fix(viral-video): 修复call_llm传list bug 改为直接chat_completion

审计发现 _step_intent_parsing(L482) 和 _step_script_generation(L867) 两处
把 messages list 当作第一个位置参数传给 call_llm(prompt:str,...)，
list被repr成字符串再套一层默认system_prompt，实际发给豆包的消息结构异常。

修复：
- 改为直接用 get_doubao_client().chat_completion(messages,...)，
  与 reviewer.py/generator.py 调用方式一致
- 增加 is_available 预检，不可用时直接走fallback（意图解析默认dict / _fallback_script）
- 不再走 call_llm 字符串包装层

本次PR同时包含#2214 timeout调优（cherry-pick）：
- fast_json _DEFAULT_TIMEOUT 12→15s
- fast_path _FAST_TIMEOUT/_FAST_JSON_TIMEOUT 12→15s
- pro fallback _DEFAULT_TIMEOUT 25→30s，外层_PRO_TIMEOUT 25→30s
- 保留环境变量覆盖能力

apps/worker/worker_app/tasks/viral_video.py

@@ -421,10 +421,14 @@ def _step_intent_parsing(job: ViralVideoJob, image_analysis: dict) -> dict:
             render_system_prompt,
             render_user_prompt,
         )
-        from packages.shared.ai_service import call_llm
+        from packages.shared.ai_client import get_doubao_client
     except ImportError:
         return {"intent": "推广产品", "key_messages": ["产品亮点"], "tone": "专业", "suggested_title": ""}
 
+    _llm_client = get_doubao_client()
+    if not _llm_client.is_available:
+        return {"intent": "推广产品", "key_messages": ["产品亮点"], "tone": "专业", "suggested_title": ""}
+
     products_summary = ""
     products = (image_analysis or {}).get("products", []) or []
     for p in products:
@@ -479,13 +483,13 @@ def _step_intent_parsing(job: ViralVideoJob, image_analysis: dict) -> dict:
     for _m, _lbl in [(_fast, "fast"), (_pro, "pro-fallback")]:
         try:
             logger.info("[爆款视频] 意图解析 model=%s label=%s", _m, _lbl)
-            raw = call_llm(
+            raw = _llm_client.chat_completion(
                 [{"role": "system", "content": system}, {"role": "user", "content": user}],
                 temperature=0.4,
                 max_tokens=1024,
                 model=_m,
                 timeout=60,
-            )  # #2180: 意图解析 LLM 实测需更长响应，原25s太紧
+            )  # #2180/#2215: 直接用 client.chat_completion 传 messages list，不再走 call_llm 字符串包装
             if not raw:
                 continue
             parsed = _parse(raw)
@@ -825,10 +829,14 @@ def _step_script_generation(job: ViralVideoJob, intent: dict, image_analysis: di
             GLOBAL_CONSTRAINTS,
             NEGATIVE_RULES,
         )
-        from packages.shared.ai_service import call_llm
+        from packages.shared.ai_client import get_doubao_client
     except ImportError:
         return _fallback_script(job)
 
+    _llm_client2 = get_doubao_client()
+    if not _llm_client2.is_available:
+        return _fallback_script(job)
+
     products_summary = _build_products_summary(image_analysis)
     dur = max(5, min(30, int(getattr(job, "duration", 15) or 15)))
 
@@ -864,7 +872,7 @@ def _step_script_generation(job: ViralVideoJob, intent: dict, image_analysis: di
 
     def _try_gen(model: str, temp: float, max_tok: int, label: str, tmo: int = 25):
         logger.info("[爆款视频] 编导脚本生成 model=%s label=%s timeout=%d", model, label, tmo)
-        raw = call_llm(
+        raw = _llm_client2.chat_completion(
             [{"role": "system", "content": system_tpl}, {"role": "user", "content": user}],
             temperature=temp,
             max_tokens=max_tok,

apps/worker/worker_app/tasks/vision/fast_path.py

@@ -23,10 +23,10 @@ logger = logging.getLogger(__name__)
 
 # 超时（可通过环境变量覆盖）
 _IMG_WORKERS = int(os.environ.get("VISION_V2_IMG_WORKERS", "8"))
-_FAST_TIMEOUT = float(os.environ.get("VISION_V2_FAST_TIMEOUT", "12"))
-_FAST_JSON_TIMEOUT = float(os.environ.get("VISION_V2_FAST_JSON_TIMEOUT", "12"))
+_FAST_TIMEOUT = float(os.environ.get("VISION_V2_FAST_TIMEOUT", "15"))
+_FAST_JSON_TIMEOUT = float(os.environ.get("VISION_V2_FAST_JSON_TIMEOUT", "15"))
 _OCR_TIMEOUT = float(os.environ.get("VISION_V2_OCR_TIMEOUT", "6"))
-_PRO_TIMEOUT = float(os.environ.get("VISION_V2_PRO_TIMEOUT", "25"))
+_PRO_TIMEOUT = float(os.environ.get("VISION_V2_PRO_TIMEOUT", "30"))
 
 _FALLBACK_RESULT = {
     "name": "未识别",

apps/worker/worker_app/tasks/vision/vlm_fallback.py

@@ -25,7 +25,7 @@ logger = logging.getLogger(__name__)
 
 _BASE_URL = "https://dashscope.aliyuncs.com/compatible-mode/v1"
 _PRO_MODEL = "qwen3.7-plus"
-_DEFAULT_TIMEOUT = 25
+_DEFAULT_TIMEOUT = 30
 _DEFAULT_MAX_TOKENS = 800
 
 

apps/worker/worker_app/tasks/vision/vlm_fast_json.py

@@ -27,7 +27,7 @@ logger = logging.getLogger(__name__)
 # DashScope OpenAI 兼容 endpoint
 _BASE_URL = "https://dashscope.aliyuncs.com/compatible-mode/v1"
 _FAST_MODEL = "qwen3.8-flash"
-_DEFAULT_TIMEOUT = 12
+_DEFAULT_TIMEOUT = 15
 _DEFAULT_MAX_TOKENS = 350