fix(host/linux): a no-channel session composites the metadata cursor — Mutter never embeds on a virtual stream

The capture-latched client (console.rs latched_mouse) never advertises
CLIENT_CAP_CURSOR, so its session resolved cursor_blend=false and asked
Mutter to EMBED the pointer. On a Mutter virtual stream that is a
fiction: since Mutter 48 (7ff5334a, hw-cursor inhibition removed) the
software cursor overlay is suppressed stage-globally whenever any
physical head realizes a HW cursor — dmabuf-recorded frames blit the
view without a pointer, and cursor-only motion schedules no re-record
either (mutter#4939). Probed on-glass on Mutter 50.3: embedded +
relative motion = frozen frame counter; SPA_META_Cursor positions kept
flowing in the same setup.

So the no-channel session now takes the path that was verified end to
end: cursor-as-metadata + the host composites, permanently — the same
arm a channel session lands in after its capture-model flip. Embedded
remains only the can't-blend fallback (libav VAAPI/NVENC, software).

- session_plan::cursor_blend_for grows the no-channel arm (codec +
  depth in, the same CUDA-payload prediction handshake makes);
  gamescope excluded so patch-2+ keeps its native-NV12 zero-copy shape
- the encode loop's composite refresh + one-shot breadcrumbs now cover
  the no-channel session; the park schedule keeps retrying while its
  composite is starved (relative-only clients cannot park themselves)
- the compositor retarget re-applies set_hw_cursor — the rebuilt
  display used to come up EMBEDDED even for channel sessions
- the GameStream virtual source takes the same rule (it never has a
  channel); its stream_body blend flag mirrors the request
- punktfunk-probe grows --cursor-nochannel (the latched-capture client,
  headless); cursor-probe grows --dump (PPM frames + a content-change
  counter, the pixel evidence the embedded A/B lacked)

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

clients/probe/src/main.rs

@@ -122,6 +122,13 @@ struct Args {
     /// pointer-lock client expecting the HOST to composite the cursor into the video. Decode the
     /// dump and look for the pointer — a cursorless dump is the bug this flag was built to catch.
     cursor_capture: bool,
+    /// `--cursor-nochannel` — the same relative wiggle WITHOUT the cursor channel: no
+    /// `CLIENT_CAP_CURSOR`, no render-mode flip. The headless reproduction of a client LATCHED
+    /// in capture mode at connect (`console.rs` `latched_mouse` — it never advertises the
+    /// channel), the shape of the 2026-07 "no cursor in Mutter capture mode" field report. The
+    /// host must composite the metadata cursor on its own; decode the dump and look for the
+    /// pointer.
+    cursor_nochannel: bool,
     /// `--discover [SECS]` — browse the LAN for native (`_punktfunk._udp`) hosts for `SECS`
     /// seconds (default 4), print what's found, and exit. No connection is made.
     discover: Option<u64>,
@@ -301,6 +308,7 @@ fn parse_args() -> Args {
             .then(|| get("--discover").and_then(|s| s.parse().ok()).unwrap_or(4)),
         clock_resync: argv.iter().any(|a| a == "--clock-resync"),
         cursor_capture: argv.iter().any(|a| a == "--cursor-capture"),
+        cursor_nochannel: argv.iter().any(|a| a == "--cursor-nochannel"),
     }
 }
 
@@ -845,32 +853,37 @@ async fn session(args: Args) -> Result<()> {
                 "SPEED TEST complete",
             );
         });
-    } else if args.cursor_capture {
+    } else if args.cursor_capture || args.cursor_nochannel {
-        // Capture-model cursor repro: flip the negotiated cursor channel to "host composites"
+        // Capture-model cursor repro. `--cursor-capture`: flip the negotiated cursor channel to
-        // and drive RELATIVE pointer motion, exactly like a pointer-lock client. Owns BOTH
+        // "host composites" and drive RELATIVE pointer motion, exactly like a pointer-lock
-        // control halves: the host may send CursorShape (0x50) messages while the channel is
+        // client. `--cursor-nochannel`: the same motion with NO channel at all (a
+        // capture-latched client) — the host must composite unprompted. Owns BOTH control
+        // halves: the host may send CursorShape (0x50) messages while a negotiated channel is
         // still in its initial client-draws state, and dropping our recv half would fail those
         // writes host-side.
+        let flip_channel = args.cursor_capture;
         let mut cs = send;
         let mut cr = recv;
         tokio::spawn(async move {
-            tokio::time::sleep(std::time::Duration::from_secs(1)).await;
+            if flip_channel {
-            match io::write_msg(
+                tokio::time::sleep(std::time::Duration::from_secs(1)).await;
-                &mut cs,
+                match io::write_msg(
-                &CursorRenderMode {
+                    &mut cs,
-                    client_draws: false,
+                    &CursorRenderMode {
-                }
+                        client_draws: false,
-                .encode(),
+                    }
-            )
+                    .encode(),
-            .await
+                )
-            {
+                .await
-                Ok(()) => tracing::info!(
+                {
-                    "cursor-capture: CursorRenderMode {{ client_draws: false }} sent — the host \
+                    Ok(()) => tracing::info!(
-                     must now composite the pointer into the video"
+                        "cursor-capture: CursorRenderMode {{ client_draws: false }} sent — the \
-                ),
+                         host must now composite the pointer into the video"
-                Err(e) => {
+                    ),
-                    tracing::error!(error = %format!("{e:#}"), "cursor-capture: render-mode write failed");
+                    Err(e) => {
-                    return;
+                        tracing::error!(error = %format!("{e:#}"), "cursor-capture: render-mode write failed");
+                        return;
+                    }
                 }
             }
             // Drain (and just log) whatever the host still sends on the control stream.

crates/pf-vdisplay/src/vdisplay/linux/mutter.rs

@@ -50,13 +50,19 @@ const APPLY_TEMPORARY: u32 = 1;
 /// the frames. The capturer always negotiates the meta (pf-capture `meta_param`) and the encoder
 /// blend composites it for sessions where the client does not draw the cursor itself — while a
 /// cursor-forwarding session strips the overlay and sends shape/state over the cursor channel.
-/// Embedded mode would leave BOTH paths blind: no metadata means nothing to forward AND nothing
+/// This is the mode for EVERY session whose host plans a composite or forward (`set_hw_cursor`
-/// to blend, and Mutter's own embedded painting is what the pre-channel path relied on.
+/// on), channel or not: Mutter's embedded painting is a fiction on a virtual stream — since
+/// Mutter 48 (commit `7ff5334a`, hw-cursor inhibition removed) the software cursor overlay is
+/// suppressed STAGE-GLOBALLY whenever any physical head realizes a hardware cursor, so
+/// dmabuf-recorded frames blit the view without a pointer, and cursor-only motion schedules no
+/// re-record either (mutter#4939). Probed on-glass (Mutter 50.3): embedded + relative motion =
+/// frozen frame counter; metadata positions kept flowing in the same setup.
 const CURSOR_METADATA: u32 = 2;
-/// `cursor-mode` embedded (mutter enum 1): Mutter composites the pointer into frames itself —
+/// `cursor-mode` embedded (mutter enum 1): Mutter composites the pointer into frames itself.
-/// zero host-side cursor work, the pre-channel path. Chosen for every session WITHOUT the
+/// Kept only as the can't-blend fallback (`set_hw_cursor` off — the resolved encode backend
-/// negotiated cursor channel (`set_hw_cursor` off — Phase B, the Windows no-regression gate
+/// cannot composite a metadata cursor, so metadata would strand the pointer in meta nothing
-/// mirrored); metadata stays the cursor-channel sessions' mode (shapes forwarded / host blend).
+/// draws). Know its limits (above): on a virtual stream it paints only into the MemFd/SHM
+/// record path (`FORCE_CURSORS`) and only refreshes on unrelated damage.
 const CURSOR_EMBEDDED: u32 = 1;
 
 /// Serializes, process-wide, every Mutter operation that adds/removes a virtual monitor or applies

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

@@ -389,12 +389,14 @@ fn run(
         return stream_body(
             &mut capturer,
             Some(&rebuild),
-            // The virtual-output source never selects cursor-as-metadata (`set_hw_cursor` is
+            // Mirrors the source's own `set_hw_cursor` request (`open_gs_virtual_source`):
-            // never called → the compositor EMBEDS the pointer where it can), so the encoder
+            // cursor-as-metadata + host blend wherever the backend composites — the
-            // is handed nothing to composite. gamescope remains the pointerless residual —
+            // compositor-EMBEDS fallback never paints on a Mutter virtual stream (the
-            // its capture carries no cursor either way (the native plane's XFixes source is
+            // native plane's no-channel rule, `session_plan::cursor_blend_for`). gamescope
-            // not wired on this plane).
+            // remains the pointerless residual — its capture carries no cursor either way
-            false,
+            // (the native plane's XFixes source is not wired on this plane).
+            compositor != crate::vdisplay::Compositor::Gamescope
+                && blend_capable_metadata_cursor(&cfg),
             &sock,
             cfg,
             running,
@@ -415,18 +417,7 @@ fn run(
     // `frame.cursor` (the caps-aware negotiation — mirror of the native plane's); otherwise ask
     // the portal to EMBED the pointer so no backend × cursor-mode combination streams
     // cursorless. Synthetic frames carry no pointer either way.
-    let metadata_cursor = {
+    let metadata_cursor = blend_capable_metadata_cursor(&cfg);
-        #[cfg(target_os = "linux")]
-        {
-            // Same CUDA-payload prediction SessionPlan/`handshake::cursor_forward` make:
-            // the NVIDIA resolution plus the zero-copy master switch.
-            let cuda_planned =
-                !crate::encode::linux_zero_copy_is_vaapi() && crate::zerocopy::enabled();
-            crate::encode::cursor_blend_capable(cfg.codec, cuda_planned, cfg.hdr)
-        }
-        #[cfg(not(target_os = "linux"))]
-        false
-    };
     // Which screen this stream must show. The host-wide pin (§5.3) applies to the compat plane too:
     // the portal chooser cannot name a head, so a pinned host MIRRORS it here the same way the
     // virtual source does via `vdisplay::open`. Without this a Moonlight client on a pinned host
@@ -643,6 +634,24 @@ fn resolve_gs_app(app: Option<&super::apps::AppEntry>) -> Option<GsApp> {
 /// entry can PIN a compositor (skips the live detect/retarget). Re-run on a mid-stream capture loss to
 /// FOLLOW a Desktop<->Game switch: it re-detects the now-live compositor and re-targets at it. Does NOT
 /// launch the app (that happens once at stream start; a rebuild must not re-spawn it).
+/// Cursor-as-metadata for this plane (GameStream has no cursor channel): only where the encode
+/// backend this session resolves to composites `frame.cursor` — the same CUDA-payload
+/// prediction `SessionPlan`/`handshake::cursor_forward` make (the NVIDIA resolution plus the
+/// zero-copy master switch). Shared by the monitor-mirror and virtual-output sources so their
+/// `set_hw_cursor` request and `stream_body`'s blend flag cannot drift.
+fn blend_capable_metadata_cursor(cfg: &StreamConfig) -> bool {
+    #[cfg(target_os = "linux")]
+    {
+        let cuda_planned = !crate::encode::linux_zero_copy_is_vaapi() && crate::zerocopy::enabled();
+        crate::encode::cursor_blend_capable(cfg.codec, cuda_planned, cfg.hdr)
+    }
+    #[cfg(not(target_os = "linux"))]
+    {
+        let _ = cfg;
+        false
+    }
+}
+
 fn open_gs_virtual_source(
     cfg: StreamConfig,
     app: Option<&super::apps::AppEntry>,
@@ -701,6 +710,16 @@ fn open_gs_virtual_source(
         }
     };
     let mut vd = crate::vdisplay::open(compositor).context("open virtual display")?;
+    // Out-of-band cursor for the virtual source (the native plane's no-channel rule, mirrored):
+    // GameStream has no cursor channel, and the compositor-EMBEDS fallback never paints on a
+    // Mutter virtual stream (stage-global overlay suppression since Mutter 48 — see
+    // pf-vdisplay `mutter.rs`), so ask for cursor-as-metadata wherever the resolved backend
+    // composites `frame.cursor`; `stream_body`'s blend flag mirrors this request. gamescope
+    // stays off: its capture carries no metadata either way, and the request would cost the
+    // native-NV12 shape for nothing.
+    vd.set_hw_cursor(
+        compositor != crate::vdisplay::Compositor::Gamescope && blend_capable_metadata_cursor(&cfg),
+    );
     // Carry the resolved launch command on the backend instance (per-session) rather than a
     // process-global env var, so concurrent sessions can't stomp each other's launch target. It is
     // the RESOLVED command, so gamescope's bare spawn nests a library title exactly like an

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

@@ -17,10 +17,13 @@ use super::*;
 /// (`encode::cursor_blend_capable`): the channel's capture-mouse flip (`CursorRenderMode`,
 /// `client_draws = false`) makes the HOST draw the pointer, and on Linux the encoder is that
 /// compositing stage — granting the channel over a backend that can't blend (libav
-/// VAAPI/NVENC, software) shipped a cursorless stream on every capture-mode flip. Denied, the
+/// VAAPI/NVENC, software) shipped a cursorless stream on every capture-mode flip. Denied — or
-/// session keeps the pre-channel path: the compositor EMBEDS the pointer and the client never
+/// never asked for (a capture-latched client, `console.rs` `latched_mouse`) — the session
-/// draws — never cursorless, never doubled. THE single predicate: the Welcome's
+/// composites host-side anyway wherever the backend can blend
-/// `HOST_CAP_CURSOR` bit is computed from it, and the session wiring reads that bit back.
+/// (`session_plan::cursor_blend_for`'s no-channel arm; the compositor-EMBEDS fallback never
+/// paints on a Mutter virtual stream), and only a can't-blend backend falls back to the
+/// compositor EMBED. THE single predicate: the Welcome's `HOST_CAP_CURSOR` bit is computed
+/// from it, and the session wiring reads that bit back.
 pub(super) fn cursor_forward(
     client_caps: u8,
     compositor: Option<crate::vdisplay::Compositor>,

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

@@ -1076,20 +1076,20 @@ pub(super) fn virtual_stream(ctx: SessionContext, prepared: Option<PreparedDispl
         ctx.bit_depth,
         ctx.chroma,
         ctx.codec,
-        // Blend CAPABILITY for cursor-FORWARD sessions (Phase B, the Windows gate mirrored):
+        // Blend CAPABILITY (the single rule in `cursor_blend_for`): cursor-FORWARD sessions
-        // their client can flip to the capture mouse model mid-stream (`CursorRenderMode`), and
+        // need it for the mid-stream capture-mouse flip (`CursorRenderMode` — WHETHER a
-        // the composite side of that flip must not need an encoder rebuild — WHETHER a frame's
+        // frame's pointer is drawn stays per-tick, the encode loop strips `frame.cursor`
-        // pointer is drawn stays per-tick (the encode loop strips `frame.cursor` while the client
+        // while the client draws locally); gamescope (Phase C) can't embed a pointer, so the
-        // draws locally, see the forwarder tick). Non-channel NON-gamescope sessions get the
+        // host always composites the XFixes-sourced cursor; and a NO-channel session gets
-        // pointer compositor-EMBEDDED (`vd.set_hw_cursor(false)` → no cursor metadata, nothing to
+        // metadata + host blend too wherever the backend composites — the compositor-EMBEDS
-        // blend), keeping the zero-cost pre-channel path. gamescope is the exception (Phase C):
+        // fallback streams cursorless on a Mutter virtual output (the overlay-visibility
-        // it can't embed the pointer, so the host ALWAYS composites the XFixes-sourced cursor —
+        // gate is stage-global since Mutter 48; see `cursor_blend_for`'s doc). Embedded
-        // the blend must be built for every gamescope session. (`cursor_forward` is already
+        // remains only the can't-blend fallback.
-        // blend-gated: `handshake::cursor_forward` grants the channel only where
-        // `encode::cursor_blend_capable` says the resolved backend composites.)
         crate::session_plan::cursor_blend_for(
             ctx.cursor_forward,
             ctx.compositor == pf_vdisplay::Compositor::Gamescope,
+            ctx.codec,
+            ctx.bit_depth,
         ),
         ctx.cursor_forward,
     );
@@ -1187,6 +1187,24 @@ pub(super) fn virtual_stream(ctx: SessionContext, prepared: Option<PreparedDispl
     if gamescope_composite {
         tracing::info!("gamescope cursor: compositing the XFixes-sourced pointer into the video");
     }
+    // No-channel metadata composite: the client never draws the pointer (it did not advertise
+    // the cursor channel — e.g. a capture-latched client, `console.rs` `latched_mouse`), and
+    // the compositor-EMBEDS fallback is a fiction on a Mutter virtual stream (the software
+    // cursor overlay is suppressed stage-globally whenever any physical head realizes a HW
+    // cursor, Mutter 48+ — dmabuf frames blit the view WITHOUT it, and cursor-only motion
+    // schedules no update either, mutter#4939). So the plan asked the backend for
+    // cursor-as-metadata and the HOST composites, permanently — the same arm a channel
+    // session lands in after its capture-model flip, minus the channel.
+    // `mut`: recomputed with `gamescope_composite` on a mid-stream compositor retarget.
+    let mut metadata_composite = cursor_fwd.is_none()
+        && plan.cursor_blend
+        && compositor != pf_vdisplay::Compositor::Gamescope;
+    if metadata_composite {
+        tracing::info!(
+            "no cursor channel — compositing the metadata cursor into the video (embedded \
+             fallback is unreliable on virtual streams)"
+        );
+    }
     if streamed_wire {
         // Client capability only — whether AUs actually stream per-slice depends on the encoder
         // backend's `supports_chunked_poll()` (today: Linux direct-NVENC only), which doesn't
@@ -1231,9 +1249,12 @@ pub(super) fn virtual_stream(ctx: SessionContext, prepared: Option<PreparedDispl
             // it with the HDR flags so nested games get HDR surfaces at all. Decided in the
             // Welcome (`capture::capturer_supports_hdr_for`), so it cannot change under us.
             vd.set_hdr(bit_depth >= 10);
-            // Cursor-forward sessions ask the backend for an out-of-band hardware cursor
+            // Out-of-band cursor request: cursor-forward sessions (Windows pf-vdisplay /
-            // (Windows pf-vdisplay / IddCx; no-op on Linux — the portal already separates it).
+            // IddCx hardware cursor; Linux metadata mode) AND no-channel host-composite
-            vd.set_hw_cursor(cursor_forward);
+            // sessions (Linux only — `metadata_composite` is `plan.cursor_blend`-gated, so
+            // it is always false on Windows). The backend keeps the pointer out of the
+            // pixels; the host blend (or the client) puts it back.
+            vd.set_hw_cursor(cursor_forward || metadata_composite);
             // Deliberate-quit wiring (Windows pf-vdisplay; no-op elsewhere): every lease the
             // backend mints — the retry-hold below AND the capturer's — carries the session's quit
             // flag, so a user "stop" (⌘D → the QUIT close code) tears the virtual monitor down the
@@ -2378,6 +2399,8 @@ pub(super) fn virtual_stream(ctx: SessionContext, prepared: Option<PreparedDispl
                                         plan.cursor_blend = crate::session_plan::cursor_blend_for(
                                             plan.cursor_forward,
                                             c == crate::vdisplay::Compositor::Gamescope,
+                                            plan.codec,
+                                            plan.bit_depth,
                                         );
                                         plan.gamescope_cursor =
                                             crate::session_plan::gamescope_cursor_for(
@@ -2385,6 +2408,16 @@ pub(super) fn virtual_stream(ctx: SessionContext, prepared: Option<PreparedDispl
                                             );
                                         gamescope_composite =
                                             plan.gamescope_cursor && cursor_fwd.is_none();
+                                        metadata_composite = cursor_fwd.is_none()
+                                            && plan.cursor_blend
+                                            && c != crate::vdisplay::Compositor::Gamescope;
+                                        // The retargeted backend starts with `hw_cursor`
+                                        // unset — without re-applying the session's
+                                        // out-of-band cursor request, the rebuilt display
+                                        // would come up EMBEDDED: double-drawn for a
+                                        // desktop-model channel client, cursorless for
+                                        // every host-composite session.
+                                        vd.set_hw_cursor(plan.cursor_forward || metadata_composite);
                                     }
                                     Err(e2) => tracing::warn!(error = %format!("{e2:#}"),
                                         "capture loss: opening the newly-detected compositor failed — retrying"),
@@ -2560,15 +2593,41 @@ pub(super) fn virtual_stream(ctx: SessionContext, prepared: Option<PreparedDispl
                     }
                 }
             }
-        } else if gamescope_composite {
+        } else if gamescope_composite || metadata_composite {
-            // gamescope (Phase C): no channel, host always composites. Refresh the (repeat or new)
+            // No channel, host always composites: gamescope (Phase C — the XFixes source
-            // frame's overlay from the capturer's LIVE cursor — the XFixes source publishes there
+            // publishes on `capturer.cursor()`) and the metadata-composite session (the portal
-            // — so pointer-only motion on a static gamescope UI re-blends at tick rate instead of
+            // `SPA_META_Cursor` live overlay publishes there too). Refresh the (repeat or new)
-            // freezing at the last damage frame (the same reason the composite arm above re-reads
+            // frame's overlay from the capturer's LIVE cursor so pointer-only motion on a
-            // it). A grabbed/hidden pointer arrives `visible: false` and is stripped just below.
+            // static desktop re-blends at tick rate instead of freezing at the last damage
+            // frame (the same reason the channel's composite arm above re-reads it). A
+            // grabbed/hidden pointer arrives `visible: false` and is stripped just below.
             #[cfg(not(target_os = "windows"))]
-            if let Some(live) = capturer.cursor() {
+            match capturer.cursor() {
-                frame.cursor = Some(live);
+                Some(live) => {
+                    if !composite_saw_overlay {
+                        composite_saw_overlay = true;
+                        tracing::info!(
+                            x = live.x,
+                            y = live.y,
+                            w = live.w,
+                            h = live.h,
+                            visible = live.visible,
+                            "host-composite: first live cursor overlay handed to the encoder \
+                             blend"
+                        );
+                    }
+                    frame.cursor = Some(live);
+                }
+                None => {
+                    if !composite_saw_none {
+                        composite_saw_none = true;
+                        tracing::info!(
+                            "host-composite active but the capture has no live cursor overlay \
+                             yet (no SPA_META_Cursor bitmap) — the stream is cursorless until \
+                             one arrives"
+                        );
+                    }
+                }
             }
         }
         // The overlay surfaces hidden pointers too (for the hint above) — strip them
@@ -2579,10 +2638,12 @@ pub(super) fn virtual_stream(ctx: SessionContext, prepared: Option<PreparedDispl
         // The seat-pointer park schedule (state + rationale at the declarations above; armed by
         // the first frame of every (re)built display and by the capture-model flip). The first
         // two attempts run unconditionally — attempt 1 can be swallowed by a cold EIS
-        // connection. Past those, only a cursor-channel session in the capture model that STILL
+        // connection. Past those, only a host-composite session that STILL has no live overlay
-        // has no live overlay keeps trying: that combination means the pointer has not reached
+        // keeps trying — a channel session in the capture model, or a no-channel
-        // the streamed output (the compositor reports cursor metadata only while it is over the
+        // metadata-composite session (both relative-only): no overlay there means the pointer
-        // recorded view), and a relative-only client cannot get it there on its own.
+        // has not reached the streamed output (the compositor reports cursor metadata only
+        // while it is over the recorded view), and a relative-only client cannot get it there
+        // on its own.
         // Armed from the loop's first tick — a static desktop may never deliver a fresh frame
         // (`parked_display` is only bookkeeping for rebuild re-arming), and the pointer must be
         // parked regardless.
@@ -2591,8 +2652,9 @@ pub(super) fn virtual_stream(ctx: SessionContext, prepared: Option<PreparedDispl
             && park_attempts < PARK_ATTEMPTS_MAX
             && std::time::Instant::now() >= next_park_at
         {
-            let composite_starved = cursor_fwd.is_some()
+            let composite_starved = ((cursor_fwd.is_some()
-                && !cursor_client_draws.load(Ordering::Relaxed)
+                && !cursor_client_draws.load(Ordering::Relaxed))
+                || metadata_composite)
                 && capturer.cursor().is_none();
             if park_attempts < 2 || composite_starved {
                 park_pointer(&input_tx, frame.width, frame.height);
@@ -3406,13 +3468,14 @@ pub(super) fn prepare_display(
         bit_depth,
         chroma,
         codec,
-        // Blend capability — must MATCH virtual_stream's resolve (Phase B: non-channel
+        // Blend capability — must MATCH virtual_stream's resolve. Windows-only path, where
-        // non-gamescope sessions get the pointer compositor-EMBEDDED, nothing to blend; the
+        // the rule is a constant `false` (the IDD capturer composites itself); passed through
-        // mid-stream `CursorRenderMode` flip strips/keeps `frame.cursor` per tick for channel
+        // the shared rule anyway so the two resolves cannot drift.
-        // sessions). gamescope (Phase C) can't embed → always composites the XFixes cursor.
         crate::session_plan::cursor_blend_for(
             cursor_forward,
             compositor == pf_vdisplay::Compositor::Gamescope,
+            codec,
+            bit_depth,
         ),
         cursor_forward,
     );

crates/punktfunk-host/src/session_plan.rs

@@ -108,7 +108,9 @@ pub struct SessionPlan {
     pub wire_chunk: Option<usize>,
     /// The session may hand the encoder cursor bitmaps to composite (cursor-as-metadata
     /// captures). Set via [`cursor_blend_for`] — the single platform rule — so it is `true` only
-    /// where the ENCODER is the compositing stage (Linux cursor-forward and gamescope sessions);
+    /// where the ENCODER is the compositing stage (Linux: cursor-forward sessions, gamescope,
+    /// AND no-channel sessions on a blend-capable backend — the compositor-EMBEDS fallback is
+    /// broken on Mutter virtual streams, see [`cursor_blend_for`]);
     /// Windows is always `false` (the IDD capturer composites the pointer itself). Encoders
     /// whose fast path cannot blend (the Vulkan EFC RGB-direct source, native NV12) stay off
     /// those shapes when this is set — see [`Self::output_format`] and
@@ -234,20 +236,46 @@ pub(crate) fn resolve_topology() -> SessionTopology {
 /// THE rule for [`SessionPlan::cursor_blend`], shared by every resolve caller (initial plan and
 /// the mid-stream compositor re-gate) so they can't drift:
 /// * **Linux**: the encoder is the compositing stage — blend for a cursor-forward session (the
-///   capture-mouse flip needs the host composite on demand) and for gamescope (its capture
+///   capture-mouse flip needs the host composite on demand), for gamescope (its capture
-///   carries no pointer at all; the XFixes-sourced cursor must be drawn into the video).
+///   carries no pointer at all; the XFixes-sourced cursor must be drawn into the video), AND
+///   for a no-channel session whenever the resolved backend can composite. The pre-channel
+///   "compositor EMBEDS the pointer" fallback is a fiction on a Mutter virtual stream:
+///   cursor-only motion never re-records the stream (probed on-glass, Mutter 50.3 — frames
+///   froze the instant motion went relative while `SPA_META_Cursor` kept updating), so a
+///   capture-latched client (which never advertises `CLIENT_CAP_CURSOR`, `console.rs`
+///   `latched_mouse`) streamed cursorless. Metadata + host blend is the path that was
+///   verified end-to-end; embedded remains only the can't-blend fallback (libav
+///   VAAPI/NVENC, software).
 /// * **Windows**: never — the IDD capturer composites the pointer itself (`cursor_blend.rs` /
 ///   DWM), and no Windows encode backend reads `frame.cursor`. Asking the encoder anyway made
 ///   `open_video`'s blends-cursor backstop fire spuriously on every cursor-channel session.
-pub(crate) fn cursor_blend_for(cursor_forward: bool, gamescope: bool) -> bool {
+pub(crate) fn cursor_blend_for(
+    cursor_forward: bool,
+    gamescope: bool,
+    codec: crate::encode::Codec,
+    bit_depth: u8,
+) -> bool {
     #[cfg(target_os = "windows")]
     {
-        let _ = (cursor_forward, gamescope);
+        let _ = (cursor_forward, gamescope, codec, bit_depth);
         false
     }
     #[cfg(not(target_os = "windows"))]
     {
-        cursor_forward || gamescope_needs_host_cursor(gamescope)
+        if gamescope {
+            // gamescope's capture carries no SPA_META_Cursor; the blend-capable term below
+            // must not apply, or a patch-2+ gamescope (composites its own pointer) would lose
+            // its native-NV12 zero-copy shape for a blend that can never receive an overlay.
+            return gamescope_needs_host_cursor(true);
+        }
+        if cursor_forward {
+            return true;
+        }
+        // No cursor channel: the same CUDA-payload prediction `handshake::cursor_forward` and
+        // the GameStream monitor mirror make — the NVIDIA resolution plus the zero-copy master
+        // switch — deciding direct-SDK NVENC (blends) vs libav NVENC (doesn't).
+        let cuda_planned = !crate::encode::linux_zero_copy_is_vaapi() && crate::zerocopy::enabled();
+        crate::encode::cursor_blend_capable(codec, cuda_planned, bit_depth == 10)
     }
 }
 

tools/cursor-probe/Cargo.toml

@@ -16,6 +16,7 @@ repository.workspace = true
 [target.'cfg(target_os = "linux")'.dependencies]
 pf-vdisplay = { path = "../../crates/pf-vdisplay" }
 pf-capture = { path = "../../crates/pf-capture" }
+pf-frame = { path = "../../crates/pf-frame" }
 pf-inject = { path = "../../crates/pf-inject" }
 punktfunk-core = { path = "../../crates/punktfunk-core", features = ["quic"] }
 anyhow = "1"

tools/cursor-probe/src/main.rs

@@ -153,6 +153,19 @@ mod linux {
             }
         });
 
+        let dump = args.iter().any(|a| a == "--dump");
+        let dump_dir = format!(
+            "/tmp/pf-probe-frames-{}",
+            if embedded { "embedded" } else { "metadata" }
+        );
+        if dump {
+            let _ = std::fs::remove_dir_all(&dump_dir);
+            let _ = std::fs::create_dir_all(&dump_dir);
+        }
+        let mut prev: Vec<u8> = Vec::new();
+        let mut changed = 0u64;
+        let mut dumped = 0u32;
+
         let deadline = Instant::now() + Duration::from_secs(secs);
         let (mut frames, mut with_overlay) = (0u64, 0u64);
         let mut last_report = Instant::now();
@@ -161,6 +174,24 @@ mod linux {
             // Damage-driven source: timeouts (gaps) are normal, hence the missing error arm.
             if let Ok(f) = cap.next_frame_within(Duration::from_millis(500)) {
                 frames += 1;
+                if let pf_frame::FramePayload::Cpu(data) = &f.payload {
+                    if !prev.is_empty()
+                        && prev.len() == data.len()
+                        && prev.as_slice() != data.as_slice()
+                    {
+                        changed += 1;
+                    }
+                    prev.clear();
+                    prev.extend_from_slice(data);
+                    if dump && frames % 15 == 1 && dumped < 16 {
+                        dumped += 1;
+                        let path = format!("{dump_dir}/frame-{frames:05}.ppm");
+                        match dump_ppm(&path, f.width, f.height, f.format, data) {
+                            Ok(()) => println!("cursor-probe: dumped {path}"),
+                            Err(e) => eprintln!("cursor-probe: dump failed: {e:#}"),
+                        }
+                    }
+                }
                 if let Some(c) = &f.cursor {
                     with_overlay += 1;
                     if with_overlay == 1 {
@@ -195,9 +226,13 @@ mod linux {
         let _ = injector.join();
 
         println!("---");
+        println!(
+            "cursor-probe: content-changed frames: {changed}/{frames} (a moving embedded \
+             cursor over a static desktop should change nearly every frame)"
+        );
         if embedded {
             println!(
-                "cursor-probe: EMBEDDED run — no metadata expected; check the stream visually. \
+                "cursor-probe: EMBEDDED run — no metadata expected; check the dumped frames. \
                  frames={frames}"
             );
         } else if live_seen || with_overlay > 0 {
@@ -214,4 +249,26 @@ mod linux {
         }
         Ok(())
     }
+
+    /// Write a packed 4-bpp frame as a PPM, swapping B/R for the BGR layouts. Enough fidelity
+    /// to answer "is the pointer in the pixels".
+    fn dump_ppm(
+        path: &str,
+        w: u32,
+        h: u32,
+        format: pf_frame::PixelFormat,
+        data: &[u8],
+    ) -> Result<()> {
+        use std::io::Write;
+        let mut out = std::io::BufWriter::new(std::fs::File::create(path)?);
+        write!(out, "P6\n{w} {h}\n255\n")?;
+        let (ri, gi, bi) = match format {
+            pf_frame::PixelFormat::Rgbx | pf_frame::PixelFormat::Rgba => (0usize, 1usize, 2usize),
+            _ => (2usize, 1usize, 0usize),
+        };
+        for px in data.chunks_exact(4).take((w as usize) * (h as usize)) {
+            out.write_all(&[px[ri], px[gi], px[bi]])?;
+        }
+        Ok(())
+    }
 }