Merging Triggers, Breaking Backdoors: Defensive Poisoning for Instruction-Tuned Language Models