diff --git a/main.py b/main.py
index df88370..0da589b 100644
--- a/main.py
+++ b/main.py
@@ -303,10 +303,13 @@ class MainWindow(QMainWindow):
         self.player.mpv.mute = checked
 
     def _seek_relative(self, delta_seconds: float):
+        # Unlike frame-stepping (which always pauses -- you're precisely
+        # inspecting a frame), a 5s skip is a coarse nudge while scrubbing
+        # around; seeking doesn't change mpv's pause state on its own, so
+        # just not touching play/pause here is enough to preserve whichever
+        # state the user was already in.
         if self.input_path is None:
             return
-        self.player.pause()
-        self.play_button.setText("Play")
         duration = self.player.mpv.duration or 0.0
         target = min(max(self.player.get_position() + delta_seconds, 0.0), duration)
         self.player.seek(target)
diff --git a/scripts/smoke_test_preserve_play_state.py b/scripts/smoke_test_preserve_play_state.py
new file mode 100644
index 0000000..6e8d82c
--- /dev/null
+++ b/scripts/smoke_test_preserve_play_state.py
@@ -0,0 +1,75 @@
+"""Verifies arrow-key skips preserve whichever play/pause state the user
+was already in, instead of always forcing a pause.
+
+Calls the real MainWindow._seek_relative() method directly rather than
+simulating the QShortcut keypress: QTest.keyClick against this shortcut was
+observed to be unreliable in this headless setup even with Qt's
+active-window state forced (see smoke_test_arrow_seek.py's note) --
+sometimes the simulated key silently doesn't reach the shortcut at all, and
+a debug script isolated that PlayerWidget.seek() itself works correctly
+while playing, so the flakiness is in the synthetic-keypress dispatch, not
+the app. The QShortcut -> _seek_relative wiring itself (a single line) was
+already independently confirmed with real xdotool keypresses against a
+live window in an earlier session.
+"""
+import sys
+import time
+from pathlib import Path
+
+from PyQt6.QtWidgets import QApplication
+
+sys.path.insert(0, str(Path(__file__).resolve().parent.parent))
+import main as app_main
+
+ROOT = Path(__file__).resolve().parent.parent
+CLIP = str(ROOT / "test_media" / "sync_test_INSYNC.mp4")
+
+
+def main():
+    qapp = QApplication(sys.argv)
+    win = app_main.MainWindow()
+    win.show()
+    qapp.processEvents()
+
+    win._load(CLIP)
+    qapp.processEvents()
+    time.sleep(0.2)
+    qapp.processEvents()
+
+    # -- Case 1: paused before the skip -> stays paused after
+    assert win.player.mpv.pause is True
+    before_pos = win.player.get_position()
+    win._seek_relative(5.0)
+    qapp.processEvents()
+    time.sleep(0.2)
+    qapp.processEvents()
+    after_pos = win.player.get_position()
+    print(f"paused case: before={before_pos:.3f}s after={after_pos:.3f}s")
+    assert win.player.mpv.pause is True, "should still be paused after skip"
+    assert win.play_button.text() == "Play"
+    assert after_pos > before_pos + 4.0, "should have actually skipped forward"
+    print("PASS: stays paused across arrow-key skip when it started paused")
+
+    # -- Case 2: playing before the skip -> stays playing after
+    win.player.seek(0.0)
+    win.player.play()
+    win.play_button.setText("Pause")
+    qapp.processEvents()
+    time.sleep(0.2)
+    before_pos = win.player.get_position()
+    win._seek_relative(5.0)
+    qapp.processEvents()
+    time.sleep(0.2)
+    qapp.processEvents()
+    assert win.player.mpv.pause is False, "should still be playing after skip"
+    assert win.play_button.text() == "Pause"
+    after_pos = win.player.get_position()
+    print(f"playing case: before={before_pos:.3f}s after={after_pos:.3f}s (should have jumped ~5s and kept playing)")
+    assert after_pos > before_pos + 4.0, "should have actually skipped forward"
+
+    win.player.shutdown()
+    print("\nPRESERVE PLAY STATE SMOKE TEST PASSED")
+
+
+if __name__ == "__main__":
+    main()