Check import order in CI and sort the files that had drifted

[jmchilton]

This PR description was drafted and posted by Claude (AI assistant) on jmchilton's behalf.

Summary

.isort.cfg and the make format target have configured isort in this repo all along — the settings are nearly identical to Galaxy's (profile=black, force_alphabetical_sort_within_sections, reverse_relative, force_grid_wrap=2; only line_length and the skip paths differ). Nothing ever ran it in CI, so 71 files had drifted out of order.

Two commits, and the second one is the boring one:

• Check import order in CI — the part worth reviewing. A format tox env running isort --check --diff ., plus a CI leg.
• Sort imports with isort — a mechanical isort . pass. No behavior change.

This came up because lib/galaxy/jobs/runners/htcondor.py failed Galaxy's isort check on galaxyproject/galaxy#23326, in an import block copied over from pulsar/managers/queued_htcondor.py. Galaxy caught it; Pulsar could not.

Notes for review

Why the format env runs on 3.14 only. The lint job matrix covers 3.7 and 3.14. isort dropped 3.7 support after 5.11, so the two legs would resolve different isort versions and could disagree about the same file. A single 3.14 leg avoids that; this also mirrors how Galaxy keeps format separate from lint.

Why it checks . rather than pulsar test. That matches what make format rewrites, so tools/ and docs/ stay sorted too and the check never disagrees with the fixer.

Dead config removed. setup.cfg carried:

import-order-style = smarkets
application-import-names = pulsar,harness,recorder

Those configure flake8-import-order, which is not in the lint deps, so they have never had any effect. They also describe a different convention from the isort profile actually in use — smarkets wants import X before from X import, which black-profile isort does not enforce. If anyone installed the plugin the two would fight, so they are better gone than latent.

Blast radius

isort --check . inspects the whole tree, so once this lands, open PRs branched from before it will fail the new check until they rebase. Worth timing accordingly.

Validation

• tox -e format passes on the sorted tree; deliberately unsorting one import block makes it exit 1, so the check demonstrably catches drift rather than just passing.
• tox -e test-unit equivalent: 241 passed, 1 failed — wsgi_app_test.py::test_standard_requests, which fails identically on unmodified master.
• flake8 and mypy clean (188 source files).