[HugoFara]

Every PR waited about 11 minutes. Nearly all of it was four Windows PHPUnit jobs: 9 minutes each, against 17 seconds on Ubuntu for the same 9,245 tests. Everything then ran a second time on merge into develop.

Changes

• Windows on main only. PRs run PHPUnit on Ubuntu with all four PHP versions. A push to main adds Windows.
• No more push runs on develop. PRs and main remain.
• Path filters on PRs. Each workflow lists the paths that can affect it. Pushes to main have no filter, so it stays the full check before a release.
• Superseded runs are cancelled. A new push to a PR replaces its running CI. Runs on main always finish.
• Xdebug off in PHPUnit. The job runs test:no-coverage, so coverage: xdebug only slowed it down. This may also shrink the Windows time; the first run on main will show.

Docs: CONTRIBUTING.md and its copy no longer say the locale check runs everywhere, and CLAUDE.md explains the setup, including adding a path to a workflow when a new test reads a file outside its list.

Choosing the filters

Two are deliberately broad, so a cross-check can't be skipped by mistake:

• PHP tests use all of src/** and locale/**, because guard tests such as PluralKeysExistTest scan TypeScript and locale files. tests/fixtures/** is shared with the frontend tests.
• Frontend uses all of src/**, because the CSS purge step scans PHP views. cypress/** is included because eslint . lints it.

Psalm uses **.php because it scans the whole repo, not just src/.

Trade-offs

• A Windows-only bug now surfaces on main, at release time.
• Two PRs that pass separately but clash together are only caught at main.
• No ruleset or branch protection requires specific checks, so skipped workflows can't block a merge. If required checks are added later, skipped ones would need handling.

Verification

This PR touches all five workflows, so it runs all of them: expect 10 checks and no Windows job. The filters themselves are confirmed by the next docs-only PR. All workflow files parse as YAML.

🤖 Generated with Claude Code