Add exclude region marking in Structure step

Users can now draw rectangles on the document image in the Structure
Detection step to mark areas (e.g. header graphics, alphabet strips)
that should be excluded from OCR results during grid building.

- Backend: PUT/DELETE endpoints for exclude regions stored in structure_result
- Backend: _build_grid_core() filters all words inside user-defined exclude regions
- Frontend: Interactive rectangle drawing with visual overlay and delete buttons
- Preserve exclude regions when re-running structure detection

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

klausur-service/backend/grid_editor_api.py

@@ -792,13 +792,44 @@ async def _build_grid_core(session_id: str, session: dict) -> dict:
     # page number ("64", "S. 12") and not real content.
     _filter_footer_words(all_words, img_h, logger, session_id)
 
-    # 2d. Filter words inside detected graphic/image regions
+    # 2d. Filter words inside user-defined exclude regions (from Structure step).
+    # These are explicitly marked by the user, so ALL words inside are removed
+    # regardless of confidence.
+    structure_result = session.get("structure_result")
+    exclude_rects = []
+    if structure_result:
+        for er in structure_result.get("exclude_regions", []):
+            exclude_rects.append({
+                "x": er["x"], "y": er["y"],
+                "w": er["w"], "h": er["h"],
+            })
+    if exclude_rects:
+        before = len(all_words)
+        filtered = []
+        for w in all_words:
+            w_cx = w["left"] + w.get("width", 0) / 2
+            w_cy = w["top"] + w.get("height", 0) / 2
+            inside = any(
+                er["x"] <= w_cx <= er["x"] + er["w"]
+                and er["y"] <= w_cy <= er["y"] + er["h"]
+                for er in exclude_rects
+            )
+            if not inside:
+                filtered.append(w)
+        removed = before - len(filtered)
+        if removed:
+            all_words = filtered
+            logger.info(
+                "build-grid session %s: removed %d words inside %d user exclude region(s)",
+                session_id, removed, len(exclude_rects),
+            )
+
+    # 2e. Filter words inside detected graphic/image regions
     # Only remove LOW-CONFIDENCE words (likely OCR artifacts from images).
     # High-confidence words are real text even if they overlap a detected
     # graphic region (e.g. colored text that graphic detection couldn't
     # fully distinguish from an image).
     _GRAPHIC_CONF_THRESHOLD = 50  # keep words with conf >= 50
-    structure_result = session.get("structure_result")
     graphic_rects = []
     if structure_result:
         for g in structure_result.get("graphics", []):