S2.40: SAM 2 as the Roto Brush object tracker

ObjectTracker interface in the segment crate (prompts on some frames, the
clip tracked both ways, each frame streamed with confidence and an occluded
flag), implemented by SAM 2. MaskModel::tracker() is None for other models.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>

crates/segment/src/lib.rs

@@ -45,6 +45,54 @@ pub trait MaskModel: Send + Sync {
     fn info(&self) -> &'static ModelInfo;
     /// Foreground probability (0–1) of each pixel of a frame (straight RGB 0–1, row-major `w×h`).
     fn segment(&self, rgb: &[[f32; 3]], w: usize, h: usize, prompt: &Prompt) -> Result<Vec<f32>>;
+    /// A video tracker built on the same weights, if the model has one.
+    fn tracker(&self) -> Option<&dyn ObjectTracker> {
+        None
+    }
+}
+
+/// One decoded frame handed to a tracker (straight RGB 0–1, row-major `w×h`).
+#[derive(Clone, Debug, Default)]
+pub struct ClipFrame {
+    pub rgb: Vec<[f32; 3]>,
+    pub w: usize,
+    pub h: usize,
+}
+
+/// What to track: the clip's frames `first..first + count` (the host's frame numbers) and the
+/// prompts of the one object, each on its frame.
+#[derive(Clone, Debug)]
+pub struct TrackRequest {
+    pub first: i64,
+    pub count: usize,
+    pub prompts: Vec<(i64, Prompt)>,
+}
+
+/// A tracker's answer for one frame.
+#[derive(Clone, Debug, Default)]
+pub struct TrackedFrame {
+    /// Foreground probability of every pixel (`w×h`); all zero when the object is absent.
+    pub prob: Vec<f32>,
+    pub w: usize,
+    pub h: usize,
+    /// How sure the tracker is (0–1), for the timeline.
+    pub confidence: f32,
+    /// The object is not in this frame (occluded or out of view).
+    pub occluded: bool,
+}
+
+/// A video tracker with memory (SAM 2): given prompts on some frames, it follows the object
+/// through the clip in both directions. Optional: a [`MaskModel`] may offer one.
+pub trait ObjectTracker: Send + Sync {
+    /// Track the object. `frames(i)` supplies frame `i` (`None`: unavailable); `cancel()` is polled
+    /// between frames; `out(i, result)` is called as soon as each frame is finished.
+    fn track(
+        &self,
+        req: &TrackRequest,
+        frames: &mut dyn FnMut(i64) -> Option<Arc<ClipFrame>>,
+        cancel: &dyn Fn() -> bool,
+        out: &mut dyn FnMut(i64, TrackedFrame),
+    ) -> Result<()>;
 }
 
 /// What a model is for.