Pretrained masked language models (MLMs) remain important backbones for natural language understanding (NLU) tasks, as they often provide a more favorable balance between task performance and computational efficiency than large decoder-only models. However, existing debiasing methods often decouple bias mitigation from downstream fine-tuning, making it difficult to preserve the mitigation effects after task adaptation. To address this, we propose ARCA (Attribution-Regularized Causal Alignment), an end-to-end framework that unifies debiasing quality and task utility. ARCA uses differential gradient attribution to identify task-specific bias words and regularize the intervention for causal alignment on counterfactual samples. This single-pipeline approach enables joint optimization of task utility and debiasing. Experiments on three mainstream NLU tasks show that ARCA achieves a strong trade-off between bias mitigation and task utility compared with representative debiasing baselines.