fix(graphical): EP hierarchical factor returns -inf for out-of-support scale (nightly crash)

[Jammy2211]

Summary

The EP guide (scripts/guides/modeling/advanced/expectation_propagation.py) crashed intermittently in the nightly workspace-validation with NormalMessage sigma cannot be negative. Root cause (pinned by deterministic reproduction): its HierarchicalFactor draws mass.slope from a GaussianPrior(mean, sigma) whose sigma is a truncated hyper-prior (lower_limit=0). The hierarchical factor's Laplace/Newton optimiser evaluates at proposed scales that dip just below 0 (the scale posterior sits at mean≈0.0005, hard against the boundary), so Factor.__call__ built GaussianPrior(sigma<0) and tripped the strict NormalMessage sigma guard.

A distribution parameterised outside its support has zero probability density, so the factor's log-value is -inf — not a hard error. Factor.__call__ now catches exc.MessageException and returns -inf. The strict guard is unchanged (still fires for genuine prior-passing misuse); its stale RelativeWidthModifier-only hint is corrected. Adds a deterministic regression test. The workspace guide is untouched (no autoimmunity — no seeding, no prior constraints).

Fixes #1363

API Changes

None — internal changes only. Behavioural: a HierarchicalFactor whose optimiser proposes an out-of-support distribution scale now yields a zero-density (-inf) factor value instead of raising MessageException. No public symbols added/removed/renamed and no signatures changed.
See full details below.

Test Plan

• Full test_autofit/ suite: 1473 passed, 1 skipped (in-worktree).
• New regression test test_autofit/graphical/hierarchical/test_negative_scale.py: the exact CI scales (−0.016, −0.17) return -inf; a valid scale returns finite; strict NormalMessage(sigma<0) still raises.
• test_autofit/graphical/ in-order: 218 passed (EP-optimisation locus green).
• EP guide in the nightly's exact mode (PYAUTO_TEST_MODE=1 PYAUTO_SMALL_DATASETS=1 PYAUTO_FAST_PLOTS=1): base crashes (sigma=-0.07… via graphical.laplace.newton); with the fix the guide runs to completion (exit 0, full MeanField output).

Full API Changes (for automation & release notes)

Changed Behaviour

• autofit.graphical.declarative.factor.hierarchical.Factor.__call__ — returns -inf (zero density) when the distribution is parameterised outside its support (raising exc.MessageException), instead of propagating the exception. Internal to the EP hierarchical-factor machinery.
• autofit.messages.normal.assert_sigma_non_negative — exception message text corrected (the RelativeWidthModifier path now uses abs(mean), so it is no longer the likely source); the sigma<0 rejection itself is unchanged.

Migration

• None.

Validation checklist (--auto run — plan was not pre-approved)

• Effective level: supervised (header: safe, cap: bug → supervised)
• Plan: on the issue (fix(graphical): EP builds strict NormalMessage from negative-variance cavity → sigma<0 crash #1363), written at start, unmodified since
• Gate: tests 1473p / 1s · smoke base-crashes / fix-clean in the nightly's exact test-mode · review CLEAN · Heart YELLOW (stale test_run 2026-07-09 + 58 stale parked scripts; human-acknowledged at the ship checkpoint as unrelated to this branch)
• Human: plan sound in hindsight?
• Human: diff matches plan (no scope creep)?
• Human: merge, amend, or reject — then log the outcome

Generated by the PyAutoLabs agent workflow.