Batch containing delete and insert leads to inconsistent results

On a single node cluster (i.e. ./bin/cassandra -f on localhost) we ran into the following. Let's consider empty keyspace with the following table:

CREATE TABLE test (
    a varchar,
    b varchar,
    PRIMARY KEY (a, b)
) WITH comment='List of a related to b - widerow';

The table is empty.

Now we issue the following batch:

BEGIN BATCH
DELETE FROM test WHERE a = 'a1' AND b = 'b1';
INSERT INTO test (a, b) VALUES ('a1', 'b1');
APPLY BATCH;

When the batch successfully finishes, the table is empty.

This is consequence of the fact tombstone wins if timestamps are the same. And they are, because the operation is batched.

I consider this a bug. Batching operations shouldn't change the semantics of batched operations.