chore(frontend): upgrade Vitest, Playwright, jsdom, Prettier

[0xMMA]

Refs #35 — E4 step 3. frontend/ devDependencies only; Angular, PrimeNG and TypeScript stay put and wait for #31.

	from	to	note
@playwright/test	1.58.2	1.63.0
jsdom	28.1.0	30.1.0	two majors
prettier	3.8.1	3.9.8
vitest	4.0.8	4.1.11	not 5 — see below

Vitest 5 was tried and rejected, on evidence

The brief asked for 4 → 5. It is not deliverable: @angular/build:unit-test declares vitest as an optional peer at ^4.0.8 — on every published 21.2.x and on 22.1.8, the newest release.

I shipped 5 anyway in the first version of this PR, on evidence that turned out to be about the wrong thing. The test run was fine — clean npm ci, 198 tests, a deliberately broken assertion still failing, and the builder's watch/rerun path working. What it missed is what developers actually do:

$ npm install --dry-run
npm error code ERESOLVE
npm error Found: vitest@5.0.1 … peerOptional vitest@"^4.0.8" from @angular/build@21.2.0

Not a warning — an error with a non-zero exit, for every package anyone tries to add, until they know to pass --legacy-peer-deps. Every CI workflow uses npm ci, which is unaffected, so a green pipeline was never evidence about this. A test-runner major is not worth breaking local installs for.

So the bump stops at 4.1.11: inside the declared range, npm install --dry-run clean, 198 tests green, and clearMocks still defaults to false as on 4.0.8 — no silent semantic change.

The conditions for moving to 5 are recorded in .claude/rules/testing.md rather than lost: the builder has to declare ^5, and the same change has to handle vitest 5 flipping clearMocks to true by default. The builder does not set it, so a mock's calls stop surviving into the next test — an assertion on call counts that should fail could start passing.

Spec files are not isolated — and the rules said the opposite

The first version of this PR added a claim that each spec file gets its own environment. It is false. The builder sets isolate: false, so files in one fork share a process, a globalThis and one jsdom window.

Measured: 11 of 14 files landed in a single fork and saw ResizeObserver and matchMedia already defined on entry, set by two files that ran before them. Which fork a file lands in changes between runs.

That also explains a leak I had reported as "observed once, did not reproduce in ten runs" — it reproduces perfectly; it depends on scheduling. The advice (every file sets its own stub) was right and the reason was wrong, which is the more damaging half: it told the next author that cross-file state cannot happen. .claude/rules/testing.md now describes the mechanism and says to run a new or changed spec on its own once, with the command.

All three load-bearing stubs were checked by removing them: settings.component.spec.ts (56 tests fail), text-enhancement.component.spec.ts (36 fail), cli-probe-startup.spec.ts (3 fail). All 14 files also pass individually, so nothing in the suite currently depends on a leaked global.

jsdom 30 changed nothing we depend on

matchMedia, ResizeObserver and IntersectionObserver are undefined in jsdom 30 — against the library and in a spec run on its own. In a full run a spec may see them anyway, per the section above; the rules file now says so, because "green suite" is not proof that your stub is there.

Prettier 3.9 reformats nothing here

3.8 and 3.9 report the same files under --list-different, and the diff of the two lists is empty. The repo is not Prettier-formatted and Prettier is not enforced anywhere — no format script, no CI step, no hook. No formatting commit, no .prettierrc pin.

Playwright 1.63

Every config key still applies, and each was confirmed by its effect rather than by reading the types: testDir (61 tests in 5 files, all from e2e/), projects ([chromium] prefix), timeout (a hanging test reports "Test timeout of 30000ms exceeded"), retries (CI=1 produces "(retry #1)" and a -retry1/ artefact dir), reporter (list output + playwright-report/, no browser opened), use.baseURL, use.screenshot (attachment on failure), use.trace (trace.zip, only on failure), and webServer.env (probe config whose server echoes the injected value).

angular.json's test target keys are all in @angular/build@21.2.0's schema. include: ["src/**/*.spec.ts"] is load-bearing: the schema default is ["**/*.spec.ts", "**/*.test.ts"] relative to the project root, which would reach the Playwright specs in e2e/.

58 passed, 3 skipped. The suite is flaky in one place — router-leave-settings fails in roughly one full run in three — and that is not from this bump: measured 3 failures in 7 runs on 1.63.0 against 1 in 4 on 1.58.2, same test, same cause. Filed with the one-line fix as #85.

Also filed

• fix(e2e): router-leave-settings races the PrimeNG tabs render #85 — the e2e flake: locator.count() taken before the tabs render.
• chore(e2e): playwright.config.ts injects NG_APP_ANTHROPIC_API_KEY, which nothing reads #86 — playwright.config.ts injects NG_APP_ANTHROPIC_API_KEY, and nothing reads it; the comment describes an @ngx-env/builder feature this project does not use.

Both predate this PR.

Verification

• npm install --dry-run — clean, exit 0 (the check the first version failed)
• npm ci from a deleted node_modules — exit 0
• npm test — 14 files, 198 tests, 0 failures; all 14 files also pass individually
• npm run build — clean
• npx playwright test — 58 passed, 3 skipped
• npx tsc -p tsconfig.spec.json --noEmit — exit 0
• git status clean after the E2E runs

🤖 Generated with Claude Code