test(desktop): fix Escape-before-mount race in channel browser e2e

[morgmart]

Why

channel-browser.spec.ts:251 › canceling section create does not affect the next global create fails intermittently in CI — it's currently red on at least two unrelated branches (#2097 and cynthiac/profile-polish), and it never reproduces locally on a fast machine.

Root cause

The test clicks quick-create and sends Escape on the very next line:

await page.getByTestId(`section-actions-${CUSTOM_SECTION.id}-quick-create`).click();
await page.keyboard.press("Escape");
await expect(page.getByTestId("channel-browser-dialog")).not.toBeVisible();

On a slow runner the dialog hasn't mounted when Escape lands, so the keypress hits no handler and is dropped. not.toBeVisible() then passes vacuously — the dialog isn't merely hidden, it doesn't exist yet. The dialog mounts after the assertion, and its data-state="open" overlay intercepts pointer events for the rest of the test, timing out the next quick-create click at 30s.

That's why the CI log shows an overlay that is open rather than animating closed:

<div data-state="open" ... data-testid="dialog-overlay" ...> intercepts pointer events

A probe confirmed the sequence directly — at the instant the spec asserts "not visible," dialog-overlay count is 0, and the overlay appears immediately afterward.

Fix

Gate on the dialog actually being visible before dismissing it — the pattern this same file already uses at line ~498 — then wait for the overlay to detach so it can't outlive the content by one exit animation and swallow the next click.

Testing

A/B under 20× CDP CPU throttling (imitating the CI runner):

sequence	result
old (no mount gate)	fails — reproduces the exact CI error, dialog-overlay ... intercepts pointer events
new (this PR)	passes

Full spec channel-browser.spec.ts: 29/29 serial. Test-only change — no product code touched.