feat(core+host+client): cursor channel — remote-desktop sweep M2a+M2b

The host cursor stops riding the video and becomes a real OS cursor on
the client (the Parsec/RDP model): pointer feel no longer pays the
capture→encode→network→decode→present round trip.

Wire (M2a):
- Hello grows a client_caps trailing byte (CLIENT_CAP_CURSOR) after the
  fixed display_hdr block — presence disambiguated by remaining length,
  which caps the post-HDR tail at 27 bytes (documented); Welcome answers
  HOST_CAP_CURSOR (capable-and-asked, the 444/clipboard precedent).
- CursorShape (0x50, control stream): serial + dims + hotspot + straight
  RGBA, ≤120px/side so the u16 frame always fits (128² would overshoot);
  client caches by serial — re-showing a known shape costs 14 bytes, not
  a bitmap (RDP pointer-cache for free).
- CursorState (0xD0 datagram): serial + visible/relative_hint flags +
  position, sent once per encode-loop tick — latest-wins, self-healing
  under loss, no refresh timer. relative_hint is reserved for M3.
- Client core: two new planes (control-task + datagram-task arms) →
  next_cursor_shape/next_cursor_state; connect() grows client_caps
  (C ABI passes 0 until the v11 cursor poll fns exist).

Host (M2b, Linux portal only):
- handshake::cursor_forward is THE predicate (client asked ∧ Linux ∧
  compositor ≠ gamescope) — Welcome bit and session wiring both read it.
- SessionPlan.cursor_blend goes false for a forwarding session; the
  encode loop ticks a CursorForwarder every iteration: shape-serial diff
  → control-task bridge (mirrors probe_result), state datagram → conn.
- CursorOverlay/capture CursorState carry the hotspot through
  (nearest-neighbor downscale backstop for XL cursors, unit-tested).

Presenter:
- CursorChannel drains both planes per loop iteration; shapes become
  SDL color cursors (from_surface + hotspot), applied while the desktop
  mouse model is engaged; visibility follows the host; capture/released
  hands back the system cursor. Sessions advertise the cap when they
  START in desktop mode.

Verified on .21: fmt + clippy -D warnings (7 crates) + tests green
(core 218 incl. new wire roundtrips, host 245 incl. e2e + forwarder
downscale tests).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

clients/apple/Sources/PunktfunkClient/Settings/SettingsView+Sections.swift

@@ -447,6 +447,14 @@ extension SettingsView {
     /// (always on macOS; an attached keyboard/mouse on iPad). Absent on tvOS (no such input path).
     @ViewBuilder var inputSection: some View {
         Section("Keyboard & mouse") {
+            #if os(macOS)
+            described(mouseModeDescription) {
+                Picker("Mouse input", selection: $mouseMode) {
+                    Text("Capture (games)").tag(MouseInputMode.capture.rawValue)
+                    Text("Desktop (absolute)").tag(MouseInputMode.desktop.rawValue)
+                }
+            }
+            #endif
             described((ModifierLayout(rawValue: modifierLayout) ?? .mac).detail) {
                 Picker("Modifier keys", selection: $modifierLayout) {
                     ForEach(ModifierLayout.allCases, id: \.self) { layout in
@@ -459,6 +467,20 @@ extension SettingsView {
             }
         }
     }
+
+    #if os(macOS)
+    /// The SELECTED mouse model explained — dynamic, like the touch-mode caption.
+    private var mouseModeDescription: String {
+        switch MouseInputMode(rawValue: mouseMode) ?? .capture {
+        case .capture:
+            return "The pointer locks to the stream and sends relative motion — best for "
+                + "games. ⌃⌥⇧M switches live; applies from the next capture otherwise."
+        case .desktop:
+            return "The pointer moves freely in and out of the stream and sends absolute "
+                + "positions — best for remote desktop work. Unavailable on gamescope hosts."
+        }
+    }
+    #endif
     #endif
 
     // MARK: - Audio

clients/apple/Sources/PunktfunkClient/Settings/SettingsView.swift

@@ -89,6 +89,7 @@ struct SettingsView: View {
     @State var customMode = false
     #endif
     #if os(macOS)
+    @AppStorage(DefaultsKey.mouseMode) var mouseMode = MouseInputMode.capture.rawValue
     @AppStorage(DefaultsKey.speakerUID) var speakerUID = ""
     @AppStorage(DefaultsKey.micUID) var micUID = ""
     @AppStorage(DefaultsKey.micChannel) var micChannel = 0

clients/apple/Sources/PunktfunkKit/Input/InputCapture.swift

@@ -110,11 +110,11 @@ public final class InputCapture {
     /// event itself is swallowed). Main queue.
     public var onToggleCapture: (() -> Void)?
 
-    /// Fired on ⌘⇧C (the client-side-cursor toggle — flips between the captured/disassociated
-    /// relative path and the visible-cursor absolute path; detected here, like ⌘⎋, so it works
-    /// regardless of the current capture state and the event itself is swallowed). macOS only;
-    /// the absolute-vs-relative forwarding lives entirely in StreamLayerView. Main queue.
-    public var onToggleCursor: (() -> Void)?
+    /// Fired on ⌃⌥⇧M (the mouse-model flip, capture ⇄ desktop — cross-client parity with the
+    /// SDL clients' Ctrl+Alt+Shift+M; detected here, like ⌘⎋, so it works regardless of the
+    /// current capture state and the event itself is swallowed). macOS only; the
+    /// absolute-vs-relative forwarding lives entirely in StreamLayerView. Main queue.
+    public var onToggleMouseMode: (() -> Void)?
 
     /// The cross-client combos (Windows/Linux parity: Ctrl+Alt+Shift+Q/D/S), fired from the macOS
     /// keyDown monitor only WHILE FORWARDING — that's the state in which the app's menu (which
@@ -245,13 +245,14 @@ public final class InputCapture {
                 self.onToggleCapture?()
                 return nil
             }
-            // ⌘⇧C toggles the client-side cursor (visible-cursor absolute path vs the
-            // captured relative path). keyCode 8 = kVK_ANSI_C; layout-independent so it
-            // fires the same on any keyboard. Suppress the C (latched like ⌘⎋'s Esc) so it
-            // doesn't type into the host, and swallow the event so it doesn't beep.
-            if event.keyCode == 8 /* C */, flags == [.command, .shift] {
-                self.suppressedVK = 0x43 // VK_C — the same physical C is en route via GC
-                self.onToggleCursor?()
+            // ⌃⌥⇧M flips the mouse model (capture ⇄ desktop — the SDL clients' identical
+            // chord). Detected in both capture states, like ⌘⎋, so the model can be set
+            // before engaging. keyCode 46 = kVK_ANSI_M; layout-independent. Suppress the M
+            // (latched like ⌘⎋'s Esc) so it doesn't type into the host, and swallow the
+            // event so it doesn't beep.
+            if event.keyCode == 46 /* M */, flags == [.control, .option, .shift] {
+                self.suppressedVK = 0x4D // VK_M — the same physical M is en route via GC
+                self.onToggleMouseMode?()
                 return nil
             }
             // The cross-client combos (Ctrl+Alt+Shift+Q/D/S — the same set every other

clients/apple/Sources/PunktfunkKit/Input/MouseInputMode.swift

@@ -0,0 +1,12 @@
+/// How a physical mouse drives the host — the cross-client mouse model (the SDL clients'
+/// `MouseMode` / `Settings::mouse_mode`, design/remote-desktop-sweep.md M1). Stored stringly
+/// under `DefaultsKey.mouseMode`.
+public enum MouseInputMode: String, CaseIterable, Sendable {
+    /// Pointer capture (disassociated, hidden cursor, relative deltas) — the game model,
+    /// and the default: the only cursor you see is the host's.
+    case capture
+    /// Absolute pointer, uncaptured: the cursor enters and leaves the stream freely and
+    /// motion is forwarded as absolute positions through the letterbox. The remote desktop
+    /// model. Requires a host injector with absolute support (not gamescope).
+    case desktop
+}

clients/apple/Sources/PunktfunkKit/Views/StreamView.swift

@@ -38,10 +38,11 @@ private let streamInputDebug =
 /// dragged deltas become the relative motion StreamLayerView forwards), and hide it.
 /// hide/unhide and associate are balanced via `captured`.
 ///
-/// In CLIENT-SIDE-CURSOR mode (gamescope, whose capture carries no host cursor) this is a
-/// no-op: the local cursor stays visible and free, and StreamLayerView forwards ABSOLUTE
-/// positions instead — the visible system cursor IS the on-screen cursor. `disassociate`
-/// selects between the two; `release()` only undoes what `capture` actually did.
+/// In the DESKTOP mouse model (absolute pointer, remote-desktop-sweep M1) this is a no-op:
+/// the pointer stays free (entering and leaving the stream at will) and StreamLayerView
+/// forwards ABSOLUTE positions instead; the local cursor is hidden only while over the view
+/// (cursor rects). `disassociate` selects between the two; `release()` only undoes what
+/// `capture` actually did.
 private final class CursorCapture {
     private var captured = false
     /// Whether the engaged capture actually disassociated+hid (false in cursor-visible mode),
@@ -207,14 +208,17 @@ public final class StreamLayerView: NSView {
     /// forwarded). Main-thread only.
     public private(set) var captured = false
 
-    /// Client-side-cursor mode: when true the local system cursor stays VISIBLE over the
-    /// stream and the mouse monitor forwards ABSOLUTE positions (the visible cursor is the
-    /// on-screen cursor — gamescope draws none, so no double cursor); when false the existing
-    /// captured/disassociated relative path runs unchanged. Initialized at session start from
-    /// the `cursorMode` setting + the host's resolved compositor, toggled live by ⌘⇧C. A live
-    /// flip re-engages capture in the new mode so disassociation + the abs/rel choice swap
-    /// atomically. Main-thread only.
-    private var cursorVisible = false
+    /// Desktop (absolute) mouse model — remote-desktop-sweep M1: when true the pointer is
+    /// never disassociated (it enters and leaves the stream freely) and the mouse monitor
+    /// forwards ABSOLUTE positions through the letterbox; the local cursor is hidden only
+    /// while over this view (cursor rects — the host's composited cursor, tracking our
+    /// sends, is the one you see) and reappears the moment it leaves. When false the
+    /// captured/disassociated relative path runs unchanged. Initialized at session start
+    /// from the `mouseMode` setting gated by the host's resolved compositor (gamescope's
+    /// EIS is relative-only — absolute sends would be dropped, so it pins to capture);
+    /// flipped live by ⌃⌥⇧M. A live flip re-engages capture in the new model so
+    /// disassociation + the abs/rel choice swap atomically. Main-thread only.
+    private var desktopMouse = false
     /// One-shot auto-engage request (stream start, trust confirmed) — attempted as soon
     /// as the view is in a window with real bounds, then dropped, so it can never fire
     /// surprisingly later (e.g. on a resize).
@@ -440,9 +444,9 @@ public final class StreamLayerView: NSView {
         // If the cursor grab is refused (e.g. the reactivating click arrives before the app is
         // frontmost), stay released so the NEXT click retries — never latch captured=true over
         // a free cursor, which would make mouseDown's `!captured` guard reject every later click.
-        // In client-side-cursor mode there is no grab (the cursor stays visible) — capture
+        // In the desktop mouse model there is no grab (the pointer stays free) — capture
         // always engages and the monitor forwards absolute positions instead.
-        guard cursorCapture.capture(in: self, disassociate: !cursorVisible) else { return }
+        guard cursorCapture.capture(in: self, disassociate: !desktopMouse) else { return }
         inputCapture?.setForwarding(true, suppressClick: fromClick)
         // Install AFTER the warp + setForwarding: the engage warp generates no forwarded
         // delta (the monitor isn't up yet), and the engage click's suppression latch is
@@ -450,6 +454,7 @@ public final class StreamLayerView: NSView {
         installMouseMonitor()
         captured = true
         window?.makeFirstResponder(self)
+        window?.invalidateCursorRects(for: self) // desktop model: hide-over-view engages
         notifyCaptureChange(true)
     }
 
@@ -459,9 +464,28 @@ public final class StreamLayerView: NSView {
         cursorCapture.release()
         inputCapture?.setForwarding(false)
         captured = false
+        window?.invalidateCursorRects(for: self)
         notifyCaptureChange(false)
     }
 
+    /// A fully transparent cursor for the desktop mouse model's hide-over-view rect —
+    /// an empty 1×1 image draws nothing.
+    private static let invisibleCursor = NSCursor(
+        image: NSImage(size: NSSize(width: 1, height: 1)), hotSpot: .zero)
+
+    /// Desktop mouse model: the local cursor is hidden while over the stream (the host's
+    /// composited cursor, tracking our absolute sends, is the one you see) and reappears
+    /// the moment it leaves the view — AppKit applies/removes the rect's cursor for us,
+    /// so there is no hide/unhide balancing to get wrong. Capture model instead hides
+    /// globally via `CursorCapture` (the pointer can't leave the view there).
+    override public func resetCursorRects() {
+        if captured && desktopMouse {
+            addCursorRect(bounds, cursor: Self.invisibleCursor)
+        } else {
+            super.resetCursorRects()
+        }
+    }
+
     /// A single local monitor for motion + buttons, installed only while captured. A local
     /// monitor is more robust than view overrides for relative motion: it sidesteps the
     /// `window.acceptsMouseMovedEvents`/tracking-area/responder-chain requirements, and
@@ -473,12 +497,12 @@ public final class StreamLayerView: NSView {
     /// via IOHID. Events are returned (not swallowed): the cursor is frozen, so they're
     /// inert locally.
     ///
-    /// In client-side-cursor mode the cursor is NOT frozen, so bare `.mouseMoved` events are
+    /// In the desktop mouse model the cursor is NOT frozen, so bare `.mouseMoved` events are
     /// only generated while `window.acceptsMouseMovedEvents` is true — we enable it here and
     /// restore it on removal so absolute hover-motion keeps flowing without a click held.
     private func installMouseMonitor() {
         guard mouseEventMonitor == nil else { return }
-        if cursorVisible {
+        if desktopMouse {
             savedAcceptsMouseMoved = window?.acceptsMouseMovedEvents
             window?.acceptsMouseMovedEvents = true
         }
@@ -490,8 +514,8 @@ public final class StreamLayerView: NSView {
             guard let self, self.captured, let ic = self.inputCapture else { return event }
             switch event.type {
             case .mouseMoved, .leftMouseDragged, .rightMouseDragged, .otherMouseDragged:
-                if self.cursorVisible {
-                    // Client-side cursor: forward the ABSOLUTE position (mapped through the
+                if self.desktopMouse {
+                    // Desktop mouse model: forward the ABSOLUTE position (mapped through the
                     // aspect-fit letterbox into host pixels), the same path the iPad pointer
                     // fallback uses. Events in the letterbox bars are dropped (nil host point).
                     if let p = self.hostPoint(from: event) {
@@ -609,14 +633,27 @@ public final class StreamLayerView: NSView {
             // be a cursor trap with dead input.
             self?.releaseCapture()
         }
-        // ⌘⇧C flips the client-side cursor live. Only the key window's stream owns it (same
-        // guard as the ⌘⎋ capture toggle). Re-engage capture in the new mode so disassociation
-        // and the absolute/relative forwarding choice swap atomically — releaseCapture restores
-        // the old mode's grab (if any), engageCapture installs the new one.
-        // ⌘⇧C would flip the client-side cursor live — NEUTERED while the feature is disabled
-        // (see the cursorVisible resolution below): toggling it on under gamescope's relative-only
-        // input traps the pointer. Restore this body when absolute/synthetic-cursor support lands.
-        capture.onToggleCursor = {}
+        // ⌃⌥⇧M flips the mouse model (capture ⇄ desktop) live — the SDL clients' identical
+        // chord. Only the key window's stream owns it (same guard as the ⌘⎋ capture toggle).
+        // Re-engage capture in the new model so disassociation and the absolute/relative
+        // forwarding choice swap atomically — releaseCapture restores the old model's grab
+        // (if any), engageCapture installs the new one. On a gamescope host the chord is a
+        // no-op: its EIS grants only a relative pointer, so the desktop model's absolute
+        // sends would be silently dropped (pointer stuck = "all input dead").
+        capture.onToggleMouseMode = { [weak self] in
+            guard let self, self.window?.isKeyWindow == true,
+                  let conn = self.connection else { return }
+            guard conn.resolvedCompositor != .gamescope else {
+                streamInputLog.info("mouse-mode chord ignored: gamescope host is relative-only")
+                return
+            }
+            let wasCaptured = self.captured
+            if wasCaptured { self.releaseCapture() }
+            self.desktopMouse.toggle()
+            if wasCaptured { self.engageCapture(fromClick: false) }
+            self.window?.invalidateCursorRects(for: self)
+            streamInputLog.info("mouse mode: \(self.desktopMouse ? "desktop" : "capture", privacy: .public)")
+        }
         // The cross-client combos (⌃⌥⇧Q/D/S — Ctrl+Alt+Shift on the other clients), delivered by
         // the monitor only while captured; the same key-window ownership rule as ⌘⎋ throughout.
         capture.onReleaseCapture = { [weak self] in
@@ -643,15 +680,18 @@ public final class StreamLayerView: NSView {
         capture.start()
         inputCapture = capture
 
-        // Client-side cursor is TEMPORARILY DISABLED. It positions the host cursor with ABSOLUTE
-        // events, but gamescope's input socket (EIS) grants only a relative pointer, so those are
-        // silently dropped — the pointer never moves and clicks/scroll land on the stuck position
-        // (looks like "all input dead"). gamescope is exactly the compositor Auto enabled it for.
-        // Forced off until per-compositor gating (KWin/GNOME/Sway have absolute) or a synthetic-
-        // cursor-over-relative path lands; the resolution logic below is kept for that. See the
-        // ⌘⇧C handler (also neutered) and the cursorMode setting (hidden).
-        cursorVisible = false
-        _ = connection.resolvedCompositor // (was: Auto → gamescope; kept to document intent)
+        // Desktop (absolute) mouse model — resolved at session start from the mouseMode
+        // setting, gated by the host's compositor: gamescope's input socket (EIS) grants
+        // only a relative pointer, so absolute sends would be silently dropped there
+        // (pointer stuck = "all input dead") — pinned to capture. ⌃⌥⇧M flips it live.
+        let mode = MouseInputMode(
+            rawValue: UserDefaults.standard.string(forKey: DefaultsKey.mouseMode) ?? ""
+        ) ?? .capture
+        let absOK = connection.resolvedCompositor != .gamescope
+        desktopMouse = mode == .desktop && absOK
+        if mode == .desktop && !absOK {
+            streamInputLog.info("desktop mouse mode unavailable on a gamescope host (relative-only) — using capture")
+        }
 
         // Presenter choice + lifecycle live in SessionPresenter (shared with iOS/tvOS): stage-2
         // (explicit VTDecompressionSession decode + a CAMetalLayer/display-link present) by

clients/apple/Sources/PunktfunkShared/DefaultsKeys.swift

@@ -94,8 +94,11 @@ public enum DefaultsKey {
     /// stays 4:2:0). Sharper text/UI at the cost of more bandwidth.
     public static let enable444 = "punktfunk.enable444"
     public static let hosts = "punktfunk.hosts"
-    /// Client-side cursor mode: "auto" (shown only in gamescope sessions), "always", "never".
-    public static let cursorMode = "punktfunk.cursorMode"
+    /// Physical-mouse model (macOS): "capture" (pointer lock + relative, the default) or
+    /// "desktop" (uncaptured absolute pointer) — the cross-client `mouse_mode`. Replaces the
+    /// never-shipped "punktfunk.cursorMode" (auto/always/never client-side-cursor setting,
+    /// which was hidden while disabled and had no readers).
+    public static let mouseMode = "punktfunk.mouseMode"
     /// Invert the scroll-wheel / two-finger-scroll direction sent to the host (both axes). Off by
     /// default: the local (natural-scrolling) sign passes through untouched. When on, the sign is
     /// negated at the single scroll sink (`InputCapture.sendScroll`), so it flips consistently across

clients/linux/src/ui_settings.rs

@@ -69,6 +69,14 @@ const TOUCH_MODE_CAPTIONS: &[&str] = &[
     "The cursor jumps to your finger — a tap clicks there",
     "Real multi-touch reaches the host — for touch-native apps",
 ];
+/// Physical-mouse model values (persisted) + labels + dynamic captions — same idiom as
+/// the touch rows. Ctrl+Alt+Shift+M flips the model live in-stream.
+const MOUSE_MODES: &[&str] = &["capture", "desktop"];
+const MOUSE_MODE_LABELS: &[&str] = &["Capture (games)", "Desktop (absolute)"];
+const MOUSE_MODE_CAPTIONS: &[&str] = &[
+    "Pointer locks to the stream — relative motion, best for games",
+    "Pointer moves freely in and out — best for remote desktop work",
+];
 
 /// punktfunk's own license (MIT OR Apache-2.0), shown on the About dialog's Legal page.
 const APP_LICENSE: &str = concat!(
@@ -542,6 +550,20 @@ pub fn show(
             set_row_subtitle(&w, TOUCH_MODE_CAPTIONS[i]);
         });
     }
+    let mouse_row = ChoiceRow::new(
+        &dialog,
+        inline,
+        "Mouse input",
+        MOUSE_MODE_CAPTIONS[0],
+        MOUSE_MODE_LABELS,
+    );
+    {
+        let w = mouse_row.widget().clone();
+        mouse_row.connect_changed(move |i| {
+            let i = (i as usize).min(MOUSE_MODE_CAPTIONS.len() - 1);
+            set_row_subtitle(&w, MOUSE_MODE_CAPTIONS[i]);
+        });
+    }
     let inhibit_row = adw::SwitchRow::builder()
         .title("Capture system shortcuts")
         .subtitle("Forward Alt+Tab, Super, … to the host while input is captured")
@@ -718,6 +740,12 @@ pub fn show(
         touch_row.set_selected(touch_i as u32);
         // set_selected never fires the changed hook, so seed the dynamic caption directly.
         set_row_subtitle(touch_row.widget(), TOUCH_MODE_CAPTIONS[touch_i]);
+        let mouse_i = MOUSE_MODES
+            .iter()
+            .position(|&m| m == s.mouse_mode)
+            .unwrap_or(0);
+        mouse_row.set_selected(mouse_i as u32);
+        set_row_subtitle(mouse_row.widget(), MOUSE_MODE_CAPTIONS[mouse_i]);
         let comp_i = COMPOSITORS
             .iter()
             .position(|&c| c == s.compositor)
@@ -788,6 +816,7 @@ pub fn show(
     touch_group.add(touch_row.widget());
     // Group titles are Pango markup — the ampersand must be an entity.
     let kbm_group = group("Keyboard & mouse", "");
+    kbm_group.add(mouse_row.widget());
     kbm_group.add(&inhibit_row);
     kbm_group.add(&invert_row);
     input.add(&touch_group);
@@ -867,6 +896,8 @@ pub fn show(
         }
         s.touch_mode =
             TOUCH_MODES[(touch_row.selected() as usize).min(TOUCH_MODES.len() - 1)].to_string();
+        s.mouse_mode =
+            MOUSE_MODES[(mouse_row.selected() as usize).min(MOUSE_MODES.len() - 1)].to_string();
         s.forward_pad = chosen_pin.borrow().clone();
         s.compositor = COMPOSITORS[(compositor_row.selected() as usize).min(COMPOSITORS.len() - 1)]
             .to_string();

clients/session/src/console.rs

@@ -158,6 +158,7 @@ pub fn run(target: Option<&str>) -> u8 {
             v => v,
         },
         touch_mode: settings_at_start.touch_mode(),
+        mouse_mode: settings_at_start.mouse_mode(),
         invert_scroll: settings_at_start.invert_scroll,
         json_status,
         on_connected: Some(Box::new(move |fingerprint: [u8; 32]| {

clients/session/src/main.rs

@@ -172,6 +172,11 @@ mod session_main {
             // defaults for Linux clients; `PUNKTFUNK_CLIENT_PEAK_NITS` (read in the session
             // pump) pins one manually.
             display_hdr: None,
+            // The presenter renders the host cursor locally in desktop mouse mode (M2 cursor
+            // channel); capture-mode sessions keep the composited cursor, so only advertise
+            // when the session STARTS in desktop mode. The host gates further (Linux portal
+            // compositors only).
+            cursor_forward: settings.mouse_mode() == trust::MouseMode::Desktop,
             mic_enabled: settings.mic_enabled,
             clipboard,
             // The Settings preference (auto → VAAPI where it exists; the presenter
@@ -429,6 +434,7 @@ mod session_main {
                 v => v,
             },
             touch_mode: settings.touch_mode(),
+            mouse_mode: settings.mouse_mode(),
             invert_scroll: settings.invert_scroll,
             json_status: true,
             on_connected: Some(Box::new(|fingerprint: [u8; 32]| {

clients/windows/src/app/settings.rs

@@ -90,6 +90,13 @@ const TOUCH_MODES: &[(&str, &str)] = &[
     ("pointer", "Direct pointer"),
     ("touch", "Touch passthrough"),
 ];
+/// Physical-mouse presets: `(stored value, display label)` — capture (pointer lock,
+/// relative, for games) vs desktop (uncaptured absolute pointer, for remote desktop
+/// work). Ctrl+Alt+Shift+M flips the model live in-stream.
+const MOUSE_MODES: &[(&str, &str)] = &[
+    ("capture", "Capture (games)"),
+    ("desktop", "Desktop (absolute)"),
+];
 /// Host compositor presets: `(stored value, display label)`. Advisory — the host falls back to
 /// auto-detect when the choice is unavailable. Only meaningful against a Linux host.
 const COMPOSITORS: &[(&str, &str)] = &[
@@ -394,6 +401,10 @@ pub(crate) fn settings_page(
     let touch_combo = setting_combo(ctx, "Touch input", touch_names, touch_i, |s, i| {
         s.touch_mode = TOUCH_MODES[i].0.to_string();
     });
+    let (mouse_names, mouse_i) = presets(MOUSE_MODES, |v| *v == s.mouse_mode);
+    let mouse_combo = setting_combo(ctx, "Mouse input", mouse_names, mouse_i, |s, i| {
+        s.mouse_mode = MOUSE_MODES[i].0.to_string();
+    });
     let invert_scroll_toggle =
         setting_toggle(ctx, "Invert scroll direction", s.invert_scroll, |s, on| {
             s.invert_scroll = on
@@ -542,6 +553,13 @@ pub(crate) fn settings_page(
             out.extend(group(
                 Some("Keyboard & mouse"),
                 vec![
+                    described(
+                        mouse_combo,
+                        "Capture locks the pointer to the stream and sends relative motion — \
+                         best for games. Desktop leaves the pointer free to enter and leave \
+                         the stream and sends absolute positions — best for remote desktop \
+                         work. Ctrl+Alt+Shift+M switches live.",
+                    ),
                     described(
                         shortcuts_toggle,
                         "Alt+Tab, the Windows key and friends reach the host while the stream \

crates/pf-capture/src/linux/mod.rs

@@ -868,6 +868,10 @@ mod pipewire {
         /// Bumps whenever the bitmap (`rgba`/`bw`/`bh`) changes — stable across position-only moves,
         /// so the GPU encoder re-uploads its cursor texture only on change.
         serial: u64,
+        /// The compositor-reported hotspot — carried on the overlay for the cursor-forward
+        /// channel (the blend path uses the pre-adjusted `x`/`y` and never reads it).
+        hot_x: i32,
+        hot_y: i32,
     }
 
     impl CursorState {
@@ -884,6 +888,8 @@ mod pipewire {
                 h: self.bh,
                 rgba: self.rgba.clone(),
                 serial: self.serial,
+                hot_x: self.hot_x.max(0) as u32,
+                hot_y: self.hot_y.max(0) as u32,
             })
         }
     }
@@ -1380,6 +1386,8 @@ mod pipewire {
         cursor.visible = true;
         cursor.x = pos_x - hot_x;
         cursor.y = pos_y - hot_y;
+        cursor.hot_x = hot_x;
+        cursor.hot_y = hot_y;
         if bmp_off == 0 {
             // Position-only update — keep the cached bitmap.
             return;

crates/pf-client-core/src/session.rs

@@ -44,6 +44,13 @@ pub struct SessionParams {
     /// Share the clipboard with this host (the per-host `KnownHost::clipboard_sync`). The
     /// bridge additionally needs the host to advertise `HOST_CAP_CLIPBOARD`.
     pub clipboard: bool,
+    /// Advertise `quic::CLIENT_CAP_CURSOR`: this embedder renders the host cursor locally
+    /// (the presenter's cursor channel, design/remote-desktop-sweep.md M2), so the host may
+    /// stop compositing the pointer into the video. Only set when the embedder actually
+    /// draws it (the SDL presenter in desktop mouse mode) — a session that advertises it
+    /// without rendering streams with NO visible cursor. The host answers `HOST_CAP_CURSOR`
+    /// when its capture can forward (Linux portal, not gamescope/Windows).
+    pub cursor_forward: bool,
     /// Video decoder preference (Settings; `PUNKTFUNK_DECODER` overrides — see
     /// `video::Decoder::new`).
     pub decoder: String,
@@ -255,6 +262,11 @@ fn pump(
         // This display's HDR volume → the host's virtual-display EDID. The env hatch wins so an
         // A/B run can pin an exact peak (PUNKTFUNK_CLIENT_PEAK_NITS=600).
         punktfunk_core::client::display_hdr_env_override().or(params.display_hdr),
+        if params.cursor_forward {
+            punktfunk_core::quic::CLIENT_CAP_CURSOR
+        } else {
+            0
+        },
         params.launch.clone(),
         params.pin,
         Some(params.identity),

crates/pf-client-core/src/trust.rs

@@ -456,6 +456,48 @@ impl TouchMode {
     }
 }
 
+/// How a physical mouse drives the host — the desktop-sweep mouse model
+/// (design/remote-desktop-sweep.md M1). Stored stringly in [`Settings::mouse_mode`] so the
+/// file stays readable; parsed with [`MouseMode::from_name`].
+#[derive(Clone, Copy, PartialEq, Eq, Debug)]
+pub enum MouseMode {
+    /// Pointer lock (relative deltas, hidden cursor) — the game model, and the default:
+    /// the only cursor you see is the host's.
+    Capture,
+    /// Absolute pointer, uncaptured: the cursor enters and leaves the stream freely and
+    /// motion goes on the wire as absolute positions through the letterbox. The remote
+    /// desktop model. Requires a host injector with absolute support (not gamescope).
+    Desktop,
+}
+
+impl MouseMode {
+    /// Cycle/picker order (also the settings pickers' option order).
+    pub const ALL: [MouseMode; 2] = [MouseMode::Capture, MouseMode::Desktop];
+
+    /// Parse the persisted name, defaulting to `Capture` for unset/unknown values.
+    pub fn from_name(s: &str) -> MouseMode {
+        match s {
+            "desktop" => MouseMode::Desktop,
+            _ => MouseMode::Capture,
+        }
+    }
+
+    /// The persisted name (the inverse of [`from_name`](Self::from_name)).
+    pub fn as_name(self) -> &'static str {
+        match self {
+            MouseMode::Capture => "capture",
+            MouseMode::Desktop => "desktop",
+        }
+    }
+
+    pub fn label(self) -> &'static str {
+        match self {
+            MouseMode::Capture => "Capture (games)",
+            MouseMode::Desktop => "Desktop (absolute)",
+        }
+    }
+}
+
 /// App settings, persisted as JSON. Stringly-typed gamepad/compositor prefs so the file
 /// stays readable; parsed with `*Pref::from_name` at connect time.
 #[derive(Clone, Serialize, Deserialize)]
@@ -490,6 +532,12 @@ pub struct Settings {
     /// stores load as trackpad.
     #[serde(default = "default_touch_mode")]
     pub touch_mode: String,
+    /// How a physical mouse drives the host: a [`MouseMode`] name — `"capture"` (default,
+    /// pointer lock + relative) or `"desktop"` (uncaptured absolute pointer). Read at
+    /// connect via [`Settings::mouse_mode`]. `default` so pre-existing stores load as
+    /// capture — today's behavior.
+    #[serde(default = "default_mouse_mode")]
+    pub mouse_mode: String,
     /// Grab compositor shortcuts (Alt+Tab, Super…) while input is captured.
     pub inhibit_shortcuts: bool,
     /// Stream the default microphone to the host's virtual mic source.
@@ -577,6 +625,10 @@ fn default_touch_mode() -> String {
     "trackpad".into()
 }
 
+fn default_mouse_mode() -> String {
+    "capture".into()
+}
+
 fn default_true() -> bool {
     true
 }
@@ -604,6 +656,10 @@ impl Settings {
         TouchMode::from_name(&self.touch_mode)
     }
 
+    pub fn mouse_mode(&self) -> MouseMode {
+        MouseMode::from_name(&self.mouse_mode)
+    }
+
     /// The `codec` setting as a `quic::CODEC_*` preference bit (`0` = auto).
     pub fn preferred_codec(&self) -> u8 {
         match self.codec.as_str() {
@@ -631,6 +687,7 @@ impl Default for Settings {
             forward_pad: String::new(),
             compositor: "auto".into(),
             touch_mode: "trackpad".into(),
+            mouse_mode: "capture".into(),
             inhibit_shortcuts: true,
             mic_enabled: false,
             audio_channels: 2,

crates/pf-console-ui/src/screens/settings.rs

@@ -10,7 +10,7 @@ use crate::screens::{Ctx, Outbox};
 use crate::theme::{Fonts, DIM, W};
 use crate::widgets::{ListMsg, MenuList, RowSpec};
 use pf_client_core::gamepad::{MenuEvent, MenuPulse};
-use pf_client_core::trust::{StatsVerbosity, TouchMode};
+use pf_client_core::trust::{MouseMode, StatsVerbosity, TouchMode};
 use skia_safe::{Canvas, Rect};
 
 /// Stable row identity — adjust/activate dispatch by id so nothing acts on a stale
@@ -29,10 +29,11 @@ enum RowId {
     Pad,
     PadType,
     Touch,
+    Mouse,
     Stats,
 }
 
-const ROWS: [RowId; 13] = [
+const ROWS: [RowId; 14] = [
     RowId::Resolution,
     RowId::Refresh,
     RowId::Bitrate,
@@ -45,6 +46,7 @@ const ROWS: [RowId; 13] = [
     RowId::Pad,
     RowId::PadType,
     RowId::Touch,
+    RowId::Mouse,
     RowId::Stats,
 ];
 
@@ -251,6 +253,7 @@ fn row_spec(id: RowId, ctx: &Ctx) -> RowSpec {
             "Touch mode",
             s.touch_mode().label().into(),
         ),
+        RowId::Mouse => (None, "Mouse mode", s.mouse_mode().label().into()),
         RowId::Stats => (
             Some("Interface"),
             "Statistics overlay",
@@ -292,6 +295,11 @@ fn detail(id: RowId) -> &'static str {
             "How the touchscreen drives the host: Trackpad (relative cursor), \
              Direct pointer (cursor jumps to your finger), or Touch passthrough (raw contacts)."
         }
+        RowId::Mouse => {
+            "How a physical mouse drives the host: Capture locks the pointer (relative, \
+             for games), Desktop leaves it free and sends absolute positions. \
+             Ctrl+Alt+Shift+M switches live while streaming."
+        }
         RowId::Stats => {
             "How much the overlay shows: Compact (one line) → Normal → Detailed. \
              Ctrl+Alt+Shift+S cycles it live while streaming."
@@ -367,6 +375,11 @@ fn adjust(id: RowId, delta: i32, wrap: bool, ctx: &mut Ctx) -> bool {
             step_option(cur, TouchMode::ALL.len(), delta, wrap)
                 .map(|i| s.touch_mode = TouchMode::ALL[i].as_name().to_string())
         }
+        RowId::Mouse => {
+            let cur = MouseMode::ALL.iter().position(|m| *m == s.mouse_mode());
+            step_option(cur, MouseMode::ALL.len(), delta, wrap)
+                .map(|i| s.mouse_mode = MouseMode::ALL[i].as_name().to_string())
+        }
         RowId::Stats => {
             let cur = StatsVerbosity::ALL
                 .iter()
@@ -510,6 +523,33 @@ mod tests {
         assert_eq!(ctx.settings.touch_mode, "trackpad");
     }
 
+    #[test]
+    fn mouse_mode_steps_and_wraps() {
+        let (mut settings, pads) = ctx_parts();
+        assert_eq!(settings.mouse_mode, "capture");
+        let library = crate::library::LibraryShared::default();
+        let mut ctx = Ctx {
+            hosts: &[],
+            library: &library,
+            settings: &mut settings,
+            pads: &pads,
+            deck: false,
+            device_name: "t",
+            t: 0.0,
+        };
+        // Capture → Desktop, then a step past the end is a boundary.
+        assert!(
+            !adjust(RowId::Mouse, -1, false, &mut ctx),
+            "already first = thud"
+        );
+        assert!(adjust(RowId::Mouse, 1, false, &mut ctx));
+        assert_eq!(ctx.settings.mouse_mode, "desktop");
+        assert!(!adjust(RowId::Mouse, 1, false, &mut ctx), "last = thud");
+        // A wraps back to the first.
+        assert!(adjust(RowId::Mouse, 1, true, &mut ctx));
+        assert_eq!(ctx.settings.mouse_mode, "capture");
+    }
+
     #[test]
     fn unknown_value_snaps_to_first() {
         let (mut settings, pads) = ctx_parts();

crates/pf-frame/src/lib.rs

@@ -196,6 +196,11 @@ pub struct CursorOverlay {
     pub rgba: std::sync::Arc<Vec<u8>>,
     /// Bumps whenever `rgba`/`w`/`h` change; stable across position-only moves.
     pub serial: u64,
+    /// Hotspot (the pixel that IS the pointer position) within `w`×`h`. The blend paths ignore
+    /// it (`x`/`y` are already hotspot-adjusted); the cursor-forward channel ships it to the
+    /// client so a locally-drawn OS cursor points with the right pixel.
+    pub hot_x: u32,
+    pub hot_y: u32,
 }
 
 /// A captured frame. [`format`](Self::format)/dimensions describe the pixels regardless of

crates/pf-presenter/src/cursor.rs

@@ -0,0 +1,119 @@
+//! Client-side cursor rendering (design/remote-desktop-sweep.md M2): the host forwards the
+//! pointer's SHAPE (reliable control stream, cached by serial) and per-frame STATE (lossy
+//! `0xD0` — position/visibility), and WE draw it as a real OS cursor — pointer feel stops
+//! paying the video round-trip (the Parsec/RDP model). Active only when the session
+//! negotiated it (`HOST_CAP_CURSOR` in the Welcome — the host stopped compositing then) and
+//! only applied while the DESKTOP mouse model is engaged: under capture the pointer is
+//! relative-locked (SDL hides it) and games draw their own cursor in-frame.
+
+use punktfunk_core::client::NativeClient;
+use punktfunk_core::quic::{CursorState, HOST_CAP_CURSOR};
+use sdl3::mouse::{Cursor, MouseUtil, SystemCursor};
+use std::collections::HashMap;
+use std::time::Duration;
+
+/// Shape serials cached at most — cursors cycle through a handful of shapes (arrow, I-beam,
+/// resize…); a runaway host can't grow the map past this (the cache resets, shapes re-arrive
+/// on the reliable stream via the serial-miss path).
+const SHAPE_CACHE_MAX: usize = 64;
+
+pub struct CursorChannel {
+    /// The Welcome carried `HOST_CAP_CURSOR` — the host forwards instead of compositing.
+    negotiated: bool,
+    /// Serial → built OS cursor. An SDL `Cursor` must outlive its `set()`, so the cache owns
+    /// every shape ever applied this session (bounded by [`SHAPE_CACHE_MAX`]).
+    shapes: HashMap<u32, Cursor>,
+    /// The serial currently installed via `Cursor::set` (`None` = default/system cursor).
+    installed: Option<u32>,
+    /// Latest `0xD0` state (latest-wins across a drained batch).
+    state: Option<CursorState>,
+}
+
+impl CursorChannel {
+    pub fn new(connector: &NativeClient) -> CursorChannel {
+        let negotiated = connector.host_caps() & HOST_CAP_CURSOR != 0;
+        if negotiated {
+            tracing::info!("cursor channel negotiated — host cursor renders locally");
+        }
+        CursorChannel {
+            negotiated,
+            shapes: HashMap::new(),
+            installed: None,
+            state: None,
+        }
+    }
+
+    /// Whether the host forwards the cursor this session (it no longer composites one).
+    pub fn negotiated(&self) -> bool {
+        self.negotiated
+    }
+
+    /// Drain the two planes and apply the newest state — once per run-loop iteration.
+    /// `desktop_active` = the desktop mouse model is engaged (captured + desktop): only then
+    /// do we own the local cursor's shape/visibility; under capture SDL's relative mode owns
+    /// it, and released the system cursor must look normal.
+    pub fn pump(&mut self, connector: &NativeClient, mouse: &MouseUtil, desktop_active: bool) {
+        if !self.negotiated {
+            return;
+        }
+        while let Ok(shape) = connector.next_cursor_shape(Duration::ZERO) {
+            if self.shapes.len() >= SHAPE_CACHE_MAX {
+                // Degenerate host: reset — live shapes re-install via the serial-miss path.
+                self.shapes.clear();
+                self.installed = None;
+            }
+            let mut data = shape.rgba;
+            let built = sdl3::surface::Surface::from_data(
+                &mut data,
+                shape.w as u32,
+                shape.h as u32,
+                shape.w as u32 * 4,
+                sdl3::pixels::PixelFormat::RGBA32,
+            )
+            .map_err(|e| e.to_string())
+            .and_then(|surf| {
+                Cursor::from_surface(&surf, shape.hot_x as i32, shape.hot_y as i32)
+                    .map_err(|e| e.to_string())
+            });
+            match built {
+                Ok(cursor) => {
+                    // A re-sent serial replaces its entry; force re-install if it's current.
+                    if self.installed == Some(shape.serial) {
+                        self.installed = None;
+                    }
+                    self.shapes.insert(shape.serial, cursor);
+                }
+                Err(e) => tracing::warn!(error = %e, w = shape.w, h = shape.h,
+                    "cursor shape rejected by SDL — keeping the previous cursor"),
+            }
+        }
+        while let Ok(st) = connector.next_cursor_state(Duration::ZERO) {
+            self.state = Some(st); // latest wins
+        }
+
+        if !desktop_active {
+            // Capture mode / released: hand the cursor back to the system default so a
+            // released pointer over the window doesn't wear the host's shape.
+            if self.installed.take().is_some() {
+                Cursor::from_system(SystemCursor::Arrow)
+                    .map(|c| c.set())
+                    .ok();
+            }
+            return;
+        }
+        let Some(st) = self.state else { return };
+        if st.visible() && self.installed != Some(st.serial) {
+            if let Some(cursor) = self.shapes.get(&st.serial) {
+                cursor.set();
+                self.installed = Some(st.serial);
+            }
+            // Serial miss: the (reliable) shape hasn't landed yet — keep the previous
+            // cursor for the RTT rather than flashing default.
+        }
+        // Visibility follows the host (a host app hid its pointer ⇒ ours hides too). Queried,
+        // not shadowed, so apply_capture's own show/hide calls can never desync us.
+        if mouse.is_cursor_showing() != st.visible() {
+            mouse.show_cursor(st.visible());
+        }
+    }
+}

crates/pf-presenter/src/input.rs

@@ -14,10 +14,17 @@
 //! Keys are SDL scancodes → VK via `keymap_sdl`, layout-independent. Motion deltas are
 //! COALESCED: one summed `MouseMove` per loop iteration (a 1000 Hz mouse would
 //! otherwise send a datagram per event).
+//!
+//! The DESKTOP mouse model (design/remote-desktop-sweep.md M1) reuses this same engage/
+//! release state but never locks the pointer: the local cursor moves freely (hidden over
+//! the window — the host's composited cursor is the one you see) and motion goes on the
+//! wire as absolute positions through the letterbox (`MouseMoveAbs`, latest-wins per loop
+//! iteration). Requires a host injector with absolute support — gamescope's EIS is
+//! relative-only, so sessions there are pinned to capture ([`Capture::new`] `abs_ok`).
 
 use crate::keymap_sdl;
 use crate::touch::{Abs, Act, Gestures};
-use pf_client_core::trust::TouchMode;
+use pf_client_core::trust::{MouseMode, TouchMode};
 use punktfunk_core::client::NativeClient;
 use punktfunk_core::input::{InputEvent, InputKind};
 use std::collections::{HashMap, HashSet};
@@ -41,6 +48,13 @@ pub struct Capture {
     held_buttons: HashSet<u32>,
     /// Relative motion not yet on the wire, summed per loop iteration.
     pending_rel: (i32, i32),
+    /// Desktop-model position not yet on the wire, latest-wins per loop iteration.
+    pending_abs: Option<Abs>,
+    /// The desktop (absolute, uncaptured) mouse model is active. Flipped live by the
+    /// Ctrl+Alt+Shift+M chord; never true unless `abs_ok`.
+    desktop: bool,
+    /// The host injector accepts `MouseMoveAbs` (any compositor but gamescope).
+    abs_ok: bool,
     /// Fractional wheel remainder per axis (x, y) in 120-unit WHEEL_DELTA space —
     /// precision surfaces deliver sub-unit deltas; truncating each event drops the tail.
     scroll_acc: (f64, f64),
@@ -70,10 +84,14 @@ fn send(connector: &NativeClient, kind: InputKind, code: u32, x: i32, y: i32, fl
 }
 
 impl Capture {
+    /// `abs_ok` = the host injector accepts absolute pointer events; without it the
+    /// desktop model is unavailable and `mouse_mode` silently resolves to capture.
     pub fn new(
         connector: Arc<NativeClient>,
         touch_mode: TouchMode,
         invert_scroll: bool,
+        mouse_mode: MouseMode,
+        abs_ok: bool,
     ) -> Capture {
         Capture {
             connector,
@@ -82,6 +100,9 @@ impl Capture {
             held_keys: HashSet::new(),
             held_buttons: HashSet::new(),
             pending_rel: (0, 0),
+            pending_abs: None,
+            desktop: abs_ok && mouse_mode == MouseMode::Desktop,
+            abs_ok,
             scroll_acc: (0.0, 0.0),
             touch_slots: HashMap::new(),
             touch_mode,
@@ -94,6 +115,24 @@ impl Capture {
         self.captured
     }
 
+    /// The desktop (absolute, uncaptured) mouse model is active.
+    pub fn desktop(&self) -> bool {
+        self.desktop
+    }
+
+    /// Flip capture ⇄ desktop (the Ctrl+Alt+Shift+M chord). `None` = the host can't take
+    /// absolute pointer events (gamescope), so the chord has nothing to offer; otherwise
+    /// the new desktop state. Motion gathered under the old model never crosses modes.
+    pub fn toggle_desktop(&mut self) -> Option<bool> {
+        if !self.abs_ok {
+            return None;
+        }
+        self.desktop = !self.desktop;
+        self.pending_rel = (0, 0);
+        self.pending_abs = None;
+        Some(self.desktop)
+    }
+
     /// Whether a regained focus should re-engage: yes unless the user released
     /// deliberately (the chord keeps its meaning across an Alt-Tab).
     pub fn should_reengage(&self) -> bool {
@@ -117,6 +156,7 @@ impl Capture {
             return false;
         }
         self.pending_rel = (0, 0); // never flush motion gathered while captured
+        self.pending_abs = None;
         for vk in self.held_keys.drain() {
             send(&self.connector, InputKind::KeyUp, vk as u32, 0, 0, 0);
         }
@@ -132,22 +172,42 @@ impl Capture {
         true
     }
 
-    /// Forward the coalesced motion delta, if any — one datagram per loop iteration.
+    /// Forward the coalesced motion, if any — one datagram per loop iteration. Only one
+    /// of the two stores is ever populated (the run loop routes by [`desktop`](Self::desktop)).
     pub fn flush_motion(&mut self) {
         let (dx, dy) = std::mem::take(&mut self.pending_rel);
         if dx != 0 || dy != 0 {
             send(&self.connector, InputKind::MouseMove, 0, dx, dy, 0);
         }
+        if let Some(a) = self.pending_abs.take() {
+            send(
+                &self.connector,
+                InputKind::MouseMoveAbs,
+                0,
+                a.x,
+                a.y,
+                Self::touch_flags(a.w, a.h),
+            );
+        }
     }
 
     /// Relative motion (SDL relative mouse mode delivers raw deltas while locked).
     pub fn on_motion(&mut self, xrel: f32, yrel: f32) {
-        if self.captured {
+        if self.captured && !self.desktop {
             self.pending_rel.0 += xrel as i32;
             self.pending_rel.1 += yrel as i32;
         }
     }
 
+    /// Desktop-model motion: the cursor's position mapped into the letterboxed content
+    /// rect. Latest-wins — intermediate positions carry no information the final one
+    /// doesn't (unlike deltas, which must sum).
+    pub fn on_motion_abs(&mut self, abs: Abs) {
+        if self.captured && self.desktop {
+            self.pending_abs = Some(abs);
+        }
+    }
+
     pub fn on_key_down(&mut self, sc: sdl3::keyboard::Scancode) {
         if !self.captured {
             return;

crates/pf-presenter/src/lib.rs

@@ -16,6 +16,8 @@
 
 #[cfg(any(target_os = "linux", windows))]
 pub mod csc;
+#[cfg(any(target_os = "linux", windows))]
+pub mod cursor;
 #[cfg(windows)]
 pub mod d3d11;
 #[cfg(target_os = "linux")]

crates/pf-presenter/src/run.rs

@@ -20,11 +20,11 @@ use crate::vk::{FrameInput, Presenter};
 use anyhow::{Context as _, Result};
 use pf_client_core::gamepad::GamepadService;
 use pf_client_core::session::{self, SessionEvent, SessionHandle, SessionParams, Stats};
-use pf_client_core::trust::{StatsVerbosity, TouchMode};
+use pf_client_core::trust::{MouseMode, StatsVerbosity, TouchMode};
 use pf_client_core::video::VulkanDecodeDevice;
 use pf_client_core::video::{DecodedFrame, DecodedImage};
 use punktfunk_core::client::NativeClient;
-use punktfunk_core::config::Mode;
+use punktfunk_core::config::{CompositorPref, Mode};
 use sdl3::event::{Event, WindowEvent};
 use sdl3::keyboard::Mod;
 use std::sync::atomic::{AtomicBool, Ordering};
@@ -48,6 +48,11 @@ pub struct SessionOpts {
     /// `Pointer` (absolute cursor), or `Touch` (real multi-touch passthrough). Latched per
     /// session — a mouse-only client leaves this at the default and never sees a finger.
     pub touch_mode: TouchMode,
+    /// Physical-mouse model: `Capture` (pointer lock + relative, the default) or `Desktop`
+    /// (uncaptured absolute pointer — design/remote-desktop-sweep.md M1). Ctrl+Alt+Shift+M
+    /// flips it live; silently resolves to capture on hosts without absolute injection
+    /// (gamescope).
+    pub mouse_mode: MouseMode,
     /// Reverse the scroll direction sent to the host ([`Settings::invert_scroll`]).
     pub invert_scroll: bool,
     /// Emit the `{"ready":true}` stdout line after the first presented frame.
@@ -228,6 +233,9 @@ struct StreamState {
     /// window-normalized position must be re-based onto the content rect). `None` until
     /// the first frame; touches before then have nothing to map onto and are dropped.
     last_video: Option<(u32, u32)>,
+    /// Client-side cursor rendering (M2 cursor channel) — created with the connector; inert
+    /// when the host didn't negotiate the channel.
+    cursor_chan: Option<crate::cursor::CursorChannel>,
 }
 
 impl StreamState {
@@ -258,6 +266,7 @@ impl StreamState {
             frames: wake_rx,
             connector: None,
             capture: None,
+            cursor_chan: None,
             force_software,
             canceled: false,
             ready_announced: false,
@@ -490,7 +499,7 @@ fn run_inner(mut opts: SessionOpts, mut mode: ModeCtl) -> Result<Option<Outcome>
                     WindowEvent::FocusLost => {
                         if let Some(cap) = stream.as_mut().and_then(|s| s.capture.as_mut()) {
                             if cap.release(false) {
-                                apply_capture(&mut window, &mouse, false);
+                                apply_capture(&mut window, &mouse, false, false);
                                 tracing::info!("focus lost — input released");
                             }
                         }
@@ -501,7 +510,7 @@ fn run_inner(mut opts: SessionOpts, mut mode: ModeCtl) -> Result<Option<Outcome>
                         if let Some(cap) = stream.as_mut().and_then(|s| s.capture.as_mut()) {
                             if cap.should_reengage() {
                                 cap.engage();
-                                apply_capture(&mut window, &mouse, true);
+                                apply_capture(&mut window, &mouse, true, cap.desktop());
                                 tracing::info!("focus gained — input recaptured");
                             }
                         }
@@ -537,20 +546,39 @@ fn run_inner(mut opts: SessionOpts, mut mode: ModeCtl) -> Result<Option<Outcome>
                         if let Some(cap) = stream.as_mut().and_then(|s| s.capture.as_mut()) {
                             if cap.captured() {
                                 cap.release(true);
-                                apply_capture(&mut window, &mouse, false);
+                                apply_capture(&mut window, &mouse, false, false);
                             } else {
                                 cap.engage();
-                                apply_capture(&mut window, &mouse, true);
+                                apply_capture(&mut window, &mouse, true, cap.desktop());
                             }
                             tracing::info!(captured = cap.captured(), "chord: release/engage");
                         }
                         continue;
                     }
+                    // Mouse model flip (capture ⇄ desktop) — applies immediately when
+                    // engaged; a released stream just changes what the next engage does.
+                    if chord && sc == Scancode::M {
+                        if let Some(cap) = stream.as_mut().and_then(|s| s.capture.as_mut()) {
+                            match cap.toggle_desktop() {
+                                Some(desktop) => {
+                                    if cap.captured() {
+                                        apply_capture(&mut window, &mouse, true, desktop);
+                                    }
+                                    tracing::info!(desktop, "chord: mouse mode");
+                                }
+                                None => tracing::info!(
+                                    "chord: mouse mode — host has no absolute pointer \
+                                     (gamescope), staying captured"
+                                ),
+                            }
+                        }
+                        continue;
+                    }
                     if chord && sc == Scancode::D {
                         if let Some(st) = &mut stream {
                             tracing::info!("chord: disconnect");
                             st.request_quit();
-                            apply_capture(&mut window, &mouse, false);
+                            apply_capture(&mut window, &mouse, false, false);
                             // The pump emits Ended(None); the end path routes per mode.
                         }
                         continue;
@@ -583,9 +611,34 @@ fn run_inner(mut opts: SessionOpts, mut mode: ModeCtl) -> Result<Option<Outcome>
                         cap.on_key_up(sc);
                     }
                 }
-                Event::MouseMotion { xrel, yrel, .. } => {
-                    if let Some(cap) = stream.as_mut().and_then(|s| s.capture.as_mut()) {
-                        cap.on_motion(xrel, yrel);
+                Event::MouseMotion {
+                    x, y, xrel, yrel, ..
+                } => {
+                    if let Some(st) = stream.as_mut() {
+                        let video = st.last_video;
+                        if let Some(cap) = st.capture.as_mut() {
+                            if cap.desktop() {
+                                // Desktop model: the cursor's window position through the
+                                // letterbox (same mapping as a pointer-mode finger).
+                                // Before the first decoded frame there is nothing to map
+                                // onto — dropped, like touch.
+                                if let Some(video) = video {
+                                    let (lw, lh) = window.size();
+                                    let nx = x / lw.max(1) as f32;
+                                    let ny = y / lh.max(1) as f32;
+                                    let (ax, ay, aw, ah) =
+                                        finger_to_content(window.size_in_pixels(), video, nx, ny);
+                                    cap.on_motion_abs(Abs {
+                                        x: ax,
+                                        y: ay,
+                                        w: aw,
+                                        h: ah,
+                                    });
+                                }
+                            } else {
+                                cap.on_motion(xrel, yrel);
+                            }
+                        }
                     }
                 }
                 Event::MouseButtonDown { mouse_btn, .. } => {
@@ -593,7 +646,7 @@ fn run_inner(mut opts: SessionOpts, mut mode: ModeCtl) -> Result<Option<Outcome>
                         if !cap.captured() {
                             // The engaging click is suppressed toward the host.
                             cap.engage();
-                            apply_capture(&mut window, &mouse, true);
+                            apply_capture(&mut window, &mouse, true, cap.desktop());
                         } else {
                             cap.on_button_down(mouse_btn);
                         }
@@ -695,6 +748,17 @@ fn run_inner(mut opts: SessionOpts, mut mode: ModeCtl) -> Result<Option<Outcome>
         if let Some(cap) = stream.as_mut().and_then(|s| s.capture.as_mut()) {
             cap.flush_motion();
         }
+        // Cursor channel (M2): drain forwarded shape/state and drive the local OS cursor —
+        // only meaningful in the desktop mouse model (capture's relative lock hides it).
+        if let Some(st) = stream.as_mut() {
+            if let (Some(chan), Some(c)) = (st.cursor_chan.as_mut(), st.connector.as_ref()) {
+                let desktop_active = st
+                    .capture
+                    .as_ref()
+                    .is_some_and(|cap| cap.captured() && cap.desktop());
+                chan.pump(c, &mouse, desktop_active);
+            }
+        }
 
         // Text input follows the overlay's editing state (edge-triggered).
         let want_text = overlay.as_ref().is_some_and(|o| o.text_input_active());
@@ -714,7 +778,7 @@ fn run_inner(mut opts: SessionOpts, mut mode: ModeCtl) -> Result<Option<Outcome>
         while escape_rx.try_recv().is_ok() {
             if let Some(cap) = stream.as_mut().and_then(|s| s.capture.as_mut()) {
                 if cap.release(true) {
-                    apply_capture(&mut window, &mouse, false);
+                    apply_capture(&mut window, &mouse, false, false);
                 }
             }
             if fullscreen && !opts.fullscreen {
@@ -727,7 +791,7 @@ fn run_inner(mut opts: SessionOpts, mut mode: ModeCtl) -> Result<Option<Outcome>
             if let Some(st) = &mut stream {
                 tracing::info!("controller chord: disconnect");
                 st.request_quit();
-                apply_capture(&mut window, &mouse, false);
+                apply_capture(&mut window, &mouse, false, false);
             }
         }
 
@@ -818,10 +882,28 @@ fn run_inner(mut opts: SessionOpts, mut mode: ModeCtl) -> Result<Option<Outcome>
                         .ok();
                     gamepad.attach(c.clone());
                     st.clock_offset = Some(c.clock_offset_shared());
-                    let mut cap = Capture::new(c.clone(), opts.touch_mode, opts.invert_scroll);
+                    // gamescope's EIS grants only a relative pointer — absolute sends
+                    // would be dropped, so the desktop model is pinned off there. Auto
+                    // (an older host that didn't say) stays allowed: Windows hosts and
+                    // pre-Welcome-compositor Linux hosts both take absolute.
+                    let abs_ok = c.resolved_compositor != CompositorPref::Gamescope;
+                    if opts.mouse_mode == MouseMode::Desktop && !abs_ok {
+                        tracing::info!(
+                            "desktop mouse mode unavailable on a gamescope host \
+                             (relative-only input) — using capture"
+                        );
+                    }
+                    let mut cap = Capture::new(
+                        c.clone(),
+                        opts.touch_mode,
+                        opts.invert_scroll,
+                        opts.mouse_mode,
+                        abs_ok,
+                    );
                     cap.engage(); // capture engages when the stream starts (ui_stream parity)
-                    apply_capture(&mut window, &mouse, true);
+                    apply_capture(&mut window, &mouse, true, cap.desktop());
                     st.capture = Some(cap);
+                    st.cursor_chan = Some(crate::cursor::CursorChannel::new(&c));
                     st.connector = Some(c);
                     if let Some(f) = opts.on_connected.as_mut() {
                         f(fingerprint);
@@ -870,7 +952,7 @@ fn run_inner(mut opts: SessionOpts, mut mode: ModeCtl) -> Result<Option<Outcome>
                         if let Some(st) = stream.take() {
                             st.shutdown();
                         }
-                        apply_capture(&mut window, &mouse, false);
+                        apply_capture(&mut window, &mouse, false, false);
                         if let Some(o) = overlay.as_mut() {
                             // A user-canceled dial ends silently — no error scene.
                             if canceled {
@@ -887,7 +969,7 @@ fn run_inner(mut opts: SessionOpts, mut mode: ModeCtl) -> Result<Option<Outcome>
                     if let Some(cap) = &mut st.capture {
                         cap.release(true);
                     }
-                    apply_capture(&mut window, &mouse, false);
+                    apply_capture(&mut window, &mouse, false, false);
                     match &mode {
                         ModeCtl::Single(_) => break 'main Some(Outcome::Ended(reason)),
                         ModeCtl::Browse(_) => {
@@ -1477,11 +1559,22 @@ impl ResizeIndicator {
 /// with a low-level keyboard hook, the same mechanism the WinUI shell's in-process
 /// client used its own WH_KEYBOARD_LL hooks for. Not engaged on Linux: the compositor
 /// shortcut-inhibit story stays the shells' concern (Settings.inhibit_shortcuts).
-fn apply_capture(window: &mut sdl3::video::Window, mouse: &sdl3::mouse::MouseUtil, on: bool) {
-    mouse.set_relative_mouse_mode(window, on);
+///
+/// The `desktop` mouse model never locks: the pointer roams (and leaves the window)
+/// freely, the local cursor is hidden over the window — the host's composited cursor,
+/// tracking our absolute sends, is the one you see (until the M2 cursor channel flips
+/// who draws it) — and system chords stay local (a remote desktop is something you
+/// Alt-Tab away from, not into). `desktop` only matters while `on`.
+fn apply_capture(
+    window: &mut sdl3::video::Window,
+    mouse: &sdl3::mouse::MouseUtil,
+    on: bool,
+    desktop: bool,
+) {
+    mouse.set_relative_mouse_mode(window, on && !desktop);
     mouse.show_cursor(!on);
     #[cfg(windows)]
-    window.set_keyboard_grab(on);
+    window.set_keyboard_grab(on && !desktop);
 }
 
 /// Is this SDL touch device a real touchscreen (DIRECT, window-relative coordinates)?
@@ -1599,7 +1692,7 @@ struct PresentedWindow {
 
 /// The capture hints (`ui_stream` parity — the words the user reads while released).
 const HINT_KEYBOARD: &str = "Click the stream to capture input · Ctrl+Alt+Shift+Q releases · \
-     Ctrl+Alt+Shift+D disconnects · Ctrl+Alt+Shift+S stats";
+     Ctrl+Alt+Shift+M mouse mode · Ctrl+Alt+Shift+D disconnects · Ctrl+Alt+Shift+S stats";
 const HINT_WITH_PAD: &str = "Click the stream to capture input · Ctrl+Alt+Shift+Q releases · \
      Ctrl+Alt+Shift+D disconnects · hold L1 + R1 + Start + Select to leave";
 

crates/punktfunk-core/src/abi.rs

@@ -1574,6 +1574,10 @@ unsafe fn connect_ex_impl(
             // themselves (EDR / MediaCodec), so the host's EDID defaults are fine there. An `ex8`
             // variant can carry it if a passthrough embedder ever needs it.
             None,
+            // No client_caps in the C ABI yet either: cursor-channel opt-in for Apple/Android
+            // arrives with the ABI v11 cursor poll fns — until an embedder can RENDER the
+            // forwarded cursor it must not ask the host to stop compositing it.
+            0,
             launch,
             pin,
             identity,

crates/punktfunk-core/src/client/mod.rs

@@ -42,8 +42,8 @@ pub use self::rumble::{ActuatorQuirks, RumbleCommand};
 use self::control::{CtrlRequest, Negotiated};
 use self::frame_channel::{DecodeLatAcc, FrameChannel, FramePop};
 use self::planes::{
-    RumbleUpdate, AUDIO_QUEUE, CLIP_EVENT_QUEUE, HDR_META_QUEUE, HIDOUT_QUEUE, HOST_TIMING_QUEUE,
-    RUMBLE_QUEUE,
+    RumbleUpdate, AUDIO_QUEUE, CLIP_EVENT_QUEUE, CURSOR_SHAPE_QUEUE, CURSOR_STATE_QUEUE,
+    HDR_META_QUEUE, HIDOUT_QUEUE, HOST_TIMING_QUEUE, RUMBLE_QUEUE,
 };
 use self::probe::ProbeState;
 use self::pump::run_pump;
@@ -93,6 +93,12 @@ pub struct NativeClient {
     /// Inbound per-AU host capture→send timings — 0xCF datagrams (the client always advertises
     /// [`quic::VIDEO_CAP_HOST_TIMING`]; an older host simply never sends any).
     host_timing: Mutex<Receiver<crate::quic::HostTiming>>,
+    /// Inbound cursor shapes (control-stream [`crate::quic::CursorShape`]) — only a session
+    /// that advertised [`quic::CLIENT_CAP_CURSOR`] against a [`quic::HOST_CAP_CURSOR`] host
+    /// ever receives any.
+    cursor_shape: Mutex<Receiver<crate::quic::CursorShape>>,
+    /// Inbound per-frame cursor state — `0xD0` datagrams (same negotiation gate as shapes).
+    cursor_state: Mutex<Receiver<crate::quic::CursorState>>,
     input_tx: tokio::sync::mpsc::UnboundedSender<InputEvent>,
     /// Outbound mic frames `(seq, pts_ns, opus)` → encoded as 0xCB datagrams by the worker.
     /// Bounded ([`MIC_QUEUE`]): a wedged worker drops fresh frames (logged) instead of queueing
@@ -316,6 +322,12 @@ impl NativeClient {
         // display's EDID so host apps tone-map to the client's real panel; `None` = unknown/SDR
         // (the host keeps its built-in EDID defaults). See [`crate::quic::Hello::display_hdr`].
         display_hdr: Option<HdrMeta>,
+        // Non-video client capabilities ([`crate::quic::Hello::client_caps`]) — set
+        // [`crate::quic::CLIENT_CAP_CURSOR`] ONLY if this embedder actually renders the host
+        // cursor locally (shape + state planes): the host stops compositing the pointer into
+        // the video for a session that advertises it, so a non-rendering embedder that sets it
+        // streams with NO visible cursor at all. `0` = today's composited behavior.
+        client_caps: u8,
         launch: Option<String>,
         pin: Option<[u8; 32]>,
         identity: Option<(String, String)>,
@@ -337,6 +349,10 @@ impl NativeClient {
         let (clip_event_tx, clip_event_rx) =
             std::sync::mpsc::sync_channel::<ClipEventCore>(CLIP_EVENT_QUEUE);
         let (clip_cmd_tx, clip_cmd_rx) = tokio::sync::mpsc::unbounded_channel::<ClipCommand>();
+        let (cursor_shape_tx, cursor_shape_rx) =
+            std::sync::mpsc::sync_channel::<crate::quic::CursorShape>(CURSOR_SHAPE_QUEUE);
+        let (cursor_state_tx, cursor_state_rx) =
+            std::sync::mpsc::sync_channel::<crate::quic::CursorState>(CURSOR_STATE_QUEUE);
         let (ready_tx, ready_rx) = std::sync::mpsc::channel::<Result<Negotiated>>();
         let shutdown = Arc::new(AtomicBool::new(false));
         let quit = Arc::new(AtomicBool::new(false));
@@ -390,6 +406,7 @@ impl NativeClient {
                     video_codecs,
                     preferred_codec,
                     display_hdr,
+                    client_caps,
                     launch,
                     pin,
                     identity,
@@ -401,6 +418,8 @@ impl NativeClient {
                     hidout_tx,
                     hdr_meta_tx,
                     host_timing_tx,
+                    cursor_shape_tx,
+                    cursor_state_tx,
                     input_rx,
                     mic_rx,
                     rich_input_rx,
@@ -445,6 +464,8 @@ impl NativeClient {
             hidout: Mutex::new(hidout_rx),
             hdr_meta: Mutex::new(hdr_meta_rx),
             host_timing: Mutex::new(host_timing_rx),
+            cursor_shape: Mutex::new(cursor_shape_rx),
+            cursor_state: Mutex::new(cursor_state_rx),
             input_tx,
             mic_tx,
             rich_input_tx,
@@ -892,6 +913,32 @@ impl NativeClient {
         }
     }
 
+    /// Pull the next host cursor shape (design/remote-desktop-sweep.md M2): RGBA bitmap +
+    /// hotspot, sent on pointer-bitmap change over the reliable control stream. The embedder
+    /// caches by `serial` and builds an OS cursor from it; [`NativeClient::next_cursor_state`]
+    /// references shapes by serial. Only a session that advertised
+    /// [`crate::quic::CLIENT_CAP_CURSOR`] against a capable host receives any. Same
+    /// timeout/closed semantics as [`NativeClient::next_hidout`].
+    pub fn next_cursor_shape(&self, timeout: Duration) -> Result<crate::quic::CursorShape> {
+        match self.cursor_shape.lock().unwrap().recv_timeout(timeout) {
+            Ok(s) => Ok(s),
+            Err(RecvTimeoutError::Timeout) => Err(PunktfunkError::NoFrame),
+            Err(RecvTimeoutError::Disconnected) => Err(PunktfunkError::Closed),
+        }
+    }
+
+    /// Pull the next per-frame cursor state (`0xD0`): position, visibility and the M3
+    /// relative-mode hint, referencing a shape by serial. Latest-wins — an embedder should
+    /// drain the queue and apply only the newest. Same negotiation gate and timeout/closed
+    /// semantics as [`NativeClient::next_cursor_shape`].
+    pub fn next_cursor_state(&self, timeout: Duration) -> Result<crate::quic::CursorState> {
+        match self.cursor_state.lock().unwrap().recv_timeout(timeout) {
+            Ok(s) => Ok(s),
+            Err(RecvTimeoutError::Timeout) => Err(PunktfunkError::NoFrame),
+            Err(RecvTimeoutError::Disconnected) => Err(PunktfunkError::Closed),
+        }
+    }
+
     /// Pull the next per-AU host timing (0xCF): the host's capture→sent duration for one access
     /// unit, correlated to the AU by `pts_ns`. Feeds the unified stats HUD's `host` / `network`
     /// split (`network = (received + clock_offset − pts) − host_us`); a stats consumer should

crates/punktfunk-core/src/client/planes.rs

@@ -35,6 +35,16 @@ pub(crate) const HOST_TIMING_QUEUE: usize = 512;
 /// a dropped fetch-request makes the serving stream time out and reset cleanly.
 pub(crate) const CLIP_EVENT_QUEUE: usize = 32;
 
+/// Cursor-shape plane depth (control-stream [`crate::quic::CursorShape`], one per pointer-bitmap
+/// change — human-paced). Overflow drops the newest (try_send); the next shape change or a
+/// serial mismatch against `0xD0` state heals it visually within a shape-change period.
+pub(crate) const CURSOR_SHAPE_QUEUE: usize = 8;
+
+/// Cursor-state plane depth (`0xD0`, one datagram per captured frame). Latest-wins state — the
+/// embedder drains per present; a tiny ring only bridges scheduling jitter. Overflow drops the
+/// newest (try_send), healed by the very next frame's datagram.
+pub(crate) const CURSOR_STATE_QUEUE: usize = 8;
+
 /// One Opus packet from the host's audio datagram stream (48 kHz stereo, 5 ms frames).
 #[derive(Clone, Debug)]
 pub struct AudioPacket {

crates/punktfunk-core/src/client/pump.rs

@@ -51,6 +51,8 @@ pub(super) async fn run_pump(args: WorkerArgs) {
         hidout_tx,
         hdr_meta_tx,
         host_timing_tx,
+        cursor_shape_tx,
+        cursor_state_tx,
         input_rx,
         mut mic_rx,
         mut rich_input_rx,
@@ -123,6 +125,7 @@ pub(super) async fn run_pump(args: WorkerArgs) {
             clock_offset: clock_offset.clone(),
             clock_gen: clock_gen.clone(),
             clip_event_tx: clip_event_tx.clone(),
+            cursor_shape_tx,
         }
         .run(),
     );
@@ -136,6 +139,7 @@ pub(super) async fn run_pump(args: WorkerArgs) {
         hidout_tx,
         hdr_meta_tx,
         host_timing_tx,
+        cursor_state_tx,
     ));
 
     // Clipboard task: the fetch-stream accept loop (host pulls what we offered) + outbound fetches

crates/punktfunk-core/src/client/pump/control_task.rs

@@ -21,6 +21,9 @@ pub(super) struct ControlTask {
     /// Clipboard metadata events (ClipState/ClipOffer) feed the same event plane the
     /// clipboard task uses for fetch data.
     pub(super) clip_event_tx: std::sync::mpsc::SyncSender<ClipEventCore>,
+    /// Host cursor shapes ([`CursorShape`], sent on pointer-bitmap change) → the embedder's
+    /// shape plane ([`NativeClient::next_cursor_shape`]).
+    pub(super) cursor_shape_tx: std::sync::mpsc::SyncSender<crate::quic::CursorShape>,
 }
 
 impl ControlTask {
@@ -36,6 +39,7 @@ impl ControlTask {
             clock_offset,
             clock_gen,
             clip_event_tx,
+            cursor_shape_tx,
         } = self;
         // Mid-stream clock re-sync (see [`ClockResync`]): a batch runs every
         // CLOCK_RESYNC_INTERVAL and whenever the pump asks (CtrlRequest::ClockResync after
@@ -167,6 +171,10 @@ impl ControlTask {
                             seq: offer.seq,
                             kinds: offer.kinds,
                         });
+                    } else if let Ok(shape) = crate::quic::CursorShape::decode(&msg) {
+                        // Pointer bitmap changed (cursor channel, only when negotiated). try_send:
+                        // an overflowing ring drops the newest shape — the next change resends.
+                        let _ = cursor_shape_tx.try_send(shape);
                     } else {
                         tracing::warn!(
                             tag = ?msg.first(),

crates/punktfunk-core/src/client/pump/datagram_task.rs

@@ -3,6 +3,9 @@
 
 use super::*;
 
+// One parameter per demuxed plane — grouping them into a struct would just move the field
+// list one hop away from the single call site.
+#[allow(clippy::too_many_arguments)]
 pub(super) async fn run(
     conn: quinn::Connection,
     audio_tx: std::sync::mpsc::SyncSender<AudioPacket>,
@@ -11,6 +14,7 @@ pub(super) async fn run(
     hidout_tx: std::sync::mpsc::SyncSender<crate::quic::HidOutput>,
     hdr_meta_tx: std::sync::mpsc::SyncSender<crate::quic::HdrMeta>,
     host_timing_tx: std::sync::mpsc::SyncSender<crate::quic::HostTiming>,
+    cursor_state_tx: std::sync::mpsc::SyncSender<crate::quic::CursorState>,
 ) {
     // Per-pad reorder gate for v2 rumble envelopes (the seq analog of the host's gamepad-state
     // gate): a datagram the network reordered must not roll a stopped motor back on. Legacy v1
@@ -73,6 +77,11 @@ pub(super) async fn run(
                     let _ = host_timing_tx.try_send(t);
                 }
             }
+            Some(&crate::quic::CURSOR_STATE_MAGIC) => {
+                if let Some(s) = crate::quic::decode_cursor_state_datagram(&d) {
+                    let _ = cursor_state_tx.try_send(s);
+                }
+            }
             _ => {} // unknown tag — a newer host; ignore
         }
     }

crates/punktfunk-core/src/client/pump/handshake.rs

@@ -143,6 +143,10 @@ pub(super) async fn connect_and_handshake(args: &WorkerArgs) -> Result<Handshake
                 // The client display's HDR volume → the host's virtual-display EDID (host apps
                 // tone-map to the client's real panel). `None` = unknown/SDR.
                 display_hdr,
+                // NOT unconditional like HOST_TIMING above: CLIENT_CAP_CURSOR makes the host
+                // stop compositing the pointer, so only an embedder that actually renders the
+                // cursor locally may set it (the embedder decides, we pass through).
+                client_caps: args.client_caps,
             }
             .encode(),
         )

crates/punktfunk-core/src/client/worker.rs

@@ -22,6 +22,7 @@ pub(crate) struct WorkerArgs {
     pub(crate) video_codecs: u8,
     pub(crate) preferred_codec: u8,
     pub(crate) display_hdr: Option<HdrMeta>,
+    pub(crate) client_caps: u8,
     pub(crate) launch: Option<String>,
     pub(crate) pin: Option<[u8; 32]>,
     pub(crate) identity: Option<(String, String)>,
@@ -38,6 +39,8 @@ pub(crate) struct WorkerArgs {
     pub(crate) hidout_tx: SyncSender<HidOutput>,
     pub(crate) hdr_meta_tx: SyncSender<HdrMeta>,
     pub(crate) host_timing_tx: SyncSender<crate::quic::HostTiming>,
+    pub(crate) cursor_shape_tx: SyncSender<crate::quic::CursorShape>,
+    pub(crate) cursor_state_tx: SyncSender<crate::quic::CursorState>,
     pub(crate) input_rx: tokio::sync::mpsc::UnboundedReceiver<InputEvent>,
     pub(crate) mic_rx: tokio::sync::mpsc::Receiver<(u32, u64, Vec<u8>)>,
     pub(crate) rich_input_rx: tokio::sync::mpsc::UnboundedReceiver<RichInput>,

crates/punktfunk-core/src/quic/caps.rs

@@ -71,6 +71,24 @@ pub const HOST_CAP_GAMEPAD_STATE: u8 = 0x01;
 /// trailing `host_caps` byte — no wire-layout change.
 pub const HOST_CAP_CLIPBOARD: u8 = 0x02;
 
+/// [`Hello::client_caps`] bit: the client renders the host cursor LOCALLY
+/// (design/remote-desktop-sweep.md M2). It consumes [`CursorShape`](super::control::CursorShape)
+/// control messages (RGBA bitmap + hotspot, cached by serial) and per-frame
+/// [`CursorState`](super::datagram::CursorState) `0xD0` datagrams (position/visibility), and
+/// draws the pointer itself — so the host must STOP compositing the cursor into the video
+/// (`SessionPlan.cursor_blend = false`) or the user sees it twice. Active only when the host
+/// answers with [`HOST_CAP_CURSOR`] (capable-and-agreed, the 444/clipboard precedent); toward
+/// an older or incapable host nothing changes.
+pub const CLIENT_CAP_CURSOR: u8 = 0x01;
+
+/// [`Welcome::host_caps`] bit: the host CAN forward the cursor out-of-band (it captures cursor
+/// metadata separately from the frame — the Linux portal `SPA_META_Cursor` path; NOT gamescope,
+/// whose capture carries no cursor, and NOT Windows yet, where DWM composites into the IDD
+/// frame). Set only when the client asked via [`CLIENT_CAP_CURSOR`]; when both bits agree the
+/// host stops blending and ships [`CursorShape`](super::control::CursorShape) +
+/// [`CursorState`](super::datagram::CursorState) instead.
+pub const HOST_CAP_CURSOR: u8 = 0x04;
+
 /// [`Hello::video_codecs`] bit: the client can decode H.264 / AVC. The GPU-less **software**
 /// encode path (openh264) emits H.264, so a client that wants to stream from a software host MUST
 /// advertise this.

crates/punktfunk-core/src/quic/control.rs

@@ -783,6 +783,83 @@ impl ClipFetchHdr {
     }
 }
 
+// --- Cursor channel (design/remote-desktop-sweep.md M2) --------------------------------------
+// The host cursor, forwarded out-of-band so the CLIENT draws it as a real OS cursor (the
+// Parsec/RDP model) instead of paying the video round-trip. Shape (rare, needs reliability)
+// rides here on the control stream; per-frame position/visibility rides the lossy `0xD0`
+// datagram plane ([`super::datagram::CursorState`]). Active only when the client's
+// [`CLIENT_CAP_CURSOR`](super::caps::CLIENT_CAP_CURSOR) met the host's
+// [`HOST_CAP_CURSOR`](super::caps::HOST_CAP_CURSOR) — the host stops compositing then.
+// ---------------------------------------------------------------------------------------------
+
+/// Type byte of [`CursorShape`] (host → client): the pointer's bitmap + hotspot changed.
+pub const MSG_CURSOR_SHAPE: u8 = 0x50;
+
+/// Per-side pixel cap for a forwarded cursor bitmap. The control-stream frame is length-prefixed
+/// with a `u16`, so a whole message must fit 65535 bytes — 128×128 RGBA (65536 B) already
+/// overshoots before the 17-byte header. 120² (57.6 KiB + header) fits with headroom and covers
+/// real cursors (typically ≤ 64 px, ≤ 96 px at HiDPI scale); the HOST downscales anything
+/// larger before forwarding, so the cap is invisible to clients.
+pub const CURSOR_SHAPE_MAX_SIDE: u16 = 120;
+
+/// `host → client` ([`MSG_CURSOR_SHAPE`]): one cursor shape, sent when the pointer's bitmap
+/// changes (never per-frame — [`super::datagram::CursorState`] carries the motion). The client
+/// caches shapes by `serial` and re-installs a cached one without any bitmap crossing again
+/// (the RDP pointer-cache idea for free: re-showing a known serial is a 14-byte
+/// [`super::datagram::CursorState`], not a resend).
+#[derive(Clone, Debug, PartialEq, Eq)]
+pub struct CursorShape {
+    /// Bitmap identity — bumped by the host's capture layer only on shape change; position
+    /// moves keep the serial stable. [`super::datagram::CursorState::serial`] references it.
+    pub serial: u32,
+    /// Bitmap dimensions in pixels, `1..=`[`CURSOR_SHAPE_MAX_SIDE`] each.
+    pub w: u16,
+    pub h: u16,
+    /// Hotspot (the pixel that IS the pointer position), within `w`×`h`.
+    pub hot_x: u16,
+    pub hot_y: u16,
+    /// Straight-alpha RGBA8, exactly `w * h * 4` bytes, no padding.
+    pub rgba: Vec<u8>,
+}
+
+impl CursorShape {
+    pub fn encode(&self) -> Vec<u8> {
+        // magic[0..4] type[4] serial[5..9] w[9..11] h[11..13] hot_x[13..15] hot_y[15..17] rgba…
+        let mut b = Vec::with_capacity(17 + self.rgba.len());
+        b.extend_from_slice(CTL_MAGIC);
+        b.push(MSG_CURSOR_SHAPE);
+        b.extend_from_slice(&self.serial.to_le_bytes());
+        b.extend_from_slice(&self.w.to_le_bytes());
+        b.extend_from_slice(&self.h.to_le_bytes());
+        b.extend_from_slice(&self.hot_x.to_le_bytes());
+        b.extend_from_slice(&self.hot_y.to_le_bytes());
+        b.extend_from_slice(&self.rgba);
+        b
+    }
+
+    pub fn decode(b: &[u8]) -> Result<CursorShape> {
+        if b.len() < 17 || &b[0..4] != CTL_MAGIC || b[4] != MSG_CURSOR_SHAPE {
+            return Err(PunktfunkError::InvalidArg("bad CursorShape"));
+        }
+        let u16at = |o: usize| u16::from_le_bytes([b[o], b[o + 1]]);
+        let (w, h) = (u16at(9), u16at(11));
+        if w == 0 || h == 0 || w > CURSOR_SHAPE_MAX_SIDE || h > CURSOR_SHAPE_MAX_SIDE {
+            return Err(PunktfunkError::InvalidArg("bad CursorShape dims"));
+        }
+        if b.len() != 17 + (w as usize) * (h as usize) * 4 {
+            return Err(PunktfunkError::InvalidArg("bad CursorShape len"));
+        }
+        Ok(CursorShape {
+            serial: u32::from_le_bytes(b[5..9].try_into().unwrap()),
+            w,
+            h,
+            hot_x: u16at(13),
+            hot_y: u16at(15),
+            rgba: b[17..].to_vec(),
+        })
+    }
+}
+
 #[cfg(test)]
 mod tests {
     use crate::config::Mode;
@@ -1147,4 +1224,42 @@ mod tests {
         assert!(ClipFetchHdr::decode(&[bytes.as_slice(), &[0]].concat()).is_err());
         assert!(ClipFetchHdr::decode(&bytes[..bytes.len() - 1]).is_err());
     }
+    #[test]
+    fn cursor_shape_roundtrip() {
+        let s = CursorShape {
+            serial: 7,
+            w: 2,
+            h: 3,
+            hot_x: 1,
+            hot_y: 2,
+            rgba: (0..2 * 3 * 4).map(|i| i as u8).collect(),
+        };
+        assert_eq!(CursorShape::decode(&s.encode()).unwrap(), s);
+        // Max-side shape still fits the u16 control frame with headroom.
+        let side = CURSOR_SHAPE_MAX_SIDE;
+        let big = CursorShape {
+            serial: u32::MAX,
+            w: side,
+            h: side,
+            hot_x: side - 1,
+            hot_y: 0,
+            rgba: vec![0xAB; side as usize * side as usize * 4],
+        };
+        let bytes = big.encode();
+        assert!(bytes.len() <= u16::MAX as usize, "must fit a control frame");
+        assert_eq!(CursorShape::decode(&bytes).unwrap(), big);
+        // Rejections: zero / oversize dims, and a length that disagrees with them.
+        let mut zero = s.encode();
+        zero[9] = 0;
+        zero[10] = 0;
+        assert!(CursorShape::decode(&zero).is_err());
+        let mut oversize = s.encode();
+        oversize[9..11].copy_from_slice(&(CURSOR_SHAPE_MAX_SIDE + 1).to_le_bytes());
+        assert!(CursorShape::decode(&oversize).is_err());
+        let mut short = s.encode();
+        short.pop();
+        assert!(CursorShape::decode(&short).is_err());
+        // Distinct from the neighboring vocabulary.
+        assert!(ClipState::decode(&s.encode()).is_err());
+    }
 }

crates/punktfunk-core/src/quic/datagram.rs

@@ -606,6 +606,75 @@ pub fn decode_host_timing_datagram(b: &[u8]) -> Option<HostTiming> {
     })
 }
 
+/// Cursor-state datagram tag, host → client (design/remote-desktop-sweep.md M2). Next tag after
+/// [`HOST_TIMING_MAGIC`]. Sent once per captured frame while the cursor channel is negotiated
+/// ([`CLIENT_CAP_CURSOR`](super::caps::CLIENT_CAP_CURSOR) ∧
+/// [`HOST_CAP_CURSOR`](super::caps::HOST_CAP_CURSOR)) — per-frame resend makes the plane
+/// self-healing under loss (latest-wins, no refresh timer). The bitmap itself rides the
+/// reliable control stream ([`CursorShape`](super::control::CursorShape)); this 14-byte
+/// datagram only moves/hides the pointer.
+pub const CURSOR_STATE_MAGIC: u8 = 0xD0;
+
+/// [`CursorState::flags`] bit: the host cursor is visible.
+pub const CURSOR_VISIBLE: u8 = 0x01;
+/// [`CursorState::flags`] bit: a host app captured/hid the pointer — the client SHOULD run
+/// relative/captured (M3 auto-flip; advisory, user override always wins).
+pub const CURSOR_RELATIVE_HINT: u8 = 0x02;
+
+/// Per-frame host-cursor state (position, visibility, mode hint). `x`/`y` are the pointer
+/// position (hotspot point, not bitmap top-left) in the host OUTPUT's pixel space — the same
+/// space the video mode describes, so the client maps through its letterbox exactly like it
+/// maps touches, in reverse.
+#[derive(Clone, Copy, Debug, PartialEq, Eq)]
+pub struct CursorState {
+    /// The [`CursorShape`](super::control::CursorShape) serial this state refers to. A client
+    /// that has no cached shape for it keeps its previous cursor until the (reliable) shape
+    /// message lands — at worst one control-stream RTT of stale shape, never a wrong position.
+    pub serial: u32,
+    /// Bitfield of [`CURSOR_VISIBLE`] / [`CURSOR_RELATIVE_HINT`].
+    pub flags: u8,
+    pub x: i32,
+    pub y: i32,
+}
+
+impl CursorState {
+    pub fn visible(&self) -> bool {
+        self.flags & CURSOR_VISIBLE != 0
+    }
+    pub fn relative_hint(&self) -> bool {
+        self.flags & CURSOR_RELATIVE_HINT != 0
+    }
+}
+
+/// Wire length of a [`CURSOR_STATE_MAGIC`] datagram: tag + u32 serial + flags + 2 × i32 = 14.
+const CURSOR_STATE_LEN: usize = 1 + 4 + 1 + 8;
+
+/// Encode a [`CursorState`] into a [`CURSOR_STATE_MAGIC`] datagram.
+pub fn encode_cursor_state_datagram(s: &CursorState) -> Vec<u8> {
+    let mut b = Vec::with_capacity(CURSOR_STATE_LEN);
+    b.push(CURSOR_STATE_MAGIC);
+    b.extend_from_slice(&s.serial.to_le_bytes());
+    b.push(s.flags);
+    b.extend_from_slice(&s.x.to_le_bytes());
+    b.extend_from_slice(&s.y.to_le_bytes());
+    b
+}
+
+/// Parse a [`CURSOR_STATE_MAGIC`] datagram → [`CursorState`]. `None` on bad tag or a short
+/// buffer (the fixed length bounds every read before it happens; a longer buffer is tolerated
+/// for append-extension, like 0xCF).
+pub fn decode_cursor_state_datagram(b: &[u8]) -> Option<CursorState> {
+    if b.len() < CURSOR_STATE_LEN || b[0] != CURSOR_STATE_MAGIC {
+        return None;
+    }
+    Some(CursorState {
+        serial: u32::from_le_bytes(b[1..5].try_into().unwrap()),
+        flags: b[5],
+        x: i32::from_le_bytes(b[6..10].try_into().unwrap()),
+        y: i32::from_le_bytes(b[10..14].try_into().unwrap()),
+    })
+}
+
 #[cfg(test)]
 mod tests {
     use crate::quic::*;
@@ -922,4 +991,32 @@ mod tests {
         )
         .is_none());
     }
+    #[test]
+    fn cursor_state_roundtrip() {
+        for (flags, x, y) in [
+            (CURSOR_VISIBLE, 0i32, 0i32),
+            (CURSOR_VISIBLE | CURSOR_RELATIVE_HINT, -5, 2160),
+            (0, i32::MIN, i32::MAX),
+        ] {
+            let s = CursorState {
+                serial: 42,
+                flags,
+                x,
+                y,
+            };
+            let d = encode_cursor_state_datagram(&s);
+            assert_eq!(decode_cursor_state_datagram(&d), Some(s));
+            assert_eq!(s.visible(), flags & CURSOR_VISIBLE != 0);
+            assert_eq!(s.relative_hint(), flags & CURSOR_RELATIVE_HINT != 0);
+            // Append-extensible like 0xCF: a longer buffer still parses the known prefix.
+            let mut ext = d.clone();
+            ext.push(0xFF);
+            assert_eq!(decode_cursor_state_datagram(&ext), Some(s));
+            // Short / wrong tag are rejected before any read.
+            assert_eq!(decode_cursor_state_datagram(&d[..d.len() - 1]), None);
+            let mut bad = d.clone();
+            bad[0] = HOST_TIMING_MAGIC;
+            assert_eq!(decode_cursor_state_datagram(&bad), None);
+        }
+    }
 }

crates/punktfunk-core/src/quic/handshake.rs

@@ -83,6 +83,15 @@ pub struct Hello {
     /// forcing the earlier placeholders. Omitted by older clients / when the client has no HDR
     /// display (decodes to `None` — the host keeps its built-in EDID defaults).
     pub display_hdr: Option<HdrMeta>,
+    /// Non-video client capabilities — a bitfield of [`CLIENT_CAP_CURSOR`] (the client renders
+    /// the host cursor locally; the host stops compositing it and forwards shape + state
+    /// instead). Appended as a single byte AFTER `display_hdr`; because that block is a fixed
+    /// [`super::datagram::HDR_META_BODY_LEN`]-byte optional with no placeholder form, presence is
+    /// disambiguated by REMAINING LENGTH at decode: fewer than `HDR_META_BODY_LEN` bytes after
+    /// `preferred_codec` ⇒ no HDR block, the tail bytes are the post-HDR fields directly. This
+    /// caps everything after `display_hdr` at `HDR_META_BODY_LEN − 1` bytes total — document any
+    /// future field here and mind the budget. Omitted when zero and by older clients (→ `0`).
+    pub client_caps: u8,
 }
 
 /// QUIC application error code a punktfunk/1 client closes the control connection with on a
@@ -244,8 +253,13 @@ impl Hello {
         let vcodecs_present = self.video_codecs != 0;
         let pref_present = self.preferred_codec != 0;
         let hdr_present = self.display_hdr.is_some();
-        let need_placeholders =
-            self.video_caps != 0 || ac_present || vcodecs_present || pref_present || hdr_present;
+        let ccaps_present = self.client_caps != 0;
+        let need_placeholders = self.video_caps != 0
+            || ac_present
+            || vcodecs_present
+            || pref_present
+            || hdr_present
+            || ccaps_present;
         match (&self.name, &self.launch) {
             (None, None) if !need_placeholders => {}
             (name, _) => {
@@ -266,21 +280,27 @@ impl Hello {
             b.push(self.video_caps);
         }
         // audio_channels: emitted when non-stereo OR a later field follows.
-        if ac_present || vcodecs_present || pref_present || hdr_present {
+        if ac_present || vcodecs_present || pref_present || hdr_present || ccaps_present {
             b.push(self.audio_channels);
         }
         // video_codecs: emitted when non-zero OR a later field follows.
-        if vcodecs_present || pref_present || hdr_present {
+        if vcodecs_present || pref_present || hdr_present || ccaps_present {
             b.push(self.video_codecs);
         }
-        // preferred_codec: emitted when non-zero OR display_hdr follows.
-        if pref_present || hdr_present {
+        // preferred_codec: emitted when non-zero OR a later field follows.
+        if pref_present || hdr_present || ccaps_present {
             b.push(self.preferred_codec);
         }
-        // display_hdr: fixed HDR_META_BODY_LEN-byte HdrMeta body. Last field; omitted when `None`.
+        // display_hdr: fixed HDR_META_BODY_LEN-byte HdrMeta body; omitted when `None` even if
+        // later fields follow (no placeholder form — the decoder disambiguates by remaining
+        // length, which caps the post-HDR tail at HDR_META_BODY_LEN − 1 bytes).
         if let Some(m) = &self.display_hdr {
             super::datagram::write_hdr_meta_body(m, &mut b);
         }
+        // client_caps: single byte after the (optional) HDR block. Emitted when non-zero.
+        if ccaps_present {
+            b.push(self.client_caps);
+        }
         b
     }
 
@@ -346,9 +366,26 @@ impl Hello {
             preferred_codec: b.get(tail + 3).copied().unwrap_or(0),
             // Optional trailing HdrMeta body (fixed length) — absent on an older client / a
             // client without an HDR display → `None` (the host keeps its EDID defaults).
-            display_hdr: b
-                .get(tail + 4..tail + 4 + super::datagram::HDR_META_BODY_LEN)
-                .map(super::datagram::read_hdr_meta_body),
+            // Presence is decided by REMAINING LENGTH (there is no placeholder form for the
+            // fixed block): ≥ HDR_META_BODY_LEN bytes after `preferred_codec` ⇒ the block is
+            // there and post-HDR fields follow it; fewer ⇒ no block, the bytes ARE the post-HDR
+            // fields. Sound as long as the post-HDR tail stays under HDR_META_BODY_LEN bytes.
+            display_hdr: (b.len().saturating_sub(tail + 4) >= super::datagram::HDR_META_BODY_LEN)
+                .then(|| {
+                    b.get(tail + 4..tail + 4 + super::datagram::HDR_META_BODY_LEN)
+                        .map(super::datagram::read_hdr_meta_body)
+                })
+                .flatten(),
+            // client_caps: the byte after the HDR block when present, else directly at tail+4.
+            client_caps: {
+                let off = if b.len().saturating_sub(tail + 4) >= super::datagram::HDR_META_BODY_LEN
+                {
+                    tail + 4 + super::datagram::HDR_META_BODY_LEN
+                } else {
+                    tail + 4
+                };
+                b.get(off).copied().unwrap_or(0)
+            },
         })
     }
 }
@@ -829,6 +866,7 @@ mod tests {
             video_codecs: CODEC_H264 | CODEC_HEVC,
             preferred_codec: CODEC_H264,
             display_hdr: None,
+            client_caps: 0,
         };
         let enc = h.encode();
         let dec = Hello::decode(&enc).unwrap();
@@ -905,6 +943,7 @@ mod tests {
             video_codecs: CODEC_H264 | CODEC_HEVC, // exercise the codec bitfield roundtrip
             preferred_codec: CODEC_HEVC,
             display_hdr: None,
+            client_caps: 0,
         };
         assert_eq!(Hello::decode(&h.encode()).unwrap(), h);
         let s = Start {
@@ -935,6 +974,7 @@ mod tests {
             video_codecs: 0,
             preferred_codec: 0,
             display_hdr: None,
+            client_caps: 0,
         };
         let enc = h.encode();
         assert_eq!(enc.len(), 26);
@@ -1052,6 +1092,7 @@ mod tests {
             video_codecs: 0,
             preferred_codec: 0,
             display_hdr: None,
+            client_caps: 0,
         };
         let enc = base.encode();
         assert_eq!(
@@ -1103,6 +1144,7 @@ mod tests {
             video_codecs: 0,
             preferred_codec: 0,
             display_hdr: None,
+            client_caps: 0,
         };
         // launch alone (no name): a zero-length name placeholder keeps the offset deterministic.
         let with_launch = Hello {
@@ -1162,6 +1204,7 @@ mod tests {
             video_codecs: 0,
             preferred_codec: 0,
             display_hdr: None,
+            client_caps: 0,
         };
         // A real client-panel volume (P3 primaries, 800-nit peak, 0.05-nit floor, 400-nit FALL).
         let vol = HdrMeta {
@@ -1229,6 +1272,7 @@ mod tests {
                 video_codecs: 0,
                 preferred_codec: 0,
                 display_hdr: None,
+                client_caps: 0,
             }
             .encode();
             assert!(PairRequest::decode(&h).is_err(), "abi {abi} parsed as pair");
@@ -1242,4 +1286,66 @@ mod tests {
         .encode();
         assert!(Hello::decode(&pr).is_err());
     }
+    #[test]
+    fn hello_client_caps_roundtrip_and_back_compat() {
+        let base = Hello {
+            abi_version: 2,
+            mode: Mode {
+                width: 1920,
+                height: 1080,
+                refresh_hz: 60,
+            },
+            compositor: CompositorPref::Auto,
+            gamepad: GamepadPref::Auto,
+            bitrate_kbps: 0,
+            name: None,
+            launch: None,
+            video_caps: 0,
+            audio_channels: 2,
+            video_codecs: 0,
+            preferred_codec: 0,
+            display_hdr: None,
+            client_caps: 0,
+        };
+        let vol = HdrMeta {
+            display_primaries: [[13250, 34500], [7500, 3000], [34000, 16000]],
+            white_point: [15635, 16450],
+            max_display_mastering_luminance: 8_000_000,
+            min_display_mastering_luminance: 500,
+            max_cll: 0,
+            max_fall: 400,
+        };
+        // caps WITHOUT an HDR block: the single byte after preferred_codec (remaining < the
+        // fixed block length, so the decoder must NOT read it as a truncated HdrMeta).
+        let caps_only = Hello {
+            client_caps: CLIENT_CAP_CURSOR,
+            ..base.clone()
+        };
+        assert_eq!(Hello::decode(&caps_only.encode()).unwrap(), caps_only);
+        // caps AND the HDR block: caps lands after the fixed block.
+        let both = Hello {
+            display_hdr: Some(vol),
+            client_caps: CLIENT_CAP_CURSOR,
+            ..base.clone()
+        };
+        assert_eq!(Hello::decode(&both.encode()).unwrap(), both);
+        // HDR without caps stays byte-identical to the pre-caps wire form and decodes caps 0.
+        let hdr_only = Hello {
+            display_hdr: Some(vol),
+            ..base.clone()
+        };
+        assert_eq!(Hello::decode(&hdr_only.encode()).unwrap(), hdr_only);
+        // An older client (no trailing byte at all) decodes to 0.
+        assert_eq!(Hello::decode(&base.encode()).unwrap().client_caps, 0);
+        // An older HOST reading a caps-bearing Hello: its decode simply never looks past the
+        // fields it knows — nothing before the caps byte moved.
+        let enc = both.encode();
+        assert_eq!(
+            Hello::decode(&enc[..enc.len() - 1]).unwrap(),
+            Hello {
+                client_caps: 0,
+                ..both.clone()
+            }
+        );
+    }
 }

crates/punktfunk-host/src/native.rs

@@ -76,6 +76,8 @@ mod handshake;
 /// The mid-stream control task (plan §W1); `serve_session` spawns `control::run` after the
 /// handshake to multiplex renegotiation / speed-test control messages onto the data-plane channels.
 mod control;
+/// Cursor-forward channel (M2): the encode loop's shape/state emission.
+mod cursor_fwd;
 
 /// The capture→encode→send data plane (plan §W1); `serve_session` dispatches the synthetic or
 /// virtual source here (`synthetic_stream` / `virtual_stream`) and hands the latter a
@@ -943,6 +945,15 @@ async fn serve_session(
     // accepted ack as "the active mode is now X" and fixes itself; old clients just log it.
     let (reconfig_result_tx, reconfig_result_rx) =
         tokio::sync::mpsc::unbounded_channel::<Reconfigured>();
+    // Cursor-forward bridge (M2): the encode loop diffs each frame's cursor serial and hands
+    // changed SHAPES here; the control task (the control stream's sole writer) sends them.
+    // Same shape as `probe_result_tx`. Wired even when the channel wasn't negotiated — it
+    // just never fires then.
+    let (cursor_shape_tx, cursor_shape_rx) =
+        tokio::sync::mpsc::unbounded_channel::<punktfunk_core::quic::CursorShape>();
+    // Negotiated cursor forwarding: MUST match the HOST_CAP_CURSOR bit the Welcome advertised
+    // (handshake::cursor_forward is the single predicate both read).
+    let cursor_forward = handshake::cursor_forward(hello.client_caps, compositor);
     // Adaptive FEC: the control task maps each client LossReport to a recovery percent and publishes
     // it here; the data-plane send loop reads + applies it per frame. Disabled (pinned) when
     // PUNKTFUNK_FEC_PCT is set. Seeded with the session's starting FEC so it's a no-op until a report.
@@ -978,6 +989,7 @@ async fn serve_session(
         probe_tx,
         probe_result_rx,
         reconfig_result_rx,
+        cursor_shape_rx,
         clip_enabled,
         clip,
     ));
@@ -1363,6 +1375,8 @@ async fn serve_session(
                         fec_target: fec_target_dp,
                         conn: conn_stream,
                         timing_conn,
+                        cursor_forward,
+                        cursor_shape_tx,
                         probe_seq,
                         streamed_au,
                         stats: stats_dp,
@@ -2029,6 +2043,7 @@ mod tests {
             0,    // video_codecs (HEVC-only)
             0,    // preferred_codec
             None, // display_hdr
+            0,    // client_caps
             None, // launch
             None, // pin (TOFU)
             None, // identity (host doesn't require pairing)
@@ -2199,6 +2214,7 @@ mod tests {
             0,    // video_codecs (0 → HEVC-only)
             0,    // preferred_codec (auto)
             None, // display_hdr
+            0,    // client_caps
             None, // launch
             None, // pin: TOFU — the operator's approval (not a PIN) authorizes this client
             Some((cert, key)),
@@ -2266,6 +2282,7 @@ mod tests {
                 0,    // video_codecs
                 0,    // preferred_codec
                 None, // display_hdr
+                0,    // client_caps
                 None, // launch
                 None,
                 None,
@@ -2295,6 +2312,7 @@ mod tests {
             0,    // video_codecs
             0,    // preferred_codec
             None, // display_hdr
+            0,    // client_caps
             None, // launch
             Some(host_fp),
             Some((cert.clone(), key.clone())),

crates/punktfunk-host/src/native/control.rs

@@ -30,6 +30,7 @@ pub(super) async fn run(
     probe_tx: std::sync::mpsc::Sender<ProbeRequest>,
     mut probe_result_rx: tokio::sync::mpsc::UnboundedReceiver<ProbeResult>,
     mut reconfig_result_rx: tokio::sync::mpsc::UnboundedReceiver<Reconfigured>,
+    mut cursor_shape_rx: tokio::sync::mpsc::UnboundedReceiver<punktfunk_core::quic::CursorShape>,
     clip_enabled: Arc<AtomicBool>,
     clip: pf_clipboard::ClipCoord,
 ) {
@@ -262,6 +263,15 @@ pub(super) async fn run(
                     break;
                 }
             }
+            shape = cursor_shape_rx.recv() => {
+                // Cursor-forward bridge (M2): the encode loop diffed a new pointer bitmap.
+                // Rare (shape changes are human-paced); ≤ ~58 KiB fits the u16 frame by
+                // construction (cursor_fwd downscales).
+                let Some(shape) = shape else { break }; // data plane gone
+                if io::write_msg(&mut ctrl_send, &shape.encode()).await.is_err() {
+                    break;
+                }
+            }
             offer = clip_offer_rx.recv(), if !clip_offer_closed => {
                 // Host copied → the coordinator minted a `ClipOffer`; forward it to the client
                 // (only while sync is on — a race with a just-received disable would otherwise

crates/punktfunk-host/src/native/cursor_fwd.rs

@@ -0,0 +1,144 @@
+//! Cursor-forward channel, host side (design/remote-desktop-sweep.md M2).
+//!
+//! When the session negotiated the cursor channel (client `CLIENT_CAP_CURSOR` met our
+//! `HOST_CAP_CURSOR`), the encoder stops blending the pointer into the video
+//! (`SessionPlan::cursor_blend = false`) and the encode loop forwards it out-of-band instead:
+//! the SHAPE (bitmap + hotspot, rare) rides the reliable control stream via the control-task
+//! bridge, per-tick STATE (position/visibility, 14 B) rides a lossy `0xD0` datagram — resent
+//! every iteration so loss self-heals with no refresh timer.
+
+use punktfunk_core::quic::{
+    encode_cursor_state_datagram, CursorShape, CursorState, CURSOR_SHAPE_MAX_SIDE, CURSOR_VISIBLE,
+};
+
+/// Per-session forward state, owned by the encode loop (the thread that binds frames).
+pub(super) struct CursorForwarder {
+    /// Serial of the last shape handed to the control-task bridge (`None` = none yet).
+    sent_serial: Option<u64>,
+    /// Last visible pointer position (hotspot point, frame px) — held across hidden spans so
+    /// a hide still states WHERE the pointer was (the M3 reappear position).
+    last_pos: (i32, i32),
+}
+
+impl CursorForwarder {
+    pub(super) fn new() -> CursorForwarder {
+        CursorForwarder {
+            sent_serial: None,
+            last_pos: (0, 0),
+        }
+    }
+
+    /// Called once per encode-loop iteration with the bound frame's overlay (also on repeat
+    /// iterations — the state datagram is the plane's loss heal, so it goes out every tick).
+    /// `None` overlay = hidden pointer (or no bitmap yet): state only, `visible` clear.
+    pub(super) fn tick(
+        &mut self,
+        cursor: Option<&pf_frame::CursorOverlay>,
+        conn: &quinn::Connection,
+        shape_tx: &tokio::sync::mpsc::UnboundedSender<CursorShape>,
+    ) {
+        let flags = match cursor {
+            Some(ov) => {
+                if self.sent_serial != Some(ov.serial) {
+                    if let Some(shape) = shape_from_overlay(ov) {
+                        // Bridge full ⇒ control task gone ⇒ session is tearing down anyway.
+                        let _ = shape_tx.send(shape);
+                        self.sent_serial = Some(ov.serial);
+                    }
+                }
+                self.last_pos = (ov.x + ov.hot_x as i32, ov.y + ov.hot_y as i32);
+                CURSOR_VISIBLE
+            }
+            None => 0,
+        };
+        let state = CursorState {
+            serial: self.sent_serial.unwrap_or(0) as u32,
+            flags,
+            x: self.last_pos.0,
+            y: self.last_pos.1,
+        };
+        let _ = conn.send_datagram(encode_cursor_state_datagram(&state).into());
+    }
+}
+
+/// Build the wire shape from a capture overlay, integer-downscaling (nearest-neighbor) anything
+/// over [`CURSOR_SHAPE_MAX_SIDE`] so the message always fits the u16-length control frame.
+/// Real cursors are far under the cap — the scale path is a correctness backstop for XL
+/// accessibility cursors, not a quality path. `None` on a malformed overlay (short buffer).
+fn shape_from_overlay(ov: &pf_frame::CursorOverlay) -> Option<CursorShape> {
+    let px = (ov.w as usize).checked_mul(ov.h as usize)?.checked_mul(4)?;
+    if ov.w == 0 || ov.h == 0 || ov.rgba.len() < px {
+        return None;
+    }
+    let max = CURSOR_SHAPE_MAX_SIDE as u32;
+    let f = ov.w.max(ov.h).div_ceil(max).max(1);
+    let (w, h) = (ov.w.div_ceil(f), ov.h.div_ceil(f));
+    let rgba = if f == 1 {
+        ov.rgba.as_ref().clone()
+    } else {
+        let mut out = Vec::with_capacity((w * h * 4) as usize);
+        for y in 0..h {
+            for x in 0..w {
+                let (sx, sy) = ((x * f).min(ov.w - 1), (y * f).min(ov.h - 1));
+                let o = ((sy * ov.w + sx) * 4) as usize;
+                out.extend_from_slice(&ov.rgba[o..o + 4]);
+            }
+        }
+        out
+    };
+    Some(CursorShape {
+        serial: ov.serial as u32,
+        w: w as u16,
+        h: h as u16,
+        hot_x: (ov.hot_x / f).min(w - 1) as u16,
+        hot_y: (ov.hot_y / f).min(h - 1) as u16,
+        rgba,
+    })
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use std::sync::Arc;
+
+    fn overlay(w: u32, h: u32, hot: (u32, u32)) -> pf_frame::CursorOverlay {
+        pf_frame::CursorOverlay {
+            x: 10,
+            y: 20,
+            w,
+            h,
+            rgba: Arc::new((0..w * h * 4).map(|i| i as u8).collect()),
+            serial: 3,
+            hot_x: hot.0,
+            hot_y: hot.1,
+        }
+    }
+
+    #[test]
+    fn small_shape_passes_through() {
+        let s = shape_from_overlay(&overlay(32, 32, (4, 5))).unwrap();
+        assert_eq!((s.w, s.h, s.hot_x, s.hot_y, s.serial), (32, 32, 4, 5, 3));
+        assert_eq!(s.rgba.len(), 32 * 32 * 4);
+        // Encodes within the u16 control-frame cap.
+        assert!(s.encode().len() <= u16::MAX as usize);
+    }
+
+    #[test]
+    fn oversize_shape_downscales_with_hotspot() {
+        // 256² → f = ceil(256/120) = 3 → 86² (256.div_ceil(3)), hotspot scales with it.
+        let s = shape_from_overlay(&overlay(256, 256, (255, 0))).unwrap();
+        assert!(s.w <= CURSOR_SHAPE_MAX_SIDE && s.h <= CURSOR_SHAPE_MAX_SIDE);
+        assert_eq!(s.rgba.len(), s.w as usize * s.h as usize * 4);
+        assert!(s.hot_x < s.w && s.hot_y < s.h);
+        assert!(s.encode().len() <= u16::MAX as usize);
+        // The scaled message must decode (dims within the cap).
+        assert_eq!(CursorShape::decode(&s.encode()).unwrap(), s);
+    }
+
+    #[test]
+    fn short_buffer_rejected() {
+        let mut ov = overlay(8, 8, (0, 0));
+        ov.rgba = Arc::new(vec![0; 8]);
+        assert!(shape_from_overlay(&ov).is_none());
+    }
+}

crates/punktfunk-host/src/native/handshake.rs

@@ -8,6 +8,22 @@
 
 use super::*;
 
+/// Whether this session forwards the cursor out-of-band (design/remote-desktop-sweep.md M2):
+/// the client asked ([`CLIENT_CAP_CURSOR`](punktfunk_core::quic::CLIENT_CAP_CURSOR)) AND the
+/// capture path can deliver cursor metadata separately from the frame — today that is the
+/// Linux portal `SPA_META_Cursor` path only: not gamescope (its capture paints no cursor at
+/// all), not Windows (DWM composites into the IDD frame — M2c). THE single predicate: the
+/// Welcome's `HOST_CAP_CURSOR` bit and the session's forwarding/blend-off wiring both read it,
+/// so they can never disagree.
+pub(super) fn cursor_forward(
+    client_caps: u8,
+    compositor: Option<crate::vdisplay::Compositor>,
+) -> bool {
+    cfg!(target_os = "linux")
+        && client_caps & punktfunk_core::quic::CLIENT_CAP_CURSOR != 0
+        && compositor.is_some_and(|c| c != crate::vdisplay::Compositor::Gamescope)
+}
+
 /// Run the Hello→Welcome→Start negotiation. Borrows the control streams (the caller keeps them for
 /// mid-stream renegotiation afterwards). `first` is the already-read first control message.
 #[allow(clippy::type_complexity, clippy::too_many_arguments)]
@@ -444,6 +460,14 @@ pub(super) async fn negotiate(
                 punktfunk_core::quic::HOST_CAP_CLIPBOARD
             } else {
                 0
+            }
+            // Cursor channel granted (client asked + this capture path can deliver cursor
+            // metadata out of the frame) — the client turns its local renderer on ONLY when
+            // it sees this bit, and serve_session wires forwarding from the same predicate.
+            | if cursor_forward(hello.client_caps, compositor) {
+                punktfunk_core::quic::HOST_CAP_CURSOR
+            } else {
+                0
             },
         // The negotiated session AEAD (resolved above) + its 32-byte key toward a ChaCha
         // client; toward everyone else cipher 0 keeps the Welcome byte-identical to the

crates/punktfunk-host/src/native/stream.rs

@@ -938,6 +938,14 @@ pub(super) struct SessionContext {
     /// thread emits one 0xCF datagram per AU (capture→sent µs) on it, so the client can split its
     /// `host+network` latency stage. `None` = older client, no emission.
     pub(super) timing_conn: Option<quinn::Connection>,
+    /// The session negotiated the cursor channel (design/remote-desktop-sweep.md M2 —
+    /// `handshake::cursor_forward`): the encoder does NOT blend the pointer into the video;
+    /// the encode loop forwards shape (via `cursor_shape_tx`) + per-tick `0xD0` state instead.
+    pub(super) cursor_forward: bool,
+    /// SHAPE bridge to the control task (the control stream's sole writer) — mirrors
+    /// `probe_result_tx`. Inert when `cursor_forward` is false.
+    pub(super) cursor_shape_tx:
+        tokio::sync::mpsc::UnboundedSender<punktfunk_core::quic::CursorShape>,
     /// The client advertised [`punktfunk_core::quic::VIDEO_CAP_PROBE_SEQ`]: speed-test bursts may
     /// run mid-session in the probe index space (its reassembler keeps a separate probe window).
     /// `false` = older client whose single-window reassembler would drop probe-space frames as
@@ -987,7 +995,10 @@ pub(super) fn virtual_stream(ctx: SessionContext, prepared: Option<PreparedDispl
         ctx.bit_depth,
         ctx.chroma,
         ctx.codec,
-        ctx.compositor != pf_vdisplay::Compositor::Gamescope,
+        // Blend the pointer into the video only where the capture HAS one (not gamescope) AND
+        // the client is not drawing it locally (the M2 cursor channel — blending too would
+        // show it twice).
+        ctx.compositor != pf_vdisplay::Compositor::Gamescope && !ctx.cursor_forward,
     );
     // PyroWave rides the datagram-aligned wire mode (§4.4): every encoder this session opens
     // packetizes at the negotiated shard payload, so a lost datagram costs blocks, not frames.
@@ -1021,6 +1032,8 @@ pub(super) fn virtual_stream(ctx: SessionContext, prepared: Option<PreparedDispl
         fec_target,
         conn,
         timing_conn,
+        cursor_forward,
+        cursor_shape_tx,
         probe_seq,
         streamed_au,
         stats,
@@ -1034,6 +1047,13 @@ pub(super) fn virtual_stream(ctx: SessionContext, prepared: Option<PreparedDispl
     // reverts to whole-AU sends without touching the encoder's slicing knobs). The third gate —
     // whether the ENCODER actually chunks — is dynamic (`supports_chunked_poll`, per AU).
     let streamed_wire = streamed_au && std::env::var("PUNKTFUNK_STREAMED_AU").as_deref() != Ok("0");
+    // Cursor-forward state (M2): shape-serial diffing + the per-tick 0xD0 state send. The
+    // encoder was told not to blend (SessionPlan above), so from the first frame the client's
+    // locally-drawn cursor is the only one.
+    let mut cursor_fwd = cursor_forward.then(super::cursor_fwd::CursorForwarder::new);
+    if cursor_forward {
+        tracing::info!("cursor channel negotiated — forwarding shape/state, encoder blend off");
+    }
     if streamed_wire {
         tracing::info!(
             "client accepts streamed AUs (VIDEO_CAP_STREAMED_AU) — chunked encoder output \
@@ -1985,6 +2005,12 @@ pub(super) fn virtual_stream(ctx: SessionContext, prepared: Option<PreparedDispl
                 );
             }
         }
+        // Cursor channel (M2): every iteration — new frame OR repeat — states the pointer
+        // (self-healing under datagram loss) and forwards a changed shape via the control
+        // bridge. `frame` is the newest bound frame either way.
+        if let Some(fwd) = cursor_fwd.as_mut() {
+            fwd.tick(frame.cursor.as_ref(), &conn, &cursor_shape_tx);
+        }
         if perf && diag_at.elapsed() >= std::time::Duration::from_secs(2) {
             let secs = diag_at.elapsed().as_secs_f64();
             tracing::info!(

include/punktfunk_core.h

@@ -498,6 +498,28 @@
 #define HOST_CAP_CLIPBOARD 2
 #endif
 
+#if defined(PUNKTFUNK_FEATURE_QUIC)
+// [`Hello::client_caps`] bit: the client renders the host cursor LOCALLY
+// (design/remote-desktop-sweep.md M2). It consumes [`CursorShape`](super::control::CursorShape)
+// control messages (RGBA bitmap + hotspot, cached by serial) and per-frame
+// [`CursorState`](super::datagram::CursorState) `0xD0` datagrams (position/visibility), and
+// draws the pointer itself — so the host must STOP compositing the cursor into the video
+// (`SessionPlan.cursor_blend = false`) or the user sees it twice. Active only when the host
+// answers with [`HOST_CAP_CURSOR`] (capable-and-agreed, the 444/clipboard precedent); toward
+// an older or incapable host nothing changes.
+#define CLIENT_CAP_CURSOR 1
+#endif
+
+#if defined(PUNKTFUNK_FEATURE_QUIC)
+// [`Welcome::host_caps`] bit: the host CAN forward the cursor out-of-band (it captures cursor
+// metadata separately from the frame — the Linux portal `SPA_META_Cursor` path; NOT gamescope,
+// whose capture carries no cursor, and NOT Windows yet, where DWM composites into the IDD
+// frame). Set only when the client asked via [`CLIENT_CAP_CURSOR`]; when both bits agree the
+// host stops blending and ships [`CursorShape`](super::control::CursorShape) +
+// [`CursorState`](super::datagram::CursorState) instead.
+#define HOST_CAP_CURSOR 4
+#endif
+
 #if defined(PUNKTFUNK_FEATURE_QUIC)
 // [`Hello::video_codecs`] bit: the client can decode H.264 / AVC. The GPU-less **software**
 // encode path (openh264) emits H.264, so a client that wants to stream from a software host MUST
@@ -750,6 +772,20 @@
 #define CLIP_FILE_INDEX_NONE UINT32_MAX
 #endif
 
+#if defined(PUNKTFUNK_FEATURE_QUIC)
+// Type byte of [`CursorShape`] (host → client): the pointer's bitmap + hotspot changed.
+#define MSG_CURSOR_SHAPE 80
+#endif
+
+#if defined(PUNKTFUNK_FEATURE_QUIC)
+// Per-side pixel cap for a forwarded cursor bitmap. The control-stream frame is length-prefixed
+// with a `u16`, so a whole message must fit 65535 bytes — 128×128 RGBA (65536 B) already
+// overshoots before the 17-byte header. 120² (57.6 KiB + header) fits with headroom and covers
+// real cursors (typically ≤ 64 px, ≤ 96 px at HiDPI scale); the HOST downscales anything
+// larger before forwarding, so the cap is invisible to clients.
+#define CURSOR_SHAPE_MAX_SIDE 120
+#endif
+
 #if defined(PUNKTFUNK_FEATURE_QUIC)
 // Datagram wire tags. Video rides UDP; everything low-rate rides QUIC datagrams,
 // demultiplexed by the first byte: input = [`crate::input::INPUT_MAGIC`] (0xC8, client→host),
@@ -838,6 +874,28 @@
 #define HOST_TIMING_MAGIC 207
 #endif
 
+#if defined(PUNKTFUNK_FEATURE_QUIC)
+// Cursor-state datagram tag, host → client (design/remote-desktop-sweep.md M2). Next tag after
+// [`HOST_TIMING_MAGIC`]. Sent once per captured frame while the cursor channel is negotiated
+// ([`CLIENT_CAP_CURSOR`](super::caps::CLIENT_CAP_CURSOR) ∧
+// [`HOST_CAP_CURSOR`](super::caps::HOST_CAP_CURSOR)) — per-frame resend makes the plane
+// self-healing under loss (latest-wins, no refresh timer). The bitmap itself rides the
+// reliable control stream ([`CursorShape`](super::control::CursorShape)); this 14-byte
+// datagram only moves/hides the pointer.
+#define CURSOR_STATE_MAGIC 208
+#endif
+
+#if defined(PUNKTFUNK_FEATURE_QUIC)
+// [`CursorState::flags`] bit: the host cursor is visible.
+#define CURSOR_VISIBLE 1
+#endif
+
+#if defined(PUNKTFUNK_FEATURE_QUIC)
+// [`CursorState::flags`] bit: a host app captured/hid the pointer — the client SHOULD run
+// relative/captured (M3 auto-flip; advisory, user override always wins).
+#define CURSOR_RELATIVE_HINT 2
+#endif
+
 #if defined(PUNKTFUNK_FEATURE_QUIC)
 // QUIC application error code a punktfunk/1 client closes the control connection with on a
 // **deliberate quit** (a user "stop", not a network drop). The host reads it off the connection's