Match the demo's segmentation parameters to the batch pipeline

The app passed retina_masks=True while scripts/run_yolo_inference.py uses the
Ultralytics defaults. Detections are identical either way, but full-resolution
masks move polygon boundaries: on a controlled crop, total projected area
shifted 2.6% and 14% of tubers crossed a USDA size threshold, so the demo
reported a different class composition than the published per-tuber CSV.

app/app.py

@@ -157,7 +157,10 @@ def _analyze_pil(image: Image.Image, source_label: str | None = None):
 
     Returns (overlay PIL image, headline markdown, class summary df, per-tuber df).
     """
-    detections = predict_image(_model, image, retina_masks=True)
+    # No retina_masks: the batch pipeline that produced the paper's tables runs
+    # with the Ultralytics defaults, and full-resolution masks shift polygon
+    # boundaries enough to move ~14% of tubers across a USDA class threshold.
+    detections = predict_image(_model, image)
     df = _per_tuber_dataframe(detections, image.width, image.height)
 
     if df.empty: