Merge windows-port-phase2: native Windows cpal/WASAPI audio port (b0fdd4e)

Brings the native Windows audio backend to main after a live host<->VM smoke
test: cpal/WASAPI capture+playback, device remap/resampling (W4/B5), cpal
RT-audit closed (B1-B5 + P3), Windows notification chimes, and Wine startup fix.

Verified on real Win11 (libvirt VM) this session: 2-way audio (host<->VM both
directions), audible join/leave/reconnect chimes, GUI renders, echo-cancel
correctly gated off. Linux unchanged (all changes cfg(windows); cargo test --lib
326/0, clippy clean). Windows build is GNU cross-compiled (b0fdd4e tester zip).

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

docs/WINDOWS.md

@@ -68,9 +68,9 @@ connections are expected and valid.
 | Echo cancellation | Linux-only PipeWire feature. The Windows UI shows it disabled as unavailable. |
 | Screen share | Requires a Windows `pixelpass.exe` on `PATH` or a configured override. |
 | Chimes | Now routed through Windows `SoundPlayer`; needs a real Windows host to audibly verify. |
-| Resampling/device format | Open. Devices must support 48 kHz, and output must support stereo; a 44.1 kHz-only/default device currently errors instead of playing. |
+| Resampling/device format | Cross-compiled. cpal/WASAPI now chooses native 48 kHz when available and otherwise resamples/remaps at the device boundary; needs real Windows hardware audio verification. |
 | Device persistence | Open. WASAPI friendly names may duplicate or change across driver/profile changes. |
-| Playback pacing | Open. The fixed playback target under WASAPI shared mode still needs real-hardware verification. |
+| Playback pacing | Cross-compiled. The fixed playback target under WASAPI shared mode still needs real-hardware verification with `audio_probe`. |
 
 Before calling Windows support done, verify a real Windows machine can create/join a room,
 capture mic audio, hear remote audio, select devices, restart with selections preserved, and

packaging/test-pack/PKGBUILD

@@ -0,0 +1,112 @@
+# Maintainer: mollusk <jitty+lc1iz0dc@protonmail.com>
+#
+# Test-pack split package: ONE `makepkg -si` builds + installs BOTH peerspeak
+# (voice chat) and pixelpass (screen sharing) from the public gitbutter repos
+# over https. pixelpass lands on /usr/bin so peerspeak's screen-share button
+# finds it. Shared version string is derived from peerspeak's git.
+#
+# Clone this repo and build from here:
+#   git clone https://gitbutter.xyz/mollusk/peerspeak.git
+#   cd peerspeak/packaging/test-pack
+#   makepkg -si
+pkgbase=peerspeak-git
+pkgname=('peerspeak-git' 'pixelpass')
+pkgver=0.1.0
+pkgrel=1
+arch=('x86_64')
+url="https://gitbutter.xyz/mollusk/peerspeak"
+license=('custom' 'MIT' 'Apache-2.0' 'OFL-1.1')
+makedepends=('git' 'cargo' 'pkgconf')
+options=('!lto' '!debug')
+source=("peerspeak::git+https://gitbutter.xyz/mollusk/peerspeak.git"
+        "pixelpass::git+https://gitbutter.xyz/mollusk/pixelpass.git#branch=main")
+sha256sums=('SKIP'
+            'SKIP')
+
+pkgver() {
+  cd "$srcdir/peerspeak"
+  # Shared across both split packages. 0.1.0.r<commits>.g<short-sha>.
+  printf '%s.r%s.g%s' \
+    "$(awk -F'\"' '/^version =/{print $2; exit}' Cargo.toml)" \
+    "$(git rev-list --count HEAD)" \
+    "$(git rev-parse --short HEAD)"
+}
+
+prepare() {
+  # Vendor deps up front so build() can run --frozen (no surprise network).
+  export CARGO_HOME="$srcdir/cargo-home"
+  local host; host="$(rustc -vV | sed -n 's/host: //p')"
+  cd "$srcdir/peerspeak";  cargo fetch --locked --target "$host"
+  cd "$srcdir/pixelpass";  cargo fetch --locked --target "$host"
+}
+
+build() {
+  export CARGO_HOME="$srcdir/cargo-home"
+  export RUSTUP_TOOLCHAIN=stable
+  export CARGO_TARGET_DIR=target
+
+  cd "$srcdir/peerspeak"
+  cargo build --frozen --release --bin peerspeak
+
+  cd "$srcdir/pixelpass"
+  # --features gui so the .desktop launcher (pixelpass --gui) works.
+  cargo build --frozen --release --features gui
+}
+
+check() {
+  export CARGO_HOME="$srcdir/cargo-home"
+  export RUSTUP_TOOLCHAIN=stable
+  # peerspeak library unit tests only — its integration suites bind real
+  # iroh/QUIC endpoints and fail in a sandboxed/offline build environment.
+  cd "$srcdir/peerspeak"
+  cargo test --frozen --release --lib
+}
+
+package_peerspeak-git() {
+  pkgdesc="Decentralized peer-to-peer voice chat (Rust/iroh/PipeWire/Opus/iced)"
+  depends=('pipewire' 'opus')
+  optdepends=('pixelpass: screen sharing inside a room'
+              'mpv: screen-share viewer (vlc is used as a fallback)')
+  provides=('peerspeak')
+  conflicts=('peerspeak')
+  license=('custom')
+
+  cd "$srcdir/peerspeak"
+  install -Dm755 "target/release/peerspeak" "$pkgdir/usr/bin/peerspeak"
+  install -Dm644 "packaging/peerspeak.desktop" \
+    "$pkgdir/usr/share/applications/peerspeak.desktop"
+
+  # Hicolor icon theme (scalable SVG + the rendered raster sizes).
+  install -Dm644 "assets/icons/peerspeak.svg" \
+    "$pkgdir/usr/share/icons/hicolor/scalable/apps/peerspeak.svg"
+  local s
+  for s in 16 24 32 48 64 128 256 512; do
+    install -Dm644 "assets/icons/peerspeak-$s.png" \
+      "$pkgdir/usr/share/icons/hicolor/${s}x${s}/apps/peerspeak.png"
+  done
+}
+
+package_pixelpass() {
+  pkgdesc='P2P screen sharing over iroh — no port forwarding, no signup'
+  depends=('gstreamer' 'gst-plugins-base' 'gst-plugins-good' 'gst-plugins-bad'
+           'gst-libav' 'gst-plugin-va' 'libpulse' 'hicolor-icon-theme'
+           'libglvnd' 'libxkbcommon' 'wayland')
+  optdepends=('mpv: recommended stream viewer (the GUI launches mpv)'
+              'vlc: alternative stream viewer'
+              'gst-plugins-ugly: software x264 encoding for `pixelpass --no-hwencode`'
+              'gst-plugin-pipewire: screen capture on Wayland sessions'
+              'xorg-xwininfo: share a single window on X11 (`pixelpass --window`)')
+  license=('MIT' 'Apache-2.0' 'OFL-1.1')
+
+  cd "$srcdir/pixelpass"
+  install -Dm0755 "target/release/pixelpass" "$pkgdir/usr/bin/pixelpass"
+  install -Dm0644 assets/pixelpass.desktop \
+    "$pkgdir/usr/share/applications/pixelpass.desktop"
+  install -Dm0644 assets/pixelpass.svg \
+    "$pkgdir/usr/share/icons/hicolor/scalable/apps/pixelpass.svg"
+  install -Dm0644 README.md "$pkgdir/usr/share/doc/pixelpass/README.md"
+  install -Dm0644 LICENSE-MIT "$pkgdir/usr/share/licenses/pixelpass/LICENSE-MIT"
+  install -Dm0644 LICENSE-APACHE "$pkgdir/usr/share/licenses/pixelpass/LICENSE-APACHE"
+  install -Dm0644 assets/NotoSans-OFL.txt \
+    "$pkgdir/usr/share/licenses/pixelpass/NotoSans-OFL.txt"
+}

packaging/test-pack/README.md

@@ -0,0 +1,52 @@
+# PeerSpeak + PixelPass — CachyOS/Arch test pack
+
+A single **split PKGBUILD** that builds the latest code from the public gitbutter
+repos and installs **both** programs at once:
+
+- `peerspeak` — decentralized P2P voice chat
+- `pixelpass` — P2P screen sharing (peerspeak launches it for the screen-share button)
+
+## Build & install (one command)
+
+```sh
+git clone https://gitbutter.xyz/mollusk/peerspeak.git
+cd peerspeak/packaging/test-pack
+makepkg -si
+```
+
+`makepkg -si` auto-installs every dependency via pacman before building —
+including the Rust toolchain itself (the `cargo` makedepend is provided by the
+`rust` package), `git`, `pkgconf`, pipewire + opus for peerspeak, and the
+gstreamer/VA-API stack for pixelpass. The only prerequisite is the `base-devel`
+group (which provides `makepkg`). If you already use `rustup`, that satisfies the
+`cargo` makedepend and the `rust` package won't be pulled in — no conflict.
+
+When it finishes you'll have `peerspeak` and `pixelpass` on your PATH at
+`/usr/bin`. To rebuild later with fresh upstream code, re-run `makepkg -si`; the
+git sources re-pull `main` and the version bumps automatically.
+
+> Skip the test step with `makepkg -si --nocheck` for a faster build.
+
+## Running the cross-internet test
+
+1. Launch `peerspeak` on both machines.
+2. One person **creates** a room and shares the room code/ticket with the other.
+3. The other **joins** with that code.
+4. iroh does NAT hole-punching automatically; if a direct path can't be made it
+   falls back to a public n0 relay — **no port forwarding required**.
+5. Allow the app through any local firewall if prompted (outbound UDP / QUIC;
+   nothing needs to be opened inbound for relay mode).
+
+### What we're smoke-testing
+- Two real humans, two networks, over the internet.
+- Mic capture + remote playback both directions, no crackle/dropouts.
+- Mute / deafen, push-to-talk.
+- Text chat in-room.
+- Avatars (presets + custom upload) show up on the other side.
+- Screen share: click the screen-share control → it launches `pixelpass`; the
+  viewer opens in `mpv` on the receiving side.
+- Notification chimes (join/leave/etc.).
+- Leave / rejoin cleanly.
+
+If anything misbehaves, grab the log path peerspeak prints on startup and the
+exact repro steps.

src/audio/mod.rs

@@ -61,6 +61,10 @@ pub mod gate;
 pub mod limiter;
 pub mod multitrack;
 pub mod pan;
+// Linear resamplers used by the Windows/cpal backend (W4). Platform-neutral and
+// pure, so it builds (and its tests run) everywhere even though only the cpal
+// backend wires it in.
+pub mod resample;
 #[cfg(target_os = "linux")]
 pub mod echo_cancel;
 #[cfg(target_os = "linux")]

src/audio/resample.rs

@@ -0,0 +1,307 @@
+//! Dep-free linear-interpolation resamplers for the Windows/cpal backend (W4).
+//!
+//! The pipeline runs internally at 48 kHz (Opus + the 20 ms frame), but a WASAPI
+//! endpoint may run at a different rate (commonly 44.1 kHz) and/or a non-stereo
+//! channel layout. These convert at the device boundary so such a device plays and
+//! captures instead of hard-erroring (the W4 limitation in the Windows port).
+//!
+//! ## Where each is used
+//! - [`PushResampler`] (single channel) converts **capture** from the device rate
+//!   to 48 kHz on the capture drain thread — off the RT callback.
+//! - [`StereoPullResampler`] converts **playback** from the internal 48 kHz stereo
+//!   bus to the device rate inside the output RT callback, pulling internal frames
+//!   from the ring on demand. It allocates nothing in `next`, so it is RT-safe.
+//!
+//! ## Quality
+//! This is plain linear interpolation with no anti-aliasing filter: correct,
+//! allocation-free, and adequate for speech, but it adds some aliasing when
+//! downsampling. The seam is intentionally tiny so a higher-quality polyphase/FIR
+//! resampler (e.g. the `rubato` crate, pending a supply-chain decision) can later
+//! replace the internals without touching the cpal backend. The matching-rate /
+//! matching-layout path in the backend bypasses these entirely and stays bit-exact.
+
+/// Linear interpolation between `a` and `b` at fractional position `frac` in `[0, 1)`.
+#[inline]
+fn lerp(a: f32, b: f32, frac: f32) -> f32 {
+    a + (b - a) * frac
+}
+
+/// Stateful single-channel **push** resampler: feed input samples at `in_rate`,
+/// receive output samples at `out_rate` through an `emit` callback. It carries the
+/// fractional read position and the previous input sample across calls, so feeding
+/// the stream block-by-block joins seamlessly. Neither [`push`](Self::push) nor
+/// [`process`](Self::process) allocates.
+pub struct PushResampler {
+    /// Input samples consumed per output sample (`in_rate / out_rate`).
+    step: f64,
+    /// Position of the next output sample, in input-sample units, measured from the
+    /// index of `prev` (the most recent input). Always advanced to stay `< 1.0`
+    /// after each input is consumed.
+    next: f64,
+    /// The previous input sample (left edge of the current interpolation segment).
+    prev: f32,
+    /// Whether any input has been seen yet (anchors the first output at input[0]).
+    started: bool,
+}
+
+impl PushResampler {
+    /// Build a resampler from `in_rate` to `out_rate` (both in Hz). Rates are
+    /// clamped to `>= 1` so `step` is always finite and non-zero: a zero `step`
+    /// would make [`push`](Self::push)'s `while self.next < 1.0` loop forever. The
+    /// cpal backend's `resolve()` also rejects such rates up front, so this is
+    /// belt-and-suspenders against a future caller (review W7).
+    pub fn new(in_rate: u32, out_rate: u32) -> Self {
+        Self {
+            step: in_rate.max(1) as f64 / out_rate.max(1) as f64,
+            next: 0.0,
+            prev: 0.0,
+            started: false,
+        }
+    }
+
+    /// Feed one input sample; `emit` is called for each output sample produced
+    /// (zero or more, depending on the rate ratio).
+    pub fn push(&mut self, cur: f32, mut emit: impl FnMut(f32)) {
+        if !self.started {
+            // First sample: just establish the left edge. Linear interpolation
+            // needs the next input as the right edge, so the first output is
+            // produced on the next push. This gives exact alignment
+            // (`output[k] == input[k]` at equal rates) with one input-sample of
+            // latency — negligible (~20 µs at 48 kHz).
+            self.started = true;
+            self.prev = cur;
+            self.next = 0.0;
+            return;
+        }
+        // `prev` sits at position 0 of this segment and `cur` at position 1; emit
+        // every output whose position falls in [0, 1).
+        while self.next < 1.0 {
+            emit(lerp(self.prev, cur, self.next as f32));
+            self.next += self.step;
+        }
+        self.next -= 1.0;
+        self.prev = cur;
+    }
+
+    /// Convenience for tests / batch callers: push a whole slice.
+    pub fn process(&mut self, input: &[f32], mut emit: impl FnMut(f32)) {
+        for &s in input {
+            self.push(s, &mut emit);
+        }
+    }
+}
+
+/// Stateful stereo **pull** resampler: produce output frames at `out_rate` by
+/// pulling input frames at `in_rate` from a closure on demand. Call
+/// [`next`](Self::next) once per output frame; it pulls as many input frames as the
+/// ratio requires and returns the interpolated `(left, right)`, or `None` when the
+/// puller runs dry (an underrun). Allocates nothing, so it is safe in an RT output
+/// callback.
+pub struct StereoPullResampler {
+    /// Input frames consumed per output frame (`in_rate / out_rate`).
+    step: f64,
+    /// Position of the next output frame within `[prev, cur)`, in `[0, 1)`.
+    frac: f64,
+    /// Left edge of the current interpolation segment.
+    prev: (f32, f32),
+    /// Right edge of the current interpolation segment.
+    cur: (f32, f32),
+    /// Whether `prev`/`cur` have been primed from the puller yet.
+    primed: bool,
+}
+
+impl StereoPullResampler {
+    /// Build a resampler from `in_rate` to `out_rate` (both in Hz). Rates are
+    /// clamped to `>= 1` so `step` is finite and non-zero — otherwise
+    /// [`next`](Self::next)'s `while self.frac >= 1.0` could spin (review W7).
+    pub fn new(in_rate: u32, out_rate: u32) -> Self {
+        Self {
+            step: in_rate.max(1) as f64 / out_rate.max(1) as f64,
+            frac: 0.0,
+            prev: (0.0, 0.0),
+            cur: (0.0, 0.0),
+            primed: false,
+        }
+    }
+
+    /// Produce the next output frame, pulling input frames via `pull` as needed.
+    /// Returns `None` if `pull` returns `None` before the frame can be formed
+    /// (underrun); the caller should substitute silence for that frame.
+    pub fn next(&mut self, mut pull: impl FnMut() -> Option<(f32, f32)>) -> Option<(f32, f32)> {
+        if !self.primed {
+            // Prime both edges from two pulls so the first output frame aligns
+            // exactly with the first input frame (`out[0] == in[0]` at equal
+            // rates). Needs two frames available to start, which the prefilled
+            // playback ring always has.
+            self.prev = pull()?;
+            self.cur = pull()?;
+            self.primed = true;
+            self.frac = 0.0;
+        }
+        // Advance the segment until the read position lands inside [prev, cur).
+        while self.frac >= 1.0 {
+            self.prev = self.cur;
+            self.cur = pull()?;
+            self.frac -= 1.0;
+        }
+        let f = self.frac as f32;
+        let out = (lerp(self.prev.0, self.cur.0, f), lerp(self.prev.1, self.cur.1, f));
+        self.frac += self.step;
+        Some(out)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    /// Equal rates align exactly: `output[k] == input[k]`. The final input lands on
+    /// the next push (one-sample streaming latency), so we get `n - 1` outputs.
+    #[test]
+    fn push_identity_when_rates_match() {
+        let mut r = PushResampler::new(48_000, 48_000);
+        let input = [0.0, 0.1, 0.2, 0.3, 0.4];
+        let mut out = Vec::new();
+        r.process(&input, |s| out.push(s));
+        assert_eq!(out.len(), input.len() - 1);
+        for (a, b) in out.iter().zip(input.iter()) {
+            assert!((a - b).abs() < 1e-6, "{a} vs {b}");
+        }
+    }
+
+    /// Upsampling 2x roughly doubles the output count and the midpoints interpolate.
+    #[test]
+    fn push_upsample_2x_interpolates_midpoints() {
+        let mut r = PushResampler::new(24_000, 48_000); // step = 0.5
+        let input = [0.0, 1.0, 2.0, 3.0];
+        let mut out = Vec::new();
+        r.process(&input, |s| out.push(s));
+        // (n - 1) segments at 2 outputs each = 6.
+        assert_eq!(out.len(), 6, "out {out:?}");
+        // A half-step between 1.0 and 2.0 must appear near 1.5.
+        assert!(
+            out.iter().any(|&s| (s - 1.5).abs() < 1e-3),
+            "expected a ~1.5 midpoint in {out:?}"
+        );
+    }
+
+    /// Downsampling drops the rate: fewer outputs than inputs, monotonic ramp preserved.
+    #[test]
+    fn push_downsample_reduces_count() {
+        let mut r = PushResampler::new(48_000, 44_100); // step ~1.088
+        let input: Vec<f32> = (0..441).map(|i| i as f32).collect();
+        let mut out = Vec::new();
+        r.process(&input, |s| out.push(s));
+        // 441 in @ 48k -> ~405 out @ 44.1k.
+        assert!(
+            (390..=410).contains(&out.len()),
+            "expected ~405 outputs, got {}",
+            out.len()
+        );
+        // Output stays within the input's value range and is non-decreasing.
+        for w in out.windows(2) {
+            assert!(w[1] >= w[0] - 1e-3, "ramp should not reverse: {w:?}");
+        }
+        assert!(*out.last().unwrap() <= 440.0 + 1e-3);
+    }
+
+    /// Pull resampler at equal rates returns each input frame in order, aligned.
+    /// Two-pull priming uses one frame of lookahead, so `n` inputs yield `n - 1`
+    /// outputs (the last frame emits once a successor arrives).
+    #[test]
+    fn pull_identity_when_rates_match() {
+        let mut r = StereoPullResampler::new(48_000, 48_000);
+        let frames = [(0.0, 9.0), (1.0, 8.0), (2.0, 7.0), (3.0, 6.0)];
+        let mut idx = 0;
+        let mut out = Vec::new();
+        while let Some(f) = r.next(|| {
+            let v = frames.get(idx).copied();
+            idx += 1;
+            v
+        }) {
+            out.push(f);
+        }
+        assert_eq!(out.len(), frames.len() - 1, "out {out:?}");
+        for (got, want) in out.iter().zip(frames.iter()) {
+            assert!((got.0 - want.0).abs() < 1e-6 && (got.1 - want.1).abs() < 1e-6);
+        }
+    }
+
+    /// Pull resampler reports underrun (`None`) once the source is exhausted.
+    #[test]
+    fn pull_returns_none_on_underrun() {
+        let mut r = StereoPullResampler::new(48_000, 44_100); // step ~1.088 -> pulls >1 per out
+        let frames = [(0.0, 0.0), (1.0, -1.0)];
+        let mut idx = 0;
+        let mut pull = || {
+            let v = frames.get(idx).copied();
+            idx += 1;
+            v
+        };
+        // First frame primes + emits; subsequent calls eventually exhaust the source.
+        let mut produced = 0;
+        let mut hit_none = false;
+        for _ in 0..10 {
+            if r.next(&mut pull).is_some() {
+                produced += 1;
+            } else {
+                hit_none = true;
+                break;
+            }
+        }
+        assert!(produced >= 1, "should produce at least the primed frame");
+        assert!(hit_none, "should report underrun once the puller is dry");
+    }
+
+    /// Downsampling via pull consumes more input frames than it emits output frames.
+    #[test]
+    fn pull_downsample_consumes_more_than_it_emits() {
+        let mut r = StereoPullResampler::new(48_000, 24_000); // step = 2.0
+        let input: Vec<(f32, f32)> = (0..100).map(|i| (i as f32, -(i as f32))).collect();
+        let mut idx = 0;
+        let mut emitted = 0;
+        for _ in 0..40 {
+            let f = r.next(|| {
+                let v = input.get(idx).copied();
+                idx += 1;
+                v
+            });
+            if f.is_some() {
+                emitted += 1;
+            } else {
+                break;
+            }
+        }
+        // At step 2.0 we consume ~2 input frames per output frame.
+        assert!(idx > emitted, "consumed {idx} input, emitted {emitted} output");
+    }
+
+    /// A zero rate must not produce a zero `step` (which would spin `push`'s inner
+    /// `while self.next < 1.0` forever). Clamping makes the call terminate (W7).
+    #[test]
+    fn push_zero_rate_does_not_spin() {
+        let mut r = PushResampler::new(0, 48_000);
+        let mut count = 0usize;
+        // Feed two samples; with a clamped non-zero step this returns promptly.
+        r.push(0.0, |_| count += 1);
+        r.push(1.0, |_| count += 1);
+        // Reaching here at all is the assertion (no hang); some output is produced.
+        assert!(count >= 1);
+    }
+
+    /// A zero output rate must not make the pull resampler's segment-advance loop
+    /// spin. Clamping keeps `step` finite so `next` terminates (W7).
+    #[test]
+    fn pull_zero_out_rate_does_not_spin() {
+        let mut r = StereoPullResampler::new(48_000, 0);
+        let frames = [(0.0, 0.0), (1.0, 1.0), (2.0, 2.0)];
+        let mut idx = 0;
+        let got = r.next(|| {
+            let v = frames.get(idx).copied();
+            idx += 1;
+            v
+        });
+        // Terminates and yields the primed frame instead of hanging.
+        assert!(got.is_some());
+    }
+}

src/bin/audio_probe.rs

@@ -1,11 +1,11 @@
 //! Audio playout diagnostic probe.
 //!
-//! Drives a phase-continuous sine tone through the *real* PipeWire playback path
-//! (`PipeWireBackend::start_playback`), using the *same* fill-paced production
-//! the production mixer uses (`core/mod.rs`): generate a frame only while the
+//! Drives a phase-continuous sine tone through the *real* playback path
+//! (PipeWire on Linux, cpal/WASAPI on Windows), using the *same* fill-paced
+//! production the production mixer uses (`core/mod.rs`): generate a frame only while the
 //! playback ring is below `PLAYBACK_TARGET_SAMPLES`, so production tracks the
-//! PipeWire hardware clock. No network, no microphone — this isolates the local
-//! output path so we can confirm the clock-paced playout is glitch-free.
+//! hardware clock. No network, no microphone — this isolates the local output
+//! path so we can confirm the clock-paced playout is glitch-free.
 //!
 //! Use your ears on the tone (any click/pop is a glitch) together with the
 //! `playout-health:` lines tailed to stdout:
@@ -18,17 +18,24 @@
 //! Run:  cargo run --bin audio_probe -- [freq_hz] [seconds] [target_node]
 //!   e.g. cargo run --release --bin audio_probe -- 440 30
 //!
-//! This probe exercises the PipeWire backend directly, so it is a Linux-only tool.
-//! On non-Linux targets `main` is a stub that explains the limitation.
+//! This probe exercises the platform playback backend directly: PipeWire on Linux
+//! and cpal/WASAPI on Windows. Other targets use a stub that explains the limitation.
 
 #[cfg(target_os = "linux")]
 fn main() {
     unix_probe::run();
 }
 
-#[cfg(not(target_os = "linux"))]
+#[cfg(windows)]
 fn main() {
-    eprintln!("audio_probe is only supported on Linux builds (it drives the PipeWire backend directly).");
+    win_probe::run();
+}
+
+#[cfg(not(any(target_os = "linux", windows)))]
+fn main() {
+    eprintln!(
+        "audio_probe is only supported on Linux and Windows builds (it drives the platform playback backend directly)."
+    );
 }
 
 #[cfg(target_os = "linux")]
@@ -88,7 +95,114 @@ mod unix_probe {
             for _ in 0..FRAME_SAMPLES {
                 let t = n as f32 / SAMPLE_RATE;
                 // 0.25 amplitude: clearly audible but not harsh.
-                let sample = (0.25 * i16::MAX as f32 * (2.0 * std::f32::consts::PI * freq * t).sin()) as i16;
+                let sample =
+                    (0.25 * i16::MAX as f32 * (2.0 * std::f32::consts::PI * freq * t).sin()) as i16;
+                // Stereo playback bus: duplicate the probe tone to L/R.
+                frame.push(sample);
+                frame.push(sample);
+                n += 1;
+            }
+            if tx.send(frame).is_err() {
+                eprintln!("playback channel closed early");
+                break;
+            }
+        }
+
+        // Let the ring drain, then stop.
+        tokio::time::sleep(Duration::from_millis(300)).await;
+        let _ = backend.stop();
+        println!("\naudio_probe: done.");
+    }
+
+    /// Open the app log, seek to the end, and echo new lines (the `playout-health:`
+    /// reports) to stdout once they appear.
+    fn spawn_log_tailer() {
+        let path = peerspeak::log_file_path();
+        std::thread::spawn(move || {
+            // Wait for the file to exist (first log_msg creates it).
+            let file = loop {
+                if let Ok(f) = std::fs::File::open(&path) {
+                    break f;
+                }
+                std::thread::sleep(Duration::from_millis(100));
+            };
+            let mut reader = BufReader::new(file);
+            let _ = reader.seek(SeekFrom::End(0));
+            loop {
+                let mut line = String::new();
+                match reader.read_line(&mut line) {
+                    Ok(0) => std::thread::sleep(Duration::from_millis(150)),
+                    Ok(_) => {
+                        if line.contains("playout-health:") {
+                            print!("{line}");
+                        }
+                    }
+                    Err(_) => std::thread::sleep(Duration::from_millis(150)),
+                }
+            }
+        });
+    }
+}
+
+#[cfg(windows)]
+mod win_probe {
+    use std::io::{BufRead, BufReader, Seek, SeekFrom};
+    use std::sync::Arc;
+    use std::sync::atomic::AtomicUsize;
+    use std::sync::mpsc;
+    use std::time::Duration;
+
+    use peerspeak::audio::AudioBackend;
+    use peerspeak::audio::cpal_impl::CpalBackend;
+    use peerspeak::core::jitter::FRAME_SAMPLES; // 960 mono frames = 20ms @ 48kHz
+
+    const SAMPLE_RATE: f32 = 48_000.0;
+
+    #[tokio::main]
+    pub async fn run() {
+        let mut args = std::env::args().skip(1);
+        let freq: f32 = args.next().and_then(|s| s.parse().ok()).unwrap_or(440.0);
+        let secs: u64 = args.next().and_then(|s| s.parse().ok()).unwrap_or(30);
+        let target_node: Option<String> = args.next();
+
+        // The playout-health logger is quiet in normal operation (it only logs
+        // glitches); ask it for the full once-per-second heartbeat so the probe can
+        // show the steady-state numbers.
+        // SAFETY: set before any playback thread starts, so no concurrent env read.
+        unsafe { std::env::set_var("PEERSPEAK_AUDIO_VERBOSE", "1") };
+
+        println!("audio_probe: {freq} Hz tone for {secs}s through the real playback path.");
+        println!("Listen for clicks/pops; watch the playout-health lines below.\n");
+
+        // Tail the app log (where playout-health lines land) to stdout in the
+        // background so it's all in one terminal.
+        spawn_log_tailer();
+
+        let backend = CpalBackend::new();
+        let (tx, rx) = mpsc::channel::<Vec<i16>>();
+        let ring_fill = Arc::new(AtomicUsize::new(0));
+        if let Err(e) = backend.start_playback(rx, target_node, ring_fill.clone()) {
+            eprintln!("failed to start playback: {e}");
+            return;
+        }
+
+        // Phase-continuous sine, generated one 20ms frame at a time, fill-paced
+        // exactly like the production mixer: only produce while the ring is below
+        // target, so production tracks the cpal/WASAPI hardware clock.
+        use std::sync::atomic::Ordering;
+        let deadline = tokio::time::Instant::now() + Duration::from_secs(secs);
+        let mut n: u64 = 0; // running sample index keeps phase continuous across frames
+        while tokio::time::Instant::now() < deadline {
+            if ring_fill.load(Ordering::Relaxed) >= peerspeak::audio::PLAYBACK_TARGET_SAMPLES {
+                tokio::time::sleep(Duration::from_millis(2)).await;
+                continue;
+            }
+            let mut frame = Vec::with_capacity(FRAME_SAMPLES * peerspeak::audio::PLAYBACK_CHANNELS);
+            for _ in 0..FRAME_SAMPLES {
+                let t = n as f32 / SAMPLE_RATE;
+                // 0.25 amplitude: clearly audible but not harsh.
+                let sample =
+                    (0.25 * i16::MAX as f32 * (2.0 * std::f32::consts::PI * freq * t).sin()) as i16;
                 // Stereo playback bus: duplicate the probe tone to L/R.
                 frame.push(sample);
                 frame.push(sample);

src/core/mod.rs

@@ -1816,17 +1816,26 @@ async fn run_core_loop(
             }
 
             CoreCommand::SetNetworkMode(mode) => {
-                network_mode = mode;
-                // Rebuild the persistent stack to the new posture immediately if
-                // idle; if a call is active, defer to the next Leave/Join so the
-                // live call isn't disrupted (preserves "applies on next join").
-                if active_session.is_none() {
-                    let lookup = net.memory_lookup.clone();
-                    net.shutdown().await;
-                    let publish = presence_mode.lock().unwrap().publishes_to_discovery();
-                    net = build_net_stack(secret_key.clone(), network_mode, lookup, friends_handler.clone(), publish).await?;
-                } else {
-                    net_rebuild_pending = true;
+                // Skip when the posture is unchanged. The GUI re-sends the saved
+                // network mode as part of its startup config-sync, and that mode
+                // usually already matches the freshly-built stack — rebuilding the
+                // iroh endpoint for an identical posture just churns the network
+                // and adds a needless ~1s teardown+rebuild bounce at every launch
+                // (seen on both Linux and Windows/Wine). A real change still
+                // rebuilds exactly as before.
+                if mode != network_mode {
+                    network_mode = mode;
+                    // Rebuild the persistent stack to the new posture immediately if
+                    // idle; if a call is active, defer to the next Leave/Join so the
+                    // live call isn't disrupted (preserves "applies on next join").
+                    if active_session.is_none() {
+                        let lookup = net.memory_lookup.clone();
+                        net.shutdown().await;
+                        let publish = presence_mode.lock().unwrap().publishes_to_discovery();
+                        net = build_net_stack(secret_key.clone(), network_mode, lookup, friends_handler.clone(), publish).await?;
+                    } else {
+                        net_rebuild_pending = true;
+                    }
                 }
             }