Fix IPA correction persistence and false-positive prefix matching

Step 5i was overwriting IPA-corrected text from Step 5c when
reconstructing cells from word_boxes. Added _ipa_corrected flag
to preserve corrections. Also tightened merged-token prefix matching
(min prefix 4 chars, min suffix 3 chars) to prevent false positives
like "sis" being extracted from "si:said".

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

klausur-service/backend/cv_ocr_engines.py

@@ -1194,9 +1194,11 @@ def _insert_missing_ipa(text: str, pronunciation: str = 'british') -> str:
                         break
                 # Merged token: dictionary word + garbled IPA stuck together.
                 # E.g. "fictionsalans'fIkfn" starts with "fiction".
-                # Extract the dictionary prefix and add it with IPA.
-                if clean_j and len(clean_j) >= 5:
-                    for pend in range(min(len(clean_j), 15), 2, -1):
+                # Extract the dictionary prefix (≥4 chars) and add it with
+                # IPA, but only if enough chars remain after the prefix (≥3)
+                # to look like garbled IPA, not just a plural 's'.
+                if clean_j and len(clean_j) >= 7:
+                    for pend in range(min(len(clean_j) - 3, 15), 3, -1):
                         prefix_j = clean_j[:pend]
                         prefix_ipa = _lookup_ipa(prefix_j, pronunciation)
                         if prefix_ipa:

klausur-service/backend/grid_editor_api.py

@@ -890,11 +890,18 @@ async def _build_grid_core(session_id: str, session: dict) -> dict:
                 if cell.get("col_type") == en_col_type:
                     cell["_orig_col_type"] = en_col_type
                     cell["col_type"] = "column_en"
+        # Snapshot text before IPA fix to detect which cells were modified
+        _pre_ipa = {id(c): c.get("text", "") for c in all_cells}
         fix_cell_phonetics(all_cells, pronunciation="british")
         for cell in all_cells:
             orig = cell.pop("_orig_col_type", None)
             if orig:
                 cell["col_type"] = orig
+            # Mark cells whose text was changed by IPA correction so that
+            # later steps (5i) don't overwrite the corrected text when
+            # reconstructing from word_boxes.
+            if cell.get("text", "") != _pre_ipa.get(id(cell), ""):
+                cell["_ipa_corrected"] = True
 
         # 5d. Fix IPA continuation cells — cells where the printed
         # phonetic transcription wraps to a line below the headword.
@@ -1296,7 +1303,9 @@ async def _build_grid_core(session_id: str, session: dict) -> dict:
                 bullet_removed += len(to_remove)
                 filtered = [wb for i, wb in enumerate(wbs) if i not in to_remove]
                 cell["word_boxes"] = filtered
-                cell["text"] = _words_to_reading_order_text(filtered)
+                # Don't overwrite text that was corrected by Step 5c IPA fix
+                if not cell.get("_ipa_corrected"):
+                    cell["text"] = _words_to_reading_order_text(filtered)
 
     # Remove cells that became empty after bullet removal
     if bullet_removed:
@@ -1477,6 +1486,11 @@ async def _build_grid_core(session_id: str, session: dict) -> dict:
         except Exception as e:
             logger.warning("Syllable insertion failed: %s", e)
 
+    # Clean up internal flags before returning
+    for z in zones_data:
+        for cell in z.get("cells", []):
+            cell.pop("_ipa_corrected", None)
+
     result = {
         "session_id": session_id,
         "image_width": img_w,