fix(native): toast coalescing, real reply target, Focus Assist mount query

- Rich toasts set a WinRT Tag (hash of the web notification tag, since
  room:thread tags exceed the 64-char limit) in a fixed Group, so a newer
  toast for the same room/thread replaces the older one instead of
  stacking; the replaced toast's keep-alive entry is dropped (#16).
- The bridge passes tag, roomId and threadId separately. The reply target
  is the real room id (+ thread), never the coalescing tag, which was
  `room:thread` for threads and `lotus-invites` for invites, so those
  replies failed silently. Toasts with no room id (invites) get no reply
  box (#17).
- New `get_focus_assist` command serves the latest Focus Assist poll so
  the web hook can hydrate on mount; the setup-time first emit was lost
  before the page listened (#15).
- Bump cinny to 3b6de2fd (the matching web-side hooks).

Closes #15, #16, #17

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PPmy3tPq869XDW4njjVaKA

src-tauri/src/native/focus_assist.rs

@@ -11,14 +11,33 @@
 //! ~5 seconds. We prefer a robust poll over hooking shell events — the poll is
 //! trivial to reason about and a 5s cadence is more than responsive enough for a
 //! notification-suppression hint. We emit **only on a boolean transition**, so the
-//! web side gets one event per change rather than a steady heartbeat; the first
-//! read always emits so the frontend learns the initial state.
+//! web side gets one event per change rather than a steady heartbeat. The latest
+//! reading is also kept in [`LAST_STATE`] and served by [`get_focus_assist`]: the
+//! first read happens during app setup, before the page has loaded, so that
+//! event is lost, and the web atom resets on every reload anyway. The web hook
+//! queries it on mount (Gitea cinny-desktop #15).
 //!
 //! Other platforms are a no-op: there's no equivalent cross-platform signal, and
 //! the web hook stays unconditional so nothing there needs guarding.
 
+use std::sync::atomic::{AtomicU8, Ordering};
 use tauri::AppHandle;
 
+/// Latest poll result: 0 = not read yet (or not Windows), 1 = inactive, 2 = active.
+static LAST_STATE: AtomicU8 = AtomicU8::new(0);
+
+/// Return the latest Focus Assist reading so the web side can hydrate
+/// `focusAssistActiveAtom` on mount. `None` until the first successful poll, and
+/// always `None` off Windows.
+#[tauri::command]
+pub fn get_focus_assist() -> Option<bool> {
+    match LAST_STATE.load(Ordering::Relaxed) {
+        1 => Some(false),
+        2 => Some(true),
+        _ => None,
+    }
+}
+
 /// Payload for the `focus-assist-changed` DOM event (`{ active: bool }`).
 #[cfg(target_os = "windows")]
 #[derive(serde::Serialize)]
@@ -81,6 +100,7 @@ fn watch_focus_assist(app: AppHandle) {
                 || state == QUNS_PRESENTATION_MODE
                 || state == QUNS_RUNNING_D3D_FULL_SCREEN
                 || state == QUNS_BUSY;
+            LAST_STATE.store(if active { 2 } else { 1 }, Ordering::Relaxed);
             if last != Some(active) {
                 last = Some(active);
                 super::emit_to_web(