China’s coast is not a passive conduit for land-derived microplastics, but an active and heterogeneous filter that alternately stores and exports particles. Here, we integrated sediment microplastic data from 17 representative coastal systems along the China coast with auxiliary water-column observations and multi-source environmental and anthropogenic predictors. Surface sediments in the upper 5 cm stored an estimated 171.4 t of microplastics, with sediment abundance ranging from 12.4 to 185.0 items kg⁻¹ and mass density from 0.12 to 8.90 kg km⁻². These stocks were concentrated in specific bays and coastal segments rather than distributed uniformly, indicating that regional trapping conditions are as important as source loading.To separate source pressure from local retention, we developed a source-normalized Coastal Retention Efficiency (CRE) framework by comparing observed sediment stocks with expected stocks reconstructed from source-pressure variables. CRE revealed a mosaic of retention-dominated systems, transitional zones and offshore-export gateways, demonstrating that China’s coast functions as a source–sink regulator rather than a simple export pathway. We further developed a spatio-temporal graph self-attention model to predict CRE from independent hydrodynamic, sedimentary, particle-property and anthropogenic drivers. The model outperformed graph-only and sequence-only baselines, while ablation and SHAP analyses identified graph connectivity, sedimentary trapping and hydrodynamic confinement as dominant controls of retention–export behavior. Future projections showed that offshore-export risk will reorganize spatially rather than increase uniformly, with several coastal systems emerging as potential export gateways. These findings redefine coastal microplastic pollution from abundance mapping to source-normalized retention–export diagnosis, providing a transferable framework for identifying coastal sinks and future offshore-export risks.