Wire up query history: localStorage session persistence + sidebar UI

apps/ai-service/bootstrap.py

@@ -137,7 +137,7 @@ def build_runtime(settings: Settings):
     effective_metrics = metrics or NullMetrics()
     configure_telemetry(settings, effective_metrics)
     if settings.embedding_provider == "disabled":
-        return None, None, PostgresTraceRepository(settings.postgres_dsn), metrics
+        return None, None, PostgresTraceRepository(settings.postgres_dsn), metrics, None
     if settings.embedding_provider != "cohere-v4":
         raise ValueError(
             "No production query embedder is configured. Supported values: "
@@ -169,8 +169,13 @@ def build_runtime(settings: Settings):
     # drug identity through `LlmQueryUnderstander` against the same catalog.
     resolver = CatalogDrugResolver(aliases)
     reranker = _build_reranker(settings)
+    # Returned on its own below (Feature-List #4/#23's section-list and
+    # verbatim-section-text endpoints) — both are plain payload-filtered
+    # Qdrant reads with no LLM/generation involved, so they read straight
+    # from this adapter rather than through `RetrievalService`/`RagAgent`.
+    qdrant_retriever = QdrantRetriever(client, settings.qdrant_collection, embedder)
     retrieval = InstrumentedRetrievalService(
-        QdrantRetriever(client, settings.qdrant_collection, embedder),
+        qdrant_retriever,
         QdrantParentStore(client, settings.qdrant_collection),
         EvidencePolicy(minimum_score=settings.evidence_minimum_score),
         section_resolver=section_resolver,
@@ -195,7 +200,7 @@ def build_runtime(settings: Settings):
         # understanding either, so there is no conversational/agent
         # capability to offer. Answer-only (retrieval-verified, no
         # generation) still works through `answers` directly.
-        return answers, None, trace_writer, metrics
+        return answers, None, trace_writer, metrics, qdrant_retriever
     agent = InstrumentedRagAgent(
         understander=InstrumentedQueryUnderstander(
             LlmQueryUnderstander(generator, _catalog_names(aliases), resolver)
@@ -208,4 +213,4 @@ def build_runtime(settings: Settings):
         store=PostgresConversationStore(settings.postgres_dsn),
         metrics=effective_metrics,
     )
-    return answers, agent, trace_writer, metrics
+    return answers, agent, trace_writer, metrics, qdrant_retriever