feat: 性能监控基础设施、解密缓存及计费简化

- 新增 PerfRecorder 性能记录工具，支持采样率与慢请求日志
- 在请求管道中埋点：auth、body_read、json_parse、context_build、authorize、handle
- 流处理器增加 parse/conversion 耗时追踪与 perf_metrics 落库
- 解密服务添加 LRU 缓存，降低高频解密 CPU 开销
- 格式转换分层开关设计：全局 OFF 时回退到端点配置，而非一刀切拒绝
- 移除 shadow billing 模块，统一使用新计费引擎
- 新增 Codex 网关请求适配器（store=false、role 映射、include 补齐）
- endpoint 创建接口支持 body_rules 参数

tests/core/api_format/conversion/test_compatibility.py

@@ -45,19 +45,38 @@ def test_cli_format_convertible_when_converter_supports_full() -> None:
     assert reason is None
 
 
-def test_global_switch_disabled_blocks_conversion() -> None:
-    """全局开关关闭时阻止转换"""
+def test_global_switch_disabled_falls_back_to_endpoint() -> None:
+    """全局开关关闭时回退到端点配置（分层开关设计）"""
+    registry = MagicMock()
+    registry.can_convert_full.return_value = True
+
+    # 全局 OFF + 端点 enabled -> 允许（端点覆盖全局默认）
     ok, needs_conv, reason = is_format_compatible(
         "claude:chat",
         "openai:chat",
         endpoint_format_acceptance_config={"enabled": True},
         is_stream=False,
         effective_conversion_enabled=False,
+        registry=registry,
+    )
+    assert ok is True
+    assert needs_conv is True
+    assert reason is None
+
+
+def test_global_switch_disabled_blocks_when_endpoint_not_configured() -> None:
+    """全局开关关闭 + 端点未配置 -> 阻止转换"""
+    ok, needs_conv, reason = is_format_compatible(
+        "claude:chat",
+        "openai:chat",
+        endpoint_format_acceptance_config=None,
+        is_stream=False,
+        effective_conversion_enabled=False,
         registry=MagicMock(),
     )
     assert ok is False
     assert needs_conv is False
-    assert reason and "格式转换已禁用" in reason
+    assert reason and "未配置" in reason
 
 
 def test_endpoint_config_none_blocks_conversion() -> None:
@@ -212,8 +231,8 @@ def test_gemini_cli_to_gemini_no_conversion_needed() -> None:
     assert reason is None
 
 
-def test_claude_cli_to_claude_blocked_when_global_switch_disabled() -> None:
-    """透传格式（CLAUDE_CLI -> CLAUDE）也受全局开关限制"""
+def test_claude_cli_to_claude_allowed_when_endpoint_enabled() -> None:
+    """透传格式（CLAUDE_CLI -> CLAUDE）全局 OFF 时回退到端点配置"""
     ok, needs_conv, reason = is_format_compatible(
         "claude:cli",
         "claude:chat",
@@ -222,8 +241,10 @@ def test_claude_cli_to_claude_blocked_when_global_switch_disabled() -> None:
         effective_conversion_enabled=False,
         registry=MagicMock(),
     )
-    assert ok is False
-    assert reason and "格式转换已禁用" in reason
+    # 全局 OFF + 端点 enabled -> 允许（同族透传无需转换）
+    assert ok is True
+    assert needs_conv is False
+    assert reason is None
 
 
 def test_claude_cli_to_claude_blocked_when_endpoint_not_configured() -> None:
@@ -312,8 +333,8 @@ def test_openai_cli_to_openai_fails_without_converter() -> None:
     assert reason and "转换器" in reason
 
 
-def test_openai_cli_to_openai_blocked_when_global_switch_disabled() -> None:
-    """同族转换（OPENAI/OPENAI_CLI）也受全局开关限制"""
+def test_openai_cli_to_openai_allowed_when_endpoint_enabled() -> None:
+    """同族转换（OPENAI/OPENAI_CLI）全局 OFF 时回退到端点配置"""
     registry = MagicMock()
     registry.can_convert_full.return_value = True
 
@@ -325,9 +346,10 @@ def test_openai_cli_to_openai_blocked_when_global_switch_disabled() -> None:
         effective_conversion_enabled=False,  # 全局开关关闭
         registry=registry,
     )
-    assert ok is False
-    assert needs_conv is False
-    assert reason and "格式转换已禁用" in reason
+    # 全局 OFF + 端点 enabled -> 允许（需要转换）
+    assert ok is True
+    assert needs_conv is True
+    assert reason is None
 
 
 def test_openai_cli_to_openai_blocked_when_endpoint_disabled() -> None:

tests/services/billing/test_shadow_billing.py

@@ -1,112 +0,0 @@
-from unittest.mock import MagicMock
-
-import pytest
-
-from src.config.settings import config
-from src.services.billing.schema import BillingSnapshot, CostResult
-from src.services.billing.shadow import CostBreakdown, ShadowBillingService
-
-
-class TestShadowBillingServiceModeResolution:
-    def test_get_engine_mode_exact_and_wildcard(self, monkeypatch: pytest.MonkeyPatch) -> None:
-        monkeypatch.setattr(config, "billing_engine", "legacy", raising=False)
-        monkeypatch.setattr(
-            config,
-            "billing_engine_overrides",
-            '{"anthropic/*": "shadow", "openai/gpt-4o": "new"}',
-            raising=False,
-        )
-
-        svc = ShadowBillingService(MagicMock())
-        assert svc.get_engine_mode("openai", "gpt-4o") == "new"
-        assert svc.get_engine_mode("anthropic", "claude-3-5-sonnet") == "shadow"
-        assert svc.get_engine_mode("other", "x") == "legacy"
-
-
-class TestShadowBillingServiceExecution:
-    def test_legacy_mode_skips_new_engine(self, monkeypatch: pytest.MonkeyPatch) -> None:
-        monkeypatch.setattr(config, "billing_engine", "legacy", raising=False)
-        monkeypatch.setattr(config, "billing_engine_overrides", "{}", raising=False)
-
-        svc = ShadowBillingService(MagicMock())
-        # Guard: if new engine calculate gets called, fail.
-        svc._new_billing = MagicMock()
-        svc._new_billing.calculate.side_effect = AssertionError(
-            "new engine should not run in legacy mode"
-        )
-
-        legacy_truth = CostBreakdown(
-            input_cost=0.1,
-            output_cost=0.2,
-            cache_creation_cost=0.0,
-            cache_read_cost=0.0,
-            request_cost=0.0,
-            total_cost=0.3,
-        )
-
-        res = svc.calculate_with_shadow(
-            provider="openai",
-            provider_id="p-1",
-            model="gpt-4o",
-            task_type="chat",
-            api_format="openai:chat",
-            input_tokens=1,
-            output_tokens=1,
-            legacy_truth=legacy_truth,
-            is_failed_request=False,
-        )
-
-        assert res.engine_mode == "legacy"
-        assert res.truth_engine == "legacy"
-        assert res.shadow_snapshot is None
-        assert res.truth_breakdown.total_cost == 0.3
-
-    def test_shadow_mode_returns_snapshot_and_keeps_legacy_truth(
-        self, monkeypatch: pytest.MonkeyPatch
-    ) -> None:
-        monkeypatch.setattr(config, "billing_engine", "shadow", raising=False)
-        monkeypatch.setattr(config, "billing_engine_overrides", "{}", raising=False)
-        monkeypatch.setattr(config, "billing_diff_threshold_usd", 0.0001, raising=False)
-
-        svc = ShadowBillingService(MagicMock())
-
-        # Stub new engine output
-        snapshot = BillingSnapshot(
-            resolved_dimensions={"input_tokens": 1},
-            resolved_variables={"input_price_per_1m": "3.0"},
-            cost_breakdown={"input_cost": 0.003},
-            total_cost=0.003,
-            status="complete",
-            calculated_at="2026-02-02T00:00:00Z",
-        )
-        svc._new_billing = MagicMock()
-        svc._new_billing.calculate.return_value = CostResult(
-            cost=0.003, status="complete", snapshot=snapshot
-        )
-
-        legacy_truth = CostBreakdown(
-            input_cost=0.004,
-            output_cost=0.0,
-            cache_creation_cost=0.0,
-            cache_read_cost=0.0,
-            request_cost=0.0,
-            total_cost=0.004,
-        )
-
-        res = svc.calculate_with_shadow(
-            provider="openai",
-            provider_id="p-1",
-            model="gpt-4o",
-            task_type="chat",
-            api_format="openai:chat",
-            input_tokens=1,
-            output_tokens=0,
-            legacy_truth=legacy_truth,
-            is_failed_request=False,
-        )
-
-        assert res.engine_mode == "shadow"
-        assert res.truth_engine == "legacy"
-        assert res.shadow_snapshot is not None
-        assert res.truth_breakdown.total_cost == 0.004
-        assert "diff_usd" in res.comparison

tests/services/test_format_conversion_candidate_selection.py

@@ -67,7 +67,11 @@ async def test_build_candidates_allows_cross_format_when_endpoint_accepts_and_ov
 
 
 @pytest.mark.asyncio
-async def test_build_candidates_blocks_cross_format_when_master_switch_off() -> None:
+async def test_build_candidates_allows_cross_format_when_global_off_but_endpoint_enabled() -> None:
+    """
+    分层开关设计：全局 OFF 时回退到端点配置
+    - 全局 OFF + 端点 enabled=True -> 允许（端点覆盖全局默认）
+    """
     register_default_normalizers()
 
     scheduler = CacheAwareScheduler()
@@ -85,6 +89,40 @@ async def test_build_candidates_blocks_cross_format_when_master_switch_off() ->
     ]
     provider.api_keys = [_mock_key("k1", ["openai:chat"])]
 
+    candidates = await scheduler._build_candidates(
+        db=MagicMock(),
+        providers=[provider],
+        client_format="claude:chat",
+        model_name="dummy-model",
+        affinity_key=None,
+        global_conversion_enabled=False,  # 全局开关关闭，但端点配置允许
+    )
+
+    # 新设计：全局 OFF 时回退到端点配置，端点 enabled=True 则允许
+    assert len(candidates) == 1
+    assert candidates[0].needs_conversion is True
+    assert candidates[0].provider_api_format == "openai:chat"
+
+
+@pytest.mark.asyncio
+async def test_build_candidates_blocks_cross_format_when_global_off_and_endpoint_not_configured() -> (
+    None
+):
+    """
+    分层开关设计：全局 OFF + 端点未配置 -> 阻止
+    """
+    register_default_normalizers()
+
+    scheduler = CacheAwareScheduler()
+    scheduler._check_model_support = AsyncMock(return_value=(True, None, None, {"m"}))  # type: ignore[method-assign]
+    scheduler._check_key_availability = MagicMock(return_value=(True, None, None))  # type: ignore[method-assign]
+
+    provider = MagicMock()
+    provider.name = "p1"
+    provider.enable_format_conversion = False
+    provider.endpoints = [_mock_endpoint("openai:chat", None)]  # 端点未配置格式接受策略
+    provider.api_keys = [_mock_key("k1", ["openai:chat"])]
+
     candidates = await scheduler._build_candidates(
         db=MagicMock(),
         providers=[provider],
@@ -94,6 +132,7 @@ async def test_build_candidates_blocks_cross_format_when_master_switch_off() ->
         global_conversion_enabled=False,  # 全局开关关闭
     )
 
+    # 全局 OFF + 端点未配置 -> 阻止
     assert candidates == []