fix(graphical): EP statistics completion — F6 truncated KL + F7(b) evidence recording

[Jammy2211]

Overview

The last two statistical legs from the #1332 audit, now that F1-F5/F7a/F7c/F8/F10 have landed (#1345/#1348/#1334/#1349/#1351):

• F6 — TruncatedNormalMessage.kl uses the untruncated Gaussian formula; the error grows as posterior mass approaches the bounds. Live in HowToFit ch3 tutorial 5's EP convergence metric (af.TruncatedGaussianPrior). Fix: compute KL from truncated moments (mean/variance of the truncated normal via the standard hazard-function expressions), with the exact same-support case first and a documented approximation policy for mismatched supports.
• F7(b) — MeanField.from_priors (the projection behind every search-driven factor update) constructs with log_norm = 0, so the sampler's per-factor evidence never enters the mean field and EPMeanField.log_evidence cannot be trusted for model comparison in sampler-driven fits. Design decision + implementation: record the search's log-evidence in the projected mean field's log_norm (recommendation to be posted after code inspection; README §5 documents the intended evidence algebra).
• Plus the README §6 sharpening from the research: EP review phase 5 — reconcile the three deterministic-variable mechanisms #1336 Recommendation-A adoption (decision recorded on research: EP review phase 5 — reconcile the three deterministic-variable mechanisms #1336, closed).

Validation: full suite; MC-vs-analytic KL tests for the truncated family (extending the #1351 direction property test); an evidence round-trip test (EP log_evidence vs analytic on the exact-conjugate toy); EP integration smoke (parity/deterministic/exact).

Autonomy: supervised (--auto chain, human directed 2026-07-10). Suggested branch feature/ep-statistics-completion.

[Jammy2211]

Shipped at PR-open — #1354

F6 (exact truncated KL, MC-verified ×3 geometries) + F7(b) (sampler evidence recorded on log_norm per the README §5 contract — all three F7 legs now fixed, evidence trustworthy with nested factor searches) + the #1336 Rec-A doc in §6. Suite 1474/14 with only the documented #1352 order-dependent failure; EP smoke 3/3 (one stale-resume-cache false alarm cleared); test_static regression caught in-suite and fixed (two-level getattr). Heart YELLOW, same acked set. Merge stays human.

[Jammy2211]

Merged — ffc04d83 via #1354. The #1332 audit's statistical findings are now fully dispersed: F1-F8 + F10 fixed/exonerated across #1345/#1348/#1334/#1349/#1351/#1354; F9 (scipy private import) remains a small chore; F7 evidence bookkeeping complete. Closing.