[leracherry]

After an HMR update, navigation can briefly load the bundled-dev fallback page. A test that edits a file before the final page reconnects can lose the update.

Remove gotoLatestBuild and wait for a DOM marker set by the fixture after an HMR round trip. This lets the test wait for the page it will actually edit. The file-restoration test also waits for the parent update before the next navigation.

Refs #23028.

Rebased onto db8b119. The build and all three HMR specs passed locally: 44 tests in bundled dev and 63 in normal dev. The delayed-registration regression timed out without the readiness wait in the earlier validation. The previous CI run passed on Linux, macOS, and Windows.

[h-a-n-a]

Hmmm, this may not be the correct way to fix it. We should probably remove gotoLatestBuild completely. But thanks for the contribution though.

[leracherry]

Thanks for reviewing. I tested replacing all three gotoLatestBuild call sites with page.goto, keeping the regression test. That produced four bundled-dev failures: reconnect logs reached later assertions, and one readiness check finished before used:foo0 arrived.

Would you prefer replacing the helper with browser-observable readiness/reload checks, or addressing the navigation/registration behavior in production first? I restored the experiment so the current patch is unchanged while I clarify the direction.

[leracherry]

Reworked this to remove gotoLatestBuild entirely. The navigations now wait for a DOM marker set after a fixture-owned HMR round trip. This handles fallback reloads and registration without inspecting private server state.

All three HMR specs pass in bundled and normal dev. The delayed-registration regression fails with the readiness wait removed. The description has the current validation results. Could you review this revised approach?

[leracherry]

Following up on the revised approach: gotoLatestBuild has been removed and replaced with fixture-owned readiness checks, as described above. The latest CI run passes and the branch is conflict-free. Could you take another look when convenient?