fix: LLM端点故障告警只数最终失败——hermes重试序列(attempt1/2的'API call failed'+'Retrying')不再计入(一次瞬时抖动=3条WARNING误报5次失败,09:54事件)

deploy/profile-scripts/mofin_health.py

@@ -1167,6 +1167,14 @@ def build_report():
                     if _ts.timestamp() < _scan_cutoff:
                         continue
                     if any(_p in _line for _p in _LLM_ERR_PATTERNS):
+                        # 2026-08-25 只数最终失败：hermes 一次抖动打3条WARNING(attempt1/2/3)，
+                        # 重试成功不算故障——attempt 1/2 的"API call failed"和"Retrying"跳过
+                        if "Retrying API call" in _line:
+                            continue
+                        import re as _re2
+                        _att = _re2.search(r"attempt (\d+)/(\d+)", _line)
+                        if _att and _att.group(1) != _att.group(2):
+                            continue  # 非最终次失败=重试中，跳过
                         _recent_errs.append({
                             "log": _lf.name, "time": _ts.strftime("%H:%M"),
                             "msg": _line[20:180],