Diagnosis from Codex (GPT-5.5 model):

• Diagnosed. The mismatch is real and is caused by Ruff isort’s package classification

depending on whether opentelemetry/__init__.py is present in the same sandbox invocation.

Key facts:

- ./pants lint src/python/pants/backend/observability/opentelemetry runs ruff check on all

13 files, including opentelemetry/__init__.py, and reports two fixable I001 issues.

- ./pants fix ... splits the files into two fix batches: init.py alone, then the other

12 files. The 12-file Ruff check --fix sandbox lacks opentelemetry/__init__.py, so Ruff no

longer treats opentelemetry as a local package and says “All checks passed!”

- Running Ruff manually in the lint sandbox with --fix fixes both issues. Running it without

opentelemetry/__init__.py reproduces the false “All checks passed!” behavior.

So yes: namespace/package shadowing around the new local

pants.backend.observability.opentelemetry package is the trigger, but the Pants-side bug is

that fix batches Ruff with less package context than lint.

Likely fix direction: ensure Ruff fix processes get the same package-context inputs as lint,

or at least include package marker files like init.py in the sandbox even if they are

not in the file batch being edited. The current by-file fix batching is too context-starved

for Ruff isort.

draft PR: https://github.com/pantsbuild/pants/pull/23298

with the PR in place,

./pants fix src/python/pants/backend/observability/opentelemetry::

correctly fixes the import sorting