Thanks for the report! I can see where the confusion is. Just to be clear, you are retrieving the refined structure, it is just in the wrong setting (P1).
The refined/primitive/conventional standard structures are generated by pymatgen’s
SpacegroupAnalyzer class (which in turn uses
spglib). When doing the refinement, sites will be moved within a given tolerance to be in their expected symmetry precisions. This has been done correctly, and the structure has been refined.
However, the problem is that the outputted CIF file, while it does have the correct symmetry-adjusted positions of the sites, is not actually outputting the CIF file in this setting, but is outputting in a P1 setting by mistake. This bug is related to Significant release of new database schema and additional data and has now been fixed internally.
You should see the correct CIFs on the website after our next database build, so likely in ~1 week.