test: unit suite for the gateway invariants plus an integration slice

Covers routing, failover, the circuit breaker, structured-output retries, the
cache-prefix validator, cost computation including cached-token pricing, the
daily budget breaker, the quota protocol, the task state machine and log
redaction. HTTP is mocked -- no unit test spends money. The integration test
runs the real slice against PostgreSQL and Redis with a faked provider.

tests/unit/test_routing.py

@@ -0,0 +1,77 @@
+"""Scene routing configuration (架构文档 §8.1: 改模型不改代码)."""
+
+from __future__ import annotations
+
+import pytest
+
+from packages.llm.errors import SceneNotRoutedError
+from packages.llm.routing import SceneRouter, build_scene_configs
+from packages.llm.types import Scene
+
+
+def test_defaults_put_the_primary_first_and_the_backup_second() -> None:
+    configs = build_scene_configs(primary_provider="deepseek", backup_provider="zhipu")
+    assert configs[Scene.AGENT_CHAT].providers == ("deepseek", "zhipu")
+
+
+def test_a_missing_backup_yields_a_single_provider_chain() -> None:
+    configs = build_scene_configs(primary_provider="deepseek", backup_provider=None)
+    assert configs[Scene.AGENT_CHAT].providers == ("deepseek",)
+
+
+def test_every_scene_is_routed_and_bound_to_a_prompt_version() -> None:
+    configs = build_scene_configs(primary_provider="deepseek")
+    assert set(configs) == set(Scene)
+    for config in configs.values():
+        assert config.prompt_id
+        assert config.prompt_version
+
+
+def test_intent_classification_is_critical_so_budget_pressure_cannot_disable_it() -> None:
+    configs = build_scene_configs(primary_provider="deepseek")
+    assert configs[Scene.INTENT_CLASSIFY].critical is True
+    assert configs[Scene.AGENT_CHAT].critical is False
+
+
+def test_agent_chat_output_budget_matches_the_token_plan() -> None:
+    # 架构文档 §7.2: 输出 ≤ 800
+    assert build_scene_configs(primary_provider="x")[Scene.AGENT_CHAT].max_tokens <= 800
+
+
+def test_overrides_change_routing_without_a_code_change() -> None:
+    configs = build_scene_configs(
+        primary_provider="deepseek",
+        backup_provider="zhipu",
+        overrides={
+            "agent_chat": {
+                "max_tokens": 1200,
+                "prompt_version": "v2",
+                "models": {"deepseek": "deepseek-v4-pro"},
+                "providers": ["zhipu", "deepseek"],
+            }
+        },
+    )
+    config = configs[Scene.AGENT_CHAT]
+    assert config.max_tokens == 1200
+    assert config.prompt_version == "v2"
+    assert config.providers == ("zhipu", "deepseek")
+    assert config.model_for("deepseek", "fallback") == "deepseek-v4-pro"
+    assert config.model_for("zhipu", "fallback") == "fallback"
+    # Untouched scenes keep their defaults.
+    assert configs[Scene.REPORT_COPY].max_tokens == 3000
+
+
+def test_structured_retry_budget_is_configurable() -> None:
+    configs = build_scene_configs(primary_provider="x", structured_max_retries=5)
+    assert configs[Scene.AGENT_CHAT].structured_max_retries == 5
+
+
+def test_router_raises_for_an_unconfigured_scene() -> None:
+    router = SceneRouter({})
+    with pytest.raises(SceneNotRoutedError):
+        router.resolve(Scene.AGENT_CHAT)
+
+
+def test_router_lists_configured_scenes() -> None:
+    router = SceneRouter(build_scene_configs(primary_provider="x"))
+    assert set(router.scenes()) == set(Scene)