fix(nodes): locate agents behind a proxy and retry missing countries [CI]

backend/src/MoShiro.Backend.Application/Agents/MoShiroFleetHooks.cs

@@ -115,10 +115,15 @@ public sealed class MoShiroNodeEventSink : INodeEventSink
                 await _events.AddAsync(id, "agent_offline", nodeEvent.AtMs, null, ct);
                 break;
             case NodeEventKind.Hello when nodeEvent.Hello is { } hello:
-                await ApplyHelloAsync(id, hello, ct);
+                await ApplyHelloAsync(id, hello, nodeEvent.RemoteIp, ct);
                 break;
             case NodeEventKind.Heartbeat when nodeEvent.Extra is { } extra:
                 await _nodes.SetVersionAsync(id.Value, NodeKinds.Agent, nodeEvent.Node.NodeVersion, ct);
+                if (_nodes.ShouldRetryGeo(id.Value) && await _agents.GetAsync(id, ct) is { } unlocated)
+                {
+                    _nodes.ObserveAgentInBackground(id.Value, GeoHostOf(unlocated, nodeEvent.RemoteIp), null);
+                }
+
                 try
                 {
                     AgentRuntimeState state = _runtime.Get(id);
@@ -145,6 +150,12 @@ public sealed class MoShiroNodeEventSink : INodeEventSink
         }
     }
 
+    /// <summary>查国家和运营商用哪个地址：对外地址优先，其次 IPv6 地址，都没有（NAT 后面、或者没有上报公网地址）才用这条连接的来源地址。</summary>
+    private static string? GeoHostOf(AgentRecord? agent, string? remoteIp) =>
+        !string.IsNullOrEmpty(agent?.PublicHost) ? agent.PublicHost
+        : !string.IsNullOrEmpty(agent?.Ipv6Host) ? agent.Ipv6Host
+        : remoteIp;
+
     /// <summary>Probe 自己连上来了：hello 里带着它的系统、架构和构建信息；上线下线只推给 SSE，不记事件。</summary>
     private async ValueTask OnProbeEventAsync(NodeId id, NodeEvent nodeEvent, CancellationToken ct)
     {
@@ -213,7 +224,7 @@ public sealed class MoShiroNodeEventSink : INodeEventSink
         }
     }
 
-    private async ValueTask ApplyHelloAsync(NodeId id, HelloParams hello, CancellationToken ct)
+    private async ValueTask ApplyHelloAsync(NodeId id, HelloParams hello, string? remoteIp, CancellationToken ct)
     {
         var kinds = new List<string>();
         foreach (string capability in hello.Capabilities)
@@ -278,7 +289,7 @@ public sealed class MoShiroNodeEventSink : INodeEventSink
 
         // 端口、证书指纹或对外地址变了，Probe 需要新的列表
         AgentRecord? after = await _agents.GetAsync(id, ct);
-        _nodes.ObserveAgentInBackground(id.Value, after?.PublicHost, meta.Platform);
+        _nodes.ObserveAgentInBackground(id.Value, GeoHostOf(after, remoteIp), meta.Platform);
         await _releases.OfferToNodeAsync(id, ct);
         await _schedule.SendToNodeAsync(id, ct);
         if (before is not null && after is not null

backend/src/MoShiro.Backend.Application/Nodes/NodeInfoService.cs

@@ -26,6 +26,7 @@ public sealed class NodeInfoService
     private readonly TimeProvider _time;
     private readonly ConcurrentDictionary<string, NodeMeta> _cache = new();
     private readonly ConcurrentDictionary<string, byte> _inflight = new();
+    private readonly ConcurrentDictionary<string, long> _lastTry = new();
     private Task? _load;
 
     public NodeInfoService(INodeMetaStore store, IGeoLookup geo, TimeProvider time)
@@ -119,6 +120,7 @@ public sealed class NodeInfoService
     public async ValueTask ObserveAddressAsync(string nodeId, string kind, string? host, CancellationToken ct)
     {
         host = host?.Trim();
+        _lastTry[nodeId] = Now();
         if (string.IsNullOrEmpty(host) || !IsPublic(host) || !_inflight.TryAdd(nodeId, 0))
         {
             return;
@@ -162,6 +164,17 @@ public sealed class NodeInfoService
         }
     }
 
+    /// <summary>国家还没查到、并且离上一次尝试已经过了重试间隔：调用方可以再查一次（心跳里调，不会每次都真去查）。</summary>
+    public bool ShouldRetryGeo(string nodeId)
+    {
+        if (_cache.TryGetValue(nodeId, out NodeMeta? meta) && (meta.Info.Country is not null || meta.Info.IspManual))
+        {
+            return false;
+        }
+
+        return !_lastTry.TryGetValue(nodeId, out long at) || Now() - at >= s_retryAfter.TotalMilliseconds;
+    }
+
     /// <summary>Agent 报告某个 Probe 连着它：在后台用 Probe 的来源地址查国家和运营商（系统信息是 Probe 自己连上面板时报的）。出错只吞掉。</summary>
     public void ObserveProbeInBackground(string probeId, string? remote) =>
         _ = Task.Run(async () =>