This is resolved.
For anyone interested: Reason was not within strapi. Instead, we used a clustered postgres backend with load balancing that did not guarantee “visibility consistency” - i.e. the results of a committed transaction were not guaranteed to be visible in an immediately following read request.