Understanding complex ocean systems requires models that not only predict future states, but also reveal the causal mechanisms shaping system-level dynamics. This work develops a causality-centered framework for complex-system modeling, causal emergence discovery, and state-transition regulation. We first propose CORAL-Net, a causality-oriented spatiotemporal graph learning framework for coastal marine ecosystems. Instead of relying only on geographic proximity or statistical correlation, CORAL-Net integrates causal discovery into representation learning. At the station level, it combines physics-informed spatial connectivity, functional similarity, and adaptive temporal dependency graphs to capture multiscale propagation across monitoring sites. At the variable level, PCMCI-based causal discovery is used to infer directed lagged dependencies among biogeochemical variables, which are then encoded through a multilayer causal aggregation module. Experiments on a long-term Hong Kong coastal monitoring dataset from 2000 to 2022, covering 76 stations, 24 variables, and multiple vertical layers, show that causal structure can provide an effective inductive bias for marine ecosystem forecasting. CORAL-Net improves prediction performance over recurrent, attention-based, and temporal convolutional baselines, while revealing interpretable pathways such as Suspended Solids → Total Volatile Solids, Kjeldahl Nitrogen → Total Nitrogen, Nitrate Nitrogen → Inorganic Nitrogen, and Chlorophyll-a → Biochemical Oxygen Demand. We then move beyond pairwise causal relations and study causal emergence using partial information decomposition. By decomposing the contribution of multiple source variables to future system states into unique, redundant, and synergistic components, this approach identifies higher-order causal structures that cannot be captured by pairwise edges alone. Its effectiveness is validated in both real-world and generative dynamical systems. We further examine its role in state-transition regulation, especially in ignition tasks where a system is driven from one basin of attraction to another. The results suggest that synergistic drivers identified through causal emergence can provide useful guidance for selecting cooperative intervention targets. Although this regulation framework has not yet been fully applied to coastal ecosystems, it points to important future applications in marine and ecological systems, such as early warning of regime shifts, coordinated control of nutrient inputs and algal blooms, oxygen-demand regulation, and the identification of multi-factor intervention strategies for maintaining desirable ecosystem states.