feat(api): 收敛对话与记忆流程边界，引入 LLM 网关与专用服务

- MemoryService 异步路径委托 MemoryIngestService / MemoryRetrievalService；富化派发经 MemoryEnrichmentScheduler
- WebSocket pipeline 经 ChatTurnService 与显式 DTO 编排单轮对话；回忆录片段入队由 MemoirIngestScheduler 封装
- 新增 LlmGateway（LlmUseCase），各 agent、任务与适配器对齐 ports
- 补充 memory 提示适配、runtime 类型、memory-retrieval 文档、ai-touchpoints 说明与扫描脚本及配套测试

Made-with: Cursor

api/app/agents/chat/orchestrator.py

@@ -4,6 +4,7 @@ ChatOrchestrator：AI 回复用户模块的编排层
 """
 
 import time
+from collections.abc import Callable
 from datetime import datetime
 from typing import TYPE_CHECKING, List, Optional
 
@@ -24,7 +25,8 @@ from app.agents.chat.stage_detection import (
 from app.agents.state_schema import MemoirStateSchema
 from app.core.agent_logging import agent_summary_enabled, log_agent_detail
 from app.core.config import settings
-from app.core.dependencies import get_llm_provider
+from app.core.dependencies import get_embedding_provider
+from app.core.llm_gateway import LlmGateway
 from app.core.logging import get_logger
 from app.features.conversation.input_normalize import normalize_chat_input_for_agent
 from app.features.memoir.state_service import (
@@ -32,18 +34,20 @@ from app.features.memoir.state_service import (
     save_interview_state_meta,
     switch_stage,
 )
+from app.features.memory.prompt_adapter import MemoryPromptAdapter
 
 
 def _llm_for_chat_input_normalize():
     try:
-        p = get_llm_provider()
-        return getattr(p, "langchain_llm", None)
+        return LlmGateway().langchain_llm_for()
     except Exception:
         return None
 
 
 if TYPE_CHECKING:
     from app.features.user.models import User
+    from app.ports.embedding import EmbeddingProvider
+    from app.ports.llm import LLMProvider
 
 logger = get_logger(__name__)
 
@@ -56,9 +60,10 @@ async def _fetch_interview_memory_bundle(
     db: AsyncSession,
     user_id: str,
     user_message: str,
+    *,
+    get_embedding_provider_fn: Callable[[], "EmbeddingProvider"],
 ) -> tuple[dict | None, object | None]:
-    """检索记忆 bundle（原始结构）；是否进主 prompt 由 `slice_interview_memory` 再筛。"""
-    from app.core.dependencies import get_embedding_provider
+    """检索记忆 bundle（原始结构）；是否进主 prompt 由 adapter 再筛。"""
     from app.features.memory.retrieval_trace import (
         chat_memory_retrieval_trace_from_bundle,
     )
@@ -76,7 +81,7 @@ async def _fetch_interview_memory_bundle(
         )
         return None, None
     try:
-        emb = get_embedding_provider()
+        emb = get_embedding_provider_fn()
         ms = MemoryService(db, embedding_provider=emb)
         top_k = settings.chat_memory_top_k
         bundle = await ms.retrieve(user_id, msg, top_k=top_k)
@@ -103,11 +108,22 @@ class ChatOrchestrator:
     """
     聊天编排器：根据用户资料完成度路由到 ProfileAgent 或 InterviewAgent。
     不直接写入 Redis/DB；由 WS pipeline / ConversationHistoryStore 落库并同步缓存。
+
+    ``get_embedding_provider_fn`` / ``llm_provider`` 供测试或脚本注入；默认使用全局依赖。
     """
 
-    def __init__(self):
-        self.profile_agent = ProfileAgent()
+    def __init__(
+        self,
+        *,
+        get_embedding_provider_fn: Callable[[], "EmbeddingProvider"] | None = None,
+        llm_provider: "LLMProvider | None" = None,
+    ):
+        self._get_embedding_provider_fn = (
+            get_embedding_provider_fn or get_embedding_provider
+        )
+        self.profile_agent = ProfileAgent(llm_provider=llm_provider)
         self.interview_agent = InterviewAgent()
+        self.memory_prompt_adapter = MemoryPromptAdapter()
 
     async def process_user_message(
         self,
@@ -272,12 +288,16 @@ class ChatOrchestrator:
             background_voice = infer_background_voice(user.occupation)
             occupation = user.occupation or ""
 
-        from app.features.memory.chat_memory_injection import slice_interview_memory
-
         memory_bundle, mem_trace = await _fetch_interview_memory_bundle(
-            db, user_id, normalized_user_message
+            db,
+            user_id,
+            normalized_user_message,
+            get_embedding_provider_fn=self._get_embedding_provider_fn,
+        )
+        mem_slices = self.memory_prompt_adapter.slice_for_interview(
+            memory_bundle,
+            normalized_user_message,
         )
-        mem_slices = slice_interview_memory(memory_bundle, normalized_user_message)
         # 场景关键词仅作为 focus planner 的辅助输入，不直接拼进记忆块，避免抢过用户明确的关系/身份线索
         scene_cues_for_planner = extract_scene_cues(normalized_user_message)