Fix garbled IPA in continuation rows using headword lookup

IPA continuation rows (phonetic transcription that wraps below the
headword) now get proper IPA by looking up headwords from the row
above. E.g. "ska:f – ska:vz" → "[skˈɑːf] – [skˈɑːvz]".

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

klausur-service/backend/cv_ocr_engines.py

@@ -1096,6 +1096,74 @@ def _insert_missing_ipa(text: str, pronunciation: str = 'british') -> str:
     return ' '.join(words)
 
 
+def fix_ipa_continuation_cell(
+    garbled_text: str,
+    headword_text: str,
+    pronunciation: str = 'british',
+) -> str:
+    """Replace garbled IPA in a continuation row with proper IPA.
+
+    Continuation rows appear below the headword and contain only the
+    printed phonetic transcription, which OCR garbles into fragments
+    like ``ska:f – ska:vz`` (should be ``[skˈɑːf] – [skˈɑːvz]``).
+
+    Args:
+        garbled_text: The OCR-garbled IPA text from the continuation row.
+        headword_text: The headword text from the previous row
+            (e.g. ``scarf – scarves``).
+        pronunciation: ``'british'`` or ``'american'``.
+
+    Returns:
+        Corrected IPA text, or the original if no fix could be applied.
+    """
+    if not IPA_AVAILABLE or not garbled_text or not headword_text:
+        return garbled_text
+
+    # Strip existing IPA brackets from headword text
+    clean_hw = re.sub(r'\[[^\]]*\]', '', headword_text).strip()
+    if not clean_hw:
+        return garbled_text
+
+    # Split headword by delimiters (– — -)
+    # "scarf – scarves" → ["scarf", "scarves"]
+    # "see - saw - seen" → ["see", "saw", "seen"]
+    parts = re.split(r'\s*[–—]\s*|\s+-\s+', clean_hw)
+    parts = [p.strip() for p in parts if p.strip()]
+
+    if not parts:
+        return garbled_text
+
+    # Look up IPA for each headword part
+    ipa_parts: List[str] = []
+    for part in parts:
+        # A part may be multi-word like "secondary school"
+        words = part.split()
+        word_ipas: List[str] = []
+        for w in words:
+            clean_w = re.sub(r'[^a-zA-Z\'-]', '', w)
+            if not clean_w or len(clean_w) < 2:
+                continue
+            # Skip grammar words like "to" at the start
+            if clean_w.lower() in _GRAMMAR_BRACKET_WORDS:
+                continue
+            ipa = _lookup_ipa(clean_w, pronunciation)
+            if ipa:
+                word_ipas.append(ipa)
+        if word_ipas:
+            ipa_parts.append('[' + ' '.join(word_ipas) + ']')
+
+    if not ipa_parts:
+        return garbled_text
+
+    # Join with delimiter
+    result = ' – '.join(ipa_parts)
+    logger.debug(
+        "fix_ipa_continuation: '%s' → '%s' (headwords: '%s')",
+        garbled_text, result, headword_text,
+    )
+    return result
+
+
 def _insert_headword_ipa(text: str, pronunciation: str = 'british') -> str:
     """Insert IPA for the first English headword in a long mixed-language line.
 

klausur-service/backend/grid_editor_api.py

@@ -22,7 +22,7 @@ from fastapi import APIRouter, HTTPException, Request
 
 from cv_box_detect import detect_boxes, split_page_into_zones
 from cv_color_detect import detect_word_colors, recover_colored_text
-from cv_ocr_engines import fix_cell_phonetics
+from cv_ocr_engines import fix_cell_phonetics, fix_ipa_continuation_cell, _text_has_garbled_ipa
 from cv_words_first import _cluster_rows, _build_cells
 from ocr_pipeline_session_store import (
     get_session_db,
@@ -1324,9 +1324,60 @@ async def _build_grid_core(session_id: str, session: dict) -> dict:
             if orig:
                 cell["col_type"] = orig
 
-        # 5d. IPA continuation rows are preserved — they contain the
-        # printed phonetic transcription that wraps to a line below the
-        # headword.  The user can manually delete them if not needed.
+        # 5d. Fix IPA continuation rows — rows where the printed
+        # phonetic transcription wraps to a line below the headword.
+        # These contain only garbled IPA in the EN column and nothing
+        # in other columns.  Replace garbled text with proper IPA
+        # looked up from the headword in the previous row.
+        ipa_cont_fixed = 0
+        for z in zones_data:
+            rows_sorted = sorted(z.get("rows", []), key=lambda r: r["index"])
+            z_cells = z.get("cells", [])
+            for idx, row in enumerate(rows_sorted):
+                ri = row["index"]
+                row_cells = [c for c in z_cells if c.get("row_index") == ri]
+                en_cells = [
+                    c for c in row_cells
+                    if c.get("col_type") == en_col_type
+                ]
+                # Other cells with ≥3 chars (ignore margin noise)
+                other_cells = [
+                    c for c in row_cells
+                    if c.get("col_type") != en_col_type
+                    and len((c.get("text") or "").strip()) >= 3
+                ]
+                if not en_cells or other_cells:
+                    continue
+                en_text = en_cells[0].get("text", "")
+                if not _text_has_garbled_ipa(en_text):
+                    continue
+                # Already has proper IPA brackets → already fixed
+                if re.search(r'\[[^\]]*[ˈˌəɪɛɒʊʌæɑɔʃʒθðŋ][^\]]*\]', en_text):
+                    continue
+                # Find headword in previous row
+                if idx == 0:
+                    continue
+                prev_ri = rows_sorted[idx - 1]["index"]
+                prev_en = [
+                    c for c in z_cells
+                    if c.get("row_index") == prev_ri
+                    and c.get("col_type") == en_col_type
+                ]
+                if not prev_en:
+                    continue
+                prev_text = prev_en[0].get("text", "")
+                fixed = fix_ipa_continuation_cell(
+                    en_text, prev_text, pronunciation="british",
+                )
+                if fixed != en_text:
+                    en_cells[0]["text"] = fixed
+                    ipa_cont_fixed += 1
+                    logger.info(
+                        "IPA continuation R%d: '%s' → '%s'",
+                        ri, en_text, fixed,
+                    )
+        if ipa_cont_fixed:
+            logger.info("Fixed %d IPA continuation rows", ipa_cont_fixed)
 
     duration = time.time() - t0