Filter pipe-character word_boxes from OCR column divider artifacts

Step 4d removes "|" and "||" word_boxes that OCR produces when reading
physical vertical divider lines between columns. Also strips stray pipe
chars from cell text.

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

klausur-service/backend/grid_editor_api.py

@@ -1709,6 +1709,42 @@ async def _build_grid_core(session_id: str, session: dict) -> dict:
                 removed_oversized, oversized_threshold, z.get("zone_index", 0),
             )
 
+    # 4d. Remove pipe-character word_boxes (column divider artifacts).
+    # OCR reads physical vertical divider lines as "|" or "||" characters.
+    # These sit at consistent x positions near column boundaries and pollute
+    # cell text.  Remove them from word_boxes and rebuild cell text.
+    _PIPE_RE = re.compile(r"^\|+$")
+    for z in zones_data:
+        removed_pipes = 0
+        for cell in z.get("cells", []):
+            wbs = cell.get("word_boxes") or []
+            filtered = [wb for wb in wbs if not _PIPE_RE.match((wb.get("text") or "").strip())]
+            if len(filtered) < len(wbs):
+                removed_pipes += len(wbs) - len(filtered)
+                cell["word_boxes"] = filtered
+                cell["text"] = " ".join(
+                    wb.get("text", "").strip()
+                    for wb in sorted(filtered, key=lambda w: (w.get("top", 0), w.get("left", 0)))
+                    if wb.get("text", "").strip()
+                )
+        # Remove cells that became empty after pipe removal
+        if removed_pipes:
+            z["cells"] = [c for c in z.get("cells", []) if (c.get("word_boxes") or c.get("text", "").strip())]
+            logger.info(
+                "build-grid: removed %d pipe-divider word_boxes from zone %d",
+                removed_pipes, z.get("zone_index", 0),
+            )
+
+    # Also strip leading/trailing pipe chars from cell text that may remain
+    # from word_boxes that contained mixed text like "word|" or "|word".
+    for z in zones_data:
+        for cell in z.get("cells", []):
+            text = cell.get("text", "")
+            if "|" in text:
+                cleaned = text.replace("|", "").strip()
+                if cleaned != text:
+                    cell["text"] = cleaned
+
     # 5. Color annotation on final word_boxes in cells
     if img_bgr is not None:
         all_wb: List[Dict] = []