Cross a clip join per track, not on the video's frame

The first cut placed every track with the cursor the video had moved. At
an overlap join the previous clip's audio is still arriving after video
has crossed, and those tail frames sit inside both clips' mark ranges —
so they took the new clip's offset, jumped forward by the overlap, and
collided with the new clip's own audio, which the muxer's monotonic nudge
then flattened. A remux confirmed it: the timeline length was already
correct and the original symptom was still there, 169 audio packets on
the tick floor.

A track's PTS only runs forward inside a clip, so its own backward step
to the next clip's IN is its crossing. That is per track, so the cursor
is too.

src/mux/demux_sink.rs

@@ -868,7 +868,7 @@ impl Stream for DemuxSink {
         let drives = Some(frame.track) == self.ref_video_track;
         // See `MkvMuxer::write_frame`: `None` is material outside the
         // playlist's clip marks and is dropped rather than emitted.
-        let Some(pts) = self.timeline.map(frame.pts, drives) else {
+        let Some(pts) = self.timeline.map(frame.pts, drives, frame.track) else {
             return Ok(());
         };
         if drives {

src/mux/mkv.rs

@@ -1411,7 +1411,7 @@ impl<W: Write + Seek> MkvMuxer<W> {
         // outside every clip's IN/OUT marks, which only a seam-plan-driven
         // title can report. Dropping it is the point: emitting it is what put
         // duplicate content on the timeline at a join.
-        let Some(pts_ns) = self.continuity.map(pts_ns, drives_epoch) else {
+        let Some(pts_ns) = self.continuity.map(pts_ns, drives_epoch, track_idx) else {
             return Ok(());
         };
         let raw_ticks = pts_ns / TIMESTAMP_SCALE_NS;

src/mux/timeline.rs

@@ -78,11 +78,27 @@ pub(crate) struct SeamClip {
 /// and material outside a clip's marks is dropped rather than emitted twice.
 pub(crate) struct SeamPlan {
     clips: Vec<SeamClip>,
-    /// Index of the clip the primary video is currently inside. Only video
+    /// Per-track position: (clip index, last raw PTS seen).
-    /// advances it, for the same reason only video drives epochs: the passive
+    ///
-    /// tracks are sparse and lag, so letting them advance the cursor would
+    /// Each track crosses a join on ITS OWN frame, not on video's. The demuxer
-    /// retire a clip while its audio was still arriving.
+    /// interleaves the tracks, so when video enters the next clip the previous
-    cursor: usize,
+    /// clip's audio and subtitle tails are still arriving — and at an OVERLAP
+    /// join those tail frames fall inside BOTH clips' mark ranges, so there is
+    /// no way to place them from the PTS alone. Sharing one cursor gave the
+    /// tail the new clip's offset, which threw it forward by the overlap and
+    /// made it collide with the new clip's own frames; the muxer's monotonic
+    /// nudge then flattened the collision onto the tick floor, which is exactly
+    /// the audio-ahead-of-picture symptom this type exists to remove.
+    ///
+    /// Indexed by track; grows on demand.
+    cursors: Vec<TrackPos>,
+}
+
+/// Where one track currently sits in the clip list.
+#[derive(Debug, Clone, Copy)]
+struct TrackPos {
+    clip: usize,
+    last_raw_ns: Option<i64>,
 }
 
 impl SeamPlan {
@@ -115,7 +131,7 @@ impl SeamPlan {
         }
         Some(Self {
             clips: out,
-            cursor: 0,
+            cursors: Vec::new(),
         })
     }
 
@@ -130,61 +146,62 @@ impl SeamPlan {
             .fold(0i64, |a, b| a.saturating_add(b))
     }
 
-    /// Place a raw PTS, advancing the cursor when `drives` (primary video).
+    /// Place a raw PTS for `track`, advancing that track's own clip cursor.
     ///
-    /// `None` means DROP: the frame lies outside every clip's marks, which is
+    /// `None` means DROP: the frame lies outside every clip's marks, so the
-    /// material the playlist does not include — most often the overlap region a
+    /// playlist does not include it.
-    /// disc stores twice. Emitting it is what produced duplicate content and a
+    fn place(&mut self, raw_ns: i64, track: usize) -> Option<i64> {
-    /// backward DTS step at the join.
+        if self.cursors.len() <= track {
-    fn place(&mut self, raw_ns: i64, drives: bool) -> Option<i64> {
+            self.cursors.resize(
-        if drives {
+                track + 1,
-            // Advance to the next clip when this frame is its opening frame.
+                TrackPos {
+                    clip: 0,
+                    last_raw_ns: None,
+                },
+            );
+        }
+        let pos = self.cursors[track];
+        let mut clip = pos.clip;
+
+        // Advance this track to the next clip when THIS track's frames say it
+        // has crossed. Two signatures, because a join is either a skip or an
+        // overlap:
         //
-            // Two signatures, because a join is either a skip or an overlap:
+        // - SKIP (next IN after this OUT): the frame is simply past this clip's
-            //
+        //   OUT mark.
-            // - SKIP (next IN is after this OUT): the frame is simply past the
+        // - OVERLAP (next IN before this OUT, the disc storing the join twice):
-            //   current clip's OUT.
+        //   the frame is still inside this clip's range, so "past OUT" never
-            // - OVERLAP (next IN is BEFORE this OUT, the disc storing the join
+        //   fires. But a track's own PTS only ever runs forward inside a clip,
-            //   twice): the frame is still inside the current clip's range, so
+        //   so the backward step to the next clip's IN is the crossing — and it
-            //   "past OUT" never fires. But the clips are concatenated in file
+        //   is per track, which is why the cursor has to be per track too.
-            //   order, so the first frame of the new clip lands on its IN mark.
-            //   Recognising that is what distinguishes the new clip's opening
-            //   from the old clip's tail, which share a PTS range.
         //
         // Bounded by the clip count, so a wild PTS cannot spin here.
-            while self.cursor + 1 < self.clips.len() {
+        while clip + 1 < self.clips.len() {
-                let cur_out = self.clips[self.cursor].out_ns;
+            let cur = self.clips[clip];
-                let next_in = self.clips[self.cursor + 1].in_ns;
+            let next_in = self.clips[clip + 1].in_ns;
-                let past_out = raw_ns > cur_out;
+            let past_out = raw_ns > cur.out_ns;
-                let at_next_in =
+            // The step must land ON the next clip's IN mark, not merely below
-                    raw_ns >= next_in && raw_ns <= next_in.saturating_add(CLIP_START_TOLERANCE_NS);
+            // the current position: a bound only from above is satisfied by
-                if past_out || at_next_in {
+            // every later clip's IN too, and the cursor would run to the end of
-                    self.cursor += 1;
+            // the list on a single backward step.
+            let stepped_back = pos.last_raw_ns.is_some_and(|last| raw_ns < last)
+                && (raw_ns.saturating_sub(next_in)).abs() <= CLIP_START_TOLERANCE_NS;
+            if past_out || stepped_back {
+                clip += 1;
             } else {
                 break;
             }
         }
-            let c = self.clips[self.cursor];
+
+        let c = self.clips[clip];
+        self.cursors[track] = TrackPos {
+            clip,
+            last_raw_ns: Some(raw_ns),
+        };
         if raw_ns < c.in_ns || raw_ns > c.out_ns {
             return None;
         }
-            return Some(raw_ns.saturating_add(c.offset_ns));
+        Some(raw_ns.saturating_add(c.offset_ns))
-        }
-        // Passive track. Try the clip video is in, then the one before it: at a
-        // join the tracks do not switch on the same frame, so a lagging audio or
-        // subtitle frame from the previous clip can arrive after video has moved
-        // on. Checking both places it correctly instead of dropping it.
-        let cur = self.clips[self.cursor];
-        if raw_ns >= cur.in_ns && raw_ns <= cur.out_ns {
-            return Some(raw_ns.saturating_add(cur.offset_ns));
-        }
-        if self.cursor > 0 {
-            let prev = self.clips[self.cursor - 1];
-            if raw_ns >= prev.in_ns && raw_ns <= prev.out_ns {
-                return Some(raw_ns.saturating_add(prev.offset_ns));
-            }
-        }
-        None
     }
 }
 
@@ -273,12 +290,12 @@ impl TimelineContinuity {
     ///
     /// Dropping only ever happens under a [`SeamPlan`]: it is material outside
     /// the playlist's marks, which the title does not include.
-    pub(crate) fn map(&mut self, raw_pts_ns: i64, drives_epoch: bool) -> Option<i64> {
+    pub(crate) fn map(&mut self, raw_pts_ns: i64, drives_epoch: bool, track: usize) -> Option<i64> {
         if self.seams.is_some() {
             // Take the plan out for the call so `place` can borrow `self`
             // mutably without fighting the borrow checker over the whole struct.
             let mut plan = self.seams.take().expect("checked is_some");
-            let placed = plan.place(raw_pts_ns, drives_epoch);
+            let placed = plan.place(raw_pts_ns, track);
             self.seams = Some(plan);
             if let Some(p) = placed {
                 // Keep the frontier meaningful for anything that reads it, and
@@ -525,13 +542,11 @@ mod tests {
             let out_ns = mpls_ticks_to_ns(c.out_time);
             // First frame of the clip lands at the running total.
             let got = plan
-                .place(in_ns, true)
+                .place(in_ns, 0)
                 .expect("clip start is inside its marks");
             assert_eq!(got, expected_start, "clip {i} start misplaced");
             // Last frame lands at the running total plus the clip's length.
-            let end = plan
+            let end = plan.place(out_ns, 0).expect("clip end is inside its marks");
-                .place(out_ns, true)
-                .expect("clip end is inside its marks");
             assert_eq!(
                 end,
                 expected_start + (out_ns - in_ns),
@@ -559,8 +574,8 @@ mod tests {
             c3_in - c2_out > 9_000_000_000,
             "fixture should contain the ~9.17s skip"
         );
-        let end_of_2 = plan.place(c2_out, true).expect("in clip 2");
+        let end_of_2 = plan.place(c2_out, 0).expect("in clip 2");
-        let start_of_3 = plan.place(c3_in, true).expect("in clip 3");
+        let start_of_3 = plan.place(c3_in, 0).expect("in clip 3");
         assert_eq!(
             start_of_3, end_of_2,
             "clip 3 must begin exactly where clip 2 ended — the skip is not content"
@@ -581,23 +596,21 @@ mod tests {
         let c1_in = mpls_ticks_to_ns(clips[1].in_time);
         assert!(c1_in < c0_out, "fixture should contain the overlap");
         // Play clip 0 through to its OUT mark.
-        let last_of_0 = plan
+        let last_of_0 = plan.place(c0_out, 0).expect("clip 0 OUT is inside clip 0");
-            .place(c0_out, true)
-            .expect("clip 0 OUT is inside clip 0");
         // The next clip opens ON its IN mark. Under the old inference this was a
         // 1.79s backward step, below the reorder threshold, so no seam was
         // recognised and the join was emitted as duplicate content whose
         // timestamps then collided. With the marks known, clip 1 is placed to
         // continue exactly where clip 0 ended: one monotonic timeline, no
         // rewind, and no collision for the muxer to flatten.
-        let first_of_1 = plan.place(c1_in, true).expect("clip 1 IN");
+        let first_of_1 = plan.place(c1_in, 0).expect("clip 1 IN");
         assert_eq!(
             first_of_1, last_of_0,
             "clip 1 must continue from clip 0's end, not rewind by the overlap"
         );
         // And the timeline keeps moving forward from there.
         let into_1 = plan
-            .place(c1_in + 1_000_000_000, true)
+            .place(c1_in + 1_000_000_000, 0)
             .expect("1s into clip 1");
         assert_eq!(
             into_1,
@@ -615,14 +628,62 @@ mod tests {
         let c0_out = mpls_ticks_to_ns(clips[0].out_time);
         let c1_in = mpls_ticks_to_ns(clips[1].in_time);
         // Video crosses into clip 1.
-        plan.place(c1_in + 500_000_000, true).expect("in clip 1");
+        plan.place(c1_in + 500_000_000, 0).expect("in clip 1");
         // A straggler from clip 0's tail arrives afterwards.
         let tail = c0_out - 50_000_000; // 50ms before clip 0's OUT
-        let placed = plan.place(tail, false).expect("straggler must be placed");
+        let placed = plan.place(tail, 1).expect("straggler must be placed");
         let expected = tail + (0i64 - mpls_ticks_to_ns(clips[0].in_time));
         assert_eq!(placed, expected, "straggler must ride clip 0's offset");
     }
 
+    /// Each track crosses a join on its OWN frame.
+    ///
+    /// This is the regression for the first attempt at this fix, which gave
+    /// every track the cursor the video had moved. At an overlap the previous
+    /// clip's audio tail is still arriving after video has crossed, and those
+    /// tail frames sit inside BOTH clips' ranges — so a shared cursor gave them
+    /// the new clip's offset, threw them forward by the overlap, and made them
+    /// collide with the new clip's own audio. Measured on a real remux: 169
+    /// audio packets flattened onto the 0.1 ms tick floor and a 1.80 s jump,
+    /// i.e. the original symptom, still present after the timeline length was
+    /// already correct.
+    #[test]
+    fn each_track_crosses_a_join_on_its_own_frame() {
+        let clips = seamless_branching_clips();
+        let mut plan = SeamPlan::from_clips(&clips).expect("plan");
+        let c0_in = mpls_ticks_to_ns(clips[0].in_time);
+        let c0_out = mpls_ticks_to_ns(clips[0].out_time);
+        let c1_in = mpls_ticks_to_ns(clips[1].in_time);
+
+        // Audio (track 1) runs up to near clip 0's OUT.
+        let tail = c0_out - 200_000_000;
+        plan.place(c0_in, 1).expect("audio start");
+        let a_tail = plan.place(tail, 1).expect("audio tail");
+
+        // Video (track 0) crosses into clip 1 first.
+        plan.place(c0_out, 0).expect("video at clip 0 OUT");
+        plan.place(c1_in, 0).expect("video at clip 1 IN");
+
+        // Audio's NEXT tail frame still belongs to clip 0 and must stay there —
+        // contiguous with the previous one, not thrown forward by the overlap.
+        let a_tail2 = plan
+            .place(tail + 10_000_000, 1)
+            .expect("audio tail continues");
+        assert_eq!(
+            a_tail2 - a_tail,
+            10_000_000,
+            "audio tail must stay on clip 0's offset while video is already in clip 1"
+        );
+
+        // When audio itself steps back to clip 1's IN, it crosses — and lands
+        // after its own tail, with no rewind and no collision.
+        let a_new = plan.place(c1_in, 1).expect("audio crosses");
+        assert!(
+            a_new > a_tail2,
+            "audio must not rewind at the join (got {a_new} after {a_tail2})"
+        );
+    }
+
     /// Clips whose marks chain contiguously must come out byte-identical to the
     /// old behaviour: a constant offset, nothing moved, nothing dropped.
     ///
@@ -654,7 +715,7 @@ mod tests {
             6_410_000_000_000,
         ] {
             assert_eq!(
-                plan.place(t, true),
+                plan.place(t, 0),
                 Some(t),
                 "contiguous clips must not move a frame (t={t})"
             );
@@ -682,9 +743,9 @@ mod tests {
     #[test]
     fn map_without_a_plan_is_the_old_behaviour() {
         let mut tc = TimelineContinuity::new();
-        assert_eq!(tc.map(0, true), Some(0));
+        assert_eq!(tc.map(0, true, 0), Some(0));
-        assert_eq!(tc.map(5 * S, true), Some(5 * S));
+        assert_eq!(tc.map(5 * S, true, 0), Some(5 * S));
-        assert_eq!(tc.map(25 * S, false), Some(25 * S));
+        assert_eq!(tc.map(25 * S, false, 1), Some(25 * S));
         assert_eq!(tc.offset_ns, 0);
     }