Reminiscence Therapy (RT), as a typical Non-pharmacological intervention (NPI), has become increasingly important in dementia care. While GenAI-powered tools have shown potential in enhancing RT, they often struggle with practical usability and effectiveness due to the complex and situated nature of real-world intervention processes. In this study, we conduct a three-stage, in-the-wild research, including a formative study with professional therapists, a therapist-involved iterative participatory design study, and an in-situ evaluation with key stakeholders, to explore how GenAI-powered tools can be effectively designed and used to support personalized, professional, and collaborative RT practices. Our findings reveal that while theory-driven tools can achieve technical and therapeutic effectiveness in the lab, the practical effectiveness is not solely determined by their theoretically grounded design, but emerges through role-stretched and negotiated use within distributed, multi-stakeholder care practices. Based on our findings, we propose a series of practical implications for future GenAI-powered intervention technologies.