fix: stop ruff format and ruff check --fix undoing each other

[realmarcin]

Closes #874.

tox -e format has been a silent no-op

poetry run tox is mandated before every commit and failed on a clean master. The cause was not what it looked like.

The format env runs ruff format then ruff check --fix. D211 (no-blank-line-before-class) is ignored, leaving its mutually exclusive opposite D203 (one-blank-line-before-class) active. So the formatter strips the blank line before a class docstring and the linter immediately restores it:

ruff format      -> 98 files changed
ruff check --fix -> 0 files changed    (all 98 reverted)

Net zero, exit 0. The env reports success while changing nothing, and ruff format --check has reported the same ~98 files unformatted indefinitely, because the fix can never stick. Ruff prints a warning naming this exact pairing on every run — it was just never acted on.

I got the diagnosis wrong first

I filed #874 blaming #862 for removing [tool.black] line-length = 100 without reformatting. The removal is real but is not the cause: the oscillation reproduces on 6be5f93^, the commit before #862 (91 files, net zero). It long predates it. Corrected on the issue.

What is #862's is the docstr-coverage half — five functions with no docstring.

The fix

D203 added to extend-ignore. With D211 already ignored, letting the formatter decide is the only self-consistent option. The 98-file reformatting here is the backlog that could never land while the two tools fought.

Verified the oscillation stops:

ruff format      -> 98 files changed
ruff check --fix -> 98 files changed   (retained)
ruff format --check -> 196 files already formatted

Plus docstrings for LocalToolkit.__init__, LegacyToolkit.__init__, guarded_getaddrinfo, guarded_connect and fail_edges — each saying what it is for rather than restating its name.

Result

format, lint, codespell and docstr-coverage all pass. Two test failures remain, neither from this change:

• test_every_referenced_curie_has_stub_node (PO:0009005) — long-standing, needs an ontologies_stubs re-run held for the MIM SSSOM
• StaleOutputGuardTest — my local gold output was built from fix(gold): bridge MeSH per term, recovering the sites #821 gave up #864's gold.py, so the content fingerprint correctly reports a mismatch. That is the guard working.

Mostly mechanical, but large (102 files). Worth landing on its own and ahead of #863/#864, both of which will need a trivial rebase after it.

🤖 Generated with Claude Code