Extreme climate events increasingly threaten terrestrial border ecosystems. However, their nonlinear threshold effects and cascading transmission mechanisms remain poorly understood. To address this gap, we quantified the ecosystem health index (EHI) of China's terrestrial borders from 2000 to 2024 utilizing an improved "Vigor-Organization-Resilience-Services" (VORS) framework. To capture complex climate disturbances, we constructed an evaluation system comprising 16 specific indices, categorized into four macro-dimensions: extreme high temperature (EHT), extreme precipitation (ERF), extreme drought (ED), and extreme low temperature (ELT). By coupling the explainable machine learning framework (XGBoost-SHAP-PDP) with partial least squares structural equation modeling (PLS-SEM), we systematically unraveled the nonlinear tipping points, compound interactive effects, and cascading transmission pathways of climate stress. The results indicate that: (1) Over the past 25 years, the region experienced significant asymmetric nighttime warming and precipitation extremization. The overall EHI improved but exhibited severe spatial polarization, with transitional health zones recently facing localized degradation. (2) Climate impacts are not linearly accumulative but are driven by distinct nonlinear thresholds. The synergistic coupling of extreme precipitation volume and intensity constitutes the dominant compound risk. (3) Direct physical impacts of climate on the EHI are marginal; degradation is primarily propelled by indirect cascading transmissions. In fragile alpine and extremely arid regions, moisture deficits trigger cascading systemic collapses abruptly via impaired ecosystem vigor (EV) and services (ES). Conversely, forest ecosystems buffer these stresses through their complex organization (EO) and resilience (ER), achieving a compensatory dynamic. This study elucidates the rigid boundaries of ecological vulnerability, emphasizing that future adaptive governance must pivot from passive restoration to active, threshold-based early warning systems.