Fine-Tuned LLMs Know They Don't Know: A Parameter-Efficient Approach to Recovering Honesty