feat(retrieval): add voyage rerank-2 cross-encoder stage (feature flag)

Stage B of voyage-upgrades-plan rewritten: instead of context-3 (which
4 POCs showed inconsistent improvement), add a cross-encoder rerank
layer on top of voyage-3. Default off (VOYAGE_RERANK_ENABLED=false).

POC validation (785-doc corpus, 12 queries, claude-haiku-4-5 judge):
- mean@3 +4.5% (4.306 → 4.500)
- practical-category queries +11.6% (3.78 → 4.22)
- latency +702ms per query
- no schema change, no re-embed, no double storage

Plumbing:
- config: VOYAGE_RERANK_ENABLED / _MODEL / _FETCH_K env vars
- embeddings.voyage_rerank() wraps voyageai client.rerank
- services/rerank.py: maybe_rerank() helper — fetches FETCH_K candidates
  via the bi-encoder then reranks to top-K. Fail-open if Voyage rerank is
  unavailable.
- tools/search.py: search_decisions, search_case_documents,
  find_similar_cases all wrapped
- services/precedent_library.search_library wrapped

Smoke-tested locally with flag on/off — produces expected behaviour and
latency profile. Ready for production rollout via Coolify env flip after
deploy.

POCs (kept under scripts/ for reference):
- voyage_context3_poc{_long}.py — context-3 evaluation (rejected)
- voyage_multimodal_poc.py — multimodal-3 (stage C, deferred)
- voyage_rerank_judge_poc.py — single-case rerank benchmark
- voyage_rerank_corpus_poc.py — full-corpus rerank validation

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

mcp-server/src/legal_mcp/services/embeddings.py

@@ -53,3 +53,26 @@ async def embed_query(query: str) -> list[float]:
     """Embed a single search query."""
     results = await embed_texts([query], input_type="query")
     return results[0]
+
+
+async def voyage_rerank(
+    query: str, documents: list[str], top_k: int | None = None,
+) -> list[tuple[int, float]]:
+    """Cross-encoder rerank via Voyage. Returns [(orig_index, score), ...]
+    sorted by relevance. Each tuple's index refers to the position in the
+    *input* documents list (not a DB row id) — caller maps it back.
+
+    Used as a second stage after bi-encoder retrieval: fetch top-N
+    candidates with cosine, then rerank to get top-K with cross-encoder
+    attention over (query, doc).
+    """
+    if not documents:
+        return []
+    client = _get_client()
+    result = client.rerank(
+        query=query,
+        documents=documents,
+        model=config.VOYAGE_RERANK_MODEL,
+        top_k=top_k,
+    )
+    return [(r.index, float(r.relevance_score)) for r in result.results]