In comparison to the sensors space data, the source space data look reasonable to me when looking at the cortical surface. However, there are also profound differences deeper within the interhemispheric fissure when looking at the 3D MRI.
It looks like your estimated sources are very similar on the cortex surface or in the volume: in the surface tab, click on the "left" of "right" buttons to show only one hemisphere, and you will probably see the same very central sources. The two approaches are expected to give similar results when using unconstrained source models.
The MNE solution reconstructs source maps involving all the dipoles of in the source space provided. Both in the surface and volume source spaces, there are some deep dipoles, which get attributed some activity. The activity of these central regions is a linear combination of the EEG data for the surrounding electrodes. If you have one electrode nearby that has an important increase in a condition A vs a condition B, then the value of these deeper sources will also increase significantly.
Event if at the source reconstruction step, the current values estimated in these regions is lower than in more superficial regions, it increase might still be significant, and show up in the normalized maps and in the final statistical analysis.
In the end, I think what you observe deep in the frontal lobe is only a projection of the high deviation observed in your right-frontal electrodes.
And this right-orbitofrontal pattern could be related with ocular artifacts. Check again carefully for eye movements and blinks in the epochs of your condition B - review them one by one with the shortcut F3.
Further, I was wondering whether the db normalization for the TF source data is reasonable? I saw that you recommend z-scores?
Yes, this is reasonable. Prefer using a type of normalization 1) you are familiar with and you understand well 2) that is commonly used in your research field and for the type of data you process.