euler_angles_ordered returns surprising results on gimbal lock (strange sign)

[lierdakil]

Consider the following example:

Rotation3::from_euler_angles(0., -90f64.to_radians(), 45f64.to_radians())
    .euler_angles_ordered(
        [
            UnitVector3::new_normalize(Vector3::z()),
            UnitVector3::new_normalize(Vector3::y()),
            UnitVector3::new_normalize(Vector3::x()),
        ],
        false,
    )

The expected result is:

([0.7853981633974483, -1.5707963267948966, 0.0], false)

However, the actual result is:

([-0.7853981633974483, -1.5707963267948966, 0.0], false)

Notice the yaw sign is inverted.

Outside gimbal lock (so at pitch = -89.9 for example), the sign is as expected, so there's a weird discontinuity here.

At apex gimbal lock (so pitch = 90.0) the angle is also as expected.

It seems the yaw is multiplied redundantly by the sign of the pitch here.