feat(tg): 思考块改为可折叠引用块（与工具调用区分）

TG 不支持按消息自定义引用块颜色（竖线颜色跟随客户端主题），按用户
选择采用官方可折叠引用块：默认收起为「💭 思考过程 ▾」，点击展开，
与工具调用的普通蓝色竖线引用块在交互上明确区分。

- rich 模式：Telegram 原生 markdown 可折叠引用语法（**> 开头 + 行尾 ||）
- HTML 模式：markdownToHtml 转换为 <blockquote expandable>
- 思考文本内的 || 全角化，避免被解析成 spoiler 实体

src/bridge.ts

@@ -210,8 +210,10 @@ const REASONING_QUOTE_MAX_CHARS = 800
 
 /**
  * Assistant message → markdown for delivery. Reasoning (thinking) blocks are
- * kept, one `> ` quote line each — Telegram renders them as the accent-bar
- * quote style the phone user expects for bash/thinking-style chatter.
+ * kept as EXPANDABLE quotes (`**>`…`||` per Telegram's native markdown spec,
+ * `<blockquote expandable>` in HTML mode) so they stay collapsed on the phone
+ * and visually distinct from tool-call quotes; ordinary quotes in assistant
+ * text remain plain blockquotes.
  */
 function contentToText(content: readonly ContentBlock[]): string {
   const parts: string[] = []
@@ -223,7 +225,13 @@ function contentToText(content: readonly ContentBlock[]): string {
       if (!thinking) continue
       if (thinking.length > REASONING_QUOTE_MAX_CHARS)
         thinking = `${thinking.slice(0, REASONING_QUOTE_MAX_CHARS)}…`
-      parts.push(thinking.split('\n').map((line) => `> ${line}`).join('\n'))
+      // `||` would open/close spoiler entities in native markdown — neutralize.
+      const lines = thinking.replace(/\|\|/g, '｜｜').split('\n')
+      const last = lines.length - 1
+      parts.push([
+        '**>💭 思考过程',
+        ...lines.map((line, i) => `> ${line}${i === last ? '||' : ''}`),
+      ].join('\n'))
     }
   }
   return parts.join('\n\n')