fix(capture-linux): measure the cursor against the window, not its mo…

electron/native-bridge/cursor/recording/pipeWireCursorAccumulator.ts

@@ -150,11 +150,15 @@ export class PipeWireCursorAccumulator {
 	addSample(payload: Extract<PipeWireHelperEvent, { event: "cursor-sample" }>) {
 		this.rememberAsset(payload.asset);
 
-		// Normalised against the stream's own dimensions, which the helper repeats
-		// on every sample. Electron's display bounds are deliberately NOT used:
-		// they are in DIPs, whereas the portal reports stream pixels, and the
-		// portal's source is whatever the user picked in its own dialog, which
-		// need not be the display the app thinks it is recording.
+		// Normalised against the RECORDED RECTANGLE, which the helper repeats on
+		// every sample: the crop for a window stream, the whole stream for a
+		// screen. It reports its own because only it knows — for a window,
+		// mutter pins the stream to the monitor and carves the window out
+		// through a crop, so the stream's dimensions describe a rectangle the
+		// file does not show. Electron's display bounds are deliberately NOT
+		// used either: they are in DIPs, whereas the portal reports pixels, and
+		// the portal's source is whatever the user picked in its own dialog,
+		// which need not be the display the app thinks it is recording.
 		const width = Math.max(1, payload.width);
 		const height = Math.max(1, payload.height);
 

electron/native/pipewire-capture/src/capture.rs

@@ -220,6 +220,17 @@ pub struct Capture {
     /// through it rather than replacing it.
     committed_width: i32,
     committed_height: i32,
+    /// The rectangle of the SOURCE STREAM the encoder last read, in stream
+    /// pixels: the live crop origin (clamped by [`Self::read_origin`]) at the
+    /// committed size.
+    ///
+    /// Exists for the cursor, not for the video. The portal reports pointer
+    /// positions in stream pixels, and for a window stream the stream is the
+    /// whole monitor — mutter pins it there and carves the window out through
+    /// SPA_META_VideoCrop. Normalising the pointer against the stream would
+    /// then place it in a rectangle the file does not show. This is the one the
+    /// file DOES show.
+    content: shim::CropRect,
 }
 
 impl Capture {
@@ -286,6 +297,7 @@ impl Capture {
                 frames_written: 0,
                 committed_width: width,
                 committed_height: height,
+                content: shim::CropRect { x: 0, y: 0, width, height },
             },
             selection,
         ))
@@ -331,6 +343,12 @@ impl Capture {
         // path subtracts the x offset from it, which is wrong for any non-zero x
         // and is latent there only because no shipping compositor sets one.
         let (x, y) = self.read_origin(frame);
+        self.content = shim::CropRect {
+            x,
+            y,
+            width: self.committed_width,
+            height: self.committed_height,
+        };
         let offset = (y as usize)
             .checked_mul(frame.stride)
             .and_then(|rows| rows.checked_add((x as usize) * BYTES_PER_SOURCE_PIXEL))
@@ -364,6 +382,11 @@ impl Capture {
         self.epoch.is_some()
     }
 
+    /// The source rectangle the file is showing. See [`Self::content`].
+    pub fn content_rect(&self) -> shim::CropRect {
+        self.content
+    }
+
     /// Encodes forward to the current clock position. Returns how many frames
     /// were written.
     pub fn advance(&mut self) -> Result<u32, String> {
@@ -658,6 +681,72 @@ mod tests {
         let _ = std::fs::remove_file(&output);
     }
 
+    /// The rectangle the cursor is measured against has to be the one the FILE
+    /// shows, which for a window is the crop and not the stream. Getting this
+    /// wrong put the pointer in monitor coordinates over window-sized footage —
+    /// a fixed offset and a wrong scale, in every window recording.
+    #[test]
+    fn the_content_rect_is_the_window_the_file_shows() {
+        let output = std::env::temp_dir().join("openscreen-capture-content.mp4");
+        let (mut capture, _) =
+            Capture::start(&output, 320, 240, 30, Some(1_000_000), Some(Backend::Software), Vec::new())
+                .expect("start");
+
+        // Before a frame is staged there is nothing cropped yet, so the whole
+        // committed size sits at the origin.
+        assert_eq!(
+            capture.content_rect(),
+            shim::CropRect { x: 0, y: 0, width: 320, height: 240 }
+        );
+
+        capture
+            .stage(&cropped_frame(
+                1920,
+                1080,
+                shim::CropRect { x: 100, y: 50, width: 320, height: 240 },
+                shim::constants().video_format_bgrx,
+            ))
+            .expect("stage");
+        assert_eq!(
+            capture.content_rect(),
+            shim::CropRect { x: 100, y: 50, width: 320, height: 240 }
+        );
+
+        // The window moved. The file follows its origin at the committed size,
+        // and so must the cursor.
+        capture
+            .stage(&cropped_frame(
+                1920,
+                1080,
+                shim::CropRect { x: 700, y: 400, width: 320, height: 240 },
+                shim::constants().video_format_bgrx,
+            ))
+            .expect("stage");
+        assert_eq!(
+            capture.content_rect(),
+            shim::CropRect { x: 700, y: 400, width: 320, height: 240 }
+        );
+
+        // An origin that would read past the buffer is clamped for the pixels,
+        // so the cursor has to be clamped with it or the two disagree about
+        // which rectangle was recorded.
+        capture
+            .stage(&cropped_frame(
+                1920,
+                1080,
+                shim::CropRect { x: 1800, y: 1000, width: 320, height: 240 },
+                shim::constants().video_format_bgrx,
+            ))
+            .expect("stage");
+        assert_eq!(
+            capture.content_rect(),
+            shim::CropRect { x: 1600, y: 840, width: 320, height: 240 }
+        );
+
+        let _ = capture.finish();
+        let _ = std::fs::remove_file(&output);
+    }
+
     /// A crop flush against the right edge leaves the last row short of a full
     /// stride. The old `stride * height` bounds check rejected exactly those —
     /// i.e. every window not touching the left edge.

electron/native/pipewire-capture/src/main.rs

@@ -808,7 +808,8 @@ fn run<W: Write>(
                     // The portal's size is in the compositor's coordinate space
                     // and can differ from the negotiated pixel size on a scaled
                     // display. Logged rather than used: cursor positions arrive
-                    // in stream pixels, so only the negotiated size normalises them.
+                    // in stream pixels, and `content_rect` places them from
+                    // there.
                     let _ = emitter.emit(&Event::Debug {
                         code: "portal-stream".to_owned(),
                         data: json_map([
@@ -1052,14 +1053,14 @@ fn run<W: Write>(
                 // A new sprite ships immediately; positions respect the sample
                 // interval so a 120fps compositor cannot flood stdout.
                 if asset_is_new || last_emit.elapsed() >= config.sample_interval {
-                    emit_sample(emitter, &cursor, size, &mut pending_asset);
+                    emit_sample(emitter, &cursor, content_rect(&capture, size), &mut pending_asset);
                     last_emit = Instant::now();
                 }
             }
 
             Err(RecvTimeoutError::Timeout) => {
                 if cursor.is_some() && last_emit.elapsed() >= config.sample_interval {
-                    emit_sample(emitter, &cursor, size, &mut pending_asset);
+                    emit_sample(emitter, &cursor, content_rect(&capture, size), &mut pending_asset);
                     last_emit = Instant::now();
                 }
                 // The heartbeat that keeps the output at a constant frame rate
@@ -1148,22 +1149,49 @@ fn finish_capture<W: Write>(
     }
 }
 
+/// The rectangle cursor positions are measured against.
+///
+/// The encoder's once a frame has been staged — that is the only rectangle the
+/// file shows. Before then, and for a cursor-only session that opens no encoder
+/// at all, the whole negotiated stream, which is what the consumer of a
+/// cursor-only recording is compositing over.
+fn content_rect(capture: &Option<Capture>, size: Option<(i32, i32)>) -> Option<shim::CropRect> {
+    match capture {
+        Some(capture) if capture.started() => Some(capture.content_rect()),
+        _ => size.map(|(width, height)| shim::CropRect { x: 0, y: 0, width, height }),
+    }
+}
+
+/// Emits one cursor sample, positioned inside `content`.
+///
+/// `content` is the sub-rectangle of the stream the recording actually shows —
+/// [`Capture::content_rect`] once pixels are being encoded, the whole stream
+/// otherwise (a cursor-only session records no video of its own). The pointer
+/// arrives in STREAM pixels, so for a window stream it is measured from the
+/// corner of the monitor while the file starts at the corner of the window;
+/// subtracting the origin is what puts the two in the same space. The consumer
+/// normalises against the `width`/`height` reported here, so those must be the
+/// content's, not the stream's.
 fn emit_sample<W: Write>(
     emitter: &mut Emitter<W>,
     cursor: &Option<CursorState>,
-    size: Option<(i32, i32)>,
+    content: Option<shim::CropRect>,
     pending_asset: &mut Option<CursorAsset>,
 ) {
-    let (Some(state), Some((width, height))) = (cursor, size) else {
+    let (Some(state), Some(content)) = (cursor, content) else {
         return;
     };
-    let visible = state.x >= 0 && state.y >= 0 && state.x < width && state.y < height;
+    let (x, y) = (state.x - content.x, state.y - content.y);
+    // A pointer outside the recorded rectangle is REPORTED, not withheld: the
+    // consumer clamps the position and carries `visible` so a renderer can hide
+    // the sprite rather than pin it to an edge.
+    let visible = x >= 0 && y >= 0 && x < content.width && y < content.height;
     let _ = emitter.emit(&Event::CursorSample {
         timestamp_ms: timestamp_ms(),
-        x: state.x,
-        y: state.y,
-        width,
-        height,
+        x,
+        y,
+        width: content.width,
+        height: content.height,
         visible,
         asset_id: state.asset_id.clone(),
         asset: pending_asset.take(),
@@ -1234,6 +1262,109 @@ fn resolve_microphone_node(label: &str, sources: &[shim::AudioSourceInfo]) -> Op
     candidates.first().map(|s| s.name.clone())
 }
 
+#[cfg(test)]
+mod cursor_sample_tests {
+    use super::*;
+
+    fn sample_json(cursor: (i32, i32), content: Option<shim::CropRect>) -> serde_json::Value {
+        let mut buffer = Vec::new();
+        let mut emitter = Emitter::new(&mut buffer, false);
+        emit_sample(
+            &mut emitter,
+            &Some(CursorState { x: cursor.0, y: cursor.1, asset_id: None }),
+            content,
+            &mut None,
+        );
+        let line = String::from_utf8(buffer).expect("utf8");
+        serde_json::from_str(line.trim()).expect("json")
+    }
+
+    /// A full-screen capture crops nothing, so the pointer keeps the numbers the
+    /// portal gave and is normalised against the whole stream.
+    #[test]
+    fn a_full_screen_capture_reports_stream_coordinates() {
+        let value = sample_json(
+            (960, 540),
+            Some(shim::CropRect { x: 0, y: 0, width: 1920, height: 1080 }),
+        );
+        assert_eq!(value["x"], 960);
+        assert_eq!(value["y"], 540);
+        assert_eq!(value["width"], 1920);
+        assert_eq!(value["height"], 1080);
+        assert_eq!(value["visible"], true);
+    }
+
+    /// THE WINDOW-CAPTURE BUG. mutter pins a window stream to the whole monitor
+    /// and carves the window out through SPA_META_VideoCrop, so the pointer
+    /// arrives measured from the monitor's corner while the file starts at the
+    /// window's. Reporting the stream's size here normalised a monitor position
+    /// against monitor dimensions and handed the compositor a fraction of the
+    /// wrong rectangle: the cursor sat at the wrong place in every window
+    /// recording, by the crop origin, at the wrong scale.
+    #[test]
+    fn a_window_capture_reports_coordinates_inside_the_window() {
+        // A 1920x1080 monitor carrying a 640x480 window at (100, 50), pointer
+        // one quarter into the window.
+        let value = sample_json(
+            (260, 170),
+            Some(shim::CropRect { x: 100, y: 50, width: 640, height: 480 }),
+        );
+        assert_eq!(value["x"], 160, "the crop origin has to come off the position");
+        assert_eq!(value["y"], 120);
+        assert_eq!(value["width"], 640, "the consumer normalises against what it is told");
+        assert_eq!(value["height"], 480);
+        assert_eq!(value["visible"], true);
+    }
+
+    /// Outside the window but still on the monitor. The old test was `x < width`
+    /// against the STREAM, which called this visible — and since the consumer
+    /// clamps to 0..1, it parked the sprite on the frame's edge for as long as
+    /// the pointer was anywhere else on screen.
+    #[test]
+    fn a_pointer_outside_the_window_is_reported_invisible() {
+        let outside = sample_json(
+            (1500, 900),
+            Some(shim::CropRect { x: 100, y: 50, width: 640, height: 480 }),
+        );
+        assert_eq!(outside["visible"], false);
+
+        // Above and to the left of the window, which goes negative rather than
+        // past the far edge.
+        let before = sample_json(
+            (10, 10),
+            Some(shim::CropRect { x: 100, y: 50, width: 640, height: 480 }),
+        );
+        assert_eq!(before["visible"], false);
+    }
+
+    /// No content rectangle means the format has not been negotiated yet. There
+    /// is nothing to measure against, so nothing is emitted — a sample stamped
+    /// with a guess would be indistinguishable from a real one downstream.
+    #[test]
+    fn nothing_is_emitted_before_the_format_is_known() {
+        let mut buffer = Vec::new();
+        let mut emitter = Emitter::new(&mut buffer, false);
+        emit_sample(
+            &mut emitter,
+            &Some(CursorState { x: 10, y: 10, asset_id: None }),
+            None,
+            &mut None,
+        );
+        assert!(buffer.is_empty(), "emitted {}", String::from_utf8_lossy(&buffer));
+    }
+
+    /// A cursor-only session opens no encoder, so it falls back to the whole
+    /// stream; once pixels are flowing the encoder's rectangle wins.
+    #[test]
+    fn the_content_rect_prefers_the_encoder_once_it_has_started() {
+        assert_eq!(
+            content_rect(&None, Some((1920, 1080))),
+            Some(shim::CropRect { x: 0, y: 0, width: 1920, height: 1080 })
+        );
+        assert_eq!(content_rect(&None, None), None);
+    }
+}
+
 #[cfg(test)]
 mod microphone_resolution_tests {
     use super::*;

technical-documentation/architecture/recording.md

@@ -52,10 +52,11 @@ Stopping is the part of that boundary that has broken repeatedly (issues #34, #1
 
 `org.freedesktop.portal.ScreenCast.SelectSources` takes exactly `(session, cursor_mode, types, multiple, restore_token, persist_mode)`. There is no window id, monitor id, or node id a caller may supply, so the compositor's own picker is the only thing that can choose a source — the app cannot ask for one and cannot override the answer. The helper reports what it was given back on `stream-started` as `sourceKind` (`"monitor"`, `"window"` or `"virtual"`); that reply is the only knowledge the app ever has about what is being recorded, and an absent `sourceKind` means unknown, not "monitor".
 
-Two consequences follow, and both were once bugs:
+Three consequences follow, and all three were once bugs:
 
 - **The HUD shows no source button on Linux.** An in-app picker cannot steer the portal, and the one that existed raised a *second* portal dialog of its own through `desktopCapturer.getSources()` whose grant was then discarded — which is why choosing a window there changed nothing.
 - **No portal restore token is persisted.** Replaying one used to suppress the picker on later runs. Because a token is bound to the source it was minted for, an approved monitor came back on every subsequent recording and the picker — the only source chooser Wayland offers — never reappeared, so "record this window" recorded the whole screen. Answering the picker each time is the cost of being able to choose at all.
+- **A window stream is a monitor stream with a crop, and the cursor has to be measured against the crop.** mutter never renegotiates the format for a window: it pins the stream to the window's monitor and carves the window out through `SPA_META_VideoCrop`, which can move on any buffer as the window does. The encoder already reads through that rect (`Capture::read_origin`), so the file holds the window — but the portal reports the pointer in *stream* pixels, measured from the monitor's corner. Normalising it against the stream's dimensions therefore described a rectangle the file does not show, and the overlay drew the cursor offset by the crop origin and scaled by the ratio of monitor to window, in every window recording. `Capture::content_rect` is the rect the file actually holds, and `emit_sample` is the one place that turns a pointer into a fraction of it.
 
 Electron resolves selected sources, devices, and paths before launching the helper. The helper does not guess a DirectShow camera: Windows receives the resolved selection. A helper error is reported explicitly rather than silently switching a Windows native feature to browser capture.