Software engineering research aspires to improve real-world software practice, yet many of the most consequential advances in modern development and operations are created and disseminated outside mainstream SE venues. We argue that this disconnect is driven less by individual choices than by structural incentives: "puzzle-solving'' on convenient artifacts, reliance on small or unrepresentative projects, disclosure norms that unintentionally exclude industry-scale studies, and slow, overloaded, high-variance peer review. Left unchanged, these forces can produce an "irrelevance cascade'' in which SE venues remain academically active but increasingly disconnected from practice. We propose near-term reforms that raise the weight of demonstrated practical impact, encourage industry-scale evaluation (including open-source "twins''), and use AI to reduce the mechanical burden and variance of review while improving reproducibility checks. We also advocate repositioning major conferences such as ICSE and FSE as impact-curated selection venues: dissemination is continuous through preprints and tool releases, authors provide impact statements, community members publicly champion or critique work, and program committees curate the most consequential results for presentation and synthesis. Finally, we sketch a longer-term vision of an AI-augmented, agentic research ecosystem that lowers the transaction costs of experimentation, replication packaging, and evaluation, enabling the community to return to harder, higher-leverage problems.