It is important to point out that the images you are seeing in the report are not exactly what you get as final output, they are “pictures” at each stage of processing that enhance certain aspects for faster visual assessment.
What does this mean? That the “smudging” effects you see in the report panels should not be present in the final data. SDC-SYN estimates a warping that, as you mention, tries to reinstate the original brain structure. The reports present one reference image derived from the BOLD timeseries with enhancements for registration and visualization. So the main focus of the SDC-SyN panels is the accuracy of that estimation. Simply, do the contours align better with data after SDC-SyN? If so, then you should use the correction (IMHO).
FMRIPREP is designed in a way that only one step of resampling is done, precisely to avoid smudging data. So, the transforms from head motion estimation and susceptibility distortion estimation are combined (with other pertinent transforms, for instance if you want data in MNI space) and applied in only one step. For this reason, it is important that the fieldmap is estimated with accuracy.
In your reports, I could see that, arguably, SDC-SyN performed “well” on most of the cases. But distortion is exaggerated for the “rest” task. In that only case, you should avoid the fieldmap estimation done with SDC-SyN.
Finally, as Chris mentioned, if distortion is not too big (I cannot say anything about how to judge this) it may not be worth including the field map, as over-correction and correction on the wrong directions are much worse than not correcting.