feat(m8): P2 pgvector + Ollama 向量化与语义检索

- PostgreSQL 切换 pgvector/pgvector:pg17 镜像;迁移 000024 建 vector 扩展、
  knowledge_chunks.embedding vector(1024) + HNSW 余弦索引,retrieval_mode 放宽三态
- OllamaEmbedder 本地 bge-m3 批量嵌入,404 惰性 pull 重试,维度/超时校验,可整体关闭
- SemanticRetriever/HybridRetriever + NewRetriever 按 retrieval_mode 分发,缺 embedder 回退 FTS
- 文档入库同步批量向量化;Ollama 故障降级入库 + embedding_failed 事件
- 修复 pgx CopyFrom 对 vector 列二进制编码误读:COPY 基础列后同事务 unnest 批量回填
- 修复降级路径 embeddings=nil 索引越界 panic(Add 与 Reprocess)
- 知识库列表 vectorized_chunk_count + 前端三态检索模式选择与向量化覆盖率
- 单测 embedder/retrievers + 集成 TestKnowledgeVectorLifecycle 全绿

Co-Authored-By: Claude <noreply@anthropic.com>

deploy/PRODUCTION.md

@@ -1,9 +1,12 @@
 # Production deployment
 
 This bundle builds the Go services and both Art Design Pro applications from
-source. PostgreSQL, two Redis roles and MinIO (object storage, M8) are included;
-ClickHouse is not required. MinIO is not a startup dependency: the gateway only
-warns and refuses file uploads until the bucket is reachable.
+source. PostgreSQL (with pgvector), two Redis roles, MinIO (object storage, M8)
+and a local Ollama container (vectorization, M8) are included; ClickHouse is not
+required. Neither MinIO nor Ollama is a startup dependency: the gateway warns
+and refuses file uploads until the bucket is reachable, and knowledge-base
+documents are still stored (with embedding set to NULL) when Ollama is down,
+with retrieval falling back to full-text search.
 
 ## Prerequisites
 
@@ -73,3 +76,15 @@ it removes persistent data.
 The bundled database URLs use `sslmode=disable` only for the private Compose
 network. When using an external PostgreSQL or Redis service, require TLS and use
 `sslmode=verify-full` / `rediss://` as supported by that service.
+
+### Vectorization and object storage
+
+- The PostgreSQL image is `pgvector/pgvector:pg17` (data-volume compatible with
+  `postgres:17-alpine`); migration `000024` creates the `vector` extension and
+  adds the HNSW embedding column. `EMBEDDING_DIM` must stay at `1024` to match
+  the `vector(1024)` column.
+- Ollama runs locally and lazily pulls `bge-m3` (~1.2 GiB) on first embedding
+  request. Set `EMBEDDINGS_ENABLED=false` to disable vectorization entirely.
+- Back up the `minio-data` and `ollama-models` volumes alongside PostgreSQL.
+- If you previously deployed with `postgres:17-alpine`, back up the PostgreSQL
+  volume before switching images.