Explainability is an important non-functional requirement for AI-enabled systems. Despite growing interest in human-centered approaches, prevailing frameworks remain largely system-centered and grounded in technical properties rather than stakeholder needs. In particular, we lack empirically derived, stakeholder-driven explainability requirements for ethically constrained autonomous decision-making. This paper addresses that gap through an empirical-first requirements engineering approach. We conducted five scenario-based focus groups with 20 participants. Participants represented four stakeholder roles, including those with experience in autonomous systems, expertise in ethics, operational experience, and community members without domain-specific expertise. Participants discussed ten ethical dilemmas involving autonomous emergency response drones. We applied Socio-Technical Grounded Theory to 282 codable segments and derived a seven-category taxonomy of explainability requirements. Four categories cover established XAI concerns such as epistemic transparency and decision rationales, while three focus on previously under-explored areas of real-time operational communication, institutional accountability, and human-machine decision-authority boundaries. From these categories, we produce an initial set of requirement patterns that link stakeholder roles and scenario ethics. In a post-study validation survey, participants confirmed that the taxonomy categories captured their expressed concerns, and taxonomy-informed explanations were preferred over generic ones across multiple evaluation dimensions.