diff --git a/src/app/components/RenderMessageContent.tsx b/src/app/components/RenderMessageContent.tsx
index 71700e82..45f9aaf5 100644
--- a/src/app/components/RenderMessageContent.tsx
+++ b/src/app/components/RenderMessageContent.tsx
@@ -65,6 +65,12 @@ type RenderMessageContentProps = {
   htmlReactParserOptions: HTMLReactParserOptions;
   linkifyOpts: Opts;
   outlineAttachment?: boolean;
+  // Threaded into `ImageContent` so its onClick can open the new
+  // atom-driven horseshoe media viewer instead of the legacy
+  // `<Overlay>` modal. Set by `RoomTimeline`; non-Room callers
+  // (pin-menu, message search) leave it undefined and stay on the
+  // legacy modal until they're migrated.
+  eventId?: string;
 };
 export function RenderMessageContent({
   displayName,
@@ -78,6 +84,7 @@ export function RenderMessageContent({
   htmlReactParserOptions,
   linkifyOpts,
   outlineAttachment,
+  eventId,
 }: RenderMessageContentProps) {
   const streamMedia = useStreamMediaContext();
   const renderUrlsPreview = (urls: string[]) => {
@@ -219,6 +226,7 @@ export function RenderMessageContent({
       <ImageContent
         {...props}
         autoPlay={mediaAutoLoad}
+        eventId={eventId}
         renderImage={(p) => <Image {...p} loading="lazy" decoding="async" />}
         renderViewer={(p) => <ImageViewer {...p} />}
       />
@@ -258,6 +266,7 @@ export function RenderMessageContent({
         body={body}
         info={info}
         {...props}
+        eventId={eventId}
         renderThumbnail={
           mediaAutoLoad
             ? () => (
diff --git a/src/app/components/message/content/ImageContent.tsx b/src/app/components/message/content/ImageContent.tsx
index 4f1d9c75..90122b24 100644
--- a/src/app/components/message/content/ImageContent.tsx
+++ b/src/app/components/message/content/ImageContent.tsx
@@ -31,6 +31,8 @@ import { decryptFile, downloadEncryptedMedia, mxcUrlToHttp } from '../../../util
 import { useMediaAuthentication } from '../../../hooks/useMediaAuthentication';
 import { ModalWide } from '../../../styles/Modal.css';
 import { validBlurHash } from '../../../utils/blurHash';
+import { useMediaViewerHost } from '../../../features/room/mediaViewerHostContext';
+import { useOpenMediaViewer } from '../../../state/hooks/mediaViewer';
 
 type RenderViewerProps = {
   src: string;
@@ -44,7 +46,17 @@ type RenderImageProps = {
   onLoad: () => void;
   onError: () => void;
   onClick: () => void;
+  onKeyDown: (e: React.KeyboardEvent<HTMLImageElement>) => void;
   tabIndex: number;
+  // `role="button"` so assistive tech announces the clickable
+  // image as a button rather than a plain image. Paired with
+  // `aria-label` and an Enter/Space `onKeyDown` to make the
+  // affordance keyboard-activatable per WAI-ARIA. Element-Web
+  // wraps in `<AccessibleButton>` — we keep the bare `<img>` to
+  // avoid relayout, which works because folds' `Image` is
+  // `as<'img'>` and just spreads these props onto the DOM node.
+  role: 'button';
+  'aria-label': string;
 };
 export type ImageContentProps = {
   body: string;
@@ -55,6 +67,13 @@ export type ImageContentProps = {
   autoPlay?: boolean;
   markedAsSpoiler?: boolean;
   spoilerReason?: string;
+  // When provided AND the `MediaViewerHostContext` is non-null,
+  // clicking the thumbnail opens the atom-driven horseshoe viewer
+  // (mobile bottom-up sheet / desktop right pane) instead of the
+  // legacy full-screen `<Overlay>` viewer. Non-Room surfaces
+  // (pin-menu, message search) leave the host context as `null` and
+  // therefore keep the legacy modal even if they pass `eventId`.
+  eventId?: string;
   renderViewer: (props: RenderViewerProps) => ReactNode;
   renderImage: (props: RenderImageProps) => ReactNode;
 };
@@ -70,6 +89,7 @@ export const ImageContent = as<'div', ImageContentProps>(
       autoPlay,
       markedAsSpoiler,
       spoilerReason,
+      eventId,
       renderViewer,
       renderImage,
       ...props
@@ -79,12 +99,37 @@ export const ImageContent = as<'div', ImageContentProps>(
     const mx = useMatrixClient();
     const useAuthentication = useMediaAuthentication();
     const blurHash = validBlurHash(info?.[MATRIX_BLUR_HASH_PROPERTY_NAME]);
+    const host = useMediaViewerHost();
+    const openMediaViewer = useOpenMediaViewer();
+    const useAtomViewer = !!(host && eventId);
 
     const [load, setLoad] = useState(false);
     const [error, setError] = useState(false);
     const [viewer, setViewer] = useState(false);
     const [blurred, setBlurred] = useState(markedAsSpoiler ?? false);
 
+    const handleOpen = () => {
+      if (useAtomViewer && host && eventId) {
+        // The viewer body re-resolves + decrypts the media itself,
+        // owning the blob-URL lifecycle so it can revoke on close.
+        // We deliberately don't pass `srcState.data` here even when
+        // it's available — pinning a blob URL into the atom would
+        // leak it (the atom outlives the timeline thumbnail).
+        openMediaViewer({
+          roomId: host.roomId,
+          eventId,
+          kind: 'image',
+          url,
+          body,
+          info,
+          encInfo,
+          mimeType,
+        });
+        return;
+      }
+      setViewer(true);
+    };
+
     const [srcState, loadSrc] = useAsyncCallback(
       useCallback(async () => {
         const mediaUrl = mxcUrlToHttp(mx, url, useAuthentication);
@@ -118,7 +163,7 @@ export const ImageContent = as<'div', ImageContentProps>(
 
     return (
       <Box className={classNames(css.RelativeBase, className)} {...props} ref={ref}>
-        {srcState.status === AsyncStatus.Success && (
+        {!useAtomViewer && srcState.status === AsyncStatus.Success && (
           <Overlay open={viewer} backdrop={<OverlayBackdrop />}>
             <OverlayCenter>
               <FocusTrap
@@ -168,15 +213,28 @@ export const ImageContent = as<'div', ImageContentProps>(
           </Box>
         )}
         {srcState.status === AsyncStatus.Success && (
-          <Box className={classNames(css.AbsoluteContainer, blurred && css.Blur)}>
+          <Box
+            className={classNames(
+              css.AbsoluteContainer,
+              blurred ? css.Blur : css.ImageClickable
+            )}
+          >
             {renderImage({
               alt: body,
               title: body,
               src: srcState.data,
               onLoad: handleLoad,
               onError: handleError,
-              onClick: () => setViewer(true),
+              onClick: handleOpen,
+              onKeyDown: (e) => {
+                if (e.key === 'Enter' || e.key === ' ') {
+                  e.preventDefault();
+                  handleOpen();
+                }
+              },
               tabIndex: 0,
+              role: 'button',
+              'aria-label': body || 'Open media',
             })}
           </Box>
         )}
diff --git a/src/app/components/message/content/VideoContent.tsx b/src/app/components/message/content/VideoContent.tsx
index b33ec272..09640f61 100644
--- a/src/app/components/message/content/VideoContent.tsx
+++ b/src/app/components/message/content/VideoContent.tsx
@@ -32,6 +32,8 @@ import {
 } from '../../../utils/matrix';
 import { useMediaAuthentication } from '../../../hooks/useMediaAuthentication';
 import { validBlurHash } from '../../../utils/blurHash';
+import { useMediaViewerHost } from '../../../features/room/mediaViewerHostContext';
+import { useOpenMediaViewer } from '../../../state/hooks/mediaViewer';
 
 type RenderVideoProps = {
   title: string;
@@ -50,6 +52,14 @@ type VideoContentProps = {
   autoPlay?: boolean;
   markedAsSpoiler?: boolean;
   spoilerReason?: string;
+  // When provided AND `MediaViewerHostContext` is non-null, tapping
+  // the thumbnail opens the atom-driven horseshoe viewer for video
+  // playback instead of loading + playing inline (which hands off to
+  // the browser's native video-element fullscreen when the user hits
+  // the controls' expand button — that's why the user used to see
+  // Chrome's default video viewer). Non-Room surfaces leave the
+  // host context as `null` and keep the inline player.
+  eventId?: string;
   renderThumbnail?: () => ReactNode;
   renderVideo: (props: RenderVideoProps) => ReactNode;
 };
@@ -65,6 +75,7 @@ export const VideoContent = as<'div', VideoContentProps>(
       autoPlay,
       markedAsSpoiler,
       spoilerReason,
+      eventId,
       renderThumbnail,
       renderVideo,
       ...props
@@ -74,6 +85,9 @@ export const VideoContent = as<'div', VideoContentProps>(
     const mx = useMatrixClient();
     const useAuthentication = useMediaAuthentication();
     const blurHash = validBlurHash(info.thumbnail_info?.[MATRIX_BLUR_HASH_PROPERTY_NAME]);
+    const host = useMediaViewerHost();
+    const openMediaViewer = useOpenMediaViewer();
+    const useAtomViewer = !!(host && eventId);
 
     const [load, setLoad] = useState(false);
     const [error, setError] = useState(false);
@@ -106,8 +120,29 @@ export const VideoContent = as<'div', VideoContentProps>(
     };
 
     useEffect(() => {
+      // Skip inline preload in atom-viewer mode — the user gets the
+      // viewer's own resolve path on tap; preloading every visible
+      // video in the timeline would burn bandwidth and decrypt CPU
+      // for videos the user never opens.
+      if (useAtomViewer) return;
       if (autoPlay) loadSrc();
-    }, [autoPlay, loadSrc]);
+    }, [autoPlay, loadSrc, useAtomViewer]);
+
+    const openAtomViewer = useCallback(() => {
+      if (!host || !eventId) return;
+      // No `resolvedSrc` — viewer body owns blob-URL lifecycle; see
+      // the rationale in `ImageContent.handleOpen`.
+      openMediaViewer({
+        roomId: host.roomId,
+        eventId,
+        kind: 'video',
+        url,
+        body,
+        info,
+        encInfo,
+        mimeType,
+      });
+    }, [host, eventId, openMediaViewer, url, body, info, encInfo, mimeType]);
 
     return (
       <Box className={classNames(css.RelativeBase, className)} {...props} ref={ref}>
@@ -129,7 +164,21 @@ export const VideoContent = as<'div', VideoContentProps>(
             {renderThumbnail()}
           </Box>
         )}
-        {!autoPlay && !blurred && srcState.status === AsyncStatus.Idle && (
+        {useAtomViewer && !blurred && (
+          <Box className={css.AbsoluteContainer} alignItems="Center" justifyContent="Center">
+            <Button
+              variant="Secondary"
+              fill="Solid"
+              radii="300"
+              size="300"
+              onClick={openAtomViewer}
+              before={<Icon size="Inherit" src={Icons.Play} filled />}
+            >
+              <Text size="B300">Watch</Text>
+            </Button>
+          </Box>
+        )}
+        {!useAtomViewer && !autoPlay && !blurred && srcState.status === AsyncStatus.Idle && (
           <Box className={css.AbsoluteContainer} alignItems="Center" justifyContent="Center">
             <Button
               variant="Secondary"
@@ -143,7 +192,7 @@ export const VideoContent = as<'div', VideoContentProps>(
             </Button>
           </Box>
         )}
-        {srcState.status === AsyncStatus.Success && (
+        {!useAtomViewer && srcState.status === AsyncStatus.Success && (
           <Box className={classNames(css.AbsoluteContainer, blurred && css.Blur)}>
             {renderVideo({
               title: body,
diff --git a/src/app/components/message/content/style.css.ts b/src/app/components/message/content/style.css.ts
index bb5d8484..dadcbe3a 100644
--- a/src/app/components/message/content/style.css.ts
+++ b/src/app/components/message/content/style.css.ts
@@ -1,6 +1,36 @@
-import { style } from '@vanilla-extract/css';
+import { globalStyle, style } from '@vanilla-extract/css';
 import { DefaultReset, config } from 'folds';
 
+// Click affordance for the timeline image thumbnail. Without this
+// the `<img>` looked decorative on web (the default cursor stays
+// as `default`) even though it's clickable to open the media
+// viewer. The subtle 0.92 brightness on hover doubles as the
+// "this is interactive" signal — same idiom as how the rest of
+// the app's clickable surfaces shift tone on hover.
+//
+// `will-change: filter` hints the compositor so the brightness
+// transition runs on the GPU instead of repainting on the CPU —
+// matters for large hi-DPI thumbnails on slower phones.
+export const ImageClickable = style({
+  cursor: 'pointer',
+  transition: 'filter 120ms ease',
+  willChange: 'filter',
+  selectors: {
+    '&:hover': { filter: 'brightness(0.92)' },
+  },
+});
+
+// `:focus-visible` outline on the inner `<img>` (which carries
+// `tabIndex` + `role="button"`, not the wrapper Box). Lives as a
+// `globalStyle` because vanilla-extract's `style({...})` only
+// permits selectors that target the class itself (`&...`) — a
+// descendant selector like `& img:focus-visible` errors at build
+// time. `globalStyle` is the documented escape valve for that.
+globalStyle(`${ImageClickable} img:focus-visible`, {
+  outline: `2px solid currentColor`,
+  outlineOffset: '-2px',
+});
+
 export const RelativeBase = style([
   DefaultReset,
   {
diff --git a/src/app/features/room/MediaViewerBody.css.ts b/src/app/features/room/MediaViewerBody.css.ts
new file mode 100644
index 00000000..b5f3a2e6
--- /dev/null
+++ b/src/app/features/room/MediaViewerBody.css.ts
@@ -0,0 +1,143 @@
+import { style } from '@vanilla-extract/css';
+import { color, config, toRem } from 'folds';
+
+// Root layout — header strip → image stage → action row. `position:
+// relative` so the prev/next chevrons can absolute-position over the
+// stage on desktop.
+export const root = style({
+  flex: 1,
+  display: 'flex',
+  flexDirection: 'column',
+  minHeight: 0,
+  minWidth: 0,
+  backgroundColor: color.Background.Container,
+  color: color.Background.OnContainer,
+});
+
+// `PageHeader` (folds `Header size="600"`) provides the strip
+// height; we only override the left gutter so the title sits at
+// the same x-offset as the profile side pane's title — keeps both
+// right-pane variants visually consistent.
+export const header = style({
+  paddingLeft: config.space.S300,
+});
+
+export const title = style({
+  minWidth: 0,
+  flex: '0 1 auto',
+});
+
+// Vertical hairline between the action cluster and the close
+// button — visually separates "do something with this image" from
+// "close the viewer", which are semantically different actions.
+// Rendered for image kind only (no zoom controls for video → only
+// download + close, which doesn't need a separator).
+export const headerSeparator = style({
+  flexShrink: 0,
+  width: '1px',
+  height: toRem(20),
+  margin: `0 ${config.space.S100}`,
+  backgroundColor: color.SurfaceVariant.Container,
+});
+
+// Image stage — fills remaining vertical space, centres the image
+// with object-fit-style logic on the wrapper. `overflow: hidden` so
+// a zoomed-in pan can't bleed past the stage. `touch-action: none`
+// blocks browser-native gestures (pinch-zoom, scroll, refresh-pull)
+// — we own the touch surface here for swipe + pan.
+export const stage = style({
+  flex: 1,
+  position: 'relative',
+  display: 'flex',
+  alignItems: 'center',
+  justifyContent: 'center',
+  overflow: 'hidden',
+  touchAction: 'none',
+  userSelect: 'none',
+});
+
+// The actual <img>. `max-*: 100%` keeps it within the stage at
+// zoom=1; transforms (scale, translate) layer above without
+// re-flowing. `will-change` hints the compositor for smooth pan /
+// swipe / zoom animations.
+export const image = style({
+  maxWidth: '100%',
+  maxHeight: '100%',
+  objectFit: 'contain',
+  willChange: 'transform',
+  // Sharp pixel-ratio handling on dense screens — keeps phone
+  // photos crisp at zoom=1, smooths only when zoomed beyond
+  // intrinsic resolution.
+  imageRendering: 'auto',
+});
+
+// Side chevrons — visibility logic:
+//   • Coarse pointer (touch devices, mobile): always visible at
+//     0.7 opacity so users have a tappable nav affordance without
+//     needing the hover state that touch lacks. Swipe still works
+//     but the buttons are the discoverability path.
+//   • Fine pointer (desktop): hidden until the stage is hovered,
+//     fading in to full opacity. Keeps the image surface clean
+//     when the cursor isn't on it.
+//   • Boundary (first/last item): the chevron is dimmed (0.25
+//     opacity, pointer-events disabled) rather than hidden, so the
+//     user gets a visual cue that prev/next isn't available
+//     instead of a missing button.
+export const chevron = style({
+  position: 'absolute',
+  top: '50%',
+  transform: 'translateY(-50%)',
+  width: toRem(40),
+  height: toRem(40),
+  borderRadius: '50%',
+  display: 'flex',
+  alignItems: 'center',
+  justifyContent: 'center',
+  backgroundColor: 'rgba(0, 0, 0, 0.55)',
+  color: color.Background.OnContainer,
+  cursor: 'pointer',
+  border: 'none',
+  outline: 'none',
+  transition: 'opacity 150ms ease, background-color 150ms ease',
+  opacity: 0,
+  '@media': {
+    '(pointer: coarse)': {
+      opacity: 0.7,
+    },
+  },
+  selectors: {
+    [`${stage}:hover &`]: { opacity: 1 },
+    '&:hover': { backgroundColor: 'rgba(0, 0, 0, 0.75)' },
+    '&[aria-disabled="true"]': { opacity: 0.25, pointerEvents: 'none' },
+  },
+});
+
+export const chevronLeft = style({
+  left: config.space.S300,
+});
+
+export const chevronRight = style({
+  right: config.space.S300,
+});
+
+// Loading / error state overlay — centred inside the stage.
+export const stateOverlay = style({
+  position: 'absolute',
+  inset: 0,
+  display: 'flex',
+  alignItems: 'center',
+  justifyContent: 'center',
+  pointerEvents: 'none',
+});
+
+// Index pill — "3 / 12" in the header. Subtle, matches the title
+// row tone.
+export const indexPill = style({
+  flexShrink: 0,
+  padding: `${toRem(2)} ${config.space.S200}`,
+  borderRadius: toRem(999),
+  backgroundColor: color.SurfaceVariant.Container,
+  color: color.SurfaceVariant.OnContainer,
+  fontSize: toRem(12),
+  fontWeight: 500,
+});
diff --git a/src/app/features/room/MediaViewerBody.tsx b/src/app/features/room/MediaViewerBody.tsx
new file mode 100644
index 00000000..a03ac2e5
--- /dev/null
+++ b/src/app/features/room/MediaViewerBody.tsx
@@ -0,0 +1,971 @@
+// Body for the room media viewer — mounted inside both the mobile
+// `MobileMediaViewerHorseshoe` panel-body and the desktop
+// `RoomViewMediaSidePanel`. Owns:
+//   • Resolving + decrypting the media URL for the current entry,
+//     including revoke-on-unmount for blob URLs we create.
+//   • Zoom for images (button row + `+`/`-`/`0` keys).
+//   • Mouse-drag pan when zoomed, with bounds clamping so the
+//     image can't be dragged off the stage.
+//   • Multi-media navigation: prev / next via on-screen chevrons,
+//     keyboard arrows, and horizontal swipe (image only — video
+//     keeps its native seekbar drag).
+//   • Programmatic `<video>` play on each entry change.
+//   • Download via `FileSaver`.
+
+import React, { MouseEventHandler, useCallback, useEffect, useMemo, useRef, useState } from 'react';
+import { Box, Icon, IconButton, Icons, Spinner, Text } from 'folds';
+import { PageHeader } from '../../components/page';
+import { ContainerColor } from '../../styles/ContainerColor.css';
+import classNames from 'classnames';
+import FileSaver from 'file-saver';
+import { MatrixEvent, MatrixEventEvent, MsgType, RoomEvent } from 'matrix-js-sdk';
+import { useTranslation } from 'react-i18next';
+import { EncryptedAttachmentInfo } from 'browser-encrypt-attachment';
+import { MediaViewerEntry } from '../../state/mediaViewer';
+import { useOpenMediaViewer } from '../../state/hooks/mediaViewer';
+import { AsyncStatus, useAsyncCallback } from '../../hooks/useAsyncCallback';
+import { useMatrixClient } from '../../hooks/useMatrixClient';
+import { useMediaAuthentication } from '../../hooks/useMediaAuthentication';
+import { useRoom } from '../../hooks/useRoom';
+import { useZoom } from '../../hooks/useZoom';
+import {
+  decryptFile,
+  downloadEncryptedMedia,
+  mxcUrlToHttp,
+} from '../../utils/matrix';
+import { FALLBACK_MIMETYPE } from '../../utils/mimeTypes';
+import {
+  IImageContent,
+  IImageInfo,
+  IThumbnailContent,
+  IVideoContent,
+  IVideoInfo,
+} from '../../../types/matrix/common';
+import * as css from './MediaViewerBody.css';
+
+// Distance past which a horizontal pointer drag commits to prev/next
+// when zoom = 1 (matches the sheet's vertical close threshold).
+const SWIPE_COMMIT_PX = 80;
+// Horizontal must dominate vertical by at least this factor before
+// the gesture is interpreted as a swipe — keeps accidental diagonal
+// drags from committing.
+const SWIPE_DOMINANCE = 1.5;
+
+// Walk the room's loaded timeline events and build the ordered set
+// of viewable image + video events. Out-of-window media (older than
+// the loaded scrollback) is intentionally not surfaced for phase 1;
+// `step()` resistance-snaps at the boundary.
+function mediaKindFor(ev: MatrixEvent): 'image' | 'video' | null {
+  if (ev.getType() !== 'm.room.message') return null;
+  if (ev.isRedacted()) return null;
+  const content = ev.getContent() as { msgtype?: string; url?: string; file?: { url?: string } };
+  if (!content) return null;
+  const hasMedia = !!content.url || !!content.file?.url;
+  if (!hasMedia) return null;
+  if (content.msgtype === MsgType.Image) return 'image';
+  if (content.msgtype === MsgType.Video) return 'video';
+  return null;
+}
+
+function eventToEntry(roomId: string, ev: MatrixEvent): MediaViewerEntry | null {
+  const kind = mediaKindFor(ev);
+  if (!kind) return null;
+  const content = ev.getContent() as (IImageContent | IVideoContent) | undefined;
+  if (!content) return null;
+  const eventId = ev.getId();
+  if (!eventId) return null;
+  const file = content.file;
+  const url = file?.url ?? content.url;
+  if (!url) return null;
+  const encInfo: EncryptedAttachmentInfo | undefined = file
+    ? (() => {
+        // Strip the `url` field — `EncryptedAttachmentInfo` doesn't
+        // include it (the wire shape `IEncryptedFile` does).
+        // eslint-disable-next-line @typescript-eslint/no-unused-vars
+        const { url: _, ...rest } = file;
+        return rest as EncryptedAttachmentInfo;
+      })()
+    : undefined;
+  return {
+    roomId,
+    eventId,
+    kind,
+    url,
+    body: content.filename ?? content.body ?? eventId,
+    info: content.info as ((IImageInfo | IVideoInfo) & IThumbnailContent) | undefined,
+    encInfo,
+    mimeType: content.info?.mimetype,
+  };
+}
+
+type MediaViewerBodyProps = {
+  entry: MediaViewerEntry;
+  requestClose: () => void;
+};
+
+export function MediaViewerBody({ entry, requestClose }: MediaViewerBodyProps) {
+  const { t } = useTranslation();
+  const mx = useMatrixClient();
+  const useAuthentication = useMediaAuthentication();
+  const room = useRoom();
+  const openMediaViewer = useOpenMediaViewer();
+
+  // `useZoom` gives us `zoom` + `setZoom`; `zoomIn`/`zoomOut` from
+  // it use a flat max=5, but we want per-image dynamic ceiling.
+  // Override with custom +/− callbacks that read `maxZoomRef`.
+  const { zoom, setZoom } = useZoom(0.25);
+
+  // Pan is local (not the shared `usePan` hook) because we need to
+  // clamp against the image bounds so the user can't drag a zoomed
+  // image past its edges — `usePan` accumulates raw `movementX/Y`
+  // without bounds, which lets the picture escape off-screen.
+  // Clamp formula: with `transform: translate3d(X, Y) scale(zoom)`,
+  // the translate is applied AFTER scaling, so X/Y are in stage
+  // pixels. Max pan in each axis = (rendered_size * zoom -
+  // stage_size) / 2. Image fits within stage at zoom=1 (object-fit:
+  // contain), so clamp is zero there and pan auto-resets via the
+  // `useEffect` below that watches `zoom`.
+  const stageRef = useRef<HTMLDivElement>(null);
+  const imgRef = useRef<HTMLImageElement>(null);
+  const videoRef = useRef<HTMLVideoElement>(null);
+  const [pan, setPan] = useState<{ x: number; y: number }>({ x: 0, y: 0 });
+  const [panCursor, setPanCursor] = useState<'grab' | 'grabbing' | 'initial'>('initial');
+
+  // Mirror `zoom` and `pan` into refs so callbacks installed in
+  // long-lived listeners can read the latest values without
+  // re-binding, and so the inlined anchor-aware zoom math in
+  // pinch / wheel can read current zoom/pan synchronously.
+  const zoomRef = useRef(zoom);
+  zoomRef.current = zoom;
+  const panRef = useRef(pan);
+  panRef.current = pan;
+
+  // Per-image zoom bounds. `minZoom` is fixed at 1.0 — with
+  // `object-fit: contain` the image already fills the stage to fit
+  // at zoom=1, and going smaller just leaves dead space. `maxZoom`
+  // is computed from the natural image size in `onImageLoad` so a
+  // hi-res screenshot can zoom up to 1:1 pixels (~ natural /
+  // displayed) and a small avatar gets a friendlier 2× ceiling.
+  // Falls back to a flat 5× for the brief window before load.
+  const MIN_ZOOM = 1;
+  const ZOOM_CEILING_FALLBACK = 5;
+  const ZOOM_CEILING_HARD_CAP = 8;
+  const [maxZoom, setMaxZoom] = useState(ZOOM_CEILING_FALLBACK);
+  // `maxZoom` mirrored into a ref for the long-lived gesture
+  // listeners installed once with `deps=[]` below (they need the
+  // latest ceiling, not the value at install time).
+  const maxZoomRef = useRef(maxZoom);
+  maxZoomRef.current = maxZoom;
+
+  // Custom zoom +/− that respect the dynamic per-image ceiling
+  // (`useZoom`'s built-in `zoomIn`/`zoomOut` would cap at the
+  // flat default of 5). Step size 0.25 matches the original hook
+  // config.
+  const zoomIn = useCallback(() => {
+    setZoom((z) => Math.min(maxZoomRef.current, z + 0.25));
+  }, [setZoom]);
+  const zoomOut = useCallback(() => {
+    setZoom((z) => Math.max(MIN_ZOOM, z - 0.25));
+  }, [setZoom]);
+
+  const onImageLoad = useCallback(() => {
+    const img = imgRef.current;
+    if (!img) return;
+    if (img.naturalWidth === 0 || img.offsetWidth === 0) return;
+    // Ratio to reach 1:1 pixels — how far we'd need to scale the
+    // displayed (object-fit-contained) image to match its native
+    // resolution. For images naturally smaller than the display
+    // (avatars), this is <1; floor to 2 so the user can still
+    // inspect detail.
+    const oneToOne = img.naturalWidth / img.offsetWidth;
+    const computed = Math.max(2, oneToOne);
+    setMaxZoom(Math.min(ZOOM_CEILING_HARD_CAP, computed));
+  }, []);
+
+  const clampPan = useCallback(
+    (raw: { x: number; y: number }, currentZoom: number) => {
+      const stage = stageRef.current;
+      const img = imgRef.current;
+      if (!stage || !img) return raw;
+      const stageRect = stage.getBoundingClientRect();
+      const maxX = Math.max(0, (img.offsetWidth * currentZoom - stageRect.width) / 2);
+      const maxY = Math.max(0, (img.offsetHeight * currentZoom - stageRect.height) / 2);
+      return {
+        x: Math.max(-maxX, Math.min(maxX, raw.x)),
+        y: Math.max(-maxY, Math.min(maxY, raw.y)),
+      };
+    },
+    []
+  );
+
+  // Anchor-aware zoom math (image-local point under anchor stays
+  // under the anchor after zoom change) is inlined in the pinch
+  // and wheel handlers below — they each need to construct the
+  // anchor from their own gesture state, so a shared helper would
+  // just be plumbing.
+
+  // Reset zoom, pan, and max-zoom-ceiling whenever the active
+  // entry changes (swipe / arrow / programmatic). Without
+  // resetting `maxZoom`, navigating from a low-res image (where
+  // `onImageLoad` set a small ceiling) to a high-res one would
+  // briefly use the stale ceiling until the new image's `onLoad`
+  // fires. Gesture state is reset in a separate effect below.
+  useEffect(() => {
+    setZoom(1);
+    setPan({ x: 0, y: 0 });
+    setMaxZoom(ZOOM_CEILING_FALLBACK);
+  }, [entry.eventId, setZoom]);
+
+  // Re-clamp the existing pan against the new bounds when zoom
+  // changes (e.g. zooming out from 3× to 1× should snap a panned
+  // image back to centre since there's nothing to pan into).
+  useEffect(() => {
+    setPan((prev) => {
+      const next = clampPan(prev, zoom);
+      if (next.x === prev.x && next.y === prev.y) return prev;
+      return next;
+    });
+    setPanCursor(zoom === 1 ? 'initial' : 'grab');
+  }, [zoom, clampPan]);
+
+  // Mouse-drag pan. Touch is handled by the stage-listener block
+  // below (which routes touch to swipe when zoom=1, ignores when
+  // zoom > 1 — pinch + touch-pan is a follow-up).
+  const onMouseDown: MouseEventHandler<HTMLImageElement> = useCallback(
+    (evt) => {
+      if (zoomRef.current === 1) return;
+      evt.preventDefault();
+      setPanCursor('grabbing');
+
+      const onMove = (mEvt: MouseEvent) => {
+        mEvt.preventDefault();
+        mEvt.stopPropagation();
+        const raw = {
+          x: panRef.current.x + mEvt.movementX,
+          y: panRef.current.y + mEvt.movementY,
+        };
+        const clamped = clampPan(raw, zoomRef.current);
+        panRef.current = clamped;
+        setPan(clamped);
+      };
+      const onUp = (mEvt: MouseEvent) => {
+        mEvt.preventDefault();
+        setPanCursor('grab');
+        document.removeEventListener('mousemove', onMove);
+        document.removeEventListener('mouseup', onUp);
+      };
+
+      document.addEventListener('mousemove', onMove);
+      document.addEventListener('mouseup', onUp);
+    },
+    [clampPan]
+  );
+
+  // Resolve + decrypt the media URL. Mirrors `ImageContent`'s
+  // pattern. Blob-URL lifecycle is managed via `lastBlobUrlRef`:
+  // every encrypted entry allocates a fresh blob INSIDE the async
+  // callback, and we revoke the previous one before storing the
+  // new one. This pattern handles two leak paths that a naive
+  // `useEffect(srcState)` cleanup misses:
+  //   1. `useAsyncCallback`'s "Request replaced!" reject path —
+  //      a rapidly-replaced loadSrc may have created a blob via
+  //      `createObjectURL` before the throw, which never reaches
+  //      `srcState` and so is never cleaned up. Here we revoke
+  //      whatever was last stamped into the ref before stamping
+  //      the new one.
+  //   2. Unmount mid-load — `useEffect` cleanup with a stale
+  //      `srcState.data` could revoke a blob the new mount is
+  //      still using. With a ref-tracked invariant of "we own
+  //      `lastBlobUrlRef.current`", the unmount cleanup is
+  //      unambiguous.
+  const lastBlobUrlRef = useRef<string | null>(null);
+  const [srcState, loadSrc] = useAsyncCallback(
+    useCallback(async () => {
+      const mediaUrl = mxcUrlToHttp(mx, entry.url, useAuthentication);
+      if (!mediaUrl) throw new Error('Invalid media URL');
+      if (entry.encInfo) {
+        const fileContent = await downloadEncryptedMedia(mediaUrl, (encBuf) =>
+          decryptFile(encBuf, entry.mimeType ?? FALLBACK_MIMETYPE, entry.encInfo!)
+        );
+        const blob = URL.createObjectURL(fileContent);
+        if (lastBlobUrlRef.current && lastBlobUrlRef.current !== blob) {
+          URL.revokeObjectURL(lastBlobUrlRef.current);
+        }
+        lastBlobUrlRef.current = blob;
+        return blob;
+      }
+      return mediaUrl;
+    }, [mx, entry.url, entry.encInfo, entry.mimeType, useAuthentication])
+  );
+
+  useEffect(() => {
+    loadSrc();
+  }, [loadSrc]);
+
+  // Revoke any blob URL we still own when the viewer body unmounts
+  // (sheet closed entirely, room change, etc).
+  useEffect(
+    () => () => {
+      if (lastBlobUrlRef.current) {
+        URL.revokeObjectURL(lastBlobUrlRef.current);
+        lastBlobUrlRef.current = null;
+      }
+    },
+    []
+  );
+
+  // Build the room's image set. Recomputed on each render — cheap
+  // (timeline events are already in memory). Not subscribed to new
+  // events for phase 1; the user navigates among the images that
+  // were visible when they opened the viewer.
+  // Bump a version counter on every live-timeline mutation in this
+  // room so `mediaSet` rebuilds when a new image / video event
+  // arrives mid-viewing — without this the user navigates a frozen
+  // snapshot. Mirror of Element-Web's FilePanel pattern. Redaction
+  // is also subscribed so a remotely-deleted image disappears from
+  // the prev/next chain.
+  //
+  // Per-event `MatrixEventEvent.Decrypted` subscription is deliberately
+  // skipped (cost of attaching N listeners). Late-arriving decrypts
+  // of pre-existing events are an accepted phase-1 limitation —
+  // documented in the comment on `mediaKindFor`.
+  const [timelineVersion, setTimelineVersion] = useState(0);
+  useEffect(() => {
+    const bump = () => setTimelineVersion((v) => v + 1);
+    // Filter Timeline events to media kinds so the mediaSet
+    // recompute doesn't fire on every reaction / edit / receipt /
+    // sync rollback (busy rooms produce dozens per second). Each
+    // recompute does an O(n) walk of `getLiveTimeline().getEvents()`.
+    const onTimeline = (ev: MatrixEvent) => {
+      if (mediaKindFor(ev) !== null) bump();
+    };
+    // Encrypted rooms (Vojo's primary surface): events arrive as
+    // `m.room.encrypted` and decrypt asynchronously after the
+    // Timeline event has fired — `mediaKindFor` returns null at
+    // that moment, so `onTimeline` above doesn't bump. When the
+    // event decrypts, `MatrixEventEvent.Decrypted` fires on the
+    // event itself. We attach via the MatrixClient so we catch
+    // every decrypt without subscribing to N per-event listeners,
+    // and filter to this room. Mirror of Element-Web's `FilePanel`
+    // pattern.
+    const onDecrypted = (ev: MatrixEvent) => {
+      if (ev.getRoomId() !== room.roomId) return;
+      if (mediaKindFor(ev) !== null) bump();
+    };
+    // Redactions: bump unconditionally — the redacted event may
+    // have been a media we're currently navigating, and we can't
+    // re-derive its kind once redacted.
+    room.on(RoomEvent.Timeline, onTimeline);
+    room.on(RoomEvent.Redaction, bump);
+    mx.on(MatrixEventEvent.Decrypted, onDecrypted);
+    return () => {
+      room.removeListener(RoomEvent.Timeline, onTimeline);
+      room.removeListener(RoomEvent.Redaction, bump);
+      mx.removeListener(MatrixEventEvent.Decrypted, onDecrypted);
+    };
+  }, [mx, room]);
+
+  const mediaSet = useMemo(() => {
+    const events = room.getLiveTimeline().getEvents();
+    const entries: MediaViewerEntry[] = [];
+    events.forEach((ev) => {
+      const e = eventToEntry(room.roomId, ev);
+      if (e) entries.push(e);
+    });
+    return entries;
+    // `timelineVersion` is the actual refresh trigger; `entry.eventId`
+    // is intentionally NOT a dep — stepping doesn't change the set.
+  }, [room, timelineVersion]);
+
+  const currentIndex = useMemo(
+    () => mediaSet.findIndex((e) => e.eventId === entry.eventId),
+    [mediaSet, entry.eventId]
+  );
+
+  const step = useCallback(
+    (delta: -1 | 1) => {
+      if (currentIndex < 0) return;
+      const target = currentIndex + delta;
+      if (target < 0 || target >= mediaSet.length) return;
+      const next = mediaSet[target];
+      if (!next) return;
+      openMediaViewer(next);
+    },
+    [currentIndex, mediaSet, openMediaViewer]
+  );
+
+  const canPrev = currentIndex > 0;
+  const canNext = currentIndex >= 0 && currentIndex < mediaSet.length - 1;
+
+  // Keyboard — ArrowLeft / Right step, +/- / 0 manage zoom. Esc is
+  // handled by the parent shell (focus-trap on desktop, explicit
+  // keydown listener on mobile).
+  useEffect(() => {
+    const onKeyDown = (e: KeyboardEvent) => {
+      if (e.defaultPrevented) return;
+      const target = e.target as HTMLElement | null;
+      if (
+        target &&
+        (target.tagName === 'INPUT' ||
+          target.tagName === 'TEXTAREA' ||
+          target.isContentEditable)
+      ) {
+        return;
+      }
+      if (e.key === 'ArrowLeft') {
+        e.preventDefault();
+        step(-1);
+      } else if (e.key === 'ArrowRight') {
+        e.preventDefault();
+        step(1);
+      } else if (e.key === '+' || e.key === '=') {
+        e.preventDefault();
+        zoomIn();
+      } else if (e.key === '-' || e.key === '_') {
+        e.preventDefault();
+        zoomOut();
+      } else if (e.key === '0') {
+        e.preventDefault();
+        setZoom(1);
+      }
+    };
+    window.addEventListener('keydown', onKeyDown);
+    return () => window.removeEventListener('keydown', onKeyDown);
+  }, [step, zoomIn, zoomOut, setZoom]);
+
+  // Multi-touch gesture state on the image stage. Three modes:
+  //   • zoom=1, 1 finger → horizontal-swipe to prev/next image.
+  //   • zoom>1, 1 finger → touch pan (clamped to image bounds; same
+  //     formula as the mouse pan path).
+  //   • 2 fingers → pinch zoom (distance ratio drives zoom; pan
+  //     resets to 0 so the image stays centered while pinching).
+  // Listeners are installed ONCE and read the latest closure values
+  // via refs — without that, a zoom-button press would re-bind every
+  // listener mid-gesture. PointerEvents only (no TouchEvents) — on
+  // Chromium both fire for touch, which double-fires every handler.
+  // The MDN reference pattern that drove this implementation:
+  // https://developer.mozilla.org/en-US/docs/Web/API/Pointer_events/Pinch_zoom_gestures
+  const [swipeOffset, setSwipeOffset] = useState(0);
+
+  const swipeStateRef = useRef({
+    active: false,
+    startX: 0,
+    startY: 0,
+    lastDx: 0,
+    claimed: 'none' as 'none' | 'swipe' | 'reject',
+  });
+
+  // Active touch pointers keyed by `pointerId`. Map (not array)
+  // because PointerEvents can arrive interleaved across multiple
+  // pointers and we need to update the per-id entry on each move.
+  const pointerCacheRef = useRef<Map<number, { x: number; y: number }>>(new Map());
+
+  // Pinch state. `null` when not pinching; set on the 2nd pointer
+  // down. `initialDistance` is the finger spread at pinch start;
+  // `initialZoom` and `initialPan` snapshot the state we'll scale
+  // from. `anchorImageX/Y` is the image-local point under the
+  // pinch midpoint at the start — held constant under the user's
+  // fingers as zoom changes (anchor-aware pinch, mirrors what
+  // Apple Photos / Google Photos / Element-Web's wheel-zoom do).
+  const pinchStateRef = useRef<{
+    initialDistance: number;
+    initialZoom: number;
+    initialPan: { x: number; y: number };
+    anchorImageX: number;
+    anchorImageY: number;
+  } | null>(null);
+
+  // Touch pan state. `null` when not panning. Stores the pan offset
+  // at the start of the gesture so we can compute the delta off the
+  // current pointer position (matches Element-Web's
+  // absolute-delta-from-anchor approach, more reliable than
+  // accumulating `movementX/Y`).
+  const touchPanStateRef = useRef<{
+    startX: number;
+    startY: number;
+    panAtStartX: number;
+    panAtStartY: number;
+  } | null>(null);
+
+  // Refs mirror the latest values for the always-installed listeners
+  // below. Without these, the listeners would close over the values
+  // from the render that installed them.
+  const canPrevRef = useRef(canPrev);
+  canPrevRef.current = canPrev;
+  const canNextRef = useRef(canNext);
+  canNextRef.current = canNext;
+  const stepRef = useRef(step);
+  stepRef.current = step;
+  const entryKindRef = useRef(entry.kind);
+  entryKindRef.current = entry.kind;
+  // `zoomRef` + `panRef` are already mirrored in the pan section
+  // above; reusing them here keeps the source of truth singular.
+  const clampPanRef = useRef(clampPan);
+  clampPanRef.current = clampPan;
+  const setZoomRef = useRef(setZoom);
+  setZoomRef.current = setZoom;
+  // `maxZoomRef` + `zoomIn` / `zoomOut` are declared higher up
+  // (right after the `maxZoom` state) because the keyboard
+  // useEffect above also depends on them.
+
+  // Reset gesture state on entry change. A finger that lifted
+  // off-stage (no `pointerup` delivered) leaves residue in
+  // `pointerCacheRef`, so without this the next entry's first
+  // `pointerdown` sees `count === 2` from nothing and starts a
+  // fake pinch on a stale midpoint. Similar concern for swipe /
+  // touch-pan state. `setPointerCapture` (added in the listeners
+  // below) should normally guarantee a pointerup delivery; this
+  // reset is the defence-in-depth.
+  useEffect(() => {
+    pointerCacheRef.current.clear();
+    pinchStateRef.current = null;
+    touchPanStateRef.current = null;
+    swipeStateRef.current.active = false;
+    swipeStateRef.current.claimed = 'none';
+    setSwipeOffset(0);
+  }, [entry.eventId]);
+
+  useEffect(() => {
+    const stageEl = stageRef.current;
+    if (!stageEl) return undefined;
+
+    const distanceBetween = (a: { x: number; y: number }, b: { x: number; y: number }) =>
+      Math.hypot(a.x - b.x, a.y - b.y);
+
+    const onPointerDown = (e: PointerEvent) => {
+      // Mouse drag at zoom>1 is handled by the image's `onMouseDown`
+      // → pan; mouse drag at zoom=1 has no useful gesture. Skip
+      // mouse here for everything.
+      if (e.pointerType === 'mouse') return;
+      if (entryKindRef.current !== 'image') return;
+
+      // Capture the pointer so subsequent `pointermove`/`pointerup`
+      // events are dispatched to the stage even when the finger
+      // leaves the stage rect (e.g. user pans onto the header).
+      // Without this, the gesture can be stranded mid-flight with
+      // a finger off-stage, leaving `pointerCacheRef` populated and
+      // breaking the next gesture. `try/catch` because some
+      // browsers reject capture on non-trusted events under
+      // automation harnesses.
+      try {
+        stageEl.setPointerCapture(e.pointerId);
+      } catch {
+        // ignore — capture is best-effort
+      }
+
+      pointerCacheRef.current.set(e.pointerId, { x: e.clientX, y: e.clientY });
+      const count = pointerCacheRef.current.size;
+
+      if (count === 2) {
+        // 2nd finger arrived — switch to pinch mode. Cancel any
+        // in-flight single-finger gesture (swipe / touch pan) so
+        // it doesn't leak state into the pinch.
+        const pts = Array.from(pointerCacheRef.current.values());
+        // Pinch midpoint → stage-center-relative anchor → derive
+        // the image-local point that's under the midpoint right
+        // now. Holding this point under the moving midpoint is
+        // what makes the zoom feel anchored.
+        const midClientX = (pts[0].x + pts[1].x) / 2;
+        const midClientY = (pts[0].y + pts[1].y) / 2;
+        const stageRect = stageEl.getBoundingClientRect();
+        const anchorX = midClientX - (stageRect.left + stageRect.width / 2);
+        const anchorY = midClientY - (stageRect.top + stageRect.height / 2);
+        const z0 = zoomRef.current;
+        const p0 = panRef.current;
+        pinchStateRef.current = {
+          initialDistance: distanceBetween(pts[0], pts[1]),
+          initialZoom: z0,
+          initialPan: { x: p0.x, y: p0.y },
+          anchorImageX: (anchorX - p0.x) / z0,
+          anchorImageY: (anchorY - p0.y) / z0,
+        };
+        if (swipeStateRef.current.active) {
+          swipeStateRef.current.active = false;
+          setSwipeOffset(0);
+        }
+        touchPanStateRef.current = null;
+        return;
+      }
+
+      if (count === 1) {
+        if (zoomRef.current === 1) {
+          // 1 finger + zoom=1 → swipe
+          const s = swipeStateRef.current;
+          s.active = true;
+          s.startX = e.clientX;
+          s.startY = e.clientY;
+          s.lastDx = 0;
+          s.claimed = 'none';
+          setSwipeOffset(0);
+        } else {
+          // 1 finger + zoom>1 → touch pan
+          touchPanStateRef.current = {
+            startX: e.clientX,
+            startY: e.clientY,
+            panAtStartX: panRef.current.x,
+            panAtStartY: panRef.current.y,
+          };
+        }
+      }
+    };
+
+    const onPointerMove = (e: PointerEvent) => {
+      if (e.pointerType === 'mouse') return;
+      const cached = pointerCacheRef.current.get(e.pointerId);
+      if (!cached) return;
+      pointerCacheRef.current.set(e.pointerId, { x: e.clientX, y: e.clientY });
+
+      // Pinch always wins when 2 fingers are down.
+      if (pinchStateRef.current && pointerCacheRef.current.size === 2) {
+        if (e.cancelable) e.preventDefault();
+        const ps = pinchStateRef.current;
+        const pts = Array.from(pointerCacheRef.current.values());
+        const dist = distanceBetween(pts[0], pts[1]);
+        const ratio = dist / ps.initialDistance;
+        const nextZoom = Math.max(
+          MIN_ZOOM,
+          Math.min(maxZoomRef.current, ps.initialZoom * ratio)
+        );
+        // Anchor-aware: hold `anchorImage*` (image-local point
+        // captured at pinch start) under the moving midpoint.
+        const midClientX = (pts[0].x + pts[1].x) / 2;
+        const midClientY = (pts[0].y + pts[1].y) / 2;
+        const stageRect = stageEl.getBoundingClientRect();
+        const anchorStageX = midClientX - (stageRect.left + stageRect.width / 2);
+        const anchorStageY = midClientY - (stageRect.top + stageRect.height / 2);
+        const rawPan = {
+          x: anchorStageX - ps.anchorImageX * nextZoom,
+          y: anchorStageY - ps.anchorImageY * nextZoom,
+        };
+        const clamped = clampPanRef.current(rawPan, nextZoom);
+        panRef.current = clamped;
+        setPan(clamped);
+        setZoomRef.current(nextZoom);
+        return;
+      }
+
+      // Single-finger paths run only if pinch isn't active.
+      if (pinchStateRef.current) return;
+
+      // Swipe path (zoom=1).
+      const s = swipeStateRef.current;
+      if (s.active) {
+        const dx = e.clientX - s.startX;
+        const dy = e.clientY - s.startY;
+        s.lastDx = dx;
+        if (s.claimed === 'none') {
+          if (Math.abs(dx) > 8 || Math.abs(dy) > 8) {
+            if (Math.abs(dx) > Math.abs(dy) * SWIPE_DOMINANCE) s.claimed = 'swipe';
+            else s.claimed = 'reject';
+          }
+        }
+        if (s.claimed !== 'swipe') return;
+        if (e.cancelable) e.preventDefault();
+        const blocked =
+          (dx > 0 && !canPrevRef.current) || (dx < 0 && !canNextRef.current);
+        setSwipeOffset(blocked ? dx / 3 : dx);
+        return;
+      }
+
+      // Touch-pan path (zoom>1).
+      const tp = touchPanStateRef.current;
+      if (tp) {
+        if (e.cancelable) e.preventDefault();
+        const raw = {
+          x: tp.panAtStartX + (e.clientX - tp.startX),
+          y: tp.panAtStartY + (e.clientY - tp.startY),
+        };
+        const clamped = clampPanRef.current(raw, zoomRef.current);
+        panRef.current = clamped;
+        setPan(clamped);
+      }
+    };
+
+    const onPointerEnd = (e: PointerEvent) => {
+      if (e.pointerType === 'mouse') return;
+      // Release the pointer capture taken on `pointerdown`. Browsers
+      // auto-release on `pointerup` / `pointercancel`, but
+      // explicitly calling it on a non-captured pointer is a no-op
+      // and the explicit release pairs the API call for clarity.
+      try {
+        if (stageEl.hasPointerCapture(e.pointerId)) {
+          stageEl.releasePointerCapture(e.pointerId);
+        }
+      } catch {
+        // ignore
+      }
+      pointerCacheRef.current.delete(e.pointerId);
+      const count = pointerCacheRef.current.size;
+
+      if (pinchStateRef.current && count < 2) {
+        // Pinch ended. Reset pinch state but DON'T immediately
+        // switch back to swipe / pan — the remaining finger (if
+        // any) needs to lift first and the user starts a fresh
+        // gesture cleanly.
+        pinchStateRef.current = null;
+        return;
+      }
+
+      if (count === 0) {
+        const s = swipeStateRef.current;
+        if (s.active) {
+          s.active = false;
+          if (s.claimed === 'swipe') {
+            if (s.lastDx > SWIPE_COMMIT_PX && canPrevRef.current) stepRef.current(-1);
+            else if (s.lastDx < -SWIPE_COMMIT_PX && canNextRef.current) stepRef.current(1);
+          }
+          setSwipeOffset(0);
+        }
+        touchPanStateRef.current = null;
+      }
+    };
+
+    // Desktop wheel zoom — anchored to the cursor (the same point
+    // under the cursor stays under the cursor as zoom changes).
+    // Element-Web's idiom: step factor proportional to `deltaY`,
+    // sign inverted so wheel-up zooms IN. `passive: false` so we
+    // can preventDefault the browser's native page scroll.
+    //
+    // Browsers send wheel events with very different `deltaMode`
+    // values (pixel vs line vs page). We normalize to pixels.
+    const onWheel = (e: WheelEvent) => {
+      if (entryKindRef.current !== 'image') return;
+      e.preventDefault();
+      const dyPx =
+        e.deltaMode === 1 ? e.deltaY * 16 // line mode → ~16px / line
+        : e.deltaMode === 2 ? e.deltaY * stageEl.clientHeight // page mode
+        : e.deltaY;
+      const factor = Math.exp(-dyPx * 0.0025);
+      const oldZoom = zoomRef.current;
+      const nextZoomRaw = oldZoom * factor;
+      const nextZoom = Math.max(MIN_ZOOM, Math.min(maxZoomRef.current, nextZoomRaw));
+      if (nextZoom === oldZoom) return;
+      // Anchor-aware: hold the image-local point under the cursor
+      // fixed across the zoom change.
+      const stageRect = stageEl.getBoundingClientRect();
+      const anchorX = e.clientX - (stageRect.left + stageRect.width / 2);
+      const anchorY = e.clientY - (stageRect.top + stageRect.height / 2);
+      const p0 = panRef.current;
+      const imageX = (anchorX - p0.x) / oldZoom;
+      const imageY = (anchorY - p0.y) / oldZoom;
+      const rawPan = {
+        x: anchorX - imageX * nextZoom,
+        y: anchorY - imageY * nextZoom,
+      };
+      const clamped = clampPanRef.current(rawPan, nextZoom);
+      panRef.current = clamped;
+      setPan(clamped);
+      setZoomRef.current(nextZoom);
+    };
+
+    stageEl.addEventListener('pointerdown', onPointerDown);
+    stageEl.addEventListener('pointermove', onPointerMove);
+    stageEl.addEventListener('pointerup', onPointerEnd);
+    stageEl.addEventListener('pointercancel', onPointerEnd);
+    stageEl.addEventListener('wheel', onWheel, { passive: false });
+
+    return () => {
+      stageEl.removeEventListener('pointerdown', onPointerDown);
+      stageEl.removeEventListener('pointermove', onPointerMove);
+      stageEl.removeEventListener('pointerup', onPointerEnd);
+      stageEl.removeEventListener('pointercancel', onPointerEnd);
+      stageEl.removeEventListener('wheel', onWheel);
+    };
+  }, []);
+
+  const effectiveSrc = srcState.status === AsyncStatus.Success ? srcState.data : undefined;
+  const isLoading =
+    !effectiveSrc &&
+    (srcState.status === AsyncStatus.Loading || srcState.status === AsyncStatus.Idle);
+  const isError = !effectiveSrc && srcState.status === AsyncStatus.Error;
+  const isReady = !!effectiveSrc;
+
+  const handleDownload = useCallback(async () => {
+    if (!effectiveSrc) return;
+    FileSaver.saveAs(effectiveSrc, entry.body);
+  }, [effectiveSrc, entry.body]);
+
+  // Video autoplay: the `<video>` element below ships with the
+  // `muted` + `autoPlay` HTML attributes. Muted autoplay is
+  // universally allowed (no gesture-activation needed) across
+  // Chrome / Firefox / WebKit / Android Capacitor / iOS Safari —
+  // see the MDN Autoplay guide. Unmuted autoplay would require a
+  // fresh user-gesture per video, which fails on iOS Safari
+  // after the first step in this viewer. Element-Web takes the
+  // same `muted` stance for the same reason. The user can unmute
+  // via the native `<video>` controls; subsequent videos in the
+  // same viewer session keep their unmuted state inherited
+  // through the element remount (key={entry.eventId} forces a
+  // fresh element, so unmute does NOT persist — accepted UX
+  // trade for "playing always works").
+
+  const transform = isReady
+    ? `translate3d(${swipeOffset + pan.x}px, ${pan.y}px, 0) scale(${zoom})`
+    : undefined;
+  const imageTransition = swipeOffset === 0 ? 'transform 200ms cubic-bezier(0.32, 0.72, 0, 1)' : 'none';
+
+  return (
+    <div className={css.root}>
+      <PageHeader
+        // `PageHeader` (= folds `Header size="600"`) so the header's
+        // bottom rule aligns to the pixel with the chat header to the
+        // left across the 12px void gap — same trick the profile
+        // side pane uses. A plain `<div height: 48px>` was visibly
+        // shorter and broke the cross-pane seam.
+        className={`${ContainerColor({ variant: 'Surface' })} ${css.header}`}
+      >
+        <Box grow="Yes" alignItems="Center" gap="200">
+          <Text className={css.title} size="H4" truncate>
+            {entry.body}
+          </Text>
+          {mediaSet.length > 1 && (
+            <span className={css.indexPill} aria-hidden="true">
+              {currentIndex >= 0 ? currentIndex + 1 : '?'} / {mediaSet.length}
+            </span>
+          )}
+        </Box>
+        <Box shrink="No" alignItems="Center" gap="100">
+          {entry.kind === 'image' && (
+            <>
+              <IconButton
+                variant="Background"
+                fill="None"
+                radii="300"
+                size="300"
+                onClick={zoomOut}
+                aria-label={t('MediaViewer.zoom_out', 'Zoom out')}
+              >
+                <Icon size="100" src={Icons.Minus} />
+              </IconButton>
+              <Text size="B300" as="span" style={{ minWidth: '3em', textAlign: 'center' }}>
+                {Math.round(zoom * 100)}%
+              </Text>
+              <IconButton
+                variant="Background"
+                fill="None"
+                radii="300"
+                size="300"
+                onClick={zoomIn}
+                aria-label={t('MediaViewer.zoom_in', 'Zoom in')}
+              >
+                <Icon size="100" src={Icons.Plus} />
+              </IconButton>
+            </>
+          )}
+          <IconButton
+            variant="Background"
+            fill="None"
+            radii="300"
+            size="300"
+            onClick={handleDownload}
+            disabled={!isReady}
+            aria-label={t('MediaViewer.download', 'Download')}
+          >
+            <Icon size="100" src={Icons.Download} />
+          </IconButton>
+          {entry.kind === 'image' && (
+            <div className={css.headerSeparator} aria-hidden="true" />
+          )}
+          <IconButton
+            variant="Background"
+            fill="None"
+            radii="300"
+            size="300"
+            onClick={requestClose}
+            aria-label={t('Common.close', 'Close')}
+          >
+            <Icon size="100" src={Icons.Cross} />
+          </IconButton>
+        </Box>
+      </PageHeader>
+
+      <div className={css.stage} ref={stageRef}>
+        {isReady && effectiveSrc && entry.kind === 'image' && (
+          <img
+            ref={imgRef}
+            src={effectiveSrc}
+            alt={entry.body}
+            className={css.image}
+            draggable={false}
+            style={{
+              transform,
+              cursor: panCursor,
+              transition: imageTransition,
+            }}
+            onMouseDown={onMouseDown}
+            onLoad={onImageLoad}
+          />
+        )}
+        {isReady && effectiveSrc && entry.kind === 'video' && (
+          <video
+            ref={videoRef}
+            key={entry.eventId}
+            src={effectiveSrc}
+            title={entry.body}
+            className={css.image}
+            controls
+            // Muted + autoPlay: see the rationale comment above
+            // — muted autoplay is the only policy-compliant way
+            // to start every video automatically including on iOS
+            // Safari after a swipe / chevron step. User unmutes
+            // via native controls.
+            muted
+            autoPlay
+            playsInline
+            style={{
+              transform: `translate3d(${swipeOffset}px, 0, 0)`,
+              transition: imageTransition,
+            }}
+          />
+        )}
+
+        {isLoading && (
+          <div className={css.stateOverlay}>
+            <Spinner variant="Secondary" size="600" />
+          </div>
+        )}
+        {isError && (
+          <div className={css.stateOverlay} style={{ pointerEvents: 'auto' }}>
+            <IconButton
+              variant="Critical"
+              fill="Soft"
+              radii="300"
+              size="400"
+              onClick={() => loadSrc()}
+              aria-label={t('Common.retry', 'Retry')}
+            >
+              <Icon size="200" src={Icons.Warning} />
+            </IconButton>
+          </div>
+        )}
+
+        {mediaSet.length > 1 && (
+          <>
+            <button
+              type="button"
+              className={classNames(css.chevron, css.chevronLeft)}
+              aria-label={t('MediaViewer.previous', 'Previous')}
+              // `aria-disabled` (not `disabled`) so the CSS dimmed-
+              // state selector still matches; the onClick already
+              // bails internally via `step`'s bounds check.
+              aria-disabled={!canPrev}
+              onClick={() => step(-1)}
+            >
+              <Icon size="200" src={Icons.ChevronLeft} />
+            </button>
+            <button
+              type="button"
+              className={classNames(css.chevron, css.chevronRight)}
+              aria-label={t('MediaViewer.next', 'Next')}
+              aria-disabled={!canNext}
+              onClick={() => step(1)}
+            >
+              <Icon size="200" src={Icons.ChevronRight} />
+            </button>
+          </>
+        )}
+      </div>
+
+    </div>
+  );
+}
diff --git a/src/app/features/room/MobileMediaViewerHorseshoe.css.ts b/src/app/features/room/MobileMediaViewerHorseshoe.css.ts
new file mode 100644
index 00000000..90978e62
--- /dev/null
+++ b/src/app/features/room/MobileMediaViewerHorseshoe.css.ts
@@ -0,0 +1,107 @@
+import { style } from '@vanilla-extract/css';
+import { color, toRem } from 'folds';
+import { VOJO_HORSESHOE_GAP_PX, VOJO_HORSESHOE_RADIUS_PX } from '../../styles/horseshoe';
+
+// Re-exported so the TSX can pick up the constants without crossing
+// the vanilla-extract / runtime boundary twice.
+export const HORSESHOE_RADIUS_PX = VOJO_HORSESHOE_RADIUS_PX;
+export const HORSESHOE_GAP_PX = VOJO_HORSESHOE_GAP_PX;
+
+// Outer container — anchor for the two absolutely-positioned panes
+// (`appBody` and `silhouette`). Same shape as the settings-sheet
+// container, see that file for the full rationale.
+export const container = style({
+  position: 'relative',
+  display: 'flex',
+  flex: 1,
+  flexDirection: 'column',
+  minWidth: 0,
+  minHeight: 0,
+  overflow: 'hidden',
+});
+
+// Holds the wrapped chat column. Stays put — clip-path carves the
+// bottom edge so virtualized timeline rows aren't re-measured
+// mid-gesture. `backgroundColor` opaque so the void colour painted
+// on the container doesn't bleed through any transparent slivers
+// in the wrapped tree (e.g. between bubble rows).
+export const appBody = style({
+  position: 'absolute',
+  top: 0,
+  left: 0,
+  right: 0,
+  bottom: 0,
+  display: 'flex',
+  flexDirection: 'column',
+  minWidth: 0,
+  minHeight: 0,
+  backgroundColor: color.Background.Container,
+  willChange: 'clip-path',
+});
+
+// The viewer sheet's surface. `Background.Container` (deepest Dawn
+// tone) so the image floats on a near-black backdrop — closest to
+// the legacy `<Overlay>` viewer's full-screen dark scrim.
+export const silhouette = style({
+  position: 'absolute',
+  bottom: 0,
+  left: 0,
+  right: 0,
+  display: 'flex',
+  flexDirection: 'column',
+  overflow: 'hidden',
+  backgroundColor: color.Background.Container,
+  willChange: 'height, border-top-left-radius, border-top-right-radius',
+});
+
+// Top-anchored — handle + viewer body reveal top-down as silhouette
+// grows. `padding-bottom: env(safe-area-inset-bottom)` keeps the
+// download / share row clear of the Android nav bar.
+export const panelContent = style({
+  position: 'absolute',
+  top: 0,
+  left: 0,
+  right: 0,
+  boxSizing: 'border-box',
+  display: 'flex',
+  flexDirection: 'column',
+  paddingBottom: 'env(safe-area-inset-bottom, 0px)',
+});
+
+// 20px drag-to-close band. The ONLY drag-to-close origin — touches
+// on the viewer body below this strip drive zoom / pan / swipe and
+// MUST NOT initiate a close gesture. `touchAction: none` blocks
+// browser-native scroll on this strip.
+export const panelHandle = style({
+  flexShrink: 0,
+  height: toRem(20),
+  display: 'flex',
+  alignItems: 'center',
+  justifyContent: 'center',
+  cursor: 'grab',
+  touchAction: 'none',
+  userSelect: 'none',
+  selectors: {
+    '&:active': { cursor: 'grabbing' },
+  },
+});
+
+// Darker than `Background.Container` would make the grabber
+// disappear — use `SurfaceVariant.Container` so the bar reads
+// against the dark viewer bg. Mirror of the settings handle's
+// step-up against `SurfaceVariant.Container` (where it uses the
+// deeper `Background.Container`).
+export const panelHandleBar = style({
+  width: toRem(36),
+  height: toRem(4),
+  borderRadius: toRem(4),
+  backgroundColor: color.SurfaceVariant.Container,
+});
+
+export const panelBody = style({
+  flex: 1,
+  display: 'flex',
+  flexDirection: 'column',
+  minHeight: 0,
+  minWidth: 0,
+});
diff --git a/src/app/features/room/MobileMediaViewerHorseshoe.tsx b/src/app/features/room/MobileMediaViewerHorseshoe.tsx
new file mode 100644
index 00000000..a54fe5b9
--- /dev/null
+++ b/src/app/features/room/MobileMediaViewerHorseshoe.tsx
@@ -0,0 +1,419 @@
+// Bottom-up «horseshoe» sheet that wraps the mobile chat column for
+// the media viewer. Sister of `MobileSettingsHorseshoe` (which wraps
+// the DM list with the same idiom for the Settings tree) — same
+// clip-path carve, same VAUL easing, same Strict-Mode-safe entry
+// animation, same `keepMounted` delayed unmount.
+//
+// Differences from the settings sheet:
+//   • Tap-to-open only. No drag-up origin — opening is driven from
+//     `useOpenMediaViewer` called by `ImageContent.onClick`. The
+//     20px handle band is the ONLY drag-sensitive area; touches on
+//     the viewer body below run zoom / pan / horizontal swipe
+//     instead.
+//   • `RAIL_FRACTION = 3 / 4` (75% of viewport) vs settings's 2/3.
+//   • Silhouette bg = `Background.Container` (deepest Dawn tone) for
+//     a dark image backdrop, vs settings's `SurfaceVariant.Container`.
+//   • Wrapped content is the chat column, not the DM list.
+//   • No `--vojo-safe-top` reset — there's no PageNav inside the
+//     viewer body that would inherit it; the viewer body has its own
+//     padding policy.
+
+import React, { ReactNode, useEffect, useLayoutEffect, useRef, useState } from 'react';
+import { createPortal } from 'react-dom';
+import { useAtomValue } from 'jotai';
+import { useTranslation } from 'react-i18next';
+import { mediaViewerAtom } from '../../state/mediaViewer';
+import { useCloseMediaViewer } from '../../state/hooks/mediaViewer';
+import { ScreenSize, useScreenSizeContext } from '../../hooks/useScreenSize';
+import { VOJO_HORSESHOE_VOID_COLOR } from '../../styles/horseshoe';
+import { MediaViewerBody } from './MediaViewerBody';
+import * as css from './MobileMediaViewerHorseshoe.css';
+
+const VAUL_EASING = 'cubic-bezier(0.32, 0.72, 0, 1)';
+const ANIMATION_MS = 250;
+// Drag distance past which release commits the close. Matches the
+// settings sheet's 80px (the only commit gesture there too).
+const COMMIT_THRESHOLD_PX = 80;
+// Sheet height as a fraction of the viewport — product spec.
+// Capped at runtime by the wrapper container's actual height
+// (see `railHeightPx` calc below) so it can't spill past the
+// chat header when chrome outside the chat column (e.g.
+// `HorseshoeContainer.bottomRail` during a call) shrinks the
+// wrapper.
+const RAIL_VIEWPORT_FRACTION = 0.8;
+// Absolute floor for tiny / split-screen viewports — the sheet
+// becomes useless below this. NB: floor cannot exceed the
+// container's ceiling (`Math.min` is applied last), so in extreme
+// squeeze cases (split-screen, oversized call rail) the sheet
+// silently shrinks below the floor rather than clipping the chat
+// header.
+const RAIL_MIN_PX = 200;
+// Headroom reserved at the TOP of the container for the chat
+// header (`PageHeader` = folds `Header size="600"`, ~64px) plus
+// the 12px horseshoe void gap above the silhouette's rounded
+// TL/TR carves. With this in place the sheet's top edge stops
+// just below the chat header even when the wrapper container
+// shrinks (e.g. `HorseshoeContainer.bottomRail` is on during a
+// call). Previously the headroom was 8px which let the sheet
+// climb over the header in tight viewports.
+const CHAT_HEADER_RESERVED_PX = 76;
+// Same emerge window as the settings sheet so the silhouette's
+// rounding + void gap finish exactly when the gesture qualifies to
+// commit.
+const HORSESHOE_EMERGE_PX = 80;
+
+// Symmetric cubic in-out — slow start, fast middle, slow finish.
+// Same curve the settings / profile horseshoes use for their emerge
+// (rationale at the top of `MobileSettingsHorseshoe.tsx`).
+const easeInOutCubic = (t: number): number =>
+  t < 0.5 ? 4 * t * t * t : 1 - ((-2 * t + 2) ** 3) / 2;
+
+type DragState = {
+  inputType: 'touch' | 'pointer';
+  startY: number;
+  deltaY: number;
+};
+
+type MobileMediaViewerHorseshoeProps = {
+  children: ReactNode;
+};
+
+function MobileMediaViewerHorseshoeImpl({ children }: MobileMediaViewerHorseshoeProps) {
+  const { t } = useTranslation();
+  const entry = useAtomValue(mediaViewerAtom);
+  const closeSheet = useCloseMediaViewer();
+
+  const [drag, setDrag] = useState<DragState | null>(null);
+  const [viewportHeight, setViewportHeight] = useState(() =>
+    typeof window === 'undefined' ? 800 : window.innerHeight
+  );
+
+  useEffect(() => {
+    const onResize = () => setViewportHeight(window.innerHeight);
+    window.addEventListener('resize', onResize);
+    return () => window.removeEventListener('resize', onResize);
+  }, []);
+
+  // Measure the wrapper container directly via `ResizeObserver` so
+  // the rail height tracks whatever vertical space is actually
+  // available to the chat column. The viewport-based fraction is
+  // the *intent* (90% of the screen — product spec), but
+  // `HorseshoeContainer.bottomRail` (the call rail) sits OUTSIDE
+  // the chat column inside the same viewport and shrinks our
+  // wrapper when active. Without the container clamp, 90% of the
+  // viewport could exceed the wrapper's height, and the sheet
+  // would render up to / past the chat header which sits inside
+  // the wrapper's appBody.
+  const containerRef = useRef<HTMLDivElement>(null);
+  const [containerHeight, setContainerHeight] = useState(() =>
+    typeof window === 'undefined' ? 800 : window.innerHeight
+  );
+
+  useLayoutEffect(() => {
+    const el = containerRef.current;
+    if (!el) return undefined;
+    setContainerHeight(el.clientHeight);
+    const ro = new ResizeObserver((entries) => {
+      const cr = entries[0]?.contentRect;
+      if (cr) setContainerHeight(cr.height);
+    });
+    ro.observe(el);
+    return () => ro.disconnect();
+  }, []);
+
+  // Final rail = viewport × fraction, then floor-then-ceiling
+  // clamped. Container (minus chat-header reserve) is the HARD
+  // ceiling: when the wrapper shrinks (call rail / split-screen)
+  // the sheet shrinks below `RAIL_MIN_PX` rather than spilling
+  // over the chat header. Order matters — `Math.min` last makes
+  // container win against floor.
+  const idealRailPx = Math.round(viewportHeight * RAIL_VIEWPORT_FRACTION);
+  const railHeightPx = Math.min(
+    Math.max(0, containerHeight - CHAT_HEADER_RESERVED_PX),
+    Math.max(RAIL_MIN_PX, idealRailPx)
+  );
+  const open = !!entry;
+
+  // Entry-animation gate — see `MobileSettingsHorseshoe` for the
+  // full rationale. `hasEnteredRef` mirror guards the atom-clearing
+  // unmount cleanup against React 18 strict-mode dev rehearsal.
+  const [hasEntered, setHasEntered] = useState(false);
+  const hasEnteredRef = useRef(false);
+  useLayoutEffect(() => {
+    const id = requestAnimationFrame(() => {
+      hasEnteredRef.current = true;
+      setHasEntered(true);
+    });
+    return () => cancelAnimationFrame(id);
+  }, []);
+
+  // Keep the viewer body mounted through the 250ms exit slide so
+  // the user sees the image slide down with the sheet instead of
+  // an empty panel collapsing.
+  const [keepMounted, setKeepMounted] = useState(open);
+  useEffect(() => {
+    if (open) {
+      setKeepMounted(true);
+      return undefined;
+    }
+    const id = window.setTimeout(() => setKeepMounted(false), ANIMATION_MS);
+    return () => window.clearTimeout(id);
+  }, [open]);
+
+  // Persist the last opened entry through the exit animation so the
+  // body keeps rendering the image as it slides down.
+  const lastEntryRef = useRef(entry);
+  if (entry) lastEntryRef.current = entry;
+
+  const baseExpanded = open && hasEntered ? railHeightPx : 0;
+  // Drag is close-only (positive deltaY); clamp negative deltas to
+  // 0 so an upward reversal cancels rather than expanding past the
+  // open height.
+  const expandedPx = drag
+    ? Math.max(0, Math.min(railHeightPx, baseExpanded - drag.deltaY))
+    : baseExpanded;
+  const expandedFraction = railHeightPx > 0 ? expandedPx / railHeightPx : 0;
+  const isDragging = drag !== null;
+  const horseshoeActive = expandedPx > 0;
+
+  const handleRef = useRef<HTMLDivElement>(null);
+
+  const dragRef = useRef<DragState | null>(null);
+  dragRef.current = drag;
+  const entryRef = useRef(entry);
+  entryRef.current = entry;
+  const closeSheetRef = useRef(closeSheet);
+  closeSheetRef.current = closeSheet;
+
+  // Clear the atom if the wrapper unmounts while the sheet is open
+  // (route change). Strict-mode rehearsal guard via `hasEnteredRef`
+  // — without it, the cleanup fires before the entry rAF and a
+  // deep-link cold-start would clear the atom mid-rehearsal.
+  useEffect(
+    () => () => {
+      if (!hasEnteredRef.current) return;
+      if (entryRef.current) closeSheetRef.current();
+    },
+    []
+  );
+
+  // Hardware Escape (also dispatched by the global Android-back
+  // handler via the portal-marker pattern below). Skip inputs /
+  // contenteditable so typing inside a future caption field doesn't
+  // dismiss the sheet.
+  useEffect(() => {
+    if (!open) return undefined;
+    const onKeyDown = (e: KeyboardEvent) => {
+      if (e.key !== 'Escape') return;
+      const target = e.target as HTMLElement | null;
+      if (
+        target &&
+        (target.tagName === 'INPUT' ||
+          target.tagName === 'TEXTAREA' ||
+          target.isContentEditable)
+      ) {
+        return;
+      }
+      closeSheetRef.current();
+    };
+    window.addEventListener('keydown', onKeyDown);
+    return () => window.removeEventListener('keydown', onKeyDown);
+  }, [open]);
+
+  // Drag-to-close only — handle band at the top of the silhouette.
+  // No document-level open-drag listener; opening is tap-only via
+  // `useOpenMediaViewer`.
+  //
+  // CLAMP, not early-return — see `MobileSettingsHorseshoe` for
+  // the bug-fix rationale. Reversing the gesture must drag deltaY
+  // back toward 0, not leave the stale value.
+  useEffect(() => {
+    const handleEl = handleRef.current;
+    if (!handleEl) return undefined;
+
+    const applyMove = (clientY: number, e: TouchEvent | PointerEvent) => {
+      const d = dragRef.current;
+      if (!d) return;
+      const rawDelta = clientY - d.startY;
+      const nextDelta = Math.max(0, rawDelta);
+      if (e.cancelable) e.preventDefault();
+      setDrag({ ...d, deltaY: nextDelta });
+    };
+
+    const applyEnd = () => {
+      const d = dragRef.current;
+      if (!d) return;
+      if (d.deltaY > COMMIT_THRESHOLD_PX) {
+        closeSheetRef.current();
+      }
+      setDrag(null);
+    };
+
+    const onHandleTouchStart = (e: TouchEvent) => {
+      if (dragRef.current) return;
+      if (!entryRef.current) return;
+      const touch = e.touches[0];
+      setDrag({ inputType: 'touch', startY: touch.clientY, deltaY: 0 });
+    };
+    const onTouchMove = (e: TouchEvent) => {
+      const d = dragRef.current;
+      if (!d || d.inputType !== 'touch') return;
+      applyMove(e.touches[0].clientY, e);
+    };
+    const onTouchEnd = () => {
+      const d = dragRef.current;
+      if (!d || d.inputType !== 'touch') return;
+      applyEnd();
+    };
+
+    const onHandlePointerDown = (e: PointerEvent) => {
+      if (e.pointerType === 'touch') return;
+      if (dragRef.current) return;
+      if (!entryRef.current) return;
+      if (e.button !== 0) return;
+      setDrag({ inputType: 'pointer', startY: e.clientY, deltaY: 0 });
+    };
+    const onPointerMove = (e: PointerEvent) => {
+      if (e.pointerType === 'touch') return;
+      const d = dragRef.current;
+      if (!d || d.inputType !== 'pointer') return;
+      applyMove(e.clientY, e);
+    };
+    const onPointerEnd = (e: PointerEvent) => {
+      if (e.pointerType === 'touch') return;
+      const d = dragRef.current;
+      if (!d || d.inputType !== 'pointer') return;
+      applyEnd();
+    };
+
+    handleEl.addEventListener('touchstart', onHandleTouchStart, { passive: true });
+    document.addEventListener('touchmove', onTouchMove, { passive: false });
+    document.addEventListener('touchend', onTouchEnd, { passive: true });
+    document.addEventListener('touchcancel', onTouchEnd, { passive: true });
+    handleEl.addEventListener('pointerdown', onHandlePointerDown);
+    document.addEventListener('pointermove', onPointerMove, { passive: false });
+    document.addEventListener('pointerup', onPointerEnd, { passive: true });
+    document.addEventListener('pointercancel', onPointerEnd, { passive: true });
+
+    return () => {
+      handleEl.removeEventListener('touchstart', onHandleTouchStart);
+      document.removeEventListener('touchmove', onTouchMove);
+      document.removeEventListener('touchend', onTouchEnd);
+      document.removeEventListener('touchcancel', onTouchEnd);
+      handleEl.removeEventListener('pointerdown', onHandlePointerDown);
+      document.removeEventListener('pointermove', onPointerMove);
+      document.removeEventListener('pointerup', onPointerEnd);
+      document.removeEventListener('pointercancel', onPointerEnd);
+    };
+  }, []);
+
+  // Geometry — same emerge ramp as the settings / profile horseshoes
+  // (rationale at the top of `MobileSettingsHorseshoe`).
+  let horseshoeRamp: number;
+  if (isDragging) {
+    horseshoeRamp = easeInOutCubic(Math.min(1, expandedPx / HORSESHOE_EMERGE_PX));
+  } else {
+    horseshoeRamp = expandedFraction > 0 ? 1 : 0;
+  }
+  const silhouetteRadiusPx = horseshoeRamp * css.HORSESHOE_RADIUS_PX;
+  const appBodyRadiusPx = horseshoeRamp * css.HORSESHOE_RADIUS_PX;
+  const appBodyGapPx = horseshoeRamp * css.HORSESHOE_GAP_PX;
+  const appBodyMaskBottomPx = expandedPx + appBodyGapPx;
+
+  const appBodyClipPath = `inset(0px 0px ${appBodyMaskBottomPx}px 0px round 0px 0px ${appBodyRadiusPx}px ${appBodyRadiusPx}px)`;
+
+  const silhouetteTransition = isDragging
+    ? 'none'
+    : `height ${ANIMATION_MS}ms ${VAUL_EASING}, border-top-left-radius ${ANIMATION_MS}ms ${VAUL_EASING}, border-top-right-radius ${ANIMATION_MS}ms ${VAUL_EASING}`;
+  const appBodyTransition = isDragging ? 'none' : `clip-path ${ANIMATION_MS}ms ${VAUL_EASING}`;
+
+  const containerStyle: React.CSSProperties = {
+    backgroundColor: horseshoeActive ? VOJO_HORSESHOE_VOID_COLOR : undefined,
+  };
+
+  const renderEntry = entry ?? lastEntryRef.current;
+  const renderBody = !!renderEntry && (keepMounted || isDragging);
+
+  // Marker portal for the global Android-back handler — when
+  // `portalContainer.firstChild` matches our `data-vojo-…-active`
+  // attribute, the back-button dispatches an Escape keydown that
+  // the `keydown` effect above catches.
+  const portalTarget =
+    typeof document !== 'undefined'
+      ? document.getElementById('portalContainer') ?? document.body
+      : null;
+
+  return (
+    <div ref={containerRef} className={css.container} style={containerStyle}>
+      {open && portalTarget
+        ? createPortal(
+            <div
+              data-vojo-media-viewer-sheet-active="true"
+              aria-hidden="true"
+              style={{ display: 'none' }}
+            />,
+            portalTarget
+          )
+        : null}
+      <div
+        className={css.appBody}
+        style={{
+          clipPath: appBodyClipPath,
+          transition: appBodyTransition,
+          overscrollBehaviorY: 'contain',
+        }}
+      >
+        {children}
+      </div>
+
+      <div
+        className={css.silhouette}
+        style={{
+          height: `${expandedPx}px`,
+          borderTopLeftRadius: `${silhouetteRadiusPx}px`,
+          borderTopRightRadius: `${silhouetteRadiusPx}px`,
+          transition: silhouetteTransition,
+          visibility: expandedPx > 0 ? 'visible' : 'hidden',
+          // Reset `--vojo-safe-top` so any descendant that paints a
+          // `padding-top: var(--vojo-safe-top)` for status-bar safe-
+          // area (e.g. PageNav-style headers, if anyone adds one
+          // here later) doesn't push the viewer header down inside
+          // the sheet — the sheet itself sits below the status bar.
+          // Mirror of the settings horseshoe.
+          ['--vojo-safe-top' as string]: '0px',
+        }}
+        role="dialog"
+        aria-label={renderEntry?.body ?? t('MediaViewer.title', 'Media viewer')}
+      >
+        <div className={css.panelContent} style={{ height: `${railHeightPx}px` }}>
+          <div
+            ref={handleRef}
+            className={css.panelHandle}
+            aria-label={t('MediaViewer.drag_to_close', 'Drag down to close')}
+          >
+            <div className={css.panelHandleBar} />
+          </div>
+          <div className={css.panelBody}>
+            {renderBody && renderEntry && (
+              <MediaViewerBody entry={renderEntry} requestClose={() => closeSheetRef.current()} />
+            )}
+          </div>
+        </div>
+      </div>
+    </div>
+  );
+}
+
+// Top-level branch — pass through on non-mobile (the desktop right
+// pane handles the same atom over there). The mobile branch owns
+// the drag / atom hooks in a sub-component so we don't run them on
+// desktop renders.
+export function MobileMediaViewerHorseshoe({
+  children,
+}: MobileMediaViewerHorseshoeProps): React.ReactElement {
+  const isMobile = useScreenSizeContext() === ScreenSize.Mobile;
+  if (!isMobile) {
+    return children as React.ReactElement;
+  }
+  return <MobileMediaViewerHorseshoeImpl>{children}</MobileMediaViewerHorseshoeImpl>;
+}
diff --git a/src/app/features/room/Room.tsx b/src/app/features/room/Room.tsx
index 96486cdb..39956cd0 100644
--- a/src/app/features/room/Room.tsx
+++ b/src/app/features/room/Room.tsx
@@ -24,6 +24,10 @@ import { callChatAtom } from '../../state/callEmbed';
 import { CallChatView } from './CallChatView';
 import { RoomViewProfilePanel } from './RoomViewProfilePanel';
 import { RoomViewProfileSidePanel } from './RoomViewProfileSidePanel';
+import { MobileMediaViewerHorseshoe } from './MobileMediaViewerHorseshoe';
+import { RoomViewMediaSidePanel } from './RoomViewMediaSidePanel';
+import { MediaViewerHostContext } from './mediaViewerHostContext';
+import { mediaViewerAtom } from '../../state/mediaViewer';
 import { useChannelsMode, ThreadDrawerOpenProvider } from '../../hooks/useChannelsMode';
 import { CHANNELS_THREAD_PATH } from '../../pages/paths';
 import { getChannelsRoomPath } from '../../pages/pathUtils';
@@ -110,7 +114,15 @@ export function Room({ renderRoomView }: RoomProps) {
   // PageRoot. The chat column applies an explicit Background bg so the
   // parent void can't bleed through any transparent slivers.
   const profileOpen = !!useAtomValue(userRoomProfileAtom);
+  const mediaOpen = !!useAtomValue(mediaViewerAtom);
   const showProfileHorseshoe = profileOpen && !isMobile && !showThreadDrawer;
+  // Media viewer side pane on desktop — same horseshoe seam idiom
+  // as the profile pane. The two are mutually exclusive in practice
+  // (the open hooks clear the other atom), so at most one shows at
+  // a time; both feed into `showAnyHorseshoe` for the chat-column
+  // bg + the void-gap render gate.
+  const showMediaHorseshoe = mediaOpen && !isMobile && !showThreadDrawer;
+  const showAnyHorseshoe = showProfileHorseshoe || showMediaHorseshoe;
 
   useKeyDown(
     window,
@@ -138,13 +150,34 @@ export function Room({ renderRoomView }: RoomProps) {
 
   const callView = room.isCallRoom();
 
+  // Disable the atom-driven media viewer when the desktop thread
+  // drawer is open — the side-pane mount block below is gated on
+  // `!showThreadDrawer`, so the new viewer's right pane wouldn't
+  // render, and an atom-only open path would silently no-op on the
+  // user's tap. With the host context cleared, `ImageContent` /
+  // `VideoContent` fall back to the legacy full-screen `<Overlay>`
+  // modal, which sits on top of every pane and still works.
+  // Mobile is unaffected: `drawerHidesChat` hides the chat column
+  // entirely on mobile + thread, so there's no media to tap anyway.
+  //
+  // `useMemo` so the value object identity is stable across
+  // unrelated re-renders — without it, every parent re-render
+  // would spread a new `{ roomId }` literal through the context,
+  // re-running every `useContext(MediaViewerHostContext)` consumer
+  // (every image / video thumbnail in the timeline).
+  const mediaHostValue = React.useMemo(
+    () => (!isMobile && showThreadDrawer ? null : { roomId: room.roomId }),
+    [isMobile, showThreadDrawer, room.roomId]
+  );
+
   return (
     <PowerLevelsContextProvider value={powerLevels}>
       <ThreadDrawerOpenProvider value={showThreadDrawer}>
+        <MediaViewerHostContext.Provider value={mediaHostValue}>
         <Box
           grow="Yes"
           style={
-            showProfileHorseshoe
+            showAnyHorseshoe
               ? { backgroundColor: VOJO_HORSESHOE_VOID_COLOR }
               : undefined
           }
@@ -154,7 +187,7 @@ export function Room({ renderRoomView }: RoomProps) {
               grow="Yes"
               direction="Column"
               className={
-                showProfileHorseshoe
+                showAnyHorseshoe
                   ? ContainerColor({ variant: 'Background' })
                   : undefined
               }
@@ -162,11 +195,13 @@ export function Room({ renderRoomView }: RoomProps) {
               // `MobileProfileHorseshoe` owns the safe-top inset and bg.
               // See the !callView twin block below for the rationale.
             >
-              <RoomViewProfilePanel header={<RoomViewHeader callView />}>
-                <Box grow="Yes">
-                  <CallView />
-                </Box>
-              </RoomViewProfilePanel>
+              <MobileMediaViewerHorseshoe>
+                <RoomViewProfilePanel header={<RoomViewHeader callView />}>
+                  <Box grow="Yes">
+                    <CallView />
+                  </Box>
+                </RoomViewProfilePanel>
+              </MobileMediaViewerHorseshoe>
             </Box>
           )}
           {!callView && !drawerHidesChat && (
@@ -174,7 +209,7 @@ export function Room({ renderRoomView }: RoomProps) {
               grow="Yes"
               direction="Column"
               className={
-                showProfileHorseshoe
+                showAnyHorseshoe
                   ? ContainerColor({ variant: 'Background' })
                   : undefined
               }
@@ -187,23 +222,26 @@ export function Room({ renderRoomView }: RoomProps) {
               // drags. Desktop branch still works because `--vojo-safe-top`
               // resolves to 0 on web.
             >
-              <RoomViewProfilePanel header={<RoomViewHeader />}>
-                <Box grow="Yes">
-                  {renderRoomView?.({ eventId }) ?? <RoomView eventId={eventId} />}
-                </Box>
-              </RoomViewProfilePanel>
+              <MobileMediaViewerHorseshoe>
+                <RoomViewProfilePanel header={<RoomViewHeader />}>
+                  <Box grow="Yes">
+                    {renderRoomView?.({ eventId }) ?? <RoomView eventId={eventId} />}
+                  </Box>
+                </RoomViewProfilePanel>
+              </MobileMediaViewerHorseshoe>
             </Box>
           )}
 
-        {/* Tablet / Desktop: profile renders as a third pane to the
-            right of the chat. Mobile uses the top horseshoe inside
-              `RoomViewProfilePanel`, so we don't mount the side pane
-              there. The 12px void gap (same as page-nav <-> chat split
-              from PageRoot) sits between the chat column and the pane
-              so the seam reads identically to the rest of the app. */}
+        {/* Tablet / Desktop: profile or media renders as a third pane
+            to the right of the chat. Mobile uses the top horseshoe
+            (profile) and the bottom horseshoe (media), so we don't
+            mount the side panes there. The 12px void gap sits between
+            the chat column and whichever pane is open — both panes
+            share the same gap geometry since they're mutually
+            exclusive via the open hooks. */}
           {!isMobile && !showThreadDrawer && (
             <>
-              {showProfileHorseshoe && (
+              {showAnyHorseshoe && (
                 <Box
                   shrink="No"
                   style={{
@@ -213,6 +251,7 @@ export function Room({ renderRoomView }: RoomProps) {
                 />
               )}
               <RoomViewProfileSidePanel />
+              <RoomViewMediaSidePanel />
             </>
           )}
 
@@ -248,6 +287,7 @@ export function Room({ renderRoomView }: RoomProps) {
             />
           )}
         </Box>
+        </MediaViewerHostContext.Provider>
       </ThreadDrawerOpenProvider>
     </PowerLevelsContextProvider>
   );
diff --git a/src/app/features/room/RoomViewMediaSidePanel.css.ts b/src/app/features/room/RoomViewMediaSidePanel.css.ts
new file mode 100644
index 00000000..a7d80426
--- /dev/null
+++ b/src/app/features/room/RoomViewMediaSidePanel.css.ts
@@ -0,0 +1,38 @@
+import { style } from '@vanilla-extract/css';
+import { color, toRem } from 'folds';
+import { VOJO_HORSESHOE_RADIUS_PX } from '../../styles/horseshoe';
+
+// Right-side media pane. Much wider than the profile pane —
+// images need room to read comfortably. `clamp(480px, 50vw, 880px)`
+// gives a generous minimum on narrow desktops while capping width
+// on ultra-wide displays.
+//
+// Left edge rounded (TL + BL) to carve across the 12px horseshoe
+// void gap rendered by `Room.tsx` — same design language as the
+// profile pane and the page-nav <-> chat split.
+//
+// `Background.Container` (#0d0e11) chosen for the dark image
+// backdrop — same logic as the mobile silhouette bg.
+export const panel = style({
+  flexShrink: 0,
+  width: `clamp(${toRem(480)}, 50vw, ${toRem(880)})`,
+  display: 'flex',
+  flexDirection: 'column',
+  backgroundColor: color.Background.Container,
+  minHeight: 0,
+  overflow: 'hidden',
+  borderTopLeftRadius: toRem(VOJO_HORSESHOE_RADIUS_PX),
+  borderBottomLeftRadius: toRem(VOJO_HORSESHOE_RADIUS_PX),
+});
+
+// Body fills the remaining vertical space below `MediaViewerBody`'s
+// own header. The pane has no separate `PageHeader` strip — the
+// viewer body owns its chrome (close button, title, action row) so
+// the header sits flush with the rounded TL corner.
+export const body = style({
+  flex: 1,
+  display: 'flex',
+  flexDirection: 'column',
+  minHeight: 0,
+  minWidth: 0,
+});
diff --git a/src/app/features/room/RoomViewMediaSidePanel.tsx b/src/app/features/room/RoomViewMediaSidePanel.tsx
new file mode 100644
index 00000000..9c6aa246
--- /dev/null
+++ b/src/app/features/room/RoomViewMediaSidePanel.tsx
@@ -0,0 +1,66 @@
+// Desktop / tablet right-side media pane. Renders the same
+// `MediaViewerBody` the mobile bottom-up horseshoe shows, but as a
+// flex sibling next to the chat column instead of a slide-up rail.
+//
+// Mounted in `Room.tsx` only on non-mobile screens; mobile uses
+// `MobileMediaViewerHorseshoe` instead.
+
+import React, { useRef } from 'react';
+import { useAtomValue } from 'jotai';
+import { useTranslation } from 'react-i18next';
+import FocusTrap from 'focus-trap-react';
+import { mediaViewerAtom } from '../../state/mediaViewer';
+import { useCloseMediaViewer } from '../../state/hooks/mediaViewer';
+import { stopPropagation } from '../../utils/keyboard';
+import { MediaViewerBody } from './MediaViewerBody';
+import * as css from './RoomViewMediaSidePanel.css';
+
+export function RoomViewMediaSidePanel() {
+  const { t } = useTranslation();
+  const entry = useAtomValue(mediaViewerAtom);
+  const close = useCloseMediaViewer();
+
+  const open = !!entry;
+
+  const entryRef = useRef(entry);
+  entryRef.current = entry;
+
+  if (!open || !entry) return null;
+
+  return (
+    <FocusTrap
+      focusTrapOptions={{
+        // Move focus into the trap on open so screen readers
+        // announce the dialog and keyboard arrow keys / Esc reach
+        // the viewer body. Falls back to first focusable element
+        // (the close button in the body header) without us pinning
+        // a specific ref — simpler than threading one down.
+        initialFocus: undefined,
+        // Outside clicks pass through to chat (`allowOutsideClick`)
+        // but don't close — clicking around in chat to scroll /
+        // compose shouldn't dismiss the viewer. Explicit close via
+        // the body's × button or `Esc`. Mirror of the profile side
+        // pane's stance.
+        clickOutsideDeactivates: false,
+        allowOutsideClick: () => true,
+        escapeDeactivates: stopPropagation,
+        onDeactivate: () => {
+          if (entryRef.current) close();
+        },
+        checkCanFocusTrap: () => Promise.resolve(),
+      }}
+      active={open}
+    >
+      <div
+        className={css.panel}
+        role="dialog"
+        aria-modal="true"
+        aria-label={entry.body || t('MediaViewer.title', 'Media viewer')}
+      >
+        <div className={css.body}>
+          <MediaViewerBody entry={entry} requestClose={close} />
+        </div>
+      </div>
+    </FocusTrap>
+  );
+}
diff --git a/src/app/features/room/mediaViewerHostContext.ts b/src/app/features/room/mediaViewerHostContext.ts
new file mode 100644
index 00000000..234d49c3
--- /dev/null
+++ b/src/app/features/room/mediaViewerHostContext.ts
@@ -0,0 +1,20 @@
+import { createContext, useContext } from 'react';
+
+// Non-null only when an atom-driven media-viewer shell (the mobile
+// bottom-up horseshoe or the desktop right side pane) is mounted in
+// the React tree above this consumer. `ImageContent` reads this to
+// decide between the new atom-open path (Room timeline) and the
+// legacy local-state `<Overlay>` viewer (pin-menu, message search
+// — surfaces where no shell is mounted). Carries `roomId` so the
+// consumer doesn't need a separate `useRoom()` call (which throws
+// outside `RoomProvider`).
+//
+// Defaults to `null` so unchanged callers keep their current
+// behaviour; `Room.tsx` provides a non-null value for the chat
+// column.
+export type MediaViewerHostValue = { roomId: string } | null;
+
+export const MediaViewerHostContext = createContext<MediaViewerHostValue>(null);
+
+export const useMediaViewerHost = (): MediaViewerHostValue =>
+  useContext(MediaViewerHostContext);
diff --git a/src/app/state/hooks/mediaViewer.ts b/src/app/state/hooks/mediaViewer.ts
new file mode 100644
index 00000000..650d48f2
--- /dev/null
+++ b/src/app/state/hooks/mediaViewer.ts
@@ -0,0 +1,28 @@
+import { useCallback } from 'react';
+import { useSetAtom } from 'jotai';
+import { mediaViewerAtom, MediaViewerEntry } from '../mediaViewer';
+import { userRoomProfileAtom } from '../userRoomProfile';
+
+export const useOpenMediaViewer = (): ((entry: MediaViewerEntry) => void) => {
+  const setMedia = useSetAtom(mediaViewerAtom);
+  // Close the profile side pane / horseshoe when opening media —
+  // having both open simultaneously fights for chat-column width on
+  // desktop and stacks two horseshoe surfaces on mobile. Mutual
+  // exclusivity is the simplest contract until the user asks for
+  // stacking.
+  const setProfile = useSetAtom(userRoomProfileAtom);
+  return useCallback(
+    (entry) => {
+      setProfile(undefined);
+      setMedia(entry);
+    },
+    [setMedia, setProfile]
+  );
+};
+
+export const useCloseMediaViewer = (): (() => void) => {
+  const setMedia = useSetAtom(mediaViewerAtom);
+  return useCallback(() => {
+    setMedia(undefined);
+  }, [setMedia]);
+};
diff --git a/src/app/state/hooks/userRoomProfile.ts b/src/app/state/hooks/userRoomProfile.ts
index 7fed0c08..00652e0e 100644
--- a/src/app/state/hooks/userRoomProfile.ts
+++ b/src/app/state/hooks/userRoomProfile.ts
@@ -2,6 +2,7 @@ import { useCallback } from 'react';
 import { useAtomValue, useSetAtom } from 'jotai';
 import { Position, RectCords } from 'folds';
 import { userRoomProfileAtom, UserRoomProfileState } from '../userRoomProfile';
+import { mediaViewerAtom } from '../mediaViewer';
 
 export const useUserRoomProfileState = (): UserRoomProfileState | undefined => {
   const data = useAtomValue(userRoomProfileAtom);
@@ -29,12 +30,20 @@ type OpenCallback = (
 ) => void;
 export const useOpenUserRoomProfile = (): OpenCallback => {
   const setUserRoomProfile = useSetAtom(userRoomProfileAtom);
+  // Bidirectional mutual exclusion with the media viewer — opening
+  // profile closes any open media viewer pane / sheet. Mirrors the
+  // clear-the-other-atom behaviour in `useOpenMediaViewer`. Without
+  // this, on desktop both `RoomViewProfileSidePanel` and
+  // `RoomViewMediaSidePanel` would mount as siblings and fight for
+  // the right-pane slot.
+  const setMediaViewer = useSetAtom(mediaViewerAtom);
 
   const open: OpenCallback = useCallback(
     (roomId, spaceId, userId, cords, position) => {
+      setMediaViewer(undefined);
       setUserRoomProfile({ roomId, spaceId, userId, cords, position });
     },
-    [setUserRoomProfile]
+    [setUserRoomProfile, setMediaViewer]
   );
 
   return open;
diff --git a/src/app/state/mediaViewer.ts b/src/app/state/mediaViewer.ts
new file mode 100644
index 00000000..d979d171
--- /dev/null
+++ b/src/app/state/mediaViewer.ts
@@ -0,0 +1,31 @@
+import { atom } from 'jotai';
+import { EncryptedAttachmentInfo } from 'browser-encrypt-attachment';
+import { IImageInfo, IVideoInfo } from '../../types/matrix/common';
+
+// Open state for the room media viewer — bottom-up sheet on mobile,
+// right-side pane on desktop. Mirror of `settingsSheetAtom`. The
+// atom-driven shell is mounted in `Room.tsx`; `ImageContent` /
+// `VideoContent` open the atom when the host context from
+// `mediaViewerHostContext` is non-null (Room timeline path);
+// otherwise they fall back to legacy local-state `<Overlay>` or
+// inline native playback (pin-menu, message search — surfaces where
+// the horseshoe shell isn't mounted).
+export type MediaViewerKind = 'image' | 'video';
+
+export type MediaViewerEntry = {
+  roomId: string;
+  eventId: string;
+  kind: MediaViewerKind;
+  // mxc:// URL (unresolved). The viewer body resolves + decrypts
+  // internally using the same code paths the inline thumbnail uses
+  // (`mxcUrlToHttp` + `downloadEncryptedMedia` + `decryptFile`), so
+  // the atom payload stays cheap to construct from a thumbnail-tap
+  // event handler without awaiting decryption.
+  url: string;
+  body: string;
+  info?: IImageInfo | IVideoInfo;
+  encInfo?: EncryptedAttachmentInfo;
+  mimeType?: string;
+};
+
+export const mediaViewerAtom = atom<MediaViewerEntry | undefined>(undefined);