They are being deleted because those taxa are not actually accepted by POWO. Instead, POWO only accepts a name for the taxon. Accepting a name and a taxon are not the same, as I told you before when I warned you about the relationships you were making being deleted.
Note the difference between the two linked pages:
× Amelasorbus jackii “This hybrid is accepted”
× Amelasorbus raciborskiana “This is an artificial hybrid”
In visiting the taxon page for the parent taxon, × Amelasorbus, you will notice only one of these two taxa listed under “Accepted Species.” This is because the taxon designated an artificial hybrid is not an accepted taxon. Again, accepting a name is not the same as accepting the taxon.
As curators, we are told to broadly avoid adding or maintaining hybrid taxa. This applies to all hybrids, especially when they are not accepted by the framework being sourced, as is the case here. Personally, I do enjoy wild hybrids, but iNaturalist is not intended to be a repository for observations of every taxon ever described. Artificial hybrid taxa fall into the bin of taxa that should not be added or maintained, with fossil taxa and cultivar names.
You are right that automated/staff-side taxon framework relationship behavior should be described in a place for curators to read somewhere. This behavior can occur under other circumstances, and is the reason for my feature request here: https://forum.inaturalist.org/t/generate-a-taxon-history-entry-when-taxon-framework-relationships-are-modified-deleted/54458
Edit: Forgot to mention that any improperly formatted taxon relationship will be deleted automatically. iNaturalist seems to detect and delete relationships that do not correspond to a real entry in the framework source. In my experience this applies to misspellings, inaccurate URLs, taxa that are not listed in the source as accepted, and possibly more.