Note the last sentence of your quote. Device-to-device variability and production yield are opposites. I can't imagine how techniques that exploit variability between devices (with no mention of temperature, lifetime drift, etc) could be reliable.
Maybe not the best example, but this was on FPGA. When making custom chips, you can make all sorts of tuning, including consistent hardly noticeable drifts from spec.
The article says they only image the metal layers, so I guess you can do anything as long as you don't visibly change the metal layers.
