I like your ideas and think this could be very useful. However, one should keep in mind that writing a complete specification of the expected behavior of a node is as complex as writing the node itself; if you had a complete mapping from input to output, you could just execute it instead of running the node. So writing a full functional mapping of input to output is probably infeasible. In your example, you’re doing something else: You specify invariants that should hold (velocity magnitude is clamped, and dv/dt corresponds with accel_max), while leaving open some degrees of freedom in the behavior of the node. I think this is the way to do it.
By the way, are you familiar with rostest? An example where it’s used are the tests for robot_pose_ekf. Here they just test some simple properties of the node’s output topics and check that the final estimated pose is correct. This is probably enough to catch 95% of regression bugs.