fix(display): order snapshot writes, retry a failed one, stop the writer on cleanup; measure refresh from the low end

From review (CodeRabbit):

- A static frame's inline save now waits for a queued write in flight and
  drops a stale queued frame, under one write lock taken by both paths, so
  the last scrolling frame can no longer land on top of the first static one.
- A background write that fails clears the recorded digest, so an unchanged
  frame is written again instead of mtime-touching a stale file into looking
  healthy.
- cleanup() stops and joins the writer thread, dropping a frame it hasn't
  started, instead of leaving it (and its DisplayManager) alive.
- Vegas measures the panel's refresh from the 10th percentile of swap gaps,
  not the median: a late swap only lengthens a gap, so with most frames late
  in the window the median tracked the loop and locked in a rate too low
  (then the scroll ran fast until restart). Samples are dropped whenever
  the pacing is re-solved, so gaps from an old frame hold aren't divided by
  a new one.

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

src/display_manager.py

@@ -199,6 +199,10 @@ class DisplayManager:
         self._snapshot_cond = threading.Condition()
         self._snapshot_pending: Optional[Image.Image] = None
         self._snapshot_thread: Optional[threading.Thread] = None
+        self._snapshot_stop = False
+        # Held for the whole of each PNG write, by the writer thread and by
+        # the inline static path, so the two land on disk in order.
+        self._snapshot_write_lock = threading.Lock()
         self._viewer_check_ts = 0.0
         self._viewer_fresh = False
         self._viewer_was_fresh = False
@@ -1274,6 +1278,8 @@ class DisplayManager:
 
     def cleanup(self):
         """Clean up resources."""
+        if hasattr(self, '_snapshot_cond'):
+            self._stop_snapshot_writer()
         if hasattr(self, 'matrix') and self.matrix is not None:
             try:
                 self.matrix.Clear()
@@ -1611,7 +1617,12 @@ class DisplayManager:
             if self.is_currently_scrolling():
                 self._queue_snapshot(self.image.copy())
             else:
-                self._save_snapshot(self.image)
+                # A scroll that just ended can leave its last frame queued or
+                # mid-write; it must not land on top of this newer one.
+                with self._snapshot_write_lock:
+                    with self._snapshot_cond:
+                        self._snapshot_pending = None
+                    self._save_snapshot(self.image)
             self._last_snapshot_ts = now
             self._last_snapshot_touch_ts = now
             self._last_snapshot_digest = digest
@@ -1696,10 +1707,35 @@ class DisplayManager:
     def _snapshot_writer(self) -> None:
         while True:
             with self._snapshot_cond:
-                while self._snapshot_pending is None:
+                while self._snapshot_pending is None and not self._snapshot_stop:
                     self._snapshot_cond.wait()
-                image, self._snapshot_pending = self._snapshot_pending, None
-            try:
-                self._save_snapshot(image)
-            except Exception as e:
-                self._log_snapshot_failure(e)
+                if self._snapshot_stop:
+                    return          # shutting down: a pending frame is dropped
+            # The write lock before the frame: whichever of this and an inline
+            # static save gets it first also writes first, and a static save
+            # clears the slot, so an older frame never lands on a newer one.
+            with self._snapshot_write_lock:
+                with self._snapshot_cond:
+                    image, self._snapshot_pending = self._snapshot_pending, None
+                if image is None:
+                    continue
+                try:
+                    self._save_snapshot(image)
+                except Exception as e:
+                    # The frame was recorded as written when it was queued.
+                    # Forget that, so an unchanged frame is written again
+                    # rather than only mtime-touching a stale file into
+                    # looking healthy.
+                    self._last_snapshot_digest = None
+                    self._log_snapshot_failure(e)
+
+    def _stop_snapshot_writer(self, timeout: float = 1.0) -> None:
+        """Stop the writer thread, dropping any frame it has not started."""
+        with self._snapshot_cond:
+            self._snapshot_stop = True
+            self._snapshot_pending = None
+            self._snapshot_cond.notify_all()
+        thread = self._snapshot_thread
+        if thread is not None and thread is not threading.current_thread():
+            thread.join(timeout)
+        self._snapshot_thread = None

src/vegas_mode/render_pipeline.py

@@ -137,6 +137,8 @@ class RenderPipeline:
         # a 95Hz panel, p99 21-28ms -- a visible hitch every few frames.
         self._crisp = None
         self._frame_hold = 1
+        # Gaps timed under the old hold would be divided by the new one.
+        self._swap_times.clear()
         if self.config.smooth_scroll and not self.config.sub_pixel_blend:
             self._crisp = solve_crisp(self.config.scroll_speed, self._refresh_hz())
             self._frame_hold = self._crisp.frame_hold
@@ -197,10 +199,15 @@ class RenderPipeline:
         refresh the panel never delivers -- 90px/s at "120Hz" is 3px every 4
         refreshes, visibly jumpy, where the real 95Hz allows 1px every refresh.
 
-        SwapOnVSync blocks for frame_hold refreshes, so the median gap between
-        swaps is frame_hold refresh periods. The median ignores the odd frame
-        that missed its vsync or waited on a recompose. Measured once: the
-        refresh only changes with the hardware config, which restarts us.
+        SwapOnVSync blocks for frame_hold refreshes, and a swap can only come
+        back late -- a missed vsync lengthens its gap by whole refreshes, never
+        shortens one -- so the low end of the gaps is frame_hold refresh
+        periods: the 10th percentile, as src/common/frame_timing.py uses. The
+        median would track the render loop instead once most frames in the
+        window were late (startup, a prefetch, a recompose), lock in a rate
+        too low, and scroll faster than configured until restart. Measured
+        once: the refresh only changes with the hardware config, which
+        restarts us.
         """
         if self._crisp is None or self._measured_hz is not None:
             return
@@ -214,11 +221,11 @@ class RenderPipeline:
 
         times = list(self._swap_times)
         gaps = sorted(b - a for a, b in zip(times, times[1:]))
-        median = gaps[len(gaps) // 2]
+        period = gaps[len(gaps) // 10]
         self._swap_times.clear()
-        if median <= 0:
+        if period <= 0:
             return
-        measured = self._frame_hold / median
+        measured = self._frame_hold / period
         cap = self._cap_hz()
         if measured >= cap * (1.0 - self.REFRESH_TOLERANCE):
             self._measured_hz = cap