If you think about it, the auto-increment number should not be transactional. If other transactions had to wait to see if the auto-number was going to be used or “rolled back”, they would be blocked by the existing transaction using the auto-number. For example, consider my psuedo code below with table A using an auto-number field for the ID column:
User 1
------------
begin transaction
insert into A ...
insert into B ...
update C ...
insert into D ...
commit
User 2
-----------
begin transaction
insert into A ...
insert into B ...
commit
If user 2’s transaction starts a millisecond after user 1’s, then their insert into table A would have to wait for user 1’s entire transaction to complete just to see if the auto-number from the first insert into A was used.
This is a feature, not a bug. I would recommend using another scheme to generate auto-numbers if you need them to be tightly sequential.