feat(audio): end-to-end 5.1/7.1 surround across the native path + all clients

Adds negotiated 5.1/7.1 surround to the punktfunk/1 protocol and every client
(previously stereo-only):

- core: new shared `audio` layout table (LAYOUT_51/71 + identity multistream
  mapping, canonical wire order FL FR FC LFE RL RR SL SR); Hello/Welcome
  `audio_channels` negotiation via the trailing-byte back-compat pattern (old
  peers fall back to stereo); C-ABI `punktfunk_connect_ex6`,
  `punktfunk_connection_audio_channels`, and in-core multistream decode
  `punktfunk_connection_next_audio_pcm` for embedders without a multistream
  Opus decoder. Real-libopus channel-identity round-trip test.
- host: native audio thread captures + Opus-(multi)stream-encodes at the
  negotiated count (with a cross-session cached-capturer channel-mismatch fix);
  GameStream surround unified onto the safe `opus::MSEncoder`, dropping
  `audiopus_sys` (~4 unsafe blocks) and un-gating Windows GameStream surround;
  WASAPI loopback capture relaxed to 2/6/8 with the correct dwChannelMask.
- clients: Linux (PipeWire), Windows (WASAPI), Android (AAudio) decode via
  `opus::MSDecoder` + render multichannel; Apple decodes in-core to PCM →
  AVAudioEngine with an explicit wire-order channel layout; each gains a
  Stereo/5.1/7.1 setting. `punktfunk-probe --audio-channels N` is the headless
  validator.

Verified on Linux: core/host/linux/probe test suites + the Android Rust
(cargo-ndk) build, clippy -D warnings, and rustfmt all green. Windows/Apple
builds, all on-glass checks, and the live native loopback are pending (CI / a
free box).

Also lands the concurrent in-tree HEVC 4:4:4 host work (PUNKTFUNK_444): it
shares the same touched files (quic.rs, punktfunk1.rs, encode/*, ...) and so
cannot be committed separately from the surround changes.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

crates/punktfunk-core/Cargo.toml

@@ -19,7 +19,7 @@ crate-type = ["lib", "cdylib", "staticlib"]
 default = []
 # Control-plane QUIC (pairing, config, reverse audio). tokio is permitted ONLY here,
 # never on the per-frame hot path. Off by default so the core stays runtime-free.
-quic = ["dep:quinn", "dep:tokio", "dep:rustls", "dep:rcgen", "dep:rustls-pki-types", "dep:sha2", "dep:hmac", "dep:spake2"]
+quic = ["dep:quinn", "dep:tokio", "dep:rustls", "dep:rcgen", "dep:rustls-pki-types", "dep:sha2", "dep:hmac", "dep:spake2", "dep:opus"]
 
 [dependencies]
 reed-solomon-simd = "3.1"      # GF(2^16) Leopard-RS, SIMD, O(n log n) — the wall-breaker (P2)
@@ -51,6 +51,12 @@ sha2 = { version = "0.10", optional = true }
 hmac = { version = "0.12", optional = true }
 spake2 = { version = "0.4", optional = true }
 tokio = { version = "1", optional = true, features = ["rt-multi-thread", "net", "sync", "macros"] }
+# In-core Opus (multistream) DECODE for the C-ABI `punktfunk_connection_next_audio_pcm` path —
+# used by embedders without a multistream-capable Opus decoder (Apple's AudioToolbox is
+# stereo-only). The Rust clients link `opus` themselves and decode the raw `next_audio` frames,
+# so this only matters when the connection API (quic) is built. Same libopus the host vendors;
+# cargo unifies the build. Multistream API: `opus::MSDecoder` (lib.rs:1187).
+opus = { version = "0.3", optional = true }
 
 # `libc` for batched UDP syscalls: `sendmmsg`/`recvmmsg` on Linux (the 1 Gbps+ lever) and the
 # `recv(MSG_DONTWAIT)` drain on the other unix (Apple/BSD) targets, which have no `recvmmsg`

crates/punktfunk-core/src/abi.rs

@@ -467,6 +467,23 @@ pub struct PunktfunkConnection {
     last: std::sync::Mutex<Option<crate::session::Frame>>,
     /// Same, for `punktfunk_connection_next_audio` (independent of the video slot).
     last_audio: std::sync::Mutex<Option<crate::client::AudioPacket>>,
+    /// Decode-in-core state for `punktfunk_connection_next_audio_pcm` (Apple / any embedder
+    /// without a multistream Opus decoder). The decoder is built lazily from the negotiated
+    /// `inner.audio_channels`; `pcm` is a fixed-capacity reusable buffer the returned pointer
+    /// borrows until the next PCM call (same contract as `last_audio`).
+    audio_pcm: std::sync::Mutex<AudioPcmState>,
+}
+
+/// Lazily-initialized in-core Opus decode state. A coupled-1-stream multistream decoder is
+/// equivalent to a plain stereo decoder, so one [`opus::MSDecoder`] handles 2/6/8 channels.
+#[cfg(feature = "quic")]
+#[derive(Default)]
+struct AudioPcmState {
+    decoder: Option<opus::MSDecoder>,
+    /// Interleaved f32 PCM, wire channel order. Pre-sized to the largest legal Opus frame
+    /// (120 ms @ 48 kHz = 5760 samples/ch) × 8 channels so decode never reallocates (which would
+    /// dangle the pointer handed to the embedder).
+    pcm: Vec<f32>,
 }
 
 /// `PunktfunkHidOutput::kind` — lightbar RGB (`r`/`g`/`b` valid).
@@ -708,12 +725,18 @@ pub const PUNKTFUNK_VIDEO_CAP_10BIT: u8 = 0x01;
 /// Video-capability bit for [`punktfunk_connect_ex5`] (`video_caps`): the client can present
 /// BT.2020 PQ HDR10 (implies 10-bit). (Mirrors `quic::VIDEO_CAP_HDR`.)
 pub const PUNKTFUNK_VIDEO_CAP_HDR: u8 = 0x02;
+/// Video-capability bit for [`punktfunk_connect_ex5`] (`video_caps`): the client can decode a
+/// full-chroma 4:4:4 HEVC stream (Range Extensions). The host emits 4:4:4 only when this is set,
+/// the host opted in, the codec is HEVC, and the GPU supports it — else the stream stays 4:2:0 and
+/// [`punktfunk_connection_chroma_format`] reports the real value. (Mirrors `quic::VIDEO_CAP_444`.)
+pub const PUNKTFUNK_VIDEO_CAP_444: u8 = 0x04;
 
 // Keep the ABI cap bits in lockstep with the wire constants (compile-time guard against drift).
 #[cfg(feature = "quic")]
 const _: () = {
     assert!(PUNKTFUNK_VIDEO_CAP_10BIT == crate::quic::VIDEO_CAP_10BIT);
     assert!(PUNKTFUNK_VIDEO_CAP_HDR == crate::quic::VIDEO_CAP_HDR);
+    assert!(PUNKTFUNK_VIDEO_CAP_444 == crate::quic::VIDEO_CAP_444);
 };
 
 // Keep the ABI gamepad constants in lockstep with the wire enum (compile-time guard against drift).
@@ -980,6 +1003,58 @@ pub unsafe extern "C" fn punktfunk_connect_ex5(
     client_cert_pem: *const std::os::raw::c_char,
     client_key_pem: *const std::os::raw::c_char,
     timeout_ms: u32,
+) -> *mut PunktfunkConnection {
+    // Delegate to the surround-aware variant requesting stereo (the pre-surround behaviour).
+    unsafe {
+        punktfunk_connect_ex6(
+            host,
+            port,
+            width,
+            height,
+            refresh_hz,
+            compositor,
+            gamepad,
+            bitrate_kbps,
+            video_caps,
+            2, // audio_channels = stereo
+            launch_id,
+            pin_sha256,
+            observed_sha256_out,
+            client_cert_pem,
+            client_key_pem,
+            timeout_ms,
+        )
+    }
+}
+
+/// Like [`punktfunk_connect_ex5`], but additionally requests the audio channel count:
+/// `2` (stereo, the default behaviour of every earlier variant), `6` (5.1) or `8` (7.1). The host
+/// clamps the request to what it can actually capture and echoes the resolved count via
+/// [`punktfunk_connection_audio_channels`]; the `0xC9` audio frames are Opus-(multi)stream encoded
+/// for that layout. A client that wants surround calls this; everything else inherits stereo.
+///
+/// # Safety
+/// Same as [`punktfunk_connect`].
+#[cfg(feature = "quic")]
+#[no_mangle]
+#[allow(clippy::too_many_arguments)]
+pub unsafe extern "C" fn punktfunk_connect_ex6(
+    host: *const std::os::raw::c_char,
+    port: u16,
+    width: u32,
+    height: u32,
+    refresh_hz: u32,
+    compositor: u32,
+    gamepad: u32,
+    bitrate_kbps: u32,
+    video_caps: u8,
+    audio_channels: u8,
+    launch_id: *const std::os::raw::c_char,
+    pin_sha256: *const u8,
+    observed_sha256_out: *mut u8,
+    client_cert_pem: *const std::os::raw::c_char,
+    client_key_pem: *const std::os::raw::c_char,
+    timeout_ms: u32,
 ) -> *mut PunktfunkConnection {
     let r = std::panic::catch_unwind(AssertUnwindSafe(|| {
         if host.is_null() {
@@ -1029,6 +1104,7 @@ pub unsafe extern "C" fn punktfunk_connect_ex5(
             gamepad,
             bitrate_kbps,
             video_caps,
+            crate::audio::normalize_channels(audio_channels),
             launch,
             pin,
             identity,
@@ -1045,6 +1121,7 @@ pub unsafe extern "C" fn punktfunk_connect_ex5(
                     inner: c,
                     last: std::sync::Mutex::new(None),
                     last_audio: std::sync::Mutex::new(None),
+                    audio_pcm: std::sync::Mutex::new(AudioPcmState::default()),
                 }))
             }
             Err(_) => std::ptr::null_mut(),
@@ -1250,6 +1327,121 @@ pub unsafe extern "C" fn punktfunk_connection_next_audio(
     })
 }
 
+/// Read the audio channel count the host resolved for this session (from its Welcome): `2`
+/// (stereo), `6` (5.1) or `8` (7.1). `*out` is filled when non-NULL. The `0xC9` Opus frames are
+/// (multistream-)encoded for this layout; an embedder decoding raw frames itself must build its
+/// decoder from THIS value (see [`crate::audio::layout_for`]) — or use
+/// [`punktfunk_connection_next_audio_pcm`], which decodes in-core. Available immediately after a
+/// successful connect (it doesn't change without a reconfigure).
+///
+/// # Safety
+/// `c` is a valid connection handle; `out` is NULL or writable for one `u8`.
+#[cfg(feature = "quic")]
+#[no_mangle]
+pub unsafe extern "C" fn punktfunk_connection_audio_channels(
+    c: *mut PunktfunkConnection,
+    out: *mut u8,
+) -> PunktfunkStatus {
+    guard(|| {
+        let c = match unsafe { c.as_ref() } {
+            Some(c) => c,
+            None => return PunktfunkStatus::NullPointer,
+        };
+        if !out.is_null() {
+            // SAFETY: `out` is non-null and the caller guarantees it is writable for one `u8`.
+            unsafe { *out = c.inner.audio_channels };
+        }
+        PunktfunkStatus::Ok
+    })
+}
+
+/// One decoded audio frame from [`punktfunk_connection_next_audio_pcm`]: interleaved 32-bit
+/// float PCM at 48 kHz, in the canonical wire channel order `FL FR FC LFE RL RR SL SR` (the
+/// first `channels` of it). `samples` points at `frame_count * channels` floats and borrows
+/// connection memory **until the next PCM call** on this handle.
+#[cfg(feature = "quic")]
+#[repr(C)]
+pub struct PunktfunkAudioPcm {
+    /// Interleaved f32 samples (wire channel order), `frame_count * channels` long.
+    pub samples: *const f32,
+    /// Samples per channel in this frame.
+    pub frame_count: u32,
+    /// Channel count (2/6/8) — the negotiated [`punktfunk_connection_audio_channels`].
+    pub channels: u8,
+    /// Source packet sequence number.
+    pub seq: u32,
+    /// Capture presentation timestamp (ns).
+    pub pts_ns: u64,
+}
+
+/// Pull the next audio frame and **decode it in-core** to interleaved f32 PCM — for embedders
+/// without a multistream-capable Opus decoder (e.g. Apple, whose AudioToolbox Opus path is
+/// stereo-only). The decoder is built once from the negotiated channel count and handles 2/6/8
+/// channels (a 1-coupled-stream multistream decoder is exactly a stereo decoder). Same
+/// timeout/closed semantics as [`punktfunk_connection_next_audio`]; `out->samples` borrows
+/// connection memory until the next PCM call on this handle. Use EITHER this or
+/// [`punktfunk_connection_next_audio`] on a given connection, from one dedicated audio thread —
+/// not both (they share the underlying queue).
+///
+/// # Safety
+/// `c` is a valid connection handle; `out` is writable. At most one thread pulls audio.
+#[cfg(feature = "quic")]
+#[no_mangle]
+pub unsafe extern "C" fn punktfunk_connection_next_audio_pcm(
+    c: *mut PunktfunkConnection,
+    out: *mut PunktfunkAudioPcm,
+    timeout_ms: u32,
+) -> PunktfunkStatus {
+    guard(|| {
+        let c = match unsafe { c.as_ref() } {
+            Some(c) => c,
+            None => return PunktfunkStatus::NullPointer,
+        };
+        if out.is_null() {
+            return PunktfunkStatus::NullPointer;
+        }
+        let channels = crate::audio::normalize_channels(c.inner.audio_channels);
+        let pkt = match c
+            .inner
+            .next_audio(std::time::Duration::from_millis(timeout_ms as u64))
+        {
+            Ok(pkt) => pkt,
+            Err(e) => return e.status(),
+        };
+        let mut state = c.audio_pcm.lock().unwrap();
+        if state.decoder.is_none() {
+            let layout = crate::audio::layout_for(channels, false);
+            match opus::MSDecoder::new(48_000, layout.streams, layout.coupled, layout.mapping) {
+                Ok(d) => {
+                    // Largest legal Opus frame is 120 ms = 5760 samples/ch.
+                    state.pcm = vec![0f32; 5760 * channels as usize];
+                    state.decoder = Some(d);
+                }
+                Err(_) => return PunktfunkStatus::Unsupported,
+            }
+        }
+        let AudioPcmState { decoder, pcm } = &mut *state;
+        let dec = decoder.as_mut().unwrap();
+        // `decode_float` divides the output buffer length by the channel count to get the
+        // per-channel capacity; an empty payload requests packet-loss concealment.
+        match dec.decode_float(&pkt.data, pcm, false) {
+            Ok(frame_count) => {
+                unsafe {
+                    *out = PunktfunkAudioPcm {
+                        samples: pcm.as_ptr(),
+                        frame_count: frame_count as u32,
+                        channels,
+                        seq: pkt.seq,
+                        pts_ns: pkt.pts_ns,
+                    };
+                }
+                PunktfunkStatus::Ok
+            }
+            Err(_) => PunktfunkStatus::BadPacket,
+        }
+    })
+}
+
 /// Pull the next rumble (force-feedback) update, waiting up to `timeout_ms`. Amplitudes
 /// are 0..0xFFFF (`low` = low-frequency motor, `high` = high-frequency), `(0, 0)` = stop.
 /// Same timeout/closed semantics as [`punktfunk_connection_next_audio`].
@@ -1414,6 +1606,33 @@ pub unsafe extern "C" fn punktfunk_connection_color_info(
     })
 }
 
+/// Read the session's resolved chroma subsampling (from the host's Welcome) as the HEVC
+/// `chroma_format_idc`: `1` = 4:2:0 (the default every pre-4:4:4 host produced), `3` = full-chroma
+/// 4:4:4. `*out` is filled when non-NULL. The in-band SPS is authoritative; this lets the embedder
+/// pre-size its decoder / pick a 4:4:4 pixel format up front. Available immediately after a
+/// successful connect (it doesn't change without a reconfigure).
+///
+/// # Safety
+/// `c` is a valid connection handle; `out` is NULL or writable for one `u8`.
+#[cfg(feature = "quic")]
+#[no_mangle]
+pub unsafe extern "C" fn punktfunk_connection_chroma_format(
+    c: *mut PunktfunkConnection,
+    out: *mut u8,
+) -> PunktfunkStatus {
+    guard(|| {
+        let c = match unsafe { c.as_ref() } {
+            Some(c) => c,
+            None => return PunktfunkStatus::NullPointer,
+        };
+        if !out.is_null() {
+            // SAFETY: `out` is non-null and the caller guarantees it is writable for one `u8`.
+            unsafe { *out = c.inner.chroma_format };
+        }
+        PunktfunkStatus::Ok
+    })
+}
+
 /// Send one input event to the host as a QUIC datagram (non-blocking enqueue).
 ///
 /// # Safety

crates/punktfunk-core/src/audio.rs

@@ -0,0 +1,298 @@
+//! Shared audio layout: the single source of truth for Opus (multi)stream surround across the
+//! host, the GameStream compatibility path, and every client decoder.
+//!
+//! **Canonical wire channel order** is `FL FR FC LFE RL RR SL SR` (the GameStream/Moonlight
+//! order, and the PipeWire/PulseAudio default map for 6/8 channels). Every host capturer
+//! delivers PCM in this order and every client decodes into it, so the Opus multistream
+//! `mapping` is the **identity** (`[0, 1, …, channels-1]`) on both ends — punktfunk owns the
+//! encoder and every decoder, so the GFE-style pre-rotation Moonlight needs over SDP
+//! (`gamestream::audio::surround_params`) is a GameStream-only concern and never touches the
+//! native `punktfunk/1` path.
+//!
+//! Channel counts the protocol negotiates: `2` (stereo), `6` (5.1) and `8` (7.1). Anything
+//! else clamps to stereo ([`normalize_channels`]).
+
+/// Canonical wire channel positions; the index is the channel's slot in the interleaved PCM
+/// frame. A count of N uses positions `0..N` (always a prefix of this 8-channel order).
+#[derive(Clone, Copy, Debug, PartialEq, Eq)]
+#[repr(u8)]
+pub enum WirePos {
+    FrontLeft = 0,
+    FrontRight = 1,
+    FrontCenter = 2,
+    Lfe = 3,
+    RearLeft = 4,
+    RearRight = 5,
+    SideLeft = 6,
+    SideRight = 7,
+}
+
+/// The full 8-channel wire order; the N-channel order is its first N entries.
+pub const WIRE_ORDER_8: [WirePos; 8] = {
+    use WirePos::*;
+    [
+        FrontLeft,
+        FrontRight,
+        FrontCenter,
+        Lfe,
+        RearLeft,
+        RearRight,
+        SideLeft,
+        SideRight,
+    ]
+};
+
+/// One Opus (multi)stream layout. `mapping` is the libopus multistream mapping we encode AND
+/// decode with — identity, since punktfunk owns both ends. `streams`/`coupled` give the
+/// normal-quality coupling (FL,FR)+(FC,LFE) [+(RL,RR) on 7.1] with the remaining channels as
+/// mono streams; high quality is one mono stream per channel. Bitrates match Sunshine's
+/// per-config values (stereo keeps punktfunk's live-validated 128 kbps).
+#[derive(Clone, Copy, Debug, PartialEq, Eq)]
+pub struct OpusLayout {
+    /// Interleaved channel count (2, 6 or 8).
+    pub channels: u8,
+    /// Number of Opus streams in the multistream packet.
+    pub streams: u8,
+    /// How many of those streams are coupled (stereo) pairs.
+    pub coupled: u8,
+    /// libopus multistream channel mapping — identity `[0, 1, …, channels-1]`.
+    pub mapping: &'static [u8],
+    /// Target Opus bitrate in bits/sec (hard CBR; constant packet size, which GameStream's
+    /// audio FEC relies on).
+    pub bitrate: i32,
+}
+
+/// Stereo: a plain coupled pair. The 128 kbps live-validated config.
+pub const LAYOUT_STEREO: OpusLayout = OpusLayout {
+    channels: 2,
+    streams: 1,
+    coupled: 1,
+    mapping: &[0, 1],
+    bitrate: 128_000,
+};
+/// 5.1 normal quality: (FL,FR)+(FC,LFE) coupled, RL+RR mono.
+pub const LAYOUT_51: OpusLayout = OpusLayout {
+    channels: 6,
+    streams: 4,
+    coupled: 2,
+    mapping: &[0, 1, 2, 3, 4, 5],
+    bitrate: 256_000,
+};
+/// 5.1 high quality: one mono stream per channel.
+pub const LAYOUT_51_HQ: OpusLayout = OpusLayout {
+    channels: 6,
+    streams: 6,
+    coupled: 0,
+    mapping: &[0, 1, 2, 3, 4, 5],
+    bitrate: 1_536_000,
+};
+/// 7.1 normal quality: (FL,FR)+(FC,LFE)+(RL,RR) coupled, SL+SR mono.
+pub const LAYOUT_71: OpusLayout = OpusLayout {
+    channels: 8,
+    streams: 5,
+    coupled: 3,
+    mapping: &[0, 1, 2, 3, 4, 5, 6, 7],
+    bitrate: 450_000,
+};
+/// 7.1 high quality: one mono stream per channel.
+pub const LAYOUT_71_HQ: OpusLayout = OpusLayout {
+    channels: 8,
+    streams: 8,
+    coupled: 0,
+    mapping: &[0, 1, 2, 3, 4, 5, 6, 7],
+    bitrate: 2_048_000,
+};
+
+/// Pick the layout for a negotiated channel count. Unknown counts fall back to stereo (clients
+/// only ever request 2/6/8). `high_quality` selects the uncoupled high-bitrate config.
+pub fn layout_for(channels: u8, high_quality: bool) -> &'static OpusLayout {
+    match (channels, high_quality) {
+        (6, false) => &LAYOUT_51,
+        (6, true) => &LAYOUT_51_HQ,
+        (8, false) => &LAYOUT_71,
+        (8, true) => &LAYOUT_71_HQ,
+        _ => &LAYOUT_STEREO,
+    }
+}
+
+/// Clamp an arbitrary (wire / requested) channel count to one the protocol negotiates. `0`,
+/// absent, or any unsupported value becomes stereo.
+pub fn normalize_channels(requested: u8) -> u8 {
+    match requested {
+        6 => 6,
+        8 => 8,
+        _ => 2,
+    }
+}
+
+// ---- per-platform channel-layout helpers (pure data; no platform deps) --------------------
+
+/// Windows `WAVEFORMATEXTENSIBLE.dwChannelMask` for the wire layout.
+///
+/// NB 7.1 == `0x63F` (FL FR FC LFE **BL BR SL SR**), NOT `0xFF` — `0xFF` selects the
+/// front-of-center pair FLC/FRC, the wrong speakers. WASAPI delivers channels in ascending
+/// mask-bit order, which equals the wire order, so the decoded PCM needs no permutation.
+pub const fn wasapi_channel_mask(channels: u8) -> u32 {
+    const FL: u32 = 0x1;
+    const FR: u32 = 0x2;
+    const FC: u32 = 0x4;
+    const LFE: u32 = 0x8;
+    const BL: u32 = 0x10; // back left  (wire RL)
+    const BR: u32 = 0x20; // back right (wire RR)
+    const SL: u32 = 0x200; // side left
+    const SR: u32 = 0x400; // side right
+    match channels {
+        6 => FL | FR | FC | LFE | BL | BR,           // 0x3F
+        8 => FL | FR | FC | LFE | BL | BR | SL | SR, // 0x63F
+        _ => FL | FR,                                // 0x3 (stereo)
+    }
+}
+
+/// PipeWire / SPA `enum spa_audio_channel` positions in wire order — identical to the host
+/// capture side (`punktfunk-host` `audio::linux::spa_positions`): FL=3 FR=4 FC=5 LFE=6 SL=7
+/// SR=8 RL=12 RR=13. Identity routing: the client sets these on its playback node so PipeWire
+/// maps each wire slot to the matching speaker (and downmixes when the sink has fewer).
+pub fn spa_positions(channels: u8) -> &'static [u32] {
+    const STEREO: [u32; 2] = [3, 4]; // FL FR
+    const C51: [u32; 6] = [3, 4, 5, 6, 12, 13]; // FL FR FC LFE RL RR
+    const C71: [u32; 8] = [3, 4, 5, 6, 12, 13, 7, 8]; // FL FR FC LFE RL RR SL SR
+    match channels {
+        6 => &C51,
+        8 => &C71,
+        _ => &STEREO,
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn layout_table_is_consistent() {
+        for l in [
+            &LAYOUT_STEREO,
+            &LAYOUT_51,
+            &LAYOUT_51_HQ,
+            &LAYOUT_71,
+            &LAYOUT_71_HQ,
+        ] {
+            // Mapping is identity and exactly `channels` entries long.
+            assert_eq!(l.mapping.len(), l.channels as usize);
+            for (i, &m) in l.mapping.iter().enumerate() {
+                assert_eq!(m as usize, i, "mapping must be identity for {l:?}");
+            }
+            // libopus invariant: total channels == coupled*2 + (streams - coupled).
+            assert_eq!(
+                l.coupled * 2 + (l.streams - l.coupled),
+                l.channels,
+                "stream/coupled accounting for {l:?}"
+            );
+            assert!(l.coupled <= l.streams);
+            assert!(l.bitrate > 0);
+        }
+    }
+
+    #[test]
+    fn layout_for_picks_expected() {
+        assert_eq!(layout_for(2, false), &LAYOUT_STEREO);
+        assert_eq!(layout_for(6, false), &LAYOUT_51);
+        assert_eq!(layout_for(6, true), &LAYOUT_51_HQ);
+        assert_eq!(layout_for(8, false), &LAYOUT_71);
+        assert_eq!(layout_for(8, true), &LAYOUT_71_HQ);
+        // Unknown / 0 → stereo.
+        assert_eq!(layout_for(0, false), &LAYOUT_STEREO);
+        assert_eq!(layout_for(3, false), &LAYOUT_STEREO);
+        assert_eq!(layout_for(7, true), &LAYOUT_STEREO);
+    }
+
+    #[test]
+    fn normalize_clamps_to_negotiable() {
+        assert_eq!(normalize_channels(2), 2);
+        assert_eq!(normalize_channels(6), 6);
+        assert_eq!(normalize_channels(8), 8);
+        for bad in [0u8, 1, 3, 4, 5, 7, 9, 255] {
+            assert_eq!(normalize_channels(bad), 2, "{bad} must clamp to stereo");
+        }
+    }
+
+    #[test]
+    fn wasapi_masks_are_correct() {
+        assert_eq!(wasapi_channel_mask(2), 0x3);
+        assert_eq!(wasapi_channel_mask(6), 0x3F);
+        assert_eq!(wasapi_channel_mask(8), 0x63F); // NOT 0xFF
+                                                   // Bit count must equal the channel count.
+        assert_eq!(wasapi_channel_mask(2).count_ones(), 2);
+        assert_eq!(wasapi_channel_mask(6).count_ones(), 6);
+        assert_eq!(wasapi_channel_mask(8).count_ones(), 8);
+    }
+
+    #[test]
+    fn spa_positions_match_wire_order() {
+        assert_eq!(spa_positions(2), &[3, 4]);
+        assert_eq!(spa_positions(6), &[3, 4, 5, 6, 12, 13]);
+        assert_eq!(spa_positions(8), &[3, 4, 5, 6, 12, 13, 7, 8]);
+        assert_eq!(spa_positions(2).len(), 2);
+        assert_eq!(spa_positions(6).len(), 6);
+        assert_eq!(spa_positions(8).len(), 8);
+    }
+
+    /// Real-libopus proof that the shared layout round-trips with channel identity: a tone fed
+    /// into wire channel N (host `opus::MSEncoder`) comes back out on channel N (client
+    /// `opus::MSDecoder`), for stereo / 5.1 / 7.1. This is the single guarantee the whole
+    /// feature rests on — encoder layout == decoder layout == identity mapping — so if a layout
+    /// constant is ever wrong, this fails. Gated on `quic` (where `opus` is a dependency).
+    #[cfg(feature = "quic")]
+    #[test]
+    fn multistream_layout_roundtrips_with_channel_identity() {
+        const SR: u32 = 48_000;
+        const SAMPLES: usize = 240; // 5 ms @ 48 kHz
+        for &channels in &[2u8, 6, 8] {
+            let l = layout_for(channels, false);
+            let ch = l.channels as usize;
+            let mut enc = opus::MSEncoder::new(
+                SR,
+                l.streams,
+                l.coupled,
+                l.mapping,
+                opus::Application::LowDelay,
+            )
+            .expect("MSEncoder");
+            enc.set_bitrate(opus::Bitrate::Bits(l.bitrate)).unwrap();
+            enc.set_vbr(false).unwrap();
+            let mut dec =
+                opus::MSDecoder::new(SR, l.streams, l.coupled, l.mapping).expect("MSDecoder");
+
+            for tone_ch in 0..ch {
+                let mut out = vec![0u8; 4000];
+                let mut energy = vec![0f64; ch];
+                // A few frames to clear the codec startup transient before measuring.
+                for f in 0..8 {
+                    let mut frame = vec![0f32; SAMPLES * ch];
+                    for t in 0..SAMPLES {
+                        let phase = (f * SAMPLES + t) as f32 * 440.0 * 2.0 * std::f32::consts::PI
+                            / SR as f32;
+                        frame[t * ch + tone_ch] = 0.5 * phase.sin();
+                    }
+                    let n = enc.encode_float(&frame, &mut out).unwrap();
+                    let mut decoded = vec![0f32; SAMPLES * ch];
+                    let got = dec.decode_float(&out[..n], &mut decoded, false).unwrap();
+                    assert_eq!(got, SAMPLES, "{channels}ch frame size");
+                    if f >= 4 {
+                        for t in 0..SAMPLES {
+                            for (c, e) in energy.iter_mut().enumerate() {
+                                *e += (decoded[t * ch + c] as f64).powi(2);
+                            }
+                        }
+                    }
+                }
+                let loudest = (0..ch)
+                    .max_by(|&a, &b| energy[a].total_cmp(&energy[b]))
+                    .unwrap();
+                assert_eq!(
+                    loudest, tone_ch,
+                    "{channels}ch: tone in channel {tone_ch} must come out on {tone_ch} (energies {energy:?})"
+                );
+            }
+        }
+    }
+}

crates/punktfunk-core/src/client.rs

@@ -40,8 +40,9 @@ enum CtrlRequest {
 /// mode, the host-resolved compositor backend, the host-resolved gamepad backend, the host's
 /// certificate fingerprint, the resolved encoder bitrate (kbps), and the host↔client clock offset
 /// (ns, host minus client; 0 = no skew correction / an old host that didn't answer the handshake).
-/// The trailing `u8` is the resolved encode bit depth (8/10) and [`ColorInfo`] the resolved colour
-/// signalling, both from the [`Welcome`].
+/// The trailing `u8`s are the resolved encode bit depth (8/10), the chroma `chroma_format_idc`
+/// (1 = 4:2:0, 3 = 4:4:4), and the resolved audio channel count (2/6/8), with [`ColorInfo`] the
+/// resolved colour signalling — all from the [`Welcome`].
 type Negotiated = (
     Mode,
     CompositorPref,
@@ -51,6 +52,8 @@ type Negotiated = (
     i64,
     u8,
     ColorInfo,
+    u8,
+    u8,
 );
 
 /// Accumulated state of an in-flight / finished speed test. The data-plane pump mirrors the
@@ -202,6 +205,17 @@ pub struct NativeClient {
     /// decoder/presenter from this. [`ColorInfo::SDR_BT709`] for an older host. The static HDR
     /// mastering metadata (when [`ColorInfo::is_hdr`]) arrives via [`NativeClient::next_hdr_meta`].
     pub color: ColorInfo,
+    /// The chroma subsampling the host resolved for this session ([`Welcome::chroma_format`]), as the
+    /// HEVC `chroma_format_idc`: [`quic::CHROMA_IDC_420`] (4:2:0, the default / older host) or
+    /// [`quic::CHROMA_IDC_444`] (full-chroma 4:4:4). The in-band SPS is authoritative; this lets the
+    /// client pre-size its decoder. `CHROMA_IDC_420` for an older host that didn't report it.
+    pub chroma_format: u8,
+    /// The audio channel count the host resolved for this session ([`Welcome::audio_channels`]):
+    /// `2` (stereo), `6` (5.1) or `8` (7.1). The client MUST build its Opus (multistream) decoder
+    /// from this value (via [`crate::audio::layout_for`]) — never from its own request — so an older
+    /// host that omits it (→ `2`) yields working stereo. The `0xC9` audio frames are encoded with the
+    /// matching layout.
+    pub audio_channels: u8,
 }
 
 /// Pin the calling thread to the user-interactive QoS class on Apple targets.
@@ -246,6 +260,9 @@ impl NativeClient {
         // VIDEO_CAP_HDR) — the host upgrades to a 10-bit / HDR encode only when the matching bit is
         // set. 0 = the 8-bit BT.709 stream every client understands.
         video_caps: u8,
+        // Requested audio channel count (2 = stereo / 6 = 5.1 / 8 = 7.1); the host clamps to what it
+        // can capture and echoes the result in [`NativeClient::audio_channels`].
+        audio_channels: u8,
         launch: Option<String>,
         pin: Option<[u8; 32]>,
         identity: Option<(String, String)>,
@@ -298,6 +315,7 @@ impl NativeClient {
                     gamepad,
                     bitrate_kbps,
                     video_caps,
+                    audio_channels,
                     launch,
                     pin,
                     identity,
@@ -329,6 +347,8 @@ impl NativeClient {
             clock_offset_ns,
             bit_depth,
             color,
+            chroma_format,
+            audio_channels,
         ) = match ready_rx.recv_timeout(timeout) {
             Ok(Ok(t)) => t,
             Ok(Err(e)) => return Err(e),
@@ -360,6 +380,8 @@ impl NativeClient {
             clock_offset_ns,
             bit_depth,
             color,
+            chroma_format,
+            audio_channels,
         })
     }
 
@@ -666,6 +688,7 @@ struct WorkerArgs {
     gamepad: GamepadPref,
     bitrate_kbps: u32,
     video_caps: u8,
+    audio_channels: u8,
     launch: Option<String>,
     pin: Option<[u8; 32]>,
     identity: Option<(String, String)>,
@@ -697,6 +720,7 @@ async fn worker_main(args: WorkerArgs) {
         gamepad,
         bitrate_kbps,
         video_caps,
+        audio_channels,
         launch,
         pin,
         identity,
@@ -763,6 +787,8 @@ async fn worker_main(args: WorkerArgs) {
                 // VIDEO_CAP_10BIT | VIDEO_CAP_HDR). The host only upgrades to a 10-bit / HDR encode
                 // when the matching bit is set, so `0` stays an 8-bit BT.709 stream.
                 video_caps,
+                // Requested surround channel count; the host echoes the resolved value in Welcome.
+                audio_channels,
             }
             .encode(),
         )
@@ -834,6 +860,8 @@ async fn worker_main(args: WorkerArgs) {
             clock_offset_ns,
             welcome.bit_depth,
             welcome.color,
+            welcome.chroma_format,
+            welcome.audio_channels,
         ))
     };
 
@@ -850,6 +878,8 @@ async fn worker_main(args: WorkerArgs) {
         clock_offset_ns,
         bit_depth,
         color,
+        chroma_format,
+        audio_channels,
     ) = match setup.await {
         Ok(t) => t,
         Err(e) => {
@@ -866,6 +896,8 @@ async fn worker_main(args: WorkerArgs) {
         clock_offset_ns,
         bit_depth,
         color,
+        chroma_format,
+        audio_channels,
     )));
 
     // Input task: embedder events → QUIC datagrams.

crates/punktfunk-core/src/lib.rs

@@ -25,6 +25,7 @@
 #![forbid(unsafe_op_in_unsafe_fn)]
 
 pub mod abi;
+pub mod audio;
 #[cfg(feature = "quic")]
 pub mod client;
 pub mod config;

crates/punktfunk-core/src/quic.rs

@@ -78,12 +78,33 @@ pub struct Hello {
     /// zero-length name/launch placeholder precedes it when those are absent so the offset stays
     /// deterministic. Omitted by older clients (decodes to `0`).
     pub video_caps: u8,
+    /// Requested audio channel count: `2` (stereo, default), `6` (5.1) or `8` (7.1). The host
+    /// resolves it against what it can capture and echoes the final count in
+    /// [`Welcome::audio_channels`], which is what both ends build their Opus (multistream)
+    /// codec from. Appended after `video_caps` as a single trailing byte; when it differs from
+    /// the stereo default the name/launch/video_caps placeholders are forced (0) so it lands at a
+    /// deterministic offset. Omitted by older clients / when `2` (decodes to `2`, i.e. stereo) so
+    /// the stereo wire form stays byte-identical to the pre-surround build.
+    pub audio_channels: u8,
 }
 
 /// [`Hello::video_caps`] bit: the client can decode a 10-bit (Main10) HEVC stream.
 pub const VIDEO_CAP_10BIT: u8 = 0x01;
 /// [`Hello::video_caps`] bit: the client can present BT.2020 PQ HDR10 (implies 10-bit).
 pub const VIDEO_CAP_HDR: u8 = 0x02;
+/// [`Hello::video_caps`] bit: the client can decode a full-chroma **4:4:4** HEVC stream (HEVC
+/// Range Extensions / Rec.ITU-T H.265 `chroma_format_idc = 3`). The host emits 4:4:4 ONLY when this
+/// bit is set, the host opted in (`PUNKTFUNK_444`), the codec is HEVC, **and** the GPU/driver
+/// actually supports a 4:4:4 encode (probed) — otherwise the session stays 4:2:0 and
+/// [`Welcome::chroma_format`] reflects the real resolved value. Independent of 10-bit/HDR (4:4:4 is a
+/// chroma decision, bit depth is a depth decision; the two may combine where the hardware allows).
+pub const VIDEO_CAP_444: u8 = 0x04;
+
+/// HEVC `chroma_format_idc` for 4:2:0 — what every pre-4:4:4 build produced and the back-compat
+/// default when a peer omits [`Welcome::chroma_format`].
+pub const CHROMA_IDC_420: u8 = 1;
+/// HEVC `chroma_format_idc` for full-chroma 4:4:4 (Range Extensions).
+pub const CHROMA_IDC_444: u8 = 3;
 
 /// Per-session colour signalling (CICP / ITU-T H.273 code points) the host resolved for the
 /// encoded video, carried on [`Welcome`]. A client configures its decoder/presenter from these
@@ -198,6 +219,22 @@ pub struct Welcome {
     /// [`ColorInfo::SDR_BT709`]. The client configures its decoder/presenter from this instead of
     /// guessing from the bitstream; the mastering metadata arrives separately on [`HDR_META_MAGIC`].
     pub color: ColorInfo,
+    /// The chroma subsampling the host actually encodes at, as the HEVC `chroma_format_idc`:
+    /// [`CHROMA_IDC_420`] (4:2:0, default / older host) or [`CHROMA_IDC_444`] (full-chroma 4:4:4,
+    /// enabled only when the client advertised [`VIDEO_CAP_444`] *and* the host could open a real
+    /// 4:4:4 encode). The client sizes its decoder/surface pool from this; the in-band SPS carries
+    /// the authoritative value, so this is a hint (and the honest-downgrade channel — if the host
+    /// requested 4:4:4 but the GPU declined, this reads `CHROMA_IDC_420`). Appended after the colour
+    /// bytes as a single trailing byte; an older host that omits it decodes to [`CHROMA_IDC_420`].
+    pub chroma_format: u8,
+    /// The audio channel count the host actually resolved and **will** send on the `0xC9` plane:
+    /// `2` (stereo, default), `6` (5.1) or `8` (7.1). Echoes [`Hello::audio_channels`] clamped to
+    /// what the host can capture (Linux PipeWire always synthesizes the count; Windows WASAPI
+    /// loopback is clamped to the render endpoint's mix-format channels). The client builds its Opus
+    /// (multistream) decoder from THIS value via [`crate::audio::layout_for`] — never from its own
+    /// request — so an older host that omits the byte (→ `2`) always yields working stereo. Appended
+    /// after `chroma_format` as a single trailing byte.
+    pub audio_channels: u8,
 }
 
 /// `client → host`: data plane is bound, begin streaming.
@@ -630,10 +667,11 @@ impl Hello {
                                                                // so a Hello with neither name nor launch stays byte-identical to the bitrate-era form
                                                                // (26 bytes). When `launch` is present we must still emit name's length byte (0 for None)
                                                                // so `launch` lands at a deterministic offset.
-                                                               // `video_caps` is the last trailing field, after `launch`; when it's present (non-zero)
-                                                               // the name/launch length bytes must still be emitted (0 for absent) so it lands at a
+                                                               // `video_caps`/`audio_channels` are the trailing fields, after `launch`; when either is
+                                                               // present (video_caps non-zero / audio_channels not stereo) the name/launch length bytes
+                                                               // AND the video_caps byte must still be emitted (0 / 0) so the later byte lands at a
                                                                // deterministic offset — the same discipline `launch` already imposes on `name`.
-        let need_placeholders = self.video_caps != 0;
+        let need_placeholders = self.video_caps != 0 || self.audio_channels != 2;
         match (&self.name, &self.launch) {
             (None, None) if !need_placeholders => {}
             (name, _) => {
@@ -648,10 +686,15 @@ impl Hello {
             b.push(l.len() as u8);
             b.extend_from_slice(l.as_bytes());
         }
-        // video_caps: single trailing byte. Last field.
-        if self.video_caps != 0 {
+        // video_caps: single trailing byte. Emitted when non-zero OR when audio_channels follows
+        // (so audio_channels lands at a deterministic offset right after it).
+        if self.video_caps != 0 || self.audio_channels != 2 {
             b.push(self.video_caps);
         }
+        // audio_channels: single trailing byte. Last field; omitted when stereo (default).
+        if self.audio_channels != 2 {
+            b.push(self.audio_channels);
+        }
         b
     }
 
@@ -714,6 +757,15 @@ impl Hello {
                 let launch_len = b.get(launch_off).copied().unwrap_or(0) as usize;
                 b.get(launch_off + 1 + launch_len).copied().unwrap_or(0)
             },
+            // Optional trailing audio-channel byte, one past video_caps. Absent on an older client
+            // → stereo. Normalized so a corrupt/unsupported value can't build a bad decoder.
+            audio_channels: {
+                let name_len = b.get(26).copied().unwrap_or(0) as usize;
+                let launch_off = 27 + name_len;
+                let launch_len = b.get(launch_off).copied().unwrap_or(0) as usize;
+                let video_caps_off = launch_off + 1 + launch_len;
+                crate::audio::normalize_channels(b.get(video_caps_off + 1).copied().unwrap_or(2))
+            },
         })
     }
 }
@@ -747,6 +799,10 @@ impl Welcome {
         b.push(self.color.transfer);
         b.push(self.color.matrix);
         b.push(self.color.full_range);
+        // Chroma subsampling at offset 64 — older clients stop before this → 4:2:0 (CHROMA_IDC_420).
+        b.push(self.chroma_format);
+        // Audio channel count at offset 65 — older clients stop before this → stereo (2).
+        b.push(self.audio_channels);
         b
     }
 
@@ -755,7 +811,8 @@ impl Welcome {
         // scheme[22] pct[23] max_data[24..26] shard[26..28] encrypt[28] key[29..45]
         // salt[45..49] frames[49..53] compositor[53] gamepad[54] bitrate_kbps[55..59]
         // bit_depth[59] color.primaries[60] color.transfer[61] color.matrix[62] color.range[63]
-        // (everything from compositor on is an optional trailing byte; an older host stops earlier).
+        // chroma_format[64] audio_channels[65] (everything from compositor on is an optional
+        // trailing byte; an older host stops earlier).
         if b.len() < 53 || &b[0..4] != MAGIC {
             return Err(PunktfunkError::InvalidArg("bad Welcome"));
         }
@@ -812,6 +869,15 @@ impl Welcome {
                 matrix: b.get(62).copied().unwrap_or(ColorInfo::MC_BT709),
                 full_range: b.get(63).copied().unwrap_or(0),
             },
+            // Optional trailing chroma byte — absent on an older host (or an explicit 0 / unknown
+            // value) → 4:2:0. Only `CHROMA_IDC_444` flips the client to a 4:4:4 decode.
+            chroma_format: match b.get(64).copied() {
+                Some(CHROMA_IDC_444) => CHROMA_IDC_444,
+                _ => CHROMA_IDC_420,
+            },
+            // Optional trailing audio-channel byte — absent on an older host → stereo. Any
+            // non-{6,8} value normalizes to stereo so a corrupt byte never builds a bad decoder.
+            audio_channels: crate::audio::normalize_channels(b.get(65).copied().unwrap_or(2)),
         })
     }
 
@@ -1809,6 +1875,8 @@ mod tests {
             bitrate_kbps: 50_000,
             bit_depth: 10,
             color: ColorInfo::HDR10_BT2020_PQ,
+            chroma_format: CHROMA_IDC_444,
+            audio_channels: 2,
         };
         assert_eq!(Welcome::decode(&w.encode()).unwrap(), w);
     }
@@ -1851,6 +1919,7 @@ mod tests {
             name: Some("Test Device".into()),
             launch: Some("steam:570".into()),
             video_caps: VIDEO_CAP_10BIT,
+            audio_channels: 2,
         };
         assert_eq!(Hello::decode(&h.encode()).unwrap(), h);
         let s = Start {
@@ -1930,6 +1999,7 @@ mod tests {
             name: None,
             launch: None,
             video_caps: 0,
+            audio_channels: 2,
         };
         let enc = h.encode();
         assert_eq!(enc.len(), 26);
@@ -1969,9 +2039,11 @@ mod tests {
             bitrate_kbps: 120_000,
             bit_depth: 10,
             color: ColorInfo::HDR10_BT2020_PQ,
+            chroma_format: CHROMA_IDC_444,
+            audio_channels: 6, // 5.1 — exercises the non-default trailing byte
         };
         let wenc = w.encode();
-        assert_eq!(wenc.len(), 64); // 60 base + 4 colour bytes
+        assert_eq!(wenc.len(), 66); // 60 base + 4 colour + 1 chroma + 1 audio-channels byte
         let legacy_w = Welcome::decode(&wenc[..53]).unwrap();
         assert_eq!(legacy_w.compositor, CompositorPref::Auto);
         assert_eq!(legacy_w.gamepad, GamepadPref::Auto);
@@ -1991,13 +2063,29 @@ mod tests {
         let pre_color_w = Welcome::decode(&wenc[..60]).unwrap();
         assert_eq!(pre_color_w.bit_depth, 10);
         assert_eq!(pre_color_w.color, ColorInfo::SDR_BT709);
+        assert_eq!(pre_color_w.chroma_format, CHROMA_IDC_420); // pre-chroma host → 4:2:0
         assert_eq!(legacy_w.color, ColorInfo::SDR_BT709);
+        assert_eq!(legacy_w.chroma_format, CHROMA_IDC_420);
+        // A pre-chroma (64-byte) Welcome carries colour but no chroma/audio bytes → 4:2:0 + stereo.
+        let pre_chroma_w = Welcome::decode(&wenc[..64]).unwrap();
+        assert_eq!(pre_chroma_w.color, ColorInfo::HDR10_BT2020_PQ);
+        assert_eq!(pre_chroma_w.chroma_format, CHROMA_IDC_420);
+        assert_eq!(pre_chroma_w.audio_channels, 2); // audio byte (offset 65) absent → stereo
+                                                    // A pre-audio (65-byte) Welcome carries chroma but no audio byte → 4:4:4 + stereo.
+        let pre_audio_w = Welcome::decode(&wenc[..65]).unwrap();
+        assert_eq!(pre_audio_w.chroma_format, CHROMA_IDC_444);
+        assert_eq!(pre_audio_w.audio_channels, 2);
         assert_eq!(Welcome::decode(&wenc).unwrap().bitrate_kbps, 120_000);
         assert_eq!(Welcome::decode(&wenc).unwrap().bit_depth, 10); // full form carries it
         assert_eq!(
             Welcome::decode(&wenc).unwrap().color,
             ColorInfo::HDR10_BT2020_PQ
         );
+        assert_eq!(
+            Welcome::decode(&wenc).unwrap().chroma_format,
+            CHROMA_IDC_444
+        ); // full form carries 4:4:4
+        assert_eq!(Welcome::decode(&wenc).unwrap().audio_channels, 6); // ...and 5.1
     }
 
     #[test]
@@ -2015,6 +2103,7 @@ mod tests {
             name: Some("Enrico's MacBook".into()),
             launch: None,
             video_caps: 0,
+            audio_channels: 2,
         };
         let enc = base.encode();
         assert_eq!(
@@ -2062,6 +2151,7 @@ mod tests {
             name: None,
             launch: None,
             video_caps: 0,
+            audio_channels: 2,
         };
         // launch alone (no name): a zero-length name placeholder keeps the offset deterministic.
         let with_launch = Hello {
@@ -2268,6 +2358,7 @@ mod tests {
                 name: None,
                 launch: None,
                 video_caps: 0,
+                audio_channels: 2,
             }
             .encode();
             assert!(PairRequest::decode(&h).is_err(), "abi {abi} parsed as pair");

crates/punktfunk-core/tests/c_abi.rs

@@ -13,8 +13,10 @@ use std::process::Command;
 fn native_libs() -> &'static [&'static str] {
     if cfg!(target_os = "macos") {
         // The workspace build unifies features into the staticlib, and `quic` pulls
-        // rustls's platform verifier → Security/CoreFoundation.
+        // rustls's platform verifier → Security/CoreFoundation, plus libopus (the in-core
+        // `next_audio_pcm` decode path) which the `abi.rs` object references.
         &[
+            "-lopus",
             "-liconv",
             "-lm",
             "-framework",
@@ -23,7 +25,17 @@ fn native_libs() -> &'static [&'static str] {
             "CoreFoundation",
         ]
     } else if cfg!(target_os = "linux") {
-        &["-lgcc_s", "-lutil", "-lrt", "-lpthread", "-lm", "-ldl"]
+        // `-lopus`: the `quic` feature pulls in-core Opus decode (`next_audio_pcm`), whose
+        // symbols the linked `abi.rs` object references. Before `-lm` (opus needs libm).
+        &[
+            "-lopus",
+            "-lgcc_s",
+            "-lutil",
+            "-lrt",
+            "-lpthread",
+            "-lm",
+            "-ldl",
+        ]
     } else {
         &[]
     }

crates/punktfunk-host/Cargo.toml

@@ -61,9 +61,10 @@ utoipa-scalar = { version = "0.3", features = ["axum"] }
 tower = { version = "0.5", features = ["util"] }
 http-body-util = "0.1"
 
-# Opus stereo encode for the host->client audio plane. The `opus` crate vendors libopus via
-# `audiopus_sys` (cmake-built from source — no system lib, no vcpkg), so it builds on Windows MSVC
-# too (needs CMake + NASM, both on the box). Both platforms that have an audio-capture backend.
+# Opus encode for the host->client audio plane — stereo (`opus::Encoder`) AND 5.1/7.1 surround
+# (`opus::MSEncoder`, the safe multistream API the crate exposes; no `audiopus_sys` needed). The
+# crate vendors libopus (cmake-built from source — no system lib, no vcpkg), so it builds on Windows
+# MSVC too (needs CMake + NASM, both on the box). Both platforms that have an audio-capture backend.
 [target.'cfg(any(target_os = "linux", target_os = "windows"))'.dependencies]
 opus = "0.3"
 
@@ -99,10 +100,6 @@ serde_json = "1"
 rusqlite = { version = "0.40", features = ["bundled"] }
 # Builds/validates the xkb keymap uploaded to the virtual keyboard + tracks modifier state.
 xkbcommon = "0.8"
-# The safe `opus` crate is stereo-only; surround (5.1/7.1) needs the libopus *multistream*
-# encoder (`opus_multistream_encoder_*`). `audiopus_sys` is the sys layer `opus` already
-# vendors (same libopus link), so this adds bindings, not a second copy of the library.
-audiopus_sys = "0.2"
 # libei (EI sender) for the portable input path on KWin/GNOME (RemoteDesktop portal).
 # The `tokio` feature wires reis's event stream into tokio's reactor.
 reis = { version = "0.6.1", features = ["tokio"] }

crates/punktfunk-host/src/audio/windows/wasapi_cap.rs

@@ -1,7 +1,9 @@
 //! WASAPI loopback capture of the default render endpoint (system output) — the Windows analogue
-//! of the PipeWire sink-monitor backend. Delivers interleaved f32 PCM at 48 kHz stereo, ready for
-//! the existing Opus path with NO resampling (WASAPI shared-mode autoconvert does any SRC). WASAPI
-//! objects are COM-apartment-bound and not `Send`, so they live on a dedicated thread (mirrors
+//! of the PipeWire sink-monitor backend. Delivers interleaved f32 PCM at 48 kHz in the requested
+//! channel count (stereo / 5.1 / 7.1, canonical wire order FL FR FC LFE RL RR SL SR via the
+//! explicit `dwChannelMask`), ready for the Opus path with NO resampling (WASAPI shared-mode
+//! autoconvert does any SRC + up/downmix to the requested layout). WASAPI objects are
+//! COM-apartment-bound and not `Send`, so they live on a dedicated thread (mirrors
 //! `linux::PwAudioCapturer`); only the channel + stop flag + join handle are in the struct.
 
 use super::{AudioCapturer, SAMPLE_RATE};
@@ -14,9 +16,6 @@ use std::thread::{self, JoinHandle};
 use std::time::Duration;
 use wasapi::{DeviceEnumerator, Direction, SampleType, StreamMode, WaveFormat};
 
-// 48 kHz stereo 32-bit float: 2 channels * 4 bytes = 8 bytes per frame.
-const BLOCK_ALIGN: usize = 2 * 4;
-
 pub struct WasapiLoopbackCapturer {
     chunks: Receiver<Vec<f32>>,
     channels: u32,
@@ -27,8 +26,8 @@ pub struct WasapiLoopbackCapturer {
 impl WasapiLoopbackCapturer {
     pub fn open(channels: u32) -> Result<WasapiLoopbackCapturer> {
         anyhow::ensure!(
-            channels == 2,
-            "WASAPI loopback backend is stereo-only (got {channels})"
+            matches!(channels, 2 | 6 | 8),
+            "WASAPI loopback backend supports 2/6/8 channels (got {channels})"
         );
         let (tx, rx) = sync_channel::<Vec<f32>>(64);
         let stop = Arc::new(AtomicBool::new(false));
@@ -39,7 +38,7 @@ impl WasapiLoopbackCapturer {
         let join = thread::Builder::new()
             .name("punktfunk-wasapi-audio".into())
             .spawn(move || {
-                if let Err(e) = capture_thread(tx, stop_t, ready_tx) {
+                if let Err(e) = capture_thread(tx, stop_t, ready_tx, channels) {
                     tracing::error!(error = format!("{e:#}"), "wasapi loopback thread failed");
                 }
             })
@@ -47,7 +46,8 @@ impl WasapiLoopbackCapturer {
         match ready_rx.recv_timeout(Duration::from_secs(3)) {
             Ok(Ok(())) => {
                 tracing::info!(
-                    "WASAPI loopback capture: 48 kHz stereo f32 (default render endpoint)"
+                    channels,
+                    "WASAPI loopback capture: 48 kHz f32 (default render endpoint)"
                 );
                 Ok(WasapiLoopbackCapturer {
                     chunks: rx,
@@ -95,7 +95,10 @@ fn capture_thread(
     tx: SyncSender<Vec<f32>>,
     stop: Arc<AtomicBool>,
     ready: SyncSender<Result<()>>,
+    channels: u32,
 ) -> Result<()> {
+    // Interleaved f32: channels * 4 bytes per frame.
+    let block_align = channels as usize * 4;
     // COM must be initialized on THIS thread (MTA), before any device call.
     if let Err(e) = wasapi::initialize_mta()
         .ok()
@@ -115,10 +118,20 @@ fn capture_thread(
             .get_default_device(&Direction::Render)
             .context("default render endpoint (loopback needs a render device)")?;
         let mut audio_client = device.get_iaudioclient().context("IAudioClient")?;
-        // 48 kHz stereo f32 interleaved; autoconvert lets WASAPI's shared-mode SRC match the engine
-        // mix format to ours, so we never resample in Rust. Loopback is implied by capturing a
-        // RENDER device with Direction::Capture in shared mode (wasapi sets STREAMFLAGS_LOOPBACK).
-        let desired = WaveFormat::new(32, 32, &SampleType::Float, SAMPLE_RATE as usize, 2, None);
+        // 48 kHz f32 interleaved in the requested channel layout; autoconvert lets WASAPI's
+        // shared-mode SRC match the engine mix format to ours (incl. up/downmix to the requested
+        // channel count), so we never resample/remix in Rust. The explicit dwChannelMask pins the
+        // wire order (FL FR FC LFE RL RR SL SR; 7.1 = 0x63F, not 0xFF). Loopback is implied by
+        // capturing a RENDER device with Direction::Capture in shared mode (STREAMFLAGS_LOOPBACK).
+        let mask = punktfunk_core::audio::wasapi_channel_mask(channels as u8);
+        let desired = WaveFormat::new(
+            32,
+            32,
+            &SampleType::Float,
+            SAMPLE_RATE as usize,
+            channels as usize,
+            Some(mask),
+        );
         let (default_period, _min_period) =
             audio_client.get_device_period().context("device period")?;
         let mode = StreamMode::EventsShared {
@@ -154,7 +167,7 @@ fn capture_thread(
                     Err(e) => return Err(anyhow!("get_next_packet_size: {e}")),
                 }
             }
-            let whole = (bytes.len() / BLOCK_ALIGN) * BLOCK_ALIGN;
+            let whole = (bytes.len() / block_align) * block_align;
             if whole == 0 {
                 continue;
             }

crates/punktfunk-host/src/capture.rs

@@ -62,6 +62,11 @@ pub struct OutputFormat {
     /// HDR: the capturer converts to 10-bit (IDD-push FP16 → `Rgb10a2`; the DDA secure-desktop HDR hint).
     /// `false` = 8-bit SDR.
     pub hdr: bool,
+    /// Full-chroma 4:4:4 session: the capturer must keep full chroma — deliver packed **RGB**
+    /// (`Bgra` / `Rgb10a2`), NOT the subsampled `Nv12`/`P010` the Windows video-engine path produces by
+    /// default — because 4:4:4 can only be recovered from a full-chroma source. NVENC then does the
+    /// RGB→YUV444 CSC at encode (chroma_format_idc=3). `false` on every 4:2:0 session.
+    pub chroma_444: bool,
 }
 
 impl OutputFormat {
@@ -73,6 +78,8 @@ impl OutputFormat {
         OutputFormat {
             gpu: gpu_encode(),
             hdr,
+            // The GameStream + spike paths are always 4:2:0 (4:4:4 is punktfunk/1-native only).
+            chroma_444: false,
         }
     }
 }
@@ -361,13 +368,16 @@ pub fn open_portal_monitor() -> Result<Box<dyn Capturer>> {
 #[cfg(target_os = "linux")]
 pub fn capture_virtual_output(
     vout: crate::vdisplay::VirtualOutput,
-    _want: OutputFormat,
+    want: OutputFormat,
     _capture: crate::session_plan::CaptureBackend,
 ) -> Result<Box<dyn Capturer>> {
-    // The Linux host stays 8-bit (HDR is blocked upstream) and the portal negotiates its own format, so
-    // the `OutputFormat` is unused here; the capture backend is always the portal (the `CaptureBackend`
-    // arg is a Windows-only dispatch — ignored here).
-    linux::PortalCapturer::from_virtual_output(vout).map(|c| Box::new(c) as Box<dyn Capturer>)
+    // The Linux host stays 8-bit (HDR is blocked upstream) and the portal negotiates its own pixel
+    // format, so only `want.gpu` is honored here: it gates GPU zero-copy capture (the capture backend
+    // is always the portal — the `CaptureBackend` arg is a Windows-only dispatch). `gpu = false`
+    // (a 4:4:4 NVENC session) forces the CPU mmap path so the encoder gets CPU-resident RGB to swscale
+    // into YUV444P — otherwise it would receive CUDA frames and bail.
+    linux::PortalCapturer::from_virtual_output(vout, want.gpu)
+        .map(|c| Box::new(c) as Box<dyn Capturer>)
 }
 
 /// `PUNKTFUNK_NO_WGC=1` forces the pure single-process DDA (Desktop Duplication) path everywhere: it
@@ -394,6 +404,14 @@ pub fn capture_virtual_output(
     })?;
     let pref = vout.preferred_mode;
     let keep = vout.keepalive;
+    // Full-chroma 4:4:4 needs a full-chroma RGB source. The IDD-push and WGC paths emit subsampled
+    // NV12/P010 by default, which can't reconstruct 4:4:4; route a 4:4:4 session to DDA, which delivers
+    // RGB (Bgra) when its `chroma_444` flag is set. (IDD-push/WGC 4:4:4 capture is a follow-up.)
+    if want.chroma_444 && capture != CaptureBackend::Dda {
+        tracing::info!("4:4:4 session — using DDA capture (RGB source) instead of {capture:?}");
+        return dxgi::DuplCapturer::open(target, pref, keep, want.gpu, false, want.chroma_444)
+            .map(|c| Box::new(c) as Box<dyn Capturer>);
+    }
     // P2 direct frame push (kill DDA): consume frames straight from the pf-vdisplay driver's shared
     // ring — no Desktop Duplication, no win32u reparenting hook. Resolved once in the `SessionPlan`
     // (was re-derived from `config().idd_push` here); `IddPush` takes the keepalive (owns the virtual
@@ -414,8 +432,15 @@ pub fn capture_virtual_output(
                     error = %format!("{e:#}"),
                     "IDD-push open/attach failed — falling back to DDA"
                 );
-                return dxgi::DuplCapturer::open(target, pref, keep, want.gpu, false)
-                    .map(|c| Box::new(c) as Box<dyn Capturer>);
+                return dxgi::DuplCapturer::open(
+                    target,
+                    pref,
+                    keep,
+                    want.gpu,
+                    false,
+                    want.chroma_444,
+                )
+                .map(|c| Box::new(c) as Box<dyn Capturer>);
             }
         }
     }
@@ -426,7 +451,7 @@ pub fn capture_virtual_output(
     // chosen backend (it owns the SudoVDA keepalive), so there's no open-time auto-fallback. The
     // backend choice (`dda`/`dxgi`/`PUNKTFUNK_NO_WGC` → DDA, else WGC) is now resolved once in the plan.
     if capture == CaptureBackend::Dda {
-        return dxgi::DuplCapturer::open(target, pref, keep, want.gpu, false)
+        return dxgi::DuplCapturer::open(target, pref, keep, want.gpu, false, want.chroma_444)
             .map(|c| Box::new(c) as Box<dyn Capturer>);
     }
     // WGC default, with a watchdog'd DDA fallback. WGC's Direct3D11CaptureFramePool::CreateFreeThreaded
@@ -461,12 +486,12 @@ pub fn capture_virtual_output(
         }
         Ok(Err(e)) => {
             tracing::warn!(error = %format!("{e:#}"), "WGC open failed — falling back to DDA");
-            dxgi::DuplCapturer::open(target, pref, keep, want.gpu, false)
+            dxgi::DuplCapturer::open(target, pref, keep, want.gpu, false, want.chroma_444)
                 .map(|c| Box::new(c) as Box<dyn Capturer>)
         }
         Err(_) => {
             tracing::warn!("WGC open timed out (CreateFreeThreaded hang on the virtual display) — falling back to DDA");
-            dxgi::DuplCapturer::open(target, pref, keep, want.gpu, false)
+            dxgi::DuplCapturer::open(target, pref, keep, want.gpu, false, want.chroma_444)
                 .map(|c| Box::new(c) as Box<dyn Capturer>)
         }
     }

crates/punktfunk-host/src/capture/linux/mod.rs

@@ -89,21 +89,29 @@ impl PortalCapturer {
             node_id,
             "ScreenCast portal session started; connecting PipeWire"
         );
-        Ok(spawn_pipewire(Some(fd), node_id, None)?.into_capturer(node_id, None))
+        // This portal path (GameStream / monitor capture) is always 4:2:0, so allow zero-copy as before.
+        Ok(spawn_pipewire(Some(fd), node_id, None, true)?.into_capturer(node_id, None))
     }
 
     /// Build a capturer from an already-created virtual output ([`crate::vdisplay::VirtualOutput`]):
     /// connect PipeWire to its node (`remote_fd` selects portal-remote vs. default-daemon) and
     /// take ownership of its keepalive so the output lives exactly as long as this capturer. This
     /// is how the client's requested resolution becomes the captured resolution without scaling.
-    pub fn from_virtual_output(vout: crate::vdisplay::VirtualOutput) -> Result<PortalCapturer> {
+    /// `allow_zerocopy` mirrors [`OutputFormat::gpu`](crate::capture::OutputFormat): `false` forces the
+    /// CPU mmap path (a 4:4:4 NVENC session needs CPU-resident RGB), `true` keeps the GPU zero-copy
+    /// path subject to `PUNKTFUNK_ZEROCOPY`.
+    pub fn from_virtual_output(
+        vout: crate::vdisplay::VirtualOutput,
+        allow_zerocopy: bool,
+    ) -> Result<PortalCapturer> {
         tracing::info!(
             node_id = vout.node_id,
+            allow_zerocopy,
             "connecting PipeWire to virtual output"
         );
         let node_id = vout.node_id;
         Ok(
-            spawn_pipewire(vout.remote_fd, node_id, vout.preferred_mode)?
+            spawn_pipewire(vout.remote_fd, node_id, vout.preferred_mode, allow_zerocopy)?
                 .into_capturer(node_id, Some(vout.keepalive)),
         )
     }
@@ -146,6 +154,12 @@ fn spawn_pipewire(
     fd: Option<OwnedFd>,
     node_id: u32,
     preferred: Option<(u32, u32, u32)>,
+    // Allow GPU zero-copy capture (dmabuf→CUDA/VA). `false` forces the CPU mmap path even when
+    // `PUNKTFUNK_ZEROCOPY` is set — a 4:4:4 NVENC session needs CPU-resident RGB (the encoder
+    // swscales RGB→YUV444P; `hevc_nvenc` can't 4:4:4 from a CUDA RGB surface), so the session plan
+    // passes `gpu = false` for it. Without this, a 4:4:4 session under `PUNKTFUNK_ZEROCOPY=1` would
+    // get CUDA frames and the encoder would bail (`want_444 && cuda`).
+    allow_zerocopy: bool,
 ) -> Result<PwHandles> {
     // Frames flow from the pipewire thread over a small bounded channel.
     let (frame_tx, frame_rx) = sync_channel::<CapturedFrame>(8);
@@ -159,7 +173,7 @@ fn spawn_pipewire(
     // sender lives on the capturer and fires in its `Drop`. Absolute `::pipewire` path — the
     // inner `mod pipewire` shadows the crate name at this scope.
     let (quit_tx, quit_rx) = ::pipewire::channel::channel::<()>();
-    let zerocopy = crate::zerocopy::enabled();
+    let zerocopy = allow_zerocopy && crate::zerocopy::enabled();
     let join = thread::Builder::new()
         .name("punktfunk-pipewire".into())
         .spawn(move || {

crates/punktfunk-host/src/capture/windows/dxgi.rs

@@ -2010,6 +2010,10 @@ pub struct DuplCapturer {
     /// first, retried (legacy DuplicateOutput can't capture HDR). Set for the secure-desktop DDA leg
     /// when the SudoVDA is in HDR; threaded into every (re)duplication incl. ACCESS_LOST recovery.
     want_hdr: bool,
+    /// Full-chroma 4:4:4 session: deliver packed RGB (`Bgra` SDR / `Rgb10a2` HDR) and SKIP the
+    /// video-engine RGB→YUV (NV12/P010) conversion — NVENC reconstructs 4:4:4 only from a full-chroma
+    /// source, so we hand it the RGB texture and it CSCs to YUV444 at encode (chroma_format_idc=3).
+    chroma_444: bool,
     /// HDR (scRGB FP16) capture state. Set when the duplication surface is `R16G16B16A16_FLOAT`
     /// (the desktop has HDR on). The frame can't be `CopyResource`d into a BGRA target, so the HDR
     /// path copies it into an FP16 SRV texture, composites the cursor, then runs [`HdrConverter`] to
@@ -2087,6 +2091,8 @@ impl DuplCapturer {
         // stage 5) so the capturer never re-derives the encode backend itself.
         gpu: bool,
         want_hdr: bool,
+        // 4:4:4 session → deliver RGB, skip the NV12/P010 video-engine conversion (see the field doc).
+        chroma_444: bool,
     ) -> Result<Self> {
         // SAFETY: runs on the capture thread that will own this `DuplCapturer`. `install_gpu_pref_hook()`
         // and the DPI-context calls take by-value handles / no args and touch only thread/process state;
@@ -2311,6 +2317,7 @@ impl DuplCapturer {
                 gpu_copy: None,
                 last_present: None,
                 want_hdr,
+                chroma_444,
                 hdr_fp16: is_hdr_init,
                 hdr_meta: hdr_meta_init,
                 fp16_src: None,
@@ -3088,7 +3095,10 @@ impl DuplCapturer {
                                                     // Video-engine path: scRGB FP16 → BT.2020 PQ P010 on the VIDEO engine (no 3D shader, and
                                                     // NVENC encodes P010 natively). Fall back to the HdrConverter pixel shader (3D) only if the
                                                     // video processor is unavailable.
-            if let Some(p010) = self.convert_to_yuv(&src, true) {
+            if let Some(p010) = (!self.chroma_444)
+                .then(|| self.convert_to_yuv(&src, true))
+                .flatten()
+            {
                 self.last_present = Some((p010.clone(), PixelFormat::P010));
                 return Ok(CapturedFrame {
                     width: self.width,
@@ -3148,7 +3158,10 @@ impl DuplCapturer {
             // conversion AND NVENC's encode stay OFF the 3D engine — the only way to keep up when a
             // game pins the 3D engine at ~100%. Fall back to handing NVENC the BGRA texture (it then
             // does RGB→YUV internally on the 3D/compute engine).
-            if let Some(nv12) = self.convert_to_yuv(&gpu, false) {
+            if let Some(nv12) = (!self.chroma_444)
+                .then(|| self.convert_to_yuv(&gpu, false))
+                .flatten()
+            {
                 self.last_present = Some((nv12.clone(), PixelFormat::Nv12));
                 return Ok(CapturedFrame {
                     width: self.width,

crates/punktfunk-host/src/config.rs

@@ -7,7 +7,7 @@
 //! **Goal-1 stages 1–2** (`design/windows-host-rewrite.md` §2.2): stage 1 stood this up; stage 2 migrated the
 //! genuinely-constant operator/dispatch knobs onto it (the dispatch-disagreement bug class: `idd_push`,
 //! `capture_backend`, `encoder_pref`, `render_adapter`, `no_wgc`, the vdisplay backend select — plus the
-//! plan-named `secure_dda`/`idd_depth`/`zerocopy`/`ten_bit` and the multi-site `perf`/`compositor`/
+//! plan-named `secure_dda`/`idd_depth`/`zerocopy`/`ten_bit`/`four_four_four` and the multi-site `perf`/`compositor`/
 //! `video_source`/`gamepad`). `SessionPlan` (stage 3) consumes it as the single owner of the
 //! capture/topology/encoder decision.
 //!
@@ -63,6 +63,10 @@ pub struct HostConfig {
     pub zerocopy: bool,
     /// `PUNKTFUNK_10BIT` — host policy gate for HEVC Main10 (only honored when the client also advertised 10-bit).
     pub ten_bit: bool,
+    /// `PUNKTFUNK_444` — host policy gate for full-chroma HEVC 4:4:4 (Range Extensions). Honored only
+    /// when the client also advertised 4:4:4, the codec is HEVC, and the GPU/driver supports a 4:4:4
+    /// encode (probed) — otherwise the session stays 4:2:0. Independent of `ten_bit` (chroma vs depth).
+    pub four_four_four: bool,
     /// `PUNKTFUNK_PERF` — per-stage timing instrumentation.
     pub perf: bool,
     /// `PUNKTFUNK_VIDEO_SOURCE` — GameStream video source select (`virtual` / `portal` / unset → synthetic).
@@ -112,6 +116,7 @@ impl HostConfig {
                 .unwrap_or(2),
             zerocopy: flag("PUNKTFUNK_ZEROCOPY"),
             ten_bit: flag("PUNKTFUNK_10BIT"),
+            four_four_four: flag("PUNKTFUNK_444"),
             perf: flag("PUNKTFUNK_PERF"),
             video_source: val("PUNKTFUNK_VIDEO_SOURCE"),
             compositor: val("PUNKTFUNK_COMPOSITOR"),

crates/punktfunk-host/src/encode.rs

@@ -29,6 +29,33 @@ pub enum Codec {
     Av1,
 }
 
+/// Chroma subsampling the encoder emits, negotiated with the client (the `PUNKTFUNK_444` gate + the
+/// client's `VIDEO_CAP_444` + a GPU probe). `Yuv420` is the universal default; `Yuv444` is HEVC-only,
+/// native-protocol-only (GameStream stays 4:2:0), and the host only ever passes it after
+/// [`can_encode_444`] confirmed the active backend supports it.
+#[derive(Clone, Copy, Debug, PartialEq, Eq, Default)]
+pub enum ChromaFormat {
+    #[default]
+    Yuv420,
+    Yuv444,
+}
+
+impl ChromaFormat {
+    /// The HEVC `chroma_format_idc` this maps to: `1` (4:2:0) or `3` (4:4:4). Also the wire value
+    /// echoed in [`punktfunk_core::quic::Welcome::chroma_format`].
+    pub fn idc(self) -> u8 {
+        match self {
+            ChromaFormat::Yuv420 => punktfunk_core::quic::CHROMA_IDC_420,
+            ChromaFormat::Yuv444 => punktfunk_core::quic::CHROMA_IDC_444,
+        }
+    }
+
+    /// True for full-chroma 4:4:4.
+    pub fn is_444(self) -> bool {
+        matches!(self, ChromaFormat::Yuv444)
+    }
+}
+
 impl Codec {
     /// The FFmpeg NVENC encoder name (selected by name, not codec id — the latter would
     /// pick the software encoder).
@@ -89,6 +116,13 @@ pub struct EncoderCaps {
     /// When `false`, `set_hdr_meta` is a no-op and no in-band grade reaches the client. Only the
     /// Windows direct-NVENC path attaches it today.
     pub supports_hdr_metadata: bool,
+    /// The opened encoder is actually producing a full-chroma 4:4:4 (`chroma_format_idc = 3`) stream.
+    /// `false` on every 4:2:0 session (the default) and on a backend that declined 4:4:4. Set by the
+    /// NVENC backends (Linux + Windows). The chroma is committed to the wire (`Welcome::chroma_format`)
+    /// from the pre-open probe, so this is a *post-open cross-check*: the session glue logs loudly if
+    /// the encoder's real chroma disagrees with what was negotiated (the in-band SPS is authoritative
+    /// for the decoder either way).
+    pub chroma_444: bool,
 }
 
 /// A hardware encoder. One per session; runs on the encode thread.
@@ -193,8 +227,21 @@ pub fn open_video(
     bitrate_bps: u64,
     cuda: bool,
     bit_depth: u8,
+    chroma: ChromaFormat,
 ) -> Result<Box<dyn Encoder>> {
     validate_dimensions(codec, width, height)?;
+    // 4:4:4 is HEVC-only. The negotiator should never pass `Yuv444` for another codec (it gates on
+    // `codec == H265`), but defend the contract here so a future caller can't silently emit a stream
+    // no decoder expects: a non-HEVC 4:4:4 request degrades to 4:2:0 with a warning.
+    let chroma = if chroma.is_444() && codec != Codec::H265 {
+        tracing::warn!(
+            ?codec,
+            "4:4:4 requested for a non-HEVC codec — encoding 4:2:0"
+        );
+        ChromaFormat::Yuv420
+    } else {
+        chroma
+    };
     #[cfg(target_os = "linux")]
     {
         // Pick the GPU encode backend. NVIDIA → NVENC/CUDA (the original path, unchanged);
@@ -203,8 +250,17 @@ pub fn open_video(
         // its errors crisply instead of silently trying the other).
         let pref = crate::config::config().encoder_pref.as_str();
         let open_vaapi = || -> Result<Box<dyn Encoder>> {
-            vaapi::VaapiEncoder::open(codec, format, width, height, fps, bitrate_bps, bit_depth)
-                .map(|e| Box::new(e) as Box<dyn Encoder>)
+            vaapi::VaapiEncoder::open(
+                codec,
+                format,
+                width,
+                height,
+                fps,
+                bitrate_bps,
+                bit_depth,
+                chroma,
+            )
+            .map(|e| Box::new(e) as Box<dyn Encoder>)
         };
         match pref {
             "nvenc" | "nvidia" | "cuda" => open_nvenc_probed(
@@ -216,6 +272,7 @@ pub fn open_video(
                 bitrate_bps,
                 cuda,
                 bit_depth,
+                chroma,
             ),
             "vaapi" | "amd" | "intel" => open_vaapi(),
             "auto" | "" => {
@@ -231,6 +288,7 @@ pub fn open_video(
                         bitrate_bps,
                         cuda,
                         bit_depth,
+                        chroma,
                     )
                 } else {
                     open_vaapi()
@@ -260,6 +318,7 @@ pub fn open_video(
                         fps,
                         bitrate_bps,
                         bit_depth,
+                        chroma,
                     )
                     .map(|e| Box::new(e) as Box<dyn Encoder>)
                 }
@@ -289,6 +348,7 @@ pub fn open_video(
                         fps,
                         bitrate_bps,
                         bit_depth,
+                        chroma,
                     )
                     .map(|e| Box::new(e) as Box<dyn Encoder>)
                 }
@@ -333,6 +393,7 @@ pub fn open_video(
             bitrate_bps,
             cuda,
             bit_depth,
+            chroma,
         );
         anyhow::bail!("video encode requires Linux or Windows")
     }
@@ -355,6 +416,7 @@ fn open_nvenc_probed(
     bitrate_bps: u64,
     cuda: bool,
     bit_depth: u8,
+    chroma: ChromaFormat,
 ) -> Result<Box<dyn Encoder>> {
     const MIN_PROBE_BPS: u64 = 50_000_000;
     let mut candidates = vec![bitrate_bps];
@@ -369,7 +431,9 @@ fn open_nvenc_probed(
     }
     let mut last: Option<anyhow::Error> = None;
     for (i, &b) in candidates.iter().enumerate() {
-        match linux::NvencEncoder::open(codec, format, width, height, fps, b, cuda, bit_depth) {
+        match linux::NvencEncoder::open(
+            codec, format, width, height, fps, b, cuda, bit_depth, chroma,
+        ) {
             Ok(enc) => {
                 if i > 0 {
                     tracing::warn!(
@@ -446,6 +510,65 @@ pub fn vaapi_codec_support() -> CodecSupport {
     })
 }
 
+/// Whether the active GPU encode backend can actually produce a full-chroma **4:4:4** HEVC stream.
+/// Resolved (and cached, once) *before* the Welcome so the host advertises the chroma it will really
+/// encode — the honest-downgrade channel. 4:4:4 is HEVC-only; the probe opens a tiny encoder on the
+/// active backend (NVENC FREXT is broad on NVIDIA, but VAAPI / AMF / QSV 4:4:4 is hardware-specific,
+/// so it must be probed, never assumed). Non-HEVC codecs are always `false`.
+#[cfg(any(target_os = "linux", target_os = "windows"))]
+pub fn can_encode_444(codec: Codec) -> bool {
+    use std::sync::OnceLock;
+    if codec != Codec::H265 {
+        return false;
+    }
+    static CACHE: OnceLock<bool> = OnceLock::new();
+    *CACHE.get_or_init(|| {
+        let supported = {
+            #[cfg(target_os = "linux")]
+            {
+                // Mirror open_video's backend dispatch: VAAPI (AMD/Intel) vs NVENC (NVIDIA).
+                if linux_zero_copy_is_vaapi() {
+                    vaapi::probe_can_encode_444(codec)
+                } else {
+                    linux::probe_can_encode_444(codec)
+                }
+            }
+            #[cfg(target_os = "windows")]
+            {
+                match windows_resolved_backend() {
+                    WindowsBackend::Nvenc => {
+                        #[cfg(feature = "nvenc")]
+                        {
+                            nvenc::probe_can_encode_444(codec)
+                        }
+                        #[cfg(not(feature = "nvenc"))]
+                        {
+                            false
+                        }
+                    }
+                    WindowsBackend::Amf | WindowsBackend::Qsv => {
+                        #[cfg(feature = "amf-qsv")]
+                        {
+                            let vendor = match windows_resolved_backend() {
+                                WindowsBackend::Qsv => ffmpeg_win::WinVendor::Qsv,
+                                _ => ffmpeg_win::WinVendor::Amf,
+                            };
+                            ffmpeg_win::probe_can_encode_444(vendor, codec)
+                        }
+                        #[cfg(not(feature = "amf-qsv"))]
+                        {
+                            false
+                        }
+                    }
+                    WindowsBackend::Software => false,
+                }
+            }
+        };
+        tracing::info!(supported, "HEVC 4:4:4 encode capability probed");
+        supported
+    })
+}
+
 // ---------------------------------------------------------------------------------------------
 // Windows backend selection (the analogue of the Linux nvidia_present / linux_zero_copy_is_vaapi
 // logic). NVIDIA → NVENC, AMD → AMF, Intel → QSV; `auto` (default) reads the DXGI adapter vendor.

crates/punktfunk-host/src/encode/linux/mod.rs

@@ -11,7 +11,7 @@
 // Every `unsafe` block in this file carries a `// SAFETY:` proof; enforce it (unsafe-proof program).
 #![deny(clippy::undocumented_unsafe_blocks)]
 
-use super::{Codec, EncodedFrame, Encoder};
+use super::{ChromaFormat, Codec, EncodedFrame, Encoder};
 use crate::capture::{CapturedFrame, FramePayload, PixelFormat};
 use anyhow::{anyhow, bail, Context, Result};
 use ffmpeg::format::Pixel;
@@ -19,9 +19,33 @@ use ffmpeg::util::frame::Video as VideoFrame;
 use ffmpeg::{codec, encoder, Dictionary, Packet, Rational};
 use ffmpeg_next as ffmpeg;
 use std::os::raw::c_int;
+use std::ptr;
 
 use ffmpeg::ffi; // = ffmpeg_sys_next
 
+/// swscale: nearest-neighbour scaler flag (`SWS_POINT`). We never rescale (src dims == dst dims), so
+/// the resampler choice only governs the colour-conversion path; POINT is the cheapest.
+const SWS_POINT: c_int = 0x10;
+/// swscale colorspace id for ITU-R BT.709 (`SWS_CS_ITU709`) — the CSC coefficients for our RGB→YUV.
+const SWS_CS_ITU709: c_int = 1;
+
+/// The swscale *source* pixel format for a captured packed RGB/BGR layout (the real byte order, not
+/// the NVENC-padded `*0` form). Used by the 4:4:4 RGB→YUV444P conversion path. Mirrors the VAAPI
+/// CPU-input mapping; YUV/10-bit inputs can't feed this path (the 4:4:4 session forces packed RGB).
+fn sws_src_pixel(format: PixelFormat) -> Result<Pixel> {
+    Ok(match format {
+        PixelFormat::Bgrx => Pixel::BGRZ, // bgr0
+        PixelFormat::Rgbx => Pixel::RGBZ, // rgb0
+        PixelFormat::Bgra => Pixel::BGRA,
+        PixelFormat::Rgba => Pixel::RGBA,
+        PixelFormat::Rgb => Pixel::RGB24,
+        PixelFormat::Bgr => Pixel::BGR24,
+        PixelFormat::Nv12 | PixelFormat::P010 | PixelFormat::Rgb10a2 => {
+            bail!("NVENC 4:4:4 CPU-input path supports packed RGB/BGR only; got {format:?}")
+        }
+    })
+}
+
 /// `AVCUDADeviceContext` (libavutil/hwcontext_cuda.h) — not in the ffmpeg-sys bindings (the
 /// crate doesn't allowlist that header), so mirror its stable 3-pointer layout. We set the
 /// first field to *our* `CUcontext` so NVENC shares the context the EGL importer maps into.
@@ -131,6 +155,10 @@ pub struct NvencEncoder {
     frame: Option<VideoFrame>,
     /// Zero-copy path: CUDA hwdevice/hwframes contexts (the encoder takes `AV_PIX_FMT_CUDA`).
     cuda: Option<CudaHw>,
+    /// 4:4:4 path only: swscale context converting the captured packed RGB/BGR → planar YUV444P
+    /// (BT.709 limited) into [`Self::frame`], because `hevc_nvenc` only emits 4:4:4 from a YUV444
+    /// *input* (RGB-in is always 4:2:0). `None` on the ordinary 4:2:0 RGB path. Freed in `Drop`.
+    sws_444: Option<*mut ffi::SwsContext>,
     src_format: PixelFormat,
     expand: bool,
     width: u32,
@@ -142,10 +170,12 @@ pub struct NvencEncoder {
     force_kf: bool,
 }
 
-// `CudaHw` holds raw `AVBufferRef`s; the encoder lives on a single thread. The CPU encoder is
-// already `Send` via ffmpeg-next; assert it for the CUDA fields too.
+// `CudaHw` holds raw `AVBufferRef`s and `sws_444` a raw `SwsContext`; the encoder lives on a single
+// thread. The CPU encoder is already `Send` via ffmpeg-next; assert it for the raw fields too.
 // SAFETY: `NvencEncoder` owns an ffmpeg-next `Encoder`/`VideoFrame` (already `Send`) plus a `CudaHw`
-// holding raw `AVBufferRef`s, which are not `Send` by default. The encoder is owned and driven by
+// holding raw `AVBufferRef`s and an optional raw `SwsContext`, none of which are `Send` by default.
+// The `SwsContext` is a self-contained swscale state object with no thread affinity, touched only
+// through `&mut self` on the one encode thread. The encoder is owned and driven by
 // exactly ONE thread — the per-session encode thread it is moved to — and is only touched through
 // `&mut self` methods, so it is never aliased or accessed concurrently. The wrapped libav contexts
 // (and the shared `CUcontext` the `CudaHw` references) have no thread affinity, so transferring
@@ -164,6 +194,7 @@ impl NvencEncoder {
         bitrate_bps: u64,
         cuda: bool,
         bit_depth: u8,
+        chroma: ChromaFormat,
     ) -> Result<Self> {
         // TODO(hdr): Linux 10-bit parity. Unlike the Windows raw-SDK path (which upconverts 8-bit
         // ARGB → Main10 via pixelBitDepthMinus8), libavcodec hevc_nvenc needs a 10-bit input pixel
@@ -175,6 +206,18 @@ impl NvencEncoder {
                 "Linux NVENC 10-bit not yet wired — encoding 8-bit"
             );
         }
+        // Full-chroma 4:4:4 (HEVC Range Extensions). `hevc_nvenc` only emits 4:4:4 from a YUV444
+        // *input* frame — feeding RGB always subsamples to 4:2:0 regardless of profile (verified on
+        // the RTX 5070 Ti). So a 4:4:4 session swscales the captured RGB → YUV444P (BT.709 limited)
+        // and feeds that with `profile=rext`. The negotiator gates this to HEVC + the single-process
+        // CPU-capture topology, so `cuda` must be false here; defend the contract.
+        let want_444 = chroma.is_444() && codec == Codec::H265;
+        if want_444 && cuda {
+            bail!(
+                "NVENC 4:4:4 needs CPU RGB frames (the session forces non-zero-copy capture for \
+                 4:4:4); got a CUDA frame — capture/encoder negotiation mismatch"
+            );
+        }
         ffmpeg::init().context("ffmpeg init")?;
         if std::env::var_os("PUNKTFUNK_FFMPEG_DEBUG").is_some() {
             // SAFETY: `av_log_set_level` sets libav's global integer log level; `48` (= AV_LOG_DEBUG)
@@ -185,7 +228,14 @@ impl NvencEncoder {
         let name = codec.nvenc_name();
         let av_codec = encoder::find_by_name(name)
             .ok_or_else(|| anyhow!("{name} not built into libavcodec"))?;
-        let (nvenc_pixel, expand) = nvenc_input(format);
+        let (rgb_pixel, rgb_expand) = nvenc_input(format);
+        // 4:4:4 feeds NVENC a planar YUV444P frame we produce by swscale; the ordinary path feeds the
+        // captured RGB straight in and lets NVENC's internal CSC subsample to 4:2:0.
+        let (nvenc_pixel, expand) = if want_444 {
+            (Pixel::YUV444P, false)
+        } else {
+            (rgb_pixel, rgb_expand)
+        };
 
         let mut video = codec::context::Context::new_with_codec(av_codec)
             .encoder()
@@ -234,12 +284,12 @@ impl NvencEncoder {
             (*video.as_mut_ptr()).gop_size = -1;
         }
 
-        // NV12 path: we did the RGB→YUV conversion ourselves as BT.709 *limited* range, so signal
-        // that in the bitstream VUI (colorspace/range/primaries/transfer) — otherwise the client
-        // decoder assumes a default and the picture comes out washed-out / wrong-contrast. The
-        // RGB-input paths leave these unset (NVENC's internal CSC writes its own VUI). Matches the
-        // Windows NV12 path's BT.709 limited-range signalling.
-        if matches!(format, PixelFormat::Nv12) {
+        // NV12 / 4:4:4 paths: we do the RGB→YUV conversion ourselves as BT.709 *limited* range
+        // (swscale), so signal that in the bitstream VUI (colorspace/range/primaries/transfer) —
+        // otherwise the client decoder assumes a default and the picture comes out washed-out /
+        // wrong-contrast. The RGB-input 4:2:0 path leaves these unset (NVENC's internal CSC writes
+        // its own VUI). Matches the Windows NV12 path's BT.709 limited-range signalling.
+        if matches!(format, PixelFormat::Nv12) || want_444 {
             // SAFETY: same `video` builder — `raw = video.as_mut_ptr()` is the non-null, properly-
             // aligned, sole-owned, not-yet-opened `AVCodecContext`. We set its four VUI colour enum
             // fields to valid `AVColorSpace`/`AVColorRange`/`AVColorPrimaries`/`AVColorTransfer-
@@ -280,6 +330,45 @@ impl NvencEncoder {
             None
         };
 
+        // 4:4:4: build the RGB→YUV444P swscale (BT.709 limited, no rescale). Mirrors the VAAPI CPU
+        // path's RGB→NV12 scaler, but the dst is full-chroma planar 4:4:4.
+        let sws_444 = if want_444 {
+            let src_av = pixel_to_av(sws_src_pixel(format)?);
+            // SAFETY: `sws_getContext` allocates a swscale context for the given src/dst dims + pixel
+            // formats. Both dims are the encoder's positive `width`/`height` as `c_int`; `src_av` is a
+            // valid `AVPixelFormat` (from the `sws_src_pixel`-validated, packed-RGB-only source), the
+            // dst is YUV444P. The trailing filter/param pointers are null = "use defaults" (documented
+            // as accepted). No Rust memory is borrowed; the returned pointer is null-checked below.
+            let sws = unsafe {
+                ffi::sws_getContext(
+                    width as c_int,
+                    height as c_int,
+                    src_av,
+                    width as c_int,
+                    height as c_int,
+                    ffi::AVPixelFormat::AV_PIX_FMT_YUV444P,
+                    SWS_POINT,
+                    ptr::null_mut(),
+                    ptr::null_mut(),
+                    ptr::null(),
+                )
+            };
+            if sws.is_null() {
+                bail!("sws_getContext(RGB→YUV444P) failed");
+            }
+            // SAFETY: `sws` is the non-null context from the call above (null-checked). The ITU-709
+            // coefficient table from `sws_getCoefficients` is a process-lifetime libswscale static,
+            // reused for src+dst matrices; `sws_setColorspaceDetails` only reads it and writes scalar
+            // CSC settings into `sws` (limited-range dst: dstRange = 0). No Rust memory is passed.
+            unsafe {
+                let cs709 = ffi::sws_getCoefficients(SWS_CS_ITU709);
+                ffi::sws_setColorspaceDetails(sws, cs709, 1, cs709, 0, 0, 1 << 16, 1 << 16);
+            }
+            Some(sws)
+        } else {
+            None
+        };
+
         // Low-latency NVENC tuning (plan §7 / linux-setup doc).
         let mut opts = Dictionary::new();
         opts.set("preset", "p1"); // fastest
@@ -288,6 +377,12 @@ impl NvencEncoder {
         opts.set("bf", "0");
         opts.set("delay", "0");
         opts.set("forced-idr", "1"); // RFI/request_keyframe → real IDR under the infinite GOP
+        if want_444 {
+            // HEVC Range Extensions — the profile that carries chroma_format_idc=3. With a YUV444P
+            // input `hevc_nvenc` auto-selects it, but pin it explicitly so the chroma is never silently
+            // dropped on a future libavcodec.
+            opts.set("profile", "rext");
+        }
 
         // Split-frame encode across both NVENC engines (GB203 has 2) when the pixel rate exceeds
         // a single engine's HEVC capacity (~1 Gpix/s); e.g. 5120x1440@240 = 1.77 Gpix/s needs it,
@@ -321,6 +416,7 @@ impl NvencEncoder {
             enc,
             frame,
             cuda: cuda_hw,
+            sws_444,
             src_format: format,
             expand,
             width,
@@ -333,6 +429,15 @@ impl NvencEncoder {
 }
 
 impl Encoder for NvencEncoder {
+    fn caps(&self) -> super::EncoderCaps {
+        super::EncoderCaps {
+            // 4:4:4 iff this session opened the RGB→YUV444P swscale path (FREXT). RFI/HDR-SEI stay
+            // unsupported on libavcodec NVENC (the trait defaults).
+            chroma_444: self.sws_444.is_some(),
+            ..super::EncoderCaps::default()
+        }
+    }
+
     fn submit(&mut self, captured: &CapturedFrame) -> Result<()> {
         anyhow::ensure!(
             captured.width == self.width && captured.height == self.height,
@@ -411,6 +516,47 @@ impl NvencEncoder {
             bytes.len(),
             src_row * h
         );
+        // 4:4:4: swscale the packed RGB straight into the planar YUV444P input frame (BT.709 limited),
+        // then send it — no byte-expand. The 4:2:0 RGB path (below) feeds NVENC packed RGB directly.
+        if let Some(sws) = self.sws_444 {
+            let frame = self
+                .frame
+                .as_mut()
+                .context("CPU frame missing (encoder opened in CUDA mode)")?;
+            // SAFETY: `format == self.src_format` and `bytes.len() >= src_row * h` (the `ensure!`s
+            // above), so `sws_scale` reads `h` rows of `src_row` bytes from `src_data[0] = bytes`
+            // (packed RGB is single-plane; the other src planes are null/0) — all in bounds. `sws` is
+            // the non-null context built in `open`. The dst is `frame`'s underlying `AVFrame`: its
+            // `data`/`linesize` in-struct arrays were sized for YUV444P by `VideoFrame::new`, and the
+            // 3 planes are each `width`×`height`. All pointers are live locals for this synchronous
+            // call; the encoder runs only on this thread (`unsafe impl Send`), so no aliasing/race.
+            unsafe {
+                let dst_av = frame.as_mut_ptr();
+                let src_data: [*const u8; 4] =
+                    [bytes.as_ptr(), ptr::null(), ptr::null(), ptr::null()];
+                let src_stride: [c_int; 4] = [src_row as c_int, 0, 0, 0];
+                let r = ffi::sws_scale(
+                    sws,
+                    src_data.as_ptr(),
+                    src_stride.as_ptr(),
+                    0,
+                    h as c_int,
+                    (*dst_av).data.as_ptr(),
+                    (*dst_av).linesize.as_ptr(),
+                );
+                if r < 0 {
+                    bail!("sws_scale(RGB→YUV444P) failed ({r})");
+                }
+            }
+            frame.set_pts(Some(pts));
+            frame.set_kind(if idr {
+                ffmpeg::picture::Type::I
+            } else {
+                ffmpeg::picture::Type::None
+            });
+            self.enc.send_frame(frame).context("send_frame(444)")?;
+            return Ok(());
+        }
         let frame = self
             .frame
             .as_mut()
@@ -526,3 +672,51 @@ impl NvencEncoder {
         Ok(())
     }
 }
+
+impl Drop for NvencEncoder {
+    fn drop(&mut self) {
+        if let Some(sws) = self.sws_444.take() {
+            // SAFETY: `sws` is the non-null `SwsContext` allocated by `sws_getContext` in `open` and
+            // owned exclusively by this encoder (taken out of the field so it can't be freed twice).
+            // `sws_freeContext` frees it; nothing else references it after this single-threaded drop.
+            unsafe { ffi::sws_freeContext(sws) };
+        }
+    }
+}
+
+/// Probe whether this NVIDIA GPU + driver + libavcodec can actually encode HEVC **4:4:4** (Range
+/// Extensions). Opens a tiny real `hevc_nvenc` 4:4:4 session — the exact path [`NvencEncoder::open`]
+/// takes for a live 4:4:4 stream — and reports whether it succeeded. HEVC-only; the result is cached
+/// by the caller ([`crate::encode::can_encode_444`]). A GPU/driver/ffmpeg without RExt 4:4:4 fails
+/// the open here, so the host resolves the session to 4:2:0 before the Welcome (honest downgrade).
+pub fn probe_can_encode_444(codec: Codec) -> bool {
+    if codec != Codec::H265 {
+        return false;
+    }
+    if ffmpeg::init().is_err() {
+        return false;
+    }
+    // Quiet ffmpeg's open error on a GPU that lacks 4:4:4 — the probe failing is an expected outcome.
+    // SAFETY: libav initialized above; `av_log_{get,set}_level` only read/write the global int level
+    // (no pointer args) and are always sound post-init.
+    let prev = unsafe {
+        let p = ffi::av_log_get_level();
+        ffi::av_log_set_level(ffi::AV_LOG_FATAL);
+        p
+    };
+    let ok = NvencEncoder::open(
+        codec,
+        PixelFormat::Bgra,
+        640,
+        480,
+        30,
+        2_000_000,
+        false, // CPU input (the 4:4:4 path never uses CUDA)
+        8,
+        ChromaFormat::Yuv444,
+    )
+    .is_ok();
+    // SAFETY: restore the saved global log level (scalar arg, no pointers).
+    unsafe { ffi::av_log_set_level(prev) };
+    ok
+}

crates/punktfunk-host/src/encode/linux/vaapi.rs

@@ -160,6 +160,18 @@ pub fn probe_can_encode(codec: Codec) -> bool {
     }
 }
 
+/// Whether the active VAAPI GPU can encode HEVC **4:4:4** (Range Extensions). **Deferred in v1 —
+/// always `false`.** VAAPI HEVC 4:4:4 encode is narrow and vendor-specific (the lab's AMD Phoenix1 /
+/// RDNA3 exposes only `VAProfileHEVCMain`/`Main10` `EncSlice`, no `Main444`), and there is no
+/// validated hardware to build + verify the 4:4:4 surface/profile path against. Returning `false`
+/// keeps the negotiation honest: a VAAPI host resolves every session to 4:2:0 before the Welcome, so
+/// the client never builds a 4:4:4 decoder it would only get 4:2:0 frames for. (Follow-up: implement
+/// + validate on an Intel Arc / RDNA4-class box that advertises a HEVC 4:4:4 encode entrypoint.)
+pub fn probe_can_encode_444(_codec: Codec) -> bool {
+    tracing::info!("VAAPI HEVC 4:4:4 encode is not implemented yet — declining (encoding 4:2:0)");
+    false
+}
+
 /// Drain the encoder for one packet (shared poll logic).
 fn poll_encoder(enc: &mut encoder::video::Encoder, fps: u32) -> Result<Option<EncodedFrame>> {
     let mut pkt = Packet::empty();
@@ -848,6 +860,7 @@ pub struct VaapiEncoder {
 unsafe impl Send for VaapiEncoder {}
 
 impl VaapiEncoder {
+    #[allow(clippy::too_many_arguments)]
     pub fn open(
         codec: Codec,
         format: PixelFormat,
@@ -856,10 +869,18 @@ impl VaapiEncoder {
         fps: u32,
         bitrate_bps: u64,
         bit_depth: u8,
+        chroma: super::ChromaFormat,
     ) -> Result<Self> {
         if bit_depth != 8 {
             tracing::warn!(bit_depth, "VAAPI 10-bit not yet wired — encoding 8-bit");
         }
+        // VAAPI 4:4:4 is deferred (see `probe_can_encode_444`): no validated AMD/Intel hardware in the
+        // lab exposes a HEVC 4:4:4 encode entrypoint, and the probe returns false so the host never
+        // negotiates 4:4:4 for a VAAPI session. If a request slips through, fall back to 4:2:0 rather
+        // than emit an unverified stream — the host signalled 4:2:0 in the Welcome anyway.
+        if chroma.is_444() {
+            tracing::warn!("VAAPI 4:4:4 encode not implemented — encoding 4:2:0");
+        }
         ffmpeg::init().context("ffmpeg init")?;
         if std::env::var_os("PUNKTFUNK_FFMPEG_DEBUG").is_some() {
             // SAFETY: `av_log_set_level` sets libav's global integer log level; `48` (= AV_LOG_DEBUG)

crates/punktfunk-host/src/encode/windows/ffmpeg_win.rs

@@ -31,7 +31,7 @@
 // Every `unsafe` block in this file carries a `// SAFETY:` proof; enforce it (unsafe-proof program).
 #![deny(clippy::undocumented_unsafe_blocks)]
 
-use super::{Codec, EncodedFrame, Encoder};
+use super::{ChromaFormat, Codec, EncodedFrame, Encoder};
 use crate::capture::{dxgi::D3d11Frame, CapturedFrame, FramePayload, PixelFormat};
 use anyhow::{anyhow, bail, Context, Result};
 use ffmpeg::format::Pixel;
@@ -241,6 +241,18 @@ unsafe fn open_win_encoder(
 /// driver/runtime rejects codecs the video engine can't do (AV1 on pre-RDNA3 AMD / pre-Arc Intel,
 /// or HEVC on a very old part). Used to build the GameStream codec advertisement so a client never
 /// negotiates a codec the encoder can't open. Torn down immediately.
+/// Whether the active AMD (AMF) / Intel (QSV) GPU can encode HEVC **4:4:4**. **Deferred in v1 —
+/// always `false`.** AMF/QSV HEVC 4:4:4 encode is narrow (AMD RDNA3+, Intel Arc/Xe2+) and the
+/// libavcodec profile/pixel-format incantation is vendor- and driver-specific — a wrong profile
+/// `avcodec_open2` *silently* falls back to 4:2:0, so a positive probe would need a verify-by-frame,
+/// and there is no AMD/Intel Windows box in the lab to build + validate that against. Returning
+/// `false` keeps the negotiation honest: an AMF/QSV host resolves every session to 4:2:0 before the
+/// Welcome. (Follow-up: implement + validate on an RDNA3+/Arc Windows box.)
+pub fn probe_can_encode_444(_vendor: WinVendor, _codec: Codec) -> bool {
+    tracing::info!("AMF/QSV HEVC 4:4:4 encode is not implemented yet — declining (encoding 4:2:0)");
+    false
+}
+
 pub fn probe_can_encode(vendor: WinVendor, codec: Codec) -> bool {
     if ffmpeg::init().is_err() {
         return false;
@@ -1096,6 +1108,7 @@ pub struct FfmpegWinEncoder {
 unsafe impl Send for FfmpegWinEncoder {}
 
 impl FfmpegWinEncoder {
+    #[allow(clippy::too_many_arguments)]
     #[allow(clippy::too_many_arguments)]
     pub fn open(
         vendor: WinVendor,
@@ -1106,7 +1119,15 @@ impl FfmpegWinEncoder {
         fps: u32,
         bitrate_bps: u64,
         bit_depth: u8,
+        chroma: ChromaFormat,
     ) -> Result<Self> {
+        // AMF/QSV 4:4:4 is deferred (see `probe_can_encode_444`): no validated AMD/Intel Windows
+        // hardware in the lab, and the AMF/QSV HEVC 4:4:4 profile/format incantations are vendor- and
+        // driver-specific (a wrong profile silently encodes 4:2:0). The probe returns false so the host
+        // never negotiates 4:4:4 for an AMF/QSV session; if a request slips through, fall back to 4:2:0.
+        if chroma.is_444() {
+            tracing::warn!("AMF/QSV 4:4:4 encode not implemented — encoding 4:2:0");
+        }
         ffmpeg::init().context("ffmpeg init")?;
         if std::env::var_os("PUNKTFUNK_FFMPEG_DEBUG").is_some() {
             // SAFETY: `ffmpeg::init()` ran on the line above, so libav is initialised; `av_log_set_level`

crates/punktfunk-host/src/encode/windows/nvenc.rs

@@ -16,7 +16,7 @@
 // Every `unsafe` block / impl in this file carries a `// SAFETY:` proof; enforce it.
 #![deny(clippy::undocumented_unsafe_blocks)]
 
-use super::{Codec, EncodedFrame, Encoder, EncoderCaps};
+use super::{ChromaFormat, Codec, EncodedFrame, Encoder, EncoderCaps};
 use crate::capture::{CapturedFrame, FramePayload, PixelFormat};
 use anyhow::{anyhow, bail, Context, Result};
 use std::collections::{HashMap, VecDeque};
@@ -57,6 +57,15 @@ pub struct NvencD3d11Encoder {
     buffer_fmt: nv::NV_ENC_BUFFER_FORMAT,
     /// Encoded bit depth (8 or 10). 10 → HEVC Main10 (NVENC upconverts the 8-bit ARGB input).
     bit_depth: u8,
+    /// Full-chroma 4:4:4 (HEVC Range Extensions, `chroma_format_idc = 3`) requested for this session.
+    /// NVENC ingests the RGB (ARGB/ABGR10) input and CSCs it to YUV444 internally — the `FREXT` profile
+    /// + `chromaFormatIDC = 3` in the encode config carry the chroma. Gated on the GPU's
+    /// `NV_ENC_CAPS_SUPPORT_YUV444_ENCODE` (cleared in `query_caps` on a card that lacks it) and on an
+    /// RGB input format (NV12/P010 capture can't reconstruct 4:4:4). HEVC-only.
+    chroma_444: bool,
+    /// `NV_ENC_CAPS_SUPPORT_YUV444_ENCODE` from the caps probe — whether this GPU can 4:4:4 encode at
+    /// all. `chroma_444` is forced off when this is false (graceful downgrade to 4:2:0).
+    yuv444_supported: bool,
     /// HDR: the capturer is delivering BT.2020 PQ 10-bit (`PixelFormat::Rgb10a2`) frames. Sets the
     /// `ABGR10` input format + the BT.2020/PQ colour VUI. Derived per-frame from the capture format
     /// (HDR can toggle mid-session); a change re-inits the session.
@@ -103,6 +112,7 @@ pub struct NvencD3d11Encoder {
 unsafe impl Send for NvencD3d11Encoder {}
 
 impl NvencD3d11Encoder {
+    #[allow(clippy::too_many_arguments)]
     pub fn open(
         codec: Codec,
         _format: PixelFormat,
@@ -111,6 +121,7 @@ impl NvencD3d11Encoder {
         fps: u32,
         bitrate_bps: u64,
         bit_depth: u8,
+        chroma: ChromaFormat,
     ) -> Result<Self> {
         Ok(Self {
             encoder: ptr::null_mut(),
@@ -122,6 +133,9 @@ impl NvencD3d11Encoder {
             bitrate_bps,
             buffer_fmt: nv::NV_ENC_BUFFER_FORMAT::NV_ENC_BUFFER_FORMAT_ARGB,
             bit_depth,
+            // 4:4:4 is HEVC-only; the GPU-support gate is applied in `query_caps`.
+            chroma_444: chroma.is_444() && codec == Codec::H265,
+            yuv444_supported: false,
             hdr: false,
             hdr_meta: None,
             regs: HashMap::new(),
@@ -209,6 +223,7 @@ impl NvencD3d11Encoder {
         let wmax = self.get_cap(enc, nv::NV_ENC_CAPS::NV_ENC_CAPS_WIDTH_MAX);
         let hmax = self.get_cap(enc, nv::NV_ENC_CAPS::NV_ENC_CAPS_HEIGHT_MAX);
         let ten_bit = self.get_cap(enc, nv::NV_ENC_CAPS::NV_ENC_CAPS_SUPPORT_10BIT_ENCODE);
+        let yuv444 = self.get_cap(enc, nv::NV_ENC_CAPS::NV_ENC_CAPS_SUPPORT_YUV444_ENCODE);
         let rfi = self.get_cap(
             enc,
             nv::NV_ENC_CAPS::NV_ENC_CAPS_SUPPORT_REF_PIC_INVALIDATION,
@@ -235,6 +250,13 @@ impl NvencD3d11Encoder {
             self.bit_depth = 8;
             self.hdr = false;
         }
+        // Same for 4:4:4: a card without YUV444 encode falls back to 4:2:0. (The host already probed
+        // this via `probe_can_encode_444` before the Welcome, so this is a belt-and-braces guard.)
+        self.yuv444_supported = yuv444 != 0;
+        if self.chroma_444 && !self.yuv444_supported {
+            tracing::warn!("NVENC: this GPU can't 4:4:4 encode — falling back to 4:2:0");
+            self.chroma_444 = false;
+        }
         self.rfi_supported = rfi != 0;
         self.custom_vbv = custom_vbv != 0;
         tracing::info!(
@@ -313,9 +335,31 @@ impl NvencD3d11Encoder {
         cfg.encodeCodecConfig.hevcConfig.tier = 1;
         cfg.encodeCodecConfig.hevcConfig.level = 0;
 
-        // 10-bit HEVC Main10 (HDR foundation): NVENC upconverts the 8-bit input; 8-bit leaves the
-        // preset default (Main) untouched.
-        if self.bit_depth == 10 {
+        // Chroma + bit depth. Full-chroma 4:4:4 (HEVC Range Extensions) takes precedence and composes
+        // with 10-bit (Main 4:4:4 10): NVENC ingests the RGB input (ARGB / ABGR10) and CSCs it to
+        // YUV444 internally when `chromaFormatIDC = 3` under the FREXT profile. Only valid on an RGB
+        // input — a subsampled NV12/P010 source can't reconstruct full chroma (so the capturer is
+        // forced to RGB for a 4:4:4 session, and we guard on the input format here too).
+        //
+        // ON-GLASS TODO (RTX box): confirm ARGB + chromaFormatIDC=3 + FREXT yields a *true* 4:4:4
+        // stream. NVENC's RGB→YUV CSC is documented to honor chromaFormatIDC (unlike libavcodec's
+        // wrapper, which always subsamples RGB to 4:2:0 — hence the Linux path feeds planar YUV444
+        // instead). If on-glass shows 4:2:0, the follow-up is a BGRA→AYUV shader feeding the native
+        // `NV_ENC_BUFFER_FORMAT_AYUV` 4:4:4 input format.
+        let rgb_input = matches!(
+            self.buffer_fmt,
+            nv::NV_ENC_BUFFER_FORMAT::NV_ENC_BUFFER_FORMAT_ARGB
+                | nv::NV_ENC_BUFFER_FORMAT::NV_ENC_BUFFER_FORMAT_ABGR10
+        );
+        if self.chroma_444 && rgb_input {
+            cfg.profileGUID = nv::NV_ENC_HEVC_PROFILE_FREXT_GUID;
+            cfg.encodeCodecConfig.hevcConfig.set_chromaFormatIDC(3);
+            if self.bit_depth == 10 {
+                cfg.encodeCodecConfig.hevcConfig.set_pixelBitDepthMinus8(2); // Main 4:4:4 10
+            }
+        } else if self.bit_depth == 10 {
+            // 10-bit HEVC Main10 (HDR foundation): NVENC upconverts the 8-bit input; 8-bit leaves the
+            // preset default (Main) untouched.
             cfg.profileGUID = nv::NV_ENC_HEVC_PROFILE_MAIN10_GUID;
             cfg.encodeCodecConfig.hevcConfig.set_pixelBitDepthMinus8(2); // 10 - 8
         }
@@ -787,6 +831,9 @@ impl Encoder for NvencD3d11Encoder {
         EncoderCaps {
             supports_rfi: self.rfi_supported,
             supports_hdr_metadata: self.hdr,
+            // Reflects what the session actually configured (cleared in `query_caps` if the GPU lacks
+            // YUV444 encode), so the glue can confirm 4:4:4 vs the negotiated request.
+            chroma_444: self.chroma_444,
         }
     }
 
@@ -904,3 +951,69 @@ impl Drop for NvencD3d11Encoder {
         unsafe { self.teardown() };
     }
 }
+
+/// Probe whether the active NVIDIA GPU can encode HEVC **4:4:4** (`NV_ENC_CAPS_SUPPORT_YUV444_ENCODE`).
+/// Creates a throwaway hardware D3D11 device + NVENC session, queries the cap, and tears down. HEVC-only;
+/// the result is cached by the caller ([`crate::encode::can_encode_444`]) and read *before* the Welcome
+/// so the host advertises the chroma it can really encode (honest downgrade to 4:2:0 on a card without it).
+pub fn probe_can_encode_444(codec: Codec) -> bool {
+    use windows::Win32::Foundation::HMODULE;
+    use windows::Win32::Graphics::Direct3D::{D3D_DRIVER_TYPE_HARDWARE, D3D_FEATURE_LEVEL_11_0};
+    use windows::Win32::Graphics::Direct3D11::{
+        D3D11CreateDevice, D3D11_CREATE_DEVICE_BGRA_SUPPORT, D3D11_SDK_VERSION,
+    };
+    if codec != Codec::H265 {
+        return false;
+    }
+    // SAFETY: a self-contained probe owning every handle it creates. `D3D11CreateDevice` (HARDWARE
+    // driver, NULL adapter) fills `device` or returns Err (→ false). `open_encode_session_ex` opens an
+    // NVENC session against that device's raw pointer (valid while `device` is held) or errors (→ false,
+    // tearing nothing down). `get_encode_caps` reads one scalar cap into `val` via the loaded API table.
+    // `destroy_encoder` frees the session exactly once; `device`/its context drop with the COM wrappers.
+    // No handle escapes this call and nothing runs concurrently.
+    unsafe {
+        let mut device: Option<ID3D11Device> = None;
+        if D3D11CreateDevice(
+            None,
+            D3D_DRIVER_TYPE_HARDWARE,
+            HMODULE::default(),
+            D3D11_CREATE_DEVICE_BGRA_SUPPORT,
+            Some(&[D3D_FEATURE_LEVEL_11_0]),
+            D3D11_SDK_VERSION,
+            Some(&mut device),
+            None,
+            None,
+        )
+        .is_err()
+        {
+            return false;
+        }
+        let Some(device) = device else { return false };
+        let mut params = nv::NV_ENC_OPEN_ENCODE_SESSION_EX_PARAMS {
+            version: nv::NV_ENC_OPEN_ENCODE_SESSION_EX_PARAMS_VER,
+            deviceType: nv::NV_ENC_DEVICE_TYPE::NV_ENC_DEVICE_TYPE_DIRECTX,
+            device: device.as_raw(),
+            apiVersion: nv::NVENCAPI_VERSION,
+            ..Default::default()
+        };
+        let mut enc: *mut c_void = ptr::null_mut();
+        if (API.open_encode_session_ex)(&mut params, &mut enc)
+            .result_without_string()
+            .is_err()
+        {
+            return false;
+        }
+        let mut param = nv::NV_ENC_CAPS_PARAM {
+            version: nv::NV_ENC_CAPS_PARAM_VER,
+            capsToQuery: nv::NV_ENC_CAPS::NV_ENC_CAPS_SUPPORT_YUV444_ENCODE,
+            reserved: [0; 62],
+        };
+        let mut val: i32 = 0;
+        let ok = (API.get_encode_caps)(enc, nv::NV_ENC_CODEC_HEVC_GUID, &mut param, &mut val)
+            .result_without_string()
+            .is_ok()
+            && val != 0;
+        let _ = (API.destroy_encoder)(enc);
+        ok
+    }
+}

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

@@ -41,8 +41,6 @@ type Aes128CbcEnc = cbc::Encryptor<aes::Aes128>;
 /// `RTP_PAYLOAD_TYPE_FEC  127`).
 const AUDIO_PACKET_TYPE: u8 = 97;
 const AUDIO_FEC_PACKET_TYPE: u8 = 127;
-/// Stereo Opus bitrate (unchanged from the live-validated stereo path).
-const OPUS_BITRATE: i32 = 128_000;
 
 /// Audio FEC geometry (moonlight-common-c `RtpAudioQueue.h`: `RTPA_DATA_SHARDS 4`,
 /// `RTPA_FEC_SHARDS 2`). Blocks are aligned: the client synthesizes the block base as
@@ -82,67 +80,20 @@ impl Default for AudioParams {
     }
 }
 
-/// One Opus (multi)stream layout. Channel order is the GameStream/Moonlight order
-/// FL FR FC LFE RL RR [SL SR]; `mapping` is the libopus multistream mapping we *encode*
-/// with — identical to Sunshine's `audio.cpp stream_configs` (verified verbatim 2026-06-10):
-/// identity mapping, so normal quality couples (FL,FR) and (FC,LFE) [+ (RL,RR) on 7.1] with
-/// the remaining channels as mono streams; high quality is one mono stream per channel.
-/// Bitrates are Sunshine's per-config values (stereo keeps punktfunk's existing 128 kbps).
-pub struct OpusLayout {
-    pub channels: u8,
-    pub streams: u8,
-    pub coupled: u8,
-    pub mapping: &'static [u8],
-    pub bitrate: i32,
-}
-
-pub const LAYOUT_STEREO: OpusLayout = OpusLayout {
-    channels: 2,
-    streams: 1,
-    coupled: 1,
-    mapping: &[0, 1],
-    bitrate: OPUS_BITRATE,
-};
-pub const LAYOUT_51: OpusLayout = OpusLayout {
-    channels: 6,
-    streams: 4,
-    coupled: 2,
-    mapping: &[0, 1, 2, 3, 4, 5],
-    bitrate: 256_000,
-};
-pub const LAYOUT_51_HQ: OpusLayout = OpusLayout {
-    channels: 6,
-    streams: 6,
-    coupled: 0,
-    mapping: &[0, 1, 2, 3, 4, 5],
-    bitrate: 1_536_000,
-};
-pub const LAYOUT_71: OpusLayout = OpusLayout {
-    channels: 8,
-    streams: 5,
-    coupled: 3,
-    mapping: &[0, 1, 2, 3, 4, 5, 6, 7],
-    bitrate: 450_000,
-};
-pub const LAYOUT_71_HQ: OpusLayout = OpusLayout {
-    channels: 8,
-    streams: 8,
-    coupled: 0,
-    mapping: &[0, 1, 2, 3, 4, 5, 6, 7],
-    bitrate: 2_048_000,
+// The Opus surround layout table (channel order FL FR FC LFE RL RR [SL SR], identity mapping,
+// Sunshine's per-config bitrates) now lives in `punktfunk_core::audio`, shared with the native
+// `punktfunk/1` path and every client decoder. Re-export the pieces the GameStream module + its
+// RTSP SDP (`rtsp.rs`) reference; the GFE-specific `surround_params` SDP rotation stays below.
+pub use punktfunk_core::audio::{
+    OpusLayout, LAYOUT_51, LAYOUT_51_HQ, LAYOUT_71, LAYOUT_71_HQ, LAYOUT_STEREO,
 };
 
-/// Pick the encoder layout for the negotiated session parameters. Unknown channel counts
-/// fall back to stereo (the client can only request 2/6/8 — `AUDIO_CONFIGURATION_*` in
+/// Pick the encoder layout for the negotiated session parameters. Thin wrapper over the shared
+/// [`punktfunk_core::audio::layout_for`] keyed on this module's [`AudioParams`] (unknown channel
+/// counts fall back to stereo; the client can only request 2/6/8 — `AUDIO_CONFIGURATION_*` in
 /// Limelight.h).
 pub fn layout_for(params: &AudioParams) -> &'static OpusLayout {
-    match (params.channels, params.high_quality) {
-        (6, false) => &LAYOUT_51,
-        (6, true) => &LAYOUT_51_HQ,
-        (8, false) => &LAYOUT_71,
-        (8, true) => &LAYOUT_71_HQ,
-        _ => &LAYOUT_STEREO,
-    }
+    punktfunk_core::audio::layout_for(params.channels, params.high_quality)
 }
 
 /// The `a=fmtp:97 surround-params=` digit string for a layout: channelCount, streams,
@@ -345,21 +296,21 @@ fn run(
 }
 
 /// Opus encoder for one session: the plain stereo encoder (the live-validated path, byte
-/// identical) or a libopus multistream encoder for 5.1/7.1.
+/// identical) or the safe `opus::MSEncoder` multistream encoder for 5.1/7.1. Both are
+/// cross-platform (Linux + Windows) — surround no longer needs `audiopus_sys`.
 #[cfg(any(target_os = "linux", target_os = "windows"))]
 enum SessionEncoder {
     Stereo(opus::Encoder),
-    // Surround needs the libopus *multistream* encoder via `audiopus_sys` (Linux-only dep).
-    #[cfg(target_os = "linux")]
-    Surround(MsEncoder),
+    Surround(opus::MSEncoder),
 }
 
 #[cfg(any(target_os = "linux", target_os = "windows"))]
 impl SessionEncoder {
     fn new(layout: &'static OpusLayout) -> Result<SessionEncoder> {
+        // RESTRICTED_LOWDELAY (`opus::Application::LowDelay`) + hard CBR, matching Sunshine — CBR
+        // keeps the Opus packet size constant, which the GameStream audio FEC (equal-length shards)
+        // relies on, and the client asserts a constant per-stream TOC.
         if layout.channels == 2 {
-            // RESTRICTED_LOWDELAY + CBR, matching Sunshine — CBR keeps the Opus TOC byte
-            // constant, which the client asserts per stream.
             let mut enc = opus::Encoder::new(
                 SAMPLE_RATE,
                 opus::Channels::Stereo,
@@ -370,138 +321,32 @@ impl SessionEncoder {
             enc.set_vbr(false).ok();
             Ok(SessionEncoder::Stereo(enc))
         } else {
-            #[cfg(target_os = "linux")]
-            {
-                Ok(SessionEncoder::Surround(MsEncoder::new(layout)?))
-            }
-            #[cfg(not(target_os = "linux"))]
-            {
-                anyhow::bail!(
-                    "surround audio ({} ch) needs the libopus multistream encoder (Linux only) — \
-                     use a stereo session",
-                    layout.channels
-                )
-            }
+            let mut enc = opus::MSEncoder::new(
+                SAMPLE_RATE,
+                layout.streams,
+                layout.coupled,
+                layout.mapping,
+                opus::Application::LowDelay,
+            )
+            .map_err(|e| anyhow::anyhow!("create Opus multistream encoder: {e}"))?;
+            enc.set_bitrate(opus::Bitrate::Bits(layout.bitrate)).ok();
+            enc.set_vbr(false).ok();
+            Ok(SessionEncoder::Surround(enc))
         }
     }
 
-    /// Encode one interleaved frame (`samples_per_channel * channels` f32s) into `out`,
-    /// returning the packet length.
-    fn encode_float(
-        &mut self,
-        frame: &[f32],
-        samples_per_channel: usize,
-        out: &mut [u8],
-    ) -> Result<usize> {
-        // `samples_per_channel` only feeds the multistream (surround) encoder; stereo infers it.
-        #[cfg(not(target_os = "linux"))]
-        let _ = samples_per_channel;
+    /// Encode one interleaved frame into `out`, returning the packet length. Both encoders infer
+    /// the per-channel sample count from `frame.len()` and their channel count.
+    fn encode_float(&mut self, frame: &[f32], out: &mut [u8]) -> Result<usize> {
         match self {
             SessionEncoder::Stereo(enc) => enc.encode_float(frame, out).context("opus encode"),
-            #[cfg(target_os = "linux")]
-            SessionEncoder::Surround(enc) => enc.encode_float(frame, samples_per_channel, out),
+            SessionEncoder::Surround(enc) => enc
+                .encode_float(frame, out)
+                .context("opus multistream encode"),
         }
     }
 }
 
-/// RAII wrapper for `OpusMSEncoder` (the safe `opus` crate is stereo-only; the multistream
-/// API comes from `audiopus_sys`, the same libopus the crate already links). Configured like
-/// the stereo path: RESTRICTED_LOWDELAY, hard CBR, per-layout bitrate.
-#[cfg(target_os = "linux")]
-struct MsEncoder {
-    st: std::ptr::NonNull<audiopus_sys::OpusMSEncoder>,
-}
-
-// SAFETY: `MsEncoder` owns a unique `OpusMSEncoder` via `NonNull` (it is neither `Clone` nor
-// `Sync`, so the pointer is never aliased). libopus's multistream encoder state is a self-contained
-// heap allocation with no thread-local or thread-affine state, so moving ownership to another thread
-// is sound; every method takes `&mut self`, keeping access single-threaded at any instant.
-#[cfg(target_os = "linux")]
-unsafe impl Send for MsEncoder {}
-
-#[cfg(target_os = "linux")]
-impl MsEncoder {
-    fn new(layout: &OpusLayout) -> Result<MsEncoder> {
-        use std::os::raw::c_int;
-        let mut err: c_int = 0;
-        // SAFETY: every scalar arg is a valid libopus input (sample rate, channel/stream/coupled
-        // counts, the RESTRICTED_LOWDELAY application constant). `layout.mapping.as_ptr()` addresses
-        // a 'static slice of exactly `layout.channels` bytes (every `OpusLayout` constant upholds
-        // that), which is the element count `opus_multistream_encoder_create` reads through it, and
-        // `&mut err` is a live local the call writes its status into. libopus copies the mapping into
-        // its own allocation, so the pointer need only be valid for the call; the returned pointer is
-        // null/`OPUS_OK`-checked below before any use.
-        let st = unsafe {
-            audiopus_sys::opus_multistream_encoder_create(
-                SAMPLE_RATE as i32,
-                layout.channels as c_int,
-                layout.streams as c_int,
-                layout.coupled as c_int,
-                layout.mapping.as_ptr(),
-                audiopus_sys::OPUS_APPLICATION_RESTRICTED_LOWDELAY,
-                &mut err,
-            )
-        };
-        let st = std::ptr::NonNull::new(st)
-            .filter(|_| err == audiopus_sys::OPUS_OK)
-            .ok_or_else(|| anyhow::anyhow!("opus_multistream_encoder_create failed ({err})"))?;
-        // SAFETY: `st` is the non-null encoder `opus_multistream_encoder_create` just returned, owned
-        // exclusively here. Each `opus_multistream_encoder_ctl` call passes a valid request constant
-        // with the single by-value `c_int` argument that request's variadic ABI expects
-        // (`OPUS_SET_BITRATE_REQUEST` → bitrate, `OPUS_SET_VBR_REQUEST` → 0). No pointer escapes the
-        // call and the encoder outlives it.
-        unsafe {
-            audiopus_sys::opus_multistream_encoder_ctl(
-                st.as_ptr(),
-                audiopus_sys::OPUS_SET_BITRATE_REQUEST,
-                layout.bitrate as c_int,
-            );
-            audiopus_sys::opus_multistream_encoder_ctl(
-                st.as_ptr(),
-                audiopus_sys::OPUS_SET_VBR_REQUEST,
-                0 as c_int, // hard CBR (constant packet size — also what audio FEC relies on)
-            );
-        }
-        Ok(MsEncoder { st })
-    }
-
-    fn encode_float(
-        &mut self,
-        frame: &[f32],
-        samples_per_channel: usize,
-        out: &mut [u8],
-    ) -> Result<usize> {
-        // SAFETY: `self.st` is the live encoder from `new`. libopus reads `samples_per_channel *
-        // channels` f32s through `frame.as_ptr()`; every caller passes a `frame` of exactly that
-        // length together with the matching `samples_per_channel` (`audio_body`'s `frame_len =
-        // samples_per_channel * layout.channels`; the round-trip tests size identically), so the read
-        // stays in bounds. `out.as_mut_ptr()` is written for at most `out.len()` bytes, which is
-        // passed as the capacity bound. Both buffers are live locals outliving this synchronous call;
-        // the return value is range-checked before being used as a length.
-        let n = unsafe {
-            audiopus_sys::opus_multistream_encode_float(
-                self.st.as_ptr(),
-                frame.as_ptr(),
-                samples_per_channel as std::os::raw::c_int,
-                out.as_mut_ptr(),
-                out.len() as i32,
-            )
-        };
-        anyhow::ensure!(n > 0, "opus_multistream_encode_float failed ({n})");
-        Ok(n as usize)
-    }
-}
-
-#[cfg(target_os = "linux")]
-impl Drop for MsEncoder {
-    fn drop(&mut self) {
-        // SAFETY: `self.st` is the encoder `opus_multistream_encoder_create` returned; this
-        // `MsEncoder` owns it uniquely and `drop` runs exactly once, so the destroy frees it once
-        // with no subsequent use.
-        unsafe { audiopus_sys::opus_multistream_encoder_destroy(self.st.as_ptr()) }
-    }
-}
-
 #[cfg(any(target_os = "linux", target_os = "windows"))]
 fn audio_body(
     cap: &mut dyn AudioCapturer,
@@ -565,7 +410,7 @@ fn audio_body(
                     *s = (*s * gain).clamp(-1.0, 1.0);
                 }
             }
-            let n = enc.encode_float(&frame, samples_per_channel, &mut out)?;
+            let n = enc.encode_float(&frame, &mut out)?;
             // AES-128-CBC the Opus payload (RTP header stays plaintext). Per-packet IV =
             // BE32(rikeyid + seq) in [0..4], zero elsewhere; PKCS7 padding.
             let iv_seq = (rikeyid as u32).wrapping_add(seq as u32);
@@ -775,41 +620,33 @@ mod tests {
     /// Real-codec proof of the 5.1 mapping math: encode with our encoder layout, decode with
     /// the mapping a stock Moonlight client derives from our advertised surround-params
     /// (parse → GFE swap), and verify a tone fed into each input channel comes out on the
-    /// same output channel.
-    #[cfg(target_os = "linux")]
+    /// same output channel. Cross-platform via the safe `opus` crate — this also guards the
+    /// (now un-gated) Windows GameStream surround build.
     #[test]
     fn multistream_51_roundtrip_channel_identity() {
         let layout = &LAYOUT_51;
         let samples = 240; // 5 ms
         let ch = layout.channels as usize;
 
-        // Client-side decoder mapping derived exactly as moonlight-common-c does.
+        // Client-side decoder mapping derived exactly as moonlight-common-c does (GFE swap).
         let s = surround_params(layout, false);
         let digits: Vec<u8> = s.bytes().map(|b| b - b'0').collect();
         let client_mapping = client_swap(&digits[3..]);
 
-        let mut err = 0i32;
-        // SAFETY: scalar args are valid libopus inputs. `client_mapping.as_ptr()` addresses a
-        // `Vec<u8>` of exactly `ch` entries (derived from the advertised surround-params), which is
-        // the element count the decoder reads through it, and `&mut err` is a live local the call
-        // writes. The returned pointer is `OPUS_OK`/non-null-checked immediately below before use.
-        let dec = unsafe {
-            audiopus_sys::opus_multistream_decoder_create(
-                SAMPLE_RATE as i32,
-                ch as i32,
-                layout.streams as i32,
-                layout.coupled as i32,
-                client_mapping.as_ptr(),
-                &mut err,
-            )
-        };
-        assert_eq!(err, audiopus_sys::OPUS_OK);
-        assert!(!dec.is_null());
+        let mut dec =
+            opus::MSDecoder::new(SAMPLE_RATE, layout.streams, layout.coupled, &client_mapping)
+                .expect("multistream decoder");
 
         for tone_ch in 0..ch {
-            let mut enc = MsEncoder::new(layout).unwrap();
+            let mut enc = opus::MSEncoder::new(
+                SAMPLE_RATE,
+                layout.streams,
+                layout.coupled,
+                layout.mapping,
+                opus::Application::LowDelay,
+            )
+            .expect("multistream encoder");
             let mut out = vec![0u8; 1400];
-            let mut decoded = vec![0f32; samples * ch];
             let mut energy = vec![0f64; ch];
             // A few frames so the codec converges past its startup transient.
             for f in 0..8 {
@@ -819,28 +656,15 @@ mod tests {
                         / SAMPLE_RATE as f32;
                     frame[t * ch + tone_ch] = 0.5 * phase.sin();
                 }
-                let n = enc.encode_float(&frame, samples, &mut out).unwrap();
+                let n = enc.encode_float(&frame, &mut out).unwrap();
                 assert!(n > 0);
-                // SAFETY: `dec` is the non-null decoder asserted above. `out.as_ptr()` is read for
-                // the `n` encoded bytes just produced by `encode_float`; `decoded.as_mut_ptr()` is
-                // written for up to `samples * ch` f32s and `decoded` is exactly that long; `samples`
-                // is the per-channel frame size. All buffers are live locals outliving the call; the
-                // return is checked to equal `samples`.
-                let got = unsafe {
-                    audiopus_sys::opus_multistream_decode_float(
-                        dec,
-                        out.as_ptr(),
-                        n as i32,
-                        decoded.as_mut_ptr(),
-                        samples as i32,
-                        0,
-                    )
-                };
-                assert_eq!(got as usize, samples);
+                let mut decoded = vec![0f32; samples * ch];
+                let got = dec.decode_float(&out[..n], &mut decoded, false).unwrap();
+                assert_eq!(got, samples);
                 if f >= 4 {
                     for t in 0..samples {
-                        for c in 0..ch {
-                            energy[c] += (decoded[t * ch + c] as f64).powi(2);
+                        for (c, e) in energy.iter_mut().enumerate() {
+                            *e += (decoded[t * ch + c] as f64).powi(2);
                         }
                     }
                 }
@@ -854,9 +678,6 @@ mod tests {
                  (energies: {energy:?})"
             );
         }
-        // SAFETY: `dec` is the decoder `opus_multistream_decoder_create` returned; the test owns it
-        // and destroys it exactly once here, after the final decode — no later use, no double free.
-        unsafe { audiopus_sys::opus_multistream_decoder_destroy(dec) };
     }
 
     /// Live 5.1 capture → multistream encode → decode, against a real PipeWire session.
@@ -869,7 +690,15 @@ mod tests {
     fn surround_capture_live() {
         let mut cap = crate::audio::open_audio_capture(6).expect("open 6ch capture");
         let layout = &LAYOUT_51;
-        let mut enc = MsEncoder::new(layout).unwrap();
+        let mut enc = opus::MSEncoder::new(
+            SAMPLE_RATE,
+            layout.streams,
+            layout.coupled,
+            layout.mapping,
+            opus::Application::LowDelay,
+        )
+        .unwrap();
+        enc.set_vbr(false).ok(); // hard CBR so packet sizes are constant (audio FEC relies on it)
         let mut out = vec![0u8; 1400];
         let mut acc: Vec<f32> = Vec::new();
         let frame_len = 240 * 6;
@@ -880,49 +709,24 @@ mod tests {
             acc.extend_from_slice(&chunk);
             while acc.len() >= frame_len && packets < 100 {
                 let frame: Vec<f32> = acc.drain(..frame_len).collect();
-                let n = enc.encode_float(&frame, 240, &mut out).unwrap();
+                let n = enc.encode_float(&frame, &mut out).unwrap();
                 sizes.insert(n);
                 packets += 1;
             }
         }
         // Hard CBR: every multistream packet must be the same size (audio FEC relies on it).
         assert_eq!(sizes.len(), 1, "CBR sizes: {sizes:?}");
-        // And a stock client's decoder must accept them.
+        // And a stock client's GFE-derived decoder must accept them.
         let s = surround_params(layout, false);
         let digits: Vec<u8> = s.bytes().map(|b| b - b'0').collect();
         let client_mapping = client_swap(&digits[3..]);
-        let mut err = 0i32;
-        // SAFETY: scalar args are valid; `client_mapping.as_ptr()` addresses a 6-entry `Vec<u8>`
-        // (matches the 6-channel layout the decoder reads through it), alive past the call, and
-        // `&mut err` is a live local. The pointer is `OPUS_OK`-checked before use.
-        let dec = unsafe {
-            audiopus_sys::opus_multistream_decoder_create(
-                48000,
-                6,
-                layout.streams as i32,
-                layout.coupled as i32,
-                client_mapping.as_ptr(),
-                &mut err,
-            )
-        };
-        assert_eq!(err, audiopus_sys::OPUS_OK);
+        let mut dec =
+            opus::MSDecoder::new(SAMPLE_RATE, layout.streams, layout.coupled, &client_mapping)
+                .unwrap();
         let mut pcm = vec![0f32; 240 * 6];
-        // SAFETY: `dec` is the non-null decoder from create. `out.as_ptr()` is read for the CBR
-        // packet length passed in (`*sizes.first()`, a real encoded packet size in `out`);
-        // `pcm.as_mut_ptr()` is written for up to `240 * 6` f32s and `pcm` is exactly that long;
-        // `240` is the per-channel frame size. All buffers are live locals outliving the call.
-        let got = unsafe {
-            audiopus_sys::opus_multistream_decode_float(
-                dec,
-                out.as_ptr(),
-                *sizes.first().unwrap() as i32,
-                pcm.as_mut_ptr(),
-                240,
-                0,
-            )
-        };
-        // SAFETY: `dec` is owned by the test; destroyed exactly once here after the final decode.
-        unsafe { audiopus_sys::opus_multistream_decoder_destroy(dec) };
+        let got = dec
+            .decode_float(&out[..*sizes.first().unwrap()], &mut pcm, false)
+            .unwrap();
         assert_eq!(got, 240);
     }
 }

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

@@ -431,6 +431,9 @@ fn stream_body(
         cfg.bitrate_kbps as u64 * 1000,
         frame.is_cuda(),
         8, // GameStream/Moonlight path: 8-bit (its own codec negotiation)
+        // GameStream/Moonlight stays 4:2:0 — stock Moonlight clients can't decode 4:4:4, and the
+        // protocol has no chroma negotiation. 4:4:4 is punktfunk/1-native only.
+        encode::ChromaFormat::Yuv420,
     )
     .context("open video encoder for stream")?;
     // FEC overhead percent (Sunshine default 20). Override with PUNKTFUNK_FEC_PCT (0 = data-only).
@@ -560,6 +563,7 @@ fn stream_body(
                     cfg.bitrate_kbps as u64 * 1000,
                     frame.is_cuda(),
                     8,
+                    encode::ChromaFormat::Yuv420, // GameStream stays 4:2:0
                 )
                 .context("reopen encoder after rebuild")?;
                 supports_rfi = enc.caps().supports_rfi;

crates/punktfunk-host/src/punktfunk1.rs

@@ -355,6 +355,15 @@ fn resolve_bitrate_kbps(requested: u32) -> u32 {
     }
 }
 
+/// Resolve the audio channel count the session will capture + encode from the client's request.
+/// Normalizes to one of 2 (stereo) / 6 (5.1) / 8 (7.1); anything else (older client, garbage)
+/// becomes stereo. Both backends can produce the requested count (PipeWire pads/upmixes positions,
+/// WASAPI loopback up/downmixes via AUTOCONVERTPCM), so no capability clamp is needed here — the
+/// surround channels just carry up/downmixed content when the host's sink has fewer real channels.
+fn resolve_audio_channels(requested: u8) -> u8 {
+    punktfunk_core::audio::normalize_channels(requested)
+}
+
 /// Static FEC override: `PUNKTFUNK_FEC_PCT`, when set, PINS the recovery percent and DISABLES
 /// adaptive FEC — so a speed test / measurement keeps a fixed, known overhead. `None` ⇒ adaptive
 /// FEC (the host sizes recovery to the loss the client reports). `0` disables FEC entirely.
@@ -623,6 +632,17 @@ async fn serve_session(
             "encoder bitrate"
         );
 
+        // Resolve the audio channel count (client request → stereo / 5.1 / 7.1). The capturer opens
+        // at this count: PipeWire synthesizes the requested positions (padding with silence when the
+        // sink has fewer), WASAPI loopback up/downmixes via AUTOCONVERTPCM — so a client always gets
+        // the channels it asked for, and the Welcome echoes the value the audio thread will encode.
+        let audio_channels = resolve_audio_channels(hello.audio_channels);
+        tracing::info!(
+            requested = hello.audio_channels,
+            resolved = audio_channels,
+            "audio channels"
+        );
+
         // Resolve the encode bit depth: HEVC Main10 only when the client advertised it AND the host
         // opted in (PUNKTFUNK_10BIT). A client that can't decode 10-bit (caps bit clear, or an older
         // client) always gets the 8-bit stream. PUNKTFUNK_10BIT is the host policy gate until a
@@ -642,6 +662,44 @@ async fn serve_session(
             "encode bit depth"
         );
 
+        // Resolve the chroma subsampling: full-chroma HEVC 4:4:4 only when ALL of — the host opted in
+        // (PUNKTFUNK_444), the client advertised VIDEO_CAP_444, the session is single-process (the
+        // two-process WGC relay encodes 4:2:0 in v1), and the active GPU/driver actually supports a
+        // 4:4:4 encode (probed, cached). The native path always encodes HEVC. We resolve this BEFORE
+        // the Welcome so `chroma_format` reflects what we'll really emit — the honest-downgrade
+        // channel: if any gate fails the client is told 4:2:0 before it builds its decoder. The probe
+        // opens a tiny encoder; it runs only when both opt-ins are set and is cached after the first.
+        let host_wants_444 = crate::config::config().four_four_four;
+        let client_supports_444 = hello.video_caps & punktfunk_core::quic::VIDEO_CAP_444 != 0;
+        let single_process = crate::session_plan::resolve_topology()
+            == crate::session_plan::SessionTopology::SingleProcess;
+        // The GPU probe opens a real (tiny) encoder on first use, so run it off the reactor like the
+        // compositor probe above (blocking probes → spawn_blocking). Short-circuit so it only runs when
+        // the cheap gates already pass. The result is cached process-wide (a negative latches until
+        // restart — acceptable: a GPU either supports HEVC 4:4:4 or it doesn't, and a transient open
+        // failure here is rare since the session's own encoder isn't open yet).
+        let gpu_supports_444 = if host_wants_444 && client_supports_444 && single_process {
+            tokio::task::spawn_blocking(|| {
+                crate::encode::can_encode_444(crate::encode::Codec::H265)
+            })
+            .await
+            .context("4:4:4 capability probe task")?
+        } else {
+            false
+        };
+        let chroma = if gpu_supports_444 {
+            crate::encode::ChromaFormat::Yuv444
+        } else {
+            crate::encode::ChromaFormat::Yuv420
+        };
+        tracing::info!(
+            chroma = ?chroma,
+            host_wants_444,
+            client_supports_444,
+            single_process,
+            "encode chroma"
+        );
+
         // Reserve a UDP port for the data plane (bind, read it back, rebind in UdpTransport).
         let probe = std::net::UdpSocket::bind("0.0.0.0:0")?;
         let udp_port = probe.local_addr()?.port();
@@ -691,6 +749,12 @@ async fn serve_session(
             } else {
                 ColorInfo::SDR_BT709
             },
+            // The chroma the encoder will actually emit (resolved + GPU-probed above) — 4:4:4 only
+            // when every gate passed, else 4:2:0. The client sizes its decoder from this.
+            chroma_format: chroma.idc(),
+            // The resolved audio channel count the audio thread will capture + Opus-(multi)stream
+            // encode (2/6/8). The client builds its decoder from this echoed value.
+            audio_channels,
         };
         io::write_msg(&mut send, &welcome.encode()).await?;
 
@@ -884,9 +948,10 @@ async fn serve_session(
         let conn = conn.clone();
         let stop = stop.clone();
         let cap = audio_cap.clone();
+        let channels = welcome.audio_channels;
         std::thread::Builder::new()
             .name("punktfunk1-audio".into())
-            .spawn(move || audio_thread(conn, stop, cap))
+            .spawn(move || audio_thread(conn, stop, cap, channels))
             .map_err(|e| tracing::error!(error = %e, "audio thread spawn failed — session continues without audio"))
             .ok()
     } else {
@@ -946,6 +1011,13 @@ async fn serve_session(
     let launch_for_dp = hello.launch.clone();
     let bitrate_kbps = welcome.bitrate_kbps; // resolved encoder bitrate (Hello clamped, or default)
     let bit_depth = welcome.bit_depth; // resolved encode bit depth (8, or 10 when negotiated)
+                                       // Resolved chroma — derive the typed value back from the wire byte the Welcome carried (so the
+                                       // session uses exactly what the client was told). `Yuv444` only when the handshake gate passed.
+    let chroma = if welcome.chroma_format == punktfunk_core::quic::CHROMA_IDC_444 {
+        crate::encode::ChromaFormat::Yuv444
+    } else {
+        crate::encode::ChromaFormat::Yuv420
+    };
     let stop_stream = stop.clone();
     let fec_target_dp = fec_target.clone(); // data-plane handle to the adaptive-FEC target
     let conn_stream = conn.clone(); // for sending the source's real HDR metadata (0xCE) mid-stream
@@ -1005,6 +1077,7 @@ async fn serve_session(
                         compositor,
                         bitrate_kbps,
                         bit_depth,
+                        chroma,
                         probe_rx,
                         probe_result_tx,
                         fec_target: fec_target_dp,
@@ -1493,33 +1566,88 @@ fn input_thread(
     }
 }
 
-/// The audio thread: desktop capture → Opus (48 kHz stereo, 5 ms, CBR — same tuning as the
-/// GameStream path) → `AUDIO_MAGIC` datagrams. QUIC already encrypts; no extra layer.
-/// The capturer comes from (and returns to) the persistent slot — see [`AudioCapSlot`].
+/// Opus encoder for the native audio plane: a plain stereo encoder (the live-validated,
+/// byte-identical path) or a libopus *multistream* encoder for 5.1/7.1, both behind one
+/// `encode_float`. Surround uses the safe `opus::MSEncoder` (no `audiopus_sys`).
 #[cfg(any(target_os = "linux", target_os = "windows"))]
-fn audio_thread(conn: quinn::Connection, stop: Arc<AtomicBool>, audio_cap: AudioCapSlot) {
-    use crate::audio::{CHANNELS, SAMPLE_RATE};
+enum NativeAudioEnc {
+    Stereo(opus::Encoder),
+    Surround(opus::MSEncoder),
+}
+
+#[cfg(any(target_os = "linux", target_os = "windows"))]
+impl NativeAudioEnc {
+    /// Build the encoder for `channels` (2/6/8), hard-CBR + RESTRICTED_LOWDELAY like the
+    /// GameStream path; bitrate from the shared layout table (stereo keeps the validated 128 kbps).
+    fn new(channels: u8) -> Result<NativeAudioEnc, opus::Error> {
+        if channels == 2 {
+            let mut e = opus::Encoder::new(
+                crate::audio::SAMPLE_RATE,
+                opus::Channels::Stereo,
+                opus::Application::LowDelay,
+            )?;
+            e.set_bitrate(opus::Bitrate::Bits(128_000)).ok();
+            e.set_vbr(false).ok();
+            Ok(NativeAudioEnc::Stereo(e))
+        } else {
+            let l = punktfunk_core::audio::layout_for(channels, false);
+            let mut e = opus::MSEncoder::new(
+                crate::audio::SAMPLE_RATE,
+                l.streams,
+                l.coupled,
+                l.mapping,
+                opus::Application::LowDelay,
+            )?;
+            e.set_bitrate(opus::Bitrate::Bits(l.bitrate)).ok();
+            e.set_vbr(false).ok();
+            Ok(NativeAudioEnc::Surround(e))
+        }
+    }
+
+    fn encode_float(&mut self, frame: &[f32], out: &mut [u8]) -> Result<usize, opus::Error> {
+        match self {
+            NativeAudioEnc::Stereo(e) => e.encode_float(frame, out),
+            NativeAudioEnc::Surround(e) => e.encode_float(frame, out),
+        }
+    }
+}
+
+/// The audio thread: desktop capture → Opus (48 kHz, 5 ms, CBR — same tuning as the GameStream
+/// path) → `AUDIO_MAGIC` datagrams, at the negotiated `channels` (2 stereo / 6 = 5.1 / 8 = 7.1,
+/// canonical wire order FL FR FC LFE RL RR SL SR). QUIC already encrypts; no extra layer. The
+/// capturer comes from (and returns to) the persistent slot — see [`AudioCapSlot`].
+#[cfg(any(target_os = "linux", target_os = "windows"))]
+fn audio_thread(
+    conn: quinn::Connection,
+    stop: Arc<AtomicBool>,
+    audio_cap: AudioCapSlot,
+    channels: u8,
+) {
+    use crate::audio::SAMPLE_RATE;
     const FRAME_MS: usize = 5;
     const SAMPLES_PER_FRAME: usize = SAMPLE_RATE as usize * FRAME_MS / 1000; // 240
+    let want = punktfunk_core::audio::normalize_channels(channels);
 
+    // Reuse the cached capturer ONLY when its channel count matches this session's; a stereo
+    // capturer left by a prior session must not feed a 5.1/7.1 session (the encoder + the client's
+    // decoder are sized for `want`, so a mismatched capturer would garble/desync the audio).
     let capturer = match audio_cap.lock().unwrap().take() {
-        Some(mut c) => {
+        Some(mut c) if c.channels() == want as u32 => {
             c.drain(); // discard audio captured between sessions
             c
         }
-        None => match crate::audio::open_audio_capture(CHANNELS as u32) {
-            Ok(c) => c,
-            Err(e) => {
-                tracing::warn!(error = %format!("{e:#}"), "punktfunk/1 audio unavailable — session continues without it");
-                return;
+        prev => {
+            drop(prev); // wrong channel count (or none): clean teardown, open fresh at `want`
+            match crate::audio::open_audio_capture(want as u32) {
+                Ok(c) => c,
+                Err(e) => {
+                    tracing::warn!(error = %format!("{e:#}"), "punktfunk/1 audio unavailable — session continues without it");
+                    return;
+                }
             }
-        },
+        }
     };
-    let mut enc = match opus::Encoder::new(
-        SAMPLE_RATE,
-        opus::Channels::Stereo,
-        opus::Application::LowDelay,
-    ) {
+    let mut enc = match NativeAudioEnc::new(want) {
         Ok(e) => e,
         Err(e) => {
             tracing::error!(error = %e, "opus encoder");
@@ -1527,12 +1655,11 @@ fn audio_thread(conn: quinn::Connection, stop: Arc<AtomicBool>, audio_cap: Audio
             return;
         }
     };
-    enc.set_bitrate(opus::Bitrate::Bits(128_000)).ok();
-    enc.set_vbr(false).ok();
 
-    let frame_len = SAMPLES_PER_FRAME * CHANNELS;
+    let frame_len = SAMPLES_PER_FRAME * want as usize;
     let mut acc: Vec<f32> = Vec::with_capacity(frame_len * 4);
-    let mut opus_buf = vec![0u8; 1500];
+    // Sized for the largest surround frame (7.1 HQ ≈ 1.3 KB at 5 ms); ample for normal quality.
+    let mut opus_buf = vec![0u8; 4096];
     let mut seq: u32 = 0;
     // Reopen-with-backoff: hold the capturer in an Option so a mid-session capture-thread death
     // (device unplug, daemon restart) reopens instead of muting the rest of a multi-hour session.
@@ -1542,14 +1669,17 @@ fn audio_thread(conn: quinn::Connection, stop: Arc<AtomicBool>, audio_cap: Audio
     // restart). The first open already happened above; failing THAT still ends the session quietly.
     let mut capturer = Some(capturer);
     let mut last_failed: Option<std::time::Instant> = None;
-    tracing::info!("punktfunk/1 audio streaming (Opus 48 kHz stereo, 5 ms datagrams)");
+    tracing::info!(
+        channels = want,
+        "punktfunk/1 audio streaming (Opus 48 kHz, 5 ms datagrams)"
+    );
     'session: while !stop.load(Ordering::SeqCst) {
         if capturer.is_none() {
             if last_failed.is_some_and(|t| t.elapsed() < INJECTOR_REOPEN_BACKOFF) {
                 std::thread::sleep(std::time::Duration::from_millis(200));
                 continue;
             }
-            match crate::audio::open_audio_capture(CHANNELS as u32) {
+            match crate::audio::open_audio_capture(want as u32) {
                 Ok(c) => {
                     tracing::info!("punktfunk/1 audio capture reopened");
                     capturer = Some(c);
@@ -1599,7 +1729,12 @@ fn audio_thread(conn: quinn::Connection, stop: Arc<AtomicBool>, audio_cap: Audio
 /// Stub — punktfunk/1 audio needs Linux (PipeWire capture + libopus); non-Linux dev builds
 /// run sessions without it, same as when the capturer fails to open.
 #[cfg(not(any(target_os = "linux", target_os = "windows")))]
-fn audio_thread(_conn: quinn::Connection, _stop: Arc<AtomicBool>, _audio_cap: AudioCapSlot) {
+fn audio_thread(
+    _conn: quinn::Connection,
+    _stop: Arc<AtomicBool>,
+    _audio_cap: AudioCapSlot,
+    _channels: u8,
+) {
     tracing::warn!("punktfunk/1 audio requires Linux or Windows — session continues without it");
 }
 
@@ -2368,6 +2503,8 @@ struct SessionContext {
     bitrate_kbps: u32,
     /// Negotiated encode bit depth (8, or 10 = HEVC Main10).
     bit_depth: u8,
+    /// Negotiated chroma subsampling (4:2:0, or 4:4:4 when the client + host + GPU all support it).
+    chroma: crate::encode::ChromaFormat,
     /// Speed-test burst requests (see [`service_probes`]).
     probe_rx: std::sync::mpsc::Receiver<ProbeRequest>,
     /// Speed-test results back to the control task.
@@ -2398,7 +2535,7 @@ fn virtual_stream(ctx: SessionContext) -> Result<()> {
     // path now reads this typed `SessionPlan` instead of re-deriving from config at each dispatch site
     // (the latent "capture and encode disagree on the backend" hazard, plan §2.4). `bit_depth` is the
     // only per-session input — capture/topology/encoder are otherwise pure functions of `HostConfig`.
-    let plan = crate::session_plan::SessionPlan::resolve(ctx.bit_depth);
+    let plan = crate::session_plan::SessionPlan::resolve(ctx.bit_depth, ctx.chroma);
     tracing::info!(?plan, "resolved session plan");
     // Windows two-process secure-desktop path: when the host runs as SYSTEM (required for the secure
     // desktop + SendInput), WGC can't activate in-process, so we capture the normal desktop via a
@@ -2420,6 +2557,8 @@ fn virtual_stream(ctx: SessionContext) -> Result<()> {
         compositor,
         bitrate_kbps,
         bit_depth,
+        // The resolved chroma is already captured in `plan` (above); ignore the duplicate here.
+        chroma: _,
         probe_rx,
         probe_result_tx,
         fec_target,
@@ -2969,6 +3108,9 @@ fn virtual_stream_relay(ctx: SessionContext) -> Result<()> {
         compositor,
         bitrate_kbps,
         bit_depth,
+        // The two-process WGC relay encodes 4:2:0 in v1 — the handshake's `single_process` gate already
+        // forced `chroma` to Yuv420 for this topology, so the helper + secure-desktop DDA stay 4:2:0.
+        chroma: _,
         probe_rx,
         probe_result_tx,
         fec_target,
@@ -3079,6 +3221,7 @@ fn virtual_stream_relay(ctx: SessionContext) -> Result<()> {
                 // stage 5) so the DDA capturer doesn't re-derive it.
                 crate::capture::gpu_encode(),
                 hdr,
+                false, // the two-process relay path is 4:2:0 in v1
             )
             .context("open DDA for secure desktop")?;
             cap.set_active(true);
@@ -3092,6 +3235,8 @@ fn virtual_stream_relay(ctx: SessionContext) -> Result<()> {
                 bitrate_kbps as u64 * 1000,
                 frame.is_cuda(),
                 bit_depth,
+                // Secure-desktop DDA on the two-process relay path: 4:2:0 in v1 (matches the helper).
+                crate::encode::ChromaFormat::Yuv420,
             )
             .context("open video encoder for DDA")?;
             Ok(DdaPipe {
@@ -3491,6 +3636,9 @@ fn is_permanent_build_error(chain: &str) -> bool {
         "could not find output", // KWin < 6.5.6: createVirtualOutput unsupported
         "must be a node id",     // PUNKTFUNK_GAMESCOPE_NODE not an integer
         "is it installed",       // gamescope / kscreen-doctor not on PATH
+        // 4:4:4 NVENC got a CUDA frame — should never happen now the Linux capturer honors gpu=false,
+        // but fail fast instead of 8× retry (~90 s) rather than wedge the session if it ever recurs.
+        "capture/encoder negotiation mismatch",
     ];
     let lower = chain.to_ascii_lowercase();
     PERMANENT.iter().any(|p| lower.contains(p))
@@ -3540,8 +3688,20 @@ fn build_pipeline(
         bitrate_kbps as u64 * 1000,
         frame.is_cuda(),
         bit_depth,
+        plan.chroma,
     )
     .context("open video encoder")?;
+    // Post-open cross-check: the Welcome already committed `chroma_format` from the pre-open probe, so
+    // warn loudly if the encoder actually opened a different chroma than negotiated (the in-band SPS is
+    // authoritative for the decoder, but a mismatch means the probe and the live open disagreed).
+    let opened_444 = enc.caps().chroma_444;
+    if opened_444 != plan.chroma.is_444() {
+        tracing::warn!(
+            negotiated_444 = plan.chroma.is_444(),
+            opened_444,
+            "encoder chroma disagrees with the negotiated Welcome — the client was told the other value"
+        );
+    }
     let interval = std::time::Duration::from_secs_f64(1.0 / effective_hz.max(1) as f64);
     Ok((capturer, enc, frame, interval))
 }
@@ -3980,6 +4140,7 @@ mod tests {
                 GamepadPref::Auto,
                 0,
                 0,    // video_caps
+                2,    // audio_channels (stereo)
                 None, // launch
                 None,
                 Some((cert.clone(), key.clone())),
@@ -4012,6 +4173,7 @@ mod tests {
             GamepadPref::Auto,
             0,
             0,    // video_caps
+            2,    // audio_channels (stereo)
             None, // launch
             None,
             Some((cert, key)),
@@ -4065,6 +4227,7 @@ mod tests {
                 GamepadPref::Auto,
                 0,
                 0,    // video_caps
+                2,    // audio_channels (stereo)
                 None, // launch
                 None,
                 None,
@@ -4090,6 +4253,7 @@ mod tests {
             GamepadPref::Auto,
             0,
             0,    // video_caps
+            2,    // audio_channels (stereo)
             None, // launch
             Some(host_fp),
             Some((cert.clone(), key.clone())),

crates/punktfunk-host/src/session_plan.rs

@@ -106,17 +106,22 @@ pub struct SessionPlan {
     /// The IDD-push HDR hint (`bit_depth >= 10`) — the want-HDR flag the capturer was passed before.
     /// Non-IDD-push Windows backends ignore it and auto-detect HDR from the monitor; Linux is 8-bit.
     pub hdr: bool,
+    /// Handshake-negotiated chroma subsampling (4:2:0, or full-chroma 4:4:4 when the client + host +
+    /// GPU all support it). Resolved before the Welcome; `Yuv420` on every backend that declined it.
+    pub chroma: crate::encode::ChromaFormat,
 }
 
 impl SessionPlan {
-    /// Resolve the whole plan once from [`config`](crate::config) + the negotiated `bit_depth`.
-    pub fn resolve(bit_depth: u8) -> Self {
+    /// Resolve the whole plan once from [`config`](crate::config) + the negotiated `bit_depth` and
+    /// `chroma`.
+    pub fn resolve(bit_depth: u8, chroma: crate::encode::ChromaFormat) -> Self {
         SessionPlan {
             capture: CaptureBackend::resolve(),
             topology: resolve_topology(),
             encoder: resolve_encoder(),
             bit_depth,
             hdr: bit_depth >= 10,
+            chroma,
         }
     }
 
@@ -124,9 +129,24 @@ impl SessionPlan {
     /// (no second backend probe), `hdr` from the plan. Handed into `capture::capture_virtual_output` so the
     /// capturer never re-derives the encode backend.
     pub fn output_format(&self) -> crate::capture::OutputFormat {
+        let gpu = self.encoder.is_gpu();
+        // Linux NVENC 4:4:4: libavcodec `hevc_nvenc` only emits 4:4:4 from a YUV444 *input* frame —
+        // RGB-in is always subsampled to 4:2:0 (verified on the RTX 5070 Ti). So the encoder does an
+        // RGB→YUV444P swscale and needs CPU-resident RGB frames; force the zero-copy GPU capture off
+        // for a 4:4:4 NVENC session. (VAAPI 4:4:4, where the hardware supports it, keeps its dmabuf
+        // path via `scale_vaapi`; Windows NVENC ingests ARGB directly and stays GPU.)
+        #[cfg(target_os = "linux")]
+        let gpu = {
+            let force_cpu_for_nvenc_444 =
+                self.chroma.is_444() && !crate::encode::linux_zero_copy_is_vaapi();
+            gpu && !force_cpu_for_nvenc_444
+        };
         crate::capture::OutputFormat {
-            gpu: self.encoder.is_gpu(),
+            gpu,
             hdr: self.hdr,
+            // 4:4:4 needs a full-chroma source: on Windows this keeps the capturer on RGB (not the
+            // default NV12/P010 video-engine output) so NVENC can CSC to 4:4:4.
+            chroma_444: self.chroma.is_444(),
         }
     }
 }
@@ -134,7 +154,7 @@ impl SessionPlan {
 /// Process topology. On Windows this is the former `punktfunk1::should_use_helper` logic verbatim; on
 /// every other platform the session is always single-process.
 #[cfg(target_os = "windows")]
-fn resolve_topology() -> SessionTopology {
+pub(crate) fn resolve_topology() -> SessionTopology {
     let cfg = crate::config::config();
     // `NO_HELPER`/`NO_WGC` force single-process; IDD-push captures in-process in Session 0 (no helper);
     // otherwise the helper runs when forced or when we're SYSTEM (in-process WGC can't activate there).
@@ -151,7 +171,7 @@ fn resolve_topology() -> SessionTopology {
 }
 
 #[cfg(not(target_os = "windows"))]
-fn resolve_topology() -> SessionTopology {
+pub(crate) fn resolve_topology() -> SessionTopology {
     SessionTopology::SingleProcess
 }
 

crates/punktfunk-host/src/spike.rs

@@ -109,7 +109,8 @@ pub fn run(opts: Options) -> Result<()> {
         opts.fps,
         opts.bitrate_bps,
         first.is_cuda(),
-        8, // spike synthetic harness: 8-bit
+        8,                            // spike synthetic harness: 8-bit
+        encode::ChromaFormat::Yuv420, // ...and 4:2:0
     )
     .context("open encoder")?;
 

crates/punktfunk-host/src/windows/wgc_helper.rs

@@ -98,6 +98,9 @@ pub fn run(opts: HelperOptions) -> Result<()> {
         opts.bitrate_kbps as u64 * 1000,
         false,          // not cuda
         opts.bit_depth, // 8, or 10 = Main10 (HDR auto-upgrades from the Rgb10a2 frame regardless)
+        // The two-process WGC relay helper encodes 4:2:0 in v1 (4:4:4 over the relay is a follow-up);
+        // the host gates 4:4:4 to the single-process topology.
+        encode::ChromaFormat::Yuv420,
     )
     .context("open NVENC")?;