Hyper-arid agricultural soils combine low organic carbon, high calcium carbonate contents, salinity, and frequent gypsum accumulation, making rapid quantification of soil organic carbon (SOC), soil inorganic carbon (SIC), and gypsum valuable but analytically demanding. We evaluated visible–near infrared (VNIR) spectroscopy for predicting these three properties in 216 samples from 18 sites spanning two oasis management systems, three topographic positions, and six soil depths, with particular emphasis on how validation design affects apparent model accuracy. Partial least squares regression (PLSR) and random forest (RF), single- and multi-target workflows, and global versus stratified calibrations were compared using site-based GroupKFold, random KFold, and leave-one-out cross-validation (LOOCV). RF outperformed PLSR across all targets. Predictability was strongly target-specific: gypsum was predicted most reliably (R² = 0.79), SOC moderately (R² = 0.46), and SIC least reliably (R² = 0.36), with stratified SIC models often failing under site-blocked validation. KFold and LOOCV consistently yielded more optimistic estimates than GroupKFold, inflating R² by up to 0.17 for SIC. VNIR spectroscopy therefore shows strong potential for gypsum, moderate utility for SOC, but limited transferability for SIC. Validation design critically determines how soil-spectroscopy performance should be interpreted for dryland monitoring.