These figures were kindly provided by Google. We could probably reproduce them by digging in CrUX data for September with BigQuery.
The offending banner: https://meta.wikimedia.org/wiki/Special:CentralNoticeBanners/edit/wlm_2019_de
The FID regression lasted for the whole duration this banner was running. Over the same period the enwiki banner for the same event (which seems to have quite different HTML/CSS/JS) didn't see a FID regression.