Crystal frequency drifts in production when load caps and PCB parasitics differ from the datasheet test circuit. How much pulling did you see lot to lot, and how did you trim it - cap selection or software compensation?