Add intelligent PR triage to replace rigid label-based routing

[jalemieux]

Problem

The current route() function uses a rigid checklist of label + state queries to find actionable work. When a PR doesn't exactly match one of the expected states, the script silently reports "No actionable work found" and sleeps — even when the fix is trivial (e.g., a closed PR that should be open, a missing comment, a stale label).

Example: PR #17 on conversation_with_ai

A bot-authored draft PR had bot:plan-proposed label and a complete implementation plan in the body, but the script couldn't find it because:

1. The PR had been closed (so gh pr list skipped it)
2. Even after reopening, it had 0 comments — and check_plan_feedback requires comment_count > 0

This required manual debugging to figure out. The script should have been able to diagnose and resolve this itself.

Other failure modes this would catch

• PR has the wrong label for its actual state
• PR is closed but shouldn't be
• Issue is assigned but the assignee's PR was abandoned
• bot:in-progress label is stale (process crashed without cleanup)
• PR has plan but no bot:plan-proposed label

Proposed Solution

Add a fallback step in the routing loop: when the rigid checks find nothing, pass all open bot PRs + unassigned issues to Claude Code for triage. Claude can inspect the actual state of each item and decide what action to take (reopen a PR, add a comment, fix a label, skip, etc.) rather than silently giving up.

This brings back Claude Code's judgment into the routing decision, so the system self-heals instead of getting stuck on small state mismatches.

Success Criteria

• Script never silently skips work that a human would recognize as actionable
• Triage step only fires when rigid routing finds nothing (no added cost in the happy path)
• Actions taken by triage are logged clearly

[jalemieux]

lgtm