fix: Phase 6联调 — 插件管理器端口修正 + 多模型配置系统整合 + 历史消息刷新修复

## 调试日志

### 1. 插件管理器启动失败
- **症状**: DevTools 显示插件管理器一直"已停止"，手动启动正常
- **排查**: 对比 process-manager.js 传入的环境变量 vs plugin-manager config.go 读取的变量
- **根因**: config.js 传入 PLUGIN_MANAGER_PORT=8094，但 config.go 读取 os.Getenv("PORT")，env 名不匹配。且 process.env 中 PORT 泄露时被误读为 9090，与 DevTools 端口冲突
- **修复**: config.js 将 PLUGIN_MANAGER_PORT → PORT，使 env 名与代码一致 (c3055f4)

### 2. 历史消息刷新后消失
- **症状**: 浏览器刷新后聊天历史清空
- **排查**: WebSocket history_response handler 中 if (msg.messages) 对空数组 [] 为 truthy
- **根因**: 后端返回空的 history_response (缓存为空) 时，空数组覆盖了 HTTP 已加载的消息
- **修复**: useWebSocket.ts 改为 if (msg.messages && msg.messages.length > 0)，空数组走 else-if 分支仅打日志，不覆盖已有消息

### 3. Phase 6 多模型配置系统
- Gateway: ModelsConfigStore (JSON文件持久化) + Admin CRUD API (providers/models/routing)
- ai-core: ModelSelector 支持按 purpose 选择 + fallback_chain，无配置时回退 .env
- DevTools: 模型配置管理面板 (Providers/Models/Routing 三Tab)、在线模型查询代理、路由表单 checkbox 多选、关键词搜索过滤
- .gitignore: models.json + platform_configs.json

### 4. 多端客户端追踪
- Hub 新增 knownClients 映射 (clientID → KnownClient)，在线/离线状态追踪
- 客户端备注持久化到 PostgreSQL
- DevTools 客户端管理面板

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

backend/ai-core/internal/llm/selector.go

@@ -0,0 +1,131 @@
+package llm
+
+import (
+	"context"
+	"fmt"
+	"sync"
+	"time"
+
+	"github.com/yourname/cyrene-ai/ai-core/internal/config"
+)
+
+// ModelPurpose identifies the kind of LLM task.
+type ModelPurpose string
+
+const (
+	PurposeChat             ModelPurpose = "chat"
+	PurposeDeepThinking     ModelPurpose = "deep_thinking"
+	PurposeIntentAnalysis   ModelPurpose = "intent_analysis"
+	PurposeToolCalling      ModelPurpose = "tool_calling"
+	PurposeMemoryExtraction ModelPurpose = "memory_extraction"
+)
+
+// ErrModelNotRequired is returned when an optional model is unavailable.
+var ErrModelNotRequired = fmt.Errorf("model not required, caller should degrade gracefully")
+
+// ModelSelector routes requests to the best available LLMProvider based on purpose.
+type ModelSelector struct {
+	loader *config.Loader
+	envCfg OpenAIConfig
+	mu     sync.RWMutex
+	cache  map[string]LLMProvider
+}
+
+// NewModelSelector creates a ModelSelector. If loader is nil or has no config,
+// all calls fall back to envCfg.
+func NewModelSelector(loader *config.Loader, envFallback OpenAIConfig) *ModelSelector {
+	return &ModelSelector{
+		loader: loader,
+		envCfg: envFallback,
+		cache:  make(map[string]LLMProvider),
+	}
+}
+
+// Select returns an LLMProvider for the given purpose. Falls back through the
+// routing fallback chain; returns the env provider if nothing matches.
+func (s *ModelSelector) Select(ctx context.Context, purpose ModelPurpose) (LLMProvider, error) {
+	if s.loader == nil || !s.loader.HasConfig() {
+		return s.envProvider(), nil
+	}
+
+	cfg := s.loader.GetConfig()
+	if cfg == nil {
+		return s.envProvider(), nil
+	}
+
+	route, ok := cfg.Routing[string(purpose)]
+	if !ok || len(route.FallbackChain) == 0 {
+		return s.envProvider(), nil
+	}
+
+	for _, modelID := range route.FallbackChain {
+		provider, err := s.getOrCreateProvider(modelID, cfg)
+		if err != nil {
+			continue
+		}
+		return provider, nil
+	}
+
+	if route.Required {
+		return nil, fmt.Errorf("all models unavailable for purpose %s", purpose)
+	}
+	return s.envProvider(), nil
+}
+
+// DefaultAdapter returns an *Adapter backed by the chat-purpose provider.
+// This is the backward-compatible entry point: all existing consumers
+// (Orchestrator, Synthesizer, BackgroundThinker, etc.) use this.
+func (s *ModelSelector) DefaultAdapter() *Adapter {
+	provider, _ := s.Select(context.Background(), PurposeChat)
+	return NewAdapter(provider)
+}
+
+func (s *ModelSelector) envProvider() LLMProvider {
+	return NewOpenAIProvider(s.envCfg)
+}
+
+func (s *ModelSelector) getOrCreateProvider(modelID string, cfg *config.ModelsConfigData) (LLMProvider, error) {
+	s.mu.RLock()
+	if p, ok := s.cache[modelID]; ok {
+		s.mu.RUnlock()
+		return p, nil
+	}
+	s.mu.RUnlock()
+
+	modelCfg, ok := cfg.Models[modelID]
+	if !ok {
+		return nil, fmt.Errorf("model %s not found", modelID)
+	}
+	if !modelCfg.Enabled {
+		return nil, fmt.Errorf("model %s is disabled", modelID)
+	}
+
+	provCfg, ok := cfg.Providers[modelCfg.Provider]
+	if !ok {
+		return nil, fmt.Errorf("provider %s not found for model %s", modelCfg.Provider, modelID)
+	}
+
+	timeout := time.Duration(provCfg.TimeoutSec) * time.Second
+	if timeout <= 0 {
+		timeout = 120 * time.Second
+	}
+	maxRetries := provCfg.MaxRetries
+	if maxRetries <= 0 {
+		maxRetries = 3
+	}
+
+	provider := NewOpenAIProvider(OpenAIConfig{
+		BaseURL:       provCfg.BaseURL,
+		APIKey:        provCfg.APIKey,
+		Model:         modelCfg.Name,
+		FallbackModel: modelCfg.Name,
+		MaxRetries:    maxRetries,
+		Timeout:       timeout,
+	})
+
+	s.mu.Lock()
+	s.cache[modelID] = provider
+	s.mu.Unlock()
+
+	return provider, nil
+}