The "agents as primary user" framing is what makes this stand out from other dashboard-as-code tools. Having agents generate dashboards is one thing, but making those dashboards reviewable and standardized is the actual hard part. Curious about the semantic layer — when an agent modifies a metric definition in semantic/, how does validation work? Does DAC flag downstream widgets that would break, or is it more of a "trust but verify" model where you catch issues at render time?