[Jammy2211]

Summary

Closes #1711. On the analytic_gaussian wave-1 ensemble (ruling R-20260910-04), 19/400 seed-legs ended with a factor that never recorded a SUCCESS update. Diagnosis on the exactly-Gaussian leg A (seed 140): the factor is the mu prior, whose Laplace optimiser met its gradient tolerance at the start point (iter=0) because the posterior mean lies within 2e-3 of the prior mean; the projection was accepted and the message updated (N(50, 22.4) → N(49.9994, 10.0)), and the final posterior matches the closed form to a = 5e-5, b = 2e-5. That is a converged update, yet optimise_quasi_newton reported it with updated=False, the same as a skipped line search. The library's STALE tracking already treated it correctly (it reads the post-projection status), so this PR aligns the Newton status with that and documents the semantics; no EP update maths, projection or Laplace numerics change, and post-fix numerics are bit-identical on the repro seeds.

The second shape in the ruling (HierarchicalFactor with only BAD_PROJECTION, leg-B seeds 29/71/106/152/178/193) is the Laplace-on-scatter boundary state (#1405) and is cured by LaplaceOptimiser(projection="moments") (#1656): seeds 29 and 71 re-run under moments give HierarchicalFactor0 {SUCCESS: 28} and {SUCCESS: 23} with no STALE warning. Two findings, two causes; only the first needed a library change.

The ensemble witness's criterion 6 ("zero SUCCESS = stale") is what conflated the two; the amended wording (SUCCESS or NO_CHANGE counts as a completed update, BAD_PROJECTION-only stays a miss) goes to the Cortex task and ensemble_seed.py separately. The seed-140 regression leg lands in autofit_workspace_test (scripts/graphical/analytic_gaussian.py).

API Changes

No public API change. optimise_quasi_newton now returns Status(updated=True, flag=StatusFlag.NO_CHANGE) when a stop condition is met before the first step with success=True (previously updated=False). NO_CHANGE rows in ep_history.csv were already written with updated=True after projection, so recorded histories do not change.
See full details below.

Test Plan

• New test_zero_iteration_convergence_is_reported_as_an_update (red before, green after) and test_factor_at_its_mode_on_first_visit_is_not_stale (guard: NO_CHANGE rows with updated=True, no STALE in log or ep_diagnostics.results, mean and width match closed form) in test_autofit/graphical/functionality/test_factor_failure_recovery.py
• pytest test_autofit/graphical -q -x: 322 passed
• Full pytest test_autofit -n 4: 3561 passed, 1 skipped, 1 xfailed
• Science repro (scratch copy of analytic_gaussian scripts, worktree PyAutoFit): seed 140 leg A and leg-B seeds 29/38/71 under mode and moments; sidecar numerics identical before and after the change

Full API Changes (for automation & release notes)

Removed

• none

Added

• none

Changed

• autofit.graphical.laplace.newton.optimise_quasi_newton — a stop condition satisfied at i == 0 with success=True now yields Status(updated=True, flag=NO_CHANGE) instead of updated=False.
• autofit.graphical.utils.StatusFlag.NO_CHANGE — documented as a converged update (start point already at the mode; projection still runs and may move the message), not a skip.
• autofit.graphical.expectation_propagation.optimiser.EPOptimiser._stale_factor_warnings — docstring: "updated" is not the same as "flagged SUCCESS"; no logic change.

Migration

• none

Generated by the PyAutoLabs agent workflow.

🤖 Generated with Claude Code

[Jammy2211]

Workspace regression: PyAutoLabs/autofit_workspace_test#115 · Cortex criterion-6 amendment: PyAutoLabs/PyAutoCortex#75 · Insight ledger reconciliation: PyAutoLabs/PyAutoInsight#38