Fix old_path and new_path being wrong for empty added or deleted files - #325

Fixes #248.

ModifiedFile.old_path returns the file's path instead of None when a commit adds an empty file, and ModifiedFile.new_path returns the path instead of None when a commit deletes one. The same object reports a change_type of ADD while also claiming a previous path, so any code that keys on old_path or new_path to detect creation or deletion silently mis-classifies the file. Empty __init__.py files make this common in Python repositories.

The cause is upstream. Git emits no hunk for an empty file, so there is no --- /dev/null or +++ /dev/null header for GitPython to read the missing side from, and _pick_best_path hands back the fallback taken from the diff --git a/X b/X line instead. The new file mode and deleted file mode lines are still emitted, which is why change_type stays correct while the paths do not. This still reproduces on GitPython 3.1.62, and gitpython-developers/GitPython#749 reports the same wrong a_path on added files and was closed as not planned.

_from_change_to_modification_type at pydriller/domain/commit.py:187 already reads diff.new_file and diff.deleted_file to decide ADD and DELETE, so having old_path and new_path consult the same two flags only makes them agree with the verdict the class already publishes, and with docs/modifiedfile.rst, which already documents both as None in these cases. The change is one guard at the top of each property in pydriller/domain/commit.py. Renames of empty files set neither flag, so renames are untouched. One downstream effect is worth naming: the seven per-file process metrics under pydriller/metrics/process/ (every one except ChangeSet) bucket by new_path, and ModifiedFile.__hash__ interpolates it too, so a deleted empty file now buckets under the None key and hashes as DELETE None .... That is not a regression, since every non-empty deletion already does both.

The bug is visible inside a bundled fixture, so no new test repository is needed. In test-repos/diff, commit 4be0402d adds the empty pydriller/tests/integration/__init__.py next to a non-empty setup.py, and only the empty one gets a bogus old_path. Commit df51c5ef deletes thirteen non-empty files, which all correctly get a new_path of None, plus that same empty file, which does not. The two new tests in tests/test_commit.py cover both cases and fail on master with assert 'pydriller/tests/integration/__init__.py' is None. Five existing tests needed two lines each: they build a ModifiedFile from a bare @patch('git.diff.Diff') mock where every MagicMock attribute is truthy, so new_file and deleted_file now have to be set false. Those mocks were always under-specified: on master they already make change_type report ADD for what the tests treat as an ordinary modification.

mypy --ignore-missing-imports pydriller/ tests/ and flake8 are clean, and the suite passes locally on Python 3.11 and macOS, 313 passed, with tests/test_memory_consumption.py excluded locally because it clones remote repositories.