Summary

The coverage job on main has never reported anything. Its log:

+ python -Iim coverage combine
No data to combine
+ python -Iim coverage report --fail-under=80
No data to report.

Two separate bugs in jupyterlab/maintainer-tools/.github/actions/report-coverage@v1:

1. It merges the matrix artifacts server-side (actions/upload-artifact/merge) but never downloads them into the workspace, so every coverage command ran against an empty directory.
2. python -Iim includes -i. Each command exits non-zero, Python then drops into an interactive interpreter, reads EOF from the empty stdin, and exits 0 — so the job went green and the --fail-under=80 gate never fired once.

This PR replaces that job with per-matrix-job uploads to Codecov, adds Codecov components, and fills in the library's largest coverage gaps.

CI wiring

• .github/workflows/tests.yml — each matrix job uploads its own coverage.xml via codecov/codecov-action@v5, tagged flags: <os>-<python-version>. Codecov merges the uploads for the commit, so the artifact round-trip and the whole coverage job are gone.
• codecov.yml — new. project: auto with a 0.5% threshold, patch: 80%, comment only when coverage changes.
• pyproject.toml — branch = true moved into [tool.coverage.run] (it was --cov-branch on the hatch script only, so any bare coverage invocation measured something different), and --cov-report xml added to cov:test.
• README.md — Codecov badge. .gitignore — coverage.xml.

Coverage

manager.py and sphinxdoc.py had no tests at all — 117 statements. The Sphinx extension imports no Sphinx, so a stub with an add_object_type method exercises setup().

application.py's ~113 uncovered lines were nearly all the help machinery, and most of it was not untested: test_help_output and friends go through check_help_output, which spawns a subprocess that coverage doesn't follow. The lines ran; nothing recorded it. The new tests call emit_alias_help, emit_flag_help, emit_options_help, emit_subcommands_help, emit_help, emit_description, emit_examples, the print_* wrappers, start_show_config, load_config_environ, boolean_flag, get_config and launch_instance in process — which measures them, and pins the generator API downstream applications build on. What still isn't covered there is error paths that log and re-raise.

93 new tests, 675 total, passing on both current pytest and the pinned minimum (7.0.1).

tests/test_typing.py is omitted from measurement: mypy type-checks it and pytest-mypy-testing never executes the bodies, so all 454 lines read as uncovered. It's a fixture for the type checker, not dead code.

Components

component_management defines traitlets (traitlets/**) and tests (tests/**), each with its own project status. A component can only report on files in the report, so coverage measures the test suite too — a test file with uncovered lines is a test that never runs. That blends the headline number upward, which is why the split matters: traitlets is the one to watch.

Coverage is measured from the repository root rather than source = ["traitlets", "tests"]. That's deliberate: coverage xml writes each path relative to its source root, so with two roots both config/__init__.py files land in the XML under the same name. They're empty today and nothing collides — the first line of code added to tests/config/__init__.py would silently merge the two on Codecov's side.

Verified, not assumed

• Codecov's API confirms the per-job uploads merge: state: complete, 22 sessions, 40 files on an earlier revision. The 40 vs 43 files measured locally is exactly the three empty __init__.py files under tests/, which coverage xml omits as zero-statement.
• coverage xml normalises the filename attribute to forward slashes on every platform (coverage/xmlreport.py:190,207), which is why the Windows uploads line up with the rest.
• codecov.yml validates against codecov.io/validate, components included.
• Rebased onto 0ef86a9.

Notes for review

• Codecov reads a couple of points below the CI number — it counts partial branches as uncovered where coverage.py gives partial credit.
• CODECOV_TOKEN is not set. Uploads take the tokenless path, which works but is rate-limited and less reliable on fork PRs. Worth adding, then flipping fail_ci_if_error: true.
• after_n_builds is deliberately absent. With one upload per matrix job it would have to equal the matrix size, and would silently report partial coverage the first time someone adds a Python version without bumping it. wait_for_ci (default true) already holds the comment and statuses until CI finishes.
• The --fail-under gate is gone with the combine job; Codecov's project status replaces it.
• The codecov/project checks compare against a November 2022 baseline (1077cfb), the last commit on main carrying a report. That comparison is measuring two different things and resolves itself once this lands and main gets a fresh report. Details in the thread below.
• If coverage was a required status check in branch protection, the check no longer exists and will need removing — I can't read branch protection settings from here.