PlanarConstraint's point-in-plane residual used a body-attached normal
(z_j) that rotates with body_j. When combined with CylindricalConstraint,
which allows rotation about the shared axis, the plane normal tilts
during Newton iteration. This allows the body to drift along the cylinder
axis while technically satisfying the rotated distance residual.
Fix: snapshot the world-frame normal at system-build time (as Const
nodes) and use it in the distance residual. The cross-product residuals
still use body-attached normals to enforce parallelism. Only the
distance residual needs the fixed reference direction.
Works for any offset value: (p_i - p_j) · z_ref - offset = 0.
When a PlanarConstraint has distance=0 (point already on plane), the
half-space tracker was using dot(z_i, z_j) as its indicator and providing
a correction function that reflects the body through the plane.
When combined with a Cylindrical joint, legitimate rotation about the
cylinder axis causes the body's planar face normal to rotate. After ~90
degrees, the dot product crosses zero, triggering the correction which
teleports the body to the other side of the plane. The C++ validator
then rejects every subsequent drag step as 'flipped orientation'.
Fix: return a tracking-only HalfSpace (no correction_fn) when the point
is already on the plane. This matches the pattern used by Cylindrical,
Revolute, and Concentric half-space trackers. The cross-product residuals
in PlanarConstraint already enforce normal alignment via the solver
itself, so the correction is redundant and harmful during drag.
