fix(gateway): 熔断计入持续 4xx + 去 assert + Protocol/transient 去重

P1-2 非可重试错误(持续 401/403)也 record_failure，坏 key 可触发熔断。
P1-9 _complete_structured 的 assert 改显式 raise ValueError（-O 安全）。
P2 GatewayRun 抽到 orchestrator/_protocols.py 单点（去 4 处重复）；
  _is_transient 抽到 adapters/base.py is_transient_by_name（去 3 处重复）；
  Gemini Protocol 改 async def；gateway._retrying 去无用 async。

packages/llm_gateway/ww_llm_gateway/adapters/anthropic.py

@@ -16,7 +16,13 @@ from pydantic import BaseModel
 
 from ..errors import TransientProviderError
 from ..types import LlmRequest
-from .base import Capabilities, ProviderResult, ProviderUsage, StreamChunk
+from .base import (
+    Capabilities,
+    ProviderResult,
+    ProviderUsage,
+    StreamChunk,
+    is_transient_by_name,
+)
 
 if TYPE_CHECKING:
     from anthropic import AsyncAnthropic
@@ -44,11 +50,7 @@ class StructuredAnthropic(Protocol):
 
 
 def _is_transient(exc: Exception) -> bool:
-    name = type(exc).__name__
-    if name in _TRANSIENT_NAMES:
-        return True
-    status = getattr(exc, "status_code", None)
-    return isinstance(status, int) and (status == 429 or status >= 500)
+    return is_transient_by_name(exc, _TRANSIENT_NAMES)
 
 
 def _system_blocks(req: LlmRequest) -> list[dict[str, Any]]:
@@ -144,7 +146,8 @@ class AnthropicAdapter:
         return ProviderResult(text=_text_from(resp), usage=usage)
 
     async def _complete_structured(self, req: LlmRequest, model: str) -> ProviderResult:
-        assert req.output_schema is not None
+        if req.output_schema is None:
+            raise ValueError("_complete_structured called without output_schema")
         kwargs: dict[str, Any] = {
             "model": model,
             "max_tokens": req.max_tokens or _DEFAULT_MAX_TOKENS,