Remove finalize_workflow tool and run coverage checks after agent with iterative reruns

[zhongkaifu]

Motivation

• Move coverage and dependency checks out of the agent-invokable tool so the system can run checks after each agent run and drive further planning externally.
• Allow automatic iterative re-runs of the planner when coverage or dependency issues are detected, driven by structured feedback.
• Simplify agent toolset by removing the finalize_workflow function and avoid relying on the agent to call a finalize tool.

Description

• Removed the finalize_workflow function_tool and removed it from the agent tools list in velvetflow/planner/structure.py.
• Added post-run logic that calls _run_coverage_check after each agent execution and, when issues exist, constructs feedback and iteratively reruns the agent using a helper _run_agent until coverage/dependency limits (max_coverage_refine_rounds / max_dependency_refine_rounds) are reached.
• Updated prompt and feedback text by changing _build_coverage_feedback_message and _build_dependency_feedback_message wording and adjusting the system prompt guidance to instruct the agent to stop calling tools when structure is done.
• Removed finalized_state usage, introduced coverage_refines/dependency_refines counters, and fallback to _prepare_skeleton_for_coverage when no latest skeleton is available.

Testing

• No automated tests were run as part of this change.
• The modified file is velvetflow/planner/structure.py and changes were validated by static inspection only.

---

Codex Task