Saline drylands, covering over one billion hectares globally, remain critically understudied with respect to the depth-resolved and fraction-specific behaviour of soil organic carbon (SOC), despite their disproportionate vulnerability to climate-driven carbon loss. This study characterised SOC dynamics across ten contrasting land uses (forest and orchard plantations, continuous cropping, fallow, and barren land) bordering the eastern shores of Lake Urmia, northwestern Iran, a basin that has lost over 90% of its volume and undergone accelerating soil salinisation over the past three decades. Soil profiles were sampled at four depth increments to 100 cm, with SOC partitioned into light-fraction (LFOC) and heavy-fraction (HFOC) pools and further resolved by particle size. A calibrated RothC biogeochemical model, driven by four downscaled CMIP6 general circulation models under SSP1-RCP2.6 and SSP5-RCP8.5, projected SOC trajectories through 2100. Total SOC declined sharply with depth across all land uses, most steeply under coniferous plantation (1.13% to 0.17%), while subsurface horizons were near-universally dominated by HFOC (up to 98%), indicating that apparent subsoil stability reflects depletion of labile carbon rather than active protection. SOC correlated strongly with pH and electrical conductivity at the surface (R² = 0.81 and 0.28, respectively) but weakened with depth, consistent with declining LFOC sensitivity to chemical disturbance. Climate projections indicated universally negative SOC trajectories under both scenarios, with losses by 2100 ranging from −1.4 to −2.8 t C ha-1, and onion cultivation, rather than the highest-stock coniferous system, emerging as the most vulnerable land use under long-term, high-emission warming. These findings indicate that vegetation continuity, rather than land use category alone, governs both present SOC fraction distribution and future carbon security in saline drylands, and suggest that standard biogeochemical models may misrepresent subsurface vulnerability in salt-affected systems unless explicitly informed by fraction-resolved data.