ER-JEPA: Experience Replay Improves Joint-Embedding Predictive Learning in Language Models | TickerVault