For me the optimal (current best thinking) way of doing continuous integration is to consider the "master" branch to be the integration branch. All tests should

For me the optimal (current best thinking) way of doing continuous integration is to consider the "master" branch to be the integration branch. All tests should always be working here. It should not be possible to deliver a feature branch which breaks the regression tests on master. Thus I don't like unnecessary merges of unfinished work, it increases the risk for breaking the tests, which will slow down development. If you rebase your feature branch continuously then you achieve what you are asking for without merging unfinished work.