Merge branch 'main' into feat-compare-fail-alert

冲突解决:compare_alert_worker.py / test_compare_alert_stuck_worker.py /
2026-08-05-compare-stuck-detection-design.md 三处均以本地(ours)为准——
保留「cancelled 可读但无卡点时回退耗时兜底(超长放弃不漏报)」的修复,
覆盖 main 带入的旧行为(可读无卡点→不报)。

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

app/core/compare_alert_worker.py

@@ -102,10 +102,12 @@ def build_hits(
     unrecognized_keywords: tuple[str, ...],
     biz_exclude_keywords: tuple[str, ...],
 ) -> list[AlertHit]:
-    """编排:cancelled 走 trace 优先(读到确认没卡则不报,读不到回退保底);failed 附卡点。
+    """编排:cancelled 先读 trace,判出原地卡点就带卡点报 T5,否则(可读没卡点/读不到)一律回退
+    耗时/步数兜底——超长放弃照报(卡点列留空);failed 命中后附末段卡点。
 
-    trace 读取限量 max_trace_reads 次/轮;任何 trace 异常都在 trace_stuck 内部降级为
-    「读不到」,cancelled 因而回退保底、failed 不附卡点,绝不影响报警发送。
+    trace 读取限量 max_trace_reads 次/轮;任何 trace 异常都在 trace_stuck 内部降级为「读不到」,
+    cancelled 因而走耗时兜底、failed 不附卡点,绝不影响报警发送。trace 只做「锦上添花」标注卡点,
+    绝不因「可读但没判出卡点」把超长放弃吞掉(线上 trace 几乎总可读,否则 total_ms 阈值形同虚设)。
     """
     base = Path(work_log_dir) if work_log_dir else None
     reads = 0
@@ -120,13 +122,13 @@ def build_hits(
                         td, threshold=stuck_threshold, max_tail=max_tail
                     )
                     reads += 1
-            if res is not None and res.readable:
-                if res.points:
-                    stuck = "、".join(_fmt_stuck(sp) for sp in res.points)
-                    hit = _dc_replace(make_hit(rec, "T5", "深度放弃"), stuck_point=stuck)
-                else:
-                    hit = None  # 读到且确认没卡 → 不报
+            if res is not None and res.readable and res.points:
+                stuck = "、".join(_fmt_stuck(sp) for sp in res.points)
+                hit = _dc_replace(make_hit(rec, "T5", "深度放弃"), stuck_point=stuck)
             else:
+                # trace 判出卡点 → 上面带卡点报。其余一律回退耗时/步数兜底:可读但没判出卡点、
+                # 读不到、无 trace,都过 total_ms/step 阈值——超长放弃照报(卡点列留空),不再因
+                # 「trace 可读但不原地卡」把超长放弃整条吞掉(线上 trace 几乎总可读,否则耗时阈值形同虚设)。
                 hit = classify_cancelled_fallback(
                     rec,
                     cancelled_ms_threshold=cancelled_ms_threshold,

docs/superpowers/specs/2026-08-05-compare-stuck-detection-design.md

@@ -97,14 +97,19 @@ def last_step(trace_dir: Path) -> StuckPoint | None:
 ```
 worker 对 cancelled 候选:
   res = read_stuck_points(dir)
-  if not res.readable:          # 读不到 trace（目录被清/生产一时读不到）→ 回退保底
-      >90s或>30步 → T5「深度放弃·等待Xs/Y步」; 否则不报
-  elif res.points:              # 读到且有卡死平台 → 报卡死
-      报 T5, reason = "卡在 " + "、".join(f"{平台}·{环节}" for res.points)
-  else:                         # 读到且没卡死（末段在推进 = 正常深度使用后退出）→ 不报
-      不报
+  if res.readable and res.points:   # 读到且有卡死平台 → 报卡死（带卡点环节）
+      报 T5, reason="深度放弃", stuck_point = "、".join(f"{平台}·{环节}" for res.points)
+  else:                             # 其余一律回退耗时/步数兜底(见下方 2026-08-06 修订)
+      >90s或>30步 → T5「深度放弃」(卡点列留空); 否则不报
 ```
 
+> **2026-08-06 修订(compare-fail-alert 排查)**:原设计「读到且没卡死 → 不报」在线上是死路——
+> **线上 trace 几乎总可读**(WORK_LOG_DIR 已配、同机直读),于是耗时兜底那条分支基本永不触发,
+> `total_ms>90s` 阈值形同虚设,**超长放弃(实测 113s / 516s)一条都报不出来**。改为:只有「判出卡点」
+> 独占带卡点的 T5;**其余(可读没卡点 / 读不到 / 无 trace)一律回退耗时兜底**,超长照报(卡点留空)。
+> 权衡:这会重新引入第 1 节「误报」——用户正常浏览 90s+ 后退出也会报。若噪音大,调高
+> `COMPARE_ALERT_CANCELLED_MS_THRESHOLD`(如 180s/300s)收敛,不动代码。
+
 ### 6.2 failed（T1/T2/T6，判定不变 + 附卡点）
 
 ```

tests/test_compare_alert_stuck_worker.py

@@ -48,8 +48,9 @@ def test_cancelled_stuck_reports_via_trace(tmp_path):
     assert "帧" in hits[0].stuck_point
 
 
-def test_cancelled_readable_not_stuck_no_report(tmp_path):
-    # trace 确认没卡(在推进);即便 total_ms/step 超阈值也不报(信 trace,不回退保底)
+def test_cancelled_readable_not_stuck_long_duration_reports(tmp_path):
+    # trace 可读但没判出原地卡点:仍过耗时/步数阈值兜底,超长(>90s)照报 T5,卡点列留空。
+    # (线上 trace 几乎总可读,若不回退则 total_ms 阈值形同虚设、超长放弃永不报——见 compare-fail-alert 排查。)
     p = tmp_path / "20260804_y" / "eleme"
     _frame(p, 0, "set_address", "home")
     for i in range(1, 6):
@@ -57,6 +58,21 @@ def test_cancelled_readable_not_stuck_no_report(tmp_path):
     rec = _Rec(status="cancelled", trace_url="https://x/traces/20260804_y/",
                total_ms=95000, step_count=40)
     hits = build_hits([rec], work_log_dir=str(tmp_path), **_KW)
+    assert len(hits) == 1
+    assert hits[0].alert_type == "T5"
+    assert hits[0].reason == "深度放弃"
+    assert hits[0].stuck_point is None  # 没卡点 → 卡片卡点列显 "-"
+
+
+def test_cancelled_readable_not_stuck_short_no_report(tmp_path):
+    # trace 可读没卡点、且耗时/步数都没超阈值 → 正常早退,不报(兜底阈值把住,不误报)。
+    p = tmp_path / "20260804_ys" / "eleme"
+    _frame(p, 0, "set_address", "home")
+    for i in range(1, 6):
+        _frame(p, i, "enter_store", "store")
+    rec = _Rec(status="cancelled", trace_url="https://x/traces/20260804_ys/",
+               total_ms=5000, step_count=3)
+    hits = build_hits([rec], work_log_dir=str(tmp_path), **_KW)
     assert hits == []