fix(chat): 会话已选 LLM 时不再被「全局未激活」误拦，提示按状态细分

发送消息的前置检查只看全局激活配置(chatAgentRef),而真正执行的
runConversationSync 会优先用会话所选/Agent 绑定的配置——两处解析分叉,导致
「会话已选一个有效配置、但全局没有激活配置」时,消息在预检查阶段就被 503
「LLM 未配置」拦下,根本走不到能用所选配置的执行阶段。

- 抽出 resolveChatAgent(conversation) 作为唯一解析入口(会话/Agent 绑定优先、
  全局兜底),send 预检查与后台执行都用它,消除分叉。
- 提示按实际状态细分(chatUnavailableReason):无任何配置→提示添加;有配置但
  未激活→提示激活或为本会话指定一个;而非一律「未配置」。

相邻的 StartTriggeredRun 仍用全局 gate(触发器场景,本次不动)。

验证: 新增两个测试——三态提示区分、会话所选配置在全局无激活时仍能解析出
ChatAgent(回归防护);连本地 PG 实跑通过,server 全量无回归。

server/server.go

@@ -492,6 +492,37 @@ func (s *Server) effectiveProfileForAgent(agentKey string, pinID *int64) *int64
 	return pinID
 }
 
+// resolveChatAgent picks the ChatAgent for one conversation: the agent's own
+// binding or this conversation's chosen profile first, the global active config
+// only as a fallback. Both the send precheck and the background runner MUST use
+// this — resolving differently in the two paths is how a conversation that had
+// picked a valid profile still got rejected with "LLM 未配置" when no global
+// config was active.
+func (s *Server) resolveChatAgent(c *db.Conversation) *agent.ChatAgent {
+	ca := s.chatAgentRef()
+	if eff := s.effectiveProfileForAgent(c.AgentKey, c.LLMProfileID); eff != nil {
+		if pa := s.chatAgentForProfile(*eff); pa != nil {
+			ca = pa
+		}
+	}
+	return ca
+}
+
+// chatUnavailableReason explains why no ChatAgent could be resolved, so the user
+// knows whether to add a config, activate one, or pick one for this conversation
+// — rather than a flat "not configured" that hides which of those it is.
+func (s *Server) chatUnavailableReason() string {
+	if s.m.pg != nil {
+		if profiles, err := s.m.pg.ListProfiles(); err == nil && len(profiles) == 0 {
+			return "尚未配置 LLM：请到 系统 → LLM 配置 添加一个配置"
+		}
+		if active, err := s.m.pg.ActiveProfile(); err == nil && active == nil {
+			return "没有已激活的 LLM 配置：请到 系统 → LLM 配置 激活一个，或在本对话为该会话指定一个配置"
+		}
+	}
+	return "LLM 未就绪，无法对话：请检查 系统 → LLM 配置是否有可用且已激活的配置"
+}
+
 // providerForProfile returns a cached provider+cfg for a profile id, so every agent
 // bound/pinned to the same profile shares one provider instance (one rate limiter).
 // ok=false when the profile is missing/invalid → caller falls back to the global pair.