feat: 添加MCP调试插件并优化流式超时处理

refactor(relay): 将Provider Key管理路由移至model_config模块
fix(saas): 修复demo_keys与provider_keys的匹配逻辑
perf(runtime): 将流式响应超时从60秒延长至180秒以适配思考型模型
docs: 新增模块化审计和上线前功能测试方案文档
chore: 添加tauri-plugin-mcp依赖及相关配置

crates/zclaw-runtime/src/loop_runner.rs

@@ -713,7 +713,11 @@ impl AgentLoop {
                 let mut text_delta_count: usize = 0;
                 let mut thinking_delta_count: usize = 0;
                 let mut stream_errored = false;
-                let chunk_timeout = std::time::Duration::from_secs(60);
+                // 180s per-chunk timeout — thinking models (Kimi, DeepSeek R1) can have
+                // long gaps between reasoning_content and content phases (observed: ~60s).
+                // The SaaS relay sends SSE heartbeat comments during idle periods, but these
+                // are filtered out by the OpenAI driver and don't yield StreamChunks.
+                let chunk_timeout = std::time::Duration::from_secs(180);
 
                 loop {
                     match tokio::time::timeout(chunk_timeout, stream.next()).await {