I reworked the temperature-compensation calibration once again to further reduce its sensitivity to the history of the temperature changes the clock encounters as it calibrates itself. Also, I changed the calibration model to be a simple linear least-squares fit across the whole temperature range.
Still testing to see if the small, sudden changes in speed noted with the last update exist. Jury's still out on that. If so, it's got to be something physical. Probably something moves with a thermally induced stick-slip motion.