The satellite position and clock correction are similar in scale as shown in figure 1 and 3. However, as described in previous sections, the accuracy of the satellite position calculated from broadcast ephemeris is about 160 cm and 5cm for the precise ephemeris.
I don't think the initial positioning of satellites matter, I'd doubt they're carried with centimeter accuracy. What you need is to get very precise measurements of relative positions once you're in orbit. Then you get much more space for small errors in the original placement.
I've looked at this recently and just created a hn account to respond. GPS satellites broadcast their position by transmitting ephemeris parameters. You can compare the accuracy of these parameters with post-processed, high accuracy orbits from igs.org. For the last few months, the real-time ephemeris has had an RMS error of about 0.5 meters (95 percentile is about 1.3 meters). (This is the range error, which includes both satellite position error and satellite clock error)
Of course, these errors cancel out if you're doing differential GPS techniques. (Where all that matters is accuracy relative to the base station)
One thing that would be nice is to access IGS's real-time streaming orbit data over the internet and have the phone's GNSS chipset use that instead of the broadcast ephemeris. However, you might still want dGPS for single frequency receivers to help with ionospheric error.
My bet is that they position the satellite in a stable orbit, or in an orbit for which they know the rate of drift. Then the satellite can adjust the position it emits to account for its variation in movement.
Then the question becomes, how to know exactly where a satellite is with precision. This can be done from a few fixed point on Earth with high accuracy, I guess using a technique like described here (using reflectors and the speed of light): https://youtu.be/dsRsap2_RAc?t=3m37s
