修复：对话触发轻量拓展分析三缺陷（帮助意图误判、Phase A 弱命中抢占、LLM 空响应重试）

apps/web/vue3/index.html

@@ -23567,7 +23567,7 @@ compare: [[flesh, peel]]`;
                         const wantsExtensionStage = lowerInput.includes('extension') || userInput.includes('拓展分析') || userInput.includes('扩展分析');
                         const asksExtensionHelp = wantsExtensionStage && (
                             lowerInput.includes('help') || userInput.includes('如何使用') || userInput.includes('怎么用') ||
-                            userInput.includes('输入') || userInput.includes('参数') || userInput.includes('功能') || userInput.includes('帮助')
+                            userInput.includes('帮助') || userInput.includes('有什么功能')
                         );
                         if (asksExtensionHelp) {
                             try {

apps/web/vue3/main.py

@@ -17944,7 +17944,9 @@ async def _run_agent_chat(
             "handoff_request": request.message,
         }
 
-    # 轻量拓展分析自然语言触发：预闸命中才查轻量 store（Phase A 命中在此不短路）
+    # 轻量拓展分析自然语言触发：预闸命中才查轻量 store。
+    # 先直接问轻量匹配器（Phase A 打分器会用 legacy 扩展的弱 alias 抢占消息，
+    # 导致轻量命不中）；轻量落空时保持原有 Phase A 链路不变。
     try:
         from gfa_core.extension_stage_lightweight import nl_intent as _lightweight_nl
 
@@ -17953,7 +17955,7 @@ async def _run_agent_chat(
         )
         if _lightweight_published and _lightweight_nl._rough_trigger(request.message or "", _lightweight_published):
             extension_intent = await asyncio.to_thread(
-                _match_extension_stage_intent, request.message, request.job_id or ""
+                _match_lightweight_extension_intent, request.message, request.job_id or ""
             )
             if extension_intent and extension_intent.get("lightweight"):
                 assistant_message = extension_intent.get("message") or "已识别为轻量拓展分析。"

gfa_core/extension_stage_lightweight/nl_intent.py

@@ -363,14 +363,20 @@ def resolve_lightweight_extension_intent(
                 raise DeepSeekSemanticError("no deepseek config for default client", code="configuration_error")
             client = _default_client(config)
         model = getattr(config, "model", None) if config is not None else None
-        response = client.chat.completions.create(
-            model=model or "deepseek-chat",
-            messages=_build_messages(message, catalog, hints),
-            temperature=0,
-            response_format={"type": "json_object"},
-            max_tokens=4000,
-        )
-        content = response.choices[0].message.content
+        content = None
+        for attempt in range(2):
+            response = client.chat.completions.create(
+                model=model or "deepseek-chat",
+                messages=_build_messages(message, catalog, hints),
+                temperature=0,
+                response_format={"type": "json_object"},
+                max_tokens=4000,
+            )
+            content = response.choices[0].message.content if response.choices else None
+            if content and str(content).strip():
+                break
+        if not content or not str(content).strip():
+            raise DeepSeekSemanticError("empty LLM response after retries", code="empty_response")
         sanitized, reason = _sanitize_resolution(json.loads(content), catalog)
         if sanitized is None:
             return LightweightIntentResolution(