More stats, yippee!
OK, looks noisier than yesterday, but look at the scale! The vertical velocity is no longer growing, and it’s nicely averaging out the noise from the accelerometer.
The problem? It lay in the calibration of the acclerometer (the gyro was fine, oddly, as they use the same code). The calibration is way out (by a factor of two) if the Accelerometer is read too often (it was at 100Hz). Slow it down to 50Hz, and all’s fine. Importantly, the calibration values between runs also seem consistent, which means I can save them to file, and restore them for use when takeoff is not on a completely horizontal surface.
So that’s the next step.