Implement a Guarded Knowledge Agent Workflow

Implement run_guarded_agent(query, tools, memory, knowledge_base, max_steps) that runs a guarded ReAct-style loop with integrated retrieval and safety checks. Parameters: - query (str): the user's question. - tools (dict): tool name to callable mapping. - memory (dict): a {"get": callable, "put": callable} interface for persistent facts. - knowledge_base (dict): a {"search": callable} interface returning list of {"content": str, "source": str} dicts. - max_steps (int): maximum loop iterations. - llm (callable): the language model to call with the scratchpad on each step. Requirements: - Check the query for prompt injection patterns before starting the loop. If detected, return an immediate safe fallback. - Retrieve relevant knowledge from knowledge_base.search(query) and inject it as context. - Retrieve relevant facts from memory.get() and merge with knowledge context. - Run a ReAct loop (up to max_steps) with Final Answer / Action parsing, tool dispatch, and observation collection. - After each tool observation, call memory.put(fact) to store any factual observation. - Support a special cite action: Action: cite(source, text) adds an inline citation. - Return {"answer": str, "citations": list[dict], "steps_used": int, "injection_blocked": bool}.