CPU fp64, standard 13-mass-component cluster model, 2 sources / 6 images):

• Source-plane (FitPositionsSource, Lenstool's default likelihood): per-step LL matches the production fit to 8 digits; total steady-state 3.1 ms, split between the multi-plane ray-trace and the Hessian magnification steps. Three production-formula facts documented while achieving the match: chi² weights residual distances by |μ|², name pairing hands the fit the model's Point centre (barycenter only as a no-profile fallback), and the noise normalization uses σ/μ.
• Image-plane (FitPositionsImagePairRepeat): triangle-tiling solve compiles in ~10.5 s per source plane and runs at ~0.32 s/call steady-state (200×200 @ 0.7″ tutorial grid, precision 0.01″); fit totals ~2 s/system. The ~100× source-vs-image-plane cost ratio is now quantified per step — the number the solver task (feature/cluster/10_solver_over_under_prediction.md) and any JAX solve optimization should anchor on.
• simulators/cluster.py brought to the current standard model (scaling tier on the #238 reference-anchored convention) — the follow-up flagged at filing, folded in.

Gate record: in-run rtol=1e-4 LL assert (source-plane) · both scripts end-to-end from clean incl. auto-simulation through should_simulate · SMOKE=1 short-circuits verified · ruff clean · review CLEAN (995 additive) · parallel-claim disjointness vs profiling-preopt-campaign re-verified · Heart-ack carried in-session. Merge stays with you.

[Jammy2211]

PR #58 merged (after a ruff-format fix commit). Cluster likelihood breakdowns live. Closing.