feat: cluster scaling relation in Lenstool convention (reference-anchored, fixed exponent, r_cut scaling)

[Jammy2211]

Overview

The cluster scaling tier currently fits b0 = scaling_factor * L^scaling_exponent with both parameters free (U(0,1) × U(0,2)) and ra/rs fixed tier-wide. Referee feedback (assessed and confirmed in PyAutoMind research, 2026-07-08) and Lenstool convention say: anchor the normalization to a reference galaxy, fix the exponent from Faber-Jackson, and scale the truncation radius too. This task moves the cluster workspace scripts (and the group-scale scaling-relation feature docs) to that convention.

Autonomy: supervised --auto with in-session one-shot ship authorization (user directive 2026-07-08: "do next two phases one shot --auto"). Judgment note: the Brain Feature Agent suggested research-first re-homing; overridden — the research is complete and encoded in the prompt (issued/-bound feature/cluster/7_scaling_relation_lenstool_convention.md).

Plan

• scripts/cluster/simulator.py: generate the scaling-tier truth on the reference-anchored relation b0_i = b0_ref * (L_i/L_ref)^0.5 and rs_i = rs_ref * (L_i/L_ref)^0.5, with the brightest member as reference. CSV schema unchanged (y, x, luminosity); prose states luminosities are relative (only L_i/L_ref matters).
• scripts/cluster/modeling.py: scaling tier composes b0_ref as the single free tier parameter (interpretable as the reference member's lens strength, prior U(0, 3) arcsec), exponent fixed at 0.5 (Faber-Jackson: b0 ∝ σ² ∝ L^0.5; Bergamini et al. 2019 kinematic calibration 2α ≈ 0.55 documented as the refinement), rs_i scaled with fixed exponent 0.5 from rs_ref. Tier dimensionality drops 2 → 1; model N=13 → 12. The old free-(factor, exponent) form stays documented as an option.
• scripts/cluster/start_here.py: mirror modeling.py.
• scripts/group/features/scaling_relation/modeling.py + modeling_for_luminosities.py: same convention + prose so group and cluster tell one story.
• Validate: simulator regeneration + both cluster scripts end-to-end in PYAUTO_TEST_MODE=2; smoke the cluster folder.

Detailed implementation plan

Affected Repositories

• autolens_workspace (primary — workspace-only; deliberately does NOT depend on the in-flight dPIEMassLenstool library PR, since b0-convention with fixed exponent 0.5 is mathematically identical to Lenstool's σ-scaling)

Branch Survey

Repository	Current Branch	Dirty?
./autolens_workspace	main	dirty (pre-existing dataset regeneration artifacts; worktree branches from origin/main)

Suggested branch: feature/cluster-scaling-lenstool
Worktree: ~/Code/PyAutoLabs-wt/cluster-scaling-lenstool/

Implementation Steps

1. simulator.py: truth relation constants → reference-anchored form; keep 10 members, luminosities log-spaced; write CSV as today.
2. modeling.py: replace the shared (scaling_factor, scaling_exponent) prior pair with b0_ref + fixed scaling_exponent = 0.5; derived per-member b0_i = b0_ref * (lum_i/lum_ref)**0.5; add rs_i = rs_ref * (lum_i/lum_ref)**0.5 (rs_ref fixed at truth, documented as freeable); update the Scaling Relation prose (referee/Lenstool rationale, α interpretation, dimensionality, Bergamini refinement, magnitudes ↔ relative luminosities).
3. start_here.py: mirror.
4. group feature scripts: same reparameterization + one-story prose.
5. Auto-sim guards already check scaling_galaxies.csv; the truth-relation change regenerates datasets on first run — verify multiple images still land at sensible positions (simulator prints).

Key Files

• autolens_workspace/scripts/cluster/{simulator,modeling,start_here}.py
• autolens_workspace/scripts/group/features/scaling_relation/{modeling,modeling_for_luminosities}.py

Original Prompt

Click to expand starting prompt

$(cat PyAutoMind/feature/cluster/7_scaling_relation_lenstool_convention.md)

[Jammy2211]

Shipped to PR-open (in-session one-shot authorization). PR: #238 (pending-release ✓, workspace-only — no library merge gate).

Gate record: cluster simulator ran end-to-end on the new truth relation (both sources still triply imaged; dataset regenerated in the diff) · structural check: model prior_count 13→12, brightest member anchors exactly (b0=b0_ref, rs=10"), faintest scales ×√(0.05/0.40) · group family arithmetic-identical at luminosity-ratio 1, dataset untouched · review surface CLEAN (26 files; deletions = stale old-truth artifacts + prose fixes) · Heart YELLOW acked in-session.

Two loose ends, both non-blocking:

1. tracer.png / galaxies_images.png deleted (old-truth artifacts); regeneration via subplot_tracer is pathologically slow on the 14-galaxy multi-plane tracer with galaxy-scale plot defaults (the PyAutoLens#578 problem) — a background render is running and will be push-appended to the branch if it completes.
2. Follow-up scope (old relation form still shown in): scripts/imaging/features/scaling_relation/, scripts/group/slam.py, guides/advanced/multi_plane.py — natural bundle for the flagship Lenstool example (task 8) or a small docs sweep.

This closes the referee-comment response: workspace scaling relations now match the Lenstool/community convention end to end. Merge stays with you.

[Jammy2211]

PR #238 merged (incl. navigator-catalogue regen follow-up commit). Scaling relations now match the Lenstool/community convention across cluster + group tiers. The two illustrative dataset PNGs regenerate on any user simulator run (background render did not complete in reasonable time — the slow-viz problem the merged PyAutoLens#578 addresses). Closing.