Upon further investigation, I believe trendline intercept and slope is optimised for least RMS error, and that leads to a different solution than optimising for R-squared value (which is less surprising now that I've thought more deeply about it).