fix: EP projection raises ProjectionException on non-finite sufficient statistics; factor_step recovers (#1653)

[Jammy2211]

Summary

AbstractMessage.project guarded its importance-weighted sufficient statistics with a bare assert np.isfinite(suff_stats).all(). One factor whose samples or log weights went non-finite (nan/inf samples, nan/+inf log weights, all weights zero, or overflow of T(x)·w) therefore raised a context-free AssertionError. That error escaped EP's factor_step recovery tuple and killed the whole graph fit. Under python -O the check vanished and the non-finite message went through silently.

This PR replaces the assert with a typed autofit.exc.ProjectionException(ValueError). The message says which input caused the failure, and the prior id and model parameter path are added on the way up. EP's factor_step now recovers from it the way it already recovers from InitializerException: the factor keeps its previous message for that sweep with StatusFlag.EXCEPTION, and max_consecutive_failures still aborts a factor that keeps failing.

API Changes

• New exception autofit.exc.ProjectionException, a subclass of ValueError.
• AbstractMessage.project, Prior.project and Result.projected_model raise it where the bare assert used to fire. The message names the failing input (non-finite samples / nan or +inf log weights / all -inf log weights / weighted-moment overflow), the prior id (Prior.project) and the parameter path (Result.projected_model). The check also survives python -O.
• EP factor_step lists it in its recovery tuple, so one bad projection becomes a failed sweep update, not a failed graph fit.
• Behaviour change: non-EP callers of Result.projected_model now get a ValueError subclass instead of AssertionError. Nothing in autofit/ catches AssertionError.
• No removals, no signature changes.
  See full details below.

Test Plan

• Red witness on the unmodified source: 10 failed, all AssertionError at autofit/messages/abstract.py:316. That covers 6 parametrised direct cases, 2 prior-id cases, the factor_step / EPOptimiser.run recovery tests and the projected_model path case.
• New/extended tests green: test_autofit/messages/test_project_nonfinite.py + test_autofit/graphical/functionality/test_factor_failure_recovery.py give 29 passed; test_autofit/graphical/test_unification.py gives 34 passed.
• Full suite after rebase on origin/main: python -m pytest test_autofit -x, 2946 passed, 2 skipped.
• Workspace smoke (smoke profile: PYAUTO_TEST_MODE=2, PYAUTO_DISABLE_JAX=1) against this branch, autofit_workspace_test: graphical/analytic_ep_minimal.py PASS (MINIMAL EP SELF-TESTS: PASS, 2 s). graphical/analytic_gaussian_collapse.py PASS (COLLAPSE CONFIG: PASS 5/5 seeds, 18 s).

Full API Changes (for automation & release notes)

Added

• autofit.exc.ProjectionException(ValueError): non-finite sufficient statistics in an importance-weighted message projection. Deliberately not a MessageException (callers turn those into silent reverts or -inf likelihoods) and not a SearchException (misconfigured search).
• AbstractMessage._nonfinite_projection_reason(samples, log_weight_list, suff_stats) (private classmethod). Called only on the failure path, it reports the first cause in propagation order.

Changed Behaviour

• autofit.messages.abstract.AbstractMessage.project: raises ProjectionException instead of assert on non-finite sufficient statistics.
• autofit.mapper.prior.abstract.Prior.project: re-raises ProjectionException with prior id=<id> (<PriorClass>) context (from e).
• autofit.non_linear.result.Result.projected_model: re-raises ProjectionException with path=<a.b.c> context (from e). Callers get ValueError rather than AssertionError.
• autofit.graphical.expectation_propagation.optimiser.factor_step: exc.ProjectionException added to the recovery tuple. The result is the previous message, a StatusFlag.EXCEPTION status, and a logged exception.

Migration

• None. Code that caught AssertionError from projected_model / project should catch ValueError (or autofit.exc.ProjectionException). No such caller exists in autofit/.

Generated by the PyAutoLabs agent workflow.

Heart RED override (development only)

• Authorised by the live human in the Claude Code session at 2026-09-30 ~11:50 BST. Question: "Do you authorise the Heart RED development override for push + pending-release PR-open on PyAutoFit#1653, and, when its implementation lands green, on PyAutoFit#1654 (moment projection)? No merge, no release; recorded in the four sinks." Answer: "Yes, both fix: EP projection raises ProjectionException on non-finite stats #1653 and feat: LaplaceOptimiser(projection="moments") for hierarchical scatter #1654 (Recommended)".
• Exact RED reasons (readiness --json, ts 2026-09-30T09:51:49Z, re-read ~11:45 BST unchanged): "PyAutoArray: 2 commit(s) behind origin"; "PyAutoLens: 2 commit(s) behind origin"; "release validation FAILED (stage integrate)". Yellow: "workspace validation not passing (0 failed, 1 timeout, cloud#36404726969: autolens_test scripts/multi_dataset/rectangular.py)"; "manifest drift: public front-door organ tables (generated) — 1 mismatch(es) vs PyAutoMind/repos.yaml".
• Branch gates passed: rebase on origin/main clean; pytest test_autofit -x 2946 passed, 2 skipped; red witness confirmed on the old source; graphical smoke 2/2 PASS.
• This PR does not claim to fix Heart. Heart stays RED for release purposes. Merge needs its own human command (/prm) and green required checks.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Up1fQmjtGAh4RhLrWHfcfg