fix(agent): retry AI Agent node on transient LLM errors (Anthropic 529)

Anthropic returns 529 "Overloaded" under load; the agent run dies ~2s in and
the user gets a hard error. n8n's engine-level retryOnFail does not apply to
chat-model sub-nodes (the sub-node retry wrapper in get-input-connection-data.ts
only covers tool invocations, see n8n#25136), and LangChain's AsyncCaller retry
does not fire on the agent's streaming path. The reliable lever is retryOnFail
on the AI Agent root node itself.

- n8n-claw-agent.json + sub-agent-runner.json: AI Agent nodes get
  retryOnFail: true, maxTries: 3, waitBetweenTries: 5000

Trade-off: a retry re-runs the whole agent turn. If the failure hits mid-loop
after a side-effecting tool call (memory save, reminder), that tool can execute
twice. Rare and mild vs. a hard user-facing failure on every 529.

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

Author: freddy-schuetz

workflows/n8n-claw-agent.json

@@ -1249,7 +1249,10 @@
       "position": [
         2832,
         -16
-      ]
+      ],
+      "retryOnFail": true,
+      "maxTries": 3,
+      "waitBetweenTries": 5000
     },
     {
       "parameters": {
@@ -2021,4 +2024,4 @@
       }
     }
   ]
-}
+}

workflows/sub-agent-runner.json

@@ -198,7 +198,10 @@
           "systemMessage": "={{ $json.systemPrompt }}",
           "maxIterations": 25
         }
-      }
+      },
+      "retryOnFail": true,
+      "maxTries": 3,
+      "waitBetweenTries": 5000
     },
     {
       "id": "sub-claude",