tl;dr At best CASSANDRA-16807 means some queries that shouldn't fail do fail, and at worst, this is a potentially dangerous consistency problem.
https://issues.apache.org/jira/browse/CASSANDRA-16807 I'm working on a solution either way, but I want to see if there are any opinions out there on whether this should be included in 4.0.0 or bumped to the first patch release. It's not strictly a regression from 3.x, although the assertion that helped us identify it only exists in 4.0.
