If we haven't even learned to avoid the one bias that we can measure super well and which is most susceptible to training, what are we even doing here?

This sounds like a job for cognitive psychology!

"Well-calibrated" should probably be improved to "well-calibrated about X"-- it's plausible that people have better and worse calibration about different subjects, and the samples in the survey only explored a tiny part of calibration space.