fix(llm): request_id 条件透传贯通网关日志（CR-H4，守 §9.3）

memory/contracts.md

@@ -29,6 +29,10 @@
   - `OpenAICompatAdapter` 不变（新增瞬时错误包装，参与回退链）。
 - **依赖**：`packages/llm_gateway/pyproject.toml` 加 `tenacity>=8.2`（已 `uv sync`；原为 instructor 传递依赖）。**`anthropic`/`google-genai` SDK 已由 @devops 加（已 `uv sync`）**——适配器仍懒导入/注入客户端，单测零依赖。
 
+#### C1 扩（CR-H4, 2026-07-08）· `LlmRequest` 加可选 `request_id`　owner @llm　状态: 稳定
+- **`LlmRequest` 加字段**（C1 形变，**仅加可选、默认 `None`，最后一位，向后兼容**）：`request_id: str | None = None`。调用方（apps/api/编排器）设置后，网关把它**条件透传**到每条 `llm_call` / `llm_provider_failed` 日志（`req.request_id is not None` 才发——避免 None 覆盖 `merge_contextvars` 在 sync/SSE 路径供的 id），贯通端到端追踪（ARCH §9.3）。
+- **无 OpenAPI 形变、无需 gen:api**：`LlmRequest` 是网关内部类型，不出现在 `packages/shared` / `schema.d.ts`；前端无感。旧调用点全部无 request_id → 默认 `None`，行为不变。
+
 #### C1 扩 follow-up #2（2026-06-19）· `build_adapter` provider→适配器工厂　owner @llm　状态: 稳定
 - 位置：`packages/llm_gateway/ww_llm_gateway/factory.py`，经包 `__init__` 导出 `build_adapter`。
 - **签名（稳定，apps/api `build_gateway_for_tier` 据此逐 provider 建适配器进 `adapters` dict）**：

packages/llm_gateway/tests/test_gateway.py

@@ -5,6 +5,7 @@ from __future__ import annotations
 import uuid
 
 import pytest
+import structlog
 from fakes import FakeAdapter, FakeLedger, fake_route
 from ww_llm_gateway.gateway import Gateway
 from ww_llm_gateway.routing import resolve_route
@@ -60,6 +61,33 @@ async def test_stream_yields_deltas_and_records_once(req: LlmRequest) -> None:
     assert ledger.records[0].output_tokens == 50
 
 
+async def test_run_logs_request_id_from_request(scope: Scope) -> None:
+    # CR-H4：调用方设了 request_id 时，网关 llm_call 日志必须带上它（贯通 §9.3 追踪）。
+    ledger = FakeLedger()
+    gw = Gateway({"deepseek": FakeAdapter()}, ledger, resolver=fake_route)
+    req = LlmRequest(tier="writer", input="x", scope=scope, request_id="rid-xyz")
+
+    with structlog.testing.capture_logs() as logs:
+        await gw.run(req)
+
+    call = next(entry for entry in logs if entry["event"] == "llm_call")
+    assert call["request_id"] == "rid-xyz"
+
+
+async def test_run_omits_request_id_when_unset(scope: Scope) -> None:
+    # 未设 request_id 时**不得**发出该键——否则 None 会覆盖 merge_contextvars 供的 id，
+    # 反倒回退了 sync/SSE 路径的追踪（条件透传锁定）。
+    ledger = FakeLedger()
+    gw = Gateway({"deepseek": FakeAdapter()}, ledger, resolver=fake_route)
+    req = LlmRequest(tier="writer", input="x", scope=scope)
+
+    with structlog.testing.capture_logs() as logs:
+        await gw.run(req)
+
+    call = next(entry for entry in logs if entry["event"] == "llm_call")
+    assert "request_id" not in call
+
+
 async def test_unknown_provider_raises_llm_unavailable(req: LlmRequest) -> None:
     gw = Gateway({}, FakeLedger(), resolver=fake_route)
 

packages/llm_gateway/tests/test_types.py

@@ -17,6 +17,7 @@ def test_llm_request_has_no_thinking_field() -> None:
         "output_schema",
         "max_tokens",
         "scope",
+        "request_id",  # CR-H4：可选端到端追踪字段（additive/optional）。
     }
     # 同名但另一处 live 字段 `Capabilities.thinking`（适配器能力矩阵）不受影响。
     assert "thinking" in Capabilities.model_fields

packages/llm_gateway/ww_llm_gateway/gateway.py

@@ -183,22 +183,26 @@ class Gateway:
     def _log_call(
         self, req: LlmRequest, usage: Usage, served_by: ServedBy, *, stream: bool
     ) -> None:
-        log.info(
-            "llm_call",
-            provider=usage.provider,
-            model=usage.model,
-            tier=req.tier,
-            input_tokens=usage.input_tokens,
-            output_tokens=usage.output_tokens,
-            cache_read_tokens=usage.cache_read_tokens,
-            cost_minor=usage.cost_minor,
-            currency=usage.currency,
-            stream=stream,
-            fell_back=served_by.fell_back,
-            degraded=served_by.degraded,
-            input_chars=_input_len(req),
-            project_id=str(req.scope.project_id) if req.scope.project_id else None,
-        )
+        fields: dict[str, object] = {
+            "provider": usage.provider,
+            "model": usage.model,
+            "tier": req.tier,
+            "input_tokens": usage.input_tokens,
+            "output_tokens": usage.output_tokens,
+            "cache_read_tokens": usage.cache_read_tokens,
+            "cost_minor": usage.cost_minor,
+            "currency": usage.currency,
+            "stream": stream,
+            "fell_back": served_by.fell_back,
+            "degraded": served_by.degraded,
+            "input_chars": _input_len(req),
+            "project_id": str(req.scope.project_id) if req.scope.project_id else None,
+        }
+        # 仅在调用方显式设了 request_id 时才发——否则 None 会覆盖 merge_contextvars
+        # 供的 id，反倒回退 sync/SSE 路径的追踪（§9.3）。
+        if req.request_id is not None:
+            fields["request_id"] = req.request_id
+        log.info("llm_call", **fields)
 
     def _retrying(self) -> AsyncRetrying:
         return AsyncRetrying(
@@ -237,6 +241,7 @@ class Gateway:
                     provider=route.provider,
                     tier=req.tier,
                     error=type(exc).__name__,
+                    **({"request_id": req.request_id} if req.request_id is not None else {}),
                 )
                 continue
             self._breaker.record_success(route.provider)
@@ -297,6 +302,7 @@ class Gateway:
                     tier=req.tier,
                     error=type(exc).__name__,
                     stream=True,
+                    **({"request_id": req.request_id} if req.request_id is not None else {}),
                 )
                 continue
             self._breaker.record_success(route.provider)

packages/llm_gateway/ww_llm_gateway/types.py

@@ -29,7 +29,11 @@ class Scope(BaseModel):
 
 
 class LlmRequest(BaseModel):
-    """统一请求。`system` 稳定块在前（断点前），`input` 易变内容在后。"""
+    """统一请求。`system` 稳定块在前（断点前），`input` 易变内容在后。
+
+    `request_id`（可选）贯通端到端追踪：调用方设置后，网关把它落到每条 LLM 调用日志，
+    使一次「写一章」的 assemble→write→审→验收全链路可 grep（ARCH §9.3）。
+    """
 
     model_config = ConfigDict(arbitrary_types_allowed=True)
 
@@ -40,6 +44,7 @@ class LlmRequest(BaseModel):
     output_schema: type[BaseModel] | None = None
     max_tokens: int | None = None
     scope: Scope
+    request_id: str | None = None
 
 
 class Usage(BaseModel):