The culprit in this approach is the leg from VARAP to D106N (which is the point 14nm from GSE along the 106 radial - reciprocal of the 286 degree inbound course).  In his picture, it's D106P, but you must be using an older database cycle.  That leg is defined as a DF (Direct To Fix) leg, which does not define an inbound course.  So the FMS will try to construct a path from the aircraft position to that point right as VARAP is sequenced.  The path consists of an entry turn and a geodesic path to the point.  The entry turn is computed using the current aircraft course as the starting direction.  The end of the turn is computed as the point along the turn where the geodesic path is tangent to the turn.  *If* the aircraft current course is within two degrees of the course to the waypoint, then the system won't bother generating the entry turn.

In this case, if the aircraft is on path at the hold fix, its course will be 320°.  The course from the hold fix to D106N is ideally 310°, but depending upon the aircraft course when the leg sequences, it could be off of that.  In your screen shots, it is actually 300°.  So even in the ideal situation, the difference between the hold inbound course and the course to D106N is more than the two degrees, so the system tries to compute an entry turn and then a geodesic to D106N.  Because the distance to D106N is so short (roughly a half mile), the system cannot compute the entry turn and a geodesic leg before getting to the waypoint.  So, it generates a nice big circle in order to hit the waypoint.

We're looking at some improvements we can make in the next release.