fix(vision): unwrap fast_json list-wrapped object

根因：
- _FAST_TIMEOUT/_FAST_JSON_TIMEOUT 设为8s，但qwen3.8-flash关thinking后单图实测8-9s，staging网络稍慢即全部超时
- 超时cancel后走pro兜底，pro timeout=20s也偏紧
- 缺少 response_format=json_object 导致qwen偶发输出中文解释而非JSON

修复：
- fast_json: _DEFAULT_TIMEOUT 8→12s
- fast_path: _FAST_TIMEOUT/_FAST_JSON_TIMEOUT 8→12s, _PRO_TIMEOUT 20→25s
- vlm_fallback: _DEFAULT_TIMEOUT 20→25s
- fast_json + fallback 都加 response_format={'type':'json_object'}强约束JSON输出

本地3图E2E: 8.57s, 3/3 fast_json命中, portrait_prompt正常输出一位年轻男性/女性...

Refs: staging P0 3/3全返回未识别/无法判断

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", "8"))
-_FAST_JSON_TIMEOUT = float(os.environ.get("VISION_V2_FAST_JSON_TIMEOUT", "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"))
 _OCR_TIMEOUT = float(os.environ.get("VISION_V2_OCR_TIMEOUT", "6"))
-_PRO_TIMEOUT = float(os.environ.get("VISION_V2_PRO_TIMEOUT", "20"))
+_PRO_TIMEOUT = float(os.environ.get("VISION_V2_PRO_TIMEOUT", "25"))
 
 _FALLBACK_RESULT = {
     "name": "未识别",

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

@@ -17,7 +17,7 @@ logger = logging.getLogger(__name__)
 
 _BASE_URL = "https://dashscope.aliyuncs.com/compatible-mode/v1"
 _PRO_MODEL = "qwen3.7-plus"
-_DEFAULT_TIMEOUT = 20
+_DEFAULT_TIMEOUT = 25
 _DEFAULT_MAX_TOKENS = 800
 
 _PRO_SYSTEM = (
@@ -113,6 +113,7 @@ def call_pro_vlm(
         "max_tokens": _DEFAULT_MAX_TOKENS,
         "stream": False,
         "enable_thinking": False,
+        "response_format": {"type": "json_object"},
     }
     try:
         r = httpx.post(
@@ -140,11 +141,11 @@ def call_pro_vlm(
             usage.get("completion_tokens", 0),
         )
         text = _strip_code_fence(raw)
-        l, r_pos = text.find("{"), text.rfind("}")
-        if l < 0 or r_pos <= l:
+        lb, r_pos = text.find("{"), text.rfind("}")
+        if lb < 0 or r_pos <= lb:
             logger.warning("[vision.v2] pro 无JSON elapsed=%.1fs head=%s", elapsed, raw[:200])
             return None
-        obj = json.loads(text[l : r_pos + 1])
+        obj = json.loads(text[lb : r_pos + 1])
         if not isinstance(obj, dict):
             return None
         scene = obj.get("scene") or "通用"

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

@@ -7,7 +7,7 @@
 - enable_thinking=false 关闭推理链（reasoning 是延迟主因）
 - system prompt 极致精简，只给字段 schema 和强约束（禁止自然语言、禁止 markdown）
 - max_tokens=350、temperature=0.1（稳定输出 JSON）
-- timeout=8s（失败由外层走 pro 兜底）
+- timeout=12s（失败由外层走 pro 兜底）
 - API Key 从环境变量 DASHSCOPE_API_KEY 读取
 """
 
@@ -24,7 +24,7 @@ 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 = 8
+_DEFAULT_TIMEOUT = 12
 _DEFAULT_MAX_TOKENS = 350
 
 # 极简 system prompt：只给字段定义 + 硬性输出要求
@@ -108,6 +108,7 @@ def call_fast_json(
         "max_tokens": max_tokens,
         "stream": False,
         "enable_thinking": False,
+        "response_format": {"type": "json_object"},
     }
     try:
         resp = httpx.post(
@@ -152,14 +153,21 @@ def call_fast_json(
             reasoning_tokens,
         )
         text = _strip_code_fence(raw)
-        l, r = text.find("{"), text.rfind("}")
-        if l >= 0 and r > l:
-            text = text[l : r + 1]
+        lb, rb = text.find("{"), text.rfind("}")
+        if lb >= 0 and rb > lb:
+            text = text[lb : rb + 1]
         try:
             obj = json.loads(text)
         except json.JSONDecodeError:
             logger.warning("[vision.v2] fast_json JSON 解析失败 elapsed=%.1fs head=%s", elapsed, raw[:200])
             return None
+        if isinstance(obj, list):
+            # 模型偶发将对象包在数组里，取第一个 dict 元素
+            dicts = [x for x in obj if isinstance(x, dict)]
+            if not dicts:
+                logger.warning("[vision.v2] fast_json list 内无 dict: %s", type(obj))
+                return None
+            obj = dicts[0]
         if not isinstance(obj, dict):
             logger.warning("[vision.v2] fast_json 非 dict: %s", type(obj))
             return None