Both faults reported came from the same wrong anchor. Character.m_eye is the
camera's eye point, not the model's: it lives outside the scaled rig, so on a
Dvergar - who renders at 78% - the quads hung where a full-height human's face
would have been, a foot above the head. And it turns with the look direction
rather than with the face, so they pointed wherever the player was looking rather
than wherever the head was.
They now hang off the player root, unscaled, and are placed onto the head bone in
world space every frame: position from the bone, orientation from the player's
own facing, and every offset multiplied by the race's visual scale so they shrink
with the body. That also means the placement constants are honest metres measured
from the base of the skull, which is what makes them tunable at all.
Driven per frame from the two places that already tick a Dvergar - the race
script for the local player, the visual controller for remote ones - because
world-space placement has to be re-applied rather than inherited.