Collapse of Irrelevant Representations (CIR) Ensures Robust and Non-Disruptive LLM Unlearning