Outer-loop tracker: ``f_syn`` reference (p, v, a) -> normalized CTBR action.
This lets synthesized plans and learned policies drive MuJoCo through the *same*
action interface (:mod:`rl_stl.envs.skydio_x2.ctbr`)::
a_cmd = a_ref + Kp (p_ref - p) + Kd (v_ref - v) (world, COM)
f_des = m (a_cmd + g e_z), horizontal part limited to tilt_max
T = f_des . z_body (projected thrust)
z_des = f_des / |f_des|
w_tilt = K_att_xy * R^T (z_body x z_des) (reduced attitude, body frame)
w_yaw = K_att_z * wrap(yaw_ref - yaw) * R^T e_z (heading, lower priority)
action = command_to_normalized(T, w_tilt + w_yaw)
The tilt and heading errors are decoupled on purpose: the full SO(3) error
``1/2 vee(R_des^T R - R^T R_des)`` loses its yaw component and flips the sign of the
tilt correction near a 180 deg heading error, which diverged from random-yaw resets.
Pure ``jax.numpy``; single sample (vmap for batches).