We need a way to test "unexpected failure" and "unexpected success" merges

[mhagger]

If the bisection heuristic fails, it can be that an autofill merge fails even though it was expected to succeed. We need a way to test such scenarios systematically. I think the easiest way would be to use an object that can mock a repository (maybe MergeState can be mocked?) and simulate arbitrary successes and failures.

Issue #3 is an example of a problem handling unexpected failures.