Modern ad delivery systems rely on multi-stage ranking pipelines in which retrieval, early ranking, and final ranking progressively refine candidate sets. While this architecture scales well, limited cross-stage coordination often leads to redundancy: early stages forward candidates that are highly filtered downstream, and lightweight signals are underutilized for efficiency optimization. We propose Cross-Stage Signal Propagation, a unified framework for negative filtering in multi-stage systems. The framework supports two complementary signal flows: forward propagation from early ranking to final ranking, and backward propagation from early ranking to retrieval. Forward propagation introduces a consistency-driven filtering gate at the final ranking stage; by detecting signal stability, the system reuses cached outcomes to bypass final ranking for low-value ads. Backward propagation reuses early-stage feedback at retrieval to suppress persistently low-value ads before they enter the ranking pipeline. Deployed on Facebook Feed, the framework achieves significant efficiency gains beyond Negative Exclusion Filtering, including 10.68% CPU reduction in final ranking, 11.20% CPU reduction in early ranking, and 13.01% GPU reduction, while maintaining stable user and advertiser outcomes. This is expected, as the filtered candidates are consistently low-value, characterized by low expected return (e.g., low bid and/or low click-through rate) and low quality, and therefore rarely survive final-stage ranking or win auctions. These results demonstrate that treating ranking signals as reusable resources substantially improves the efficiency of large-scale ad delivery systems.