[modacker]

Why the committed baselines were tofu boxes

#236 regenerated the 14 Linux screenshots on the GitHub-hosted ubuntu-latest
runner and merged them. That runner had no CJK font installed at all: the gate
ran playwright install chromium, which downloads a browser and nothing else.
The WebUI is Chinese, so every Chinese glyph in those baselines was drawn as an
empty box. The baselines matched what CI rendered, so the gate was green — but
nobody could read the Chinese in a diff, and the visual gate no longer proved
that Chinese text renders correctly. It only proved that boxes render
consistently.

What installs the fonts

playwright install --with-deps chromium additionally installs Playwright's own
system dependency set for the platform. For ubuntu24.04-x64 that set is the
tools list in packages/playwright-core/src/server/registry/nativeDeps.ts,
and it includes fonts-wqy-zenhei (WenQuanYi Zen Hei), plus fonts-unifont,
fonts-ipafont-gothic and fonts-liberation. The Chinese surfaces then render
real glyphs.

The command lives in the test:webui-browser npm script, not in a workflow
file. Both Source verification (.github/workflows/ci.yml) and
WebUI npm preview (.github/workflows/webui-npm-preview.yml) reach the
browser gate through pnpm verify --profile full -> test:webui-browser, so
the script is the single place that decides what the gate installs. Putting the
install step in one workflow would let the other path drift, and the npm
preview path is the one that publishes.

The change is Linux-only in practice: registry.installDeps has branches for
win32 and linux only, so on macOS and Windows the call returns without
installing anything. Local development on those platforms is unaffected.

Reproducibility of the regenerated baselines

Regenerated on ubuntu-latest with playwright install --with-deps chromium
and --update-snapshots, against the same commit.

On the branch's final base, two consecutive runs are byte-identical in all
14 images, and both are byte-identical to the committed files.

On an earlier base, three runs were compared instead, and they are identical
under Playwright's own comparison but not byte-for-byte:

comparison	raw byte-level differing pixels	Playwright comparison (pixelmatch, threshold 0.1)
run 1 vs run 2	1	0
run 1 vs run 3	9	0
run 2 vs run 3	10	0

Under Playwright's own comparison the runs are identical: zero differing pixels
in all 14 images, in every pair of runs. At the byte level two images differ
between runs, by one and by nine pixels of sub-threshold anti-aliasing jitter
(at most 2/765 on a channel). The gate's verdict therefore does not depend on
which run supplied the committed file.

Nothing was weakened

No assertion, expectation, pixel threshold, skip or retry was added, removed or
relaxed. The screenshot oracle is still the default toHaveScreenshot
comparison with the default zero-pixel tolerance. The 14 file names and the
file count are unchanged, so release/public-source.json and
test/vitest-suites.json are untouched. The DOM and geometry assertions in
visual-baselines.spec.mjs and plugin-catalogue-layout.spec.mjs are
unchanged, including the settle-wait that #236 brought in — it is kept as is.

The only difference the baselines now carry is that the Chinese is legible.

CI

verify (ubuntu-latest, 24): PASS, 14m25s. The browser gate reports
184 passed (4.0m) — 184 tests in 34 files, all of them compared against the
new baselines, none skipped, none retried. The test:webui gate before it
reports 213 files / 3301 tests passed. verify (macos-latest, 24),
verify (windows-latest, 24), build, performance and
source-history-artifact all pass.

The apt step is visible in the job log: Installing dependencies... at
06:09:33 and Setting up fonts-wqy-zenhei at 06:09:43, so the font install
costs about 10-25s. That is not separable from the runner's own variation:
across five Source verification runs on the base branch without this change
the same gate measured 191.6s, 251.0s, 255.3s, 255.8s and 289.1s, and the two
runs of this branch measured 210.1s and 264.0s. The whole verification stays
well inside the job's 25-minute ceiling (837.6s elapsed).

Remaining trade-offs

• The gate now depends on apt on the runner. The apt update/install is
  normally ~17s, but it can be slow when the runner's package mirror is
  contended: one measured run of the install step took 5m24s instead of 24s.
  The verify (ubuntu-latest, 24) job keeps its 25-minute ceiling.
• The baselines now record the runner's font set. A different font set or a
  different Chromium revision on the runner will turn the gate red. That is the
  intended behaviour — a font change is a rendering change — but it is a state
  the repository holds, not a guarantee that the gate stays green on an
  arbitrary runner.
• The screenshots remain the only visual oracle for these surfaces. A reviewer
  now has to trust that the fonts on the runner are the ones the baselines were
  rendered with; the regeneration procedure is documented in
  packages/webui/docs/verification/browser-suite.md.