fix: Result.instance falls back to a valid sample; persist samples before materialisation

[Jammy2211]

Overview

Six of seven finished Phase 8B MultiStartProdigy arms on the RAL A100 (autolens_profiling jobs 341874/341875, 1.8–4.0 h each) completed all 3000 steps and then died in analysis.save_results with SamplesException ← ModelParameterException: ell_comps must satisfy e0**2+e1**2 < 1. The best lane sits outside the ellipticity unit disk (box prior; validate_ell_comps is a no-op on JAX tracers), and the rejection surfaces only on host materialisation. Because updater._save_samples returns early on the exception, the run also loses samples.csv and .completed — the mechanism #1487 suspects. Companion PyAutoLens change widens the save_results catch.

Plan

• Result.instance: on SamplesException from the summary path, fall back to Samples.max_log_likelihood() (the recovering path added by fix(autofit): reconstructing a stored sample raises through ignore_assertions=True #1486); WARNING log.
• updater._save_samples: write samples.csv / summary before instance materialisation so a rejected best point never costs a run its data (also lets the weight-threshold prune run — Weight-threshold prune does not remove zero-weight samples with checks enabled #1487).
• PyAutoLens save_results / save_results_combined: catch af.exc.SamplesException as well as AttributeError.
• Tests for both behaviours; no change to validate_ell_comps tracer behaviour (the physics gap is a separate follow-up).
• After merge: HPCPullPyAuto on RAL so the pending 8B arms write normally.

Detailed implementation plan

Work Classification

Library (PyAutoFit primary, PyAutoLens companion) — library-first gate.

Affected Repositories

• PyAutoFit (primary)
• PyAutoLens

Branch Survey

Repository	Current Branch	Dirty?
./PyAutoFit	main	clean
./PyAutoLens	main	clean

Suggested branch: feature/result-instance-fallback
Worktree root: ~/Code/PyAutoLabs-wt/result-instance-fallback/

Implementation Steps

1. autofit/non_linear/result.py:116-121 — Result.instance: try: self.samples_summary.instance except exc.SamplesException: log + return self.samples.max_log_likelihood().
2. autofit/non_linear/search/updater.py:211-224 — reorder: save_samples (and a summary without the instance) before the samples_summary.instance attempt; keep the exception handling for the instance-dependent outputs only.
3. autofit/non_linear/samples/summary.py:58 — leave the raising policy (single sample); document why.
4. autolens/analysis/analysis/dataset.py:186-191 (+ save_results_combined) — except (AttributeError, af.exc.SamplesException).
5. Tests: PyAutoFit — model whose instance_from_vector rejects the stored best vector → Result.instance returns best valid sample; samples.csv exists; .completed written via a stub search. PyAutoLens — save_results survives a raising result.

Key Files

• autofit/non_linear/result.py, autofit/non_linear/search/updater.py, autofit/non_linear/samples/{summary,interface,samples}.py
• autolens/analysis/analysis/dataset.py
• Root-cause evidence: RAL logs error.341875_*.err; audit https://claude.ai/code/artifact/d9f4b0f3-52a1-4830-a9ad-11a225b77507

Original Prompt

Click to expand starting prompt

@PyAutoFit Bug: a rejected best point kills the run at results-write and loses samples.csv

Type: bug
Target: autofit
Difficulty: small
Autonomy: supervised
Priority: high
Status: formalised
Issued: 2026-08-27

Original request (verbatim): "Look at the JAX profiling gradinet epic, download results
overnight from A100, and do a major assessment of the results so far are, maybe double
checking some conclusiojns." → audit → "yep do all that" (ranked action 1).

Problem

Six of seven finished Phase 8B MultiStartProdigy arms on the RAL A100 (autolens_profiling
jobs 341874/341875, 1.8–4.0 h each) completed 3000/3000 steps and then died inside
analysis.save_results with

autofit.exc.SamplesException: The stored parameters returned by max_log_likelihood
cannot be reconstructed as a model instance because the current model rejected them
← autogalaxy.exc.ModelParameterException: ell_comps must satisfy e0**2+e1**2 < 1

The best lane sits outside the ellipticity unit disk (the ell_comps prior is an
independent TruncatedGaussian(-1,1) box; validate_ell_comps is a no-op on JAX tracers,
so the jitted objective is finite there). The rejection only surfaces on host
materialisation, and then:

1. @PyAutoFit/autofit/non_linear/result.py:116-121 Result.instance goes through
   SamplesSummary.max_log_likelihood (samples/summary.py:58 → samples/interface.py:122,
   raising policy). The summary holds ONE sample so it cannot "next_valid"; the recovery
   added by fix(autofit): reconstructing a stored sample raises through ignore_assertions=True #1486 lives only on Samples.max_log_likelihood (samples/samples.py:378-413)
   and never applies to the path that runs. (This is the open Weight-threshold prune does not remove zero-weight samples with checks enabled #1487 suspect.)
2. @PyAutoFit/autofit/non_linear/search/updater.py:211-224 catches the exception and
   returns early — save_samples_summary and save_samples are SKIPPED, so the run has
   no samples.csv, no .completed, and the only surviving artifact is
   search_internal.dill (which the crash coincidentally spares from deletion).
3. @PyAutoLens/autolens/analysis/analysis/dataset.py:186-191 save_results catches
   only AttributeError, so the exception escapes start_resume_fit
   (abstract_search.py:777) before paths.completed() at :781.

Fix (library-first; this prompt covers PyAutoFit + the PyAutoLens catch)

• Result.instance: try the summary; on SamplesException fall back to
  self.samples.max_log_likelihood() (the recovering path). Log at WARNING.
• updater._save_samples: persist samples (and a summary marked invalid / without
  instance) BEFORE instance materialisation, so a rejected best point never costs the
  run its data. This also removes the early return that Weight-threshold prune does not remove zero-weight samples with checks enabled #1487 identifies as the reason
  the weight-threshold prune never runs.
• PyAutoLens save_results / save_results_combined: also catch
  af.exc.SamplesException (belt-and-braces; never let a tracer/plot failure kill a fit).
• Tests: a model whose stored best vector is rejected by instance_from_vector →
  Result.instance returns the best VALID sample; samples.csv is written; no raise.
• Do NOT make validate_ell_comps fire on tracers. The physics gap (joint disk
  constraint / reparameterisation) is a separate Phase-3 follow-up prompt.

After merge: HPCPullPyAuto on RAL so the ~22 pending 8B arms write normally.
Audit: https://claude.ai/code/artifact/d9f4b0f3-52a1-4830-a9ad-11a225b77507

[Jammy2211]

Library PRs Created — 2026-08-27

• PyAutoFit: fix: Result.instance falls back to a valid sample; persist samples before materialisation (#1535) #1536 (Result.instance fallback + samples written before materialisation; also closes the Weight-threshold prune does not remove zero-weight samples with checks enabled #1487 early-return)
• PyAutoLens: fix: save_results survives a rejected max-likelihood sample (PyAutoFit#1535) PyAutoLens#713 (widened save_results catch)

Shipped under the human-authorised Heart RED override for release validation FAILED (stage integrate) (nightly 33073386315, unrelated scripts). Companion results PR: PyAutoLabs/autolens_profiling#182.

API changes: none removed/renamed; Result.instance no longer raises on a rejected best point (WARNING + next-valid sample); _save_samples always writes samples.csv / samples_summary.json.

Follow-up (not in this PR): PyAutoGalaxy autogalaxy/analysis/analysis/dataset.py:178 has the identical except AttributeError-only catch (shadowed for lens fits).

Next: merge (human) → HPCPullPyAuto on RAL so pending Phase 8B arms 341874/341875 write normally.

[Jammy2211]

Shipped — 2026-08-27

PRs

• fix: Result.instance falls back to a valid sample; persist samples before materialisation (#1535) #1536
• fix: save_results survives a rejected max-likelihood sample (PyAutoFit#1535) PyAutoLens#713

Summary

Merged after green CI on every leg (library-first). Human-authorised Heart RED override for release validation FAILED (stage integrate) (unrelated nightly failures). Companion issue: PyAutoLabs/autolens_profiling#182.

Session Notes

Audit: https://claude.ai/code/artifact/d9f4b0f3-52a1-4830-a9ad-11a225b77507. RAL mirror pulled after merge. Follow-ups: PyAutoGalaxy dataset.py:178 same unguarded catch; F2 reference ruling before any 8B verdict; ell_comps joint disk constraint (Mind draft feature/autogalaxy/ell_comps_joint_disk_constraint.md).