arXiv:2608.06660v2 Announce Type: replace Abstract: Physics-Informed Kolmogorov-Arnold Networks (PIKANs) offer effective neural solvers for partial differential equations (PDEs), but their conventional formulations rely on Cartesian domains and scattered collocation points, often resulting in high computational costs, approximate boundary enforcement, and no rigorous error guarantees. We introduce CT-PIKAN, which maps curvilinear domains to the unit square and computes all geometric quantities, including those required by the Laplace-Beltrami operator, through automatic differentiation. Building on this transformation, Curvilinear Separable PIKAN (CS-PIKAN) represents solutions as rank-R sums of products of one-dimensional KANs with topology-adapted bases (Fourier for periodic and Chebyshev for bounded coordinates and time). A separable envelope and transfinite lift enforce boundary and initial conditions exactly, while a single sqrt{g}-weighted residual is minimized on a stratified tensor grid, reducing network cost from the product to the sum of grid sizes. We derive a metric-explicit Poincar'e constant, establish computable a-posteriori error bounds for elliptic, parabolic, and hyperbolic problems, and prove that the loss provides an unbiased estimator of the squared residual norm underlying these bounds. Nine benchmarks cover advection, Poisson, and heat conduction on Cartesian, polar, wavy, annular, star-shaped, and sampled-boundary domains. Across three seeds, CS-PIKAN reduces the relative L2 error of CT-PIKAN by factors of 7-126 at equal or lower training cost (up to 10 times faster), with all certified bounds satisfied. Ablations highlight the importance of exact periodicity and boundary constraints; KAN factors outperform MLPs for transport-dominated problems but not smooth elliptic ones. Nevertheless, classical body-fitted solvers remain substantially cheaper at comparable accuracy for these two-dimensional linear problems.