fix(venv): group venv prefixes by path component, not raw path (#3333)

When files overlap between packages, and dist-info directories are
present, it results in a prefix list like `foo foo-bar foo/bar`. When
sorted as raw strings, hyphen sorts before slash, so the continuity of
path prefixes is violated and they are grouped separately. An error then
occurs because both `foo/` and `foo/bar` are created, but the latter is
a sub-path of the former.

To fix, change the sort key to a tuple of path components. This makes
`foo foo-bar foo/bar` sort as `(foo,) (foo, bar), (foo-bar, )`,
resulting in the correct order.

Fixes #3204

Author: rickeylev

python/private/venv_runfiles.bzl

@@ -138,7 +138,9 @@ def _group_venv_path_entries(entries):
     """
 
     # Sort so order is top-down, ensuring grouping by short common prefix
-    entries = sorted(entries, key = lambda e: e.venv_path)
+    # Split it into path components so `foo foo-bar foo/bar` sorts as
+    # `foo foo/bar foo-bar`
+    entries = sorted(entries, key = lambda e: tuple(e.venv_path.split("/")))
 
     groups = []
     current_group = None

tests/venv_site_packages_libs/app_files_building/app_files_building_tests.bzl

@@ -61,6 +61,7 @@ _tests.append(_test_conflict_merging)
 def _test_conflict_merging_impl(env, _):
     entries = [
         _entry("a", "+pypi_a/site-packages/a", ["a.txt"]),
+        _entry("a-1.0.dist-info", "+pypi_a/site-packages/a-1.0.dist-info", ["METADATA"]),
         _entry("a/b", "+pypi_a_b/site-packages/a/b", ["b.txt"]),
         _entry("x", "_main/src/x", ["x.txt"]),
         _entry("x/p", "_main/src-dev/x/p", ["p.txt"]),
@@ -72,6 +73,7 @@ def _test_conflict_merging_impl(env, _):
 
     actual = build_link_map(_ctx(), entries)
     expected_libs = {
+        "a-1.0.dist-info": "+pypi_a/site-packages/a-1.0.dist-info",
         "a/a.txt": _file("../+pypi_a/site-packages/a/a.txt"),
         "a/b/b.txt": _file("../+pypi_a_b/site-packages/a/b/b.txt"),
         "duplicate/d.py": _file("../+dupe_a/site-packages/duplicate/d.py"),