Entity-relation extraction aims to jointly identify entities and their semantic relations from unstructured text, yet it remains hindered by relation sparsity, long-tail distributions, and optimization conflicts between entity and relation subtasks. We propose SPADE, a unified framework that addresses these challenges through two core components: (1) a Symmetry-aware and Self-referential Relation Augmentation strategy that injects structurally valid training instances to improve relational diversity and robustness; and (2) a Joint Optimization with Priority Constraints method that enforces hierarchical learning dynamics via Lagrangian multipliers, prioritizing entity boundary detection to stabilize multi-task training. Extensive experiments on five benchmark datasets across multiple domains show that SPADE consistently outperforms strong baselines, achieving up to +3.1% absolute improvement in triplet-level F1 on low-resource settings. Our code and models will be released to support future research.