fix(graphical): EP statistics fix batch — F1/F2/F4/F8 from #1332

[Jammy2211]

Overview

Executes the remainder of the #1332 audit's recommended fix batch, now that the #1331 decision hub is resolved and shipped (#1345, #1348) and F3/F10 (#1349), F5 (#1334), F7(c) (#1345) have landed. Prompt: PyAutoMind/bug/autofit/ep_statistics_fix_batch.md.

Plan

• F1 / F7(a) — MeanField.__truediv__ / __pow__ pass log_norm into the plates ctor slot (evidence silently dropped; a float lands in _plates). Fix the ctor calls; fix __pow__'s meaningless log_norm * other.log_norm branch.
• F2 (+extended) — GammaMessage.kl and BetaMessage.kl compute KL(other‖self) while Normal/TruncatedNormal compute KL(self‖other); EPHistory.kl_divergence sums per-variable KLs, so mixed graphs mix directions. Contract: self.kl(other) = KL(self‖other), documented once on AbstractMessage.kl, aligned family-wide, enforced with a numerical property test (sample-based KL estimate per family).
• F4 — AbstractMessage.update_invalid scalar branch is self-flagged broken (the BAD_PROJECTION recovery path). Fix + test both branches.
• F8 — delete dead/suspect quasi-Newton variants in laplace/newton.py (diag_sr1_bfgs_update → returns None; bfgs1_update sign-disputed vs exported bfgs_update; diag_sr1_update_ unused).

Out of scope (parked with notes on #1332): F6 truncated-KL moments, F7(b) factor-evidence recording design, F9 private scipy import.

Validation: full suite; new unit tests per fix; autofit_workspace_test/scripts/graphical/ep_{parity,deterministic,exact}.py as integration smoke.

Branch survey

Repository	Current Branch	Dirty?
./PyAutoFit	main (post #1349)	clean

Suggested branch: feature/ep-statistics-fix-batch. PyAutoFit claim-free (verified).

Autonomy: supervised (--auto chain, human directed "roll straight into the #1332 fix batch" 2026-07-10; heart-ack 6-reason set binds).

🤖 Generated with Claude Code

[Jammy2211]

Shipped at PR-open — #1351

All four findings executed; one verdict flipped honestly: F4 is an exoneration, not a fix — the self-flagged scalar branch is empirically correct (messages iterate over parameters; all-or-nothing swap valid for ndim==0); the stale TODO is replaced and both branches pinned by tests.

Gate: suite 1465/14 with one pre-existing failure (test_full_hierachical — fails on clean main with this batch fully stashed, bisect-proven ×3; filed as bug/autofit/hierarchical_ep_test_flaky.md — notably the #1349 diagnostics artifacts are the right instrument for it) · EP integration smoke 3/3 PASS · KL direction certified per-family by Monte Carlo · Heart YELLOW, identical 6-reason acked set.

Out-of-scope reminders parked on #1332: F6 (truncated KL moments), F7(b) (factor-evidence recording design), F9 (private scipy import). Merge stays human.

[Jammy2211]

Merged — c5c50cca via #1351 (merged-unchanged). Closing.