Type: feature Target: workspaces Repos:
- workspaces Difficulty: large Autonomy: supervised Priority: normal Status: formalised Epic: cluster-strong-lensing Phase: 10 Parent: draft/feature/autolens/source_cluster_arc.md Filed: 2026-08-19 (backfilled from git)
Part of the Source & Cluster arc (phase 10 of 12), gated on phases 4 and 9. User request (verbatim): "Ability to use parametric pixelized source reconstructions on clusters after model is inferred, perhaps include a few extra mass components for refining it. Requires us to deal with masking challenges, that is every source needs to be paired to a mask. You could either give it one image for the whole cluster, which is huge, and would end up fitting the wrong stuff. Instead, each source needs its own mask, thus we need some sort of system to ensure that happens. Obvious API is to use AnalysisFactor in a similar fashion to multi_dataset, e.g. one Analysis for each source, and to use the same source naming API to only evaluate each source light profile for each mask. This requires users to use a GUI to make custom masks, which is fine but annoying."
Audit: the mechanism fully exists — the gap is purely the example/wiring layer.
af.AnalysisFactor(prior_model, analysis)binds one Analysis (one masked dataset) to a shared prior_model; cluster start_here.py:419-434 already uses one factor per SOURCE (point datasets); multi_dataset/start_here.py:207-233,321-350 is the exact per-dataset-mask template; multi_dataset/features/imaging_and_point_source/modeling.py is the nearest precedent for mixing point positions + pixelized imaging under one mass model (galaxy-scale today).- draft/docs/workspaces/cluster_regime_narrative.md already plans a features/extended_source/ cluster follow-up (one A2744 arc, pixelized, cluster mass model as start) — this phase implements that plan's machinery and supersedes its gap-filling bullet; keep the narrative work there.
Work:
- Example:
scripts/cluster/features/extended_source/— post-inference refinement: fix (or narrow priors from) the inferred cluster mass model, add a few free local mass components near the refined arc, one AnalysisImaging per source with its own mask, one AnalysisPoint per remaining source, all in one FactorGraphModel. - Masking system: per-source mask pairing convention (source-name → mask file), GUI custom-mask flow documented (existing mask GUI), loud failure when a source lacks a mask (no silent guards).
- Source-naming API: ensure per-factor model evaluation only computes that source's light for that mask (verify what FactorGraphModel already gives; only extend the library if evaluation actually crosses sources).
- Decide and document: whole-cluster single mask explicitly rejected (fits the wrong stuff) — say so in prose.