fix(llm): 收编波五件——L4/L5 累积降级链+配平起点避字符串内左括号+响应体单次解析连接复用+400 自愈不占重试预算+usage 上游未给 None 化 (审计 P3-1..5 收编) [vps] [no-doc]

tests/api/test_llm_client.py

@@ -203,3 +203,81 @@ class TestWantUsage:
         client, _, _ = _make(lambda r: _ok())
         out = asyncio.run(client.chat_json(MSGS))
         assert out == {"a": 1} and not isinstance(out, tuple)
+
+
+class TestEfficiency:
+    """P3-3: 同一 200 响应体只解析一次 + 整轮对话(含重试/strict)复用单个
+    AsyncClient(此前每 attempt 新建连接池,响应体 json.loads 两遍)."""
+
+    def test_200_body_parsed_once(self, monkeypatch):
+        import sanguo_api.llm.client as client_mod
+
+        class _Shim:
+            count = 0
+
+            @staticmethod
+            def loads(*a, **k):
+                _Shim.count += 1
+                return real_loads(*a, **k)
+
+        real_loads = client_mod.json.loads
+        monkeypatch.setattr(client_mod, "json", _Shim)
+        client, _, _ = _make(lambda r: _ok())
+        assert asyncio.run(client.chat_json(MSGS)) == {"a": 1}
+        assert _Shim.count == 1
+
+    def test_async_client_reused_across_attempts(self, monkeypatch):
+        import sanguo_api.llm.client as client_mod
+
+        real_cls = client_mod.httpx.AsyncClient
+        built = []
+
+        class _Counting(real_cls):
+            def __init__(self, *a, **k):
+                built.append(1)
+                super().__init__(*a, **k)
+
+        monkeypatch.setattr(client_mod.httpx, "AsyncClient", _Counting)
+
+        def h(request):
+            calls.append(request)
+            if len(calls) == 1:
+                return httpx.Response(429, json={"error": "rate"})
+            return _ok()
+
+        client, calls, _ = _make(h)
+        assert asyncio.run(client.chat_json(MSGS)) == {"a": 1}
+        assert len(calls) == 2 and len(built) == 1
+
+
+class TestHealBudget:
+    def test_400_self_heal_does_not_consume_retry_budget(self):
+        """P3-4: 参数自愈重发不计 429/5xx 重试预算——自愈吃掉 attempts 后
+        真 429 提前「重试耗尽」,退避语义混杂."""
+        def h(request):
+            calls.append(request)
+            import json as j
+            if "response_format" in j.loads(request.content):
+                return httpx.Response(400,
+                                      text='{"error":"response_format x"}')
+            if len(calls) < 4:
+                return httpx.Response(429, json={"error": "rate"})
+            return _ok()
+
+        client, calls, sleeps = _make(h)
+        assert asyncio.run(client.chat_json(MSGS)) == {"a": 1}
+        assert len(calls) == 4 and sleeps == [2.0, 4.0]
+
+
+class TestUsageMissing:
+    def test_missing_usage_returns_none_not_empty_dict(self):
+        """P3-5: 上游没给 usage 返 None,与「给了但 0 token」可区分
+        (漏斗计量口径);None 不得伪装成 {}."""
+        def h(request):
+            import json as j
+            body = {"choices": [{"message": {"content": '{"a": 1}'}}]}
+            return httpx.Response(200, text=j.dumps(body))
+
+        client, _, _ = _make(h)
+        obj, usage = asyncio.run(client.chat_json(MSGS, want_usage=True))
+        assert obj == {"a": 1} and usage is None

tests/api/test_llm_json_utils.py

@@ -49,3 +49,15 @@ def test_non_str_input_raises_json_parse_error():
     for bad in (None, 42, {"a": 1}, ["a"], b"{}"):
         with pytest.raises(JSONParseError):
             robust_json_parse(bad)
+
+
+def test_l4_l5_fixes_cumulative():
+    """P3-1: L4 反斜杠修复与 L5 尾逗号清洗须累积叠加——并行候选下「LaTeX
+    残留+尾逗号」同现时两级各自单独修都不够,整段降级链全败(假阴性 502)."""
+    assert robust_json_parse('{"s": "\\(x\\)",}') == {"s": "(x)"}
+
+
+def test_balanced_start_skips_brace_inside_string():
+    """P3-2: 首个 { 落在字符串内时旧 find("{") 起点错位,配平块从头就坏,
+    后续 L4/L5 全被拖死——起点扫描须字符串感知(复用 in_str/esc 状态机)."""
+    assert robust_json_parse('他说"{不是json" 然后 {"a": 1}') == {"a": 1}