[Jammy2211]

Summary

Closes #1713. On analytic_gaussian ensemble seeds 25 and 172 (ruling R-20260910-04), EPMeanField.log_evidence emitted RuntimeWarning: invalid value encountered in add repeatedly and wrote NaN into ep_history.csv from sweep 4 onwards. Diagnosis (reproduced on main, mode projection only): one HierarchicalFactor member sends the scatter sigma an accepted TruncatedNormalMessage with mean −7374 and sigma 384 on [0, 100], so its support sits ~19 sigma into the tail; TruncatedNormalMessage.log_partition computes log(Φ(b) − Φ(a)) directly, which underflows to −inf (true value ≈ −187.9), the variable's evidence becomes +inf, every adjacent factor term −inf, and the sum NaN. The posterior is bit-identical to the committed sidecars; evidence_tol is None by default so stopping is KL-only; only the evidence readout, its plot and the factor log line were affected.

This PR is reporting only: the non-finite term is detected before summing, named once per fit in a logger.warning and in ep_diagnostics.results, and the evidence is returned as an explicit NaN with no numpy warnings; evidence_divergence returns inf on a NaN so an evidence-based stopping rule can never claim convergence on it. No message arithmetic, projection or optimiser changes. The cure of the term itself (a tail-stable log_partition) is message maths and is filed separately as PyAutoMind/draft/bug/autofit/truncated_normal_log_partition_tail_underflow.md for the human's decision.

Workspace regression: autofit_workspace_test (seed-0 legs assert no ep_mean_field RuntimeWarning and finite evidence; new seed-25 mode-projection leg asserts the term is named).

API Changes

No public API change. EPMeanField.log_evidence now returns float("nan") (instead of a NaN produced with numpy warnings) when any term is non-finite and logs the offending variables/factors once per fit; new read-only property EPMeanField.nonfinite_evidence_terms ({label: kind}); EPHistory.evidence_divergence returns inf instead of NaN when an evidence is NaN; ep_diagnostics.results gains an evidence_nonfinite: line when it happens.
See full details below.

Test Plan

• New test_autofit/graphical/functionality/test_evidence_nonfinite.py (7 tests; 6 fail on the previous code, 7 pass now): no RuntimeWarning, one named warning, NaN evidence, healthy evidence unchanged from the old formula, divergence inf on NaN / 0 when healthy, diagnostics line
• pytest test_autofit/graphical -q -x: 329 passed
• Full pytest test_autofit serially: 3890 passed, 1 skipped, 1 xfailed (under -n 4 the suite shows pre-existing xdist races on the shared test_autofit/output/ — unrelated to this change)
• Science repro (scratch copy of analytic_gaussian scripts): seeds 25 and 172 under -W error::RuntimeWarning exit 0; ep_diagnostics.results names variable 'truncatedgaussianprior_387' (+inf, first at step 39) (step 40 on seed 172); posteriors bit-identical to the committed sidecars

Full API Changes (for automation & release notes)

Removed

• none

Added

• autofit.graphical.expectation_propagation.ep_mean_field.EPMeanField.nonfinite_evidence_terms — property, {label: kind} of non-finite evidence terms (empty when finite)
• autofit.graphical.expectation_propagation.diagnostics.check_evidence_nonfinite() — the evidence_nonfinite: results line; EPDiagnostics.evidence_nonfinite records first step and kind per term

Changed

• EPMeanField.log_evidence — explicit float("nan") on any non-finite term, one logger.warning per fit naming it, no numpy RuntimeWarnings; finite path unchanged
• EPHistory.evidence_divergence — inf when the difference is NaN
• EPOptimiser — appends check_evidence_nonfinite to the ep_diagnostics.results warnings

Migration

• none

Generated by the PyAutoLabs agent workflow.

🤖 Generated with Claude Code

[Jammy2211]

Workspace regression: PyAutoLabs/autofit_workspace_test#116 · new-scope prompt (message maths, human decision): PyAutoMind draft/bug/autofit/truncated_normal_log_partition_tail_underflow.md