Cut pytest-xdist per-worker import cost on Windows unit runs

[cboos]

Part of #248.

Problem

On Windows, pytest-xdist spawns each worker and re-imports the whole
module graph per worker (Linux forks and skips this). Measured:

• module import alone: ~12.9s serial
• a single trivial test: -n0 4.5s vs -n auto 18.6s → ~14s of pure
  worker-spawn + re-import (12 workers)

Two heavy dependencies were imported on every unit worker even though
unit tests never use them:

• pytest-playwright auto-loads via its entry point (~1s/worker)
• browser/TUI test modules import playwright/textual at module top
  level, and pytest imports every collected module before marker
  deselection (~0.8s/worker for textual)

Change

• test/conftest.py: a pytest_ignore_collect hook that skips
  *_browser.py / test_tui*.py modules only when they would be
  deselected anyway, evaluated via pytest's own marker Expression
  (so -m browser, -m tui, -m "not slow", and unmarked full runs all
  still collect them). Verified no unmarked test hides in those modules,
  so no coverage is lost.
• -p no:playwright on every unit-invoking command (justfile
test/test-all/test-cov + the CI unit step); browser/TUI steps are
  untouched so they still load the plugin.

Results (local Windows, 12-core)

• collect-only wall (best-of-5, -n auto): 4.2s → 3.1s (~26%)
• selection unchanged end-to-end (unit / -m browser / -m tui counts
  identical; browser + TUI runs still execute)

Worker count (deliberately not changed)

The spawn/import tax scales with worker count, but the right fix is to
make each worker cheaper, not to cap workers, so -n auto is kept. The
optimal CI worker count depends on the runner (GitHub windows-latest
has ~4 vCPUs vs a 12-core dev box), so no CI -n value is hard-committed
here — it should be A/B'd on the real runner. Full measurements in
work/xdist-import-cost.md.

Note: the collect-skip matcher keys off the filename pattern; a one-line
comment near it documents that coupling (a browser-marked test in a
non-matching filename stays collected on unit runs — harmless, just a
missed optimization).

Summary by CodeRabbit

• Tests
  • Reduced unit test collection overhead by skipping browser and TUI tests earlier when they aren’t selected.
  • Disabled an unused browser testing plugin during unit test runs, keeping the same test coverage and filters.
• Chores
  • Updated test commands and CI test steps to use the faster unit-test setup consistently.
  • Added notes documenting the test runtime improvement approach.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro

Run ID: 5ee1ce0e-88a4-4619-893e-ad43966807e3

📥 Commits

Reviewing files that changed from the base of the PR and between 4bd633f and 78b17f0.

📒 Files selected for processing (4)

• .github/workflows/ci.yml
• justfile
• test/conftest.py
• work/xdist-import-cost.md

---

📝 Walkthrough

Walkthrough

Two optimizations reduce pytest startup and collection overhead for unit test runs: a new pytest_ignore_collect hook in conftest.py skips browser and TUI modules when their markers are excluded by the active -m expression, and -p no:playwright is added to all unit test invocations in justfile and the CI workflow.

Changes

Pytest collection and plugin optimizations

Layer / File(s)	Summary
pytest_ignore_collect hook test/conftest.py	Adds Optional import and a new hook that infers a module's marker from its filename, evaluates the active -m expression, and returns True to skip the module when the marker cannot be selected.
-p no:playwright flag in all unit test invocations justfile, .github/workflows/ci.yml, work/xdist-import-cost.md	Adds -p no:playwright to the test, test-all, and test-cov targets in justfile and to the CI coverage step; includes a WIP investigation note documenting the root cause and measurements.

Estimated code review effort

🎯 2 (Simple) | ⏱️ ~10 minutes

Poem

🐇 No playwright here, no browser to load,
Just markers and hooks along the fast road.
We skip what we must, collect only true,
And Windows xdist runs faster now too!
Hip hop hooray, the imports are few! 🎉

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly summarizes the main change: reducing pytest-xdist per-worker import overhead on Windows unit test runs.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

✨ Finishing Touches

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch dev/xdist-import-cost

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}