Next-release planning and feature selection remain critical challenges in the software deployment cycle, especially with the rapid introduction of updates and AI-driven functionalities. Selecting a subset of high-impact features that balances customer satisfaction with stakeholder business goals requires scalable models capable of producing near real-time predictions from requirements descriptions. However, uncertainty about which features truly align with customer needs makes this task difficult. This paper proposes a scalable feature prioritization framework that leverages rules derived from feature perplexity, probabilistic relevance, and customer-oriented criteria. We introduce a dual deep-reinforcement learning (DDRL) architecture that recommends high-value requirements through a hierarchical strategy. The first level applies graph reinforcement learning (GRL) guided by rule-based reasoning to structure tasks and reduce search space, while the second level uses deep reinforcement learning (DRL) to refine predictions. Experimental results on Zoom, Webex, and Teams datasets show that the approach maintains linear complexity ($O(m)$) compared to the cubic complexity ($O(m^3)$) of traditional graph-based methods, where $m$ is the number of edges. The model achieves up to 85.56% accuracy and superior decision quality (false negatives at just 11.28%), outperforming GraphRL, REST, and NSGA-II in stakeholder-driven next-release planning.