Low-latency model loses yaw responsiveness in Planner VR3PT mode

[LEECHOONGHO]

Hellow,

We're seeing a difference in yaw turning behavior between the standard model
and the low-latency SONIC model:

• Standard model: yaw works well in both Pose and Planner, VR3PT mode
• Low-latency model: yaw works well in Other mode, but barely responds in
  Planner VR3PT mode — only exaggerated direction inputs** produce visible turning

** facing = self.yaw_accumulator.update(rx * 3, self.dt)

Looking at GetObservationRegistry() in g1_deploy_onnx_ref.cpp, the anchor
orientation observation has three modes: full base quaternion (mode 0),
heading-only/yaw-isolated (mode 1), and refheading (mode 2). The 10-frame and
2-frame variants register all three, but the 4-frame (low-latency) variant
only registers mode 0:

{"smpl_anchor_orientation_4frame_step1", ...}  // no heading/refheading version

So the low-latency model can only observe the full quaternion (roll+pitch+yaw
combined), not a yaw-isolated signal. Since Planner mode's planner_motion_
carries more roll/pitch noise than a fixed Pose-mode reference, this may be
diluting the yaw signal enough to cause the weak turning response.

Could you confirm whether this is intentional, and whether adding
heading/refheading variants for the 4-frame observation set would need
retraining, or could be tested against the existing checkpoint?

Happy to share logs or a repro video if helpful.