fix: detect red/coloured markers in OCR pipeline

Two fixes for marker column content (e.g. red "!" marks):

1. Skip _clean_cell_text() noise filter for column_marker — it
   requires 2+ consecutive letters, which drops punctuation-only
   markers like "!" or "*".

2. For marker columns, detect coloured pixels via HSV saturation
   check (S>80) in addition to grayscale darkness. Create a
   binarized image where both dark AND saturated pixels become
   black foreground, so Tesseract can see red markers that appear
   near-white in standard grayscale conversion.

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

klausur-service/backend/cv_vocab_pipeline.py

@@ -4621,10 +4621,39 @@ def _ocr_single_cell(
                 # in an otherwise empty cell.
                 dark_ratio = float(np.count_nonzero(crop < 180)) / crop.size
                 _run_fallback = dark_ratio > 0.005
+        # For marker columns, also check the colour image — red/coloured
+        # markers appear near-white in grayscale but have high saturation.
+        if not _run_fallback and img_bgr is not None and col.type == 'column_marker':
+            bgr_crop = img_bgr[cell_y:cell_y + cell_h, cell_x:cell_x + cell_w]
+            if bgr_crop.size > 0:
+                hsv = cv2.cvtColor(bgr_crop, cv2.COLOR_BGR2HSV)
+                # Saturation > 80 indicates coloured (non-grey) pixels
+                sat_ratio = float(np.count_nonzero(hsv[:, :, 1] > 80)) / (hsv.shape[0] * hsv.shape[1])
+                if sat_ratio > 0.005:
+                    _run_fallback = True
     if _run_fallback:
+        # For marker columns with coloured content (e.g. red "!"), convert
+        # the BGR crop to a binarized grayscale that preserves saturated pixels.
+        _marker_ocr_img = ocr_img
+        if col.type == 'column_marker' and img_bgr is not None:
+            bgr_crop_full = img_bgr[cell_y:cell_y + cell_h, cell_x:cell_x + cell_w]
+            if bgr_crop_full.size > 0:
+                hsv = cv2.cvtColor(bgr_crop_full, cv2.COLOR_BGR2HSV)
+                # Create mask: either dark (V<180) or saturated (S>60)
+                dark_mask = hsv[:, :, 2] < 180
+                sat_mask = hsv[:, :, 1] > 60
+                combined = dark_mask | sat_mask
+                # Build grayscale: foreground=0 (black), background=255 (white)
+                marker_gray = np.full(combined.shape, 255, dtype=np.uint8)
+                marker_gray[combined] = 0
+                # Place into a full-size image at the crop position
+                _marker_full = np.full_like(ocr_img, 255)
+                _marker_full[cell_y:cell_y + cell_h, cell_x:cell_x + cell_w] = marker_gray
+                _marker_ocr_img = _marker_full
+
         # For narrow columns, upscale the crop before OCR
-        if is_narrow and ocr_img is not None:
-            _crop_slice = ocr_img[cell_y:cell_y + cell_h, cell_x:cell_x + cell_w]
+        if is_narrow and _marker_ocr_img is not None:
+            _crop_slice = _marker_ocr_img[cell_y:cell_y + cell_h, cell_x:cell_x + cell_w]
             _upscaled = _ensure_minimum_crop_size(_crop_slice)
             if _upscaled is not _crop_slice:
                 # Build a temporary full-size image with the upscaled crop
@@ -4653,7 +4682,7 @@ def _ocr_single_cell(
                 )
                 _cell_psm = _select_psm_for_column(col.type, col.width, row.height)
                 cell_lang = lang_map.get(col.type, lang)
-                fallback_words = ocr_region(ocr_img, cell_region,
+                fallback_words = ocr_region(_marker_ocr_img, cell_region,
                                             lang=cell_lang, psm=_cell_psm)
         else:
             cell_region = PageRegion(
@@ -4670,7 +4699,7 @@ def _ocr_single_cell(
             else:
                 _cell_psm = _select_psm_for_column(col.type, col.width, row.height)
                 cell_lang = lang_map.get(col.type, lang)
-                fallback_words = ocr_region(ocr_img, cell_region,
+                fallback_words = ocr_region(_marker_ocr_img, cell_region,
                                             lang=cell_lang, psm=_cell_psm)
 
         if fallback_words:
@@ -4694,7 +4723,7 @@ def _ocr_single_cell(
                 width=cell_w, height=cell_h,
             )
             cell_lang = lang_map.get(col.type, lang)
-            psm7_words = ocr_region(ocr_img, _fb_region, lang=cell_lang, psm=7)
+            psm7_words = ocr_region(_marker_ocr_img, _fb_region, lang=cell_lang, psm=7)
             if psm7_words:
                 psm7_words = [w for w in psm7_words if w.get('conf', 0) >= _MIN_WORD_CONF]
             if psm7_words:
@@ -4739,7 +4768,9 @@ def _ocr_single_cell(
                         used_engine = 'row_strip_rapid'
 
     # --- NOISE FILTER: clear cells that contain only OCR artifacts ---
-    if text.strip():
+    # Skip noise filter for marker columns — they legitimately contain
+    # only punctuation like "!" or "*" which _clean_cell_text would remove.
+    if text.strip() and col.type != 'column_marker':
         text = _clean_cell_text(text)
         if not text:
             avg_conf = 0.0