test: keep cancelled embedding probe rounds incomplete

research-wiki/findings/2026-09-09-134-thinking-contracts-validation.md

@@ -169,3 +169,18 @@ date: 2026-09-09
 pi-lens 仍报非 conda 解释器缺 httpx／pytest／dotenv／pydantic 及旧 StrEnum 噪音；按任务授权记录，不添加 ignore、不改枚举、不扩环境修复范围。实际 conda 解释器为 `/home/iomgaa/miniconda3/envs/PolyGateway/bin/python`，本会话导入四依赖成功（httpx 0.28.1、pytest 9.1.1、python-dotenv 1.2.3、pydantic 2.13.4）。conda 启动器自身另有 base Python 3.13 的 RequestsDependencyWarning；未静音，不宣称输出零告警，测试进程与静态门实际退出0。
 
 本修复不补写真正缺失的历史报告、不改变 embedding 能力判据；真实服务、slow、下游与发布证据仍由后续验收负责。
+
+## 独立审查补正：取消不计完成轮（起点 7f6a824）
+
+独立 verifier 指出：probe 的 finally 无条件写 `completed_rounds=1`，但 CancelledError 穿透时仍是 `FAIL／轮次未完成`，分母记录自相矛盾。已对照源码并在真实消费者复现，接受该问题；上节“已收尾即完成”的措辞不适用于取消，本节修正为**取得正常成功或普通异常分类终态才算完成**，不是 finally 执行过就完成。
+
+最小修复仅在 probe 初始化 `completed_rounds=0`，成功判定或普通异常分类返回后置1；finally 写实际计数。取消仍穿透、计数保留0，不新增捕获 BaseException、不动生产 API／版本／分类器。扩展原消费者参数化测试增加两型号取消节点：真实 task 在 MockTransport 进入等待后由调用方 cancel，断言 CancelledError 穿透、task.cancelled、报告 FAIL／未完成、planned=1／completed=0、原调用关联及客户端关闭；其他8例保持完成1。所有节点只替换外部 HTTP，不联网、不付费、不跑 slow。
+
+| 命令（pytest 前缀为 `conda run --no-capture-output -n PolyGateway`） | 本轮实际证据（tests/outputs/134/，各有 .log／.exit） |
+| --- | --- |
+| `pytest tests/unit/test_live_evidence.py -k 'embed_probe_report and cancelled' -q`，修复前 | `embed-cancel-red`：2 failed／137 deselected，exit1；两例均先验证取消穿透、报告存在及资源关闭，再因 `completed_rounds` 实际1而期望0失败 |
+| `pytest tests/unit/test_live_evidence.py -k embed_probe_report -q`，修复后 | `embed-cancel-green`：10 passed／129 deselected，exit0；覆盖原8例与新增2例 |
+| `pytest tests/unit/ -q` | `embed-cancel-unit`：1385 passed，4.08秒，exit0 |
+| `make check` | `embed-cancel-check`：格式／ruff通过，import-linter 1 kept／0 broken，exit0 |
+
+既有非 conda LSP 误报继续只记录（本次额外将 `asyncio.timeout` 误判为缺属性）；conda pytest 实际可执行，base RequestsDependencyWarning 未静音。此处是针对独立审查问题的实现及自验，修复后独立复核仍交父会话；不冒称审查门或版本验收已通过。原 slow 日志不改写。

tests/e2e/test_embed_probe.py

@@ -53,6 +53,7 @@ async def test_probe_real_gateway_embeddings():
     transport = ObservedTransport(real, capture)
     run_id, parent, call_id = uuid4().hex, uuid4().hex, uuid4().hex
     verdict = LiveVerdict("FAIL", "轮次未完成")
+    completed_rounds = 0
     try:
         with capture.round_context(session_id=run_id, parent_call_id=parent):
             try:
@@ -65,10 +66,12 @@ async def test_probe_real_gateway_embeddings():
                 assert len(events) == 1 and request_is_valid(events[0])
                 assert result.dim > 0 and len(result.vectors) == 1
                 verdict = LiveVerdict("PASS", "向量形状与实发请求合格")
+                completed_rounds = 1
             except Exception as error:
                 verdict = classify_live_failure(
                     error, capture.attempts(session_id=run_id, parent_call_id=parent)
                 )
+                completed_rounds = 1
             finally:
                 write_live_round(
                     Path("tests/outputs/134/live"),
@@ -79,7 +82,7 @@ async def test_probe_real_gateway_embeddings():
                         "requested_model": source.model,
                         "provider": source.provider,
                         "planned_rounds": 1,
-                        "completed_rounds": 1,
+                        "completed_rounds": completed_rounds,
                         "status": verdict.status,
                         "reason": verdict.reason,
                         "session_id": run_id,

tests/unit/test_live_evidence.py

@@ -692,6 +692,7 @@ async def test_round_consumer_keeps_first_success_when_second_assertion_fails(tm
         ("503", "FAIL", "TransientError"),
         ("404", "UNCOVERED", "RequestRejectedError"),
         ("request_error", "FAIL", "TransientError"),
+        ("cancelled", "FAIL", None),
     ],
 )
 async def test_embed_probe_report_keeps_actual_source_and_round_identity(
@@ -731,18 +732,22 @@ async def test_embed_probe_report_keeps_actual_source_and_round_identity(
     original_factory = LiveCapture.client_factory
     clients = []
     calls = []
+    reached = asyncio.Event()
 
     def factory(capture, source):
         """保留真实取证 hooks、源与逻辑 ID，只隔离网络出口。"""
         client = original_factory(capture, source)
 
-        def handler(request):
+        async def handler(request):
             """提供完整成功／错误样本，敏感回显不得进入报告。"""
             payload = json.loads(request.content)
             assert payload["model"] == model == source.model
             assert source.model != env["LLM__MINIMAX__1__MODEL"]
             assert request.headers["Authorization"] == f"Bearer {_SECRET}"
             calls.append((source, capture._round.get(), capture._attempt.get().call_id))
+            if outcome == "cancelled":
+                reached.set()
+                await asyncio.Future()
             if outcome == "request_error":
                 raise httpx.ConnectError(_SECRET + _PROMPT, request=request)
             if outcome == "success":
@@ -765,7 +770,19 @@ async def test_embed_probe_report_keeps_actual_source_and_round_identity(
 
     monkeypatch.setattr(LiveCapture, "client_factory", factory)
     probe = namespace["test_probe_real_gateway_embeddings"]
-    if status == "PASS":
+    if outcome == "cancelled":
+        task = asyncio.create_task(probe())
+        try:
+            async with asyncio.timeout(5):
+                await reached.wait()
+                task.cancel()
+                with pytest.raises(asyncio.CancelledError):
+                    await task
+            assert task.cancelled()
+        finally:
+            task.cancel()
+            await asyncio.gather(task, return_exceptions=True)
+    elif status == "PASS":
         await probe()
     elif status == "UNCOVERED":
         with pytest.raises(pytest.skip.Exception):
@@ -785,7 +802,10 @@ async def test_embed_probe_report_keeps_actual_source_and_round_identity(
     assert row["status"] == status
     assert row["requested_model"] == source.model == model
     assert row["provider"] == source.provider == "minimax"
-    assert row["planned_rounds"] == row["completed_rounds"] == 1
+    assert row["planned_rounds"] == 1
+    assert row["completed_rounds"] == (0 if outcome == "cancelled" else 1)
+    if outcome == "cancelled":
+        assert row["reason"] == "轮次未完成"
     assert row["session_id"] == paths[0].parent.name == session_id
     assert row["parent_call_id"] == parent_call_id
     assert len({session_id, parent_call_id, call_id}) == 3
@@ -804,12 +824,13 @@ async def test_embed_probe_report_keeps_actual_source_and_round_identity(
             "503": 503,
             "404": 404,
             "request_error": 0,
+            "cancelled": 0,
         }[outcome]
     )
     if outcome in {"503", "404"}:
         assert event["error_body_complete"] is True
         assert event["machine_type"] == "model_not_found"
-    elif outcome == "request_error":
+    elif outcome in {"request_error", "cancelled"}:
         assert "无可配对响应" in row["evidence_notes"]