feat(monitor): auto-complete monitor tasks from the user's own messages

[seasidespace]

Summary

Auto-completes a monitor-created task when the user later does the thing and says so in the same conversation, by watching their own outgoing Telegram messages.

• When a monitor task exists for a chat, outgoing self-messages in that chat are judged for completion intent (e.g. 拿了 / 搞定了 / done); a confident match flips the matching task to completed with completed_via:"self_report".
• 省token: an LLM call only fires when the conversation actually has an open monitor task — a cheap task-store gate runs first and returns with zero tokens otherwise.
• 不影响性能: the normal incoming-message triage path is untouched. Outgoing self-messages are emitted on a dedicated message_self kind and short-circuited in the router before triage, preserving the #569 "self-messages never trigger triage" guarantee.
• The completion judge uses the monitor's configured model (routed via the workflow runner's new complete_text), not a hardcoded model.
• Completion phrases = built-in multilingual defaults + per-monitor configurable, passed to the model as hints (never a hard gate). Multiple open tasks in one chat → the model disambiguates which task.
• New RPCs: monitor_completion_phrase_add / monitor_completion_phrase_delete.

Test plan

• cargo check -p puffer-subscriptions -p puffer-cli
• cargo test -p puffer-subscriptions self_report (4 passing: self-message detection by kind/payload, remote handler)
• cargo test -p puffer-cli monitor_self_report (11 passing: zero-token gating, single/multi-task judging, phrase-hint merging, completed/ignored tasks excluded)
• outgoing message_self kind mapping test
• Manual e2e: create a monitor task from an incoming message, send a self-message "搞定了" in the same chat → task shows completed (completed_via:"self_report"); unrelated self-message does not complete it; self-message in a chat with no open task makes zero LLM calls.

🤖 Generated with Claude Code

[vercel]

@seasidespace is attempting to deploy a commit to the fuzzland Team on Vercel.

A member of the Team first needs to authorize it.