feat(build): avoid settling seen preload dependencies for performance

[StirStudios]

perf(build): avoid settling empty preload dependency results

The runtime preload helper maps every dependency result through its allSettled compatibility implementation. Already-seen dependencies, existing stylesheet links and newly inserted JavaScript preload links return undefined, so each still creates a resolved promise and settlement callback even though there is no pending work to await.

Filter only undefined entries before the settlement mapper. The preload loop still creates the same links. Pending CSS loads remain awaited, and the existing error event/cancellation behavior and browser compatibility implementation are preserved.

The build regression loads the CSS-bearing hello fixture twice and checks that these empty results are not passed to Promise.resolve. It fails against stock Vite 8.2.2 (three calls) and passes with this change.

Validation against current main (68aeb8a):

• Unit tests: 985 passed, 3 skipped.
• Selected preload, CSS splitting/dynamic import, CSP and chunk import map build-browser tests: 47 passed, 5 skipped.
• JS source-map build tests: 8 passed. Updated the preload snapshot for the helper length change; decoded mappings preserve all original source locations, with generated columns shifting by 22 characters.
• Vite build, type checking and targeted lint passed.
• Representative Nuxt production consumer: three interleaved mobile Lighthouse runs, median TBT 607 → 453 ms; LCP 2.04 → 1.99 s; 162 script requests in both variants. This is a local case study, not a general performance guarantee.
• CSS success/failure, canceled preload errors, duplicate dependencies, module rejection and CSP nonce behavior were also exercised in the diagnostic.

Related work checked: #19805 adds the compatibility implementation retained here; #19722 and #19757 address different preload injection/chunk-mapping concerns. Latest stable 8.2.2 and main were checked on September 6, 2026; no equivalent fix was found.

The source regression was first verified failing/passing against v8.2.2. The change was then applied to current main and the checks above rerun. Full cross-platform CI and unrelated browser playground suites were not run locally.

[sapphi-red]

• Representative Nuxt production consumer: three interleaved mobile Lighthouse runs, median TBT 607 → 453 ms; LCP 2.04 → 1.99 s; 162 script requests in both variants. This is a local case study, not a general performance guarantee.

Is there a project I can test this with?

[StirStudios]

Yes @sapphi-red — I've prepared a standalone standard Nuxt reproduction here:
https://github.com/StirStudios/vite/tree/repro/nuxt-preload-23446

It uses Nuxt 4.5.2 with 48 ordinary lazy Vue components and scoped CSS, without Drupal, our layer, custom chunking or artificial delays. stock/ and patched/ have identical app source and resolved package versions; the latter applies this change to Vite 8.2.2 through pnpm. The README includes build/run instructions, a browser regression test, and an optional interleaved Lighthouse runner.

The browser test observed 144 Promise.resolve(undefined) calls on stock versus 0 patched, with 52 script requests and 48 stylesheet links in both, and successful hydration/interactivity. The instrumentation is only used for that mechanism check, not the timing runs.

To distinguish this from the quoted production result: this is a reduced synthetic consumer, not the original site. Its three-run local median TBT was 90.15 → 80.90 ms, and LCP 660.72 → 651.29 ms (Lighthouse 13.4.1, applied mobile throttling). All six samples are included. This does not reproduce the original 154 ms reduction or establish a general performance guarantee; it does give you a runnable Nuxt example of the extra settlements. The existing Vite playground regression in this PR remains the focused correctness test.

[sapphi-red]

Thank you!

[sapphi-red]

/ecosystem-ci run

[pkg-pr-new]

Open in StackBlitz

@vitejs/plugin-legacy

pnpm add https://pkg.pr.new/@vitejs/plugin-legacy@23446 -D

npm i https://pkg.pr.new/@vitejs/plugin-legacy@23446 -D

yarn add https://pkg.pr.new/@vitejs/plugin-legacy@23446.tgz -D

vite

pnpm add https://pkg.pr.new/vite@23446 -D

npm i https://pkg.pr.new/vite@23446 -D

yarn add https://pkg.pr.new/vite@23446.tgz -D

commit: 6aa4fa1

[vite-ecosystem-ci]

📝 Ran ecosystem CI on abf4a55: Open

suite	result	latest scheduled
react-router	❌ failure	❌ failure
netlify-vite-plugin	❌ failure	❌ failure
qwik	❌ failure	❌ failure
sveltekit	❌ failure	❌ failure
storybook	❌ failure	✅ success
vitest	❌ failure	❌ failure
tanstack-start	⏹️ cancelled	⏹️ cancelled
vite-plugin-rsc	❌ failure	❌ failure

✅ module-federation, vite-environment-examples, analogjs, vike, waku, nitro, nuxt, vite-plugin-cloudflare, vite-plugin-vue, laravel, vinext, vite-plugin-react, quasar, marko, unocss, vite-plugin-pwa, vite-setup-catalogue, vite-plugin-svelte, vuepress, vitepress, astro