Discovering Transformer Circuits via a Hybrid Attribution and Pruning Framework