Accurate remaining useful life (RUL) prediction of lithium-ion batteries is essential for safe, reliable, and cost-effective battery management. However, many existing data-driven methods rely on continuous long-term cycling histories and may generate physically inconsistent predictions under limited or irregularly sampled degradation data. This study proposes a physics-constrained deep operator network (PC-DON) for battery RUL prediction without requiring continuous cycling histories. The proposed framework formulates RUL prediction as an operator-learning problem, in which early-cycle incremental capacity (IC) curves are used as cell-specific degradation signatures, while current-cycle electrochemical descriptors are used to characterize the instantaneous degradation state. By combining these two types of information, the model can predict RUL at different aging stages using limited early-life information and current-cycle measurements, rather than relying on complete historical degradation trajectories. To improve physical consistency, three candidate physics-based constraints are evaluated, including cycle-number monotonicity, SOH–RUL consistency, and RUL–SOH curvature. Results show that the first-order monotonicity and SOH–RUL consistency constraints provide the most effective regularization. On a dataset of 124 commercial cells, the optimized PC-DON achieves an RMSE of 104.14 cycles, MAE of 68.96 cycles, MAPE of 18.75%, and F1-score of 0.72, outperforming both the unconstrained model and the benchmark PINN. Additional analyses of physics-loss combinations, early-cycle window length, latent-feature correlation, SHAP-based feature importance, transferability, and economic implications further demonstrate the accuracy, interpretability, and practical potential of the proposed framework for battery health prognosis.