According to the people at Imaging Resource, they "micro-gel" (see link) their HMI lights. That would cause their lights to lose some brightness over time, though the effect may be too slight to matter.
And importantly, that link makes no promise of lighting intensity consistency from one test to another, only of consistency across the frame of any single test shot to within 0.1 stop (and even that isn't so much a promise as a statement about the lighting at the time it was measured in 2006).
So I guess it's likely the lighting intensity isn't kept perfectly consistent over time. (That the old 5D Mark II metered the same as the 5D Mark III doesn't prove the lighting was the same, as you know, since different camera models definitely have metered differently from each other over the years.)
And since you mentioned it, I had a look at the softbox reflections in the bottles. The top-right reflection, in particular, does look dimmer in the 5D Mark III photos compared to the D800 shots taken a week later. Did they replace that HMI globe, perhaps?
DxOMark will hopefully test these sensors soon. DPReview and Imaging Resource don't seem to realise that exposure consistence is critical to comparing noise performance. The exact ISO calibration and metering behaviour isn't really important; what's important is shooting a scene of identical brightness, at identical f-stop and shutter speed, and developing the resulting images to the same brightness. At that point, you can compare the images. Otherwise you're comparing ISO calibration and meters, something that might be interesting for actually using the cameras, but something which doesn't tell us much about image sensor performance.