feat(FR-3230): add agent watcher start/stop/restart actions to AgentDetailDrawer (#8087)

Resolves #8085 (FR-3230)

## Summary

Migrates the agent node lifecycle controls (start / stop / restart) from the Control Panel into the WebUI. Superadmins can now start, stop, and restart an agent node directly from the **Agent detail drawer**, next to the existing Settings action.

## Test server

A superadmin login is required for both.

- **Normal flow** (start / stop / restart on a healthy watcher): use the `10.82.0.130` test server (manager API on `:8090`).
- **Error handling** (watcher down, agent not loaded with systemctl, etc.): use the `10.82.0.151` test server (manager API on `:8090`).

## Changes

- **`AgentLifeCycleControlModal.tsx`** (new): a single modal that owns the start / stop / restart confirmation and request. It takes `open` + a `lifeCycleType` (`'start' | 'stop' | 'restart'`) and branches the title / success message internally; the body shows a shared warning plus the target agent's name in a surface box (borrowed from `BAIDeleteConfirmModal`). The three watcher calls run through `useTanMutation`.
- **`AgentActionButtons.tsx`**: renders Settings + Restart / Start / Stop trigger buttons in the agent detail drawer. Each watcher button sets a single `lifeCycleType` state that drives the modal (`open={!!lifeCycleType}`). Status-aware disabling: Start disabled when the agent is `ALIVE`, Stop disabled otherwise; Restart always available. Icons/colors follow `SessionActionButtons` — Stop uses `BAITerminateIcon` + `colorError`, Restart uses the `RefreshCw` (lucide) icon.
- **`backend.ai-client` (`client.ts`)**: added `start_watcher_agent` / `stop_watcher_agent` / `restart_watcher_agent` SDK methods (with matching `BackendAIClient` type signatures) so the watcher REST calls are encapsulated and reusable.
- **`AgentResources.tsx`** (drawer crash guard): a stopped agent's `live_stat` no longer contains `node.mem`, and three call sites accessed `parsedLiveStat?.node?.mem.capacity` without optional chaining after `mem`, throwing a `TypeError` that crashed the whole drawer. Added the missing `?.` at all three sites, and the **Utilization** section now renders `-` when `live_stat.node` is empty instead of misleading 0% / NaN values.
- **`AgentDetailDrawerContent.tsx`**: wrapped `<AgentResources>` with `ErrorBoundaryWithNullFallback` so a stat-parsing failure can never take down the entire drawer again.
- **i18n**: added `agent.Watcher*` keys (button labels, success toasts, a single action warning) across all 21 locales.

## API

Uses the manager REST watcher endpoints via the new client SDK methods. These were chosen over the legacy `/api/agents/watcher_agent_*` paths (removed on current backend.ai `main`):

- `POST /resource/watcher/agent/start`
- `POST /resource/watcher/agent/stop`
- `POST /resource/watcher/agent/restart`

Body: `{ agent_id }` (the agent's `row_id`).

## Error handling

The manager proxies these to the agent's watcher daemon, so several non-trivial outcomes are handled explicitly in the modal:

- **200 + `{ result: "ok" }`** → success toast.
- **200 + a body message** (e.g. `"Agent is not loaded with systemctl."`) → surfaced as a warning. HTTP 200 does **not** guarantee success here. The client also hands back a plain string for `text/*` responses, which is handled too.
- **403 (non-superadmin)** → the component is only rendered on superadmin-only pages and the endpoints are superadmin-only on the manager.
- **400 (invalid params) / 403 (watcher token mismatch) / 500 (agent not found, watcher unreachable, timeout, etc.)** → error toast resolved through `getErrorMessage`, logged via `useBAILogger`. A missing `row_id` short-circuits before the request.

## Post-action data refresh

- `AgentLifeCycleControlModalFragment` is now `@refetchable` (`id`, `status`, `status_changed`). After a successful watcher request the modal refetches it with `network-only`, so every component reading the same `AgentNode` record (table row, detail drawer) picks up the new status automatically via the normalized store.
- The refetch runs inside `startTransition` — without it the suspend during refetch collapsed the drawer's Suspense boundary to its skeleton fallback.
- `AgentList`'s manual reload button is replaced with `BAIFetchKeyButton` (`autoUpdateDelay` 15s, matching `StorageProxyList`). Node-level refetch cannot change connection membership/count (e.g. a stopped agent leaving the ALIVE tab), and the watcher actions are async on the backend, so the periodic list refresh is what makes the list itself converge.

## Verification

`bash scripts/verify.sh` → **Relay / Lint / Format / TypeScript all PASS**. (The single terminology warning is a pre-existing key, `error.UserHasNoGroup`, unrelated to this change.)

## Screenshots

_TODO: capture from the `10.82.0.130` test server with a superadmin login._

Author: ironAiken2

packages/backend.ai-client/src/client.ts

@@ -1281,6 +1281,51 @@ export class Client {
     return this._wrapWithPromise(rqst);
   }
 
+  /**
+   * Start the agent's watcher daemon. (superadmin only)
+   *
+   * @param {string} agentId - the agent id (row id) to control
+   */
+  async start_watcher_agent(agentId: string): Promise<any> {
+    const rqst = this.newSignedRequest(
+      'POST',
+      '/resource/watcher/agent/start',
+      { agent_id: agentId },
+      null,
+    );
+    return this._wrapWithPromise(rqst);
+  }
+
+  /**
+   * Stop the agent's watcher daemon. (superadmin only)
+   *
+   * @param {string} agentId - the agent id (row id) to control
+   */
+  async stop_watcher_agent(agentId: string): Promise<any> {
+    const rqst = this.newSignedRequest(
+      'POST',
+      '/resource/watcher/agent/stop',
+      { agent_id: agentId },
+      null,
+    );
+    return this._wrapWithPromise(rqst);
+  }
+
+  /**
+   * Restart the agent's watcher daemon. (superadmin only)
+   *
+   * @param {string} agentId - the agent id (row id) to control
+   */
+  async restart_watcher_agent(agentId: string): Promise<any> {
+    const rqst = this.newSignedRequest(
+      'POST',
+      '/resource/watcher/agent/restart',
+      { agent_id: agentId },
+      null,
+    );
+    return this._wrapWithPromise(rqst);
+  }
+
   /**
    * Create a `compute session` if the session for the given sessionId does not exist.
    * It returns the information for the existing session otherwise, without error.

react/src/__generated__/AgentActionButtonsFragment.graphql.ts

@@ -8,6 +8,7 @@ import AgentActionButtons from './AgentNodeItems/AgentActionButtons';
 import AgentComputePlugins from './AgentNodeItems/AgentComputePlugins';
 import AgentResources from './AgentNodeItems/AgentResources';
 import AgentStatusTag from './AgentNodeItems/AgentStatusTag';
+import BAIErrorBoundary from './BAIErrorBoundary';
 import { CheckOutlined, CloseOutlined } from '@ant-design/icons';
 import { Descriptions, Grid, Tabs, theme, Typography } from 'antd';
 import {
@@ -150,7 +151,11 @@ const AgentDetailDrawerContent: React.FC<AgentDetailDrawerContentProps> = ({
           {
             key: 'resources',
             label: t('agent.Resources'),
-            children: <AgentResources agentNodeFrgmt={agent} />,
+            children: (
+              <BAIErrorBoundary>
+                <AgentResources agentNodeFrgmt={agent} />
+              </BAIErrorBoundary>
+            ),
           },
         ]}
       />