Emit __start signal at call_model entry for instant UI feedback

call_model now emits writer({"__start": True}) before LLM inference.
stream_response() converts it to {"__status": label} — distinct from
__meta so the UI shows it immediately without accumulating in the log.
Removes the 10-second silent wait before the first progress message.

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

services/agent/agent_service.py

@@ -115,6 +115,9 @@ class AgentService:
                 writer = get_stream_writer()
             except Exception:
                 writer = None
+            # LLM 추론 시작 직전에 즉시 신호 emit — UI에 "분석 중" 표시
+            if writer:
+                writer({"__start": True})
             # 체크박스 값을 모델의 enable_thinking으로 전달 (런타임 오버라이드)
             show_thinking = config.get("configurable", {}).get("show_thinking", False)
             _llm = llm_with_tools.bind(enable_thinking=show_thinking) if show_thinking != chat_model.enable_thinking else llm_with_tools
@@ -249,6 +252,11 @@ class AgentService:
 
             # ── custom 이벤트 ────────────────────────────────────────────
             if mode == "custom":
+                if isinstance(data, dict) and "__start" in data:
+                    # call_model 시작 즉시 emit — LLM 추론 전에 상태 표시
+                    label = "검색 결과를 분석하고 있습니다..." if prev_node == "tools" else "질문을 분석하고 있습니다..."
+                    yield {"__status": label}
+                    continue
                 if isinstance(data, dict) and "__query_rewrite" in data:
                     info = data["__query_rewrite"]
                     if lg or self._rag_verbose: