feat: LLM 调用日志 + ModelSelector 优化 + devtools.bat 编码修复

- 新增 call_log.go: 全局环形缓冲区记录每次 LLM 调用(模型/Token/耗时/错误)
- OpenAIProvider.doChat/ChatStreamWithTools 自动记录调用日志
- ai-core 暴露 GET /api/v1/llm-calls 端点, DevTools 代理 + UI 面板
- ModelSelector.envProvider 改为单例缓存, 避免重复创建 HTTP Client
- 新增 PurposeToolCalling 适配器, 后台思考工具调用走专用路由
- envFallback 超时 120s→180s, 显式设置 MaxRetries
- devtools.bat 全英文, 解决 Windows CMD GBK 编码乱码问题

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

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

@@ -25,10 +25,11 @@ var ErrModelNotRequired = fmt.Errorf("model not required, caller should degrade
 
 // 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
+	loader    *config.Loader
+	envCfg    OpenAIConfig
+	mu        sync.RWMutex
+	cache     map[string]LLMProvider
+	cachedEnv LLMProvider // cached env fallback, created once
 }
 
 // NewModelSelector creates a ModelSelector. If loader is nil or has no config,
@@ -81,7 +82,12 @@ func (s *ModelSelector) DefaultAdapter() *Adapter {
 }
 
 func (s *ModelSelector) envProvider() LLMProvider {
-	return NewOpenAIProvider(s.envCfg)
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	if s.cachedEnv == nil {
+		s.cachedEnv = NewOpenAIProvider(s.envCfg)
+	}
+	return s.cachedEnv
 }
 
 func (s *ModelSelector) getOrCreateProvider(modelID string, cfg *config.ModelsConfigData) (LLMProvider, error) {