This is broken at the protocol-level and in the implementation.
At the protocol-level the <flags> component of the batch comes after the queries. That means the protocol parser would need to read ahead (and back track) to determine the values encoding and correctly read the values from the query entries. Also, a batch-level setting for named values forces all queries to use the same encoding. Should batches force a single, homogenous query value encoding? (This is confusing)
In the implementation, values are indiscriminately read using CBUtil.readValueList(), and the batch flags are never checked (for (Flag.NAMES_FOR_VALUES) to see if CBUtil.readNameAndValueList() should be called instead: https://github.com/apache/cassandra/blob/cassandra-2.1/src/java/org/apache/cassandra/transport/messages/BatchMessage.java#L64
Proposed solution: CASSANDRA-10247