[h-a-n-a]

Stop HMR propagation at modules that are no longer imported in the module graph.

A module that ran in the page, and whose importers all removed their import of it, is an orphan. Previously, if a module whose importer is an orphan and it's not a hot boundary, then it reloads the page because there's no boundary can be found.

This also fixes the flaky test case on main.

[h-a-n-a]

Why the tests were flaky

delete file should not break hmr changes two files together:

// index.js
import { displayCount } from './re-export.js' // → from './display.js'
// display.js
export const displayCount = (count) => `count is ${count}` // → `count is ${count}!`

The watcher does not always report the two changes together. Sometimes they arrive as one HMR update, sometimes as two. Rolldown's file debouncer sends each event on its own, and a poll scan can also run between the two writes.

• One update: the client walks the old graph, display.js → re-export.js → index.js. index.js accepts itself, so the update is applied, and the test passes.
• Two updates: update 1 makes index.js import display.js directly. Nothing imports re-export.js any more, but it is still listed as an importer of display.js. Update 2 (display.js) walks up to both importers: re-export.js has no importer, so the client reports no hmr boundary found and reloads the page. The count goes back to 1, and the test fails.

Locally, every failing run had the two changes split into two updates, and every passing run had them in one.

import.meta.hot?.accept failed because of the test before it. The delete test checked its last steps with toMatch, which also matches the old text ('count is 3' matches 'count is 3!'). So the test could end before its last change was handled, and that change then reached the next test.

[h-a-n-a]

• delete file should not break hmr: https://github.com/vitejs/vite/actions/runs/39244590997/job/118033392583
• import.meta.hot?.accept: https://github.com/vitejs/vite/actions/runs/39207790462/job/117914359589

[h-a-n-a]

I wrote this previously but it turned out to be a mistake. Previously, bundled-dev didn't keep the page as is and the reason it was kept is because of the one update bug I mentioned in the comment section above so it sometime passes the test case (i.e. keeping the page as-is).

[sapphi-red]

instead of using this.orphans, I wonder if we can change this to

if (isEntrypoint) { /* full-reload */ } else { /* no-op */ }

[h-a-n-a]

Yes! That's what I wanted to do. But this would cause a breaking change to the current Vite user as it requires Rolldown to expose the new entry data in runtime. Once we expose the entry data in next minor, we can update this line above: this.orphans.has(id) && !this.hasExecutedImporter(id) and the line you suggested as well.

[h-a-n-a]

Hmmm, maybe we can do this in the current Rolldown version, as exposing entry data is not a breaking change to the current Vite user and Vite user that uses the latest Vite which includes this PR can leverage the data from Rolldown.

[h-a-n-a]

I will start implementing the fix in Rolldown right now. But I'm not sure if we're going to hold this flaky issue until Rolldown gets released and updated.

[sapphi-red]

Ah, right, we need a new Rolldown patch release. Let's merge this one first 👍

[h-a-n-a]

I will get the Rolldown patch released before tomorrow's scheduled release so we can update the Vite implementation.