This is usually done by driving the motor very slowly open loop at startup. The idea is that the rotor will follow the magnetic field. Once you think the rotor is in sync with the drive, you speed up the drive and look for the back EMF. Since the magnitude of the back EMF is proportional to speed, you need to get the motor going at least somewhat before sensing the rotor position is possible.
It may not be obvious this is happening when watching a motor start up. The initial slow open-loop part may be only a fraction of a rotation. Sometimes you can see these types of motors go a little bit in the opposite direction when starting up. There is a 50% chance of that happening with a random unknown starting position.
So you start with setting up the stator field to one of the six possible combinations and trust that the stall torque would be lower than the one you are creating with the field?
You don't just start with a single magnetic orientation. That could result in low torque for the position the rotor happens to be in. Instead, you slowly rotate the field, knowing that the rotor will lock to the rotation in less than 180°. From there, you increase the rotation speed to high enough to be able to sense the back EMF, and off you go with position feedback.
If the rotor won't lock at slow speed (when there is essentially no back EMF cancelling out the drive, and hence maximum torque), then you've got no chance of running the motor anyway.