fix: draft max_tokens 512→2048 (thought 블록이 토큰 소진하여 빈 초안 문제)

src/inference/api_server.py

@@ -1043,7 +1043,7 @@ class vLLMEngineManager:
 
             gen_request = GenerateCivilResponseRequest(
                 prompt=working_query,
-                max_tokens=512,
+                max_tokens=2048,
                 temperature=0.7,
                 use_rag=False,
             )