fix(aiops): route backup decisions through ssh

docs/adr/ADR-058-host-auto-repair-ssh-whitelist.md

@@ -140,6 +140,48 @@ MoWoooWorkDown → Jaccard 匹配 momo-app-down-repair → SSH ollama@192.168.0.
 
 ---
 
+## Appendix B — Backup Failure Route Parity (2026-05-01)
+
+`HostBackupFailed` / backup 類告警的 `alert_category` 是 `backup_failure`。它必須在所有 host-layer 自動化路徑與 `host_resource` 同級處理：
+
+| Layer | 必須行為 |
+|-------|----------|
+| Alertmanager rule-first | YAML `SSH_DIAGNOSE` / `NO_ACTION` 不進 LLM 覆寫 |
+| AutoRepairService | `backup_failure` 視為 host-layer，拒絕 K8s Playbook fallback |
+| DecisionManager | 非 `kubectl` 動作在 kubectl parser 前路由 SSH MCP |
+| DecisionManager K8s guard | `backup_failure` 產生 `kubectl` 時降級 emergency escalation |
+| Telegram buttons | `backup_failure` 顯示只讀診斷按鈕：主機磁碟、備份 Job、Velero 狀態 |
+
+2026-05-01 根因：DecisionManager SSH route 只含 `infrastructure` / `host_resource`，漏掉 `backup_failure`，導致 `ssh 192.168.0.110 '...;...'` 類只讀診斷動作掉進 `parse_kubectl_action()`，被 `forbidden_shell_metachar` 擋下。
+
+同日按鈕審計也發現 category button 的 friendly provider name 會漂移：`callback_action_spec.yaml` 使用 `k8s` / `ssh`，但 MCP registry 實際名稱是 `kubernetes` / `ssh_host`。Dispatcher 必須正規化 provider alias，否則卡片按鈕會顯示但執行時變成 `provider_not_found`。
+
+### Runtime 權限基準
+
+- K8s Secret:
+  - `awoooi-repair-ssh-key` 掛載 `/etc/repair-ssh/`
+  - `awoooi-repair-known-hosts` 掛載 `/etc/repair-known-hosts/`
+  - `ssh-mcp-key` 掛載 `/run/secrets/ssh_mcp_key` 與 `/etc/ssh-mcp/known_hosts`
+- Remote `authorized_keys`:
+  - `wooo@192.168.0.110`
+  - `wooo@192.168.0.120`
+  - `wooo@192.168.0.121`
+  - `ollama@192.168.0.188`
+- NetworkPolicy egress:
+  - `192.168.0.110:22`
+  - `192.168.0.120:22`
+  - `192.168.0.121:22`
+  - `192.168.0.188:22`
+- `awoooi-executor` RBAC:
+  - read `jobs.batch`, `cronjobs.batch`
+  - read `persistentvolumeclaims`
+  - read Velero `backups`, `backupstoragelocations`, `backuprepositories`, `podvolumebackups`, `podvolumerestores`, `restores`, `schedules`
+  - patch `statefulsets.apps` / `daemonsets.apps` only for safe rollout restart
+
+If SSH MCP fails, the incident must not silently become a manual approval card; it must raise the emergency intervention path with the exact SSH failure reason when available.
+
+---
+
 ## 首席架構師 Review 記錄 (2026-04-05)
 
 評分：**72/100 → 修正後 88/100**