If a file has a removed line that looks vaguely like a file header (i.e. the

line is "-- XX (Y)"), the entry for the removed line in the diff would get

filtered out by our code that tries to mangle the diffs to make renames and

copies work properly.
This change makes the code that detects and mangles file headers much more

specific, looking at each chunk of 5 lines instead of lines one by one.

Created a reproduction case and verified that rbt diff and svn diff

  produced the same output.
Wrote a unit test and checked that the test failed before the change and

  succeeded afterward.
Ran entire test suite.