merge: T1 四档分级降级熔断（T-P1-25b，NFR-COST-02）

tests/integration/test_vertical_slice.py

@@ -12,6 +12,8 @@ from __future__ import annotations
 
 import os
 import uuid
+from collections.abc import AsyncIterator
+from contextlib import asynccontextmanager
 from decimal import Decimal
 from typing import Any
 
@@ -31,40 +33,48 @@ pytestmark = [
 ]
 
 
-@pytest.fixture
-async def client() -> Any:
+class FakeProvider:
+    """替换真实 provider：集成用例验证配额与计量链路，不该花钱也不该受网络影响。"""
+
+    name = "deepseek"
+    default_model = "deepseek-v4-flash"
+
+    async def complete(self, request: Any) -> Any:
+        from packages.llm.types import ProviderResponse, Usage
+
+        return ProviderResponse(
+            text="建议白衬衫配西裤。",
+            model="deepseek-v4-flash",
+            usage=Usage(tokens_in=800, tokens_out=120, tokens_cached=600),
+        )
+
+    async def aclose(self) -> None:
+        return None
+
+
+@asynccontextmanager
+async def fake_client(app: Any) -> AsyncIterator[Any]:
+    """起 app lifespan + ASGI 客户端，并把 deepseek 换成 FakeProvider。"""
     import httpx
 
-    from apps.api.main import create_app
-    from packages.llm.types import ProviderResponse, Usage
-    from packages.shared.config import get_settings
-
-    app = create_app(get_settings())
-
-    class FakeProvider:
-        name = "deepseek"
-        default_model = "deepseek-v4-flash"
-
-        async def complete(self, request: Any) -> ProviderResponse:
-            return ProviderResponse(
-                text="建议白衬衫配西裤。",
-                model="deepseek-v4-flash",
-                usage=Usage(tokens_in=800, tokens_out=120, tokens_cached=600),
-            )
-
-        async def aclose(self) -> None:
-            return None
-
     transport = httpx.ASGITransport(app=app)
     async with (
         httpx.AsyncClient(transport=transport, base_url="http://test") as http,
         app.router.lifespan_context(app),
     ):
-        # 替换掉真实 provider：集成用例验证配额与计量链路，不该花钱也不该受网络影响
         app.state.llm_gateway._providers["deepseek"] = FakeProvider()
         yield http
 
 
+@pytest.fixture
+async def client() -> Any:
+    from apps.api.main import create_app
+    from packages.shared.config import get_settings
+
+    async with fake_client(create_app(get_settings())) as http:
+        yield http
+
+
 async def test_chat_slice_reserves_commits_and_records_cost(client: Any) -> None:
     token_resp = await client.post(
         "/api/v1/auth/dev-token", json={"device_key": f"it-{uuid.uuid4().hex[:8]}"}
@@ -99,3 +109,67 @@ async def test_unauthenticated_requests_get_the_unified_envelope(client: Any) ->
     resp = await client.post("/api/v1/agent/chat", json={"message": "hi"})
     assert resp.status_code == 401
     assert resp.json()["code"] == 40100
+
+
+async def test_block_all_budget_rejects_and_refunds_the_reservation(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    """T-P1-25b：reserve 之后才触发熔断 → gateway 拒绝 → refund → 净扣减 = 0。
+
+    这是四档降级（架构 §13.3）的验收路径：把 block_all 档的 allow 名单清空
+    （连纯文本对话也拦），让 agent_chat 走 100% 档的既有 60003 语义，从而
+    断言配额三段式在「先预扣、后熔断」时的补偿行为。
+    """
+    import json
+
+    from apps.api.main import create_app
+    from packages.llm.budget import DailyBudgetGuard, InMemoryCostCounter
+    from packages.shared.config import get_settings
+    from packages.shared.redis_client import get_redis
+
+    monkeypatch.setenv("LLM_DAILY_BUDGET_CENTS", "1")  # 预算 1 分钱
+    monkeypatch.setenv(
+        "LLM_BUDGET_TIERS_JSON",
+        json.dumps(
+            [
+                {"threshold": 0.6, "name": "warn", "blocked": []},
+                {"threshold": 0.8, "name": "degrade_secondary", "blocked": ["hd_export"]},
+                {"threshold": 0.9, "name": "degrade_generation", "blocked": ["hairstyle"]},
+                # allow 清空：agent_chat 也被 block_all 拦下（演练配置）
+                {"threshold": 1.0, "name": "block_all", "blocked": "*", "allow": []},
+            ]
+        ),
+    )
+    get_settings.cache_clear()
+    try:
+        app = create_app(get_settings())
+        async with fake_client(app) as http:
+            redis = get_redis()
+            # 注入当日成本到预算以上（键格式照 DailyBudgetGuard.current_key()）
+            key_helper = DailyBudgetGuard(InMemoryCostCounter(), daily_budget_cents=Decimal("1"))
+            cost_key = key_helper.current_key()
+            await redis.incrbyfloat(cost_key, 5)
+
+            token_resp = await http.post(
+                "/api/v1/auth/dev-token", json={"device_key": f"it-{uuid.uuid4().hex[:8]}"}
+            )
+            headers = {"Authorization": f"Bearer {token_resp.json()['data']['access_token']}"}
+
+            resp = await http.post(
+                "/api/v1/agent/chat", json={"message": "面试穿什么"}, headers=headers
+            )
+            assert resp.json()["code"] == 60003  # LLM_DAILY_BUDGET_EXCEEDED
+
+            # 被拒请求的配额净扣减 = 0：consume(-1, refunded) + refund(+1, committed)
+            rows = (await http.get("/api/v1/quota/ledger", headers=headers)).json()["data"]
+            consume = [r for r in rows if r["tx_type"] == "consume"]
+            refunds = [r for r in rows if r["tx_type"] == "refund"]
+            assert len(consume) == 1 and consume[0]["status"] == "refunded"
+            assert len(refunds) == 1 and refunds[0]["delta"] == 1
+            assert sum(r["delta"] for r in (*consume, *refunds)) == 0
+
+            await redis.delete(cost_key)
+    finally:
+        # 恢复全局 settings 缓存，避免本用例的环境覆盖渗进其它用例
+        monkeypatch.undo()
+        get_settings.cache_clear()