Closed-loop learned locomotion has matured on commercial quadrupeds but remains uncommon at the sub-kilogram scale. We present a single learned policy that handles throw recovery, multi-posture walking, and hardware variation on the MiNI-Q, a 270 g quadruped whose belt-driven legs rotate continuously through 360°. The same trained network switches between posture targets on the joint torus T⁸, walks upright and inverted under a gravity-conditioned reference, and recovers from arbitrary spawn orientations, all without state machines or phase switching. We identify the Dynamixel XL330 actuators across their operating regimes, cross-validate the trained policy in MuJoCo, Isaac Gym, and Isaac Lab, and deploy a compact sub-100k-parameter network onboard at 50 Hz on a dual-microcontroller stack. We further train under foot-shape domain randomization and verify on hardware that the same policy walks across multiple printed lower-link variants swapped at test time. The result is a single learned controller for throw recovery, multi-posture walking, and hardware variation, all on the same miniature robot.