fix: increase color recovery occupancy padding to prevent gap artifacts

Colored-pixel fragments in narrow inter-word gaps were being recovered
as false characters (e.g., "!" between "lend" and "sb."), disrupting
word order. Use adaptive padding based on median word height instead
of fixed 4px.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

klausur-service/backend/cv_color_detect.py

@@ -210,9 +210,14 @@ def recover_colored_text(
     ih, iw = img_bgr.shape[:2]
     max_area = int(ih * iw * 0.005)
 
-    # --- Build occupancy mask from existing words (with 4px padding) ---
+    # --- Build occupancy mask from existing words (adaptive padding) ---
+    # Pad word boxes generously to prevent colored-pixel artifacts in
+    # narrow inter-word gaps from being recovered as false characters.
+    heights = [wb["height"] for wb in existing_words if wb.get("height", 0) > 0]
+    median_h = int(np.median(heights)) if heights else 20
+    pad = max(8, int(median_h * 0.35))
+
     occupied = np.zeros((ih, iw), dtype=np.uint8)
-    pad = 4
     for wb in existing_words:
         x1 = max(0, int(wb["left"]) - pad)
         y1 = max(0, int(wb["top"]) - pad)