fix: dedup three-bucket projection + honest C3 error-handling note

2026-07-15 12:29:58 -04:00
parent aa10485b9f
commit 84b0b30213
2 changed files with 159 additions and 51 deletions
@@ -1,8 +1,11 @@
 """离线诊断编排集成测试（LLM 类：MD 产出到 tests/outputs/）。

 用 fake run_diagnosis + fake deps 覆盖编排契约，不实际调 LLM/VLM：
-    1. 续跑幂等 —— 已落盘题跳过，第二次无剩余则 run_diagnosis 收到空列表。
+    1. 续跑幂等 —— 已落盘题跳过；第二次 remaining 为空则直接早返回、不调用
+       run_diagnosis。
    2. 投影正确 —— defect→T2、lapse→T1，evolution_target 由 error_type 派生。
+    3. 三桶去重 —— 同题同时在 degraded 与 attributions 时按 degraded>infra>
+       attribution 优先级只落 1 行、tier=uncertain。

 测试结束把编排过程（remaining、各 tier 计数、投影样例）写入
 tests/outputs/test_baseline_diagnosis/<test>_<固定 ts>.md（CLAUDE.md §4.6）。
@@ -262,3 +265,90 @@ async def test_infra_and_degraded_projection(tmp_path, monkeypatch):
    )
    assert md_path.exists()
    store.close()
+
+
+@pytest.mark.asyncio
+async def test_degraded_overrides_attribution(tmp_path, monkeypatch):
+    """三桶去重：同题同时在 degraded 与 attributions 时只落 1 行且 tier=uncertain。
+
+    judge 解析失败会生成 degraded metrics，若仍是错题还会建 attribution，
+    故同一 question_id 可同时出现在两桶。按 degraded>infra>attribution 优先级，
+    该题必须只落 1 行、tier=uncertain（degraded 置位），attribution 行被跳过。
+    """
+    calls: list[tuple[str, ...]] = []
+
+    async def fake_run_diagnosis(
+        run_id,
+        questions,
+        tree_data,
+        llm,
+        run_log,
+        skill_store,
+        prompts,
+        *,
+        concurrency,
+        question_ids=None,
+        task_types=None,
+        only_incorrect=False,
+    ):
+        calls.append(tuple(question_ids or []))
+        from core.evolution.types import DiagnosisResult, ErrorAttribution
+
+        # q5 同时出现在 attributions（judge 解析失败仍建归因）与 degraded_question_ids。
+        return DiagnosisResult(
+            run_id=run_id,
+            error_attributions=[ErrorAttribution("q5", "reasoning_failure", None, "defect")],
+            infra_question_ids=[],
+            degraded_question_ids=["q5"],
+        )
+
+    monkeypatch.setattr("app.harness.baseline_diagnosis.run_diagnosis", fake_run_diagnosis)
+    deps = DiagnosisDeps(
+        run_log=_FakeRunLog(),
+        llm=_FakeLLM(),
+        skill_store=_FakeSkillStore(),
+        prompts=object(),
+        tree_data={},
+        concurrency=2,
+    )
+    store = SqliteDiagnosisSignalStore(str(tmp_path / "h.db"))
+    q_by_id = {"q5": _mk_q("q5")}
+
+    await run_baseline_diagnosis(
+        baseline_run_id="infer_adhoc",
+        diag_fingerprint="fp",
+        wrong_ids=["q5"],
+        questions=q_by_id,
+        store=store,
+        deps=deps,
+    )
+
+    all_rows = store.load("infer_adhoc", "fp")
+    assert len(all_rows) == 1  # 只落 1 行（无同 PK 双写）
+    row = all_rows[0]
+    assert row.question_id == "q5"
+    assert row.tier == "uncertain"  # degraded 优先级最高
+    assert row.degraded is True
+    assert row.error_type is None  # 走 degraded 投影而非 attribution
+    assert row.evolution_target is None
+
+    md_path = _write_md(
+        "test_degraded_overrides_attribution",
+        [
+            "# 离线诊断编排：三桶优先级去重",
+            "",
+            "## 任务描述",
+            "q5 同时出现在 error_attributions（defect）与 degraded_question_ids，"
+            "验证按 degraded>infra>attribution 优先级只落 1 行。",
+            "",
+            "## 落库结果",
+            "| question_id | 落库行数 | tier | degraded | error_type |",
+            "|---|---|---|---|---|",
+            f"| q5 | {len(all_rows)} | {row.tier} | {row.degraded} | {row.error_type} |",
+            "",
+            "## 结论",
+            "degraded 优先级最高，attribution 行被跳过 → 每题恰写一行、counts 恰计一次。",
+        ],
+    )
+    assert md_path.exists()
+    store.close()