Upgrade testem to latest

[NullVoxPopuli]

Upstream fix testem/testem#2013

[github-actions]

📊 Size report

Tarball size — 1.2 MB → 1.2 MB

dist/dev   No changes

File	Before (Size / Brotli)	After (Size / Brotli)
./packages/shared-chunks/template-{hash}.js	1.1 kB / 401 B	-54%↓491 B / -49.4%↓203 B
Total (Includes all files)	2 MB / 478.1 kB	2 MB / 478.1 kB

---

dist/prod   No changes

File	Before (Size / Brotli)	After (Size / Brotli)
./packages/shared-chunks/api-{hash}.js	8.3 kB / 1.8 kB	-28.6%↓5.9 kB / -17.3%↓1.5 kB
Total (Includes all files)	1.8 MB / 436 kB	1.8 MB / 436 kB

---

smoke-tests/v2-app-hello-world-template/dist   No changes

File	Before (Size / Brotli)	After (Size / Brotli)
Total (Includes all files)	235 kB / 64.4 kB	235 kB / 64.4 kB

_{🤖 This report was automatically generated by wyvox/pkg-size}

[kategengler]

Serializing questions here for memory:

• Why was the testem issue not an issue on Chrome and why did it only show up on Firefox?
• How is this not an issue for the tutorial/new app generation that use testem 3.20?

[NullVoxPopuli-ai-agent]

Why was the testem issue not an issue on Chrome and why did it only show up on Firefox?

it isn't firefox-only -- the renovate bump #21394 hits the exact same Browser failed to connect within 120s. testem.js not loaded? on Basic Test (Chrome). firefox just looks exclusive because browser-test has needs: [basic-test, lint, types] and skips when chrome already failed.

both browsers fetch qunit (and friends) at /assets/node_modules/.pnpm/qunit@2.25.0/... -- that's coming from preserveModules: true in our vite.config.mjs on main, which mirrors the pnpm layout into dist/index.html <script> tags. testem 3.20's send@1.x 500s on any dot-prefixed intermediate segment (testem/testem#2012, fix in testem/testem#2013).

How is this not an issue for the tutorial/new app generation that use testem 3.20?

neither of them sets preserveModules. super-rentals' vite.config.mjs is the embroider defaults with no rollup overrides, and its test script is just vite build --mode development && ember test --path dist; same shape in ember-new-output. bundled output is /assets/main-*.js / /assets/app-*.js -- no dot-prefixed segments hit testem's static server, so the dotfiles: 'ignore' default never matches. we're the odd one out because the ember-source test build keeps modules unbundled.

[NullVoxPopuli]

Over here I'm testing that hypothesis via setting preserveModules to false on the renovate upgrade PR: #21394

[NullVoxPopuli-ai-agent]

ran it locally to confirm -- flip preserveModules to false and drop the patch, and testem 3.20 connects fine.

dist/index.html becomes the plain bundled shape:

<script type="module" crossorigin src="/assets/index-BW7JjjBU.js"></script>
<link rel="stylesheet" crossorigin href="/assets/index-K4R3Fv9v.css">

no dot-prefixed segments, so send@1.x never rejects. pnpm test runs to completion: # tests 9325 / # pass 9305 / # skip 17 / # fail 3.

the 3 fails aren't testem -- they're (eval) runtime-template-compiler tests (e.g. keyword helper: array (runtime): implicit scope, keyword helper: element (runtime): implicit scope (eval)) that need @ember/component / @glimmer/runtime resolvable as separate modules at runtime. rollup warns about them at build time:

packages/@ember/component/index.ts is dynamically imported by ... runtime-template-compiler-explicit-test.ts ... but also statically imported by ... dynamic import will not move module into another chunk.

so the real tradeoff: keep preserveModules: true + the patch (this PR), or flip to bundled and lose runtime-template-compiler test coverage. patch is the smaller blast radius.

[NullVoxPopuli]

we can't turn off prserveModules because then some of our runtime compiler tests fail