I wrote two sources, which I've yet to hear back from. It may be the "vacation" week. So to provide a speedier -- and still expert -- answer I posted the question in Bruno's forum here. His answer is as follows:
IMNSHO, balanced cable connectors should've never had three pins to begin with. This is what caused the confusion in the first place. Equipment is supposed to respond only to the difference voltage between the hot and cold pins. The ground pin should be used only roughly to equalise the ground/chassis potentials of the interconnected equipment. Had XLR connectors had only two pins, the function of the shell would have obviously been to connect the shield to the chassis.
Unfortunately, a profusion of different wiring practices has spawned quite a few boxes whose performance degrades when current flows into pin 1. In a typical bad case, pin 1 is wired to the circuit board, allowing all current flowing through the cable shield to make its way into the internal ground of the circuit. This is why some people have resorted to the practice of leaving the shield open on the receiving end or connecting it only through a capacitor. At least this cuts out hum in pin 1 sensitive gear, although RFI is still free to come in through the front door.
A new standard, AES48, finally normalises the situation (unfortunately compliance is optional) by requiring pin 1 to be directly bonded to the chassis and the XLR shell. Two pieces of AES48 compliant equipment will always interface correctly because no amount of current into pin 1 can cause currents anywhere in the audio circuit. Ground loops are a fact of life in any sufficiently complex setup and AES48 compliance insures that they are allowed to exist without impacting audio performance.
So, if all equipment were correctly designed, the answer to your question would've been a definitive "both pin 1 and shell". Pin 1 would suffice for low frequencies, the shell would help shield the inevitable pig-tail ground connection. (For best results, Pin 1 ought to be connected to chassis ground rather than signal ground.) But lacking AES48 compliance across the board, the answer is, unfortunately, "depends".
I believe this then to be the answer the people in this thread arrived at. It would be nice to connect to the shield but for those of us who have to work with what we find, connecting to the shell isn't practical.
Barry