Code and run records for "When do physics-informed neural networks fail for Darcy flow in random heterogeneous media? Coefficient representation, correlation length and interface sharpness"

Code, run configurations, and complete run records supporting the article "When do physics-informed neural networks fail for Darcy flow in random heterogeneous media? Coefficient representation, correlation length, and interface sharpness" (submitted to Advances in Water Resources). The package contains the finite-volume reference solver and random-field generators (Gaussian, two-facies, graded and discontinuous conductivity; analytic spectral fields), the PINN implementations (conservative and non-conservative strong forms, mixed head-flux form, mixed form with Fourier features, with and without residual-based adaptive sampling), JSON-lines records of every run (configuration, seeds, error metrics, loss histories, gradient norms, run time and memory), the derived tables, the scripts that reproduce every figure and statistic, and the reference-convergence studies. Raw borehole records from Binh Thuan Province, Vietnam, are not redistributed; only derived statistics and the processing script are included. See README.md for contents and requirements.