slack: drop the working pill when a codex turn ends (task_complete)

[rogeriochaves]

Summary

Codex writes a task_complete event_msg to its rollout the moment a turn finishes; the assistant's final text is the agent_message immediately before it. The bridge only watched for the out-of-credits flavour of task_complete (where last_agent_message == null), so any normal completed turn left the working pill lit on the assistant's last post forever — the refresh loop kept re-setting it every minute and the channel showed "is working…" on a turn that ended hours ago.

Two paths cover the in-batch and cross-batch cases:

1. format.ts: when a non-out-of-credits task_complete lands in the same poll batch as the final agent_message, mark that text post terminal. The bridge's existing terminal-text path skips the new pill and the previous anchor's pill is cleared in the normal flow. This is the common case — agent_message and task_complete are written back-to-back in the rollout, so a single poll tick sees both.
2. bridge.ts: cross-batch safety net. If a poll tick sees a task_complete but emits no terminal text (because the agent_message was already posted in a prior tick), clear the active pill directly.

Test plan

• All 242 existing CLI tests pass.
• Deploy to the dev box, restart agents-slack-bridge.service.
• Trigger a normal pr-reviewer turn end to end via Slack. The "is working…" pill clears within pollMs of the agent's last text post.
• No clear pill on codex task_complete errors in journalctl -u agents-slack-bridge.service.