Mathematics anxiety (MA) is reliably associated with lower achievement, but how it shapes engagement during learner interaction with coordinated representations in digital environments is less known. This study examined MA-related differences in emotional and cognitive engagement dynamics while learning with siMMath, an interactive simulation that integrates dynamic manipulatives with explanatory text to support proportional reasoning for medication calculations. Seventy undergraduate nursing students completed a lower-complexity unit-conversion phase and a higher-complexity concentration and infusion-rate phase, with synchronized multimodal data collection (PANAS, automated facial expression analysis, and eye-tracking analytics). As task complexity increased, visual attention shifted from predominantly processing digital manipulatives to a more balanced allocation across representations, including more frequent revisits of the explanatory text. Higher MA was associated with lower test scores, and more negative self-reported affect during learning mediated the MA-performance link. These findings suggest that emotional engagement is a key pathway through which MA undermines performance, and they provide design-relevant evidence on how learners coordinate interactive and textual representations under increasing task demands. Future research should examine how specific design features of dynamic digital representations can better support learning in higher-complexity learning tasks.