Point-source defaults campaign: solved centres, tensor, pairing

[Jammy2211]

Overview

End goal: a full workspace docs update — the point_source guides and modeling scripts present ALL likelihood options with performance/robustness evidence — with the complete evidence base built in autolens_profiling (truth-anchored result JSONs + a results/notes synthesis). Adopts solved-centre demonstrated defaults everywhere, tensor source-plane weighting, and (pending the phase-B discriminator) all-to-all image-plane pairing as the AnalysisPoint default. Follow-on to the #657 series; SUPERSEDES the same-day galaxy/cluster split decision; COORDINATES with autolens_workspace#436 (cluster swap, in flight — phase D reconciles, never duplicates).

Plan

• Phase A — library prerequisites (PyAutoLens, small PR, FIRST): log-space rework of the FitPositionsImagePairAll mixture chi-squared (values identical where currently finite, finite where currently -inf at ≳38σ) + a weighting option on free-centre FitPositionsSource ("magnification" default, back-compat; "jacobian" opts into the tensor). wst jax_likelihood literals re-run as the invariance gate; merge + HPCPullPyAuto before phase B.
• Phase B — evidence campaign (autolens_profiling, RAL A100s): truth-anchored matrix on the galaxy-scale simple quad AND a ported cluster tier; five experiments — source-plane tensor flavours, image-plane solved-vs-free re-run post-fix, pairing discriminator (missing/extra image), posterior-width honesty, near-caustic stress. JSONs committed to results/searches/** + results/notes/point_source_defaults_campaign.md synthesis.
• Phase C — defaults change (PyAutoLens): AnalysisPoint.fit_positions_cls → FitPositionsImagePairAllSolved (expected, pending B); FitPointDataset default reconciled; release notes with ## API Changes heading.
• Phase D — workspace docs (autolens_workspace, THE END GOAL): solved-centre demonstrated defaults in point_source modeling scripts; free-centre clearly documented with its use-cases (informative centre priors, linked centres across bands/epochs, centre-as-science, glSNe standard-candle flux caveat); guides pairing doc becomes the full option matrix with phase-B numbers; cluster scripts reconciled with test: mask padding likelihood sanity check #436; lenstool mirror stays free-centre scalar; notebooks regenerated; ship behind the library-first gate.
• Explicitly deferred: the time-delay + solved-fluxes free-H0 search arm.

Detailed implementation plan

Affected Repositories

• PyAutoLens (primary; phases A + C)
• autolens_profiling (phase B)
• autolens_workspace (phase D)
• autolens_workspace_test (read-only — invariance-gate runs only)

Branch Survey (2026-07-31)

Repository	Current Branch	Dirty?
./PyAutoLens	main	clean — claimed concurrently by potential-correction-validation (#672); scopes disjoint (autolens/point/fit/ vs pixelization); pre-merge origin/main before PR
./autolens_profiling	main	untracked dataset/imaging/jwst_lw/ (unrelated)
./autolens_workspace	main	clean — claimed by 3 in-flight tasks incl. #436; phase-D claim deferred until D starts
./autolens_workspace_test	main	clean

Suggested branch: feature/point-source-defaults-campaign
Worktree root: ~/Code/PyAutoLabs-wt/point-source-defaults-campaign/ (PyAutoLens at phase A; worktree_add_repo autolens_profiling at phase B)
Work Classification: Both (phased: library → profiling evidence → library → workspace)

Implementation Steps

Phase A (PyAutoLens):

1. autolens/point/fit/positions/image/pair_all.py:96-133 — all_permutations_log_likelihoods computes log(sum(exp(log_p))) literally; rework as a shifted logsumexp (max + log(sum(exp(x − max))), xp-threaded so numpy and JAX paths agree). Values identical wherever currently finite; finite where currently -inf; NaN-padded solver rows must drop out exactly as before; the no_image_residual / has_image fallback in chi_squared (:136-188) unchanged.
2. autolens/point/fit/positions/source/separations.py — weighting class attribute on FitPositionsSource (default "magnification", byte-identical existing behaviour — Lenstool comparisons and literals untouched; "jacobian" computes the quadratic-form chi-squared + matching normalization around the FREE centre, reusing precision_tensor_components_from from fit/solved.py).
3. Numpy unit tests in test_autolens/point/ (house rule: no JAX in unit tests): old-vs-new equality at moderate σ; finiteness at extreme (≳38σ) mismatch; free-tensor vs solved-tensor consistency when evaluated at the solved centre.
4. Invariance gate: re-run the autolens_workspace_test jax_likelihood point_source literal scripts — unchanged at truth.
5. Small PR; after merge, HPCPullPyAuto syncs the RAL mirrors before any phase-B submission.

Phase B (autolens_profiling):
6. Galaxy tier: existing simple quad anchors (results/searches/**, PR#99 methodology — evaluate each flavour AT SIMULATOR TRUTH; compare Nautilus + Prodigy best-logL / recovery / wall against it).
7. Cluster tier: port the autolens_workspace/cluster test case (family-CSV multi-plane system — 2 sources at different redshifts, dPIE members + host halo, point_datasets.csv conventions). Pin cosmology in every solver-chained gradient cell (free cosmology cannot cross the solver custom_jvp boundary — Tracer aux). Absorbs draft/research/autolens_profiling/cluster_gradient_search_benchmark.md (banner in file).
8. Experiments (both tiers): (i) source-plane tensor-solved vs free-centre-tensor (new, A.2) vs scalar flavours; (ii) image-plane solved vs free re-run AFTER A.1 (tests the -inf-plateau hypothesis for the 256-start failure); (iii) pairing discriminator — quad with one image REMOVED and one spurious EXTRA position, PairAllSolved vs PairRepeatSolved posterior bias (Nautilus) — this decides the phase-C default; (iv) posterior WIDTH free vs solved image-plane (profile-vs-marginal honesty — if solved posteriors are materially overconfident, the docs' free-centre section must say so); (v) near-caustic stress test (where tensor source-plane breaks and image-plane must take over → domain-of-validity prose).
9. Execution: hpc/sync push-submit gpu <script> (SLURM gpu-partition array); JAX_ENABLE_X64 set INSIDE scripts/sbatch (sbatch does NOT inherit it — silent float32 poisons the anchors); submit detached (/mnt/ral is NFS-slow); hpc/sync pull JSONs back; device blocks recorded so CPU-laptop vs A100 rows are distinguishable (CPU spot-checks are smoke only, never committed evidence).
10. Synthesis: results/notes/point_source_defaults_campaign.md.

Phase C (PyAutoLens):
11. autolens/point/model/analysis.py:44 fit_positions_cls=FitPositionsImagePairRepeat → FitPositionsImagePairAllSolved (expected, pending step 8.iii); autolens/point/fit/dataset.py default (FitPositionsImagePair, the superseded Hungarian) decided consistently. Release notes with ## API Changes heading (classify_pr matches headings).

Phase D (autolens_workspace):
12. point_source modeling.py scripts → solved-centre demonstrated defaults; free-centre composition as the clearly-documented alternative with use-cases; cluster scripts touched only to reconcile with what #436 lands.
13. The point-source pairing guide → full option matrix with phase-B numbers (per option: params, wall/eval cost, gradient behaviour, robustness to missing/extra images, when to prefer).
14. point_source/fit.py, cluster guides, lenstool mapping updated consistently (lenstool stays free-centre scalar deliberately — its table already notes the solved/tensor siblings); notebooks regenerated; ship_workspace behind the library-first merge gate.

Key Files

• autolens/point/fit/positions/image/pair_all.py — A.1 logsumexp rework
• autolens/point/fit/positions/source/separations.py + autolens/point/fit/solved.py — A.2 free-centre tensor option
• autolens/point/model/analysis.py, autolens/point/fit/dataset.py — phase-C defaults
• autolens_profiling/results/searches/**, results/notes/point_source_defaults_campaign.md — phase-B evidence
• autolens_workspace/scripts/point_source/** + guides — phase D

Context literals (2026-07-31, #657 + profiling PR#99)

• Truth anchors (galaxy-scale simple quad, |mu| at obs = [8.2, 45.9, 366.6, 28.3]): image_plane free +7.20 / solved +7.74; source_plane scalar −33788.4 (the mu=367 image's radial noise mis-mapped), tensor-solved +0.6.
• Tensor-vs-solved isolation: solved+scalar still prefers wrong models (truth −2338 vs −848/−610); solved+tensor ranks truth first by >1500 logL — the WEIGHTING is the fix; the solved centre is the orthogonal dimensionality/marginalization win.
• Benchmark: nautilus image_plane +9.56 (740s) at truth; prodigy solved +2.37 (982s, 64×300) truth recovered THROUGH the solver; prodigy free −47.7 (3516s, 256×300) basin found but plateaued — the -inf plateau is the prime suspect (A.1).
• Constraint: free cosmology cannot cross the solver custom_jvp boundary — pin cosmology in solver-chained cells.

Exit criteria

Phase-A PR merged (literals invariant); phase-B notes doc + JSONs committed; phase-C default change merged with release notes; phase-D workspace PR merged — guides/scripts present all options with evidence-backed defaults (solved centres, tensor source-plane, all-to-all image-plane if phase B confirms) and a clear, use-case-driven free-centre section.

Original Prompt

Click to expand starting prompt

Point-source defaults campaign: tensor/solved defaults, pairing decision, full docs + performance evidence

Human-scoped 2026-07-31 (#657 wrap-up chat; run in a fresh session). End goal: a full
workspace docs update — the point_source guides and modeling scripts present ALL
likelihood options with performance/robustness information — with the complete evidence
base built in autolens_profiling (truth-anchored result JSONs + a results/notes summary
doc; "full information building", not one-off runs).

Target defaults to demonstrate and then adopt:

1. Source-plane chi-squared: tensor weighting with SOLVED centres
   (FitPositionsSourceSolved, weighting="jacobian"), while also demonstrating the
   tensor works with FREE centres (requires phase A below — no free-centre tensor exists
   in the API yet).
2. Image-plane chi-squared: SOLVED centres (free centres demonstrably need sampler
   muscle: 2026-07-31 benchmark — free-centre Prodigy needed 256 starts vs solved
   converging at 64; Nautilus handles both).
3. Workspace demonstrated defaults move to solved centres everywhere, with the free
   option CLEARLY documented (modeling.py scripts + point_source guides), including WHEN
   free is required: informative centre priors, linked centres across bands/epochs,
   centre-as-science, and the standard-candle flux caveat (PointSolved forces
   FitFluxesSolved, whose flat-prior F* discards standardizable-candle priors — glSNe).
4. Image-plane pairing: expected outcome is all-to-all (FitPositionsImagePairAll*)
   as default — smooth in the pairings (gradient-friendly), principled Occam mixture —
   pending the phase-B robustness discriminator. This is an AnalysisPoint DEFAULT
   CHANGE (currently FitPositionsImagePairRepeat) → release notes ## API Changes.

SUPERSEDES the 2026-07-31 morning split decision (galaxy-scale free / cluster solved)
— the human widened it to solved demonstrated defaults everywhere. COORDINATE, don't
duplicate: the cluster swap is ALREADY IN FLIGHT as its own task
(active/cluster_default_point_solved.md, issue autolens_workspace#436, routed to
start_workspace 2026-07-31) — phase D here covers the galaxy-scale point_source scripts
and the guides; touch cluster scripts only to reconcile with whatever #436 lands (it may
merge first). This prompt also ABSORBS the ideas.md PairAll-logsumexp entry (phase A).
EXPLICITLY DEFERRED: the time-delay + solved-fluxes free-H0 search arm (revisit after
this campaign).

Phase A — library prerequisites (PyAutoLens; small PR, do FIRST)

1. PairAll log-sum-exp stabilization: FitPositionsImagePairAll.chi_squared
   exponentiates before logging → underflows to -inf at ≳38σ worst-image mismatch
   (measured 2026-07-31), strangling gradient flow (the -inf plateau is the prime
   suspect for the 64-start image-plane failure). Rework in log-space: values identical
   wherever currently finite, finite where currently -inf; re-run the wst
   jax_likelihood literals (unchanged at truth) as the invariance gate. Prerequisite —
   every phase-B result changes if run against the unfixed objective.
2. Free-centre tensor option: weighting class attribute on FitPositionsSource
   (default "magnification" — back-compat, Lenstool comparisons and literals untouched;
   "jacobian" opts into the tensor + matching normalization, reusing
   precision_tensor_components_from from solved.py). Numpy unit tests per house rule.

Phase B — evidence campaign (autolens_profiling; truth-anchored; A100s on RAL)

Methodology from the 2026-07-31 runs (results/searches/**, PR#99): evaluate each
likelihood flavour AT THE SIMULATOR TRUTH as the anchor; compare Nautilus + Prodigy
best-logL/recovery/wall against it. Write every run's JSON + a
results/notes/point_source_defaults_campaign.md synthesis.

Two dataset tiers — run the full matrix on BOTH:

• Galaxy-scale: the existing profiling simple quad (anchors + 2026-07-31 results
  already in results/searches/).
• Cluster-scale: port the autolens_workspace/cluster test case (the family-CSV
  multi-plane system — 2 sources at different redshifts, dPIE members + host halo,
  point_datasets.csv conventions; the good test case per the human). Expect MUCH
  slower (image-plane forward solve ~0.3 s/call at cluster scale; multi-plane tracer).
  This is where the solved-centre dimensionality win compounds (−2 params/source) and
  where the defaults matter most. Constraint: pin cosmology in every solver-chained
  gradient cell (free cosmology cannot cross the custom_jvp boundary — Tracer aux).
  This tier ABSORBS draft/research/autolens_profiling/cluster_gradient_search_benchmark.md.

Execution environment — A100s on RAL (the human's call: the profiling runs that guide
all of this run on GPU):

• Drive with the project's hpc/sync CLI: hpc/sync push-submit gpu <script> (SLURM
  gpu-partition array, JAX auto-uses the A100), then hpc/sync jobs / tail gpu /
  pull. Venv: /mnt/ral/jnightin/PyAuto via its activate.sh (PYAUTO_HPC_BASE).
• Phase A must be MERGED and synced to RAL first — refresh the mirrored library
  mains with HPCPullPyAuto before submitting anything.
• GOTCHA (recorded): sbatch does NOT inherit JAX_ENABLE_X64 — set x64 explicitly
  inside the scripts/sbatch or everything runs float32 silently, poisoning the truth
  anchors and FD checks. /mnt/ral is NFS-slow — submit detached, don't babysit.
• hpc/sync pull the result JSONs back into results/searches/** so the
  information-building convention (committed JSONs + notes synthesis) holds; record
  device blocks (the JSONs carry device.backend) so CPU-laptop vs A100 rows are
  distinguishable. CPU spot-checks on the laptop are fine for smoke, not for the
  committed evidence.

1. Source-plane tensor-solved cells (Nautilus + Prodigy source_plane_solved — fit
   dispatch already exists) vs the free-centre TENSOR variant (new, phase A.2) vs the
   scalar flavours (already run: scalar truth −33788 vs wrong models −110/−313 — the
   bias showcase to reproduce in the notes).
2. Image-plane solved vs free (extend the existing 2026-07-31 results: solved 64×300
   converges +2.37/truth+7.74; free needs 256 starts, plateaus at −47.7/truth+7.20).
   Re-run the free-centre arm AFTER phase A.1 to test the plateau hypothesis.
3. Pairing discriminator (the open decision): simulate a quad with (a) one image
   REMOVED from the dataset and (b) one spurious EXTRA position; compare posterior bias
   of PairAllSolved vs PairRepeatSolved (Nautilus). This decides all-to-all vs
   repeat as default; document whichever caveats fall out.
4. Posterior WIDTH comparison (profile-vs-marginal honesty): same dataset, Nautilus
   free-centre vs solved image-plane — compare mass-parameter error bars. The image-plane
   solved variants are plug-in profiles (no marginalization term); if solved posteriors
   are materially overconfident, the docs' free-centre section must say so.
5. Near-caustic stress test (tensor domain of validity): source hugging the caustic
   so the linearization A at observed positions degrades; establish where tensor
   source-plane itself breaks and image-plane must take over → domain-of-validity prose
   for the pairing guide.

Phase C — defaults decision + library change (PyAutoLens)

Informed by phase B: AnalysisPoint fit_positions_cls default →
FitPositionsImagePairAllSolved (expected); decide FitPointDataset default
consistently (currently FitPositionsImagePair — the Hungarian, already superseded).
Release notes with ## API Changes heading (classify_pr matches headings).

Phase D — workspace docs update (autolens_workspace; the END GOAL)

• All point_source + cluster modeling.py scripts demonstrate solved-centre defaults;
  free-centre composition shown as the clearly-documented alternative WITH its
  use-cases (see 3 above).
• guides/point_source_pairing.py becomes the full option matrix WITH performance and
  robustness numbers from phase B (per-option: params, wall/eval cost, gradient
  behaviour, robustness to missing/extra images, when to prefer).
• point_source/fit.py, cluster guides, lenstool mapping updated consistently
  (lenstool script stays free-centre scalar deliberately — its table already notes the
  solved/tensor siblings).
• Notebooks regenerated; ship_workspace behind the library-first gate.

Context literals (2026-07-31, all in autolens_profiling results/searches + #657)

• Truth anchors (galaxy-scale simple quad, |mu| at obs = [8.2, 45.9, 366.6, 28.3]):
  image_plane free +7.20 / solved +7.74; source_plane scalar −33788.4 (the mu=367 image's
  radial noise mis-mapped), tensor-solved +0.6.
• Tensor-vs-solved isolation: solved+scalar still prefers wrong models (truth −2338 vs
  −848/−610); solved+tensor ranks truth first by >1500 logL → the weighting is the
  fix; the solved centre is the orthogonal dimensionality/marginalization win.
• Benchmark: nautilus image_plane +9.56 (740s) at truth; prodigy solved +2.37 (982s,
  64×300) truth recovered THROUGH the solver; prodigy free −47.7 (3516s, 256×300)
  basin found but plateaued.
• Constraint: free cosmology cannot cross the solver custom_jvp boundary (Tracer aux) —
  pin cosmology in solver-chained cells (follow-up filed separately).

Exit criteria

Phase-A PR merged (literals invariant); phase-B notes doc + JSONs committed; phase-C
default change merged with release notes; phase-D workspace PR merged: guides/scripts
present all options with evidence-backed defaults (solved centres, tensor source-plane,
all-to-all image-plane if phase B confirms) and a clear, use-case-driven free-centre
section.

[Jammy2211]

Phase A: Library PR Created

PR: #679 (pending-release, awaiting human merge)

Gate evidence:

• Full test_autolens/ suite: 498 passed (worktree, post-merge of origin/main — already up to date with the feat: PointSolver implicit-diff gradients (phase 5 of #657) #677 merge base).
• Existing pair_all literal tests pass with their exact == asserts unchanged — value identity in the finite regime, no re-pinning needed.
• wst invariance gate: all four scripts/point_source/jax_likelihood/ scripts pass with pinned literals matched exactly (point.py −82.33883111107943 NumPy=JIT; image_plane.py all-solved −82.33883111107943 / repeat-solved −89.71129441664627 NumPy=JIT; source_plane.py eager −94.70750992850252; fluxes_time_delays.py −133.0115150576916 / H0-matched −137.67455435685469 eager=vmap). The source_plane.py/fluxes_time_delays.py jit-path BLOCKER prints are the pre-existing tracked PointSolver-pytree limitation, not a regression.

Heart note: shipped under explicit human authorization with Heart RED from the unrelated overnight release-validation failure (release validation FAILED (stage integrate), nightly v2026.7.31.1.dev69201) — orthogonal to this branch; feature-scoped evidence above is fully green.

Workspace impact: additive + value-preserving — no immediate migration. All workspace consumers of the touched symbols are the point_source/cluster scripts already scheduled as phase D (post-evidence), plus the wst parity scripts validated by the gate. No no_run.yaml exposure.

Next: after this PR merges — HPCPullPyAuto to sync the RAL mirrors, then phase B (evidence campaign on the A100s).

[Jammy2211]

Phase B underway (autolens_profiling)

Phase A merged (#679) and RAL mirrors synced (HPCPullPyAuto — diff confirms #677+#679 on the HPC checkouts). autolens_profiling claimed in the task worktree on feature/point-source-defaults-campaign.

Survey findings that adjust the plan:

• The cluster tier is half-ported already: scripts/misc/simulators/cluster.py + dataset/cluster/simple/ (2 dPIE mains + 10 scaling members + NFW host, sources z=1.0/2.0, point_datasets.csv) exist with likelihood_breakdown cells. Phase B adds the missing searches support (dataset_class="cluster" dispatch, factor-graph analysis per the workspace cluster/modeling.py pattern, leaf cells, sweep rows) — cosmology pinned (solver custom_jvp constraint).
• Deviation from the prompt: hpc/sync does not exist in autolens_profiling (it lives in the assistant repos). The repo's documented RAL flow is git-based: commit cells + hpc/batch_gpu/submit_* scripts → branch checkout on the HPC clone → sbatch → commit result JSONs back. Following the in-repo convention instead of porting sync. Submit headers already set JAX_ENABLE_X64=True unconditionally (the sbatch-inheritance gotcha is covered).

Build in progress (delegated, spec-reviewed):

• Framework: truth.json + truth-anchored truth_log_likelihood/delta_max_ll_vs_truth in every point-source result JSON (the campaign's anchor methodology, new for search cells), posterior_stats (per-param mean/std) for the posterior-width experiment, source_plane_tensor (free-centre tensor, Point-source phase A: PairAll log-sum-exp stabilization + free-centre tensor option #679's weighting="jacobian") and image_plane_repeat_solved model types.
• Instrument variants for the discriminator + stress tests: simple_missing (one image dropped), simple_extra (one spurious position), near_caustic (source at 0.95× the tangential caustic).
• 25 A100 submit scripts + a one-shot submit_all_678_phase_b.sh driver covering: exp 1 source-plane flavours (both tiers), exp 2 image-plane solved-vs-free re-run post-logsumexp, exp 3 pairing discriminator (PairAllSolved vs PairRepeatSolved on missing/extra), exp 4 posterior width (from nautilus posteriors), exp 5 near-caustic domain-of-validity.

[Jammy2211]

Phase B discovery: cluster simulators solve every source to the FINAL plane

The new truth-anchor step caught a real defect on its first cluster run. PointSolver.solve(plane_redshift=None) defaults to the tracer's final plane, and both cluster simulators never pass plane_redshift:

• autolens_profiling/scripts/misc/simulators/cluster.py (fixed on this branch)
• autolens_workspace/scripts/cluster/simulator.py:482 (NOT yet fixed — see heads-up on autolens_workspace#436)

Consequence: the z=1.0 source's simulated positions are images of a hypothetical z=2.0 source — unphysical for the source they're attributed to. The z=2.0 source is coincidentally correct (final plane).

Evidence chain:

• Truth anchor at cluster truth: solved logL = −4,230,422.58 — byte-identical to the committed results/breakdown/cluster/source_plane_breakdown_v2026.7.23.1.json, i.e. this is the pre-existing committed state (feat: point-source likelihood variants design (arXiv:2406.15280) #657 phase 3 compared relative timings, so it went unnoticed).
• Per-position: point_0 (z=1.0) back-traces to (−2.79,−5.95), (0.21,0.31), (0.79,7.95) on its own plane — but to (0.3, 0.5) exactly (spread 0.0004″) on the z=2.0 plane. point_1 (z=2.0): perfect.
• After passing plane_redshift per source and regenerating: truth anchor = +26.06 (χ²≈0.017/0.016 per system, solved centres exact). The fit-side plane handling was always correct — only simulated datasets are affected, not fits of real data (a2744/smacs0723 unaffected).

Note: the profiling sim's 0.005″ position noise is a deliberate documented choice (5 mas HST centroiding), not part of this bug.

Implications: the committed cluster breakdown/summary artifacts are tainted as accuracy references (timing conclusions stand); phase D must fold the workspace simulator fix + dataset regeneration into the cluster reconcile (or #436 takes it — human's call).

[Jammy2211]

Phase B: 25 A100 jobs queued on RAL

Branch feature/point-source-defaults-campaign (autolens_profiling ef7da54) checked out on the RAL clone (rebuilt as a real https clone — the old checkout was a dead rsync'd worktree pointer). Datasets pre-generated on the login node with the plane_redshift-fixed simulator (stale tainted cluster dataset wiped first); all 25 campaign cells verified PENDING on the gpu partition (search_*, fp64, x64 forced in every submit).

Cells in flight: galaxy tier ×11 (nautilus + prodigy across image/source × free/solved/tensor/repeat-solved), discriminator ×4 (PairAllSolved vs PairRepeatSolved on simple_missing/simple_extra), near-caustic ×3, cluster tier ×7 (factor-graph, cosmology pinned; image-plane cells on 12 h walls). Jobs mail on END/FAIL; next step when they land: pull result JSONs into results/searches/**, write results/notes/point_source_defaults_campaign.md, and take the exp-3 discriminator verdict into phase C.

[Jammy2211]

Phase B checkpoint (end of day)

22/25 result JSONs written on the RAL clone (feature/point-source-defaults-campaign); remaining cluster long-wall cells run overnight. Zero job failures after the shadow-package fix. Early validation: the first four galaxy cells reproduce the truth-anchor literals exactly (image_plane free 7.20 / solved 7.74, source_plane scalar −33788.4, tensor-solved +0.60) with backend=gpu and posterior_stats populated.

First scientific signal (exp 3): both simple_extra arms — image_plane_solved (PairAllSolved) and image_plane_repeat_solved — timed out at their 2 h walls, versus 147 s for the same solved fit on the clean quad. One spurious extra position makes both pairing likelihoods dramatically harder; the missing-image arms completed normally. The timeout is itself robustness evidence, but both cells will be resubmitted with 8 h walls to obtain posteriors before the discriminator verdict is called.

Known caveat to verify at results-pull: the two free-centre galaxy cells recorded total_wall_s=1 — persistent JAX compile cache or a silent PyAutoFit resume of pre-logsumexp search state (identifiers ignore data). If resumed, their search output gets wiped and re-run — exp 2's post-fix comparison depends on it.

Full resume runbook is in PyAutoMind/active.md (this task's entry). Next session: finish/resubmit cells → pull JSONs (laptop-side pull; RAL cannot push) → results/notes synthesis → phase C keyed on the exp-3 verdict.

[Jammy2211]

Phase B checkpoint (2026-08-01): A100 sweep complete at 22/25 cells; the failing MSP cluster cell is root-caused with two fixes in PR.

Bug — multi_start_prodigy/cluster/image_plane_solved crash (job 331843, masked as sacct COMPLETED): two independent defects, both reproduced and fixed locally on CPU:

1. PyAutoFit#1441 — derived instance attributes (NFWMCRLudlowSph.scale_radius from free mass_at_200) landed in pytree aux data as traced values; the PointSolver custom_jvp rule's inner jax.jvp then received them as stale tracers → UnexpectedTracerError. Fix: flatten promotes any JAX-valued attribute to a dynamic child. Regression script: autofit_workspace_test#81.
2. fix: sanitize padded rows in the PointSolver implicit-diff rule for reverse mode #685 — with (1) fixed, all gradients were NaN: padded solver rows were zeroed to (0,0) = the cluster profile centres, where the NFW deflection Jacobian is NaN; reverse mode row-sums cotangents, so one NaN padded row poisoned every parameter. Fix: anchor padded rows at a real solved image + sanitize the padded solve inputs. After both fixes: 8/8 random cluster draws give finite value and gradient, FOMs byte-identical to before (forward path untouched).

Once merged + HPCPullPyAuto, the cell resubmits on the A100s.

Timeout cells: the two simple_extra exp-3 arms (nautilus image_plane_solved / image_plane_repeat_solved) resubmitted with 8h walls (jobs 331885/331886, in flight).

Early phase-B evidence (details land in results/notes/): exp-3 missing-image arm is decisive for all-to-all — PairAllSolved recovers truth on simple_missing (Δ=+1.3) while PairRepeatSolved is catastrophically biased (truth logL −183389, Δ=+183402). Tensor source-plane ranks truth highest at both tiers and holds at 0.95× the tangential caustic; image-plane free-centre Prodigy still plateaus below truth post-logsumexp (solved converges), and solved posteriors are not overconfident vs free on einstein_radius (σ 0.038 vs 0.027).

[Jammy2211]

Phases C and D are up as PRs:

• feat!: point-source position fit defaults to FitPositionsImagePairAllSolved (#678 phase C) #686 (phase C, library): AnalysisPoint + FitPointDataset default → FitPositionsImagePairAllSolved, with ## API Changes migration notes. Also carries the register_tracer_classes Galaxy-redshift aux fix needed by multi-plane jitted solves. Merge gate: the spurious-extra-position exp-3 arms are rerunning at 8h walls (331885/331886) — hold merge until they land; the missing-image arm already rules out repeat pairing (Δ=+1.8×10⁵ mis-ranking).
• autolens_workspace#453 (phase D, PENDING RELEASE): all point_source + cluster examples demonstrate ps.PointSolved + the solved all-to-all default; free-centre kept as the documented alternative with use-cases (incl. the glSNe standard-candle flux caveat); guides/point_source_pairing.py now carries the full truth-anchored benchmark evidence; cluster/simulator.py per-source plane_redshift bug fixed (the test: mask padding likelihood sanity check #436 taint) and the simulator validated end-to-end. Merge after feat!: point-source position fit defaults to FitPositionsImagePairAllSolved (#678 phase C) #686 ships in a release.

Remaining before closing the campaign: extra-arm JSONs (in flight), the multi_start_prodigy/cluster/image_plane_solved rerun on the merged gradient fixes (331887), and the notes-doc refresh with those final numbers.