A minimum of 100 cycles. I would overshoot by maybe 5% to 10% to ensure that what you measure does not include any period variation from the output transient results (so 105 to 110 cycles). That is, make sure you're not seeing any variation in the period for a few cycles. Does that make sense?
For example: Let's say you've calculated that the lock time is 100 nSec. Set the simulation time to 105 to 110 nSec. Look at the results and see if your measurements are consistent.
Play with the simulation time until you're satisfied that the output is completely settled. What happens as you vary corners is that the period will vary (as expected). You will want to capture enough cycles to make sure you have an accurate measure of this change over corners.
Hope that makes sense.
fd