gRPC

Deploys that survive a dropped stream or a control-plane restart

How I moved each Railyard build into an agent-side session that outlives the gRPC stream, so a ping timeout or a control-plane release no longer kills a customer's deploy, and the Go context bug that came with it.