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

@@ -1030,6 +1030,15 @@ def _text_has_garbled_ipa(text: str) -> bool:
         # Contains IPA special characters
         if any(c in w for c in 'əɪɛɒʊʌæɑɔʃʒθðŋ'):
             return True
+        # Embedded apostrophe suggesting merged garbled IPA with stress mark.
+        # E.g. "Scotland'skotland" — OCR reads ˈ as '.
+        # Guard: apostrophe must be after ≥3 chars and before ≥3 lowercase
+        # chars to avoid contractions (don't, won't, o'clock).
+        if "'" in w and not w.startswith("'"):
+            apos_idx = w.index("'")
+            after = w[apos_idx + 1:]
+            if apos_idx >= 3 and len(after) >= 3 and after[0].islower():
+                return True
     return False
 
 
@@ -1183,6 +1192,19 @@ def _insert_missing_ipa(text: str, pronunciation: str = 'british') -> str:
                     if _lookup_ipa(clean_j, pronunciation):
                         kept.extend(words[j:])
                         break
+                # Merged token: dictionary word + garbled IPA stuck together.
+                # E.g. "fictionsalans'fIkfn" starts with "fiction".
+                # 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:
+                            kept.append(f"{prefix_j} [{prefix_ipa}]")
+                            break
+                    break  # rest of this token is garbled
                 # Otherwise — likely garbled phonetics, skip
             words = kept
             break

klausur-service/backend/grid_editor_api.py

@@ -858,6 +858,7 @@ async def _build_grid_core(session_id: str, session: dict) -> dict:
     # Single/two-column layouts are continuous text, not vocab tables.
     all_cells = [cell for z in zones_data for cell in z.get("cells", [])]
     total_cols = sum(len(z.get("columns", [])) for z in zones_data)
+    en_col_type = None
     if total_cols >= 3:
         # Find the column that contains IPA brackets → English headwords.
         # Count cells with bracket patterns per col_type.  The column with
@@ -872,7 +873,6 @@ async def _build_grid_core(session_id: str, session: dict) -> dict:
             if ct.startswith("column_") and '[' in txt:
                 col_bracket_count[ct] = col_bracket_count.get(ct, 0) + 1
         # Pick column with most bracket IPA patterns
-        en_col_type = None
         if col_bracket_count:
             en_col_type = max(col_bracket_count, key=col_bracket_count.get)
         else:
@@ -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.
@@ -1105,6 +1112,10 @@ async def _build_grid_core(session_id: str, session: dict) -> dict:
     slash_ipa_fixed = 0
     for z in zones_data:
         for cell in z.get("cells", []):
+            # Only process English headword column — avoid converting
+            # German text like "der/die/das" to IPA.
+            if en_col_type and cell.get("col_type") != en_col_type:
+                continue
             text = cell.get("text", "")
             if "/" not in text:
                 continue
@@ -1292,7 +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:
@@ -1473,6 +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,