diff --git a/apps/desktop/src/components/InCallPanel.tsx b/apps/desktop/src/components/InCallPanel.tsx
index b6d2607..94942a9 100644
--- a/apps/desktop/src/components/InCallPanel.tsx
+++ b/apps/desktop/src/components/InCallPanel.tsx
@@ -1,7 +1,7 @@
 import type { ConversationSummary } from '@chat-app/shared/chat';
 import type { ConnectionQuality, RemoteParticipant, Room } from 'livekit-client';
 import { RoomEvent, Track } from 'livekit-client';
-import { useEffect, useMemo, useRef, useState } from 'react';
+import { useEffect, useMemo, useState } from 'react';
 import { useTranslation } from 'react-i18next';
 
 import { useAuth } from '../context/AuthContext';
@@ -195,22 +195,6 @@ export function InCallPanel({ conversation }: Props) {
     if (soundboardCount === 0) setSoundboardOpen(false);
   }, [soundboardCount]);
 
-  // Active-speaker auto-focus uses "who most recently started speaking"
-  // rather than "exactly one speaker" — matches Discord more closely and
-  // handles the case where two people talk briefly without the focus
-  // collapsing to nobody.
-  const [lastStartedSpeakerId, setLastStartedSpeakerId] = useState<string | null>(null);
-  const prevActiveSpeakersRef = useRef<Set<string>>(new Set());
-  useEffect(() => {
-    for (const id of activeSpeakers) {
-      if (!prevActiveSpeakersRef.current.has(id)) {
-        setLastStartedSpeakerId(id);
-        break;
-      }
-    }
-    prevActiveSpeakersRef.current = new Set(activeSpeakers);
-  }, [activeSpeakers]);
-
   // Single right-click dispatcher for all tiles. User-tiles open the volume
   // menu; screen-tiles open the share-specific menu (volume + mute + stop
   // watching). Self-tiles get no menu — no volume to control, and you can
@@ -385,24 +369,13 @@ export function InCallPanel({ conversation }: Props) {
     }));
 
   if (callMode === 'fullscreen') {
-    // In fullscreen, a "manual focus" = user explicitly picked someone OR
-    // the person who most recently started speaking (tracked in
-    // lastStartedSpeakerId). Screen shares are no longer an auto-focus
-    // trigger; they stay as equal-size grid tiles until the user clicks
-    // one. "Most recent speaker" beats "exactly one currently speaking"
-    // because two people briefly overlapping shouldn't kick us out of
-    // auto-focus.
-    const autoSpeaker =
-      focusedId === null && lastStartedSpeakerId !== null
-        ? tiles.find(
-            (t) =>
-              t.kind === 'user' &&
-              !t.self &&
-              t.userId === lastStartedSpeakerId,
-          )
-        : undefined;
-    const hasFocus = focusedId !== null || autoSpeaker !== undefined;
-    const effectiveSpeaker = hasFocus ? speaker ?? autoSpeaker : undefined;
+    // Discord-style: without an explicit pin, fullscreen stays as a stable
+    // equal-size grid. Focus only kicks in when the user pins a tile
+    // (focusedId !== null). Auto-promoting the latest speaker caused the
+    // whole layout to flip on every utterance, which is not what users
+    // expect from a "cinema" view.
+    const hasFocus = focusedId !== null;
+    const effectiveSpeaker = hasFocus ? speaker : undefined;
     return (
       <>
         {micError && (
@@ -1032,7 +1005,7 @@ function CallStage({
             {others.map((p) => (
               <div
                 key={p.id}
-                className="h-full w-[240px] shrink-0 [&>div]:h-full"
+                className="aspect-video h-full shrink-0 [&>div]:h-full [&>div]:w-full"
               >
                 <TileRender
                   tile={p}
@@ -1058,9 +1031,16 @@ function CallStage({
   const gridClass = gridColsFor(tiles.length);
   return (
     <div className={'min-h-0 flex-1 p-3 ' + (compact ? '' : 'p-4')}>
-      <div className={'grid h-full gap-2 ' + gridClass}>
+      <div
+        className={
+          'grid h-full max-h-full gap-2 place-content-center ' + gridClass
+        }
+      >
         {tiles.map((p) => (
-          <div key={p.id} className="[&>div]:h-full [&>div]:w-full">
+          <div
+            key={p.id}
+            className="aspect-video min-h-0 w-full [&>div]:h-full [&>div]:w-full"
+          >
             <TileRender
               tile={p}
               activeSpeakers={activeSpeakers}
@@ -1112,17 +1092,16 @@ function FocusedTile({
 const GRID_PAGE_SIZE = 12;
 
 function gridColsFor(n: number): string {
-  // Explicit `grid-rows-*` so cells get a defined height (1fr of available
-  // space). Without this, implicit rows default to auto → they size to
-  // content, and a video element's intrinsic size blows the tile past the
-  // container bounds (overlapping the toolbar below).
-  if (n <= 1) return 'grid-cols-1 grid-rows-1';
-  if (n === 2) return 'grid-cols-2 grid-rows-1';
-  if (n === 3) return 'grid-cols-3 grid-rows-1';
-  if (n === 4) return 'grid-cols-2 grid-rows-2';
-  if (n <= 6) return 'grid-cols-3 grid-rows-2';
-  if (n <= 9) return 'grid-cols-3 grid-rows-3';
-  return 'grid-cols-4 grid-rows-3';
+  // Discord-style: column count only. Cells are `aspect-video` so their
+  // height follows from their width, and the container centers them
+  // vertically when the row stack is shorter than the available area.
+  if (n <= 1) return 'grid-cols-1';
+  if (n === 2) return 'grid-cols-2';
+  if (n === 3) return 'grid-cols-3';
+  if (n === 4) return 'grid-cols-2';
+  if (n <= 6) return 'grid-cols-3';
+  if (n <= 9) return 'grid-cols-3';
+  return 'grid-cols-4';
 }
 
 // Promote self + active speakers to the front of the tile list. Stable
@@ -1217,11 +1196,13 @@ function FullscreenCall({
   const others = hasFocus ? tiles.filter((p) => p.id !== speaker!.id) : [];
 
   // Active-speaker reorder + paginate. When more than GRID_PAGE_SIZE tiles
-  // exist, slice them into pages. Reset to page 0 if the page count drops
-  // below the current page (someone left).
+  // exist, slice them into pages and prioritize active speakers onto page 1.
+  // Otherwise keep a stable order (Discord-style) so tiles don't shuffle
+  // whenever someone speaks.
+  const needsPagination = tiles.length > GRID_PAGE_SIZE;
   const sortedGridTiles = useMemo(
-    () => prioritizeTiles(tiles, activeSpeakers),
-    [tiles, activeSpeakers],
+    () => (needsPagination ? prioritizeTiles(tiles, activeSpeakers) : tiles),
+    [tiles, activeSpeakers, needsPagination],
   );
   const pageCount = Math.max(1, Math.ceil(sortedGridTiles.length / GRID_PAGE_SIZE));
   useEffect(() => {