<p>Understanding how land-cover configurations influence spatial carbon performance (SCP) is critical, yet existing studies often overlook the scale-dependent and nonlinear threshold effects across urban–rural gradients. Taking Tianjin as a case study, we construct a four-tier gradient framework. By coupling Multi-scale Geographically Weighted Regression (MGWR) with the LightGBM-SHAP machine learning approach, this study decodes the spatially heterogeneous, nonlinear, and interactive mechanisms driving SCP. The results reveal a fundamental shift in land-cover impacts from the urban core to the rural periphery. Crucially, we identify explicit morphological and ecological thresholds: in high-density urban cores, building density exhibits a pronounced inverted-U response, with SCP peaking only when density ranges between 2.5 and 3.5, beyond which heat-island penalties and energy consumption outweigh agglomeration benefits. Conversely, in rural and peri-urban zones, ecological drivers dominate; notably, green-space coverage demonstrates a critical threshold at 15%, beyond which its marginal contribution to SCP shows a substantial synergistic increase. Furthermore, green-space connectivity triggers significant carbon-sink benefits once exceeding an index of 0.45. Furthermore, complex interactions, such as the coupling of road network density with topographic variability and functional diversity, are quantified. These findings advance the transition from "one-size-fits-all" emission policies to gradient-specific, threshold-guided spatial optimization, offering actionable evidence for sustainable territorial governance.</p>