style(vision): rename ambiguous l/r vars to lb/rb (ruff E741)

根因：上版同时传 thinking={type:disabled} + reasoning_effort=low，方舟返回400
Invalid combination of reasoning_effort and thinking type。降级重试分支
把两个参数都pop掉，模型回到默认thinking开启→响应9-12s超时，fast全败。

修复：
1. 只传 thinking={"type":"disabled"}，去掉互斥的 reasoning_effort
2. 400降级只pop thinking，保留精简payload重试（不pop多个）
3. 先单独验证 lite 单次HTTP 200且reasoning_tokens=0再跑E2E，省一轮部署

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

@@ -187,10 +187,10 @@ def call_pro_vlm(
         return None
 
     text = _strip_code_fence(raw)
-    l, r = text.find("{"), text.rfind("}")
-    if l >= 0 and r > l:
+    lb, rb = text.find("{"), text.rfind("}")
+    if lb >= 0 and rb > lb:
         try:
-            obj = json.loads(text[l : r + 1])
+            obj = json.loads(text[lb : rb + 1])
             if isinstance(obj, dict):
                 logger.info("[vision.vlm] 图片 #%d pro VLM JSON 完成 elapsed=%.1fs", idx, elapsed)
                 return {

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

@@ -117,14 +117,9 @@ def call_fast_json(
             "max_tokens": max_tokens,
             "stream": False,
         }
-        # 关键：关闭 thinking（避免产生 reasoning_tokens 拖慢响应）
-        # 方舟/豆包 2.x 模型支持 thinking.type=disabled
-        try:
-            payload["thinking"] = {"type": "disabled"}
-        except Exception:
-            pass
-        # 部分模型用 reasoning_effort 控制思考深度
-        payload["reasoning_effort"] = "low"
+        # 关键：关闭 thinking（reasoning_tokens 是延迟主因，单次要10-12s）
+        # 方舟/豆包 Seed 2.x 支持 thinking={type:"disabled"}，且不要和 reasoning_effort 同时传（两者互斥会400）
+        payload["thinking"] = {"type": "disabled"}
 
         resp = httpx.post(
             url,
@@ -133,15 +128,12 @@ def call_fast_json(
             timeout=timeout,
         )
         elapsed = time.time() - t0
-        if resp.status_code != 200:
-            logger.warning(
-                "[vision.v2] fast_json HTTP %d elapsed=%.1fs body=%s", resp.status_code, elapsed, resp.text[:200]
-            )
-            # 如果400说明不支持thinking参数，降级重试一次
-            if resp.status_code == 400 and "thinking" in resp.text.lower():
+        # 400 说明模型不支持 thinking 参数（极少数旧模型），重试一次不带 thinking
+        if resp.status_code == 400:
+            body_preview = resp.text[:300].lower()
+            logger.warning("[vision.v2] fast_json HTTP 400 elapsed=%.1fs body=%s", elapsed, resp.text[:200])
+            if "thinking" in body_preview or "reasoning" in body_preview:
                 payload.pop("thinking", None)
-                payload.pop("reasoning_effort", None)
-                time.time()
                 resp = httpx.post(
                     url,
                     headers={"Authorization": f"Bearer {api_key}", "Content-Type": "application/json"},
@@ -150,10 +142,15 @@ def call_fast_json(
                 )
                 elapsed = time.time() - t0
                 if resp.status_code != 200:
-                    logger.warning("[vision.v2] fast_json 降级后 HTTP %d elapsed=%.1fs", resp.status_code, elapsed)
+                    logger.warning("[vision.v2] fast_json 降级重试 HTTP %d elapsed=%.1fs", resp.status_code, elapsed)
                     return None
             else:
                 return None
+        elif resp.status_code != 200:
+            logger.warning(
+                "[vision.v2] fast_json HTTP %d elapsed=%.1fs body=%s", resp.status_code, elapsed, resp.text[:200]
+            )
+            return None
         data = resp.json()
         raw = (data.get("choices") or [{}])[0].get("message", {}).get("content")
         if raw is None:
@@ -175,10 +172,10 @@ def call_fast_json(
 
         text = _strip_code_fence(raw)
         # 截到第一个 { 和最后一个 } 之间，容忍前后偶发文字
-        l = text.find("{")
-        r = text.rfind("}")
-        if l >= 0 and r > l:
-            text = text[l : r + 1]
+        lb = text.find("{")
+        rb = text.rfind("}")
+        if lb >= 0 and rb > lb:
+            text = text[lb : rb + 1]
         try:
             obj = json.loads(text)
         except json.JSONDecodeError: