Fix live multi-turn: pass the resolved drug, stop did-you-mean garbage

apps/ai-service/rag/ports.py

@@ -26,6 +26,27 @@ class AnswerGenerationUnavailable(RuntimeError):
     """
 
 
+class RerankUnavailable(RuntimeError):
+    """The reranker could not be reached.
+
+    Same fail-open contract as the other provider errors, but softer: losing the
+    reranker only means the candidates keep their original order, so the caller
+    catches this and proceeds rather than abstaining. Rerank is an ordering
+    improvement on the fallback, never a precondition for an answer.
+    """
+
+
+class Reranker(Protocol):
+    """Reorders candidate texts by joint relevance to the query.
+
+    Returns indices into `documents`, most relevant first. A cross-encoder pass
+    that recovers precision the bi-encoder embedding cannot; applied only to the
+    similarity/overview fallback, never to the deterministic section route.
+    """
+
+    def rerank(self, query: str, documents: list[str], top_n: int | None = None) -> list[int]: ...
+
+
 class AnswerGenerator(Protocol):
     """Rewrites retrieved evidence into prose. Never a source of facts.