Beam-tilt based focus measurement is based on the coma effect created by tilting the beam. It does have limits. You can learn the details in Typke and Dierksen (1995) Optik 99:155-166. To summary it, the aberration that gives images shift for a defocused image under different beam tilts also gives apparent defocus and astigmatism change. As a result, the two images that are correlated in this measurement are not related only by a translation, and hence the double peak you see in your correlation image. There is a similar type of distortion in the correlation image shown in [Optimizing_Autofocus]
At high magnification (or ,to use a better term, smaller pixel size), the additional aberration is more obvious because the details of the image are more resolved. The split of the real peak does make peak detection harder, so I am not surprise of your problem. It would have been true on any microscope. However, I can also see streak at x=0 and y=0 in your correlation map which may be related to the filtering function you added since I've never seen that before. The double-channel gain correction used in Leginon usually take care of the origin peak. You may want to check your gain corrections in this case.
As for the defocus measurement been off by 30 um, even when the correct peak was found, you should check if the defocus calibration done in Beam_Tilt node is reasonable at this mag. During the calibration, two defocii are used so you should see the correlation peak position change between them. For example, if you are well focused to start with and use 2e-6 m and 4e-6 m for the two defocii, the correlation peak distance to the origin should roughly double. If both are so close to the origin so that a difference can not be tell, the calibration won't be usable. In such a case, the measured image shift would be transformed into defocus incorrectly.
Another point is that you did the Z-focus at 2.5 kx, which is 50 times smaller in magnification which might not be good enough to bring the defocus variation within the limit of linearity for a measurement at 120 kx. If you need to do the focus at such a high magnification, you might want to add an intermediate mag for a medium defocus or z-height correction to bridge the gap.
Regarding possible JEOL1230-specific configuration problem, the only thing I can say is that you should check it yourself. We wrote pyscope assuming the returning value from tem subclass is in SI unit. Whether JEOL-1230.py and JEOL-1230lib.py written by Minghui makes the conversion from JEOL-1230 readout to SI unit or not is not what I can tell from the reading the code alone.
