feat(v1.3): partial Bayesian-weighted ranking — log-additive (closes #146)

[robotrocketscience]

Summary

Implements issue #146 — partial Bayesian-weighted ranking per the adopted Path B contract in docs/bayesian_ranking.md:

score = log(-bm25_raw) + posterior_weight * log(posterior_mean(α, β))

• Default posterior_weight = 0.5 — synthetic-graph optimum from the v1.3 calibration. Configurable via [retrieval] posterior_weight in .aelfrice.toml and AELFRICE_POSTERIOR_WEIGHT=<float> env var. Negative values clamp to 0.0.
• Byte-identical at posterior_weight = 0.0 — regression-tested against MemoryStore.search_beliefs directly (AC2) and against the size-16 benchmark corpus's 16 queries.
• Reuses scoring.posterior_mean (Jeffreys prior α / (α+β)). The Laplace (α+1)/(α+β+2) form sketched in [retrieval] Posterior-weighted ranking via Beta-Bernoulli prior (log-additive, weight 0.5) #151 is explicitly rejected at this layer per the spec rationale.
• Locks (L0) bypass scoring entirely. L2.5 entity-index hits and L3 BFS expansions are unaffected — the weight only reranks the L1 BM25 candidate set.
• bm25 == 0 (FTS5 non-match) is floored at PARTIAL_BAYESIAN_BM25_FLOOR = 1e-12 so log(0) cannot raise.
• RetrievalCache key gains posterior_weight (rounded to four decimals via POSTERIOR_WEIGHT_KEY_PRECISION) — additive, so existing cached entries with the old key shape are still functional but cannot collide with new-key entries. Cache invalidation is unchanged: apply_feedback → store.update_belief → _fire_invalidation → cache wipe. No new hook in apply_feedback.

Out of scope (per spec § "Out of scope")

10-round MRR uplift eval, ECE calibration scorer, BM25F + heat-kernel composition, real-feedback retest — all deferred to v2.0.0. Issue #151 should be re-scoped to that v2.0 follow-up per the spec's disposition recommendation.

README copy update is deferred to a follow-up docs sweep PR (orchestrator's Phase G).

Atomic commits

1. feat(retrieval): add MemoryStore.search_beliefs_scored returning (belief, bm25) — exposes the FTS5 BM25 score; existing search_beliefs unchanged.
2. feat(scoring): add partial_bayesian_score combining BM25 + posterior log-additively — pure scoring helper.
3. feat(retrieval): wire posterior_weight into retrieve / cache key — retrieve() / retrieve_with_tiers() / retrieve_v2() plumbing + RetrievalCache key extension.
4. test: 22 acceptance tests for partial Bayesian-weighted ranking (#146) — one test per spec acceptance criterion plus calibration + precedence pins.
5. docs: document posterior_weight in CONFIG.md and CHANGELOG.md (#146) — TOML reference + Unreleased entry. (docs/LIMITATIONS.md already carried the v1.3.0 paragraph; docs/ROADMAP.md already linked the spec — no edits needed.)

Test plan

• uv run pytest -q passes (1378 tests; pre-existing test_serve_raises_clear_error_when_fastmcp_missing flake unrelated to this PR is deselected locally — it also fails on a clean main).
• uv run pyright src/aelfrice/scoring.py src/aelfrice/retrieval.py src/aelfrice/store.py tests/test_bayesian_ranking.py — strict-clean (0 errors).
• Byte-identical ordering verified at posterior_weight = 0.0 against MemoryStore.search_beliefs directly and against the size-16 benchmark corpus's 16 queries.
• Calibration regression: rank-3 belief promotes to rank 1 after one apply_feedback(b, valence=+1.0) at default weight 0.5.
• Lock-bypass invariance verified at weights {0.0, 0.5, 1.0}.
• Cache hit/miss matrix verified for posterior_weight axis.
• Cache wipe through store callback (no direct cache.invalidate() from apply_feedback).
• bm25 == 0 non-match path returns [] without raising.
• Full-process CI green across the 8 jobs.