I haven't read the code, but the associated document explaining the method makes no mention of accounting for line renumbering as a result of changes.
The accuracy with which changes are identified depends on the diff algorithm. I imagine it is more likely to generate false positives than false negatives (which is preferable), but I am sure there are cases where a line is unchanged textually but is different semantically (e.g. by the insertion or removal of an 'else' above it), and the diff program will consider it unchanged.
In my prototype this caused effectively any change outside of a function body to end up disabling the optimisation entirely :(
Tests are there to fight our natural tendency toward wishful thinking. Tests that fail for no reason enable people to keep thinking there was nothing wrong with their code. Meanwhile other people are twiddling waiting for them to fix something they don't think they broke.
> Test selection algorithm is not perfect yet (it may produce false-positives and false-negatives)
The problem is you're going to need to run the entire test suite before deploying anyway unless the algorithm is perfected. If you have to do this anyway, while it's helpful to have more confidence going into an hour long test build that it's going to pass, it's probably more helpful to just build an infrastructure on which the entire test suite runs more quickly. I've worked on sites with extremely large test suites and it was never an issue to run the tests in < 10 minutes because they could be run on an infrastructure that allowed tests to be run in parallel. I think that's the solution rather than a complex and flakey algorithm.