Implement an End-to-End Agent Runtime

Implement run_agent_runtime(goal, config) that runs a complete agent lifecycle. config is a dict with: - planner (callable): generates a plan (planner(goal) returns list of step dicts with {"id": str, "tool": str, "params": any, "depends_on": list[str]}). - tools (dict): tool name to callable mapping. - max_steps (int): maximum execution steps. - memory_limit (int): max entries in working memory. - guardrails (list[callable]): each guardrail(step, context) returns True (allow) or False (block). - approval_rules (dict): {tool_name: bool} where True means the step needs human approval. Requirements: - Start by calling the planner to generate a step plan. - Execute plan steps considering dependencies (like the supervisor pattern). - Before each step, check guardrails -- if any guardrail returns False, mark the step as blocked and continue. - If a step's tool is in approval_rules and the rule is True, mark it as pending approval rather than executing immediately; simulate approval by setting a flag. - Maintain a working memory dict (bounded by memory_limit) that stores results keyed by step ID. - Record a trace: [{"step": str, "tool": str, "status": str, "result": any, "duration_ms": int}]. - If memory exceeds memory_limit, evict oldest entries (LRU-style). - Return {"answer": str, "trace": list, "steps_completed": int, "memory_used": int, "blocked_steps": list[str]}.