Every other sample hands the runtime a display rig: the canvas is a portal onto a virtual display, and the scene is authored to fit it. This one hands over its own camera instead — position, orientation and vertical FOV — and the runtime perturbs that frustum with the viewer's eyes. Drag to orbit; the framing stays yours, the stereo stays the runtime's.
It is the distance where content sits on the glass — nearer pops
out, further recedes. Left at the orbit distance the cube straddles the display. Wind it
in and the cube leaps forward; wind it out and the whole scene sinks behind the panel.
The runtime's comfort rule is
ipd × metersToVirtual × diopters × 0.5 ≤ 1; the readout above turns red past
it, which is where far content starts asking your eyes to diverge.
The display rig it switches to is built to frame the cube identically at the home angle, so the cut is invisible standing still. Then orbit: the camera rig follows you around the cube, the display rig does not — it is a fixed window onto a fixed scene. That difference is the choice between the two.
attach sends an identity-posed rig and parents the eye cameras under the app camera instead, so three composes this frame's world pose. Toggle it while whipping the orbit around: without it the stereo trails the render by one frame.