Cirque morphology in High Mountain Asia (HMA) is a key paleoclimatic proxy, yet quantifying how much it reflects paleoclimatic legacy versus present-day environmental context remains challenging. Here we compile a strictly quality-controlled database of 5828 HMA cirques and apply an interpretable stacked ensemble machine-learning framework with SHAP-based attribution to quantify model-based associations between predictors and cirque morphology. On an independent 20% hold-out test set, the model predicts cirque vertical scale well (Height, R2 = 0.80), whereas planimetric shape (Length/Width ratio) is poorly predicted (R2 = 0.09), suggesting that the available macro-scale predictors explain substantially more variance in vertical-scale metrics than in planform ratios, which likely depend on additional local controls not represented here. SHAP attributions further indicate a pronounced spatial transition in dominant predictor regimes. Along the monsoon-influenced Himalayas, Paleo-ELA and elevation-context variables emerge as the strongest predictors, consistent with an "energy-limited" setting where equilibrium constraints and relief boundaries are important. Toward the semi-arid continental interior (e.g., Gangdese), tectonic-enabling proxies (seismic density metrics) rank highest, with local topography (slope) playing a secondary role, consistent with a shift toward a "substrate-limited" regime. The Tian Shan shows an intermediate pattern where Paleo-ELA remains influential despite increasing aridity. Overall, these results emphasize spatially shifting predictor regimes and help reconcile the apparent mismatch between modern climate fields and relict landforms by jointly considering paleoclimatic proxies and present-day covariates within a unified, interpretable modeling framework.