fix(probe): end an Agent session that has gone silent and redial [CI]

backend/probe/src/MoShiro.Probe/Link/AgentConnection.cs

@@ -23,6 +23,9 @@ public sealed class AgentConnection : IDisposable
 {
     private static readonly TimeSpan s_flushInterval = TimeSpan.FromSeconds(ProtocolLimits.ChunkSeconds);
     private static readonly TimeSpan s_clockInterval = TimeSpan.FromSeconds(30);
+
+    /// <summary>Agent 的任何帧（时钟回复、任务集）都算存活；比时钟同步间隔长得多，容得下差线路上的黑洞，但不会让一个 Agent 已经丢掉的会话永远挂着。</summary>
+    private static readonly TimeSpan s_idleLimit = TimeSpan.FromSeconds(120);
     private static readonly RpcRouter<ConnectionSession> s_router = BuildRouter();
 
     /// <summary>对等面入站方法的路由器（只读，测试用）。</summary>
@@ -376,6 +379,13 @@ public sealed class AgentConnection : IDisposable
         long lastClockUs = _env.Clock.NowUs();
         while (await flush.WaitForNextTickAsync(ct))
         {
+            if (rpc.IdleFor > s_idleLimit)
+            {
+                // Agent 已经不认这个会话（它重启、重连、或者空闲超时关了会话，而 FIN 没有到达）：Probe 这边结束它，让外层按退避重拨
+                _env.Log.Warn("agent session is silent, reconnecting", ("agent", Endpoint.AgentId), ("idle_s", (long)rpc.IdleFor.TotalSeconds));
+                throw new IOException("the Agent stopped answering on this session");
+            }
+
             session.FlushLineBlocks();
             if (_env.Clock.NowUs() - lastClockUs >= s_clockInterval.TotalMicroseconds)
             {