After three sessions more or less successful, it seems that the cause for the car going really bad is the RTK station mount point chosen: there are significant differences.
On the first session all went bad. Changing the ground plane, connecting it to electrical ground or not, disabling the internal compass and changing parameters didn’t make significant changes. As suggested, I finally left ATC_STR_RAT_P=0.3 (previously 0.5), which seemed to stop steering oscillations.
On the second session I got two good runs:
Including the hud undocked in these videos with the tlogs may give clues.
The third session was dedicated to the five possibilities for the RTK station mount point. They are described elsewere and are physically different, and one of them is virtual (which gave the worst results). It must be taken into account that they are meant for different needs and which can be very far away from the RTK physical station, but for these tests the distance is 14 Km.
This is the tlogs video with the hud at 1x speed:
These were the trajectories obtained with two of the mount points, which were unacceptable (the second one is virtual):
This other was a bit better:
This other was almost acceptable:
Finally, these were two runs with a mount point with good results:
So the conclusion is first of all if having several mount points, try all and choose the best one. I wonder if there is any parameter (as hdop and satellites for the GPS) to evaluate the quality of the mount point connected, since what appears in Mission Planner RTK screen showed good data for all above mount points.