feat: persistent chat message storage + Android pull-to-load history

Backend:
- Add ChatMessage model + Alembic migration 024
- Add on_message callback to AgentRuntime for persisting messages during SSE streaming
- Plumb session_id from ChatRequest to AgentContext in all 4 chat endpoints
- Add GET /agent-chat/{id}/sessions and /sessions/{sid}/messages with cursor pagination

Android:
- Add DTOs/ApiService/MessageDao for server-side chat history
- ChatRepository: fetchOlderMessages (API + Room cache), offline fallback
- ChatViewModel: loadMoreHistory with isLoadingMore/hasMoreMessages state
- ChatScreen: scroll-to-top detection + top loading indicator

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

backend/app/agent_runtime/core.py

@@ -105,6 +105,7 @@ class AgentRuntime:
         execution_logger: Optional[Any] = None,
         on_tool_executed: Optional[Callable[[str], Any]] = None,
         on_llm_call: Optional[Callable[[Dict[str, Any]], Any]] = None,
+        on_message: Optional[Callable[[Dict[str, Any]], Any]] = None,
         hook_manager: Optional[HookManager] = None,
         streamlined: bool = False,
     ):
@@ -141,6 +142,7 @@ class AgentRuntime:
         self.execution_logger = execution_logger
         self.on_tool_executed = on_tool_executed
         self.on_llm_call = on_llm_call
+        self.on_message = on_message
         self._memory_context_loaded = False
         self._llm_invocations = 0
         # 自主学习作用域：bare 聊天用 "bare"，Agent 用 "agent"
@@ -808,6 +810,12 @@ class AgentRuntime:
 
         # 2. 追加用户消息
         self.context.add_user_message(user_input)
+        if self.on_message:
+            self.on_message({
+                "role": "user", "content": user_input,
+                "session_id": self.context.session_id,
+                "iteration": 0,
+            })
 
         # 2.5 计划模式 (P2) — 流式生成执行计划
         plan: Optional[Plan] = None
@@ -969,6 +977,12 @@ class AgentRuntime:
                 # LLM 直接返回文本 → 结束
                 self.context.add_assistant_message(content)
                 final_text = content or "（模型未返回有效内容）"
+                if self.on_message:
+                    self.on_message({
+                        "role": "assistant", "content": final_text,
+                        "session_id": self.context.session_id,
+                        "iteration": self.context.iteration,
+                    })
                 review_score = 0.0
 
                 # 输出质量自检（默认关闭）
@@ -1027,6 +1041,12 @@ class AgentRuntime:
 
             # 有工具调用 → 先记录 assistant 消息
             self.context.add_assistant_message(content or "", tool_calls, reasoning)
+            if self.on_message:
+                self.on_message({
+                    "role": "assistant", "content": content or "",
+                    "session_id": self.context.session_id,
+                    "iteration": self.context.iteration,
+                })
 
             # yield think 事件
             tc_names = [tc["function"]["name"] for tc in tool_calls]
@@ -1087,6 +1107,15 @@ class AgentRuntime:
                     result = json.dumps({"error": hook_res.reason}, ensure_ascii=False)
                     yield {"type": "tool_result", "name": tname, "result": result, "iteration": self.context.iteration}
                     self.context.add_tool_result(tcid, tname, result)
+                    if self.on_message:
+                        self.on_message({
+                            "role": "tool", "content": result,
+                            "session_id": self.context.session_id,
+                            "iteration": self.context.iteration,
+                            "tool_name": tname,
+                            "tool_input": json.dumps(targs, ensure_ascii=False) if targs else None,
+                            "tool_output": result,
+                        })
                     continue
                 if hook_res.modified_input:
                     targs = hook_res.modified_input
@@ -1119,11 +1148,29 @@ class AgentRuntime:
                         result = f"[审批拒绝] 工具 {tname} 需要人工审批但被拒绝。"
                         yield {"type": "tool_result", "name": tname, "result": result, "iteration": self.context.iteration}
                         self.context.add_tool_result(tcid, tname, result)
+                        if self.on_message:
+                            self.on_message({
+                                "role": "tool", "content": result,
+                                "session_id": self.context.session_id,
+                                "iteration": self.context.iteration,
+                                "tool_name": tname,
+                                "tool_input": json.dumps(targs, ensure_ascii=False) if targs else None,
+                                "tool_output": result,
+                            })
                         continue
                     elif decision == "skip":
                         result = f"[审批跳过] 工具 {tname} 被跳过。"
                         yield {"type": "tool_result", "name": tname, "result": result, "iteration": self.context.iteration}
                         self.context.add_tool_result(tcid, tname, result)
+                        if self.on_message:
+                            self.on_message({
+                                "role": "tool", "content": result,
+                                "session_id": self.context.session_id,
+                                "iteration": self.context.iteration,
+                                "tool_name": tname,
+                                "tool_input": json.dumps(targs, ensure_ascii=False) if targs else None,
+                                "tool_output": result,
+                            })
                         continue
                     # decision == "approved" → 继续执行
 
@@ -1154,6 +1201,15 @@ class AgentRuntime:
                 ))
 
                 self.context.add_tool_result(tcid, tname, result)
+                if self.on_message:
+                    self.on_message({
+                        "role": "tool", "content": result,
+                        "session_id": self.context.session_id,
+                        "iteration": self.context.iteration,
+                        "tool_name": tname,
+                        "tool_input": json.dumps(targs, ensure_ascii=False) if targs else None,
+                        "tool_output": result,
+                    })
                 self.context.tool_calls_made += 1
 
                 # Hook: PostToolUse — 工具执行后处理 (流式)