feat: 接入 BGE-M3 本地嵌入默认配置 + 新增 Qwen3-Reranker 重排

- .env.example 嵌入默认改为本地 BGE-M3（EMBEDDING_PROVIDER=local, dim=1024），NAS 全新部署即用
- 新增 app/services/reranker.py：基于 Ollama /api/rerank 的 cross-encoder 重排服务
- retriever 在 chunk 候选阶段接入语义精排，调用失败优雅降级为 RRF 顺序
- docker-compose 拉取 qwen3-reranker:0.6b，OLLAMA_MAX_LOADED_MODELS 提至 3
- .gitignore 排除 .workbuddy/ 与 .trae/（防止误提交项目记忆与 IDE spec）
- 新增 reranker 单元与集成测试，全量测试 529 passed

tests/test_retriever.py

@@ -82,15 +82,29 @@ def _route(
     )
 
 
-def _make_retriever(qdrant: FakeQdrant, route: RouteDecision) -> Retriever:
+def _make_retriever(qdrant: FakeQdrant, route: RouteDecision, reranker: Any = None) -> Retriever:
     return Retriever(
         qdrant=qdrant,  # type: ignore[arg-type]
         query_parser=FakeParser(route),  # type: ignore[arg-type]
         embedding=FakeEmbedding(),
         sparse_encoder=SparseEncoder(),
+        reranker=reranker,
     )
 
 
+class FakeReranker:
+    """假 RerankerService：返回预设分数或抛出预设异常"""
+
+    def __init__(self, scores: list[float] | None = None, exc: Exception | None = None) -> None:
+        self.scores = scores if scores is not None else []
+        self.exc = exc
+
+    async def rerank(self, query: str, documents: list[str]) -> list[float]:
+        if self.exc is not None:
+            raise self.exc
+        return self.scores
+
+
 def _calls(qdrant: FakeQdrant, collection: str) -> list[dict[str, Any]]:
     return [c for c in qdrant.calls if c["collection"] == collection]
 
@@ -348,3 +362,52 @@ class TestSearchApi:
         resp = client.get("/api/v1/health")
         assert resp.status_code == 200
         assert resp.json() == {"status": "ok"}
+
+
+class TestReranker:
+    """重排启用时对 chunk 候选精排；关闭或失败则退化为 RRF 顺序"""
+
+    @staticmethod
+    def _full_qdrant() -> FakeQdrant:
+        return FakeQdrant(
+            {
+                COLLECTION_L1: [[_point("l1a", "d1")]],
+                COLLECTION_L2: [[_point("l2a", "d1", "章节A")]],
+                COLLECTION_L3: [[_point("l3a", "d1", "章节A")]],
+                COLLECTION_CHUNKS: [
+                    [
+                        _point("c1", "d1", "章节A"),
+                        _point("c2", "d1", "章节A"),
+                        _point("c3", "d1", "章节A"),
+                    ]
+                ],
+            }
+        )
+
+    async def test_enabled_reorders_by_relevance(self):
+        """重排按相关性分数降序重排，并将 score 写回相关性分数"""
+        reranker = FakeReranker(scores=[0.1, 0.9, 0.5])  # c1<c3<c2
+        retriever = _make_retriever(self._full_qdrant(), _route(), reranker=reranker)
+
+        resp = await retriever.search(SearchRequest(query="测试查询"))
+
+        # 重排后顺序为 c2, c3, c1
+        assert [h.text for h in resp.hits] == ["text-c2", "text-c3", "text-c1"]
+        assert resp.hits[0].score == 0.9
+
+    async def test_disabled_keeps_rrf_order(self):
+        """重排未注入（关闭）时保持 RRF 融合原始顺序"""
+        retriever = _make_retriever(self._full_qdrant(), _route(), reranker=None)
+
+        resp = await retriever.search(SearchRequest(query="测试查询"))
+
+        assert [h.text for h in resp.hits] == ["text-c1", "text-c2", "text-c3"]
+
+    async def test_failure_falls_back_to_rrf(self):
+        """重排调用异常时退化为 RRF 顺序，检索不中断"""
+        reranker = FakeReranker(exc=RuntimeError("ollama 不可用"))
+        retriever = _make_retriever(self._full_qdrant(), _route(), reranker=reranker)
+
+        resp = await retriever.search(SearchRequest(query="测试查询"))
+
+        assert [h.text for h in resp.hits] == ["text-c1", "text-c2", "text-c3"]