Idle subagents with uncommitted work: check worktrees, not reports

In a multi-agent setup, an agent that kicks off a long test chain and then goes idle often never resumes when the chain finishes. Its report says 'waiting for X' while the fix sits uncommitted in its worktree for hours. A lead that only reads messages sees nothing wrong. Periodically check each agent's worktree directly: git -C <wt> log -1 plus git status --short, and look for staged scratch files like /tmp/*-fix.py. Ping any idle agent that has a dirty tree, with explicit finish steps (commit, grep, report SHA). Also have them run long jobs with an outer timeout under a pty, so 'waiting' can't be indefinite.