feat(media): stabilize hdmi upstream timing

Cargo.lock

@@ -1642,7 +1642,7 @@ checksum = "09edd9e8b54e49e587e4f6295a7d29c3ea94d469cb40ab8ca70b288248a81db2"
 
 [[package]]
 name = "lesavka_client"
-version = "0.12.4"
+version = "0.13.0"
 dependencies = [
  "anyhow",
  "async-stream",
@@ -1676,7 +1676,7 @@ dependencies = [
 
 [[package]]
 name = "lesavka_common"
-version = "0.12.4"
+version = "0.13.0"
 dependencies = [
  "anyhow",
  "base64",
@@ -1688,7 +1688,7 @@ dependencies = [
 
 [[package]]
 name = "lesavka_server"
-version = "0.12.4"
+version = "0.13.0"
 dependencies = [
  "anyhow",
  "base64",

client/Cargo.toml

@@ -4,7 +4,7 @@ path                    = "src/main.rs"
 
 [package]
 name                    = "lesavka_client"
-version                 = "0.12.4"
+version                 = "0.13.0"
 edition                 = "2024"
 
 [dependencies]

client/src/app/uplink_media.rs

@@ -19,11 +19,10 @@ impl LesavkaClientApp {
             let telemetry_thread = telemetry.clone();
             let queue_thread = queue.clone();
             std::thread::spawn(move || {
-                let mut age_tracker = PacketAgeTracker::default();
                 while stop_rx.try_recv().is_err() {
                     if let Some(pkt) = mic_clone.pull() {
                         trace!("🎤📤 cli  {} bytes → gRPC", pkt.data.len());
-                        let enqueue_age = age_tracker.packet_age(pkt.pts);
+                        let enqueue_age = crate::live_capture_clock::packet_age(pkt.pts);
                         let stats = queue_thread.push(pkt, enqueue_age);
                         if stats.dropped_queue_full > 0 {
                             telemetry_thread.record_queue_full_drop(stats.dropped_queue_full);
@@ -102,7 +101,6 @@ impl LesavkaClientApp {
                 let telemetry = telemetry.clone();
                 let queue = queue.clone();
                 move || {
-                    let mut age_tracker = PacketAgeTracker::default();
                     loop {
                         if stop_rx.try_recv().is_ok() {
                             break;
@@ -118,7 +116,7 @@ impl LesavkaClientApp {
                             tracing::trace!("📸 cli  frame#{n}  {} B", pkt.data.len());
                         }
                         tracing::trace!("📸⬆️ sent webcam AU pts={} {} B", pkt.pts, pkt.data.len());
-                        let enqueue_age = age_tracker.packet_age(pkt.pts);
+                        let enqueue_age = crate::live_capture_clock::packet_age(pkt.pts);
                         let stats = queue.push(pkt, enqueue_age);
                         if stats.dropped_queue_full > 0 {
                             telemetry.record_queue_full_drop(stats.dropped_queue_full);
@@ -205,21 +203,3 @@ fn queue_depth_u32(depth: usize) -> u32 {
 fn duration_ms(duration: Duration) -> f32 {
     duration.as_secs_f32() * 1_000.0
 }
-
-#[cfg(not(coverage))]
-#[derive(Default)]
-struct PacketAgeTracker {
-    origin: Option<Instant>,
-}
-
-#[cfg(not(coverage))]
-impl PacketAgeTracker {
-    fn packet_age(&mut self, pts_us: u64) -> Duration {
-        let pts = Duration::from_micros(pts_us);
-        let now = Instant::now();
-        let origin = self
-            .origin
-            .get_or_insert_with(|| now.checked_sub(pts).unwrap_or(now));
-        now.saturating_duration_since(*origin + pts)
-    }
-}

client/src/bin/lesavka-sync-analyze.rs

@@ -0,0 +1,98 @@
+#[cfg(any(not(coverage), test))]
+use anyhow::{Context, Result, bail};
+
+#[cfg(not(coverage))]
+use lesavka_client::sync_probe::analyze::{SyncAnalysisOptions, analyze_capture};
+
+#[cfg(not(coverage))]
+fn main() -> Result<()> {
+    let (capture_path, emit_json) = parse_args(std::env::args().skip(1))?;
+    let report = analyze_capture(&capture_path, &SyncAnalysisOptions::default())
+        .with_context(|| format!("analyzing sync capture {}", capture_path.display()))?;
+
+    if emit_json {
+        println!(
+            "{}",
+            serde_json::to_string_pretty(&report).context("serializing JSON report")?
+        );
+    } else {
+        println!("A/V sync report for {}", capture_path.display());
+        println!("- video onsets: {}", report.video_event_count);
+        println!("- audio onsets: {}", report.audio_event_count);
+        println!("- paired pulses: {}", report.paired_event_count);
+        println!(
+            "- first skew: {:+.1} ms (audio after video is positive)",
+            report.first_skew_ms
+        );
+        println!("- last skew: {:+.1} ms", report.last_skew_ms);
+        println!("- mean skew: {:+.1} ms", report.mean_skew_ms);
+        println!("- median skew: {:+.1} ms", report.median_skew_ms);
+        println!("- max abs skew: {:.1} ms", report.max_abs_skew_ms);
+        println!("- drift: {:+.1} ms", report.drift_ms);
+    }
+
+    Ok(())
+}
+
+#[cfg(any(not(coverage), test))]
+fn parse_args<I, S>(args: I) -> Result<(std::path::PathBuf, bool)>
+where
+    I: IntoIterator<Item = S>,
+    S: Into<String>,
+{
+    let args = args.into_iter().map(Into::into).collect::<Vec<_>>();
+    if args.is_empty() || args.iter().any(|arg| arg == "--help" || arg == "-h") {
+        println!("Usage: lesavka-sync-analyze <capture.mkv> [--json]");
+        std::process::exit(0);
+    }
+
+    let mut emit_json = false;
+    let mut capture_path = None::<std::path::PathBuf>;
+    for arg in args {
+        if arg == "--json" {
+            emit_json = true;
+            continue;
+        }
+        if capture_path.is_some() {
+            bail!("unexpected extra argument `{arg}`");
+        }
+        capture_path = Some(std::path::PathBuf::from(arg));
+    }
+
+    let capture_path = capture_path.context("capture path is required")?;
+    Ok((capture_path, emit_json))
+}
+
+#[cfg(coverage)]
+fn main() {}
+
+#[cfg(test)]
+mod tests {
+    use super::parse_args;
+
+    #[test]
+    fn parse_args_accepts_capture_path_and_json_flag() {
+        let (path, json) = parse_args(["capture.mkv", "--json"]).expect("args");
+        assert_eq!(path, std::path::PathBuf::from("capture.mkv"));
+        assert!(json);
+    }
+
+    #[test]
+    fn parse_args_rejects_extra_positional_arguments() {
+        assert!(parse_args(["one.mkv", "two.mkv"]).is_err());
+    }
+
+    #[test]
+    fn parse_args_requires_a_capture_path() {
+        let error = parse_args(["--json"]).expect_err("missing capture path should fail");
+        assert!(
+            error.to_string().contains("capture path is required"),
+            "unexpected error: {error:#}"
+        );
+    }
+
+    #[test]
+    fn coverage_main_stub_is_non_panicking() {
+        let _ = super::main();
+    }
+}

client/src/bin/lesavka-sync-probe.rs

@@ -0,0 +1,19 @@
+#[cfg(not(coverage))]
+use anyhow::Result;
+
+#[cfg(not(coverage))]
+#[tokio::main(flavor = "current_thread")]
+async fn main() -> Result<()> {
+    lesavka_client::sync_probe::run_sync_probe_from_args(std::env::args().skip(1)).await
+}
+
+#[cfg(coverage)]
+fn main() {}
+
+#[cfg(test)]
+mod tests {
+    #[test]
+    fn coverage_main_stub_is_non_panicking() {
+        let _ = super::main();
+    }
+}

client/src/input/camera/capture_pipeline.rs

@@ -1,3 +1,21 @@
+#[cfg(any(coverage, test))]
+fn shared_capture_pts_us() -> u64 {
+    use std::sync::OnceLock;
+    use std::time::Instant;
+
+    static CAPTURE_ORIGIN: OnceLock<Instant> = OnceLock::new();
+    CAPTURE_ORIGIN
+        .get_or_init(Instant::now)
+        .elapsed()
+        .as_micros()
+        .min(u64::MAX as u128) as u64
+}
+
+#[cfg(not(any(coverage, test)))]
+fn shared_capture_pts_us() -> u64 {
+    crate::live_capture_clock::capture_pts_us()
+}
+
 impl CameraCapture {
     pub fn new(device_fragment: Option<&str>, cfg: Option<CameraConfig>) -> anyhow::Result<Self> {
         gst::init().ok();
@@ -234,7 +252,7 @@ impl CameraCapture {
         let sample = self.sink.pull_sample().ok()?;
         let buf = sample.buffer()?;
         let map = buf.map_readable().ok()?;
-        let pts = buf.pts().unwrap_or(gst::ClockTime::ZERO).nseconds() / 1_000;
+        let pts = shared_capture_pts_us();
         static FIRST_CAMERA_PACKET: AtomicBool = AtomicBool::new(false);
         if !FIRST_CAMERA_PACKET.swap(true, Ordering::Relaxed) {
             tracing::info!(

client/src/input/microphone.rs

@@ -20,6 +20,24 @@ use tracing::{debug, warn};
 #[cfg(not(coverage))]
 use tracing::{error, info, trace};
 
+#[cfg(any(coverage, test))]
+fn shared_capture_pts_us() -> u64 {
+    use std::sync::OnceLock;
+    use std::time::Instant;
+
+    static CAPTURE_ORIGIN: OnceLock<Instant> = OnceLock::new();
+    CAPTURE_ORIGIN
+        .get_or_init(Instant::now)
+        .elapsed()
+        .as_micros()
+        .min(u64::MAX as u128) as u64
+}
+
+#[cfg(not(any(coverage, test)))]
+fn shared_capture_pts_us() -> u64 {
+    crate::live_capture_clock::capture_pts_us()
+}
+
 const MIC_GAIN_ENV: &str = "LESAVKA_MIC_GAIN";
 const MIC_GAIN_CONTROL_ENV: &str = "LESAVKA_MIC_GAIN_CONTROL";
 const MIC_LEVEL_TAP_ENV: &str = "LESAVKA_UPLINK_MIC_LEVEL";
@@ -122,7 +140,7 @@ impl MicrophoneCapture {
             Ok(sample) => {
                 let buf = sample.buffer().unwrap();
                 let map = buf.map_readable().unwrap();
-                let pts = buf.pts().unwrap_or(gst::ClockTime::ZERO).nseconds() / 1_000;
+                let pts = shared_capture_pts_us();
                 #[cfg(not(coverage))]
                 {
                     static CNT: AtomicU64 = AtomicU64::new(0);

client/src/launcher/tests/ui_runtime.rs

@@ -289,9 +289,9 @@ fn server_chip_state_tracks_connection_not_just_reachability() {
     assert_eq!(server_version_label(&state), "-");
 
     state.set_server_available(true);
-    state.set_server_version(Some("0.12.4".to_string()));
+    state.set_server_version(Some("0.13.0".to_string()));
     assert_eq!(server_light_state(&state, false), StatusLightState::Live);
-    assert_eq!(server_version_label(&state), "v0.12.4");
+    assert_eq!(server_version_label(&state), "v0.13.0");
 
     assert_eq!(
         server_light_state(&state, true),

client/src/lib.rs

@@ -12,8 +12,10 @@ pub mod handshake;
 pub mod input;
 pub mod launcher;
 pub mod layout;
+pub(crate) mod live_capture_clock;
 pub mod output;
 pub mod paste;
+pub mod sync_probe;
 pub(crate) mod uplink_fresh_queue;
 pub(crate) mod uplink_latency_harness;
 pub(crate) mod uplink_telemetry;

client/src/live_capture_clock.rs

@@ -0,0 +1,56 @@
+#![forbid(unsafe_code)]
+
+use std::sync::OnceLock;
+use std::time::{Duration, Instant};
+
+static CAPTURE_ORIGIN: OnceLock<Instant> = OnceLock::new();
+
+fn origin() -> Instant {
+    *CAPTURE_ORIGIN.get_or_init(Instant::now)
+}
+
+/// Return the shared live-capture timestamp for upstream camera/mic packets.
+///
+/// Inputs: none.
+/// Outputs: microseconds elapsed since the relay child first stamped live media.
+/// Why: camera and microphone capture pipelines run independently, so they need
+/// one explicit common origin before the server can keep them on the same live
+/// call timeline.
+#[must_use]
+pub fn capture_pts_us() -> u64 {
+    origin().elapsed().as_micros().min(u64::MAX as u128) as u64
+}
+
+/// Measure how old one shared capture timestamp is right now.
+///
+/// Inputs: a packet timestamp previously produced by `capture_pts_us`.
+/// Outputs: the elapsed age as a `Duration`.
+/// Why: upstream freshness telemetry should use the same shared live clock as
+/// packet timestamps so queue-age calculations stay honest.
+#[must_use]
+pub fn packet_age(pts_us: u64) -> Duration {
+    Duration::from_micros(capture_pts_us().saturating_sub(pts_us))
+}
+
+#[cfg(test)]
+mod tests {
+    use super::{capture_pts_us, packet_age};
+    use std::time::Duration;
+
+    #[test]
+    fn capture_pts_us_monotonically_advances() {
+        let first = capture_pts_us();
+        std::thread::sleep(Duration::from_millis(2));
+        let second = capture_pts_us();
+        assert!(second >= first);
+    }
+
+    #[test]
+    fn packet_age_is_small_for_recent_packets() {
+        let pts = capture_pts_us();
+        std::thread::sleep(Duration::from_millis(2));
+        let age = packet_age(pts);
+        assert!(age >= Duration::from_millis(1));
+        assert!(age < Duration::from_secs(1));
+    }
+}

client/src/sync_probe/analyze.rs

@@ -0,0 +1,86 @@
+//! Analyze captured upstream sync-probe media for audio/video skew and drift.
+
+mod media_extract;
+mod onset_detection;
+mod report;
+#[cfg(test)]
+pub(super) mod test_support;
+
+use anyhow::Result;
+use std::path::Path;
+
+use media_extract::{extract_audio_samples, extract_video_brightness, extract_video_timestamps};
+use onset_detection::{
+    DEFAULT_AUDIO_SAMPLE_RATE_HZ, correlate_segments, detect_audio_segments, detect_video_segments,
+};
+
+pub use onset_detection::{detect_audio_onsets, detect_video_onsets};
+pub use report::{SyncAnalysisOptions, SyncAnalysisReport};
+
+/// Analyzes a captured upstream sync-probe file by extracting video and audio
+/// pulses, then correlating them into skew and drift metrics.
+pub fn analyze_capture(
+    capture_path: &Path,
+    options: &SyncAnalysisOptions,
+) -> Result<SyncAnalysisReport> {
+    let timestamps = extract_video_timestamps(capture_path)?;
+    let brightness = extract_video_brightness(capture_path)?;
+    let video_segments = detect_video_segments(&timestamps, &brightness)?;
+
+    let audio_samples = extract_audio_samples(capture_path)?;
+    let audio_segments = detect_audio_segments(
+        &audio_samples,
+        DEFAULT_AUDIO_SAMPLE_RATE_HZ,
+        options.audio_window_ms,
+    )?;
+
+    correlate_segments(
+        &video_segments,
+        &audio_segments,
+        options.pulse_period_s,
+        options.pulse_width_s,
+        options.marker_tick_period,
+        options.max_pair_gap_s,
+    )
+}
+
+#[cfg(test)]
+mod tests {
+    use super::test_support::{
+        audio_samples_to_bytes, click_track_samples, frame_json, with_fake_media_tools,
+    };
+    use super::{SyncAnalysisOptions, analyze_capture};
+
+    #[test]
+    fn analyze_capture_runs_against_fake_media_tools() {
+        let timestamps = (0..15).map(|index| index as f64 / 10.0).collect::<Vec<_>>();
+        let brightness = timestamps
+            .iter()
+            .enumerate()
+            .map(|(index, _)| if matches!(index, 0 | 5 | 10) { 250 } else { 5 })
+            .collect::<Vec<_>>();
+        let audio = click_track_samples(&[0.05, 0.55, 1.05], 53_000);
+
+        with_fake_media_tools(
+            &frame_json(&timestamps),
+            &brightness,
+            &audio_samples_to_bytes(&audio),
+            |capture_path| {
+                let report = analyze_capture(
+                    capture_path,
+                    &SyncAnalysisOptions {
+                        pulse_period_s: 0.5,
+                        ..SyncAnalysisOptions::default()
+                    },
+                )
+                .expect("analysis report");
+                assert_eq!(report.video_event_count, 3);
+                assert_eq!(report.audio_event_count, 3);
+                assert_eq!(report.paired_event_count, 3);
+                assert_eq!(report.skews_ms.len(), 3);
+                assert!((report.first_skew_ms - 50.0).abs() < 10.0);
+                assert!(report.max_abs_skew_ms < 120.0);
+            },
+        );
+    }
+}

client/src/sync_probe/analyze/media_extract.rs

@@ -0,0 +1,240 @@
+use anyhow::{Context, Result, bail};
+use serde::Deserialize;
+use std::path::Path;
+use std::process::Command;
+
+#[derive(Debug, Deserialize)]
+struct ProbeFrameResponse {
+    #[serde(default)]
+    frames: Vec<ProbeFrameEntry>,
+}
+
+#[derive(Debug, Deserialize)]
+struct ProbeFrameEntry {
+    best_effort_timestamp_time: Option<String>,
+}
+
+pub(super) fn extract_video_timestamps(capture_path: &Path) -> Result<Vec<f64>> {
+    let output = run_command(
+        Command::new("ffprobe")
+            .arg("-hide_banner")
+            .arg("-loglevel")
+            .arg("error")
+            .arg("-select_streams")
+            .arg("v:0")
+            .arg("-show_frames")
+            .arg("-show_entries")
+            .arg("frame=best_effort_timestamp_time")
+            .arg("-of")
+            .arg("json")
+            .arg(capture_path),
+        "ffprobe video timestamps",
+    )?;
+    let response: ProbeFrameResponse =
+        serde_json::from_slice(&output).context("parsing ffprobe frame JSON")?;
+    let timestamps = response
+        .frames
+        .into_iter()
+        .filter_map(|entry| entry.best_effort_timestamp_time)
+        .map(|value| value.parse::<f64>().context("parsing frame timestamp"))
+        .collect::<Result<Vec<_>>>()?;
+    if timestamps.is_empty() {
+        bail!("ffprobe did not return any video frame timestamps");
+    }
+    Ok(timestamps)
+}
+
+pub(super) fn extract_video_brightness(capture_path: &Path) -> Result<Vec<u8>> {
+    let output = run_command(
+        Command::new("ffmpeg")
+            .arg("-hide_banner")
+            .arg("-loglevel")
+            .arg("error")
+            .arg("-i")
+            .arg(capture_path)
+            .arg("-map")
+            .arg("0:v:0")
+            .arg("-vf")
+            .arg("scale=1:1,format=gray")
+            .arg("-f")
+            .arg("rawvideo")
+            .arg("-pix_fmt")
+            .arg("gray")
+            .arg("-"),
+        "ffmpeg video brightness extraction",
+    )?;
+    if output.is_empty() {
+        bail!("ffmpeg did not emit any video brightness data");
+    }
+    Ok(output)
+}
+
+pub(super) fn extract_audio_samples(capture_path: &Path) -> Result<Vec<i16>> {
+    let output = run_command(
+        Command::new("ffmpeg")
+            .arg("-hide_banner")
+            .arg("-loglevel")
+            .arg("error")
+            .arg("-i")
+            .arg(capture_path)
+            .arg("-map")
+            .arg("0:a:0")
+            .arg("-ac")
+            .arg("1")
+            .arg("-ar")
+            .arg(super::onset_detection::DEFAULT_AUDIO_SAMPLE_RATE_HZ.to_string())
+            .arg("-f")
+            .arg("s16le")
+            .arg("-acodec")
+            .arg("pcm_s16le")
+            .arg("-"),
+        "ffmpeg audio extraction",
+    )?;
+    if output.len() < 2 {
+        bail!("ffmpeg did not emit enough audio data to analyze");
+    }
+    Ok(output
+        .chunks_exact(2)
+        .map(|chunk| i16::from_le_bytes([chunk[0], chunk[1]]))
+        .collect())
+}
+
+pub(super) fn run_command(command: &mut Command, description: &str) -> Result<Vec<u8>> {
+    let output = command
+        .output()
+        .with_context(|| format!("running {description}"))?;
+    if !output.status.success() {
+        let stderr = String::from_utf8_lossy(&output.stderr);
+        bail!("{description} failed: {}", stderr.trim());
+    }
+    Ok(output.stdout)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::{
+        extract_audio_samples, extract_video_brightness, extract_video_timestamps, run_command,
+    };
+    use crate::sync_probe::analyze::test_support::{
+        audio_samples_to_bytes, frame_json, with_fake_media_tools,
+    };
+    use std::process::Command;
+
+    #[test]
+    fn extract_video_timestamps_reads_fake_ffprobe_output() {
+        let timestamps = vec![0.0, 0.5, 1.0];
+        with_fake_media_tools(
+            &frame_json(&timestamps),
+            &[1, 2, 3],
+            &[1, 0],
+            |capture_path| {
+                let parsed = extract_video_timestamps(capture_path).expect("video timestamps");
+                assert_eq!(parsed, timestamps);
+            },
+        );
+    }
+
+    #[test]
+    fn extract_video_timestamps_rejects_empty_and_invalid_outputs() {
+        with_fake_media_tools(br#"{"frames":[]}"#, &[1], &[1, 0], |capture_path| {
+            let error = extract_video_timestamps(capture_path).expect_err("empty frames fail");
+            assert!(
+                error
+                    .to_string()
+                    .contains("did not return any video frame timestamps")
+            );
+        });
+
+        with_fake_media_tools(
+            br#"{"frames":[{"best_effort_timestamp_time":"bad"}]}"#,
+            &[1],
+            &[1, 0],
+            |capture_path| {
+                let error =
+                    extract_video_timestamps(capture_path).expect_err("invalid timestamp fails");
+                assert!(error.to_string().contains("parsing frame timestamp"));
+            },
+        );
+    }
+
+    #[test]
+    fn extract_video_brightness_reads_fake_ffmpeg_output() {
+        let brightness = vec![5u8, 100, 250];
+        with_fake_media_tools(
+            br#"{"frames":[{"best_effort_timestamp_time":"0.0"}]}"#,
+            &brightness,
+            &[1, 0],
+            |capture_path| {
+                let parsed = extract_video_brightness(capture_path).expect("video brightness");
+                assert_eq!(parsed, brightness);
+            },
+        );
+    }
+
+    #[test]
+    fn extract_video_brightness_rejects_empty_output() {
+        with_fake_media_tools(
+            br#"{"frames":[{"best_effort_timestamp_time":"0.0"}]}"#,
+            &[],
+            &[1, 0],
+            |capture_path| {
+                let error = extract_video_brightness(capture_path).expect_err("empty brightness");
+                assert!(
+                    error
+                        .to_string()
+                        .contains("did not emit any video brightness data")
+                );
+            },
+        );
+    }
+
+    #[test]
+    fn extract_audio_samples_reads_fake_ffmpeg_output() {
+        let samples = vec![1i16, -2, 32_000];
+        with_fake_media_tools(
+            br#"{"frames":[{"best_effort_timestamp_time":"0.0"}]}"#,
+            &[1],
+            &audio_samples_to_bytes(&samples),
+            |capture_path| {
+                let parsed = extract_audio_samples(capture_path).expect("audio samples");
+                assert_eq!(parsed, samples);
+            },
+        );
+    }
+
+    #[test]
+    fn extract_audio_samples_rejects_too_short_output() {
+        with_fake_media_tools(
+            br#"{"frames":[{"best_effort_timestamp_time":"0.0"}]}"#,
+            &[1],
+            &[7],
+            |capture_path| {
+                let error = extract_audio_samples(capture_path).expect_err("short audio");
+                assert!(
+                    error
+                        .to_string()
+                        .contains("did not emit enough audio data to analyze")
+                );
+            },
+        );
+    }
+
+    #[test]
+    fn run_command_reports_success_and_failure() {
+        let output = run_command(
+            Command::new("sh").arg("-c").arg("printf 'ok'"),
+            "success command",
+        )
+        .expect("success output");
+        assert_eq!(output, b"ok");
+
+        let error = run_command(
+            Command::new("sh")
+                .arg("-c")
+                .arg("printf 'boom' >&2; exit 7"),
+            "failing command",
+        )
+        .expect_err("failing command should error");
+        assert!(error.to_string().contains("failing command failed: boom"));
+    }
+}

client/src/sync_probe/analyze/onset_detection.rs

@@ -0,0 +1,229 @@
+use anyhow::{Result, bail};
+
+mod correlation;
+#[cfg(test)]
+mod tests;
+
+pub(crate) use correlation::correlate_segments;
+
+pub(super) const DEFAULT_AUDIO_SAMPLE_RATE_HZ: u32 = 48_000;
+const MIN_VIDEO_CONTRAST: u8 = 16;
+
+#[derive(Clone, Copy, Debug, PartialEq)]
+pub(crate) struct PulseSegment {
+    pub start_s: f64,
+    pub end_s: f64,
+    pub duration_s: f64,
+}
+
+pub fn detect_video_onsets(timestamps_s: &[f64], brightness: &[u8]) -> Result<Vec<f64>> {
+    Ok(detect_video_segments(timestamps_s, brightness)?
+        .into_iter()
+        .map(|segment| segment.start_s)
+        .collect())
+}
+
+pub(crate) fn detect_video_segments(
+    timestamps_s: &[f64],
+    brightness: &[u8],
+) -> Result<Vec<PulseSegment>> {
+    let frame_count = timestamps_s.len().min(brightness.len());
+    if frame_count == 0 {
+        bail!("capture did not contain any video frames");
+    }
+
+    let slice = &brightness[..frame_count];
+    let min = *slice.iter().min().expect("non-empty brightness slice");
+    let max = *slice.iter().max().expect("non-empty brightness slice");
+    if max.saturating_sub(min) < MIN_VIDEO_CONTRAST {
+        bail!("video flash contrast is too low to detect sync pulses");
+    }
+    let threshold = ((u16::from(min) + u16::from(max)) / 2) as u8;
+    let frame_step_s = median_frame_step_seconds(&timestamps_s[..frame_count]).max(1.0 / 120.0);
+    let mut segments = Vec::new();
+    let mut previous_active = false;
+    let mut segment_start = 0.0_f64;
+    let mut previous_timestamp = None;
+    let mut last_active_timestamp = None;
+    for (timestamp, level) in timestamps_s.iter().copied().zip(slice.iter().copied()) {
+        let active = level >= threshold;
+        if active && !previous_active {
+            segment_start = previous_timestamp
+                .map(|prior| edge_midpoint(prior, timestamp))
+                .unwrap_or(timestamp);
+        }
+        if active {
+            last_active_timestamp = Some(timestamp);
+        }
+        if previous_active && !active {
+            let end_s = edge_midpoint(
+                last_active_timestamp.unwrap_or(timestamp - frame_step_s),
+                timestamp,
+            )
+            .max(segment_start + frame_step_s / 2.0);
+            segments.push(PulseSegment {
+                start_s: segment_start,
+                end_s,
+                duration_s: end_s - segment_start,
+            });
+        }
+        previous_active = active;
+        previous_timestamp = Some(timestamp);
+    }
+    if previous_active {
+        let last_timestamp = timestamps_s[frame_count - 1];
+        let end_s = last_timestamp + frame_step_s / 2.0;
+        segments.push(PulseSegment {
+            start_s: segment_start,
+            end_s,
+            duration_s: end_s - segment_start,
+        });
+    }
+
+    Ok(segments)
+}
+
+pub fn detect_audio_onsets(
+    samples: &[i16],
+    sample_rate_hz: u32,
+    window_ms: u32,
+) -> Result<Vec<f64>> {
+    Ok(detect_audio_segments(samples, sample_rate_hz, window_ms)?
+        .into_iter()
+        .map(|segment| segment.start_s)
+        .collect())
+}
+
+pub(crate) fn detect_audio_segments(
+    samples: &[i16],
+    sample_rate_hz: u32,
+    window_ms: u32,
+) -> Result<Vec<PulseSegment>> {
+    if samples.is_empty() {
+        bail!("capture did not contain any audio samples");
+    }
+    if sample_rate_hz == 0 {
+        bail!("audio sample rate must stay positive");
+    }
+    if window_ms == 0 {
+        bail!("audio analysis window must stay positive");
+    }
+
+    let window_samples = ((sample_rate_hz as usize * window_ms as usize) / 1000).max(1);
+    let envelope = samples
+        .chunks(window_samples)
+        .map(|chunk| {
+            let total: u64 = chunk
+                .iter()
+                .map(|sample| i32::from(*sample).unsigned_abs() as u64)
+                .sum();
+            total as f64 / chunk.len() as f64
+        })
+        .collect::<Vec<_>>();
+    let peak = envelope.iter().copied().fold(0.0_f64, f64::max);
+    if peak < 50.0 {
+        bail!("audio probe peaks are too quiet to detect sync pulses");
+    }
+    let baseline = median(envelope.clone());
+    let threshold = baseline + ((peak - baseline) * 0.45);
+    let sample_abs = samples
+        .iter()
+        .map(|sample| i32::from(*sample).unsigned_abs() as f64)
+        .collect::<Vec<_>>();
+    let sample_peak = sample_abs.iter().copied().fold(0.0_f64, f64::max);
+    let sample_baseline = median(sample_abs.clone());
+    let sample_threshold = sample_baseline + ((sample_peak - sample_baseline) * 0.35);
+    let mut segments = Vec::new();
+    let mut previous_active = false;
+    let mut segment_start = 0usize;
+    for (index, level) in envelope.iter().copied().enumerate() {
+        let active = level >= threshold;
+        if active && !previous_active {
+            segment_start = index;
+        }
+        if previous_active && !active {
+            segments.push(window_segment(
+                samples,
+                sample_rate_hz,
+                window_samples,
+                segment_start,
+                index,
+                sample_threshold,
+            ));
+        }
+        previous_active = active;
+    }
+    if previous_active {
+        segments.push(window_segment(
+            samples,
+            sample_rate_hz,
+            window_samples,
+            segment_start,
+            envelope.len(),
+            sample_threshold,
+        ));
+    }
+
+    Ok(segments)
+}
+
+pub(super) fn edge_midpoint(previous_s: f64, current_s: f64) -> f64 {
+    previous_s + ((current_s - previous_s) / 2.0)
+}
+
+pub(super) fn window_segment(
+    samples: &[i16],
+    sample_rate_hz: u32,
+    window_samples: usize,
+    start_window_index: usize,
+    end_window_index_exclusive: usize,
+    sample_threshold: f64,
+) -> PulseSegment {
+    let start_sample = start_window_index.saturating_mul(window_samples);
+    let end_sample = end_window_index_exclusive
+        .saturating_mul(window_samples)
+        .min(samples.len());
+
+    let refined_start_sample = samples[start_sample..end_sample]
+        .iter()
+        .position(|sample| i32::from(*sample).unsigned_abs() as f64 >= sample_threshold)
+        .map(|offset| start_sample + offset)
+        .unwrap_or(start_sample);
+    let refined_end_sample = samples[start_sample..end_sample]
+        .iter()
+        .rposition(|sample| i32::from(*sample).unsigned_abs() as f64 >= sample_threshold)
+        .map(|offset| start_sample + offset + 1)
+        .unwrap_or(end_sample);
+
+    let start_s = refined_start_sample as f64 / f64::from(sample_rate_hz);
+    let end_s = refined_end_sample.max(refined_start_sample + 1) as f64 / f64::from(sample_rate_hz);
+    PulseSegment {
+        start_s,
+        end_s,
+        duration_s: end_s - start_s,
+    }
+}
+
+pub(super) fn median_frame_step_seconds(timestamps_s: &[f64]) -> f64 {
+    let diffs = timestamps_s
+        .windows(2)
+        .filter_map(|pair| {
+            let diff = pair[1] - pair[0];
+            (diff.is_finite() && diff > 0.0).then_some(diff)
+        })
+        .collect::<Vec<_>>();
+    median(diffs)
+}
+
+pub(super) fn median(mut values: Vec<f64>) -> f64 {
+    if values.is_empty() {
+        return 0.0;
+    }
+    values.sort_by(|left, right| left.total_cmp(right));
+    let mid = values.len() / 2;
+    if values.len() % 2 == 0 {
+        (values[mid - 1] + values[mid]) / 2.0
+    } else {
+        values[mid]
+    }
+}

client/src/sync_probe/analyze/onset_detection/correlation.rs

@@ -0,0 +1,310 @@
+use anyhow::{Result, bail};
+use std::collections::BTreeMap;
+
+use crate::sync_probe::analyze::report::SyncAnalysisReport;
+
+use super::{PulseSegment, median};
+
+const MARKER_WIDTH_MULTIPLIER: f64 = 1.5;
+
+#[cfg_attr(not(test), allow(dead_code))]
+pub(super) fn correlate_onsets(
+    video_onsets_s: &[f64],
+    audio_onsets_s: &[f64],
+    pulse_period_s: f64,
+    max_pair_gap_s: f64,
+) -> Result<SyncAnalysisReport> {
+    if video_onsets_s.is_empty() {
+        bail!("video onset list is empty");
+    }
+    if audio_onsets_s.is_empty() {
+        bail!("audio onset list is empty");
+    }
+    if max_pair_gap_s <= 0.0 {
+        bail!("max pair gap must stay positive");
+    }
+    if pulse_period_s <= 0.0 {
+        bail!("pulse period must stay positive");
+    }
+
+    let video_pulses = index_onsets_by_spacing(video_onsets_s, pulse_period_s);
+    let audio_pulses = index_onsets_by_spacing(audio_onsets_s, pulse_period_s);
+    let offset_candidates = candidate_index_offsets(&video_pulses, &audio_pulses);
+    let mut skews_ms = best_skews_for_index_offsets(
+        &video_pulses,
+        &audio_pulses,
+        &offset_candidates,
+        max_pair_gap_s,
+    );
+
+    if skews_ms.is_empty() && video_onsets_s.len() == 1 && audio_onsets_s.len() == 1 {
+        let video_phase_s = estimate_phase(video_onsets_s, pulse_period_s);
+        let audio_phase_s = estimate_phase(audio_onsets_s, pulse_period_s);
+        let phase_skew_ms =
+            shortest_wrapped_difference(audio_phase_s - video_phase_s, pulse_period_s) * 1000.0;
+        if phase_skew_ms.abs() <= max_pair_gap_s * 1000.0 {
+            skews_ms.push(phase_skew_ms);
+        }
+    }
+
+    if skews_ms.is_empty() {
+        bail!("no audio/video pulse pairs were close enough to compare");
+    }
+
+    Ok(sync_report_from_skews(
+        video_onsets_s,
+        audio_onsets_s,
+        skews_ms,
+    ))
+}
+
+pub(crate) fn correlate_segments(
+    video_segments: &[PulseSegment],
+    audio_segments: &[PulseSegment],
+    pulse_period_s: f64,
+    pulse_width_s: f64,
+    marker_tick_period: u32,
+    max_pair_gap_s: f64,
+) -> Result<SyncAnalysisReport> {
+    let video_onsets_s = video_segments
+        .iter()
+        .map(|segment| segment.start_s)
+        .collect::<Vec<_>>();
+    let audio_onsets_s = audio_segments
+        .iter()
+        .map(|segment| segment.start_s)
+        .collect::<Vec<_>>();
+    if video_onsets_s.is_empty() {
+        bail!("video onset list is empty");
+    }
+    if audio_onsets_s.is_empty() {
+        bail!("audio onset list is empty");
+    }
+    if pulse_period_s <= 0.0 {
+        bail!("pulse period must stay positive");
+    }
+    if pulse_width_s <= 0.0 {
+        bail!("pulse width must stay positive");
+    }
+    if marker_tick_period == 0 {
+        bail!("marker tick period must stay positive");
+    }
+    if max_pair_gap_s <= 0.0 {
+        bail!("max pair gap must stay positive");
+    }
+
+    let video_marker_onsets = marker_onsets(video_segments, pulse_width_s);
+    let audio_marker_onsets = marker_onsets(audio_segments, pulse_width_s);
+    let video_indexed = index_onsets_by_spacing(&video_onsets_s, pulse_period_s);
+    let audio_indexed = index_onsets_by_spacing(&audio_onsets_s, pulse_period_s);
+    let offset_candidates = marker_index_offsets(
+        &video_indexed,
+        &audio_indexed,
+        &video_marker_onsets,
+        &audio_marker_onsets,
+    );
+    let mut skews_ms = best_skews_for_index_offsets(
+        &video_indexed,
+        &audio_indexed,
+        &offset_candidates,
+        max_pair_gap_s,
+    );
+
+    if skews_ms.is_empty() && video_onsets_s.len() == 1 && audio_onsets_s.len() == 1 {
+        let video_phase_s = estimate_phase(&video_onsets_s, pulse_period_s);
+        let audio_phase_s = estimate_phase(&audio_onsets_s, pulse_period_s);
+        let phase_skew_ms =
+            shortest_wrapped_difference(audio_phase_s - video_phase_s, pulse_period_s) * 1000.0;
+        if phase_skew_ms.abs() <= max_pair_gap_s * 1000.0 {
+            skews_ms.push(phase_skew_ms);
+        }
+    }
+
+    if skews_ms.is_empty() {
+        bail!("no audio/video pulse pairs were close enough to compare");
+    }
+
+    Ok(sync_report_from_skews(
+        &video_onsets_s,
+        &audio_onsets_s,
+        skews_ms,
+    ))
+}
+
+pub(super) fn estimate_phase(onsets_s: &[f64], pulse_period_s: f64) -> f64 {
+    let (sum_sin, sum_cos) =
+        onsets_s
+            .iter()
+            .copied()
+            .fold((0.0_f64, 0.0_f64), |(sum_sin, sum_cos), onset| {
+                let wrapped = onset.rem_euclid(pulse_period_s);
+                let angle = (wrapped / pulse_period_s) * std::f64::consts::TAU;
+                (sum_sin + angle.sin(), sum_cos + angle.cos())
+            });
+
+    let mean_angle = sum_sin.atan2(sum_cos).rem_euclid(std::f64::consts::TAU);
+    (mean_angle / std::f64::consts::TAU) * pulse_period_s
+}
+
+pub(super) fn index_onsets_by_spacing(onsets_s: &[f64], pulse_period_s: f64) -> BTreeMap<i64, f64> {
+    let mut indexed = BTreeMap::new();
+    let Some(first_onset) = onsets_s.first().copied() else {
+        return indexed;
+    };
+
+    let mut pulse_index = 0_i64;
+    let mut previous_onset = first_onset;
+    indexed.insert(pulse_index, first_onset);
+    for onset in onsets_s.iter().copied().skip(1) {
+        let pulse_steps = ((onset - previous_onset) / pulse_period_s).round().max(1.0) as i64;
+        pulse_index += pulse_steps;
+        indexed.insert(pulse_index, onset);
+        previous_onset = onset;
+    }
+    indexed
+}
+
+pub(super) fn candidate_index_offsets(
+    video_indexed: &BTreeMap<i64, f64>,
+    audio_indexed: &BTreeMap<i64, f64>,
+) -> Vec<i64> {
+    if video_indexed.is_empty() || audio_indexed.is_empty() {
+        return Vec::new();
+    }
+
+    let video_min = *video_indexed
+        .keys()
+        .next()
+        .expect("non-empty indexed video onset map has a first key");
+    let video_max = *video_indexed
+        .keys()
+        .next_back()
+        .expect("non-empty indexed video onset map has a last key");
+    let audio_min = *audio_indexed
+        .keys()
+        .next()
+        .expect("non-empty indexed audio onset map has a first key");
+    let audio_max = *audio_indexed
+        .keys()
+        .next_back()
+        .expect("non-empty indexed audio onset map has a last key");
+
+    (audio_min - video_max..=audio_max - video_min).collect()
+}
+
+pub(super) fn marker_index_offsets(
+    video_indexed: &BTreeMap<i64, f64>,
+    audio_indexed: &BTreeMap<i64, f64>,
+    video_marker_onsets: &[f64],
+    audio_marker_onsets: &[f64],
+) -> Vec<i64> {
+    let mut offsets = Vec::new();
+    if !video_marker_onsets.is_empty() && !audio_marker_onsets.is_empty() {
+        let video_markers = pulse_indices_for_onsets(video_indexed, video_marker_onsets);
+        let audio_markers = pulse_indices_for_onsets(audio_indexed, audio_marker_onsets);
+        for video_marker in &video_markers {
+            for audio_marker in &audio_markers {
+                offsets.push(audio_marker - video_marker);
+            }
+        }
+    }
+
+    offsets.extend(candidate_index_offsets(video_indexed, audio_indexed));
+    offsets.sort_unstable();
+    offsets.dedup();
+    offsets
+}
+
+fn pulse_indices_for_onsets(indexed: &BTreeMap<i64, f64>, marker_onsets: &[f64]) -> Vec<i64> {
+    marker_onsets
+        .iter()
+        .filter_map(|marker_onset| {
+            indexed.iter().find_map(|(pulse_index, onset)| {
+                ((onset - marker_onset).abs() < 0.000_001).then_some(*pulse_index)
+            })
+        })
+        .collect()
+}
+
+fn best_skews_for_index_offsets(
+    video_indexed: &BTreeMap<i64, f64>,
+    audio_indexed: &BTreeMap<i64, f64>,
+    offset_candidates: &[i64],
+    max_pair_gap_s: f64,
+) -> Vec<f64> {
+    let max_pair_gap_ms = max_pair_gap_s * 1000.0;
+    let mut best: Option<(usize, f64, Vec<f64>)> = None;
+
+    for offset in offset_candidates.iter().copied() {
+        let skews_ms = video_indexed
+            .iter()
+            .filter_map(|(pulse_index, video_time)| {
+                audio_indexed
+                    .get(&(pulse_index + offset))
+                    .map(|audio_time| (audio_time - video_time) * 1000.0)
+            })
+            .filter(|skew_ms| skew_ms.abs() <= max_pair_gap_ms)
+            .collect::<Vec<_>>();
+        if skews_ms.is_empty() {
+            continue;
+        }
+
+        let score =
+            skews_ms.iter().map(|skew_ms| skew_ms.abs()).sum::<f64>() / skews_ms.len() as f64;
+        match &best {
+            Some((best_count, best_score, _))
+                if skews_ms.len() < *best_count
+                    || (skews_ms.len() == *best_count && score >= *best_score) => {}
+            _ => best = Some((skews_ms.len(), score, skews_ms)),
+        }
+    }
+
+    best.map(|(_, _, skews)| skews).unwrap_or_default()
+}
+
+pub(super) fn marker_onsets(segments: &[PulseSegment], pulse_width_s: f64) -> Vec<f64> {
+    let threshold = pulse_width_s * MARKER_WIDTH_MULTIPLIER;
+    segments
+        .iter()
+        .filter(|segment| segment.duration_s >= threshold)
+        .map(|segment| segment.start_s)
+        .collect()
+}
+
+pub(super) fn shortest_wrapped_difference(delta_s: f64, pulse_period_s: f64) -> f64 {
+    let half_period = pulse_period_s / 2.0;
+    ((delta_s + half_period).rem_euclid(pulse_period_s)) - half_period
+}
+
+fn sync_report_from_skews(
+    video_onsets_s: &[f64],
+    audio_onsets_s: &[f64],
+    skews_ms: Vec<f64>,
+) -> SyncAnalysisReport {
+    let mut sorted_skews = skews_ms.clone();
+    sorted_skews.sort_by(|left, right| left.total_cmp(right));
+    let first_skew_ms = *skews_ms.first().expect("paired skew list is not empty");
+    let last_skew_ms = *skews_ms.last().expect("paired skew list is not empty");
+    let mean_skew_ms = skews_ms.iter().sum::<f64>() / skews_ms.len() as f64;
+    let median_skew_ms = median(sorted_skews);
+    let max_abs_skew_ms = skews_ms
+        .iter()
+        .copied()
+        .map(f64::abs)
+        .fold(0.0_f64, f64::max);
+
+    SyncAnalysisReport {
+        video_event_count: video_onsets_s.len(),
+        audio_event_count: audio_onsets_s.len(),
+        paired_event_count: skews_ms.len(),
+        first_skew_ms,
+        last_skew_ms,
+        mean_skew_ms,
+        median_skew_ms,
+        max_abs_skew_ms,
+        drift_ms: last_skew_ms - first_skew_ms,
+        skews_ms,
+        video_onsets_s: video_onsets_s.to_vec(),
+        audio_onsets_s: audio_onsets_s.to_vec(),
+    }
+}

client/src/sync_probe/analyze/onset_detection/tests.rs

@@ -0,0 +1,309 @@
+use super::correlation::{
+    candidate_index_offsets, correlate_onsets, estimate_phase, index_onsets_by_spacing,
+    marker_index_offsets, marker_onsets, shortest_wrapped_difference,
+};
+use super::{
+    PulseSegment, correlate_segments, detect_audio_onsets, detect_audio_segments,
+    detect_video_onsets, detect_video_segments, median,
+};
+use crate::sync_probe::analyze::report::SyncAnalysisReport;
+use std::collections::BTreeMap;
+
+#[test]
+fn detect_video_onsets_finds_bright_transitions() {
+    let timestamps = (0..60).map(|idx| idx as f64 / 10.0).collect::<Vec<_>>();
+    let brightness = timestamps
+        .iter()
+        .enumerate()
+        .map(|(idx, _)| {
+            if idx == 0 || idx == 10 || idx == 20 {
+                250
+            } else {
+                5
+            }
+        })
+        .collect::<Vec<_>>();
+
+    let onsets = detect_video_onsets(&timestamps, &brightness).expect("video onsets");
+    assert_eq!(onsets, vec![0.0, 0.95, 1.95]);
+}
+
+#[test]
+fn detect_audio_onsets_finds_click_bursts() {
+    let mut samples = vec![0i16; 48_000];
+    for start in [0usize, 48_000 / 2] {
+        for sample in samples.iter_mut().skip(start).take(300) {
+            *sample = 18_000;
+        }
+    }
+
+    let onsets = detect_audio_onsets(&samples, 48_000, 5).expect("audio onsets");
+    assert_eq!(onsets.len(), 2);
+    assert!((onsets[0] - 0.0).abs() < 0.01);
+    assert!((onsets[1] - 0.5).abs() < 0.02);
+}
+
+#[test]
+fn detect_video_segments_keeps_regular_and_marker_durations_distinct() {
+    let timestamps = (0..30).map(|idx| idx as f64 / 30.0).collect::<Vec<_>>();
+    let brightness = [
+        0, 255, 255, 255, 0, 0, 0, 0, 0, 0, 255, 255, 255, 255, 255, 255, 0, 0, 0, 0, 0, 0, 0, 0,
+        0, 0, 0, 0, 0, 0,
+    ];
+    let segments = detect_video_segments(&timestamps, &brightness).expect("video segments");
+    assert_eq!(segments.len(), 2);
+    assert!(segments[1].duration_s > segments[0].duration_s);
+}
+
+#[test]
+fn detect_audio_segments_keeps_regular_and_marker_durations_distinct() {
+    let mut samples = vec![0i16; 48_000];
+    for sample in samples.iter_mut().take(3_000) {
+        *sample = 18_000;
+    }
+    for sample in samples.iter_mut().skip(24_000).take(6_000) {
+        *sample = 18_000;
+    }
+    let segments = detect_audio_segments(&samples, 48_000, 5).expect("audio segments");
+    assert_eq!(segments.len(), 2);
+    assert!(segments[1].duration_s > segments[0].duration_s);
+}
+
+#[test]
+fn detect_video_segments_closes_a_pulse_that_stays_active_until_the_last_frame() {
+    let timestamps = [0.0, 0.1, 0.2, 0.3];
+    let brightness = [0, 0, 255, 255];
+    let segments = detect_video_segments(&timestamps, &brightness).expect("trailing video segment");
+    assert_eq!(segments.len(), 1);
+    assert!(segments[0].end_s > segments[0].start_s);
+    assert!(segments[0].end_s >= 0.3);
+}
+
+#[test]
+fn detect_audio_segments_closes_a_click_that_stays_active_until_the_capture_ends() {
+    let mut samples = vec![0i16; 4_800];
+    let midpoint = samples.len() / 2;
+    for sample in samples.iter_mut().skip(midpoint) {
+        *sample = 18_000;
+    }
+    let segments = detect_audio_segments(&samples, 48_000, 5).expect("trailing audio segment");
+    assert_eq!(segments.len(), 1);
+    assert!(segments[0].end_s > segments[0].start_s);
+}
+
+#[test]
+fn correlate_onsets_reports_skew_and_drift() {
+    let report = correlate_onsets(&[0.0, 1.0, 2.0, 3.0], &[0.05, 1.04, 2.03, 3.02], 1.0, 0.2)
+        .expect("correlated report");
+
+    assert_sync_report_shape(&report, 4);
+    assert!((report.first_skew_ms - 50.0).abs() < 0.001);
+    assert!((report.last_skew_ms - 20.0).abs() < 0.001);
+    assert!((report.drift_ms + 30.0).abs() < 0.001);
+    assert!(report.max_abs_skew_ms >= 50.0);
+}
+
+#[test]
+fn correlate_onsets_single_pulse_uses_phase_fallback() {
+    let report = correlate_onsets(&[0.95], &[0.05], 1.0, 0.2).expect("single-pulse fallback");
+    assert_eq!(report.paired_event_count, 1);
+    assert!((report.first_skew_ms - 100.0).abs() < 0.001);
+}
+
+#[test]
+fn detect_video_onsets_rejects_empty_low_contrast_and_missing_edges() {
+    assert!(detect_video_onsets(&[], &[]).is_err());
+    assert!(detect_video_onsets(&[0.0, 0.1], &[10, 12]).is_err());
+    assert!(detect_video_onsets(&[0.0, 0.1, 0.2], &[255, 255, 255]).is_err());
+}
+
+#[test]
+fn detect_audio_onsets_rejects_empty_invalid_and_too_quiet_inputs() {
+    assert!(detect_audio_onsets(&[], 48_000, 5).is_err());
+    assert!(detect_audio_onsets(&[1, 2, 3], 0, 5).is_err());
+    assert!(detect_audio_onsets(&[1, 2, 3], 48_000, 0).is_err());
+    assert!(detect_audio_onsets(&vec![1i16; 4_800], 48_000, 5).is_err());
+}
+
+#[test]
+fn correlate_onsets_rejects_empty_inputs_invalid_gap_and_unpairable_events() {
+    assert!(correlate_onsets(&[], &[0.0], 1.0, 0.2).is_err());
+    assert!(correlate_onsets(&[0.0], &[], 1.0, 0.2).is_err());
+    assert!(correlate_onsets(&[0.0], &[0.0], 1.0, 0.0).is_err());
+    assert!(correlate_onsets(&[0.0, 1.0], &[2.0, 3.0], 1.0, 0.1).is_err());
+    assert!(correlate_onsets(&[0.0], &[0.0], 0.0, 0.1).is_err());
+}
+
+#[test]
+fn correlate_segments_validate_inputs_and_support_single_pulse_fallback() {
+    let video = [PulseSegment {
+        start_s: 0.95,
+        end_s: 1.05,
+        duration_s: 0.1,
+    }];
+    let audio = [PulseSegment {
+        start_s: 0.05,
+        end_s: 0.15,
+        duration_s: 0.1,
+    }];
+    let report =
+        correlate_segments(&video, &audio, 1.0, 0.1, 3, 0.2).expect("single segment fallback");
+    assert_eq!(report.paired_event_count, 1);
+    assert!((report.first_skew_ms - 100.0).abs() < 0.001);
+
+    assert!(correlate_segments(&[], &audio, 1.0, 0.1, 3, 0.2).is_err());
+    assert!(correlate_segments(&video, &[], 1.0, 0.1, 3, 0.2).is_err());
+    assert!(correlate_segments(&video, &audio, 0.0, 0.1, 3, 0.2).is_err());
+    assert!(correlate_segments(&video, &audio, 1.0, 0.0, 3, 0.2).is_err());
+    assert!(correlate_segments(&video, &audio, 1.0, 0.1, 0, 0.2).is_err());
+    assert!(correlate_segments(&video, &audio, 1.0, 0.1, 3, 0.0).is_err());
+    assert!(correlate_segments(&video, &audio, 1.0, 0.1, 3, 0.05).is_err());
+}
+
+#[test]
+fn phase_estimation_and_indexing_stay_stable_when_pulses_are_missing() {
+    let video_phase = estimate_phase(&[4.0, 5.0, 7.0, 8.0, 10.0], 1.0);
+    let audio_phase = estimate_phase(&[4.018, 5.017, 6.019, 8.018, 9.018], 1.0);
+    assert!((video_phase - 0.0).abs() < 0.02);
+    assert!((audio_phase - 0.018).abs() < 0.02);
+
+    let video_indexed = index_onsets_by_spacing(&[4.0, 5.0, 7.0, 8.0, 10.0], 1.0);
+    let audio_indexed = index_onsets_by_spacing(&[4.018, 5.017, 6.019, 8.018, 9.018], 1.0);
+    assert_eq!(
+        video_indexed.keys().copied().collect::<Vec<_>>(),
+        vec![0, 1, 3, 4, 6]
+    );
+    assert_eq!(
+        audio_indexed.keys().copied().collect::<Vec<_>>(),
+        vec![0, 1, 2, 4, 5]
+    );
+}
+
+#[test]
+fn correlation_helpers_cover_empty_index_sets_and_wrapped_phase_math() {
+    assert!(index_onsets_by_spacing(&[], 1.0).is_empty());
+    assert!(candidate_index_offsets(&BTreeMap::new(), &BTreeMap::new()).is_empty());
+
+    let mut video_only = BTreeMap::new();
+    video_only.insert(0, 1.0);
+    assert!(candidate_index_offsets(&video_only, &BTreeMap::new()).is_empty());
+
+    let mut audio_only = BTreeMap::new();
+    audio_only.insert(0, 1.0);
+    assert!(candidate_index_offsets(&BTreeMap::new(), &audio_only).is_empty());
+
+    let mut video_indexed = BTreeMap::new();
+    video_indexed.insert(2, 2.0);
+    let mut audio_indexed = BTreeMap::new();
+    audio_indexed.insert(5, 5.0);
+    assert_eq!(
+        candidate_index_offsets(&video_indexed, &audio_indexed),
+        vec![3]
+    );
+
+    assert!((shortest_wrapped_difference(0.6, 1.0) + 0.4).abs() < 0.000_001);
+    assert!((shortest_wrapped_difference(-0.6, 1.0) - 0.4).abs() < 0.000_001);
+}
+
+#[test]
+fn marker_index_offsets_include_marker_alignment_and_general_fallback() {
+    let video_indexed = index_onsets_by_spacing(&[4.0, 5.0, 7.0, 8.0, 10.0], 1.0);
+    let audio_indexed = index_onsets_by_spacing(&[5.018, 6.017, 7.019, 9.018, 10.018], 1.0);
+    let offsets = marker_index_offsets(&video_indexed, &audio_indexed, &[10.0], &[10.018]);
+    assert!(offsets.contains(&1));
+    assert!(offsets.contains(&0));
+}
+
+#[test]
+fn correlate_onsets_ignores_missing_pulses_and_preserves_stable_skew() {
+    let report = correlate_onsets(
+        &[4.0, 5.0, 7.0, 8.0, 10.0],
+        &[4.018, 5.017, 6.019, 8.018, 9.018],
+        1.0,
+        0.2,
+    )
+    .expect("correlated report");
+
+    assert_eq!(report.paired_event_count, 3);
+    assert!((report.mean_skew_ms - 17.666).abs() < 5.0);
+    assert!(report.max_abs_skew_ms < 30.0);
+}
+
+#[test]
+fn correlate_segments_uses_markers_to_break_period_aliasing() {
+    let video = vec![
+        PulseSegment {
+            start_s: 3.3,
+            end_s: 3.55,
+            duration_s: 0.25,
+        },
+        PulseSegment {
+            start_s: 4.266667,
+            end_s: 4.4,
+            duration_s: 0.133333,
+        },
+        PulseSegment {
+            start_s: 5.3,
+            end_s: 5.433333,
+            duration_s: 0.133333,
+        },
+    ];
+    let audio = vec![
+        PulseSegment {
+            start_s: 3.35,
+            end_s: 3.59,
+            duration_s: 0.24,
+        },
+        PulseSegment {
+            start_s: 4.316667,
+            end_s: 4.436667,
+            duration_s: 0.12,
+        },
+        PulseSegment {
+            start_s: 5.35,
+            end_s: 5.47,
+            duration_s: 0.12,
+        },
+    ];
+
+    let report =
+        correlate_segments(&video, &audio, 1.0, 0.12, 5, 0.2).expect("marker-correlated report");
+    assert_eq!(report.paired_event_count, 3);
+    assert!((report.mean_skew_ms - 50.0).abs() < 10.0);
+}
+
+#[test]
+fn marker_detection_finds_wider_segments_only() {
+    let markers = marker_onsets(
+        &[
+            PulseSegment {
+                start_s: 1.0,
+                end_s: 1.12,
+                duration_s: 0.12,
+            },
+            PulseSegment {
+                start_s: 5.0,
+                end_s: 5.24,
+                duration_s: 0.24,
+            },
+        ],
+        0.12,
+    );
+    assert_eq!(markers, vec![5.0]);
+}
+
+#[test]
+fn median_handles_empty_even_and_odd_inputs() {
+    assert_eq!(median(Vec::new()), 0.0);
+    assert_eq!(median(vec![1.0, 3.0, 2.0]), 2.0);
+    assert_eq!(median(vec![4.0, 1.0, 3.0, 2.0]), 2.5);
+}
+
+fn assert_sync_report_shape(report: &SyncAnalysisReport, paired_events: usize) {
+    assert_eq!(report.video_event_count, paired_events);
+    assert_eq!(report.audio_event_count, paired_events);
+    assert_eq!(report.paired_event_count, paired_events);
+    assert_eq!(report.skews_ms.len(), paired_events);
+    assert_eq!(report.video_onsets_s.len(), paired_events);
+    assert_eq!(report.audio_onsets_s.len(), paired_events);
+}

client/src/sync_probe/analyze/report.rs

@@ -0,0 +1,59 @@
+use serde::Serialize;
+
+const DEFAULT_AUDIO_WINDOW_MS: u32 = 5;
+const DEFAULT_MAX_PAIR_GAP_S: f64 = 0.5;
+const DEFAULT_PULSE_PERIOD_S: f64 = 1.0;
+const DEFAULT_PULSE_WIDTH_S: f64 = 0.12;
+const DEFAULT_MARKER_TICK_PERIOD: u32 = 5;
+
+#[derive(Clone, Debug, PartialEq, Serialize)]
+pub struct SyncAnalysisReport {
+    pub video_event_count: usize,
+    pub audio_event_count: usize,
+    pub paired_event_count: usize,
+    pub first_skew_ms: f64,
+    pub last_skew_ms: f64,
+    pub mean_skew_ms: f64,
+    pub median_skew_ms: f64,
+    pub max_abs_skew_ms: f64,
+    pub drift_ms: f64,
+    pub skews_ms: Vec<f64>,
+    pub video_onsets_s: Vec<f64>,
+    pub audio_onsets_s: Vec<f64>,
+}
+
+#[derive(Clone, Debug, PartialEq)]
+pub struct SyncAnalysisOptions {
+    pub audio_window_ms: u32,
+    pub max_pair_gap_s: f64,
+    pub pulse_period_s: f64,
+    pub pulse_width_s: f64,
+    pub marker_tick_period: u32,
+}
+
+impl Default for SyncAnalysisOptions {
+    fn default() -> Self {
+        Self {
+            audio_window_ms: DEFAULT_AUDIO_WINDOW_MS,
+            max_pair_gap_s: DEFAULT_MAX_PAIR_GAP_S,
+            pulse_period_s: DEFAULT_PULSE_PERIOD_S,
+            pulse_width_s: DEFAULT_PULSE_WIDTH_S,
+            marker_tick_period: DEFAULT_MARKER_TICK_PERIOD,
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::SyncAnalysisOptions;
+
+    #[test]
+    fn default_options_match_live_probe_expectations() {
+        let options = SyncAnalysisOptions::default();
+        assert_eq!(options.audio_window_ms, 5);
+        assert!((options.max_pair_gap_s - 0.5).abs() < f64::EPSILON);
+        assert!((options.pulse_period_s - 1.0).abs() < f64::EPSILON);
+        assert!((options.pulse_width_s - 0.12).abs() < f64::EPSILON);
+        assert_eq!(options.marker_tick_period, 5);
+    }
+}

client/src/sync_probe/analyze/test_support.rs

@@ -0,0 +1,85 @@
+use std::env;
+use std::fs;
+use std::os::unix::fs::PermissionsExt;
+use std::path::Path;
+
+use temp_env::with_var;
+use tempfile::tempdir;
+
+pub(super) fn with_fake_media_tools<T>(
+    ffprobe_output: &[u8],
+    ffmpeg_video_output: &[u8],
+    ffmpeg_audio_output: &[u8],
+    test: impl FnOnce(&Path) -> T,
+) -> T {
+    let temp_dir = tempdir().expect("tempdir");
+    fs::write(temp_dir.path().join("ffprobe.out"), ffprobe_output).expect("write ffprobe");
+    fs::write(
+        temp_dir.path().join("ffmpeg-video.out"),
+        ffmpeg_video_output,
+    )
+    .expect("write ffmpeg video");
+    fs::write(
+        temp_dir.path().join("ffmpeg-audio.out"),
+        ffmpeg_audio_output,
+    )
+    .expect("write ffmpeg audio");
+    write_executable(
+        temp_dir.path(),
+        "ffprobe",
+        "#!/bin/sh\ncat \"$(dirname \"$0\")/ffprobe.out\"\n",
+    );
+    write_executable(
+        temp_dir.path(),
+        "ffmpeg",
+        "#!/bin/sh\ncase \" $* \" in\n  *\" -map 0:v:0 \"*) cat \"$(dirname \"$0\")/ffmpeg-video.out\" ;;\n  *\" -map 0:a:0 \"*) cat \"$(dirname \"$0\")/ffmpeg-audio.out\" ;;\n  *) printf 'unexpected ffmpeg args: %s\\n' \"$*\" >&2; exit 64 ;;\nesac\n",
+    );
+
+    let prior_path = env::var("PATH").unwrap_or_default();
+    let merged_path = if prior_path.is_empty() {
+        temp_dir.path().display().to_string()
+    } else {
+        format!("{}:{prior_path}", temp_dir.path().display())
+    };
+    let capture_path = temp_dir.path().join("capture.mkv");
+    fs::write(&capture_path, b"fake-capture").expect("write capture");
+    with_var("PATH", Some(merged_path.as_str()), || test(&capture_path))
+}
+
+pub(super) fn frame_json(timestamps: &[f64]) -> Vec<u8> {
+    let frames = timestamps
+        .iter()
+        .map(|timestamp| {
+            serde_json::json!({
+                "best_effort_timestamp_time": format!("{timestamp:.3}")
+            })
+        })
+        .collect::<Vec<_>>();
+    serde_json::to_vec(&serde_json::json!({ "frames": frames })).expect("frame json")
+}
+
+pub(super) fn click_track_samples(click_times_s: &[f64], total_samples: usize) -> Vec<i16> {
+    let mut samples = vec![0i16; total_samples];
+    for click_time_s in click_times_s {
+        let start = (*click_time_s * 48_000.0).round() as usize;
+        for sample in samples.iter_mut().skip(start).take(300) {
+            *sample = 18_000;
+        }
+    }
+    samples
+}
+
+pub(super) fn audio_samples_to_bytes(samples: &[i16]) -> Vec<u8> {
+    samples
+        .iter()
+        .flat_map(|sample| sample.to_le_bytes())
+        .collect()
+}
+
+fn write_executable(dir: &Path, name: &str, contents: &str) {
+    let path = dir.join(name);
+    fs::write(&path, contents).expect("write script");
+    let mut permissions = fs::metadata(&path).expect("script metadata").permissions();
+    permissions.set_mode(0o755);
+    fs::set_permissions(&path, permissions).expect("script permissions");
+}

client/src/sync_probe/capture.rs

@@ -0,0 +1,165 @@
+//! Shared-clock synthetic A/V source for the upstream sync probe.
+
+#[cfg(any(not(coverage), test))]
+use anyhow::{Context, Result, bail};
+#[cfg(any(not(coverage), test))]
+use gst::prelude::*;
+#[cfg(any(not(coverage), test))]
+use gstreamer as gst;
+#[cfg(any(not(coverage), test))]
+use gstreamer_app as gst_app;
+#[cfg(any(not(coverage), test))]
+use lesavka_common::lesavka::{AudioPacket, VideoPacket};
+#[cfg(any(not(coverage), test))]
+use std::sync::{
+    Arc,
+    atomic::{AtomicBool, Ordering},
+};
+#[cfg(any(not(coverage), test))]
+use std::thread::{self, JoinHandle};
+#[cfg(any(not(coverage), test))]
+use std::time::{Duration, Instant};
+#[cfg(any(not(coverage), test))]
+use std::{f64::consts::TAU, mem::size_of};
+
+#[cfg(any(not(coverage), test))]
+use crate::input::camera::{CameraCodec, CameraConfig};
+#[cfg(any(not(coverage), test))]
+use crate::sync_probe::schedule::PulseSchedule;
+#[cfg(any(not(coverage), test))]
+use crate::uplink_fresh_queue::{FreshPacketQueue, FreshQueueConfig};
+
+#[cfg(coverage)]
+mod coverage_stub;
+#[cfg(not(coverage))]
+mod runtime;
+#[cfg(test)]
+mod tests;
+
+#[cfg(coverage)]
+pub use coverage_stub::SyncProbeCapture;
+#[cfg(not(coverage))]
+pub use runtime::SyncProbeCapture;
+
+#[cfg(any(not(coverage), test))]
+const PROBE_VIDEO_QUEUE: FreshQueueConfig = FreshQueueConfig {
+    capacity: 8,
+    max_age: Duration::from_millis(350),
+};
+
+#[cfg(any(not(coverage), test))]
+const PROBE_AUDIO_QUEUE: FreshQueueConfig = FreshQueueConfig {
+    capacity: 32,
+    max_age: Duration::from_millis(400),
+};
+
+#[cfg(any(not(coverage), test))]
+const AUDIO_SAMPLE_RATE: i32 = 48_000;
+#[cfg(any(not(coverage), test))]
+const AUDIO_CHANNELS: usize = 2;
+#[cfg(any(not(coverage), test))]
+const AUDIO_CHUNK_MS: u64 = 10;
+#[cfg(any(not(coverage), test))]
+const AUDIO_PULSE_FREQUENCY_HZ: f64 = 1_800.0;
+#[cfg(any(not(coverage), test))]
+const AUDIO_PULSE_AMPLITUDE: f64 = 24_000.0;
+
+#[cfg(any(not(coverage), test))]
+fn build_dark_probe_frame(width: usize, height: usize) -> Vec<u8> {
+    vec![16u8; width.saturating_mul(height).saturating_mul(3)]
+}
+
+#[cfg(any(not(coverage), test))]
+fn build_regular_probe_frame(width: usize, height: usize) -> Vec<u8> {
+    let mut frame = build_dark_probe_frame(width, height);
+    let x0 = width / 4;
+    let x1 = width.saturating_sub(x0);
+    let y0 = height / 4;
+    let y1 = height.saturating_sub(y0);
+    fill_rect(&mut frame, width, x0, y0, x1, y1, 255);
+    frame
+}
+
+#[cfg(any(not(coverage), test))]
+fn build_marker_probe_frame(width: usize, height: usize) -> Vec<u8> {
+    let mut frame = build_dark_probe_frame(width, height);
+    let x0 = width / 5;
+    let x1 = width.saturating_sub(x0);
+    let y0 = height / 5;
+    let y1 = height.saturating_sub(y0);
+    fill_rect(&mut frame, width, x0, y0, x1, y1, 255);
+
+    let cross_half_w = (width / 48).max(6);
+    let cross_half_h = (height / 48).max(6);
+    let cx = width / 2;
+    let cy = height / 2;
+    fill_rect(
+        &mut frame,
+        width,
+        cx.saturating_sub(cross_half_w),
+        y0,
+        (cx + cross_half_w).min(width),
+        y1,
+        255,
+    );
+    fill_rect(
+        &mut frame,
+        width,
+        x0,
+        cy.saturating_sub(cross_half_h),
+        x1,
+        (cy + cross_half_h).min(height),
+        255,
+    );
+    frame
+}
+
+#[cfg(any(not(coverage), test))]
+fn fill_rect(
+    frame: &mut [u8],
+    width: usize,
+    x0: usize,
+    y0: usize,
+    x1: usize,
+    y1: usize,
+    value: u8,
+) {
+    let height = frame.len() / width.saturating_mul(3);
+    let x1 = x1.min(width);
+    let y1 = y1.min(height);
+    for y in y0.min(height)..y1 {
+        for x in x0.min(width)..x1 {
+            let offset = (y * width + x) * 3;
+            frame[offset] = value;
+            frame[offset + 1] = value;
+            frame[offset + 2] = value;
+        }
+    }
+}
+
+#[cfg(any(not(coverage), test))]
+fn render_audio_chunk(
+    schedule: &PulseSchedule,
+    chunk_pts: Duration,
+    samples_per_chunk: usize,
+) -> Vec<u8> {
+    let sample_step = Duration::from_nanos(1_000_000_000u64 / AUDIO_SAMPLE_RATE as u64);
+    let mut pcm = Vec::with_capacity(samples_per_chunk * AUDIO_CHANNELS * size_of::<i16>());
+    for sample_index in 0..samples_per_chunk {
+        let sample_pts = chunk_pts + sample_step.saturating_mul(sample_index as u32);
+        let amplitude = if schedule.flash_active(sample_pts) {
+            let phase = TAU * AUDIO_PULSE_FREQUENCY_HZ * sample_pts.as_secs_f64();
+            (phase.sin() * AUDIO_PULSE_AMPLITUDE) as i16
+        } else {
+            0
+        };
+        for _ in 0..AUDIO_CHANNELS {
+            pcm.extend_from_slice(&amplitude.to_le_bytes());
+        }
+    }
+    pcm
+}
+
+fn probe_pts_exceeds_duration(pts_usecs: u64, duration: std::time::Duration) -> bool {
+    pts_usecs > duration.as_micros() as u64
+}

client/src/sync_probe/capture/coverage_stub.rs

@@ -0,0 +1,34 @@
+#[derive(Default)]
+pub struct SyncProbeCapture;
+
+impl SyncProbeCapture {
+    pub fn new(
+        _camera: crate::input::camera::CameraConfig,
+        _schedule: crate::sync_probe::schedule::PulseSchedule,
+        _duration: std::time::Duration,
+    ) -> anyhow::Result<Self> {
+        Ok(Self)
+    }
+
+    pub fn video_queue(
+        &self,
+    ) -> crate::uplink_fresh_queue::FreshPacketQueue<lesavka_common::lesavka::VideoPacket> {
+        crate::uplink_fresh_queue::FreshPacketQueue::new(
+            crate::uplink_fresh_queue::FreshQueueConfig {
+                capacity: 1,
+                max_age: std::time::Duration::from_millis(1),
+            },
+        )
+    }
+
+    pub fn audio_queue(
+        &self,
+    ) -> crate::uplink_fresh_queue::FreshPacketQueue<lesavka_common::lesavka::AudioPacket> {
+        crate::uplink_fresh_queue::FreshPacketQueue::new(
+            crate::uplink_fresh_queue::FreshQueueConfig {
+                capacity: 1,
+                max_age: std::time::Duration::from_millis(1),
+            },
+        )
+    }
+}

client/src/sync_probe/capture/runtime.rs

@@ -0,0 +1,330 @@
+use super::*;
+
+pub struct SyncProbeCapture {
+    pipeline: gst::Pipeline,
+    running: Arc<AtomicBool>,
+    video_queue: FreshPacketQueue<VideoPacket>,
+    audio_queue: FreshPacketQueue<AudioPacket>,
+    video_thread: Option<JoinHandle<()>>,
+    audio_thread: Option<JoinHandle<()>>,
+}
+
+impl SyncProbeCapture {
+    pub fn new(camera: CameraConfig, schedule: PulseSchedule, duration: Duration) -> Result<Self> {
+        gst::init().context("gst init")?;
+
+        let pipeline = build_pipeline(camera, &schedule)?;
+        let video_src = pipeline
+            .by_name("sync_probe_video_src")
+            .context("missing sync probe video appsrc")?
+            .downcast::<gst_app::AppSrc>()
+            .expect("video appsrc");
+        let video_sink = pipeline
+            .by_name("sync_probe_video_sink")
+            .context("missing sync probe video appsink")?
+            .downcast::<gst_app::AppSink>()
+            .expect("video appsink");
+        let audio_src = pipeline
+            .by_name("sync_probe_audio_src")
+            .context("missing sync probe audio appsrc")?
+            .downcast::<gst_app::AppSrc>()
+            .expect("audio appsrc");
+        let audio_sink = pipeline
+            .by_name("sync_probe_audio_sink")
+            .context("missing sync probe audio appsink")?
+            .downcast::<gst_app::AppSink>()
+            .expect("audio appsink");
+
+        pipeline
+            .set_state(gst::State::Playing)
+            .context("starting sync probe pipeline")?;
+
+        let running = Arc::new(AtomicBool::new(true));
+        let probe_start = Instant::now();
+        let video_queue = FreshPacketQueue::new(PROBE_VIDEO_QUEUE);
+        let audio_queue = FreshPacketQueue::new(PROBE_AUDIO_QUEUE);
+
+        let video_thread = spawn_video_thread(
+            video_src,
+            video_sink,
+            camera,
+            schedule.clone(),
+            duration,
+            probe_start,
+            running.clone(),
+            video_queue.clone(),
+        );
+        let audio_thread = spawn_audio_thread(
+            audio_src,
+            audio_sink,
+            schedule,
+            duration,
+            probe_start,
+            running.clone(),
+            audio_queue.clone(),
+        );
+
+        Ok(Self {
+            pipeline,
+            running,
+            video_queue,
+            audio_queue,
+            video_thread: Some(video_thread),
+            audio_thread: Some(audio_thread),
+        })
+    }
+
+    pub fn video_queue(&self) -> FreshPacketQueue<VideoPacket> {
+        self.video_queue.clone()
+    }
+
+    pub fn audio_queue(&self) -> FreshPacketQueue<AudioPacket> {
+        self.audio_queue.clone()
+    }
+}
+
+impl Drop for SyncProbeCapture {
+    fn drop(&mut self) {
+        self.running.store(false, Ordering::Release);
+        self.video_queue.close();
+        self.audio_queue.close();
+        let _ = self.pipeline.set_state(gst::State::Null);
+        if let Some(handle) = self.video_thread.take() {
+            let _ = handle.join();
+        }
+        if let Some(handle) = self.audio_thread.take() {
+            let _ = handle.join();
+        }
+    }
+}
+
+fn build_pipeline(camera: CameraConfig, _schedule: &PulseSchedule) -> Result<gst::Pipeline> {
+    let video_caps = format!(
+        "video/x-raw,format=RGB,width={},height={},framerate={}/1",
+        camera.width,
+        camera.height,
+        camera.fps.max(1)
+    );
+    let video_branch = match camera.codec {
+        CameraCodec::Mjpeg => format!(
+            "appsrc name=sync_probe_video_src is-live=true format=time do-timestamp=false caps={video_caps} ! \
+             queue max-size-buffers=4 leaky=downstream ! videoconvert ! \
+             jpegenc quality=90 ! image/jpeg,parsed=true,width={},height={},framerate={}/1 ! \
+             appsink name=sync_probe_video_sink emit-signals=false sync=false max-buffers=4 drop=true",
+            camera.width,
+            camera.height,
+            camera.fps.max(1),
+        ),
+        CameraCodec::H264 => format!(
+            "appsrc name=sync_probe_video_src is-live=true format=time do-timestamp=false caps={video_caps} ! \
+             queue max-size-buffers=4 leaky=downstream ! videoconvert ! \
+             {} ! h264parse config-interval=-1 ! video/x-h264,stream-format=byte-stream,alignment=au ! \
+             appsink name=sync_probe_video_sink emit-signals=false sync=false max-buffers=4 drop=true",
+            pick_h264_encoder(camera.fps.max(1))?
+        ),
+    };
+
+    let audio_branch = format!(
+        "appsrc name=sync_probe_audio_src is-live=true format=time do-timestamp=false \
+         caps=audio/x-raw,format=S16LE,layout=interleaved,channels={},rate={} ! \
+         queue max-size-buffers=8 leaky=downstream ! \
+         audioconvert ! audioresample ! audio/x-raw,channels=2,rate={} ! \
+         {} ! aacparse ! capsfilter caps=audio/mpeg,stream-format=adts,rate={},channels=2 ! \
+         appsink name=sync_probe_audio_sink emit-signals=false sync=false max-buffers=32 drop=true",
+        AUDIO_CHANNELS,
+        AUDIO_SAMPLE_RATE,
+        AUDIO_SAMPLE_RATE,
+        pick_aac_encoder()?,
+        AUDIO_SAMPLE_RATE,
+    );
+
+    let desc = format!("{video_branch} {audio_branch}");
+    gst::parse::launch(&desc)
+        .with_context(|| format!("building sync probe pipeline: {desc}"))?
+        .downcast::<gst::Pipeline>()
+        .map_err(|_| anyhow::anyhow!("sync probe description did not build a pipeline"))
+}
+
+fn pick_h264_encoder(fps: u32) -> Result<String> {
+    if gst::ElementFactory::find("x264enc").is_some() {
+        return Ok(format!(
+            "x264enc tune=zerolatency speed-preset=ultrafast bitrate=2500 key-int-max={}",
+            fps.max(1)
+        ));
+    }
+    if gst::ElementFactory::find("openh264enc").is_some() {
+        return Ok("openh264enc bitrate=2500000".to_string());
+    }
+    if gst::ElementFactory::find("v4l2h264enc").is_some() {
+        return Ok("v4l2h264enc".to_string());
+    }
+    bail!("no usable H.264 encoder found for sync probe")
+}
+
+fn pick_aac_encoder() -> Result<&'static str> {
+    [
+        "avenc_aac bitrate=128000",
+        "fdkaacenc bitrate=128000",
+        "faac bitrate=128000",
+    ]
+    .into_iter()
+    .find(|entry| {
+        let name = entry.split_ascii_whitespace().next().unwrap_or_default();
+        gst::ElementFactory::find(name).is_some()
+    })
+    .ok_or_else(|| anyhow::anyhow!("no usable AAC encoder found for sync probe"))
+}
+
+fn spawn_video_thread(
+    src: gst_app::AppSrc,
+    sink: gst_app::AppSink,
+    camera: CameraConfig,
+    schedule: PulseSchedule,
+    duration: Duration,
+    probe_start: Instant,
+    running: Arc<AtomicBool>,
+    queue: FreshPacketQueue<VideoPacket>,
+) -> JoinHandle<()> {
+    thread::spawn(move || {
+        let dark_frame = build_dark_probe_frame(camera.width as usize, camera.height as usize);
+        let regular_pulse_frame =
+            build_regular_probe_frame(camera.width as usize, camera.height as usize);
+        let marker_pulse_frame =
+            build_marker_probe_frame(camera.width as usize, camera.height as usize);
+        let frame_step = Duration::from_nanos(1_000_000_000u64 / u64::from(camera.fps.max(1)));
+        let mut frame_index = 0u64;
+
+        while running.load(Ordering::Acquire) {
+            let pts = schedule.frame_pts(frame_index, camera.fps.max(1));
+            if pts > duration {
+                break;
+            }
+
+            let deadline = probe_start + pts;
+            if let Some(remaining) = deadline.checked_duration_since(Instant::now())
+                && !remaining.is_zero()
+            {
+                thread::sleep(remaining);
+            }
+
+            let frame = if schedule.flash_active(pts) && schedule.pulse_is_marker(pts) {
+                &marker_pulse_frame
+            } else if schedule.flash_active(pts) {
+                &regular_pulse_frame
+            } else {
+                &dark_frame
+            };
+            let mut buffer = gst::Buffer::from_slice(frame.clone());
+            if let Some(meta) = buffer.get_mut() {
+                let pts_time = gst::ClockTime::from_nseconds(pts.as_nanos() as u64);
+                meta.set_pts(Some(pts_time));
+                meta.set_dts(Some(pts_time));
+                meta.set_duration(Some(gst::ClockTime::from_nseconds(
+                    frame_step.as_nanos() as u64
+                )));
+            }
+            if src.push_buffer(buffer).is_err() {
+                break;
+            }
+
+            if let Some(sample) = sink.try_pull_sample(gst::ClockTime::from_mseconds(250))
+                && let Some(buffer) = sample.buffer()
+                && let Ok(map) = buffer.map_readable()
+            {
+                let packet = VideoPacket {
+                    id: 2,
+                    pts: buffer.pts().unwrap_or(gst::ClockTime::ZERO).nseconds() / 1_000,
+                    data: map.as_slice().to_vec(),
+                    ..Default::default()
+                };
+                let _ = queue.push(packet, Duration::ZERO);
+            }
+
+            frame_index = frame_index.saturating_add(1);
+        }
+
+        let _ = src.end_of_stream();
+        queue.close();
+    })
+}
+
+fn spawn_audio_thread(
+    src: gst_app::AppSrc,
+    sink: gst_app::AppSink,
+    schedule: PulseSchedule,
+    duration: Duration,
+    probe_start: Instant,
+    running: Arc<AtomicBool>,
+    queue: FreshPacketQueue<AudioPacket>,
+) -> JoinHandle<()> {
+    thread::spawn(move || {
+        let chunk_duration = Duration::from_millis(AUDIO_CHUNK_MS);
+        let samples_per_chunk =
+            (AUDIO_SAMPLE_RATE as usize * AUDIO_CHUNK_MS as usize / 1_000).max(1);
+        let mut chunk_index = 0u64;
+
+        while running.load(Ordering::Acquire) {
+            let pts = chunk_duration.saturating_mul(chunk_index as u32);
+            if pts > duration {
+                break;
+            }
+
+            let deadline = probe_start + pts;
+            if let Some(remaining) = deadline.checked_duration_since(Instant::now())
+                && !remaining.is_zero()
+            {
+                thread::sleep(remaining);
+            }
+
+            let chunk = render_audio_chunk(&schedule, pts, samples_per_chunk);
+            let mut buffer = gst::Buffer::from_slice(chunk);
+            if let Some(meta) = buffer.get_mut() {
+                let pts_time = gst::ClockTime::from_nseconds(pts.as_nanos() as u64);
+                meta.set_pts(Some(pts_time));
+                meta.set_dts(Some(pts_time));
+                meta.set_duration(Some(gst::ClockTime::from_nseconds(
+                    chunk_duration.as_nanos() as u64,
+                )));
+            }
+            if src.push_buffer(buffer).is_err() {
+                break;
+            }
+
+            drain_audio_samples(&sink, &queue, duration, gst::ClockTime::ZERO);
+            chunk_index = chunk_index.saturating_add(1);
+        }
+
+        let _ = src.end_of_stream();
+        drain_audio_samples(&sink, &queue, duration, gst::ClockTime::from_mseconds(100));
+        queue.close();
+    })
+}
+
+fn drain_audio_samples(
+    sink: &gst_app::AppSink,
+    queue: &FreshPacketQueue<AudioPacket>,
+    duration: Duration,
+    timeout: gst::ClockTime,
+) {
+    while let Some(sample) = sink.try_pull_sample(timeout) {
+        let Some(buffer) = sample.buffer() else {
+            continue;
+        };
+        let pts_usecs = buffer.pts().unwrap_or(gst::ClockTime::ZERO).nseconds() / 1_000;
+        if probe_pts_exceeds_duration(pts_usecs, duration) {
+            break;
+        }
+        let Ok(map) = buffer.map_readable() else {
+            continue;
+        };
+        let packet = AudioPacket {
+            id: 0,
+            pts: pts_usecs,
+            data: map.as_slice().to_vec(),
+        };
+        let _ = queue.push(packet, Duration::ZERO);
+        if timeout == gst::ClockTime::ZERO {
+            continue;
+        }
+    }
+}

client/src/sync_probe/capture/tests.rs

@@ -0,0 +1,131 @@
+use super::{
+    SyncProbeCapture, build_dark_probe_frame, build_marker_probe_frame, build_regular_probe_frame,
+};
+use crate::input::camera::{CameraCodec, CameraConfig};
+use crate::sync_probe::analyze::detect_audio_onsets;
+use crate::sync_probe::schedule::PulseSchedule;
+use lesavka_common::lesavka::{AudioPacket, VideoPacket};
+use std::time::Duration;
+
+fn stub_camera() -> CameraConfig {
+    CameraConfig {
+        codec: CameraCodec::Mjpeg,
+        width: 1280,
+        height: 720,
+        fps: 30,
+    }
+}
+
+#[tokio::test]
+async fn coverage_stub_exposes_live_video_and_audio_queues() {
+    let capture = SyncProbeCapture::new(
+        stub_camera(),
+        PulseSchedule::new(
+            Duration::from_secs(4),
+            Duration::from_secs(1),
+            Duration::from_millis(100),
+            5,
+        ),
+        Duration::from_secs(2),
+    )
+    .expect("stub capture");
+
+    let video_queue = capture.video_queue();
+    let audio_queue = capture.audio_queue();
+    let _ = video_queue.push(
+        VideoPacket {
+            id: 2,
+            pts: 1,
+            data: vec![1, 2, 3],
+            ..Default::default()
+        },
+        Duration::ZERO,
+    );
+    let _ = audio_queue.push(
+        AudioPacket {
+            id: 0,
+            pts: 2,
+            data: vec![4, 5, 6],
+        },
+        Duration::ZERO,
+    );
+
+    let video = video_queue.pop_fresh().await;
+    let audio = audio_queue.pop_fresh().await;
+    assert_eq!(video.packet.expect("video packet").data, vec![1, 2, 3]);
+    assert_eq!(audio.packet.expect("audio packet").data, vec![4, 5, 6]);
+}
+
+#[test]
+fn probe_pts_duration_cutoff_matches_video_boundary_rule() {
+    assert!(!super::probe_pts_exceeds_duration(
+        Duration::from_secs(2).as_micros() as u64,
+        Duration::from_secs(2),
+    ));
+    assert!(super::probe_pts_exceeds_duration(
+        Duration::from_secs(2).as_micros() as u64 + 1,
+        Duration::from_secs(2),
+    ));
+}
+
+#[test]
+fn synthesized_audio_pulses_track_video_flash_boundaries() {
+    let schedule = PulseSchedule::new(
+        Duration::from_secs(4),
+        Duration::from_secs(1),
+        Duration::from_millis(120),
+        5,
+    );
+    let chunk_samples =
+        (super::AUDIO_SAMPLE_RATE as usize * super::AUDIO_CHUNK_MS as usize / 1_000).max(1);
+    let mut samples = Vec::<i16>::new();
+    let chunk_duration = Duration::from_millis(super::AUDIO_CHUNK_MS);
+    for chunk_index in 0..600u64 {
+        let pts = chunk_duration.saturating_mul(chunk_index as u32);
+        if pts > Duration::from_secs(6) {
+            break;
+        }
+        let chunk = super::render_audio_chunk(&schedule, pts, chunk_samples);
+        samples.extend(
+            chunk
+                .chunks_exact(std::mem::size_of::<i16>())
+                .map(|bytes| i16::from_le_bytes([bytes[0], bytes[1]])),
+        );
+    }
+
+    let mono = samples
+        .chunks_exact(super::AUDIO_CHANNELS)
+        .map(|frame| frame[0])
+        .collect::<Vec<_>>();
+    let onsets =
+        detect_audio_onsets(&mono, super::AUDIO_SAMPLE_RATE as u32, 5).expect("audio onsets");
+    assert!(
+        onsets.len() >= 2,
+        "expected at least 2 onsets, got {onsets:?}"
+    );
+    assert!(
+        onsets
+            .iter()
+            .zip([4.0, 5.0])
+            .all(|(actual, expected)| { (*actual - expected).abs() <= 0.02 })
+    );
+}
+
+#[test]
+fn probe_video_frames_render_distinct_idle_regular_and_marker_patterns() {
+    let dark = build_dark_probe_frame(64, 36);
+    let regular = build_regular_probe_frame(64, 36);
+    let marker = build_marker_probe_frame(64, 36);
+
+    assert_eq!(dark.len(), regular.len());
+    assert_eq!(dark.len(), marker.len());
+    assert!(
+        regular.iter().map(|byte| u64::from(*byte)).sum::<u64>()
+            != dark.iter().map(|byte| u64::from(*byte)).sum::<u64>()
+    );
+    assert!(
+        marker.iter().map(|byte| u64::from(*byte)).sum::<u64>()
+            != dark.iter().map(|byte| u64::from(*byte)).sum::<u64>()
+    );
+    assert_ne!(regular, marker);
+}

client/src/sync_probe/config.rs

@@ -0,0 +1,214 @@
+//! CLI parsing for the upstream A/V sync probe.
+
+use anyhow::{Context, Result, bail};
+use std::time::Duration;
+
+use crate::app_support::DEFAULT_SERVER_ADDR;
+
+#[derive(Clone, Debug, Eq, PartialEq)]
+pub struct ProbeConfig {
+    pub server: String,
+    pub duration: Duration,
+    pub warmup: Duration,
+    pub pulse_period: Duration,
+    pub pulse_width: Duration,
+    pub marker_tick_period: u32,
+}
+
+#[derive(Debug, Eq, PartialEq)]
+pub enum ParseOutcome {
+    Run(ProbeConfig),
+    Help,
+}
+
+pub fn usage() -> &'static str {
+    "Usage: lesavka-sync-probe [--server http://HOST:50051] [--duration-seconds 10] [--warmup-seconds 4] [--pulse-period-ms 1000] [--pulse-width-ms 120] [--marker-tick-period 5]"
+}
+
+pub fn parse_args_outcome_from<I, S>(args: I) -> Result<ParseOutcome>
+where
+    I: IntoIterator<Item = S>,
+    S: Into<String>,
+{
+    let mut args = args.into_iter().map(Into::into);
+    let mut server = DEFAULT_SERVER_ADDR.to_string();
+    let mut duration_seconds = 10u64;
+    let mut warmup_seconds = 4u64;
+    let mut pulse_period_ms = 1_000u64;
+    let mut pulse_width_ms = 120u64;
+    let mut marker_tick_period = 5u32;
+
+    while let Some(arg) = args.next() {
+        match arg.as_str() {
+            "--server" => {
+                server = args
+                    .next()
+                    .context("missing value after --server")?
+                    .trim()
+                    .to_string();
+            }
+            "--duration-seconds" => {
+                duration_seconds = parse_u64_arg(
+                    args.next(),
+                    "--duration-seconds",
+                    "duration must be at least one second",
+                )?;
+                if duration_seconds == 0 {
+                    bail!("duration must be at least one second\n{}", usage());
+                }
+            }
+            "--warmup-seconds" => {
+                warmup_seconds = parse_u64_arg(
+                    args.next(),
+                    "--warmup-seconds",
+                    "warmup must stay non-negative",
+                )?;
+            }
+            "--pulse-period-ms" => {
+                pulse_period_ms = parse_u64_arg(
+                    args.next(),
+                    "--pulse-period-ms",
+                    "pulse period must be positive",
+                )?;
+                if pulse_period_ms == 0 {
+                    bail!("pulse period must be positive\n{}", usage());
+                }
+            }
+            "--pulse-width-ms" => {
+                pulse_width_ms = parse_u64_arg(
+                    args.next(),
+                    "--pulse-width-ms",
+                    "pulse width must be positive",
+                )?;
+                if pulse_width_ms == 0 {
+                    bail!("pulse width must be positive\n{}", usage());
+                }
+            }
+            "--marker-tick-period" => {
+                marker_tick_period = parse_u32_arg(
+                    args.next(),
+                    "--marker-tick-period",
+                    "marker tick period must be positive",
+                )?;
+                if marker_tick_period == 0 {
+                    bail!("marker tick period must be positive\n{}", usage());
+                }
+            }
+            "--help" | "-h" => return Ok(ParseOutcome::Help),
+            _ => bail!("unexpected argument `{arg}`\n{}", usage()),
+        }
+    }
+
+    if pulse_width_ms >= pulse_period_ms {
+        bail!(
+            "pulse width must stay smaller than the pulse period\n{}",
+            usage()
+        );
+    }
+
+    Ok(ParseOutcome::Run(ProbeConfig {
+        server,
+        duration: Duration::from_secs(duration_seconds),
+        warmup: Duration::from_secs(warmup_seconds),
+        pulse_period: Duration::from_millis(pulse_period_ms),
+        pulse_width: Duration::from_millis(pulse_width_ms),
+        marker_tick_period,
+    }))
+}
+
+fn parse_u64_arg(value: Option<String>, flag: &str, context: &str) -> Result<u64> {
+    value
+        .context(format!("missing value after {flag}"))?
+        .trim()
+        .parse::<u64>()
+        .with_context(|| format!("{context}\n{}", usage()))
+}
+
+fn parse_u32_arg(value: Option<String>, flag: &str, context: &str) -> Result<u32> {
+    value
+        .context(format!("missing value after {flag}"))?
+        .trim()
+        .parse::<u32>()
+        .with_context(|| format!("{context}\n{}", usage()))
+}
+
+#[cfg(test)]
+mod tests {
+    use super::{DEFAULT_SERVER_ADDR, ParseOutcome, parse_args_outcome_from};
+    use std::time::Duration;
+
+    #[test]
+    fn parse_args_uses_operational_defaults() {
+        let outcome = parse_args_outcome_from(std::iter::empty::<&str>()).expect("defaults");
+        let ParseOutcome::Run(config) = outcome else {
+            panic!("expected run config");
+        };
+
+        assert_eq!(config.server, DEFAULT_SERVER_ADDR);
+        assert_eq!(config.duration, Duration::from_secs(10));
+        assert_eq!(config.warmup, Duration::from_secs(4));
+        assert_eq!(config.pulse_period, Duration::from_millis(1_000));
+        assert_eq!(config.pulse_width, Duration::from_millis(120));
+        assert_eq!(config.marker_tick_period, 5);
+    }
+
+    #[test]
+    fn parse_args_accepts_explicit_probe_settings() {
+        let outcome = parse_args_outcome_from([
+            "--server",
+            "http://lab:50051",
+            "--duration-seconds",
+            "14",
+            "--warmup-seconds",
+            "6",
+            "--pulse-period-ms",
+            "750",
+            "--pulse-width-ms",
+            "90",
+            "--marker-tick-period",
+            "3",
+        ])
+        .expect("configured run");
+        let ParseOutcome::Run(config) = outcome else {
+            panic!("expected run config");
+        };
+
+        assert_eq!(config.server, "http://lab:50051");
+        assert_eq!(config.duration, Duration::from_secs(14));
+        assert_eq!(config.warmup, Duration::from_secs(6));
+        assert_eq!(config.pulse_period, Duration::from_millis(750));
+        assert_eq!(config.pulse_width, Duration::from_millis(90));
+        assert_eq!(config.marker_tick_period, 3);
+    }
+
+    #[test]
+    fn parse_args_rejects_invalid_timing_relationships() {
+        assert!(parse_args_outcome_from(["--pulse-width-ms", "1000"]).is_err());
+        assert!(parse_args_outcome_from(["--pulse-period-ms", "0"]).is_err());
+        assert!(parse_args_outcome_from(["--duration-seconds", "0"]).is_err());
+        assert!(parse_args_outcome_from(["--marker-tick-period", "0"]).is_err());
+    }
+
+    #[test]
+    fn parse_args_supports_help() {
+        let outcome = parse_args_outcome_from(["--help"]).expect("help");
+        assert_eq!(outcome, ParseOutcome::Help);
+    }
+
+    #[test]
+    fn parse_args_rejects_missing_values_and_unknown_flags() {
+        assert!(parse_args_outcome_from(["--server"]).is_err());
+        assert!(parse_args_outcome_from(["--duration-seconds"]).is_err());
+        assert!(parse_args_outcome_from(["--marker-tick-period"]).is_err());
+        assert!(parse_args_outcome_from(["--wat"]).is_err());
+    }
+
+    #[test]
+    fn parse_args_rejects_non_numeric_values() {
+        assert!(parse_args_outcome_from(["--duration-seconds", "abc"]).is_err());
+        assert!(parse_args_outcome_from(["--warmup-seconds", "abc"]).is_err());
+        assert!(parse_args_outcome_from(["--pulse-period-ms", "abc"]).is_err());
+        assert!(parse_args_outcome_from(["--pulse-width-ms", "abc"]).is_err());
+        assert!(parse_args_outcome_from(["--marker-tick-period", "abc"]).is_err());
+    }
+}

client/src/sync_probe/mod.rs

@@ -0,0 +1,15 @@
+//! End-to-end upstream A/V sync probe tooling.
+//!
+//! This module exists to answer one narrow operational question: when the
+//! client captures audio and video at the same moment, do those streams arrive
+//! on Tethys in sync? The probe emits deterministic marker pulses through the
+//! normal Lesavka relay path so the downstream consumer can be measured
+//! precisely instead of judged by eye.
+
+pub mod analyze;
+mod capture;
+mod config;
+mod runner;
+mod schedule;
+
+pub use runner::run_sync_probe_from_args;

client/src/sync_probe/runner.rs

@@ -0,0 +1,173 @@
+//! Runtime entrypoint for the shared-clock upstream sync probe.
+
+use anyhow::{Context, Result, bail};
+
+use crate::app_support;
+use crate::handshake;
+use crate::sync_probe::capture::SyncProbeCapture;
+use crate::sync_probe::config::{ParseOutcome, ProbeConfig, parse_args_outcome_from, usage};
+use crate::sync_probe::schedule::PulseSchedule;
+
+#[cfg(not(coverage))]
+use lesavka_common::lesavka::relay_client::RelayClient;
+#[cfg(not(coverage))]
+use tonic::{Request, transport::Channel};
+
+pub async fn run_sync_probe_from_args<I, S>(args: I) -> Result<()>
+where
+    I: IntoIterator<Item = S>,
+    S: Into<String>,
+{
+    match parse_args_outcome_from(args)? {
+        ParseOutcome::Run(config) => run_sync_probe(config).await,
+        ParseOutcome::Help => {
+            println!("{}", usage());
+            Ok(())
+        }
+    }
+}
+
+#[cfg(not(coverage))]
+async fn run_sync_probe(config: ProbeConfig) -> Result<()> {
+    let caps = handshake::negotiate(config.server.as_str()).await;
+    if !caps.camera || !caps.microphone {
+        bail!("server does not advertise both camera and microphone support");
+    }
+    let camera = app_support::camera_config_from_caps(&caps)
+        .context("server handshake did not include a complete camera profile")?;
+
+    let schedule = PulseSchedule::new(
+        config.warmup,
+        config.pulse_period,
+        config.pulse_width,
+        config.marker_tick_period,
+    );
+    let capture = SyncProbeCapture::new(camera, schedule, config.duration)?;
+
+    tracing::info!(
+        server = %config.server,
+        duration_s = config.duration.as_secs(),
+        codec = ?camera.codec,
+        width = camera.width,
+        height = camera.height,
+        fps = camera.fps,
+        "🧪 A/V sync probe starting"
+    );
+
+    let video_channel = connect(config.server.as_str()).await?;
+    let audio_channel = connect(config.server.as_str()).await?;
+    let video_queue = capture.video_queue();
+    let audio_queue = capture.audio_queue();
+
+    let video_task = tokio::spawn(async move {
+        let mut client = RelayClient::new(video_channel);
+        let outbound = async_stream::stream! {
+            loop {
+                let next = video_queue.pop_fresh().await;
+                if let Some(packet) = next.packet {
+                    yield packet;
+                    continue;
+                }
+                break;
+            }
+        };
+        let mut response = client
+            .stream_camera(Request::new(outbound))
+            .await
+            .context("starting sync probe camera stream")?;
+        while response.get_mut().message().await.transpose().is_some() {}
+        Ok::<(), anyhow::Error>(())
+    });
+
+    let audio_task = tokio::spawn(async move {
+        let mut client = RelayClient::new(audio_channel);
+        let outbound = async_stream::stream! {
+            loop {
+                let next = audio_queue.pop_fresh().await;
+                if let Some(packet) = next.packet {
+                    yield packet;
+                    continue;
+                }
+                break;
+            }
+        };
+        let mut response = client
+            .stream_microphone(Request::new(outbound))
+            .await
+            .context("starting sync probe microphone stream")?;
+        while response.get_mut().message().await.transpose().is_some() {}
+        Ok::<(), anyhow::Error>(())
+    });
+
+    let (video_result, audio_result) =
+        tokio::try_join!(video_task, audio_task).context("joining sync probe streams")?;
+    video_result.context("sync probe camera task failed")?;
+    audio_result.context("sync probe microphone task failed")?;
+
+    tracing::info!("🧪 A/V sync probe finished");
+    Ok(())
+}
+
+#[cfg(not(coverage))]
+async fn connect(server_addr: &str) -> Result<Channel> {
+    Channel::from_shared(server_addr.to_string())
+        .context("invalid relay server address")?
+        .tcp_nodelay(true)
+        .connect()
+        .await
+        .with_context(|| format!("connecting to relay at {server_addr}"))
+}
+
+#[cfg(coverage)]
+async fn run_sync_probe(_config: ProbeConfig) -> Result<()> {
+    Ok(())
+}
+
+#[cfg(test)]
+mod tests {
+    use super::parse_args_outcome_from;
+    #[cfg(coverage)]
+    use super::run_sync_probe_from_args;
+    use crate::sync_probe::config::ParseOutcome;
+
+    #[test]
+    fn help_passthrough_stays_stable() {
+        let outcome = parse_args_outcome_from(["--help"]).expect("help");
+        assert_eq!(outcome, ParseOutcome::Help);
+    }
+
+    #[cfg(coverage)]
+    #[test]
+    fn coverage_run_path_accepts_default_probe_args() {
+        let rt = tokio::runtime::Runtime::new().expect("runtime");
+        rt.block_on(async {
+            run_sync_probe_from_args(std::iter::empty::<&str>())
+                .await
+                .expect("coverage run path");
+        });
+    }
+
+    #[cfg(coverage)]
+    #[test]
+    fn coverage_run_path_accepts_custom_probe_args() {
+        let rt = tokio::runtime::Runtime::new().expect("runtime");
+        rt.block_on(async {
+            run_sync_probe_from_args([
+                "--server",
+                "http://lab:50051",
+                "--duration-seconds",
+                "2",
+                "--warmup-seconds",
+                "5",
+                "--pulse-period-ms",
+                "700",
+                "--pulse-width-ms",
+                "80",
+                "--marker-tick-period",
+                "4",
+            ])
+            .await
+            .expect("configured coverage run path");
+        });
+    }
+}

client/src/sync_probe/schedule.rs

@@ -0,0 +1,234 @@
+#![cfg_attr(not(test), allow(dead_code))]
+//! Timing helpers for the shared-clock sync probe.
+
+use std::time::Duration;
+
+#[derive(Clone, Debug, Eq, PartialEq)]
+pub struct PulseSchedule {
+    warmup: Duration,
+    pulse_period: Duration,
+    pulse_width: Duration,
+    marker_tick_period: u32,
+}
+
+impl PulseSchedule {
+    pub fn new(
+        warmup: Duration,
+        pulse_period: Duration,
+        pulse_width: Duration,
+        marker_tick_period: u32,
+    ) -> Self {
+        assert!(!pulse_period.is_zero(), "pulse period must stay positive");
+        assert!(!pulse_width.is_zero(), "pulse width must stay positive");
+        assert!(
+            pulse_width < pulse_period,
+            "pulse width must stay smaller than the pulse period"
+        );
+        assert!(
+            marker_tick_period > 0,
+            "marker tick period must stay positive"
+        );
+
+        Self {
+            warmup,
+            pulse_period,
+            pulse_width,
+            marker_tick_period,
+        }
+    }
+
+    pub fn warmup(&self) -> Duration {
+        self.warmup
+    }
+
+    pub fn pulse_period(&self) -> Duration {
+        self.pulse_period
+    }
+
+    pub fn marker_tick_period(&self) -> u32 {
+        self.marker_tick_period
+    }
+
+    pub fn pulse_width(&self) -> Duration {
+        self.pulse_width
+    }
+
+    pub fn pulse_index(&self, pts: Duration) -> u64 {
+        if pts < self.warmup_boundary() {
+            return 0;
+        }
+        let pts = pts - self.warmup_boundary();
+        let period_ns = self.pulse_period.as_nanos().max(1);
+        (pts.as_nanos() / period_ns) as u64
+    }
+
+    pub fn pulse_offset(&self, pts: Duration) -> Duration {
+        if pts < self.warmup_boundary() {
+            return pts;
+        }
+        let pts = pts - self.warmup_boundary();
+        let period_ns = self.pulse_period.as_nanos().max(1);
+        let offset_ns = (pts.as_nanos() % period_ns) as u64;
+        Duration::from_nanos(offset_ns)
+    }
+
+    pub fn pulse_is_marker(&self, pts: Duration) -> bool {
+        pts >= self.warmup_boundary()
+            && self
+                .pulse_index(pts)
+                .is_multiple_of(u64::from(self.marker_tick_period.max(1)))
+    }
+
+    pub fn marker_pulse_width(&self) -> Duration {
+        let widened = self.pulse_width.saturating_mul(2);
+        widened.min(self.pulse_period.saturating_sub(Duration::from_millis(1)))
+    }
+
+    pub fn flash_active(&self, pts: Duration) -> bool {
+        if pts < self.warmup_boundary() {
+            return false;
+        }
+        let width = if self.pulse_is_marker(pts) {
+            self.marker_pulse_width()
+        } else {
+            self.pulse_width
+        };
+        self.pulse_offset(pts) < width
+    }
+
+    pub fn warmup_boundary(&self) -> Duration {
+        if self.warmup.is_zero() {
+            return Duration::ZERO;
+        }
+        let period_ns = self.pulse_period.as_nanos().max(1) as u64;
+        let warmup_ns = self.warmup.as_nanos() as u64;
+        let rounded = ((warmup_ns + period_ns - 1) / period_ns) * period_ns;
+        Duration::from_nanos(rounded)
+    }
+
+    pub fn audio_gate_open_at(&self) -> Duration {
+        let lead = self.pulse_width.min(Duration::from_millis(200));
+        self.warmup_boundary().saturating_sub(lead)
+    }
+
+    pub fn frame_pts(&self, frame_index: u64, fps: u32) -> Duration {
+        let frame_step_ns = 1_000_000_000u64 / u64::from(fps.max(1));
+        Duration::from_nanos(frame_index.saturating_mul(frame_step_ns))
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::PulseSchedule;
+    use std::time::Duration;
+
+    #[test]
+    fn flash_windows_follow_the_expected_pulse_boundaries() {
+        let schedule = PulseSchedule::new(
+            Duration::from_secs(4),
+            Duration::from_millis(1_000),
+            Duration::from_millis(120),
+            5,
+        );
+
+        assert!(!schedule.flash_active(Duration::from_millis(3_999)));
+        assert!(schedule.flash_active(Duration::from_millis(4_000)));
+        assert!(schedule.flash_active(Duration::from_millis(4_200)));
+        assert!(!schedule.flash_active(Duration::from_millis(4_240)));
+        assert!(!schedule.flash_active(Duration::from_millis(4_999)));
+        assert!(schedule.flash_active(Duration::from_millis(5_000)));
+        assert!(!schedule.flash_active(Duration::from_millis(5_120)));
+    }
+
+    #[test]
+    fn pulse_index_and_offset_repeat_cleanly() {
+        let schedule = PulseSchedule::new(
+            Duration::from_secs(2),
+            Duration::from_millis(750),
+            Duration::from_millis(90),
+            3,
+        );
+
+        assert_eq!(schedule.pulse_index(Duration::from_millis(0)), 0);
+        assert_eq!(schedule.warmup_boundary(), Duration::from_millis(2_250));
+        assert_eq!(schedule.pulse_index(Duration::from_millis(2_999)), 0);
+        assert_eq!(schedule.pulse_index(Duration::from_millis(3_000)), 1);
+        assert_eq!(
+            schedule.pulse_offset(Duration::from_millis(2_320)),
+            Duration::from_millis(70)
+        );
+        assert_eq!(
+            schedule.pulse_offset(Duration::from_millis(3_750)),
+            Duration::from_millis(0)
+        );
+    }
+
+    #[test]
+    fn frame_pts_respects_requested_framerate() {
+        let schedule = PulseSchedule::new(
+            Duration::ZERO,
+            Duration::from_secs(1),
+            Duration::from_millis(100),
+            5,
+        );
+
+        assert_eq!(schedule.frame_pts(0, 30), Duration::from_nanos(0));
+        assert_eq!(schedule.frame_pts(1, 30), Duration::from_nanos(33_333_333));
+        assert_eq!(
+            schedule.frame_pts(30, 30),
+            Duration::from_nanos(999_999_990)
+        );
+    }
+
+    #[test]
+    fn getters_and_zero_fps_fallback_stay_stable() {
+        let schedule = PulseSchedule::new(
+            Duration::from_secs(3),
+            Duration::from_millis(800),
+            Duration::from_millis(90),
+            7,
+        );
+
+        assert_eq!(schedule.warmup(), Duration::from_secs(3));
+        assert_eq!(schedule.warmup_boundary(), Duration::from_millis(3_200));
+        assert_eq!(schedule.audio_gate_open_at(), Duration::from_millis(3_110));
+        assert_eq!(schedule.pulse_period(), Duration::from_millis(800));
+        assert_eq!(schedule.pulse_width(), Duration::from_millis(90));
+        assert_eq!(schedule.marker_tick_period(), 7);
+        assert_eq!(schedule.frame_pts(1, 0), Duration::from_secs(1));
+    }
+
+    #[test]
+    fn marker_pulses_are_detectably_wider_than_regular_pulses() {
+        let schedule = PulseSchedule::new(
+            Duration::from_secs(1),
+            Duration::from_millis(1_000),
+            Duration::from_millis(120),
+            5,
+        );
+
+        assert!(schedule.pulse_is_marker(Duration::from_millis(1_000)));
+        assert_eq!(schedule.marker_pulse_width(), Duration::from_millis(240));
+        assert!(schedule.flash_active(Duration::from_millis(1_200)));
+        assert!(!schedule.flash_active(Duration::from_millis(1_241)));
+        assert!(!schedule.pulse_is_marker(Duration::from_millis(2_000)));
+        assert!(!schedule.flash_active(Duration::from_millis(2_200)));
+    }
+
+    #[test]
+    #[should_panic(expected = "pulse period must stay positive")]
+    fn constructor_rejects_zero_period() {
+        let _ = PulseSchedule::new(Duration::ZERO, Duration::ZERO, Duration::from_millis(50), 1);
+    }
+
+    #[test]
+    #[should_panic(expected = "pulse width must stay smaller than the pulse period")]
+    fn constructor_rejects_width_not_smaller_than_period() {
+        let _ = PulseSchedule::new(
+            Duration::ZERO,
+            Duration::from_millis(100),
+            Duration::from_millis(100),
+            1,
+        );
+    }
+}

common/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name                    = "lesavka_common"
-version                 = "0.12.4"
+version                 = "0.13.0"
 edition                 = "2024"
 build                   = "build.rs"
 

docs/operational-env.md

@@ -181,6 +181,11 @@ Hardware-facing assumptions belong near the code that uses them; this file is th
 | `LESAVKA_SONAR_ENFORCE` | CI gate enforcement override |
 | `LESAVKA_SUPPLY_CHAIN_ENFORCE_TOOLS` | CI gate enforcement override |
 | `LESAVKA_TAP_AUDIO` | client media capture/playback override |
+| `LESAVKA_UAC_BUFFER_TIME_US` | server audio sink latency override |
+| `LESAVKA_UAC_COMPENSATION_US` | server audio sink latency override |
+| `LESAVKA_UAC_DEV` | server hardware/device override |
+| `LESAVKA_UAC_HDMI_COMPENSATION_US` | server audio sink latency override |
+| `LESAVKA_UAC_LATENCY_TIME_US` | server audio sink latency override |
 | `LESAVKA_TEST_CAM_U32` | test/build contract variable; not runtime operator config |
 | `LESAVKA_TEST_CAP_CAMERA` | test/build contract variable; not runtime operator config |
 | `LESAVKA_TEST_CAP_MIC` | test/build contract variable; not runtime operator config |

scripts/ci/quality_gate_baseline.json

@@ -16,6 +16,14 @@
       "line_percent": 100.0,
       "loc": 304
     },
+    "client/src/bin/lesavka-sync-analyze.rs": {
+      "line_percent": 95.0,
+      "loc": 98
+    },
+    "client/src/bin/lesavka-sync-probe.rs": {
+      "line_percent": 100.0,
+      "loc": 19
+    },
     "client/src/handshake.rs": {
       "line_percent": 100.0,
       "loc": 381
@@ -29,8 +37,8 @@
       "loc": 69
     },
     "client/src/input/camera/capture_pipeline.rs": {
-      "line_percent": 99.28,
-      "loc": 254
+      "line_percent": 99.32,
+      "loc": 272
     },
     "client/src/input/camera/device_selection.rs": {
       "line_percent": 97.62,
@@ -86,7 +94,7 @@
     },
     "client/src/input/microphone.rs": {
       "line_percent": 100.0,
-      "loc": 398
+      "loc": 416
     },
     "client/src/input/mouse.rs": {
       "line_percent": 98.85,
@@ -130,7 +138,7 @@
     },
     "client/src/launcher/ui.rs": {
       "line_percent": 100.0,
-      "loc": 184
+      "loc": 182
     },
     "client/src/launcher/ui/session_preview_coverage.rs": {
       "line_percent": 100.0,
@@ -140,6 +148,10 @@
       "line_percent": 97.56,
       "loc": 78
     },
+    "client/src/live_capture_clock.rs": {
+      "line_percent": 100.0,
+      "loc": 56
+    },
     "client/src/main.rs": {
       "line_percent": 100.0,
       "loc": 101
@@ -168,6 +180,50 @@
       "line_percent": 100.0,
       "loc": 82
     },
+    "client/src/sync_probe/analyze.rs": {
+      "line_percent": 97.92,
+      "loc": 86
+    },
+    "client/src/sync_probe/analyze/media_extract.rs": {
+      "line_percent": 98.36,
+      "loc": 240
+    },
+    "client/src/sync_probe/analyze/onset_detection.rs": {
+      "line_percent": 96.77,
+      "loc": 274
+    },
+    "client/src/sync_probe/analyze/onset_detection/correlation.rs": {
+      "line_percent": 100.0,
+      "loc": 310
+    },
+    "client/src/sync_probe/analyze/report.rs": {
+      "line_percent": 100.0,
+      "loc": 59
+    },
+    "client/src/sync_probe/analyze/test_support.rs": {
+      "line_percent": 98.44,
+      "loc": 85
+    },
+    "client/src/sync_probe/capture.rs": {
+      "line_percent": 100.0,
+      "loc": 449
+    },
+    "client/src/sync_probe/capture/coverage_stub.rs": {
+      "line_percent": 100.0,
+      "loc": 34
+    },
+    "client/src/sync_probe/config.rs": {
+      "line_percent": 98.03,
+      "loc": 214
+    },
+    "client/src/sync_probe/runner.rs": {
+      "line_percent": 95.65,
+      "loc": 173
+    },
+    "client/src/sync_probe/schedule.rs": {
+      "line_percent": 98.74,
+      "loc": 234
+    },
     "client/src/uplink_fresh_queue.rs": {
       "line_percent": 100.0,
       "loc": 288
@@ -222,7 +278,7 @@
     },
     "server/src/audio/voice_input.rs": {
       "line_percent": 100.0,
-      "loc": 204
+      "loc": 358
     },
     "server/src/bin/lesavka_uvc/control_payloads.rs": {
       "line_percent": 100.0,
@@ -244,6 +300,10 @@
       "line_percent": 100.0,
       "loc": 471
     },
+    "server/src/camera/selection.rs": {
+      "line_percent": 97.67,
+      "loc": 372
+    },
     "server/src/camera_runtime.rs": {
       "line_percent": 95.52,
       "loc": 211
@@ -278,7 +338,7 @@
     },
     "server/src/main.rs": {
       "line_percent": 100.0,
-      "loc": 95
+      "loc": 96
     },
     "server/src/main/entrypoint.rs": {
       "line_percent": 100.0,
@@ -294,15 +354,15 @@
     },
     "server/src/main/handler_startup.rs": {
       "line_percent": 100.0,
-      "loc": 130
+      "loc": 131
     },
     "server/src/main/relay_service.rs": {
       "line_percent": 100.0,
-      "loc": 242
+      "loc": 289
     },
     "server/src/main/relay_service_coverage.rs": {
-      "line_percent": 98.31,
-      "loc": 138
+      "line_percent": 100.0,
+      "loc": 179
     },
     "server/src/main/rpc_helpers.rs": {
       "line_percent": 100.0,
@@ -328,6 +388,10 @@
       "line_percent": 100.0,
       "loc": 90
     },
+    "server/src/upstream_media_runtime.rs": {
+      "line_percent": 100.0,
+      "loc": 369
+    },
     "server/src/uvc_runtime.rs": {
       "line_percent": 98.48,
       "loc": 241
@@ -346,15 +410,15 @@
     },
     "server/src/video_sinks/hdmi_sink.rs": {
       "line_percent": 100.0,
-      "loc": 354
+      "loc": 393
     },
     "server/src/video_sinks/webcam_sink.rs": {
       "line_percent": 100.0,
       "loc": 199
     },
     "server/src/video_support.rs": {
-      "line_percent": 97.48,
-      "loc": 236
+      "line_percent": 97.74,
+      "loc": 263
     }
   }
 }

scripts/manual/browser_consumer_probe.py

@@ -0,0 +1,306 @@
+#!/usr/bin/env python3
+import argparse
+import http.server
+import json
+import socketserver
+import threading
+import time
+from pathlib import Path
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description="Serve a local browser consumer probe page")
+    parser.add_argument("--host", default="127.0.0.1")
+    parser.add_argument("--port", type=int, default=18443)
+    parser.add_argument("--output", default="/tmp/lesavka-browser-av-sync.webm")
+    parser.add_argument("--status", default="/tmp/lesavka-browser-av-sync-status.json")
+    parser.add_argument("--duration-seconds", type=int, default=15)
+    return parser.parse_args()
+
+
+class ProbeState:
+    def __init__(self, output_path: Path, status_path: Path, duration_seconds: int) -> None:
+        self.output_path = output_path
+        self.status_path = status_path
+        self.duration_seconds = duration_seconds
+        self.lock = threading.Lock()
+        self.start_token = 0
+        self.status = {
+            "booted_at": time.time(),
+            "ready": False,
+            "recording": False,
+            "uploaded": False,
+            "last_error": None,
+            "selected_video": None,
+            "selected_audio": None,
+            "devices": [],
+            "page_message": "booting",
+            "last_update": time.time(),
+        }
+        self.write_status()
+
+    def write_status(self) -> None:
+        self.status_path.parent.mkdir(parents=True, exist_ok=True)
+        tmp = self.status_path.with_suffix(".tmp")
+        tmp.write_text(json.dumps(self.status, indent=2, sort_keys=True), encoding="utf-8")
+        tmp.replace(self.status_path)
+
+    def update(self, payload: dict) -> None:
+        with self.lock:
+            self.status.update(payload)
+            self.status["last_update"] = time.time()
+            self.write_status()
+
+    def snapshot(self) -> dict:
+        with self.lock:
+            snap = dict(self.status)
+            snap["start_token"] = self.start_token
+            snap["duration_seconds"] = self.duration_seconds
+            return snap
+
+    def request_start(self) -> dict:
+        with self.lock:
+            self.start_token += 1
+            self.status.update({
+                "recording": False,
+                "uploaded": False,
+                "last_error": None,
+                "page_message": "start requested",
+                "start_requested_at": time.time(),
+            })
+            self.write_status()
+            return self.snapshot()
+
+    def store_upload(self, blob: bytes) -> dict:
+        with self.lock:
+            self.output_path.parent.mkdir(parents=True, exist_ok=True)
+            self.output_path.write_bytes(blob)
+            self.status.update({
+                "uploaded": True,
+                "recording": False,
+                "page_message": f"capture uploaded to {self.output_path}",
+                "upload_size": len(blob),
+                "uploaded_at": time.time(),
+            })
+            self.write_status()
+            return self.snapshot()
+
+
+def page_html(duration_seconds: int) -> str:
+    duration_ms = duration_seconds * 1000
+    return f"""<!doctype html>
+<html>
+<head>
+<meta charset=\"utf-8\">
+<title>Lesavka Browser Sync Probe</title>
+<style>
+body {{ margin: 0; background: #0f131a; color: #e7edf6; font: 16px/1.4 system-ui, sans-serif; }}
+.wrap {{ padding: 16px; display: grid; grid-template-columns: 1.1fr 0.9fr; gap: 16px; min-height: 100vh; box-sizing: border-box; }}
+.panel {{ background: #1b212c; border: 1px solid #2a3342; border-radius: 12px; padding: 12px; box-sizing: border-box; }}
+h1 {{ margin: 0 0 10px; font-size: 22px; }}
+h2 {{ margin: 0 0 8px; font-size: 16px; }}
+video {{ width: 100%; aspect-ratio: 16 / 9; background: black; border-radius: 10px; object-fit: contain; }}
+pre {{ white-space: pre-wrap; word-break: break-word; margin: 0; font: 13px/1.35 ui-monospace, monospace; }}
+.meter {{ height: 18px; background: #111722; border-radius: 999px; overflow: hidden; border: 1px solid #2a3342; }}
+.bar {{ height: 100%; width: 0%; background: linear-gradient(90deg, #2fbf71, #ffd36e, #ff6b6b); transition: width 120ms linear; }}
+.row {{ margin: 10px 0; }}
+</style>
+</head>
+<body>
+<div class=\"wrap\">
+  <div class=\"panel\">
+    <h1>Lesavka Browser Sync Probe</h1>
+    <video id=\"video\" autoplay playsinline muted></video>
+    <div class=\"row\"><div class=\"meter\"><div id=\"bar\" class=\"bar\"></div></div></div>
+  </div>
+  <div class=\"panel\">
+    <h2>Status</h2>
+    <pre id=\"status\">booting…</pre>
+  </div>
+</div>
+<script>
+const videoEl = document.getElementById('video');
+const statusEl = document.getElementById('status');
+const barEl = document.getElementById('bar');
+let stream = null;
+let recorder = null;
+let chunks = [];
+let startToken = 0;
+let analyser = null;
+let audioCtx = null;
+let recording = false;
+let heartbeatCounter = 0;
+
+function setStatus(lines) {{ statusEl.textContent = lines.join('\\n'); }}
+async function postJson(path, payload) {{
+  await fetch(path, {{ method: 'POST', headers: {{ 'Content-Type': 'application/json' }}, body: JSON.stringify(payload) }});
+}}
+async function postBlob(path, blob) {{
+  await fetch(path, {{ method: 'POST', headers: {{ 'Content-Type': 'application/octet-stream' }}, body: blob }});
+}}
+function fmtDevice(d) {{ return `${{d.kind}}: ${{d.label || '(unlabeled)'}}`; }}
+function meterLoop() {{
+  if (!analyser) return;
+  const data = new Uint8Array(analyser.fftSize);
+  const tick = () => {{
+    if (!analyser) return;
+    analyser.getByteTimeDomainData(data);
+    let peak = 0;
+    for (let i = 0; i < data.length; i++) peak = Math.max(peak, Math.abs(data[i] - 128));
+    const pct = Math.min(100, Math.round((peak / 128) * 100));
+    barEl.style.width = pct + '%';
+    requestAnimationFrame(tick);
+  }};
+  requestAnimationFrame(tick);
+}}
+function attachMeter(track) {{
+  if (!track) return;
+  try {{
+    audioCtx = new AudioContext();
+    const source = audioCtx.createMediaStreamSource(new MediaStream([track]));
+    analyser = audioCtx.createAnalyser();
+    analyser.fftSize = 2048;
+    source.connect(analyser);
+    meterLoop();
+  }} catch (err) {{
+    console.warn('meter setup failed', err);
+  }}
+}}
+async function initStream() {{
+  const lines = [];
+  try {{
+    await postJson('/status', {{ page_message: 'page loaded' }});
+    lines.push('requesting permission…');
+    await postJson('/status', {{ page_message: 'requesting permission' }});
+    const warm = await navigator.mediaDevices.getUserMedia({{ video: true, audio: true }});
+    warm.getTracks().forEach(track => track.stop());
+    await postJson('/status', {{ page_message: 'permission granted' }});
+    const devices = await navigator.mediaDevices.enumerateDevices();
+    await postJson('/status', {{ page_message: 'devices enumerated', devices: devices.map(fmtDevice) }});
+    const videoIn = devices.find(d => d.kind === 'videoinput' && /UGREEN/i.test(d.label)) || devices.find(d => d.kind === 'videoinput');
+    const audioIn = devices.find(d => d.kind === 'audioinput' && /(Multifunction Composite Gadget|Lesavka Composite)/i.test(d.label)) || devices.find(d => d.kind === 'audioinput');
+    stream = await navigator.mediaDevices.getUserMedia({{
+      video: videoIn ? {{ deviceId: {{ exact: videoIn.deviceId }} }} : true,
+      audio: audioIn ? {{ deviceId: {{ exact: audioIn.deviceId }} }} : true,
+    }});
+    await postJson('/status', {{ page_message: 'media stream opened' }});
+    videoEl.srcObject = stream;
+    attachMeter(stream.getAudioTracks()[0]);
+    const payload = {{
+      ready: true,
+      selected_video: videoIn ? fmtDevice(videoIn) : null,
+      selected_audio: audioIn ? fmtDevice(audioIn) : null,
+      devices: devices.map(fmtDevice),
+      page_message: 'ready for start',
+      last_error: null,
+    }};
+    await postJson('/status', payload);
+    lines.push('ready');
+    lines.push('video: ' + payload.selected_video);
+    lines.push('audio: ' + payload.selected_audio);
+    setStatus(lines);
+  }} catch (err) {{
+    const message = String(err && (err.stack || err));
+    await postJson('/status', {{ ready: false, last_error: message, page_message: 'permission or stream setup failed' }});
+    setStatus(['consumer status: FAIL', message]);
+  }}
+}}
+async function maybeStartRecording() {{
+  if (!stream || recording) return;
+  try {{
+    const response = await fetch('/command').then(r => r.json());
+    if (response.start_token === startToken) return;
+    startToken = response.start_token;
+    await postJson('/status', {{ page_message: 'start token observed', observed_start_token: startToken }});
+    recording = true;
+    chunks = [];
+    const preferredMime = 'video/webm;codecs=vp8,opus';
+    const options = MediaRecorder.isTypeSupported(preferredMime) ? {{ mimeType: preferredMime }} : undefined;
+    recorder = options ? new MediaRecorder(stream, options) : new MediaRecorder(stream);
+    await postJson('/status', {{ recording: true, uploaded: false, page_message: 'recording' }});
+    recorder.ondataavailable = event => {{ if (event.data && event.data.size > 0) chunks.push(event.data); }};
+    recorder.onstop = async () => {{
+      const blob = new Blob(chunks, {{ type: recorder.mimeType || 'video/webm' }});
+      await postBlob('/upload', blob);
+      recording = false;
+    }};
+    recorder.start(250);
+    setTimeout(() => recorder && recorder.state !== 'inactive' && recorder.stop(), response.duration_seconds * 1000);
+  }} catch (err) {{
+    recording = false;
+    const message = String(err && (err.stack || err));
+    await postJson('/status', {{ last_error: message, page_message: 'recording setup failed', recording: false }});
+  }}
+}}
+setInterval(() => {{
+  heartbeatCounter += 1;
+  void postJson('/status', {{ heartbeat_counter: heartbeatCounter, page_message: stream ? 'ready heartbeat' : 'boot heartbeat' }});
+}}, 1000);
+setInterval(() => {{ void maybeStartRecording(); }}, 250);
+void initStream();
+</script>
+</body>
+</html>"""
+
+
+class ProbeHandler(http.server.BaseHTTPRequestHandler):
+    state: ProbeState
+
+    def _send(self, code: int, body: bytes, content_type: str = "application/json") -> None:
+        self.send_response(code)
+        self.send_header("Content-Type", content_type)
+        self.send_header("Content-Length", str(len(body)))
+        self.end_headers()
+        self.wfile.write(body)
+
+    def do_GET(self) -> None:
+        if self.path in ("/", "/index.html"):
+            snap = self.state.snapshot()
+            self.state.update({
+                "page_message": "html served",
+                "html_served_count": int(snap.get("html_served_count", 0)) + 1,
+            })
+            self._send(200, page_html(self.state.duration_seconds).encode("utf-8"), "text/html; charset=utf-8")
+            return
+        if self.path == "/command":
+            self._send(200, json.dumps(self.state.snapshot()).encode("utf-8"))
+            return
+        if self.path == "/status":
+            self._send(200, json.dumps(self.state.snapshot()).encode("utf-8"))
+            return
+        self._send(404, b"not found", "text/plain; charset=utf-8")
+
+    def do_POST(self) -> None:
+        length = int(self.headers.get("Content-Length", "0"))
+        body = self.rfile.read(length)
+        if self.path == "/status":
+            payload = json.loads(body.decode("utf-8"))
+            self.state.update(payload)
+            self._send(200, json.dumps(self.state.snapshot()).encode("utf-8"))
+            return
+        if self.path == "/start":
+            self._send(200, json.dumps(self.state.request_start()).encode("utf-8"))
+            return
+        if self.path == "/upload":
+            self._send(200, json.dumps(self.state.store_upload(body)).encode("utf-8"))
+            return
+        self._send(404, b"not found", "text/plain; charset=utf-8")
+
+    def log_message(self, fmt: str, *args) -> None:
+        pass
+
+
+def main() -> None:
+    args = parse_args()
+    state = ProbeState(Path(args.output), Path(args.status), args.duration_seconds)
+
+    class Handler(ProbeHandler):
+        pass
+
+    Handler.state = state
+    with socketserver.TCPServer((args.host, args.port), Handler) as httpd:
+        httpd.serve_forever()
+
+
+if __name__ == "__main__":
+    main()

scripts/manual/run_upstream_av_sync.sh

@@ -0,0 +1,102 @@
+#!/usr/bin/env bash
+# scripts/manual/run_upstream_av_sync.sh
+#
+# Manual: capture the real Tethys webcam/mic endpoints while the shared-clock
+# sync probe streams upstream media through Lesavka, then analyze the skew.
+
+set -euo pipefail
+
+SCRIPT_DIR="$(cd -- "$(dirname "${BASH_SOURCE[0]}")" >/dev/null 2>&1 && pwd)"
+REPO_ROOT="$(cd -- "${SCRIPT_DIR}/../.." >/dev/null 2>&1 && pwd)"
+
+TETHYS_HOST=${TETHYS_HOST:-tethys}
+LESAVKA_SERVER_ADDR=${LESAVKA_SERVER_ADDR:-http://38.28.125.112:50051}
+PROBE_DURATION_SECONDS=${PROBE_DURATION_SECONDS:-10}
+PROBE_WARMUP_SECONDS=${PROBE_WARMUP_SECONDS:-4}
+LEAD_IN_SECONDS=${LEAD_IN_SECONDS:-8}
+TAIL_SECONDS=${TAIL_SECONDS:-2}
+CAPTURE_SECONDS=${CAPTURE_SECONDS:-$((PROBE_DURATION_SECONDS + PROBE_WARMUP_SECONDS + LEAD_IN_SECONDS + TAIL_SECONDS))}
+REMOTE_CAPTURE=${REMOTE_CAPTURE:-/tmp/lesavka-upstream-av-sync.mkv}
+LOCAL_OUTPUT_DIR=${LOCAL_OUTPUT_DIR:-"${REPO_ROOT}/tmp"}
+VIDEO_SIZE=${VIDEO_SIZE:-1280x720}
+VIDEO_FPS=${VIDEO_FPS:-30}
+VIDEO_FORMAT=${VIDEO_FORMAT:-mjpeg}
+SSH_OPTS=${SSH_OPTS:-"-o BatchMode=yes -o ConnectTimeout=5"}
+
+mkdir -p "${LOCAL_OUTPUT_DIR}"
+STAMP="$(date +%Y%m%d-%H%M%S)"
+LOCAL_CAPTURE="${LOCAL_OUTPUT_DIR}/lesavka-upstream-av-sync-${STAMP}.mkv"
+
+echo "==> starting Tethys capture on ${TETHYS_HOST}"
+ssh ${SSH_OPTS} "${TETHYS_HOST}" bash -s -- \
+  "${REMOTE_CAPTURE}" \
+  "${CAPTURE_SECONDS}" \
+  "${VIDEO_SIZE}" \
+  "${VIDEO_FPS}" \
+  "${VIDEO_FORMAT}" <<'REMOTE_CAPTURE_SCRIPT' &
+set -euo pipefail
+remote_capture=$1
+capture_seconds=$2
+video_size=$3
+video_fps=$4
+video_format=$5
+
+rm -f "${remote_capture}"
+video_args=(-f video4linux2 -framerate "${video_fps}" -video_size "${video_size}")
+if [[ -n "${video_format}" ]]; then
+  video_args+=(-input_format "${video_format}")
+fi
+
+ffmpeg -hide_banner -loglevel error -y \
+  -thread_queue_size 1024 \
+  "${video_args[@]}" \
+  -i /dev/video0 \
+  -thread_queue_size 1024 \
+  -f alsa -ac 2 -ar 48000 \
+  -i hw:3,0 \
+  -t "${capture_seconds}" \
+  -c:v ffv1 -level 3 -g 1 \
+  -c:a pcm_s16le \
+  "${remote_capture}"
+REMOTE_CAPTURE_SCRIPT
+capture_pid=$!
+
+sleep "${LEAD_IN_SECONDS}"
+
+echo "==> running local Lesavka sync probe against ${LESAVKA_SERVER_ADDR}"
+probe_status=0
+(
+  cd "${REPO_ROOT}"
+  cargo run -p lesavka_client --bin lesavka-sync-probe -- \
+    --server "${LESAVKA_SERVER_ADDR}" \
+    --duration-seconds "${PROBE_DURATION_SECONDS}" \
+    --warmup-seconds "${PROBE_WARMUP_SECONDS}"
+) || probe_status=$?
+
+capture_status=0
+wait "${capture_pid}" || capture_status=$?
+
+if ssh ${SSH_OPTS} "${TETHYS_HOST}" "test -f '${REMOTE_CAPTURE}'"; then
+  echo "==> fetching capture back to ${LOCAL_CAPTURE}"
+  scp ${SSH_OPTS} "${TETHYS_HOST}:${REMOTE_CAPTURE}" "${LOCAL_CAPTURE}"
+fi
+
+if [[ "${probe_status}" -ne 0 ]]; then
+  echo "sync probe failed with status ${probe_status}" >&2
+  [[ -f "${LOCAL_CAPTURE}" ]] && echo "partial capture preserved at ${LOCAL_CAPTURE}" >&2
+  exit "${probe_status}"
+fi
+if [[ "${capture_status}" -ne 0 ]]; then
+  echo "Tethys capture failed with status ${capture_status}" >&2
+  [[ -f "${LOCAL_CAPTURE}" ]] && echo "partial capture preserved at ${LOCAL_CAPTURE}" >&2
+  exit "${capture_status}"
+fi
+
+echo "==> analyzing capture"
+(
+  cd "${REPO_ROOT}"
+  cargo run -p lesavka_client --bin lesavka-sync-analyze -- "${LOCAL_CAPTURE}"
+)
+
+echo "==> done"
+echo "capture: ${LOCAL_CAPTURE}"

scripts/manual/run_upstream_browser_av_sync.sh

@@ -0,0 +1,168 @@
+#!/usr/bin/env bash
+# scripts/manual/run_upstream_browser_av_sync.sh
+#
+# Drive a real browser consumer on Tethys, record the combined MediaStream,
+# pull the capture back, and analyze it with the Lesavka sync analyzer.
+
+set -euo pipefail
+
+SCRIPT_DIR="$(cd -- "$(dirname "${BASH_SOURCE[0]}")" >/dev/null 2>&1 && pwd)"
+REPO_ROOT="$(cd -- "${SCRIPT_DIR}/../.." >/dev/null 2>&1 && pwd)"
+
+TETHYS_HOST=${TETHYS_HOST:-tethys}
+LESAVKA_SERVER_ADDR=${LESAVKA_SERVER_ADDR:-http://38.28.125.112:50051}
+PROBE_DURATION_SECONDS=${PROBE_DURATION_SECONDS:-15}
+BROWSER_PORT=${BROWSER_PORT:-18443}
+REMOTE_SCRIPT=${REMOTE_SCRIPT:-/tmp/lesavka-browser-consumer-probe.py}
+REMOTE_CAPTURE=${REMOTE_CAPTURE:-/tmp/lesavka-browser-av-sync.webm}
+REMOTE_STATUS=${REMOTE_STATUS:-/tmp/lesavka-browser-av-sync-status.json}
+REMOTE_PROFILE_DIR=${REMOTE_PROFILE_DIR:-/tmp/lesavka-browser-probe-profile}
+LOCAL_OUTPUT_DIR=${LOCAL_OUTPUT_DIR:-"${REPO_ROOT}/tmp"}
+SSH_OPTS=${SSH_OPTS:-"-o BatchMode=yes -o ConnectTimeout=5"}
+DISPLAY_ENV=${DISPLAY_ENV:-":0"}
+REMOTE_RUNTIME_DIR=${REMOTE_RUNTIME_DIR:-/run/user/1000}
+REMOTE_DBUS_ADDRESS=${REMOTE_DBUS_ADDRESS:-}
+REMOTE_XAUTHORITY=${REMOTE_XAUTHORITY:-}
+READY_TIMEOUT_SECONDS=${READY_TIMEOUT_SECONDS:-120}
+
+mkdir -p "${LOCAL_OUTPUT_DIR}"
+STAMP="$(date +%Y%m%d-%H%M%S)"
+LOCAL_CAPTURE="${LOCAL_OUTPUT_DIR}/lesavka-browser-av-sync-${STAMP}.webm"
+
+scp ${SSH_OPTS} "${REPO_ROOT}/scripts/manual/browser_consumer_probe.py" "${TETHYS_HOST}:${REMOTE_SCRIPT}"
+
+ssh ${SSH_OPTS} "${TETHYS_HOST}" bash -s -- \
+  "${REMOTE_SCRIPT}" \
+  "${REMOTE_CAPTURE}" \
+  "${REMOTE_STATUS}" \
+  "${REMOTE_PROFILE_DIR}" \
+  "${PROBE_DURATION_SECONDS}" \
+  "${BROWSER_PORT}" \
+  "${DISPLAY_ENV}" \
+  "${REMOTE_RUNTIME_DIR}" <<'REMOTE_SETUP'
+set -euo pipefail
+remote_script=$1
+remote_capture=$2
+remote_status=$3
+remote_profile_dir=$4
+duration=$5
+port=$6
+display_env=$7
+runtime_dir=$8
+dbus_address=""
+xauthority_path=""
+firefox_pid="$(pgrep -n -x firefox-esr || true)"
+if [[ -n "${firefox_pid}" && -r "/proc/${firefox_pid}/environ" ]]; then
+  while IFS='=' read -r key value; do
+    case "$key" in
+      DBUS_SESSION_BUS_ADDRESS) dbus_address="$value" ;;
+      XAUTHORITY) xauthority_path="$value" ;;
+      DISPLAY) [[ -z "${display_env}" || "${display_env}" == ":0" ]] && display_env="$value" ;;
+    esac
+  done < <(tr '\0' '\n' <"/proc/${firefox_pid}/environ")
+fi
+[[ -z "${dbus_address}" ]] && dbus_address="unix:path=${runtime_dir}/bus"
+fuser -k "${port}/tcp" >/dev/null 2>&1 || true
+pkill -f "firefox.*${remote_profile_dir}" >/dev/null 2>&1 || true
+for _ in $(seq 1 20); do
+  if ! pgrep -f "firefox.*${remote_profile_dir}" >/dev/null 2>&1; then
+    break
+  fi
+  sleep 0.25
+done
+rm -f "$remote_capture" "$remote_status"
+rm -rf "$remote_profile_dir"
+mkdir -p "$remote_profile_dir"
+cat >"${remote_profile_dir}/user.js" <<'FIREFOX_PREFS'
+user_pref("media.navigator.permission.disabled", true);
+user_pref("permissions.default.camera", 1);
+user_pref("permissions.default.microphone", 1);
+user_pref("media.autoplay.default", 0);
+user_pref("media.autoplay.blocking_policy", 0);
+user_pref("toolkit.telemetry.reportingpolicy.firstRun", false);
+user_pref("browser.shell.checkDefaultBrowser", false);
+user_pref("browser.tabs.warnOnClose", false);
+user_pref("browser.startup.page", 1);
+user_pref("browser.startup.homepage_override.mstone", "ignore");
+user_pref("startup.homepage_welcome_url", "");
+user_pref("startup.homepage_welcome_url.additional", "");
+user_pref("browser.aboutwelcome.enabled", false);
+user_pref("trailhead.firstrun.didSeeAboutWelcome", true);
+FIREFOX_PREFS
+printf 'user_pref("browser.startup.homepage", "http://127.0.0.1:%s/");\n' "$port" >>"${remote_profile_dir}/user.js"
+nohup python3 "$remote_script" --port "$port" --output "$remote_capture" --status "$remote_status" --duration-seconds "$duration" >/tmp/lesavka-browser-consumer-probe.log 2>&1 &
+if [[ -n "${xauthority_path}" ]]; then
+  nohup env DISPLAY="$display_env" XDG_RUNTIME_DIR="$runtime_dir" DBUS_SESSION_BUS_ADDRESS="$dbus_address" XAUTHORITY="$xauthority_path" \
+    firefox --new-instance --no-remote --profile "$remote_profile_dir" \
+    >/tmp/lesavka-browser-consumer-firefox.log 2>&1 &
+else
+  nohup env DISPLAY="$display_env" XDG_RUNTIME_DIR="$runtime_dir" DBUS_SESSION_BUS_ADDRESS="$dbus_address" \
+    firefox --new-instance --no-remote --profile "$remote_profile_dir" \
+    >/tmp/lesavka-browser-consumer-firefox.log 2>&1 &
+fi
+REMOTE_SETUP
+
+echo "==> waiting for browser consumer to become ready on ${TETHYS_HOST}"
+deadline=$(( $(date +%s) + READY_TIMEOUT_SECONDS ))
+while true; do
+  status_json=$(ssh ${SSH_OPTS} "${TETHYS_HOST}" "test -f '${REMOTE_STATUS}' && cat '${REMOTE_STATUS}'" || true)
+  if [[ -n "${status_json}" ]]; then
+    if STATUS_JSON="${status_json}" python3 -c 'import json, os, sys; status = json.loads(os.environ["STATUS_JSON"]); sys.exit(0 if status.get("ready") else 1)'
+    then
+      echo "==> browser consumer ready"
+      break
+    fi
+  fi
+  if (( $(date +%s) >= deadline )); then
+    echo "browser consumer did not become ready before timeout" >&2
+    [[ -n "${status_json:-}" ]] && echo "last status: ${status_json}" >&2
+    exit 1
+  fi
+  sleep 1
+
+done
+
+echo "==> triggering browser recording"
+ssh ${SSH_OPTS} "${TETHYS_HOST}" "curl -fsS -X POST http://127.0.0.1:${BROWSER_PORT}/start >/dev/null"
+
+sleep 1
+
+echo "==> running local Lesavka sync probe against ${LESAVKA_SERVER_ADDR}"
+(
+  cd "${REPO_ROOT}"
+  cargo run -p lesavka_client --bin lesavka-sync-probe -- \
+    --server "${LESAVKA_SERVER_ADDR}" \
+    --duration-seconds "${PROBE_DURATION_SECONDS}"
+)
+
+echo "==> waiting for browser recording upload"
+dealine_upload=$(( $(date +%s) + PROBE_DURATION_SECONDS + 60 ))
+while true; do
+  status_json=$(ssh ${SSH_OPTS} "${TETHYS_HOST}" "test -f '${REMOTE_STATUS}' && cat '${REMOTE_STATUS}'" || true)
+  if [[ -n "${status_json}" ]]; then
+    if STATUS_JSON="${status_json}" python3 -c 'import json, os, sys; status = json.loads(os.environ["STATUS_JSON"]); sys.exit(0 if status.get("uploaded") else 1)'
+    then
+      echo "==> browser recording uploaded"
+      break
+    fi
+  fi
+  if (( $(date +%s) >= dealine_upload )); then
+    echo "browser recording was not uploaded before timeout" >&2
+    [[ -n "${status_json:-}" ]] && echo "last status: ${status_json}" >&2
+    exit 1
+  fi
+  sleep 1
+
+done
+
+echo "==> fetching capture back to ${LOCAL_CAPTURE}"
+scp ${SSH_OPTS} "${TETHYS_HOST}:${REMOTE_CAPTURE}" "${LOCAL_CAPTURE}"
+
+echo "==> analyzing browser capture"
+(
+  cd "${REPO_ROOT}"
+  cargo run -p lesavka_client --bin lesavka-sync-analyze -- "${LOCAL_CAPTURE}"
+)
+
+echo "==> done"
+echo "capture: ${LOCAL_CAPTURE}"

server/Cargo.toml

@@ -10,7 +10,7 @@ bench                   = false
 
 [package]
 name                    = "lesavka_server"
-version                 = "0.12.4"
+version                 = "0.13.0"
 edition                 = "2024"
 autobins                = false
 

server/src/audio/voice_input.rs

@@ -62,6 +62,47 @@ fn voice_input_caps() -> gst::Caps {
         .build()
 }
 
+fn voice_sink_buffer_time_us() -> i64 {
+    positive_voice_sink_timing_env("LESAVKA_UAC_BUFFER_TIME_US", 20_000)
+}
+
+fn voice_sink_latency_time_us() -> i64 {
+    positive_voice_sink_timing_env("LESAVKA_UAC_LATENCY_TIME_US", 5_000)
+}
+
+fn voice_sink_compensation_us() -> i64 {
+    std::env::var("LESAVKA_UAC_COMPENSATION_US")
+        .ok()
+        .and_then(|value| value.trim().parse::<i64>().ok())
+        .filter(|value| *value >= 0)
+        .unwrap_or_else(default_voice_sink_compensation_us)
+}
+
+fn default_voice_sink_compensation_us() -> i64 {
+    let cfg = crate::camera::current_camera_config();
+    if cfg.output == crate::camera::CameraOutput::Hdmi {
+        non_negative_voice_sink_timing_env("LESAVKA_UAC_HDMI_COMPENSATION_US", 105_000)
+    } else {
+        0
+    }
+}
+
+fn positive_voice_sink_timing_env(name: &str, default: i64) -> i64 {
+    std::env::var(name)
+        .ok()
+        .and_then(|value| value.trim().parse::<i64>().ok())
+        .filter(|value| *value > 0)
+        .unwrap_or(default)
+}
+
+fn non_negative_voice_sink_timing_env(name: &str, default: i64) -> i64 {
+    std::env::var(name)
+        .ok()
+        .and_then(|value| value.trim().parse::<i64>().ok())
+        .filter(|value| *value >= 0)
+        .unwrap_or(default)
+}
+
 impl Voice {
     #[cfg(coverage)]
     pub async fn new(_alsa_dev: &str) -> anyhow::Result<Self> {
@@ -133,11 +174,32 @@ impl Voice {
         let alsa_sink = gst::ElementFactory::make("alsasink")
             .build()
             .context("make alsasink")?;
+        let delay_queue = gst::ElementFactory::make("queue")
+            .build()
+            .context("make voice delay queue")?;
+        let buffer_time_us = voice_sink_buffer_time_us();
+        let latency_time_us = voice_sink_latency_time_us();
+        let compensation_us = voice_sink_compensation_us();
 
         alsa_sink.set_property("device", alsa_dev);
         alsa_sink.set_property("sync", false);
         alsa_sink.set_property("async", false);
         alsa_sink.set_property("enable-last-sample", false);
+        alsa_sink.set_property("provide-clock", false);
+        alsa_sink.set_property("buffer-time", buffer_time_us);
+        alsa_sink.set_property("latency-time", latency_time_us);
+        let compensation_ns = (compensation_us.max(0) as u64).saturating_mul(1_000);
+        delay_queue.set_property("max-size-buffers", 0u32);
+        delay_queue.set_property("max-size-bytes", 0u32);
+        delay_queue.set_property("max-size-time", compensation_ns);
+        delay_queue.set_property("min-threshold-time", compensation_ns);
+        tracing::info!(
+            %alsa_dev,
+            buffer_time_us,
+            latency_time_us,
+            compensation_us,
+            "🎤 UAC sink low-latency timing armed"
+        );
 
         pipeline.add_many([
             appsrc.upcast_ref(),
@@ -145,10 +207,11 @@ impl Voice {
             &convert,
             &resample,
             &capsfilter,
+            &delay_queue,
             &alsa_sink,
         ])?;
         appsrc.link(&decodebin)?;
-        gst::Element::link_many([&convert, &resample, &capsfilter, &alsa_sink])?;
+        gst::Element::link_many([&convert, &resample, &capsfilter, &delay_queue, &alsa_sink])?;
 
         /*------------ decodebin autolink ----------------*/
         let convert_sink = convert
@@ -202,3 +265,94 @@ impl Voice {
         let _ = self.appsrc.end_of_stream();
     }
 }
+
+#[cfg(test)]
+mod voice_sink_timing_tests {
+    use crate::camera::update_camera_config;
+    use super::{voice_sink_buffer_time_us, voice_sink_latency_time_us};
+    use super::{default_voice_sink_compensation_us, voice_sink_compensation_us};
+
+    #[test]
+    fn voice_sink_timing_defaults_stay_live_call_friendly() {
+        temp_env::with_var_unset("LESAVKA_UAC_BUFFER_TIME_US", || {
+            temp_env::with_var_unset("LESAVKA_UAC_LATENCY_TIME_US", || {
+                temp_env::with_var_unset("LESAVKA_UAC_COMPENSATION_US", || {
+                    temp_env::with_var_unset("LESAVKA_UAC_HDMI_COMPENSATION_US", || {
+                        temp_env::with_var("LESAVKA_CAM_OUTPUT", Some("uvc"), || {
+                            update_camera_config();
+                            assert_eq!(voice_sink_buffer_time_us(), 20_000);
+                            assert_eq!(voice_sink_latency_time_us(), 5_000);
+                            assert_eq!(voice_sink_compensation_us(), 0);
+                        });
+                    });
+                });
+            });
+        });
+    }
+
+    #[test]
+    fn voice_sink_timing_env_accepts_positive_overrides_only() {
+        temp_env::with_var("LESAVKA_CAM_OUTPUT", Some("uvc"), || {
+            update_camera_config();
+            temp_env::with_var("LESAVKA_UAC_BUFFER_TIME_US", Some("42000"), || {
+                temp_env::with_var("LESAVKA_UAC_LATENCY_TIME_US", Some("7000"), || {
+                    assert_eq!(voice_sink_buffer_time_us(), 42_000);
+                    assert_eq!(voice_sink_latency_time_us(), 7_000);
+                    assert_eq!(voice_sink_compensation_us(), 0);
+                });
+            });
+        });
+
+        temp_env::with_var("LESAVKA_CAM_OUTPUT", Some("uvc"), || {
+            update_camera_config();
+            temp_env::with_var("LESAVKA_UAC_BUFFER_TIME_US", Some("0"), || {
+                temp_env::with_var("LESAVKA_UAC_LATENCY_TIME_US", Some("-5"), || {
+                    temp_env::with_var("LESAVKA_UAC_COMPENSATION_US", Some("166667"), || {
+                        assert_eq!(voice_sink_buffer_time_us(), 20_000);
+                        assert_eq!(voice_sink_latency_time_us(), 5_000);
+                        assert_eq!(voice_sink_compensation_us(), 166_667);
+                    });
+                });
+            });
+        });
+
+        temp_env::with_var("LESAVKA_CAM_OUTPUT", Some("uvc"), || {
+            update_camera_config();
+            temp_env::with_var("LESAVKA_UAC_COMPENSATION_US", Some("-5"), || {
+                temp_env::with_var("LESAVKA_UAC_BUFFER_TIME_US", Some("0"), || {
+                    temp_env::with_var("LESAVKA_UAC_LATENCY_TIME_US", Some("-5"), || {
+                        assert_eq!(voice_sink_buffer_time_us(), 20_000);
+                        assert_eq!(voice_sink_latency_time_us(), 5_000);
+                        assert_eq!(voice_sink_compensation_us(), 0);
+                    });
+                });
+            });
+        });
+    }
+
+    #[test]
+    fn hdmi_sink_compensation_defaults_to_hdmi_specific_delay() {
+        temp_env::with_var_unset("LESAVKA_UAC_COMPENSATION_US", || {
+            temp_env::with_var_unset("LESAVKA_UAC_HDMI_COMPENSATION_US", || {
+                temp_env::with_var("LESAVKA_CAM_OUTPUT", Some("hdmi"), || {
+                    update_camera_config();
+                    assert_eq!(default_voice_sink_compensation_us(), 105_000);
+                    assert_eq!(voice_sink_compensation_us(), 105_000);
+                });
+            });
+        });
+    }
+
+    #[test]
+    fn explicit_compensation_override_wins_over_hdmi_default() {
+        temp_env::with_var("LESAVKA_CAM_OUTPUT", Some("hdmi"), || {
+            update_camera_config();
+            temp_env::with_var("LESAVKA_UAC_HDMI_COMPENSATION_US", Some("120000"), || {
+                temp_env::with_var("LESAVKA_UAC_COMPENSATION_US", Some("90000"), || {
+                    assert_eq!(default_voice_sink_compensation_us(), 120_000);
+                    assert_eq!(voice_sink_compensation_us(), 90_000);
+                });
+            });
+        });
+    }
+}

server/src/camera.rs

@@ -2,11 +2,11 @@
 
 #![cfg_attr(coverage, allow(dead_code, unused_imports, unused_variables))]
 
-use gstreamer as gst;
-use std::collections::HashMap;
-use std::fs;
+mod selection;
+
 use std::sync::{OnceLock, RwLock};
-use tracing::{info, warn};
+
+use selection::select_camera_config;
 
 #[derive(Clone, Copy, Debug, PartialEq, Eq)]
 pub enum CameraOutput {
@@ -76,15 +76,15 @@ impl CameraConfig {
         }
 
         if let (Some(width), Some(height)) = (
-            read_u32_from_env("LESAVKA_HDMI_WIDTH"),
-            read_u32_from_env("LESAVKA_HDMI_HEIGHT"),
+            selection::read_u32_from_env("LESAVKA_HDMI_WIDTH"),
+            selection::read_u32_from_env("LESAVKA_HDMI_HEIGHT"),
         ) {
             return (width, height);
         }
 
         self.hdmi
             .as_ref()
-            .and_then(|hdmi| preferred_hdmi_mode(&hdmi.modes))
+            .and_then(|hdmi| selection::preferred_hdmi_mode(&hdmi.modes))
             .map(|mode| (mode.width, mode.height))
             .unwrap_or((self.width, self.height))
     }
@@ -127,345 +127,6 @@ pub fn current_camera_config() -> CameraConfig {
     update_camera_config()
 }
 
-#[cfg(coverage)]
-fn select_camera_config() -> CameraConfig {
-    let output_override = std::env::var("LESAVKA_CAM_OUTPUT")
-        .ok()
-        .as_deref()
-        .and_then(parse_camera_output);
-
-    match output_override.unwrap_or(CameraOutput::Uvc) {
-        CameraOutput::Hdmi => select_hdmi_config(detect_hdmi_connector(false)),
-        CameraOutput::Uvc => select_uvc_config(),
-    }
-}
-
-#[cfg(not(coverage))]
-fn select_camera_config() -> CameraConfig {
-    let output_env = std::env::var("LESAVKA_CAM_OUTPUT").ok();
-    let output_override = output_env.as_deref().and_then(parse_camera_output);
-
-    let require_connected = output_override != Some(CameraOutput::Hdmi);
-    let hdmi = detect_hdmi_connector(require_connected);
-    if output_override == Some(CameraOutput::Hdmi) && hdmi.is_none() {
-        warn!("📷 HDMI output forced but no connector detected");
-    }
-    let output = match output_override {
-        Some(v) => v,
-        None => {
-            if hdmi.is_some() {
-                CameraOutput::Hdmi
-            } else {
-                CameraOutput::Uvc
-            }
-        }
-    };
-
-    let cfg = match output {
-        CameraOutput::Hdmi => select_hdmi_config(hdmi),
-        CameraOutput::Uvc => select_uvc_config(),
-    };
-
-    let (display_width, display_height) = cfg.hdmi_display_size();
-    info!(
-        output = cfg.output.as_str(),
-        codec = cfg.codec.as_str(),
-        width = cfg.width,
-        height = cfg.height,
-        fps = cfg.fps,
-        display_width,
-        display_height,
-        hdmi = cfg.hdmi.as_ref().map(|h| h.name.as_str()).unwrap_or("none"),
-        "📷 camera output selected"
-    );
-
-    cfg
-}
-
-fn parse_camera_output(raw: &str) -> Option<CameraOutput> {
-    match raw.trim().to_ascii_lowercase().as_str() {
-        "uvc" => Some(CameraOutput::Uvc),
-        "hdmi" => Some(CameraOutput::Hdmi),
-        "auto" | "" => None,
-        _ => None,
-    }
-}
-
-fn select_hdmi_config(hdmi: Option<HdmiConnector>) -> CameraConfig {
-    let hw_decode = has_hw_h264_decode();
-    let (default_width, default_height) = if hw_decode { (1920, 1080) } else { (1280, 720) };
-    let width = read_u32_from_env("LESAVKA_CAM_WIDTH").unwrap_or(default_width);
-    let height = read_u32_from_env("LESAVKA_CAM_HEIGHT").unwrap_or(default_height);
-    let fps = read_u32_from_env("LESAVKA_CAM_FPS").unwrap_or(30).max(1);
-    #[cfg(not(coverage))]
-    if !hw_decode {
-        if width == default_width && height == default_height {
-            warn!(
-                "📷 HDMI output: hardware H264 decoder not detected; requesting 720p30 camera uplink"
-            );
-        } else {
-            warn!(
-                width,
-                height,
-                fps,
-                "📷 HDMI output: hardware H264 decoder not detected; using configured camera uplink size"
-            );
-        }
-    }
-    CameraConfig {
-        output: CameraOutput::Hdmi,
-        codec: CameraCodec::H264,
-        width,
-        height,
-        fps,
-        hdmi,
-    }
-}
-
-#[cfg(coverage)]
-fn select_uvc_config() -> CameraConfig {
-    let width = read_u32_from_env("LESAVKA_UVC_WIDTH").unwrap_or(1280);
-    let height = read_u32_from_env("LESAVKA_UVC_HEIGHT").unwrap_or(720);
-    let fps = read_u32_from_env("LESAVKA_UVC_FPS")
-        .or_else(|| {
-            read_u32_from_env("LESAVKA_UVC_INTERVAL").and_then(|interval| {
-                if interval == 0 {
-                    None
-                } else {
-                    Some(10_000_000 / interval)
-                }
-            })
-        })
-        .unwrap_or(25);
-
-    CameraConfig {
-        output: CameraOutput::Uvc,
-        codec: CameraCodec::Mjpeg,
-        width,
-        height,
-        fps,
-        hdmi: None,
-    }
-}
-
-#[cfg(not(coverage))]
-fn select_uvc_config() -> CameraConfig {
-    let mut uvc_env = HashMap::new();
-    if let Ok(text) = fs::read_to_string("/etc/lesavka/uvc.env") {
-        uvc_env = parse_env_file(&text);
-    }
-
-    let width = read_u32_from_env("LESAVKA_UVC_WIDTH")
-        .or_else(|| read_u32_from_map(&uvc_env, "LESAVKA_UVC_WIDTH"))
-        .unwrap_or(1280);
-    let height = read_u32_from_env("LESAVKA_UVC_HEIGHT")
-        .or_else(|| read_u32_from_map(&uvc_env, "LESAVKA_UVC_HEIGHT"))
-        .unwrap_or(720);
-    let fps = read_u32_from_env("LESAVKA_UVC_FPS")
-        .or_else(|| read_u32_from_map(&uvc_env, "LESAVKA_UVC_FPS"))
-        .or_else(|| {
-            read_u32_from_env("LESAVKA_UVC_INTERVAL")
-                .or_else(|| read_u32_from_map(&uvc_env, "LESAVKA_UVC_INTERVAL"))
-                .and_then(|interval| {
-                    if interval == 0 {
-                        None
-                    } else {
-                        Some(10_000_000 / interval)
-                    }
-                })
-        })
-        .unwrap_or(25);
-
-    CameraConfig {
-        output: CameraOutput::Uvc,
-        codec: CameraCodec::Mjpeg,
-        width,
-        height,
-        fps,
-        hdmi: None,
-    }
-}
-
-#[cfg(coverage)]
-fn has_hw_h264_decode() -> bool {
-    std::env::var("LESAVKA_HW_H264").is_ok()
-}
-
-#[cfg(not(coverage))]
-fn has_hw_h264_decode() -> bool {
-    if gst::init().is_err() {
-        return false;
-    }
-    for name in ["v4l2h264dec", "v4l2slh264dec", "omxh264dec"] {
-        if gst::ElementFactory::find(name).is_some() {
-            return true;
-        }
-    }
-    false
-}
-
-#[cfg(coverage)]
-fn detect_hdmi_connector(require_connected: bool) -> Option<HdmiConnector> {
-    let _ = require_connected;
-    std::env::var("LESAVKA_HDMI_CONNECTOR")
-        .ok()
-        .map(|name| HdmiConnector {
-            name,
-            id: None,
-            modes: std::env::var("LESAVKA_HDMI_MODES")
-                .ok()
-                .map(|raw| parse_hdmi_modes(&raw))
-                .unwrap_or_default(),
-        })
-}
-
-#[cfg(not(coverage))]
-fn detect_hdmi_connector(require_connected: bool) -> Option<HdmiConnector> {
-    let preferred = std::env::var("LESAVKA_HDMI_CONNECTOR").ok();
-    let entries = fs::read_dir("/sys/class/drm").ok()?;
-    let mut connectors = Vec::new();
-
-    for entry in entries.flatten() {
-        let name = entry.file_name().to_string_lossy().into_owned();
-        if !name.contains("HDMI-A-") {
-            continue;
-        }
-        let status_path = entry.path().join("status");
-        let status = fs::read_to_string(&status_path)
-            .ok()
-            .map(|v| v.trim().to_string())
-            .unwrap_or_default();
-        let id = fs::read_to_string(entry.path().join("connector_id"))
-            .ok()
-            .and_then(|v| v.trim().parse::<u32>().ok());
-        let modes = fs::read_to_string(entry.path().join("modes"))
-            .ok()
-            .map(|raw| parse_hdmi_modes(&raw))
-            .unwrap_or_default();
-        connectors.push((name, status, id, modes));
-    }
-    connectors.sort_by(|a, b| a.0.cmp(&b.0));
-
-    let matches_preferred =
-        |name: &str, preferred: &str| name == preferred || name.ends_with(preferred);
-
-    if let Some(pref) = preferred.as_deref() {
-        for (name, status, id, modes) in &connectors {
-            if matches_preferred(name, pref) && (!require_connected || status == "connected") {
-                return Some(HdmiConnector {
-                    name: name.clone(),
-                    id: *id,
-                    modes: modes.clone(),
-                });
-            }
-        }
-    }
-
-    // Keep the previously-selected connector stable when no explicit override is set.
-    // This prevents connector flapping when multiple HDMI outputs are simultaneously connected.
-    if preferred.is_none() {
-        let previous = LAST_CONFIG
-            .get()
-            .and_then(|lock| lock.read().ok())
-            .and_then(|cfg| cfg.hdmi.as_ref().map(|h| h.name.clone()));
-        if let Some(prev) = previous {
-            for (name, status, id, modes) in &connectors {
-                if *name == prev && (!require_connected || status == "connected") {
-                    return Some(HdmiConnector {
-                        name: name.clone(),
-                        id: *id,
-                        modes: modes.clone(),
-                    });
-                }
-            }
-        }
-    }
-
-    for (name, status, id, modes) in connectors {
-        if !require_connected || status == "connected" {
-            return Some(HdmiConnector { name, id, modes });
-        }
-    }
-
-    None
-}
-
-fn parse_hdmi_modes(raw: &str) -> Vec<HdmiMode> {
-    raw.lines()
-        .flat_map(|line| line.split(','))
-        .filter_map(parse_hdmi_mode)
-        .collect()
-}
-
-fn parse_hdmi_mode(raw: &str) -> Option<HdmiMode> {
-    let raw = raw.trim();
-    let (width, rest) = raw.split_once('x')?;
-    let width = width.trim().parse::<u32>().ok()?;
-    let height_digits: String = rest
-        .trim()
-        .chars()
-        .take_while(|ch| ch.is_ascii_digit())
-        .collect();
-    let height = height_digits.parse::<u32>().ok()?;
-    (width > 0 && height > 0).then_some(HdmiMode { width, height })
-}
-
-fn preferred_hdmi_mode(modes: &[HdmiMode]) -> Option<HdmiMode> {
-    for preferred in [
-        HdmiMode {
-            width: 1920,
-            height: 1080,
-        },
-        HdmiMode {
-            width: 1280,
-            height: 720,
-        },
-    ] {
-        if modes.contains(&preferred) {
-            return Some(preferred);
-        }
-    }
-
-    modes
-        .iter()
-        .copied()
-        .filter(|mode| mode.width.saturating_mul(9) == mode.height.saturating_mul(16))
-        .filter(|mode| mode.width.saturating_mul(mode.height) <= 1920 * 1080)
-        .max_by_key(|mode| mode.width.saturating_mul(mode.height))
-        .or_else(|| modes.first().copied())
-}
-
-#[cfg(not(coverage))]
-fn parse_env_file(text: &str) -> HashMap<String, String> {
-    let mut out = HashMap::new();
-    for line in text.lines() {
-        let line = line.trim();
-        if line.is_empty() || line.starts_with('#') {
-            continue;
-        }
-        let mut parts = line.splitn(2, '=');
-        let key = match parts.next() {
-            Some(v) => v.trim(),
-            None => continue,
-        };
-        let val = match parts.next() {
-            Some(v) => v.trim(),
-            None => continue,
-        };
-        out.insert(key.to_string(), val.to_string());
-    }
-    out
-}
-
-fn read_u32_from_env(key: &str) -> Option<u32> {
-    std::env::var(key).ok().and_then(|v| v.parse::<u32>().ok())
-}
-
-#[cfg(not(coverage))]
-fn read_u32_from_map(map: &HashMap<String, String>, key: &str) -> Option<u32> {
-    map.get(key).and_then(|v| v.parse::<u32>().ok())
-}
-
 #[cfg(test)]
 #[path = "tests/camera.rs"]
 mod tests;

server/src/camera/selection.rs

@@ -0,0 +1,372 @@
+use super::{CameraCodec, CameraConfig, CameraOutput, HdmiConnector, HdmiMode, LAST_CONFIG};
+use gstreamer as gst;
+use std::collections::HashMap;
+use std::fs;
+use tracing::{info, warn};
+
+#[cfg(coverage)]
+pub(super) fn select_camera_config() -> CameraConfig {
+    let output_override = std::env::var("LESAVKA_CAM_OUTPUT")
+        .ok()
+        .as_deref()
+        .and_then(parse_camera_output);
+
+    match output_override.unwrap_or(CameraOutput::Uvc) {
+        CameraOutput::Hdmi => select_hdmi_config(detect_hdmi_connector(false)),
+        CameraOutput::Uvc => select_uvc_config(),
+    }
+}
+
+#[cfg(not(coverage))]
+pub(super) fn select_camera_config() -> CameraConfig {
+    let output_env = std::env::var("LESAVKA_CAM_OUTPUT").ok();
+    let output_override = output_env.as_deref().and_then(parse_camera_output);
+
+    let require_connected = output_override != Some(CameraOutput::Hdmi);
+    let hdmi = detect_hdmi_connector(require_connected);
+    if output_override == Some(CameraOutput::Hdmi) && hdmi.is_none() {
+        warn!("📷 HDMI output forced but no connector detected");
+    }
+    let output = match output_override {
+        Some(v) => v,
+        None => {
+            if hdmi.is_some() {
+                CameraOutput::Hdmi
+            } else {
+                CameraOutput::Uvc
+            }
+        }
+    };
+
+    let cfg = match output {
+        CameraOutput::Hdmi => select_hdmi_config(hdmi),
+        CameraOutput::Uvc => select_uvc_config(),
+    };
+
+    let (display_width, display_height) = cfg.hdmi_display_size();
+    info!(
+        output = cfg.output.as_str(),
+        codec = cfg.codec.as_str(),
+        width = cfg.width,
+        height = cfg.height,
+        fps = cfg.fps,
+        display_width,
+        display_height,
+        hdmi = cfg.hdmi.as_ref().map(|h| h.name.as_str()).unwrap_or("none"),
+        "📷 camera output selected"
+    );
+
+    cfg
+}
+
+fn parse_camera_output(raw: &str) -> Option<CameraOutput> {
+    match raw.trim().to_ascii_lowercase().as_str() {
+        "uvc" => Some(CameraOutput::Uvc),
+        "hdmi" => Some(CameraOutput::Hdmi),
+        "auto" | "" => None,
+        _ => None,
+    }
+}
+
+fn parse_camera_codec(raw: &str) -> Option<CameraCodec> {
+    match raw.trim().to_ascii_lowercase().as_str() {
+        "h264" => Some(CameraCodec::H264),
+        "mjpeg" | "mjpg" | "jpeg" => Some(CameraCodec::Mjpeg),
+        _ => None,
+    }
+}
+
+fn select_hdmi_codec(hw_decode: bool) -> CameraCodec {
+    std::env::var("LESAVKA_CAM_CODEC")
+        .ok()
+        .as_deref()
+        .and_then(parse_camera_codec)
+        .unwrap_or_else(|| {
+            if hw_decode {
+                CameraCodec::H264
+            } else {
+                CameraCodec::Mjpeg
+            }
+        })
+}
+
+fn select_hdmi_config(hdmi: Option<HdmiConnector>) -> CameraConfig {
+    let hw_decode = has_hw_h264_decode();
+    let (default_width, default_height) = if hw_decode { (1920, 1080) } else { (1280, 720) };
+    let width = read_u32_from_env("LESAVKA_CAM_WIDTH").unwrap_or(default_width);
+    let height = read_u32_from_env("LESAVKA_CAM_HEIGHT").unwrap_or(default_height);
+    let fps = read_u32_from_env("LESAVKA_CAM_FPS").unwrap_or(30).max(1);
+    let codec = select_hdmi_codec(hw_decode);
+    #[cfg(not(coverage))]
+    if !hw_decode {
+        if matches!(codec, CameraCodec::Mjpeg) {
+            warn!(
+                width,
+                height,
+                fps,
+                "📷 HDMI output: hardware H264 decoder not detected; preferring MJPEG uplink"
+            );
+        } else if width == default_width && height == default_height {
+            warn!(
+                "📷 HDMI output: hardware H264 decoder not detected; forcing H264 uplink at requested size"
+            );
+        } else {
+            warn!(
+                width,
+                height,
+                fps,
+                "📷 HDMI output: hardware H264 decoder not detected; using configured camera uplink size"
+            );
+        }
+    }
+    CameraConfig {
+        output: CameraOutput::Hdmi,
+        codec,
+        width,
+        height,
+        fps,
+        hdmi,
+    }
+}
+
+#[cfg(coverage)]
+fn select_uvc_config() -> CameraConfig {
+    let width = read_u32_from_env("LESAVKA_UVC_WIDTH").unwrap_or(1280);
+    let height = read_u32_from_env("LESAVKA_UVC_HEIGHT").unwrap_or(720);
+    let fps = read_u32_from_env("LESAVKA_UVC_FPS")
+        .or_else(|| {
+            read_u32_from_env("LESAVKA_UVC_INTERVAL").and_then(|interval| {
+                if interval == 0 {
+                    None
+                } else {
+                    Some(10_000_000 / interval)
+                }
+            })
+        })
+        .unwrap_or(25);
+
+    CameraConfig {
+        output: CameraOutput::Uvc,
+        codec: CameraCodec::Mjpeg,
+        width,
+        height,
+        fps,
+        hdmi: None,
+    }
+}
+
+#[cfg(not(coverage))]
+fn select_uvc_config() -> CameraConfig {
+    let mut uvc_env = HashMap::new();
+    if let Ok(text) = fs::read_to_string("/etc/lesavka/uvc.env") {
+        uvc_env = parse_env_file(&text);
+    }
+
+    let width = read_u32_from_env("LESAVKA_UVC_WIDTH")
+        .or_else(|| read_u32_from_map(&uvc_env, "LESAVKA_UVC_WIDTH"))
+        .unwrap_or(1280);
+    let height = read_u32_from_env("LESAVKA_UVC_HEIGHT")
+        .or_else(|| read_u32_from_map(&uvc_env, "LESAVKA_UVC_HEIGHT"))
+        .unwrap_or(720);
+    let fps = read_u32_from_env("LESAVKA_UVC_FPS")
+        .or_else(|| read_u32_from_map(&uvc_env, "LESAVKA_UVC_FPS"))
+        .or_else(|| {
+            read_u32_from_env("LESAVKA_UVC_INTERVAL")
+                .or_else(|| read_u32_from_map(&uvc_env, "LESAVKA_UVC_INTERVAL"))
+                .and_then(|interval| {
+                    if interval == 0 {
+                        None
+                    } else {
+                        Some(10_000_000 / interval)
+                    }
+                })
+        })
+        .unwrap_or(25);
+
+    CameraConfig {
+        output: CameraOutput::Uvc,
+        codec: CameraCodec::Mjpeg,
+        width,
+        height,
+        fps,
+        hdmi: None,
+    }
+}
+
+#[cfg(coverage)]
+fn has_hw_h264_decode() -> bool {
+    std::env::var("LESAVKA_HW_H264").is_ok()
+}
+
+#[cfg(not(coverage))]
+fn has_hw_h264_decode() -> bool {
+    if gst::init().is_err() {
+        return false;
+    }
+    for name in ["v4l2h264dec", "v4l2slh264dec", "omxh264dec"] {
+        if gst::ElementFactory::find(name).is_some() {
+            return true;
+        }
+    }
+    false
+}
+
+#[cfg(coverage)]
+fn detect_hdmi_connector(require_connected: bool) -> Option<HdmiConnector> {
+    let _ = require_connected;
+    std::env::var("LESAVKA_HDMI_CONNECTOR")
+        .ok()
+        .map(|name| HdmiConnector {
+            name,
+            id: None,
+            modes: std::env::var("LESAVKA_HDMI_MODES")
+                .ok()
+                .map(|raw| parse_hdmi_modes(&raw))
+                .unwrap_or_default(),
+        })
+}
+
+#[cfg(not(coverage))]
+fn detect_hdmi_connector(require_connected: bool) -> Option<HdmiConnector> {
+    let preferred = std::env::var("LESAVKA_HDMI_CONNECTOR").ok();
+    let entries = fs::read_dir("/sys/class/drm").ok()?;
+    let mut connectors = Vec::new();
+
+    for entry in entries.flatten() {
+        let name = entry.file_name().to_string_lossy().into_owned();
+        if !name.contains("HDMI-A-") {
+            continue;
+        }
+        let status_path = entry.path().join("status");
+        let status = fs::read_to_string(&status_path)
+            .ok()
+            .map(|v| v.trim().to_string())
+            .unwrap_or_default();
+        let id = fs::read_to_string(entry.path().join("connector_id"))
+            .ok()
+            .and_then(|v| v.trim().parse::<u32>().ok());
+        let modes = fs::read_to_string(entry.path().join("modes"))
+            .ok()
+            .map(|raw| parse_hdmi_modes(&raw))
+            .unwrap_or_default();
+        connectors.push((name, status, id, modes));
+    }
+    connectors.sort_by(|a, b| a.0.cmp(&b.0));
+
+    let matches_preferred =
+        |name: &str, preferred: &str| name == preferred || name.ends_with(preferred);
+
+    if let Some(pref) = preferred.as_deref() {
+        for (name, status, id, modes) in &connectors {
+            if matches_preferred(name, pref) && (!require_connected || status == "connected") {
+                return Some(HdmiConnector {
+                    name: name.clone(),
+                    id: *id,
+                    modes: modes.clone(),
+                });
+            }
+        }
+    }
+
+    if preferred.is_none() {
+        let previous = LAST_CONFIG
+            .get()
+            .and_then(|lock| lock.read().ok())
+            .and_then(|cfg| cfg.hdmi.as_ref().map(|h| h.name.clone()));
+        if let Some(prev) = previous {
+            for (name, status, id, modes) in &connectors {
+                if *name == prev && (!require_connected || status == "connected") {
+                    return Some(HdmiConnector {
+                        name: name.clone(),
+                        id: *id,
+                        modes: modes.clone(),
+                    });
+                }
+            }
+        }
+    }
+
+    for (name, status, id, modes) in connectors {
+        if !require_connected || status == "connected" {
+            return Some(HdmiConnector { name, id, modes });
+        }
+    }
+
+    None
+}
+
+pub(crate) fn parse_hdmi_modes(raw: &str) -> Vec<HdmiMode> {
+    raw.lines()
+        .flat_map(|line| line.split(','))
+        .filter_map(parse_hdmi_mode)
+        .collect()
+}
+
+pub(crate) fn parse_hdmi_mode(raw: &str) -> Option<HdmiMode> {
+    let raw = raw.trim();
+    let (width, rest) = raw.split_once('x')?;
+    let width = width.trim().parse::<u32>().ok()?;
+    let height_digits: String = rest
+        .trim()
+        .chars()
+        .take_while(|ch| ch.is_ascii_digit())
+        .collect();
+    let height = height_digits.parse::<u32>().ok()?;
+    (width > 0 && height > 0).then_some(HdmiMode { width, height })
+}
+
+pub(crate) fn preferred_hdmi_mode(modes: &[HdmiMode]) -> Option<HdmiMode> {
+    for preferred in [
+        HdmiMode {
+            width: 1920,
+            height: 1080,
+        },
+        HdmiMode {
+            width: 1280,
+            height: 720,
+        },
+    ] {
+        if modes.contains(&preferred) {
+            return Some(preferred);
+        }
+    }
+
+    modes
+        .iter()
+        .copied()
+        .filter(|mode| mode.width.saturating_mul(9) == mode.height.saturating_mul(16))
+        .filter(|mode| mode.width.saturating_mul(mode.height) <= 1920 * 1080)
+        .max_by_key(|mode| mode.width.saturating_mul(mode.height))
+        .or_else(|| modes.first().copied())
+}
+
+#[cfg(not(coverage))]
+fn parse_env_file(text: &str) -> HashMap<String, String> {
+    let mut out = HashMap::new();
+    for line in text.lines() {
+        let line = line.trim();
+        if line.is_empty() || line.starts_with('#') {
+            continue;
+        }
+        let mut parts = line.splitn(2, '=');
+        let key = match parts.next() {
+            Some(v) => v.trim(),
+            None => continue,
+        };
+        let val = match parts.next() {
+            Some(v) => v.trim(),
+            None => continue,
+        };
+        out.insert(key.to_string(), val.to_string());
+    }
+    out
+}
+
+pub(crate) fn read_u32_from_env(key: &str) -> Option<u32> {
+    std::env::var(key).ok().and_then(|v| v.parse::<u32>().ok())
+}
+
+#[cfg(not(coverage))]
+fn read_u32_from_map(map: &HashMap<String, String>, key: &str) -> Option<u32> {
+    map.get(key).and_then(|v| v.parse::<u32>().ok())
+}

server/src/lib.rs

@@ -12,6 +12,7 @@ pub mod gadget;
 pub mod handshake;
 pub mod paste;
 pub mod runtime_support;
+pub mod upstream_media_runtime;
 pub mod uvc_runtime;
 pub mod video;
 pub(crate) mod video_sinks;

server/src/main.rs

@@ -25,8 +25,8 @@ use lesavka_common::lesavka::{
 
 use lesavka_server::{
     camera, camera_runtime::CameraRuntime, capture_power::CapturePowerManager, gadget::UsbGadget,
-    handshake::HandshakeSvc, paste, runtime_support, runtime_support::init_tracing, uvc_runtime,
-    video,
+    handshake::HandshakeSvc, paste, runtime_support, runtime_support::init_tracing,
+    upstream_media_runtime::UpstreamMediaRuntime, uvc_runtime, video,
 };
 
 /*──────────────── constants ────────────────*/
@@ -65,6 +65,7 @@ struct Handler {
     gadget: UsbGadget,
     did_cycle: Arc<AtomicBool>,
     camera_rt: Arc<CameraRuntime>,
+    upstream_media_rt: Arc<UpstreamMediaRuntime>,
     capture_power: CapturePowerManager,
     eye_hubs: Arc<Mutex<HashMap<EyeHubKey, Arc<EyeHub>>>>,
 }

server/src/main/handler_startup.rs

@@ -45,6 +45,7 @@ impl Handler {
             gadget,
             did_cycle: Arc::new(AtomicBool::new(false)),
             camera_rt: Arc::new(CameraRuntime::new()),
+            upstream_media_rt: Arc::new(UpstreamMediaRuntime::new()),
             capture_power: CapturePowerManager::new(),
             eye_hubs: Arc::new(Mutex::new(HashMap::new())),
         })

server/src/main/relay_service.rs

@@ -110,23 +110,55 @@ impl Relay for Handler {
         req: Request<tonic::Streaming<AudioPacket>>,
     ) -> Result<Response<Self::StreamMicrophoneStream>, Status> {
         let rpc_id = runtime_support::next_stream_id();
-        info!(rpc_id, "🎤 stream_microphone opened");
+        let lease = self.upstream_media_rt.activate_microphone();
+        info!(rpc_id, session_id = lease.session_id, "🎤 stream_microphone opened");
+        let Some(microphone_sink_permit) = self
+            .upstream_media_rt
+            .reserve_microphone_sink(lease.generation)
+            .await
+        else {
+            info!(
+                rpc_id,
+                session_id = lease.session_id,
+                "🎤 stream_microphone stood down before the sink became available"
+            );
+            return Err(Status::aborted(
+                "microphone stream superseded before sink became available",
+            ));
+        };
         // 1 ─ build once, early
         let uac_dev = std::env::var("LESAVKA_UAC_DEV").unwrap_or_else(|_| "hw:UAC2Gadget,0".into());
         info!(%uac_dev, "🎤 stream_microphone using UAC sink");
         let mut sink = runtime_support::open_voice_with_retry(&uac_dev)
             .await
-            .map_err(|e| Status::internal(format!("{e:#}")))?;
+            .map_err(|e| {
+                self.upstream_media_rt.close_microphone(lease.generation);
+                Status::internal(format!("{e:#}"))
+            })?;
 
         // 2 ─ dummy outbound stream (same trick as before)
         let (tx, rx) = tokio::sync::mpsc::channel(1);
+        let upstream_media_rt = self.upstream_media_rt.clone();
 
         // 3 ─ drive the sink in a background task
         tokio::spawn(async move {
+            let _microphone_sink_permit = microphone_sink_permit;
             let mut inbound = req.into_inner();
             static CNT: std::sync::atomic::AtomicU64 = std::sync::atomic::AtomicU64::new(0);
 
-            while let Some(pkt) = inbound.next().await.transpose()? {
+            loop {
+                if !upstream_media_rt.is_microphone_active(lease.generation) {
+                    info!(rpc_id, session_id = lease.session_id, "🎤 stream_microphone session superseded");
+                    break;
+                }
+                let next_packet = tokio::select! {
+                    packet = inbound.next() => packet,
+                    _ = tokio::time::sleep(Duration::from_millis(50)) => continue,
+                };
+                let Some(mut pkt) = next_packet.transpose()? else {
+                    break;
+                };
+                pkt.pts = upstream_media_rt.map_audio_pts(pkt.pts);
                 let n = CNT.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
                 if n < 5 || n.is_multiple_of(3_000) {
                     tracing::info!(rpc_id, "🎤⬇ srv  pkt#{n}  {} bytes", pkt.data.len());
@@ -134,8 +166,9 @@ impl Relay for Handler {
                 sink.push(&pkt);
             }
             sink.finish(); // flush on EOS
+            upstream_media_rt.close_microphone(lease.generation);
             let _ = tx.send(Ok(Empty {})).await;
-            info!(rpc_id, "🎤 stream_microphone closed");
+            info!(rpc_id, session_id = lease.session_id, "🎤 stream_microphone closed");
             Ok::<(), Status>(())
         });
 
@@ -160,22 +193,36 @@ impl Relay for Handler {
             "🎥 stream_camera output selected"
         );
 
+        let upstream_lease = self.upstream_media_rt.activate_camera();
         let (session_id, relay) = self.camera_rt.activate(&cfg).await?;
         let camera_rt = self.camera_rt.clone();
+        let upstream_media_rt = self.upstream_media_rt.clone();
         info!(rpc_id, session_id, "🎥 stream_camera opened");
+        let frame_step_us = (1_000_000u64 / u64::from(cfg.fps.max(1))).max(1);
 
         // dummy outbound (same pattern as other streams)
         let (tx, rx) = tokio::sync::mpsc::channel(1);
 
         tokio::spawn(async move {
             let mut s = req.into_inner();
-            while let Some(pkt) = s.next().await.transpose()? {
-                if !camera_rt.is_active(session_id) {
+            loop {
+                if !camera_rt.is_active(session_id)
+                    || !upstream_media_rt.is_camera_active(upstream_lease.generation)
+                {
                     info!(rpc_id, session_id, "🎥 stream_camera session superseded");
                     break;
                 }
+                let next_packet = tokio::select! {
+                    packet = s.next() => packet,
+                    _ = tokio::time::sleep(Duration::from_millis(50)) => continue,
+                };
+                let Some(mut pkt) = next_packet.transpose()? else {
+                    break;
+                };
+                pkt.pts = upstream_media_rt.map_video_pts(pkt.pts, frame_step_us);
                 relay.feed(pkt); // ← all logging inside video.rs
             }
+            upstream_media_rt.close_camera(upstream_lease.generation);
             tx.send(Ok(Empty {})).await.ok();
             info!(rpc_id, session_id, "🎥 stream_camera closed");
             Ok::<(), Status>(())

server/src/main/relay_service_coverage.rs

@@ -55,18 +55,45 @@ impl Relay for Handler {
         &self,
         req: Request<tonic::Streaming<AudioPacket>>,
     ) -> Result<Response<Self::StreamMicrophoneStream>, Status> {
+        let lease = self.upstream_media_rt.activate_microphone();
+        let Some(microphone_sink_permit) = self
+            .upstream_media_rt
+            .reserve_microphone_sink(lease.generation)
+            .await
+        else {
+            return Err(Status::aborted(
+                "microphone stream superseded before sink became available",
+            ));
+        };
         let uac_dev = std::env::var("LESAVKA_UAC_DEV").unwrap_or_else(|_| "hw:UAC2Gadget,0".into());
         let mut sink = runtime_support::open_voice_with_retry(&uac_dev)
             .await
-            .map_err(|e| Status::internal(format!("{e:#}")))?;
+            .map_err(|e| {
+                self.upstream_media_rt.close_microphone(lease.generation);
+                Status::internal(format!("{e:#}"))
+            })?;
         let (tx, rx) = tokio::sync::mpsc::channel(1);
+        let upstream_media_rt = self.upstream_media_rt.clone();
 
         tokio::spawn(async move {
+            let _microphone_sink_permit = microphone_sink_permit;
             let mut inbound = req.into_inner();
-            while let Some(pkt) = inbound.next().await.transpose()? {
+            loop {
+                if !upstream_media_rt.is_microphone_active(lease.generation) {
+                    break;
+                }
+                let next_packet = tokio::select! {
+                    packet = inbound.next() => packet,
+                    _ = tokio::time::sleep(Duration::from_millis(25)) => continue,
+                };
+                let Some(mut pkt) = next_packet.transpose()? else {
+                    break;
+                };
+                pkt.pts = upstream_media_rt.map_audio_pts(pkt.pts);
                 sink.push(&pkt);
             }
             sink.finish();
+            upstream_media_rt.close_microphone(lease.generation);
             let _ = tx.send(Ok(Empty {})).await;
             Ok::<(), Status>(())
         });
@@ -79,18 +106,32 @@ impl Relay for Handler {
         req: Request<tonic::Streaming<VideoPacket>>,
     ) -> Result<Response<Self::StreamCameraStream>, Status> {
         let cfg = camera::current_camera_config();
+        let upstream_lease = self.upstream_media_rt.activate_camera();
         let (session_id, relay) = self.camera_rt.activate(&cfg).await?;
         let camera_rt = self.camera_rt.clone();
+        let upstream_media_rt = self.upstream_media_rt.clone();
         let (tx, rx) = tokio::sync::mpsc::channel(1);
+        let frame_step_us = (1_000_000u64 / u64::from(cfg.fps.max(1))).max(1);
 
         tokio::spawn(async move {
             let mut s = req.into_inner();
-            while let Some(pkt) = s.next().await.transpose()? {
-                if !camera_rt.is_active(session_id) {
+            loop {
+                if !camera_rt.is_active(session_id)
+                    || !upstream_media_rt.is_camera_active(upstream_lease.generation)
+                {
                     break;
                 }
+                let next_packet = tokio::select! {
+                    packet = s.next() => packet,
+                    _ = tokio::time::sleep(Duration::from_millis(25)) => continue,
+                };
+                let Some(mut pkt) = next_packet.transpose()? else {
+                    break;
+                };
+                pkt.pts = upstream_media_rt.map_video_pts(pkt.pts, frame_step_us);
                 relay.feed(pkt);
             }
+            upstream_media_rt.close_camera(upstream_lease.generation);
             tx.send(Ok(Empty {})).await.ok();
             Ok::<(), Status>(())
         });

server/src/tests/camera.rs

@@ -1,6 +1,7 @@
+use super::selection::{parse_hdmi_mode, parse_hdmi_modes, preferred_hdmi_mode};
 use super::{
     CameraCodec, CameraConfig, CameraOutput, HdmiConnector, HdmiMode, current_camera_config,
-    parse_hdmi_mode, parse_hdmi_modes, preferred_hdmi_mode, update_camera_config,
+    update_camera_config,
 };
 use serial_test::serial;
 use temp_env::with_var;
@@ -35,21 +36,54 @@ fn camera_config_env_override_prefers_uvc_values() {
 #[serial]
 fn hdmi_camera_profile_honors_installed_1080p_override() {
     with_var("LESAVKA_CAM_OUTPUT", Some("hdmi"), || {
-        with_var("LESAVKA_CAM_WIDTH", Some("1920"), || {
-            with_var("LESAVKA_CAM_HEIGHT", Some("1080"), || {
-                with_var("LESAVKA_CAM_FPS", Some("30"), || {
-                    let cfg = update_camera_config();
-                    assert_eq!(cfg.output, CameraOutput::Hdmi);
-                    assert_eq!(cfg.codec, CameraCodec::H264);
-                    assert_eq!(cfg.width, 1920);
-                    assert_eq!(cfg.height, 1080);
-                    assert_eq!(cfg.fps, 30);
+        with_var("LESAVKA_CAM_CODEC", Some("h264"), || {
+            with_var("LESAVKA_CAM_WIDTH", Some("1920"), || {
+                with_var("LESAVKA_CAM_HEIGHT", Some("1080"), || {
+                    with_var("LESAVKA_CAM_FPS", Some("30"), || {
+                        let cfg = update_camera_config();
+                        assert_eq!(cfg.output, CameraOutput::Hdmi);
+                        assert_eq!(cfg.codec, CameraCodec::H264);
+                        assert_eq!(cfg.width, 1920);
+                        assert_eq!(cfg.height, 1080);
+                        assert_eq!(cfg.fps, 30);
+                    });
                 });
             });
         });
     });
 }
 
+#[test]
+#[serial]
+fn hdmi_camera_profile_prefers_mjpeg_without_hardware_decoder() {
+    with_var("LESAVKA_CAM_OUTPUT", Some("hdmi"), || {
+        with_var("LESAVKA_HW_H264", None::<&str>, || {
+            with_var("LESAVKA_CAM_CODEC", None::<&str>, || {
+                let cfg = update_camera_config();
+                assert_eq!(cfg.output, CameraOutput::Hdmi);
+                assert_eq!(cfg.codec, CameraCodec::Mjpeg);
+                assert_eq!(cfg.width, 1280);
+                assert_eq!(cfg.height, 720);
+                assert_eq!(cfg.fps, 30);
+            });
+        });
+    });
+}
+
+#[test]
+#[serial]
+fn hdmi_camera_profile_honors_explicit_codec_override() {
+    with_var("LESAVKA_CAM_OUTPUT", Some("hdmi"), || {
+        with_var("LESAVKA_HW_H264", None::<&str>, || {
+            with_var("LESAVKA_CAM_CODEC", Some("h264"), || {
+                let cfg = update_camera_config();
+                assert_eq!(cfg.output, CameraOutput::Hdmi);
+                assert_eq!(cfg.codec, CameraCodec::H264);
+            });
+        });
+    });
+}
+
 #[test]
 fn hdmi_mode_parsing_accepts_sysfs_and_override_shapes() {
     assert_eq!(

server/src/upstream_media_runtime.rs

@@ -0,0 +1,369 @@
+#![forbid(unsafe_code)]
+
+use std::sync::atomic::{AtomicU64, Ordering};
+use std::sync::{Arc, Mutex};
+use tokio::sync::{OwnedSemaphorePermit, Semaphore};
+
+/// Logical upstream media kinds that share one live-call session timeline.
+#[derive(Clone, Copy, Debug, Eq, PartialEq)]
+pub enum UpstreamMediaKind {
+    /// Webcam uplink frames destined for the UVC/HDMI sink path.
+    Camera,
+    /// Microphone uplink packets destined for the UAC sink path.
+    Microphone,
+}
+
+/// Lease returned when one upstream media stream becomes the active owner.
+#[derive(Clone, Copy, Debug, Eq, PartialEq)]
+pub struct UpstreamStreamLease {
+    /// Shared session id for the current upstream live-call window.
+    pub session_id: u64,
+    /// Per-kind generation used to supersede older streams of the same kind.
+    pub generation: u64,
+}
+
+#[derive(Debug, Default)]
+struct UpstreamClockState {
+    session_id: u64,
+    active_camera_generation: Option<u64>,
+    active_microphone_generation: Option<u64>,
+    base_remote_pts_us: Option<u64>,
+    last_video_local_pts_us: Option<u64>,
+    last_audio_local_pts_us: Option<u64>,
+}
+
+/// Coordinate upstream stream ownership and keep audio/video on one timeline.
+///
+/// Inputs: stream-open/close events plus remote packet timestamps.
+/// Outputs: active-stream leases and rebased local PTS values.
+/// Why: live calls need one current webcam owner, one current microphone owner,
+/// and one shared media clock so reconnects do not leave old sinks alive or let
+/// audio/video drift onto separate timing islands.
+#[derive(Debug)]
+pub struct UpstreamMediaRuntime {
+    next_session_id: AtomicU64,
+    next_camera_generation: AtomicU64,
+    next_microphone_generation: AtomicU64,
+    microphone_sink_gate: Arc<Semaphore>,
+    state: Mutex<UpstreamClockState>,
+}
+
+impl UpstreamMediaRuntime {
+    /// Build an empty upstream runtime.
+    #[must_use]
+    pub fn new() -> Self {
+        Self {
+            next_session_id: AtomicU64::new(0),
+            next_camera_generation: AtomicU64::new(0),
+            next_microphone_generation: AtomicU64::new(0),
+            microphone_sink_gate: Arc::new(Semaphore::new(1)),
+            state: Mutex::new(UpstreamClockState::default()),
+        }
+    }
+
+    /// Activate a camera stream as the current owner for the session.
+    #[must_use]
+    pub fn activate_camera(&self) -> UpstreamStreamLease {
+        self.activate(UpstreamMediaKind::Camera)
+    }
+
+    /// Activate a microphone stream as the current owner for the session.
+    #[must_use]
+    pub fn activate_microphone(&self) -> UpstreamStreamLease {
+        self.activate(UpstreamMediaKind::Microphone)
+    }
+
+    /// Reserve the single live microphone sink slot for one generation.
+    ///
+    /// Inputs: the microphone lease generation that wants to own the UAC sink.
+    /// Outputs: an owned semaphore permit while that generation still owns the
+    /// microphone slot, or `None` if a newer stream superseded it before the
+    /// previous sink fully stood down.
+    /// Why: ALSA only allows one live owner of the UAC playback device, so a
+    /// replacement stream must wait for the old owner to release the sink
+    /// before opening a new playback pipeline.
+    pub async fn reserve_microphone_sink(&self, generation: u64) -> Option<OwnedSemaphorePermit> {
+        let permit = self
+            .microphone_sink_gate
+            .clone()
+            .acquire_owned()
+            .await
+            .ok()?;
+        self.is_microphone_active(generation).then_some(permit)
+    }
+
+    fn activate(&self, kind: UpstreamMediaKind) -> UpstreamStreamLease {
+        let generation = match kind {
+            UpstreamMediaKind::Camera => {
+                self.next_camera_generation.fetch_add(1, Ordering::SeqCst) + 1
+            }
+            UpstreamMediaKind::Microphone => {
+                self.next_microphone_generation
+                    .fetch_add(1, Ordering::SeqCst)
+                    + 1
+            }
+        };
+
+        let mut state = self
+            .state
+            .lock()
+            .expect("upstream media state mutex poisoned");
+        if state.active_camera_generation.is_none() && state.active_microphone_generation.is_none()
+        {
+            state.session_id = self.next_session_id.fetch_add(1, Ordering::SeqCst) + 1;
+            state.base_remote_pts_us = None;
+            state.last_video_local_pts_us = None;
+            state.last_audio_local_pts_us = None;
+        }
+        match kind {
+            UpstreamMediaKind::Camera => state.active_camera_generation = Some(generation),
+            UpstreamMediaKind::Microphone => state.active_microphone_generation = Some(generation),
+        }
+        UpstreamStreamLease {
+            session_id: state.session_id,
+            generation,
+        }
+    }
+
+    /// Return whether the supplied camera lease is still the active owner.
+    #[must_use]
+    pub fn is_camera_active(&self, generation: u64) -> bool {
+        self.is_active(UpstreamMediaKind::Camera, generation)
+    }
+
+    /// Return whether the supplied microphone lease is still the active owner.
+    #[must_use]
+    pub fn is_microphone_active(&self, generation: u64) -> bool {
+        self.is_active(UpstreamMediaKind::Microphone, generation)
+    }
+
+    fn is_active(&self, kind: UpstreamMediaKind, generation: u64) -> bool {
+        let state = self
+            .state
+            .lock()
+            .expect("upstream media state mutex poisoned");
+        match kind {
+            UpstreamMediaKind::Camera => state.active_camera_generation == Some(generation),
+            UpstreamMediaKind::Microphone => state.active_microphone_generation == Some(generation),
+        }
+    }
+
+    /// Mark a camera stream as closed if it still owns the camera slot.
+    pub fn close_camera(&self, generation: u64) {
+        self.close(UpstreamMediaKind::Camera, generation);
+    }
+
+    /// Mark a microphone stream as closed if it still owns the microphone slot.
+    pub fn close_microphone(&self, generation: u64) {
+        self.close(UpstreamMediaKind::Microphone, generation);
+    }
+
+    fn close(&self, kind: UpstreamMediaKind, generation: u64) {
+        let mut state = self
+            .state
+            .lock()
+            .expect("upstream media state mutex poisoned");
+        match kind {
+            UpstreamMediaKind::Camera if state.active_camera_generation == Some(generation) => {
+                state.active_camera_generation = None;
+            }
+            UpstreamMediaKind::Microphone
+                if state.active_microphone_generation == Some(generation) =>
+            {
+                state.active_microphone_generation = None;
+            }
+            _ => return,
+        }
+        if state.active_camera_generation.is_none() && state.active_microphone_generation.is_none()
+        {
+            state.base_remote_pts_us = None;
+            state.last_video_local_pts_us = None;
+            state.last_audio_local_pts_us = None;
+        }
+    }
+
+    /// Rebase one upstream video packet timestamp onto the shared session clock.
+    #[must_use]
+    pub fn map_video_pts(&self, remote_pts_us: u64, frame_step_us: u64) -> u64 {
+        self.map_pts(
+            UpstreamMediaKind::Camera,
+            remote_pts_us,
+            frame_step_us.max(1),
+        )
+    }
+
+    /// Rebase one upstream audio packet timestamp onto the shared session clock.
+    #[must_use]
+    pub fn map_audio_pts(&self, remote_pts_us: u64) -> u64 {
+        self.map_pts(UpstreamMediaKind::Microphone, remote_pts_us, 1)
+    }
+
+    fn map_pts(&self, kind: UpstreamMediaKind, remote_pts_us: u64, min_step_us: u64) -> u64 {
+        let mut state = self
+            .state
+            .lock()
+            .expect("upstream media state mutex poisoned");
+        let base_remote = *state.base_remote_pts_us.get_or_insert(remote_pts_us);
+        let mut local_pts_us = remote_pts_us.saturating_sub(base_remote);
+        let last_slot = match kind {
+            UpstreamMediaKind::Camera => &mut state.last_video_local_pts_us,
+            UpstreamMediaKind::Microphone => &mut state.last_audio_local_pts_us,
+        };
+        if let Some(last_pts_us) = *last_slot
+            && local_pts_us <= last_pts_us
+        {
+            local_pts_us = last_pts_us.saturating_add(min_step_us.max(1));
+        }
+        *last_slot = Some(local_pts_us);
+        local_pts_us
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::{UpstreamMediaKind, UpstreamMediaRuntime};
+    use std::sync::Arc;
+    use std::time::Duration;
+
+    #[test]
+    fn first_stream_starts_a_new_shared_session() {
+        let runtime = UpstreamMediaRuntime::new();
+        let camera = runtime.activate_camera();
+        let microphone = runtime.activate_microphone();
+
+        assert_eq!(camera.session_id, 1);
+        assert_eq!(microphone.session_id, 1);
+        assert!(runtime.is_camera_active(camera.generation));
+        assert!(runtime.is_microphone_active(microphone.generation));
+    }
+
+    #[test]
+    fn replacing_one_kind_keeps_the_session_but_preempts_the_old_owner() {
+        let runtime = UpstreamMediaRuntime::new();
+        let first = runtime.activate_microphone();
+        let second = runtime.activate_microphone();
+
+        assert_eq!(first.session_id, second.session_id);
+        assert!(!runtime.is_microphone_active(first.generation));
+        assert!(runtime.is_microphone_active(second.generation));
+    }
+
+    #[test]
+    fn closing_the_last_stream_resets_the_next_session_anchor() {
+        let runtime = UpstreamMediaRuntime::new();
+        let camera = runtime.activate_camera();
+        let microphone = runtime.activate_microphone();
+        runtime.close_camera(camera.generation);
+        runtime.close_microphone(microphone.generation);
+
+        let next = runtime.activate_camera();
+        assert_eq!(next.session_id, 2);
+    }
+
+    #[test]
+    fn shared_clock_rebases_audio_and_video_against_the_same_origin() {
+        let runtime = UpstreamMediaRuntime::new();
+        let _camera = runtime.activate_camera();
+        let _microphone = runtime.activate_microphone();
+
+        let video_first = runtime.map_video_pts(1_000_000, 16_666);
+        let audio_first = runtime.map_audio_pts(1_000_000);
+        let audio_next = runtime.map_audio_pts(1_010_000);
+        let video_next = runtime.map_video_pts(1_033_333, 16_666);
+
+        assert_eq!(video_first, 0);
+        assert_eq!(audio_first, 0);
+        assert_eq!(audio_next, 10_000);
+        assert_eq!(video_next, 33_333);
+    }
+
+    #[test]
+    fn shared_clock_keeps_each_kind_monotonic_when_remote_pts_repeat() {
+        let runtime = UpstreamMediaRuntime::new();
+        let _camera = runtime.activate_camera();
+
+        let first = runtime.map_video_pts(50_000, 16_666);
+        let repeated = runtime.map_video_pts(50_000, 16_666);
+
+        assert_eq!(first, 0);
+        assert_eq!(repeated, 16_666);
+    }
+
+    #[test]
+    fn close_ignores_superseded_generation_values() {
+        let runtime = UpstreamMediaRuntime::new();
+        let first = runtime.activate_camera();
+        let second = runtime.activate_camera();
+        runtime.close_camera(first.generation);
+
+        assert!(runtime.is_camera_active(second.generation));
+        runtime.close(UpstreamMediaKind::Camera, second.generation);
+        let next = runtime.activate_camera();
+        assert_eq!(next.session_id, 2);
+    }
+
+    #[tokio::test(flavor = "current_thread")]
+    async fn new_microphone_owner_waits_for_the_previous_sink_to_release() {
+        let runtime = Arc::new(UpstreamMediaRuntime::new());
+        let first = runtime.activate_microphone();
+        let first_permit = runtime
+            .reserve_microphone_sink(first.generation)
+            .await
+            .expect("first owner should acquire the sink gate");
+        let second = runtime.activate_microphone();
+
+        let waiter = tokio::spawn({
+            let runtime = runtime.clone();
+            async move {
+                runtime
+                    .reserve_microphone_sink(second.generation)
+                    .await
+                    .is_some()
+            }
+        });
+
+        tokio::time::sleep(Duration::from_millis(25)).await;
+        assert!(!waiter.is_finished());
+
+        drop(first_permit);
+        assert!(waiter.await.expect("waiter task should finish"));
+    }
+
+    #[tokio::test(flavor = "current_thread")]
+    async fn superseded_microphone_waiter_stands_down_before_opening_a_sink() {
+        let runtime = Arc::new(UpstreamMediaRuntime::new());
+        let first = runtime.activate_microphone();
+        let first_permit = runtime
+            .reserve_microphone_sink(first.generation)
+            .await
+            .expect("first owner should acquire the sink gate");
+        let second = runtime.activate_microphone();
+
+        let superseded_waiter = tokio::spawn({
+            let runtime = runtime.clone();
+            async move {
+                runtime
+                    .reserve_microphone_sink(second.generation)
+                    .await
+                    .is_some()
+            }
+        });
+
+        tokio::time::sleep(Duration::from_millis(25)).await;
+        let third = runtime.activate_microphone();
+        drop(first_permit);
+
+        assert!(
+            !superseded_waiter
+                .await
+                .expect("superseded waiter task should finish"),
+            "older waiter should stand down instead of opening a sink after supersession"
+        );
+
+        let third_permit = runtime
+            .reserve_microphone_sink(third.generation)
+            .await
+            .expect("latest owner should acquire the sink gate");
+        drop(third_permit);
+    }
+}

server/src/video_sinks/hdmi_sink.rs

@@ -12,6 +12,11 @@ pub struct HdmiSink {
     frame_step_us: u64,
 }
 
+#[cfg(any(not(coverage), test))]
+fn hdmi_queue_buffers() -> u32 {
+    crate::video_support::env_u32("LESAVKA_HDMI_QUEUE_BUFFERS", 1).max(1)
+}
+
 impl HdmiSink {
     /// Build a new HDMI sink pipeline.
     ///
@@ -76,13 +81,21 @@ impl HdmiSink {
             .property("caps", &raw_caps)
             .build()?;
 
+        let queue_depth = hdmi_queue_buffers();
         let queue = gst::ElementFactory::make("queue")
-            .property("max-size-buffers", 4u32)
+            .property("max-size-buffers", queue_depth)
+            .property("max-size-bytes", 0u32)
+            .property("max-size-time", 0u64)
             .build()?;
         let convert = gst::ElementFactory::make("videoconvert").build()?;
         let rate = gst::ElementFactory::make("videorate").build()?;
         let scale = gst::ElementFactory::make("videoscale").build()?;
         let sink = build_hdmi_sink(cfg)?;
+        tracing::info!(
+            target: "lesavka_server::video",
+            queue_depth,
+            "📺 HDMI sink queue depth armed"
+        );
 
         if (display_width, display_height) != (cfg.width, cfg.height) {
             tracing::info!(
@@ -190,7 +203,8 @@ impl HdmiSink {
     pub fn push(&self, pkt: VideoPacket) {
         let mut buf = gst::Buffer::from_slice(pkt.data);
         if let Some(meta) = buf.get_mut() {
-            let pts_us = next_local_pts(&self.next_pts_us, self.frame_step_us);
+            let pts_us =
+                crate::video_support::reserve_local_pts(&self.next_pts_us, pkt.pts, self.frame_step_us);
             let ts = gst::ClockTime::from_useconds(pts_us);
             meta.set_pts(Some(ts));
             meta.set_dts(Some(ts));
@@ -352,3 +366,28 @@ fn read_bool_env(name: &str) -> Option<bool> {
         _ => None,
     }
 }
+
+#[cfg(test)]
+mod hdmi_queue_tests {
+    use super::hdmi_queue_buffers;
+
+    #[test]
+    fn hdmi_queue_depth_defaults_to_one_frame() {
+        temp_env::with_var_unset("LESAVKA_HDMI_QUEUE_BUFFERS", || {
+            assert_eq!(hdmi_queue_buffers(), 1);
+        });
+    }
+
+    #[test]
+    fn hdmi_queue_depth_accepts_positive_env_override_only() {
+        temp_env::with_var("LESAVKA_HDMI_QUEUE_BUFFERS", Some("3"), || {
+            assert_eq!(hdmi_queue_buffers(), 3);
+        });
+        temp_env::with_var("LESAVKA_HDMI_QUEUE_BUFFERS", Some("0"), || {
+            assert_eq!(hdmi_queue_buffers(), 1);
+        });
+        temp_env::with_var("LESAVKA_HDMI_QUEUE_BUFFERS", Some("nope"), || {
+            assert_eq!(hdmi_queue_buffers(), 1);
+        });
+    }
+}

server/src/video_sinks/webcam_sink.rs

@@ -9,7 +9,7 @@ use std::sync::atomic::AtomicU64;
 use tracing::warn;
 
 use crate::camera::{CameraCodec, CameraConfig};
-use crate::video_support::{contains_idr, dev_mode_enabled, next_local_pts, pick_h264_decoder};
+use crate::video_support::{contains_idr, dev_mode_enabled, pick_h264_decoder, reserve_local_pts};
 
 /// Push H.264 or MJPEG frames into the USB UVC gadget.
 ///
@@ -180,7 +180,7 @@ impl WebcamSink {
     pub fn push(&self, pkt: VideoPacket) {
         let mut buf = gst::Buffer::from_slice(pkt.data);
         if let Some(meta) = buf.get_mut() {
-            let pts_us = next_local_pts(&self.next_pts_us, self.frame_step_us);
+            let pts_us = reserve_local_pts(&self.next_pts_us, pkt.pts, self.frame_step_us);
             let ts = gst::ClockTime::from_useconds(pts_us);
             meta.set_pts(Some(ts));
             meta.set_dts(Some(ts));

server/src/video_support.rs

@@ -166,11 +166,30 @@ pub fn next_local_pts(counter: &AtomicU64, frame_step_us: u64) -> u64 {
     counter.fetch_add(frame_step_us, Ordering::Relaxed)
 }
 
+/// Reserve a monotonic local timestamp while preferring a caller-provided value.
+///
+/// Inputs: the shared counter, a preferred local timestamp, and the minimum
+/// step to enforce between consecutive values.
+/// Outputs: a timestamp that never goes backwards for the current sink.
+/// Why: upstream media can now arrive with a shared session timeline, so sink
+/// playback should honor that timing when possible while still guarding against
+/// repeated values that would destabilize live playback.
+#[must_use]
+pub fn reserve_local_pts(counter: &AtomicU64, preferred_pts_us: u64, frame_step_us: u64) -> u64 {
+    let next_allowed_pts_us = counter.load(Ordering::Relaxed);
+    let chosen_pts_us = preferred_pts_us.max(next_allowed_pts_us);
+    counter.store(
+        chosen_pts_us.saturating_add(frame_step_us.max(1)),
+        Ordering::Relaxed,
+    );
+    chosen_pts_us
+}
+
 #[cfg(test)]
 mod tests {
     use super::{
         adjust_effective_fps, contains_idr, default_eye_fps, env_u32, env_usize, next_local_pts,
-        should_send_frame,
+        reserve_local_pts, should_send_frame,
     };
     use serial_test::serial;
     use std::sync::atomic::AtomicU64;
@@ -212,6 +231,14 @@ mod tests {
         assert_eq!(next_local_pts(&counter, 40_000), 40_000);
     }
 
+    #[test]
+    fn reserve_local_pts_prefers_preferred_value_but_stays_monotonic() {
+        let counter = AtomicU64::new(0);
+        assert_eq!(reserve_local_pts(&counter, 0, 40_000), 0);
+        assert_eq!(reserve_local_pts(&counter, 10_000, 40_000), 40_000);
+        assert_eq!(reserve_local_pts(&counter, 120_000, 40_000), 120_000);
+    }
+
     #[test]
     #[serial]
     fn env_helpers_parse_values_and_fallbacks() {

testing/tests/client_microphone_include_contract.rs

@@ -382,7 +382,7 @@ JSON
     }
 
     #[test]
-    fn pull_returns_packet_when_appsink_has_buffered_sample() {
+    fn pull_returns_packet_when_appsink_has_buffered_sample_with_shared_capture_clock_pts() {
         gst::init().ok();
         let pipeline = gst::Pipeline::new();
         let src = gst::ElementFactory::make("appsrc")
@@ -406,21 +406,33 @@ JSON
         src.link(&sink).expect("link appsrc->appsink");
         pipeline.set_state(gst::State::Playing).ok();
 
-        let mut buf = gst::Buffer::from_slice(vec![1_u8, 2, 3, 4]);
-        buf.get_mut()
+        let mut first = gst::Buffer::from_slice(vec![1_u8, 2, 3, 4]);
+        first
+            .get_mut()
             .expect("buffer mut")
             .set_pts(Some(gst::ClockTime::from_useconds(321)));
-        src.push_buffer(buf).expect("push sample");
+        src.push_buffer(first).expect("push first sample");
+
+        let mut second = gst::Buffer::from_slice(vec![5_u8, 6, 7, 8]);
+        second
+            .get_mut()
+            .expect("buffer mut")
+            .set_pts(Some(gst::ClockTime::from_useconds(999_999)));
+        src.push_buffer(second).expect("push second sample");
 
         let cap = MicrophoneCapture {
             pipeline,
             sink,
             level_tap_running: None,
         };
-        let pkt = cap.pull().expect("audio packet");
-        assert_eq!(pkt.id, 0);
-        assert_eq!(pkt.pts, 321);
-        assert_eq!(pkt.data, vec![1, 2, 3, 4]);
+        let first_pkt = cap.pull().expect("first audio packet");
+        let second_pkt = cap.pull().expect("second audio packet");
+        assert_eq!(first_pkt.id, 0);
+        assert_eq!(first_pkt.data, vec![1, 2, 3, 4]);
+        assert_eq!(second_pkt.data, vec![5, 6, 7, 8]);
+        assert!(second_pkt.pts >= first_pkt.pts);
+        assert_ne!(first_pkt.pts, 321);
+        assert_ne!(second_pkt.pts, 999_999);
     }
 
     #[test]

testing/tests/handshake_camera_contract.rs

@@ -208,18 +208,46 @@ fn handshake_uses_uvc_interval_when_fps_is_unset() {
 
 #[test]
 #[serial]
-fn handshake_returns_hdmi_caps_with_h264_codec() {
+fn handshake_returns_hdmi_caps_with_mjpeg_when_h264_decode_is_unavailable() {
     with_var("LESAVKA_CAM_OUTPUT", Some("hdmi"), || {
         with_var("LESAVKA_DISABLE_UAC", Some("1"), || {
-            let rt = Runtime::new().expect("create runtime");
-            let caps = rt.block_on(negotiate_against_local_server());
-            assert_eq!(caps.camera_output, Some(String::from("hdmi")));
-            assert_eq!(caps.camera_codec, Some(String::from("h264")));
-            assert_eq!(caps.camera_fps, Some(30));
-            assert!(!caps.microphone);
-            assert!(caps.camera);
-            assert!(matches!(caps.camera_width, Some(1280) | Some(1920)));
-            assert!(matches!(caps.camera_height, Some(720) | Some(1080)));
+            with_var("LESAVKA_HW_H264", None::<&str>, || {
+                with_var("LESAVKA_CAM_CODEC", None::<&str>, || {
+                    let _ = lesavka_server::camera::update_camera_config();
+                    let rt = Runtime::new().expect("create runtime");
+                    let caps = rt.block_on(negotiate_against_local_server());
+                    assert_eq!(caps.camera_output, Some(String::from("hdmi")));
+                    assert_eq!(caps.camera_codec, Some(String::from("mjpeg")));
+                    assert_eq!(caps.camera_fps, Some(30));
+                    assert!(!caps.microphone);
+                    assert!(caps.camera);
+                    assert_eq!(caps.camera_width, Some(1280));
+                    assert_eq!(caps.camera_height, Some(720));
+                });
+            });
+        });
+    });
+}
+
+#[test]
+#[serial]
+fn handshake_honors_explicit_hdmi_h264_override() {
+    with_var("LESAVKA_CAM_OUTPUT", Some("hdmi"), || {
+        with_var("LESAVKA_DISABLE_UAC", Some("1"), || {
+            with_var("LESAVKA_HW_H264", None::<&str>, || {
+                with_var("LESAVKA_CAM_CODEC", Some("h264"), || {
+                    let _ = lesavka_server::camera::update_camera_config();
+                    let rt = Runtime::new().expect("create runtime");
+                    let caps = rt.block_on(negotiate_against_local_server());
+                    assert_eq!(caps.camera_output, Some(String::from("hdmi")));
+                    assert_eq!(caps.camera_codec, Some(String::from("h264")));
+                    assert_eq!(caps.camera_fps, Some(30));
+                    assert!(!caps.microphone);
+                    assert!(caps.camera);
+                    assert_eq!(caps.camera_width, Some(1280));
+                    assert_eq!(caps.camera_height, Some(720));
+                });
+            });
         });
     });
 }
@@ -237,7 +265,10 @@ fn handshake_auto_mode_falls_back_to_a_valid_camera_configuration() {
                 assert!(matches!(caps.camera_height, Some(720) | Some(360)));
             }
             Some("hdmi") => {
-                assert_eq!(caps.camera_codec.as_deref(), Some("h264"));
+                assert!(matches!(
+                    caps.camera_codec.as_deref(),
+                    Some("mjpeg") | Some("h264")
+                ));
                 assert!(matches!(caps.camera_width, Some(1280) | Some(1920)));
                 assert!(matches!(caps.camera_height, Some(720) | Some(1080)));
             }

testing/tests/server_audio_include_contract.rs

@@ -6,6 +6,8 @@
 //! Why: audio pipeline setup is branchy and should stay stable without requiring
 //! physical ALSA/UAC hardware in CI.
 
+pub use lesavka_server::camera;
+
 #[path = "../../server/src/audio.rs"]
 #[allow(warnings)]
 mod server_audio_contract;

testing/tests/server_camera_contract.rs

@@ -75,21 +75,25 @@ fn camera_config_zero_interval_falls_back_to_default_fps() {
 
 #[test]
 #[serial]
-fn camera_config_forced_hdmi_tracks_cached_state() {
+fn camera_config_forced_hdmi_defaults_to_mjpeg_without_hardware_decode() {
     with_var("LESAVKA_CAM_OUTPUT", Some("hdmi"), || {
-        let cfg = update_camera_config();
-        assert_eq!(cfg.output, CameraOutput::Hdmi);
-        assert_eq!(cfg.codec, CameraCodec::H264);
-        assert_eq!(cfg.fps, 30);
-        assert!(matches!(
-            (cfg.width, cfg.height),
-            (1920, 1080) | (1280, 720)
-        ));
+        with_var("LESAVKA_HW_H264", None::<&str>, || {
+            with_var("LESAVKA_CAM_CODEC", None::<&str>, || {
+                let cfg = update_camera_config();
+                assert_eq!(cfg.output, CameraOutput::Hdmi);
+                assert_eq!(cfg.codec, CameraCodec::Mjpeg);
+                assert_eq!(cfg.width, 1280);
+                assert_eq!(cfg.height, 720);
+                assert_eq!(cfg.fps, 30);
 
-        let cached = current_camera_config();
-        assert_eq!(cached.output, CameraOutput::Hdmi);
-        assert_eq!(cached.codec, CameraCodec::H264);
-        assert_eq!(cached.fps, 30);
+                let cached = current_camera_config();
+                assert_eq!(cached.output, CameraOutput::Hdmi);
+                assert_eq!(cached.codec, CameraCodec::Mjpeg);
+                assert_eq!(cached.width, 1280);
+                assert_eq!(cached.height, 720);
+                assert_eq!(cached.fps, 30);
+            });
+        });
     });
 }
 
@@ -97,21 +101,42 @@ fn camera_config_forced_hdmi_tracks_cached_state() {
 #[serial]
 fn camera_config_forced_hdmi_honors_1080p_uplink_override() {
     with_var("LESAVKA_CAM_OUTPUT", Some("hdmi"), || {
-        with_var("LESAVKA_CAM_WIDTH", Some("1920"), || {
-            with_var("LESAVKA_CAM_HEIGHT", Some("1080"), || {
-                with_var("LESAVKA_CAM_FPS", Some("30"), || {
-                    let cfg = update_camera_config();
-                    assert_eq!(cfg.output, CameraOutput::Hdmi);
-                    assert_eq!(cfg.codec, CameraCodec::H264);
-                    assert_eq!(cfg.width, 1920);
-                    assert_eq!(cfg.height, 1080);
-                    assert_eq!(cfg.fps, 30);
+        with_var("LESAVKA_HW_H264", None::<&str>, || {
+            with_var("LESAVKA_CAM_WIDTH", Some("1920"), || {
+                with_var("LESAVKA_CAM_HEIGHT", Some("1080"), || {
+                    with_var("LESAVKA_CAM_FPS", Some("30"), || {
+                        with_var("LESAVKA_CAM_CODEC", None::<&str>, || {
+                            let cfg = update_camera_config();
+                            assert_eq!(cfg.output, CameraOutput::Hdmi);
+                            assert_eq!(cfg.codec, CameraCodec::Mjpeg);
+                            assert_eq!(cfg.width, 1920);
+                            assert_eq!(cfg.height, 1080);
+                            assert_eq!(cfg.fps, 30);
+                        });
+                    });
                 });
             });
         });
     });
 }
 
+#[test]
+#[serial]
+fn camera_config_forced_hdmi_honors_explicit_h264_override() {
+    with_var("LESAVKA_CAM_OUTPUT", Some("hdmi"), || {
+        with_var("LESAVKA_HW_H264", None::<&str>, || {
+            with_var("LESAVKA_CAM_CODEC", Some("h264"), || {
+                let cfg = update_camera_config();
+                assert_eq!(cfg.output, CameraOutput::Hdmi);
+                assert_eq!(cfg.codec, CameraCodec::H264);
+                assert_eq!(cfg.width, 1280);
+                assert_eq!(cfg.height, 720);
+                assert_eq!(cfg.fps, 30);
+            });
+        });
+    });
+}
+
 #[test]
 #[serial]
 fn camera_config_output_override_is_case_insensitive() {
@@ -184,7 +209,7 @@ fn camera_config_invalid_output_falls_back_to_detected_policy() {
                 assert!(cfg.fps > 0);
             }
             CameraOutput::Hdmi => {
-                assert_eq!(cfg.codec, CameraCodec::H264);
+                assert!(matches!(cfg.codec, CameraCodec::Mjpeg | CameraCodec::H264));
                 assert_eq!(cfg.fps, 30);
             }
         }

testing/tests/server_main_binary_contract.rs

@@ -49,6 +49,7 @@ mod server_main_binary {
                 gadget: UsbGadget::new("lesavka"),
                 did_cycle: std::sync::Arc::new(std::sync::atomic::AtomicBool::new(false)),
                 camera_rt: std::sync::Arc::new(CameraRuntime::new()),
+                upstream_media_rt: std::sync::Arc::new(UpstreamMediaRuntime::new()),
                 capture_power: CapturePowerManager::new(),
                 eye_hubs: std::sync::Arc::new(tokio::sync::Mutex::new(
                     std::collections::HashMap::new(),

testing/tests/server_main_binary_extra_contract.rs

@@ -118,6 +118,7 @@ mod server_main_binary_extra {
                 gadget: UsbGadget::new("lesavka"),
                 did_cycle: std::sync::Arc::new(std::sync::atomic::AtomicBool::new(false)),
                 camera_rt: std::sync::Arc::new(CameraRuntime::new()),
+                upstream_media_rt: std::sync::Arc::new(UpstreamMediaRuntime::new()),
                 capture_power: CapturePowerManager::new(),
                 eye_hubs: std::sync::Arc::new(tokio::sync::Mutex::new(
                     std::collections::HashMap::new(),

testing/tests/server_main_media_extra_contract.rs

@@ -88,6 +88,7 @@ mod server_main_media_extra {
                 gadget: UsbGadget::new("lesavka"),
                 did_cycle: std::sync::Arc::new(std::sync::atomic::AtomicBool::new(false)),
                 camera_rt: std::sync::Arc::new(CameraRuntime::new()),
+                upstream_media_rt: std::sync::Arc::new(UpstreamMediaRuntime::new()),
                 capture_power: CapturePowerManager::new(),
                 eye_hubs: std::sync::Arc::new(tokio::sync::Mutex::new(
                     std::collections::HashMap::new(),

testing/tests/server_main_rpc_contract.rs

@@ -48,6 +48,7 @@ mod server_main_rpc {
             gadget: UsbGadget::new("lesavka"),
             did_cycle: std::sync::Arc::new(std::sync::atomic::AtomicBool::new(false)),
             camera_rt: std::sync::Arc::new(CameraRuntime::new()),
+            upstream_media_rt: std::sync::Arc::new(UpstreamMediaRuntime::new()),
             capture_power: CapturePowerManager::new(),
             eye_hubs: std::sync::Arc::new(
                 tokio::sync::Mutex::new(std::collections::HashMap::new()),

testing/tests/server_main_rpc_reset_contract.rs

@@ -45,6 +45,7 @@ mod server_main_rpc_reset {
             gadget: UsbGadget::new("lesavka"),
             did_cycle: std::sync::Arc::new(std::sync::atomic::AtomicBool::new(false)),
             camera_rt: std::sync::Arc::new(CameraRuntime::new()),
+            upstream_media_rt: std::sync::Arc::new(UpstreamMediaRuntime::new()),
             capture_power: CapturePowerManager::new(),
             eye_hubs: std::sync::Arc::new(
                 tokio::sync::Mutex::new(std::collections::HashMap::new()),
@@ -102,6 +103,7 @@ mod server_main_rpc_reset {
                                 false,
                             )),
                             camera_rt: std::sync::Arc::new(CameraRuntime::new()),
+                            upstream_media_rt: std::sync::Arc::new(UpstreamMediaRuntime::new()),
                             capture_power: CapturePowerManager::new(),
                             eye_hubs: std::sync::Arc::new(tokio::sync::Mutex::new(
                                 std::collections::HashMap::new(),

testing/tests/server_main_usb_recovery_contract.rs

@@ -118,6 +118,7 @@ mod server_main_binary_extra {
                 gadget: UsbGadget::new("lesavka"),
                 did_cycle: std::sync::Arc::new(std::sync::atomic::AtomicBool::new(false)),
                 camera_rt: std::sync::Arc::new(CameraRuntime::new()),
+                upstream_media_rt: std::sync::Arc::new(UpstreamMediaRuntime::new()),
                 capture_power: CapturePowerManager::new(),
                 eye_hubs: std::sync::Arc::new(tokio::sync::Mutex::new(
                     std::collections::HashMap::new(),
@@ -151,6 +152,7 @@ mod server_main_binary_extra {
             gadget: UsbGadget::new("lesavka"),
             did_cycle: std::sync::Arc::new(std::sync::atomic::AtomicBool::new(false)),
             camera_rt: std::sync::Arc::new(CameraRuntime::new()),
+            upstream_media_rt: std::sync::Arc::new(UpstreamMediaRuntime::new()),
             capture_power: CapturePowerManager::new(),
             eye_hubs: std::sync::Arc::new(
                 tokio::sync::Mutex::new(std::collections::HashMap::new()),
@@ -214,6 +216,7 @@ echo noop core helper >&2
                                 false,
                             )),
                             camera_rt: std::sync::Arc::new(CameraRuntime::new()),
+                            upstream_media_rt: std::sync::Arc::new(UpstreamMediaRuntime::new()),
                             capture_power: CapturePowerManager::new(),
                             eye_hubs: std::sync::Arc::new(tokio::sync::Mutex::new(
                                 std::collections::HashMap::new(),
@@ -284,6 +287,7 @@ printf 'configured\n' > "$LESAVKA_GADGET_SYSFS_ROOT/class/udc/fake-ctrl.usb/stat
                                 false,
                             )),
                             camera_rt: std::sync::Arc::new(CameraRuntime::new()),
+                            upstream_media_rt: std::sync::Arc::new(UpstreamMediaRuntime::new()),
                             capture_power: CapturePowerManager::new(),
                             eye_hubs: std::sync::Arc::new(tokio::sync::Mutex::new(
                                 std::collections::HashMap::new(),

testing/tests/server_upstream_media_contract.rs

@@ -60,6 +60,7 @@ mod server_upstream_media {
                 gadget: UsbGadget::new("lesavka"),
                 did_cycle: std::sync::Arc::new(std::sync::atomic::AtomicBool::new(false)),
                 camera_rt: std::sync::Arc::new(CameraRuntime::new()),
+                upstream_media_rt: std::sync::Arc::new(UpstreamMediaRuntime::new()),
                 capture_power: CapturePowerManager::new(),
                 eye_hubs: std::sync::Arc::new(tokio::sync::Mutex::new(
                     std::collections::HashMap::new(),
@@ -127,6 +128,70 @@ mod server_upstream_media {
         });
     }
 
+    #[test]
+    #[serial]
+    fn stream_microphone_supersedes_the_previous_owner_cleanly() {
+        let rt = tokio::runtime::Runtime::new().expect("runtime");
+        with_var("LESAVKA_CAPTURE_POWER_UNIT", Some("none"), || {
+            rt.block_on(async {
+                let (_dir, handler) = build_handler_for_tests();
+                let (server, mut cli) = serve_handler(handler).await;
+                let (first_tx, first_rx) = tokio::sync::mpsc::channel(1);
+                let (_second_tx, second_rx) = tokio::sync::mpsc::channel(1);
+
+                let mut first = cli
+                    .stream_microphone(tonic::Request::new(
+                        tokio_stream::wrappers::ReceiverStream::new(first_rx),
+                    ))
+                    .await
+                    .expect("first microphone stream")
+                    .into_inner();
+
+                let _second = cli
+                    .stream_microphone(tonic::Request::new(
+                        tokio_stream::wrappers::ReceiverStream::new(second_rx),
+                    ))
+                    .await
+                    .expect("second microphone stream supersedes first");
+
+                drop(first_tx);
+                let ack = tokio::time::timeout(std::time::Duration::from_secs(1), first.message())
+                    .await
+                    .expect("superseded microphone ack timeout")
+                    .expect("superseded microphone ack grpc")
+                    .expect("superseded microphone ack item");
+                assert_eq!(ack, Empty {});
+
+                server.abort();
+            });
+        });
+    }
+
+    #[test]
+    #[serial]
+    fn stream_microphone_surfaces_internal_error_when_sink_open_fails() {
+        let rt = tokio::runtime::Runtime::new().expect("runtime");
+        with_var("LESAVKA_CAPTURE_POWER_UNIT", Some("none"), || {
+            with_var("LESAVKA_TEST_FORCE_PIPELINE_START_ERROR", Some("1"), || {
+                rt.block_on(async {
+                    let (_dir, handler) = build_handler_for_tests();
+                    let (server, mut cli) = serve_handler(handler).await;
+                    let (_tx, rx) = tokio::sync::mpsc::channel(1);
+
+                    let err = cli
+                        .stream_microphone(tonic::Request::new(
+                            tokio_stream::wrappers::ReceiverStream::new(rx),
+                        ))
+                        .await
+                        .expect_err("missing sink should fail the stream");
+                    assert_eq!(err.code(), tonic::Code::Internal);
+
+                    server.abort();
+                });
+            });
+        });
+    }
+
     #[test]
     #[serial]
     fn stream_camera_accepts_upstream_video_packets() {