It seems to be an issue with this specific dataset. I made new set of segmentations (Prostate label and AIF label) and stored them each as DICOM segmentation. And now I find that the segmentation nodes are found as expected. It needs to be determined if the old data set is somehow corrupted or if it is a backwards compatibility issue (the masks were initially created in 2019).

I compared the geometry of the dynamic image, the old mask (which could not be found anymore) and a new mask (is found), using the Details info box in MITK. I found small differences in the geometries between the old mask and the dynamic image, but not between the new mask and the dynamic image. Namely:
Spacing:
dynamic image: [1.0156, 1.0156, 5.99998]
old mask: [1.0156, 1.0156, 5.99997]
new mask: [1.0156, 1.0156, 5.99998]

Matrix: (an in analogy inverse matrix)
dynamic image:

[1.01547 -0.0150323 0.0325525 
 0.015929 0.983792 -1.48688 
 -0.00161227 0.251736 5.81274]

old mask:

[1.01547 -0.0150323 0.0325525 
0.015929 0.983792 -1.48688 
-0.00161227 0.251736 5.81273]

new mask:

[1.01547 -0.0150323 0.0325525 
 0.015929 0.983792 -1.48688 
 -0.00161227 0.251736 5.81274]

Normal:
dynamic image: [0.0325525, -1.48688, 5.81274]
old mask: [0.0325525, -1.48688, 5.81273]
new mask: [0.0325525, -1.48688, 5.81274]

Discussion points: Could this be the reason for not being able to select the mask? And if yes, would it maybe make sense to make the comparison less strict?

In the MitkNodePredicateGeometry to precision is set to 1e-6. So this is at least one of the issues why the old masks are not found. It could be an option to relax the precision, but since the rotation matrix might have a large effect even for small inaccuracies, it needs to be thoroughly tested first how big the effect is if. e.g 1e-4 is accepted.
For now, the new segmentation masks are provided as test data instead of the old masks.