Fix Merge tool results into single message to avoid per-user rate-limit interruption

[aalinyu]

Summary

Buffer and merge per-tool-result messages into a single delivery to stay
within the WeChat API per-user rate limit, preventing session interruption.

Problem

The WeChat Bot API enforces a per-user rate limit: when the same user
receives more than 10 messages within a 60-second window, the API
rejects further sends and the session is interrupted.

This was triggered during AI agent turns that involved multiple tool
calls. Each tool result (search, browse, compute, etc.) was dispatched
as an individual sendMessage API call. A typical complex query could
produce 8–10 tool results in rapid succession, exhausting the per-minute
budget before the final AI conclusion could be delivered.

Two additional issues compounded the problem:

1. API errors were silently ignored — the sendMessage function
   previously discarded the API response body, so rate-limit rejections
   went undetected. Response validation (added in an earlier commit)
   surfaced the errors but couldn't prevent them.

2. Dispatch aborted on failure — the deliver callback re-threw any
   send error, causing the framework's reply dispatcher to stop entirely.
   The final AI conclusion (which always arrives after all tool results)
   was frequently lost.

Per-message retry and throttle were attempted but only added latency
without addressing the root cause: the per-user, per-minute message
budget was being consumed by individual tool result echoes.

Solution

Intercept and merge tool results into a single summary message,
eliminating the burst pattern that triggers rate limiting.

The deliver callback now checks deliveryCtx.kind:

• Tool result → appended to an in-memory buffer, return immediately
  (no API call).
• Non-tool message (e.g. final conclusion) → flush the buffer as one
  merged summary, then send this message.
• Dispatch end (finally block) → final flush for any remaining
  buffered results.

The merged message is formatted as a step list:

⚙ 执行步骤 (3 步)：

搜索：重庆四代宅近三年价格…
浏览：正在获取网页内容…
计算：价格分析完成…

Other changes:

• Removed retry/throttle from send.ts — with only ~2 messages per
  turn instead of ~10, the per-minute budget is no longer a concern.
• Deliver no longer re-throws — send failures send an error notice
  without aborting the dispatch, so one bad message doesn't block the
  rest.

Effect

	Before	After
Messages per turn (per user)	~10 (individual tool results + conclusion)	~2 (1 merged summary + conclusion)
Per-minute budget usage	Exhausted in seconds	Well under limit
Rate-limit risk	Guaranteed for multi-tool turns	Eliminated
Final conclusion delivery	Frequently lost (dispatch aborted)	Reliable
Latency	Added by retry + throttle	No artificial delays

[niruiyu]

I am having the similar issue: if a request triggers 2-3 tools call, the final summary message can be delivered to wechat successfully. But if a request triggers ~10 tools call, the final summary message can NOT be delivered.