feat(ocr-pipeline): line breaks, hyphen rejoin & oversized row splitting

- Preserve \n between visual lines within cells (instead of joining with space)
- Rejoin hyphenated words split across line breaks (e.g. Fuß-\nboden → Fußboden)
- Split oversized rows (>1.5× median height) into sub-entries when EN/DE
  line counts match — deterministic fix for missed Step 4 row boundaries
- Frontend: render \n as <br/>, use textarea for multiline editing

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

admin-lehrer/components/ocr-pipeline/StepWordRecognition.tsx

@@ -5,6 +5,16 @@ import type { WordResult, WordEntry, WordGroundTruth } from '@/app/(admin)/ai/oc
 
 const KLAUSUR_API = '/klausur-api'
 
+/** Render text with \n as line breaks */
+function MultilineText({ text }: { text: string }) {
+  if (!text) return <span className="text-gray-300 dark:text-gray-600">—</span>
+  const lines = text.split('\n')
+  if (lines.length === 1) return <>{text}</>
+  return <>{lines.map((line, i) => (
+    <span key={i}>{line}{i < lines.length - 1 && <br />}</span>
+  ))}</>
+}
+
 interface StepWordRecognitionProps {
   sessionId: string | null
   onNext: () => void
@@ -318,13 +328,13 @@ export function StepWordRecognition({ sessionId, onNext, goToStep }: StepWordRec
                       >
                         <td className="py-1 pr-2 text-gray-400">{idx + 1}</td>
                         <td className="py-1 pr-2 font-mono text-gray-700 dark:text-gray-300 cursor-pointer">
-                          {entry.english || <span className="text-gray-300 dark:text-gray-600">—</span>}
+                          <MultilineText text={entry.english} />
                         </td>
                         <td className="py-1 pr-2 font-mono text-gray-700 dark:text-gray-300 cursor-pointer">
-                          {entry.german || <span className="text-gray-300 dark:text-gray-600">—</span>}
+                          <MultilineText text={entry.german} />
                         </td>
-                        <td className="py-1 pr-2 font-mono text-gray-500 dark:text-gray-400 cursor-pointer max-w-[200px] truncate">
+                        <td className="py-1 pr-2 font-mono text-gray-500 dark:text-gray-400 cursor-pointer max-w-[200px]">
-                          {entry.example || <span className="text-gray-300 dark:text-gray-600">—</span>}
+                          <MultilineText text={entry.example} />
                         </td>
                         <td className={`py-1 text-right font-mono ${confColor(entry.confidence)}`}>
                           {entry.confidence}%
@@ -428,30 +438,30 @@ export function StepWordRecognition({ sessionId, onNext, goToStep }: StepWordRec
             <div className="space-y-2">
               <div>
                 <label className="text-[10px] font-medium text-gray-500 dark:text-gray-400">English</label>
-                <input
+                <textarea
-                  ref={enRef}
+                  ref={enRef as any}
-                  type="text"
+                  rows={Math.max(1, (editedEntries[activeIndex]?.english || '').split('\n').length)}
                   value={editedEntries[activeIndex]?.english || ''}
                   onChange={(e) => updateEntry(activeIndex, 'english', e.target.value)}
-                  className="w-full px-2 py-1.5 text-sm border rounded dark:bg-gray-700 dark:border-gray-600 font-mono"
+                  className="w-full px-2 py-1.5 text-sm border rounded dark:bg-gray-700 dark:border-gray-600 font-mono resize-none"
                 />
               </div>
               <div>
                 <label className="text-[10px] font-medium text-gray-500 dark:text-gray-400">Deutsch</label>
-                <input
+                <textarea
-                  type="text"
+                  rows={Math.max(1, (editedEntries[activeIndex]?.german || '').split('\n').length)}
                   value={editedEntries[activeIndex]?.german || ''}
                   onChange={(e) => updateEntry(activeIndex, 'german', e.target.value)}
-                  className="w-full px-2 py-1.5 text-sm border rounded dark:bg-gray-700 dark:border-gray-600 font-mono"
+                  className="w-full px-2 py-1.5 text-sm border rounded dark:bg-gray-700 dark:border-gray-600 font-mono resize-none"
                 />
               </div>
               <div>
                 <label className="text-[10px] font-medium text-gray-500 dark:text-gray-400">Example</label>
-                <input
+                <textarea
-                  type="text"
+                  rows={Math.max(1, (editedEntries[activeIndex]?.example || '').split('\n').length)}
                   value={editedEntries[activeIndex]?.example || ''}
                   onChange={(e) => updateEntry(activeIndex, 'example', e.target.value)}
-                  className="w-full px-2 py-1.5 text-sm border rounded dark:bg-gray-700 dark:border-gray-600 font-mono"
+                  className="w-full px-2 py-1.5 text-sm border rounded dark:bg-gray-700 dark:border-gray-600 font-mono resize-none"
                 />
               </div>
             </div>
@@ -503,7 +513,7 @@ export function StepWordRecognition({ sessionId, onNext, goToStep }: StepWordRec
                       'bg-gray-300 dark:bg-gray-600'
                     }`} />
                     <span className="truncate text-gray-600 dark:text-gray-400 font-mono">
-                      {entry.english || '—'} → {entry.german || '—'}
+                      {(entry.english || '—').replace(/\n/g, ' ')} → {(entry.german || '—').replace(/\n/g, ' ')}
                     </span>
                   </div>
                 ))}

klausur-service/backend/cv_vocab_pipeline.py

@@ -2173,20 +2173,65 @@ def analyze_layout_by_words(ocr_img: np.ndarray, dewarped_bgr: np.ndarray) -> Li
 # Pipeline Step 5: Word Grid from Columns × Rows
 # =============================================================================
 
-def _words_to_reading_order_text(words: List[Dict], y_tolerance_px: int = 15) -> str:
+def _words_to_reading_order_lines(words: List[Dict], y_tolerance_px: int = 15) -> List[str]:
-    """Join OCR words into text in correct reading order.
+    """Group OCR words into visual lines in reading order.
 
-    Groups words into visual lines by Y-tolerance, sorts each line by X,
+    Returns a list of line strings (one per visual line in the cell).
-    then joins lines with spaces. This fixes multi-line cell reading order.
     """
     if not words:
-        return ''
+        return []
 
     lines = _group_words_into_lines(words, y_tolerance_px=y_tolerance_px)
-    line_texts = []
+    return [' '.join(w['text'] for w in line) for line in lines]
-    for line in lines:
+
-        line_texts.append(' '.join(w['text'] for w in line))
+
-    return ' '.join(line_texts)
+def _rejoin_hyphenated(lines: List[str]) -> List[str]:
+    """Rejoin words split by line-break hyphenation.
+
+    E.g. ['Fuß-', 'boden'] → ['Fußboden']
+         ['some text-', 'thing here'] → ['something here']
+    """
+    if len(lines) <= 1:
+        return lines
+
+    result = []
+    i = 0
+    while i < len(lines):
+        line = lines[i]
+        # If line ends with '-' and there's a next line, rejoin
+        if i + 1 < len(lines) and line.rstrip().endswith('-'):
+            stripped = line.rstrip()
+            # Get the word fragment before hyphen (last word)
+            prefix = stripped[:-1]  # remove trailing hyphen
+            next_line = lines[i + 1]
+            # Join: last word of this line + first word of next line
+            prefix_words = prefix.rsplit(' ', 1)
+            next_words = next_line.split(' ', 1)
+            if len(prefix_words) > 1:
+                joined = prefix_words[0] + ' ' + prefix_words[1] + next_words[0]
+            else:
+                joined = prefix_words[0] + next_words[0]
+            remainder = next_words[1] if len(next_words) > 1 else ''
+            if remainder:
+                result.append(joined + ' ' + remainder)
+            else:
+                result.append(joined)
+            i += 2
+        else:
+            result.append(line)
+            i += 1
+    return result
+
+
+def _words_to_reading_order_text(words: List[Dict], y_tolerance_px: int = 15) -> str:
+    """Join OCR words into text in correct reading order, preserving line breaks.
+
+    Groups words into visual lines by Y-tolerance, sorts each line by X,
+    rejoins hyphenated words, then joins lines with newlines.
+    """
+    lines = _words_to_reading_order_lines(words, y_tolerance_px)
+    lines = _rejoin_hyphenated(lines)
+    return '\n'.join(lines)
 
 
 # --- RapidOCR integration (PaddleOCR models on ONNX Runtime) ---
@@ -2279,6 +2324,119 @@ def ocr_region_rapid(
     return words
 
 
+def _split_oversized_entries(
+    entries: List[Dict[str, Any]],
+    content_rows: List[RowGeometry],
+    img_w: int,
+    img_h: int,
+) -> List[Dict[str, Any]]:
+    """Split entries from oversized rows into multiple entries.
+
+    If a row is >1.5× the median height, it likely contains multiple vocabulary
+    entries that Step 4 failed to separate. We split based on line count:
+    if EN and DE have the same number of newline-separated lines, each line
+    becomes its own entry.
+
+    This is a deterministic plausibility check — no LLM needed.
+    """
+    if len(entries) < 3:
+        return entries
+
+    # Calculate median row height from pixel heights
+    row_heights = [r.height for r in content_rows]
+    row_heights_sorted = sorted(row_heights)
+    median_h = row_heights_sorted[len(row_heights_sorted) // 2]
+
+    if median_h <= 0:
+        return entries
+
+    height_threshold = median_h * 1.5
+    result: List[Dict[str, Any]] = []
+    split_count = 0
+
+    for entry in entries:
+        # Get pixel height from bbox percent
+        entry_h_px = entry['bbox']['h'] / 100.0 * img_h
+
+        if entry_h_px <= height_threshold:
+            result.append(entry)
+            continue
+
+        # This row is oversized — check if we can split
+        en_lines = entry['english'].split('\n') if entry['english'] else ['']
+        de_lines = entry['german'].split('\n') if entry['german'] else ['']
+        ex_lines = entry['example'].split('\n') if entry['example'] else ['']
+
+        # Filter empty lines
+        en_lines = [l for l in en_lines if l.strip()] or ['']
+        de_lines = [l for l in de_lines if l.strip()] or ['']
+        ex_lines = [l for l in ex_lines if l.strip()] or ['']
+
+        # Determine split count: EN and DE must agree (or one is empty)
+        n_en = len(en_lines)
+        n_de = len(de_lines)
+        n_ex = len(ex_lines)
+
+        can_split = False
+        n_split = 1
+
+        if n_en > 1 and n_de > 1 and n_en == n_de:
+            n_split = n_en
+            can_split = True
+        elif n_en > 1 and n_de <= 1:
+            # Only EN has multiple lines — still split, DE goes to first
+            n_split = n_en
+            can_split = True
+        elif n_de > 1 and n_en <= 1:
+            # Only DE has multiple lines
+            n_split = n_de
+            can_split = True
+
+        if not can_split or n_split <= 1:
+            result.append(entry)
+            continue
+
+        # Split into n_split sub-entries
+        orig_y = entry['bbox']['y']
+        orig_h = entry['bbox']['h']
+        sub_h = orig_h / n_split
+
+        for k in range(n_split):
+            sub_entry = {
+                'row_index': entry['row_index'],
+                'english': en_lines[k] if k < len(en_lines) else '',
+                'german': de_lines[k] if k < len(de_lines) else '',
+                'example': ex_lines[k] if k < len(ex_lines) else '',
+                'confidence': entry['confidence'],
+                'bbox': {
+                    'x': entry['bbox']['x'],
+                    'y': round(orig_y + k * sub_h, 2),
+                    'w': entry['bbox']['w'],
+                    'h': round(sub_h, 2),
+                },
+                'bbox_en': entry['bbox_en'],
+                'bbox_de': entry['bbox_de'],
+                'bbox_ex': entry['bbox_ex'],
+                'ocr_engine': entry.get('ocr_engine', ''),
+                'split_from_row': entry['row_index'],
+            }
+            result.append(sub_entry)
+
+        split_count += 1
+        logger.info(f"split_oversized: row {entry['row_index']} "
+                    f"(h={entry_h_px:.0f}px > {height_threshold:.0f}px) "
+                    f"→ {n_split} sub-entries")
+
+    if split_count > 0:
+        # Re-number row indices
+        for i, e in enumerate(result):
+            e['row_index'] = i
+        logger.info(f"split_oversized: {split_count} rows split, "
+                    f"{len(entries)} → {len(result)} entries")
+
+    return result
+
+
 def build_word_grid(
     ocr_img: np.ndarray,
     column_regions: List[PageRegion],
@@ -2433,6 +2591,9 @@ def build_word_grid(
         if entry['english'] or entry['german'] or entry['example']:
             entries.append(entry)
 
+    # --- Post-processing: split oversized rows ---
+    entries = _split_oversized_entries(entries, content_rows, img_w, img_h)
+
     logger.info(f"build_word_grid: {len(entries)} entries from "
                 f"{len(content_rows)} content rows × {len(relevant_cols)} columns "
                 f"(engine={engine_name})")