I've only read the ars writeup, and not the paper itself, but here's the impression I get:
Using your stated setup, the apparent "violation of causality" from this experiment comes in that measurements showing that A & D are entangled happen *before* the entangling of B & C. So in a completely arbitrary sort of timeline-form:
0:00 - A & B are entangled. C & D are entangled.
0:01 - A is sent to Alice. D is sent to Bob. B & C are sent to Victor.
0:02 - Alice measures A. Bob measures D.
0:03 - Victor decides whether or not to entangle B & C.
In the cases where at 0:03 Victor decides to entangle B&C, the measurement taken at 0:02 shows A & D as entangled.
In the cases where at 0:03 Victor decides *not* to entangle B&C,the measurement taken at 0:02 doesn't seem to show A&D as entangled.
From this, we get the apparent break in causality, in that a measurement taken at 0:02 is apparently affected by something that doesn't happen until 0:03.