fix(BIZ-26): 限流范围收窄到 NVIDIA 网关

- 新增网关识别逻辑：只识别 nvidia / nvidiavx18088980513 为限流目标
- volcengine-plan、siliconflow、deepseek 等非 NVIDIA 网关默认不进入令牌桶
- RequestScheduler 增加 gateway/model 参数与 _should_rate_limit 判断
- 未知网关默认不限流，避免误伤其他通道
- 补充网关范围测试与使用文档说明

Co-authored-by: multica-agent <github@multica.ai>

scripts/rate_limiter.py

@@ -25,6 +25,57 @@ from enum import IntEnum
 from datetime import datetime, timedelta
 
 
+# ============================================================================
+# 网关识别：只对 NVIDIA 网关限流
+# ============================================================================
+
+NVIDIA_GATEWAY_ALIASES = {
+    "nvidia",
+    "nvidia-gateway",
+    "nvidia_gateway",
+    "nvidiavx18088980513",
+}
+
+UNLIMITED_GATEWAY_ALIASES = {
+    "volcengine",
+    "volcengine-plan",
+    "siliconflow",
+    "deepseek",
+    "deepseek-api",
+}
+
+
+def normalize_gateway_name(value: Optional[str]) -> Optional[str]:
+    """
+    归一化网关/模型名称。
+
+    输入可以是：
+    - provider: nvidia / volcengine-plan / siliconflow / deepseek
+    - model: nvidiavx18088980513/deepseek-ai/deepseek-v4-pro
+    - model: volcengine-plan/ark-code-latest
+
+    返回 provider 前缀的小写形式。未知则返回 None。
+    """
+    if not value:
+        return None
+    text = str(value).strip().lower()
+    if not text:
+        return None
+    return text.split("/", 1)[0]
+
+
+def is_nvidia_gateway(value: Optional[str]) -> bool:
+    """判断请求是否走 NVIDIA 网关。未知网关默认不限流。"""
+    provider = normalize_gateway_name(value)
+    if provider is None:
+        return False
+    if provider in NVIDIA_GATEWAY_ALIASES:
+        return True
+    if provider in UNLIMITED_GATEWAY_ALIASES:
+        return False
+    return provider.startswith("nvidia")
+
+
 # ============================================================================
 # 优先级枚举
 # ============================================================================
@@ -50,6 +101,8 @@ class Request:
     payload: Any = field(compare=False)
     callback: Optional[Callable] = field(compare=False, default=None)
     fallback_model: Optional[str] = field(compare=False, default=None)
+    gateway: Optional[str] = field(compare=False, default=None)
+    model: Optional[str] = field(compare=False, default=None)
     
     def __post_init__(self):
         if self.timestamp is None:
@@ -69,7 +122,10 @@ class Request:
 
 class TokenBucket:
     """
-    令牌桶限流器
+    NVIDIA 网关专用令牌桶限流器
+
+    注意：令牌桶本身只负责节流算法；是否启用由 RequestScheduler._should_rate_limit()
+    按 gateway/model 判断。volcengine-plan、siliconflow、DeepSeek 等非 NVIDIA 网关不会进入此桶。
     
     参数：
         rate: 令牌生成速率（个/秒），默认 40 RPM = 0.67 个/秒
@@ -362,6 +418,28 @@ class RequestScheduler:
                 # 记录错误但不中断工作线程
                 print(f"[RequestScheduler] Worker error: {e}")
     
+    def _extract_gateway_hint(self, request: Request) -> Optional[str]:
+        """从 request.gateway / request.model / payload 中提取网关提示。"""
+        if request.gateway:
+            return request.gateway
+        if request.model:
+            return request.model
+        if isinstance(request.payload, dict):
+            for key in ("gateway", "provider", "model", "model_id"):
+                value = request.payload.get(key)
+                if value:
+                    return str(value)
+        return None
+
+    def _should_rate_limit(self, request: Request) -> bool:
+        """
+        只对 NVIDIA 网关请求启用令牌桶。
+
+        设计原则：未知网关默认不限制，避免误伤 volcengine-plan / siliconflow / DeepSeek
+        等其他 API 网关。要被限流，调用方必须显式传 gateway/model，且能识别为 NVIDIA。
+        """
+        return is_nvidia_gateway(self._extract_gateway_hint(request))
+
     def _process_request(self, request: Request) -> None:
         """
         处理单个请求
@@ -372,7 +450,12 @@ class RequestScheduler:
         """
         self.stats["total_requests"] += 1
         
-        # 尝试获取令牌
+        # 只对 NVIDIA 网关请求启用令牌桶；其他网关直接执行
+        if not self._should_rate_limit(request):
+            self._execute_request(request)
+            return
+
+        # NVIDIA 网关请求：尝试获取令牌
         if request.priority <= Priority.HIGH:
             # 高优先级：无限等待
             got_token = self.token_bucket.wait_for_token(timeout=None)
@@ -419,7 +502,9 @@ class RequestScheduler:
         priority: Priority = Priority.NORMAL,
         callback: Optional[Callable] = None,
         fallback_model: Optional[str] = None,
-        request_id: Optional[str] = None
+        request_id: Optional[str] = None,
+        gateway: Optional[str] = None,
+        model: Optional[str] = None
     ) -> str:
         """
         提交请求到调度队列
@@ -440,7 +525,9 @@ class RequestScheduler:
             request_id=request_id,
             payload=payload,
             callback=callback,
-            fallback_model=fallback_model
+            fallback_model=fallback_model,
+            gateway=gateway,
+            model=model
         )
         
         self.request_queue.put(req)