Fix NIC e2e test flake in safari (#3198)

Test failure in #3195 repros locally no problem:

<img width="747" height="409" alt="image"
src="https://github.com/user-attachments/assets/d1fed181-f26c-455b-a0ab-c8f8e962c1ae"
/>

Claude's speculation on why it suddenly became a problem after #3178 is
good enough for me. This is the second or third time lately we've run
into small issues because we're not awaiting
`queryClient.invalidateEndpoint` in the mutation handlers. Something to
keep an eye on.

> The race: `NetworkingTab` gates the Delete row action on `nic.primary
&& multipleNics`, where `multipleNics = nics.length > 1`. When the
`my-nic` delete mutation succeeds, `onSuccess` synchronously calls
`queryClient.invalidateEndpoint` (not awaited) and closes the confirm
modal. If the test opens nic-3's row-actions menu before the refetch
lands, `nics.length` is still 2, so Delete renders disabled and wrapped
in a `<Wrap with={<Tooltip/>}>`. When the refetch then lands mid-click,
the `<Wrap>` unwraps, the `Menu.Item` remounts (old DOM detached), and
on Safari/Firefox base-ui's Menu treats that focus/DOM churn as a close
— so Playwright's retry never resolves a clickable Delete and hits the
60s timeout.
> 
> When it was introduced: the disable rule landed April 2025 in #2806
("Disable primary NIC delete when multiple NICs present"). The
`DropdownMenu.Item` `<Wrap>`-with-Tooltip pattern has been in place
since ~March 2025. So the race has been latent for about a year.
> 
> Why it's surfacing now: first CI appearance was on the #3178
"Responsive grid layout" PR (2026-04-21), which landed on trunk
2026-04-22; every workflow run after that hits it. #3178 moved
TopBar/Sidebar from CSS-grid cells to fixed-position elements, switched
to document-level scroll, removed `overflow-y-hidden` on body, and set
`height: 100%` on html/body/#root. None of that touches the NIC flow,
but it changed paint/layout timing around modal open/close and table
reflow. That shifted the window such that `expect(my-nic
cell).toBeHidden()` now reliably resolves while the page is briefly
inert during the modal's close transition — before the refetch has
actually landed — so the test races into opening nic-3's menu in the
stale state. The layout PR didn't introduce the bug; it just removed the
timing cushion that had been hiding it.

Author: david-crespo

test/e2e/instance-networking.e2e.ts

@@ -100,7 +100,13 @@ test('Instance networking tab — NIC table', async ({ page }) => {
   await clickRowAction(page, 'my-nic', 'Delete')
   await expect(page.getByText('Are you sure you want to delete my-nic?')).toBeVisible()
   await page.getByRole('button', { name: 'Confirm' }).click()
-  await expect(page.getByRole('cell', { name: 'my-nic' })).toBeHidden()
+  // Wait for the NIC list refetch to land before opening nic-3's menu.
+  // Otherwise `multipleNics` is still true, Delete renders disabled (wrapped
+  // in a tooltip), and the unwrap when the refetch lands detaches the
+  // menuitem and closes the menu on Safari/FF. Row count (vs. checking my-nic
+  // is gone) because any absence check passes transiently while the confirm
+  // modal closes and the page is inert.
+  await expect(nicTable.getByRole('row')).toHaveCount(2) // header + nic-3
 
   // Now the primary NIC is deletable
   await clickRowAction(page, 'nic-3', 'Delete')