Fix Ruff import sorting when fix batches omit package markers

[tdyas]

Problem

As observed with CI test failures in #23284, Ruff’s isort rules can classify imports differently depending on whether or not ancestor __init__.py files are present in the sandbox. The fix goal may split files into batches, which can leave a Ruff batch without the package marker files that were present during lint. In that case, pants lint can report a fixable I001, while pants fix reports "All checks passed!."

Solution

This change adds a custom partitioner to the Ruff fix rules to record the necessary __init__.py files as per-partition metadata. When the fix goal executes on a partition, it includes the __init__.py files set in the metadata back into the input digest so Ruff sees the correct package behavior.

Also adds a regression test covering a local package whose name shadows a third-party package.

AI Disclousre

Codex (GPT-5.5) was used to diagnose the problem observed in #23284, write a reproduction test, and write the fix with design guidance from me.

[sureshjoshi]

Bear with me, as I haven't had my coffee - but why do we have this files? It's an alias for snapshot.files in several of these cases.

So, it's files for the files we want fixed, while snapshot is files + init.py?

[tdyas]

Exactly. RunRuffRequest.files is the Python files to fix while snapshot contains those files plus any __init__.py which needed to be added back.

[tdyas]

In terms of naming, files is consistent with the AbstractFmtRequest.files property.

[sureshjoshi]

I have a vague recollection that there was an idea to not have ruff partitions, as it would be faster to just run it over the largest set of code, than to go through the overhead to run it otherwise.

However, I don't see how that would have worked in multi-py-version repos, as each has it's own python level.

[tdyas]

I have a vague recollection that there was an idea to not have ruff partitions, as it would be faster to just run it over the largest set of code, than to go through the overhead to run it otherwise.

However, I don't see how that would have worked in multi-py-version repos, as each has it's own python level.

I don't have any context on the Ruff design choices. The custom partitioner here is putting everything into a single partition. The fix goal, however, is capable of splitting partitions; for example, by size. The __init__.py files don't get special treatment in that case and could be separated from other Python files. This PR makes sure they get added back.