[Re-bin]

Summary

• introduce a small shared hook interface for AgentRunner
• replace scattered runner callbacks with a single lifecycle hook contract
• migrate loop and subagent onto the new interface while preserving existing behavior

Why

AgentRunner had started to accumulate several ad-hoc callback entry points,
which made the execution surface harder to extend and harder to read.

This change keeps the runner focused on the core execution loop and moves
lifecycle customization behind one small shared abstraction.

It also lays the groundwork for team mode, where team workers will need a
cleaner way to plug per-iteration behavior into the shared runner without
pushing team-specific logic down into the runner itself.

Test plan

• pytest tests/agent/test_runner.py tests/agent/test_task_cancel.py
• pytest tests/cli/test_cli_input.py