test(ui): tighten connection-indicator regression tests (follow-up to #273)

[sanity]

Follow-ups from the multi-model review of PR #273 (Bug #5: WebSocket connection indicator on no-rooms screen). All Important / Suggestion level — non-blocking, but worth picking up.

Items

Important

1. Assert exactly ONE pill is visible at each viewport. Both placements (RoomList always-visible + Conversation no-room inline) share data-testid="connection-status-indicator". The current spec uses :visible + .first(), which masks the regression where md:hidden accidentally inverts (or the Tailwind v4 @source "../src/**/*.rs" directive gets dropped — see AGENTS.md). Add await expect(page.locator(PILL_TESTID)).toHaveCount(1) at desktop (1280px) AND mobile (375x812).

2. State-transition coverage. All four current tests assert "Connecting..." at initial paint. A regression that replaces try_read() with peek() (no subscription at all per Dioxus 0.7.x) would leave the pill stuck and not be caught. Either assert the dot's bg-color class so a stuck-state regression is detectable, or drive a state transition via test fixture.

Suggestion

3. Tighten waitForApp selector — expect(page.locator(\"aside, .app-root button\")).not.toHaveCount(0) matches any button. Should wait for the rail itself.

4. Run the new spec across webkit and mobile-safari projects too — Firefox-mobile is where the try_read() rule originated.

5. Consider a data-testid=\"rooms-rail\" instead of hasText: 'Rooms' for the rail-membership assertion — text changes break it brittle-ly.

Context

• Original bug report: Ivvor on Matrix 2026-05-17 — "I had no connection indicator in the bottom right to signify my setup wasn't fully connected to the node."
• Fixed in PR fix(ui): show WebSocket connection indicator on no-rooms screen (Bug #5) #273 (merged 2026-05-17).
• These follow-ups don't gate behavior, just harden the regression test surface.

[AI-assisted - Claude]