[SamZhu19921116]

1. 当 LLM 服务出现临时性错误（如网络波动、超时、429限流等）时， base.py 中的 _run_with_retry 方法会启动重试机制。
2. 在重试等待期间， _sleep_with_heartbeat 方法会周期性调用 on_retry_wait 回调函数，发送类似 'Model request failed, retry in 1s (attempt 1)' 的心跳消息。
3. 之前 on_retry_wait 参数被错误地绑定到 _bus_progress ，导致这些内部诊断消息被当作普通进度消息发送到飞书客户端。
4. manager.py 的消息分发器没有过滤这类重试心跳消息。

修复方案：

1. loop.py - 新增重试等待回调- 新增独立的 _on_retry_wait 回调函数，为重试消息添加 _retry_wait: True 元数据标识- 在 AgentRunSpec 中传入 retry_wait_callback 参数。
2. runner.py - 支持重试回调参数- 在 AgentRunSpec 数据类中新增 retry_wait_callback 字段- 在 _build_request_kwargs 中将 on_retry_wait 参数从 progress_callback 改为 retry_wait_callback。
3. manager.py - 过滤重试心跳消息- 在 _dispatch_outbound 方法中新增过滤逻辑，丢弃所有带 _retry_wait 标识的消息，确保重试心跳不会发送到任何客户端。

[Re-bin]

This is the right kind of fix.

The regression comes from fbedf7ad ("feat: harden agent runtime for long-running tasks"), where AgentRunSpec.progress_callback ended up wired into the provider's on_retry_wait parameter. That meant every transient-retry heartbeat ("Model request failed, retry in 1s (attempt 1).") got published as a normal progress message and delivered to end-user channels — exactly the kind of internal diagnostic a user should never see.

The fix splits the two signals cleanly:

• AgentRunSpec grows a dedicated retry_wait_callback, and _build_request_kwargs forwards on_retry_wait to it instead of progress_callback.
• AgentLoop._process_message builds a distinct _on_retry_wait closure that tags outbound messages with _retry_wait: True.
• ChannelManager._dispatch_outbound drops any message carrying that tag before it reaches a channel.

I like that retry-wait events still enter the bus instead of being silently dropped at source — future diagnostics/metrics subscribers can observe them without any further plumbing, and only the end-user delivery path is gated.

What the PR was missing was a regression test; given this is exactly the "silently re-bindable callback" class of bug, I pushed two focused ones on top of the existing changes:

• tests/agent/test_runner.py::test_runner_binds_on_retry_wait_to_retry_callback_not_progress
  pins _build_request_kwargs to forward retry_wait_callback (not progress_callback) as on_retry_wait.
• tests/channels/test_channel_manager_delta_coalescing.py::TestRetryWaitFiltering::test_retry_wait_message_dropped
  exercises _dispatch_outbound end-to-end and asserts _retry_wait: True messages never reach channel.send.

Both tests fail on origin/main and pass with this PR's fix. Full suite is green locally (2032 passed) and CI is already green across 3.11/3.12/3.13.

From my side this is clean, narrow, and ready to merge.