Mountain-glacier retreat is commonly described as a gradual response to warming — yet whether all retreating glaciers follow the same path remains an open question. Using a deep-learning segmentation model across five satellite sensors (Landsat 5/7/8/9, Sentinel-2), we reconstruct 786,014 glacier-year observations of end-of-season snow fraction across 22,155 glaciers in Alaska and western North America (1984–2025), and show that three structurally distinct decline regimes coexist under the same regional forcing — with no way back for many.

Family A (48%) maintains buffered snow persistence through topographic sheltering, with intact accumulation-zone capacity. Family B (35%) was indistinguishable from Family A until a collective regime shift concentrated between 2013 and 2020 pushed each glacier past its own internal threshold. Family C (15%) has been in structural decline since 1984, governed by regional climatic position rather than local geometry.

We introduce glacier decapitation — the persistent, satellite-detectable absence of end-of-season snow across the entire visible glacier surface. Of 21,748 glaciers, 9.1% have experienced at least one episode, with more than half occurring in the last five years (Mann–Kendall τ = 0.432, p < 0.001). Reversibility is asymmetric: Family A glaciers typically recover, while 66% of decapitated Family B glaciers never do. Static attributes alone predict decapitation timing (Cox C = 0.749). If the Family-B post-break state proves irreversible — as current persistence rates suggest — Families B and C combined represent nearly half the region's glaciers on a trajectory of sustained bare-ice exposure, with diminished capacity to buffer seasonal water supply downstream.