Another solution would be to create the symlinks in the test code. This is simpler, but should we add a note in the test somewhere stating that these test files are not present in the sdist?

I prefer this, both because it's simpler and because I regularly use test packages directly rather than via a pytest invocation, so I prefer it if they were complete without dynamic additions.

Added the note in the test case itself, that is indeed useful.

CI failure is unrelated, looks like a change in Meson master:

FAILED tests/test_editable.py::test_mesonpy_meta_finder - AssertionError: assert 'D:/a/meson-python/meson-python/tests/packages/complex/complex\\__init__.py' == 'D:\\a\\meson-python\\meson-python\\tests\\packages\\complex\\complex\\__init__.py'
  
  - D:\a\meson-python\meson-python\tests\packages\complex\complex\__init__.py
  ?   ^ ^            ^            ^     ^        ^       ^
  + D:/a/meson-python/meson-python/tests/packages/complex/complex\__init__.py
  ?   ^ ^            ^            ^     ^        ^       ^
FAILED tests/test_editable.py::test_importlib_resources - AssertionError: assert 'D:/a/meson-python/meson-python/tests/packages/simple/__init__.py' == 'D:\\a\\meson-python\\meson-python\\tests\\packages\\simple\\__init__.py'

mesonbuild/meson#12194 probably.

mesonbuild/meson#12194 probably

Most likely. That patch changes the format of the introspection data used by meson-python. So far meson-python assumes that the introspection data uses native path separators. Now, some data uses native path separators, and some uses POSIX path separators independently of the platform. For supporting Meson version before and after the change, meson-python will need to normalize all paths read from introspection data.

I opened gh-778 to keep track of the new failure. This PR is ready to merge I'd say.

I moved the note in a place where it makes more sense.