fix(auto-repair): prefer exact playbooks and fail failed steps

apps/api/src/services/auto_repair_service.py

@@ -321,7 +321,16 @@ class AutoRepairService:
             )
 
         # 4. 檢查最佳匹配
-        best_match = recommendations[0]
+        best_match = self._select_best_recommendation(recommendations, symptoms)
+        if best_match is not recommendations[0]:
+            logger.warning(
+                "auto_repair_exact_match_prioritized",
+                incident_id=incident.incident_id,
+                selected_playbook_id=best_match.playbook.playbook_id,
+                original_playbook_id=recommendations[0].playbook.playbook_id,
+                selected_similarity=best_match.similarity_score,
+                original_similarity=recommendations[0].similarity_score,
+            )
 
         # 2026-04-07 Claude Code: 統帥指令「直接全部跳成自動修復」
         # 移除: 相似度門檻、is_high_quality 門檻、冷啟動機制、風險等級門檻
@@ -416,6 +425,8 @@ class AutoRepairService:
                 executed_steps.append(
                     f"Step {step.step_number}: {step.command[:50]}... -> {step_result}"
                 )
+                if self._is_step_failure_result(step_result):
+                    raise RuntimeError(f"Step {step.step_number} failed: {step_result}")
 
             # 更新 Playbook 統計
             await self._playbook_service.record_execution(
@@ -697,6 +708,44 @@ class AutoRepairService:
             keywords=keywords[:10],
         )
 
+    def _select_best_recommendation(
+        self,
+        recommendations,
+        symptoms: SymptomPattern,
+    ):
+        """Prefer deterministic alert/service matches over fuzzy similarity only.
+
+        A higher fuzzy score must not outrank a playbook that explicitly names the
+        firing alert or affected service. Live-fire T16 proved that this can route
+        a safe K8s canary into an unrelated host diagnostic playbook.
+        """
+
+        symptom_alerts = {str(name) for name in (symptoms.alert_names or []) if name}
+        symptom_services = {
+            str(service) for service in (symptoms.affected_services or []) if service
+        }
+
+        def _priority(recommendation) -> tuple[int, int, float]:
+            pattern = recommendation.playbook.symptom_pattern
+            playbook_alerts = {
+                str(name) for name in (pattern.alert_names or []) if name
+            }
+            playbook_services = {
+                str(service) for service in (pattern.affected_services or []) if service
+            }
+            alert_exact = int(bool(symptom_alerts & playbook_alerts))
+            service_exact = int(bool(symptom_services & playbook_services))
+            return (alert_exact, service_exact, float(recommendation.similarity_score or 0.0))
+
+        return max(recommendations, key=_priority)
+
+    @staticmethod
+    def _is_step_failure_result(step_result: str) -> bool:
+        """Treat executor-declared failures as failed auto-repair executions."""
+
+        normalized = (step_result or "").strip().upper()
+        return normalized.startswith("FAILED:") or normalized == "UNKNOWN_ACTION_TYPE"
+
     def _get_max_risk_level(self, playbook: Playbook) -> RiskLevel:
         """取得 Playbook 中最高的風險等級"""
         risk_order = {