[Remove] site registry formatter alias - #438

## Why

The regular journeys use local fixtures, so they cannot catch failures
in a real WordPress download, dependency installation or full build.
Automate that occasional manual check without adding network setup time
to every PR.

## What changes

- Add a manual-only GitHub Actions workflow with macOS, Windows or both
as the platform choice.
- Drive the source app through site creation, its automatic
clone/install/build chain, and dev-server startup; verify the WordPress
login form over HTTP.
- Reuse the isolated Electron session harness, keep logs and reports for
seven days, and exclude generated Playwright output from ESLint.
- Use a separate Playwright config and explicit environment opt-in. No
push, PR or scheduled trigger; no automatic retries or packaged-app
changes.

## How to test this

**Platforms:** macOS and Windows. This changes test tooling only, with
no application UI changes.

**Starting state:** this branch checked out with its own dependencies
installed (`npm ci`) and renderer built (`npm run build:once`). Allow
network access, several GB of disk and potentially tens of minutes.

1. Run `npx --no-install playwright test --list`: the real-setup test
must not appear. Run `npx --no-install playwright test --config
tests/e2e/real-setup.config.js` without the opt-in: it must report one
skipped test.
2. On macOS run `TOOLKIT_REAL_SETUP=1 npx --no-install playwright test
--config tests/e2e/real-setup.config.js`. On PowerShell first run
`$env:TOOLKIT_REAL_SETUP = '1'`, then the same Playwright command, and
remove the variable afterwards with `Remove-Item
Env:TOOLKIT_REAL_SETUP`.
3. The app opens with an empty temporary profile, uses **Create
WordPress Core site**, and automatically downloads, installs and builds.
The test then starts the dev server, verifies HTTP 200 and the WordPress
login form, stops the processes and closes the app. Expect one passing
test and an HTML report under `playwright-report/real-setup`.
4. Once the workflow is on `trunk`, open **Actions → Real WordPress
setup → Run workflow**, select the branch and platform, and run it.
Expect one job per selected platform and downloadable
`playwright-real-setup-<platform>` artifacts.

**What must not have happened:** existing sites or the normal site
registry changed; the real-setup job started on a push or PR; the
network test joined the regular test suites.

## Risks and limitations

The real source-app flow passed locally on macOS in 51.2 seconds; this
is not a CI runtime estimate. Windows and hosted workflow execution
remain unverified. The workflow becomes manually discoverable after it
reaches the default branch.

Upstream WordPress, npm and network failures can fail the run
independently of Toolkit. The test is limited to 45 minutes and the job
to 60 minutes, with no retries. Cleanup is best-effort through the
existing fixture; forced cancellation can interrupt it. This does not
exercise the signed installer or native folder chooser, which the
harness answers automatically.

## Related

Follow-up to the audit's real-setup coverage gap and the test guardrails
in #443.

---

<details>
<summary>Review outcome (required — see AGENTS.md)</summary>

0 [fix here] · 0 [follow-up]. No findings across architecture, security,
performance, cross-platform and tests.

- **Review:** completed — fresh agent `review_manual_setup`, following
the repository self-review skill and canonical review standard. Reviewed
HEAD `c1533bde92f403ecfd7f097974462baf4065a848` plus all five
uncommitted files, against confirmed no-PR target `trunk` at
`d06630e471939877201a8941c18548f08705b683`. Inspected relevant harness
and application code. The intervening base commit only removes an unused
external-URL predicate and does not interfere with this change.
- **Since review:** committed the reviewed work unchanged as
`ef87625a3ca3cbd52920a524f5e7fbda57342964`; verified the five-file,
171-line diff and clean working tree. No later code changes.
- **Validation:** `npm run lint`, all 1,251 `npm test` tests, `git diff
--check`, YAML parsing/manual-only trigger check, default/separate
Playwright discovery, explicit opt-out skip, and a real macOS setup/HTTP
pass. Windows and hosted Actions execution have not run.

- **Pre-merge update:** Codex inspected
`ef87625a3ca3cbd52920a524f5e7fbda57342964 →
dbfaef2` after merging current `trunk`
(`f7f6069d0e842342edb5d9f15df6d651994ffa44`) without conflicts. The
feature diff has the same stable patch ID before and after; the new base
changes only remove unused wrappers and consolidate/remove tests, with
no effect on the setup flow. No new findings. Lint and all 1,248 tests
pass on the updated head; the count changed because of those upstream
test cleanups. Hosted CI is rerunning for this head.
- **CodeRabbit:** review did not run due to its usage limit, despite its
successful status check. The completed independent review above is the
review evidence; the status is not counted as a review.

- **Further base updates:** Codex inspected the subsequent merges
through head `be75d73fb17359ed2029e7221a0c62c5047233b2`, base
`f85e6e28eb7cf0219b92b2ce5f6f5e73923862ed`. Upstream #438 removes a
redundant site-log formatter alias; #441 removes unused patch exports.
Neither changes this flow. The feature patch remains identical (stable
patch ID `657ce39a26f1baeb1f3fd28a681792fdeb6a7960`), with no conflicts
or new findings. Lint and all 1,245 current tests pass locally; hosted
checks rerun for this final head. The real-setup workflow itself still
awaits its first manual hosted run.

</details>