Test case:

recreate table t (col1 int, col2 blob);

insert into t values (1, '1');
insert into t values (2, '2');
insert into t values (3, '3');
commit;

select * from t;

        COL1 COL2
============ =================
           1 84:0
           2 84:1
           3 84:2

update t set col1 = -col1;
select * from t;

        COL1 COL2
============ =================
          -1 84:0
          -2 84:1
          -3 84:2

-- So far so good: COL1 is negated, COL2 has the same BLOB IDs

rollback;

alter table t add col3 date;

select * from t;

        COL1 COL2 COL3
============ ================= ===========
           1 84:0 <null>
           2 84:1 <null>
           3 84:2 <null>

update t set col1 = -col1;
select * from t;

        COL1 COL2 COL3
============ ================= ===========
          -1 84:3 <null>
          -2 84:4 <null>
          -3 84:5 <null>

-- BUG: COL2 has different BLOB IDs, i.e. the whole blobs were copied rather than just their IDs

rollback;

This is a regression introduced while fixing CORE-5600. v3.0.3 and later versions are affected.