Fix feed OOM: cap follow feed at 30, WebKit software rendering, dedup notification fetches

- followNotes capped at 30 (was 80) — following feed was rendering 2.7x more
  notes than global, causing 4GB+ spike on media-heavy follow content
- fetchFollowFeed limit 80→30 to match
- WEBKIT_FORCE_SOFTWARE_RENDERING=1 replaces WEBKIT_DISABLE_COMPOSITING_MODE=1
  (compositing mode killed Wayland path → blank window on Hyprland)
- HardwareAccelerationPolicy::Never → OnDemand (Never also caused blank screen)
- set_enable_page_cache(false) — SPA never navigates, bfcache is pure waste
- Removed duplicate fetchNotifications calls on login (was firing 3x in 8s)
- First notification poll delayed 8s→90s to avoid competing with feed load
- Result: login 3600MB→453MB, following feed crash→737MB, plateau at ~950MB

src-tauri/Cargo.lock

@@ -5320,7 +5320,7 @@ checksum = "accd4ea62f7bb7a82fe23066fb0957d48ef677f6eeb8215f372f52e48bb32426"
 
 [[package]]
 name = "vega"
-version = "0.12.6"
+version = "0.12.7"
 dependencies = [
  "futures-util",
  "hex",

src-tauri/src/lib.rs

@@ -436,7 +436,7 @@ pub fn run() {
                 }
             }
 
-            // ── WebKit GPU workaround for Linux (webkit2gtk 2.50+ black screen) ──
+            // ── WebKit memory tuning for Linux (webkit2gtk) ──────────────────
             #[cfg(target_os = "linux")]
             {
                 let main_window = app.get_webview_window("main").unwrap();
@@ -444,15 +444,20 @@ pub fn run() {
                     use webkit2gtk::{CacheModel, SettingsExt, WebContextExt, WebViewExt};
                     let wv = webview.inner();
                     if let Some(settings) = wv.settings() {
+                        // OnDemand: use GPU if available, CPU fallback otherwise.
+                        // HardwareAccelerationPolicy::Never + WEBKIT_DISABLE_COMPOSITING_MODE=1
+                        // both kill the Wayland compositor path → blank window on Hyprland.
+                        // WEBKIT_FORCE_SOFTWARE_RENDERING=1 (set in main.rs) forces CPU
+                        // rasterization without disrupting the Wayland surface.
                         settings.set_hardware_acceleration_policy(
-                            webkit2gtk::HardwareAccelerationPolicy::Never,
+                            webkit2gtk::HardwareAccelerationPolicy::OnDemand,
                         );
+                        // Vega is a SPA — no back/forward navigation, page cache is pure waste.
+                        settings.set_enable_page_cache(false);
                     }
-                    // Minimize WebKit's in-memory content cache (decoded images, scripts, etc.)
-                    // Default is WebBrowser which caches aggressively. DocumentViewer is the
-                    // minimum: no back/forward page cache, smallest memory footprint.
-                    // This is safe for Vega — it's a single-page app, never navigates between pages.
                     if let Some(ctx) = wv.context() {
+                        // DocumentViewer: smallest in-memory content cache footprint.
+                        // No back/forward page cache, only the active document cached.
                         ctx.set_cache_model(CacheModel::DocumentViewer);
                     }
                 }).ok();

src-tauri/src/main.rs

@@ -7,12 +7,11 @@ fn main() {
     #[cfg(target_os = "linux")]
     {
         std::env::set_var("WEBKIT_DISABLE_DMABUF_RENDERER", "1");
-        // Required on Linux with large RAM/swap: WebKitGTK compositor pre-allocates
-        // ~25% of total virtual memory (RAM+swap) for its tile cache. On a 14GB RAM +
-        // 19GB swap system this is ~4 GB, filling all RAM and freezing the machine.
-        // Software rendering is slower but memory-safe. Fix: reduce swap or implement
-        // virtual scrolling (fewer compositor layers).
-        std::env::set_var("WEBKIT_DISABLE_COMPOSITING_MODE", "1");
+        // Force CPU-only rasterization while keeping the Wayland compositor path intact.
+        // WEBKIT_DISABLE_COMPOSITING_MODE=1 kills the GPU process but also breaks Wayland
+        // rendering (blank window on Hyprland). WEBKIT_FORCE_SOFTWARE_RENDERING=1 cuts GPU
+        // RAM without disrupting the Wayland surface — the right tradeoff on this machine.
+        std::env::set_var("WEBKIT_FORCE_SOFTWARE_RENDERING", "1");
     }
 
     vega_lib::run()

src/components/feed/Feed.tsx

@@ -89,7 +89,7 @@ export function Feed() {
     try {
       await ensureConnected();
       const events = await diagWrapFetch("follow_fetch", () => fetchFollowFeed(follows));
-      setFollowNotes(events);
+      setFollowNotes(events.slice(0, 30));
       const prev = useFeedStore.getState().lastUpdated;
       useFeedStore.setState({ lastUpdated: { ...prev, following: Date.now() } });
     } finally {

src/components/feed/NoteCard.tsx

@@ -71,7 +71,7 @@ export const NoteCard = memo(function NoteCard({ event, focused, onReplyInThread
     <article
       ref={cardRef}
       data-note-id={event.id}
-      className={`border-b border-border px-4 py-3 hover:bg-bg-hover transition-colors cursor-pointer group/card [content-visibility:auto] [contain-intrinsic-size:auto_120px]${focused ? " bg-accent/10 border-l-2 border-l-accent" : ""}`}
+      className={`border-b border-border px-4 py-3 hover:bg-bg-hover transition-colors cursor-pointer group/card${focused ? " bg-accent/10 border-l-2 border-l-accent" : ""}`}
       onClick={(e) => {
         // Don't navigate if clicking on interactive elements
         const target = e.target as HTMLElement;

src/lib/nostr/notes.ts

@@ -20,7 +20,7 @@ export async function fetchMediaFeed(limit: number = 500): Promise<NDKEvent[]> {
   return Array.from(events).sort((a, b) => (b.created_at ?? 0) - (a.created_at ?? 0));
 }
 
-export async function fetchFollowFeed(pubkeys: string[], limit = 80): Promise<NDKEvent[]> {
+export async function fetchFollowFeed(pubkeys: string[], limit = 30): Promise<NDKEvent[]> {
   if (pubkeys.length === 0) return [];
   const instance = getNDK();
   const since = Math.floor(Date.now() / 1000) - 24 * 3600; // last 24h for follows

src/lib/notificationPoller.ts

@@ -44,9 +44,9 @@ async function pollOnce(pubkey: string) {
         const name = await getProfileName(e.pubkey);
         notifyMention(name, e.content?.slice(0, 120) || "mentioned you").catch(() => {});
       }
+      // Only refresh the full store when there's actually something new to show
+      useNotificationsStore.getState().fetchNotifications(pubkey).catch(() => {});
     }
-    // Also update the notifications store
-    useNotificationsStore.getState().fetchNotifications(pubkey).catch(() => {});
   } catch { /* non-critical */ }
 
   // Zaps
@@ -107,18 +107,15 @@ export function startNotificationPoller(pubkey: string) {
   // Instant: load cached notifications from DB (no flicker)
   useNotificationsStore.getState().loadFromDb(pubkey);
 
-  // Then connect to relays and fetch new data in background
-  (async () => {
-    try {
-      const connected = await ensureConnected();
-      debug.log("notif:poller ensureConnected →", connected);
-    } catch { /* continue anyway */ }
-    debug.log("notif:poller initial fetch for", pubkey.slice(0, 8));
-    useNotificationsStore.getState().fetchNotifications(pubkey).catch(() => {});
-  })();
+  // The full fetchNotifications() sweep is already called by the login flow
+  // (loginWithNsec / loginWithPubkey / restoreSession) before this function runs.
+  // Starting it again here would fire two concurrent 7-day sweeps on every login.
+  // We only need the incremental pollOnce() loop from here on.
 
-  // Run first full poll after a longer delay (give relays more time)
-  setTimeout(() => pollOnce(pubkey).catch(() => {}), 8000);
+  // Delay the first poll to give the app and relays time to fully initialize.
+  // 8s was too aggressive — it would fire a heavy fetchNotifications during the
+  // initial feed load, contributing to the login memory spike.
+  setTimeout(() => pollOnce(pubkey).catch(() => {}), 90_000);
   intervalId = setInterval(() => pollOnce(pubkey).catch(() => {}), POLL_INTERVAL);
 }