fix: broaden phonetic bracket regex to catch Tesseract-garbled IPA

Tesseract mangles IPA square brackets into curly braces or parentheses
(e.g. China [ˈtʃaɪnə] → China {'tfatno]). The previous regex only
matched [...], missing all garbled variants.

- Match any bracket type: [...], {...}, (...) including mixed pairs
- Add _is_meaningful_bracket_content() to preserve legitimate German
  prefixes like (zer)brechen and Tanz(veranstaltung)
- Trigger IPA replacement on any bracket character, not just [

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

klausur-service/backend/cv_vocab_pipeline.py

@@ -4201,9 +4201,11 @@ def _attach_example_sentences(entries: List[Dict[str, Any]]) -> List[Dict[str, A
 
 # --- D. Phonetic Bracket IPA Replacement ---
 
-# Pattern: word [phonetic] or word (phonetic) — capture the word before brackets
+# Pattern: word followed by any bracket type containing phonetic content.
+# Tesseract often garbles IPA brackets: [ˈdɑːns] → {'tfatno] or (cy) etc.
+# Match any opener ([, {, () with any closer (], }, )) — even mixed pairs.
 _PHONETIC_BRACKET_RE = re.compile(
-    r'(\b[a-zA-ZäöüÄÖÜß]+)\s*\[([^\]]*)\]'
+    r'(\b[a-zA-ZäöüÄÖÜß]+)\s*[\[\{\(]([^\]\}\)]*?)[\]\}\)]'
 )
 
 
@@ -4274,25 +4276,64 @@ def _fix_phonetic_brackets(
     for entry in entries:
         for field in ('english', 'german', 'example'):
             text = entry.get(field, '') or ''
-            if '[' not in text:
+            # Check for any bracket type — Tesseract garbles [ into { or (
+            if not any(ch in text for ch in '[{('):
                 continue
             entry[field] = _replace_phonetics_in_text(text, pronunciation)
 
     return entries
 
 
+# German prefixes/words commonly in parentheses in vocab tables: (zer)brechen, Tanz(veranstaltung)
+_GERMAN_BRACKET_PREFIXES = frozenset({
+    'ab', 'an', 'auf', 'aus', 'be', 'bei', 'dar', 'ein', 'emp', 'ent',
+    'er', 'ge', 'her', 'hin', 'los', 'mit', 'nach', 'um', 'un', 'unter',
+    'ver', 'vor', 'weg', 'zer', 'zu', 'zurück',
+})
+
+
+def _is_meaningful_bracket_content(content: str) -> bool:
+    """Return True if bracket content is a meaningful word/prefix, not garbled IPA.
+
+    Meaningful: (zer)brechen, (veranstaltung), (Haupt)stadt
+    Garbled IPA: {'tfatno, (cy, 1u], 'daens
+    """
+    if not content:
+        return False
+    # Must be pure letters (no digits, punctuation, IPA symbols)
+    if not re.match(r'^[a-zA-ZäöüÄÖÜßé]+$', content):
+        return False
+    # Known German prefix
+    if content.lower() in _GERMAN_BRACKET_PREFIXES:
+        return True
+    # Long enough to be a real word (not 1-2 char garbled IPA like "cy")
+    if len(content) >= 4:
+        return True
+    return False
+
+
 def _replace_phonetics_in_text(text: str, pronunciation: str = 'british') -> str:
-    """Replace [phonetic] after words with dictionary IPA."""
+    """Replace [phonetic] / {phonetic} / (phonetic) after words with dictionary IPA.
+
+    Tesseract garbles IPA brackets, e.g. China [ˈtʃaɪnə] → China {'tfatno].
+    We match any bracket type and replace with dictionary IPA if found.
+    Legitimate parenthetical content like (zer)brechen or (veranstaltung) is preserved.
+    """
     if not IPA_AVAILABLE:
         return text
 
     def replacer(match):
         word = match.group(1)
-        ocr_phonetic = match.group(2)
+        bracket_content = match.group(2).strip()
 
-        # Skip if bracket content looks like regular text (has spaces + capitals)
+        # Skip if bracket content looks like regular text (multiple words)
-        if len(ocr_phonetic.split()) > 3:
+        if len(bracket_content.split()) > 3:
-            return match.group(0)  # Keep original
+            return match.group(0)
+
+        # Skip if bracket content is a meaningful word/prefix — e.g. (zer)brechen,
+        # Tanz(veranstaltung). These are real German morphemes, not garbled IPA.
+        if _is_meaningful_bracket_content(bracket_content):
+            return match.group(0)
 
         # Look up in IPA dictionary
         ipa = _lookup_ipa(word, pronunciation)
@@ -4886,13 +4927,15 @@ def build_cell_grid_v2(
     Drop-in replacement for build_cell_grid() — same signature & return type.
     No full-page word assignment; each cell is OCR'd from its own crop.
     """
-    # Resolve engine
+    # Resolve engine — default to Tesseract for cell-first OCR.
+    # Tesseract excels at isolated text crops (binarized, upscaled).
+    # RapidOCR is optimized for full-page scene-text and produces artifacts
+    # on small cell crops (extra chars, missing punctuation, garbled IPA).
     use_rapid = False
     if ocr_engine in ("trocr-printed", "trocr-handwritten", "lighton"):
         engine_name = ocr_engine
     elif ocr_engine == "auto":
-        use_rapid = RAPIDOCR_AVAILABLE and img_bgr is not None
+        engine_name = "tesseract"
-        engine_name = "rapid" if use_rapid else "tesseract"
     elif ocr_engine == "rapid":
         if not RAPIDOCR_AVAILABLE:
             logger.warning("RapidOCR requested but not available, falling back to Tesseract")
@@ -5034,13 +5077,15 @@ def build_cell_grid_v2_streaming(
     Yields:
         (cell_dict, columns_meta, total_cells)
     """
-    # Resolve engine
+    # Resolve engine — default to Tesseract for cell-first OCR.
+    # Tesseract excels at isolated text crops (binarized, upscaled).
+    # RapidOCR is optimized for full-page scene-text and produces artifacts
+    # on small cell crops (extra chars, missing punctuation, garbled IPA).
     use_rapid = False
     if ocr_engine in ("trocr-printed", "trocr-handwritten", "lighton"):
         engine_name = ocr_engine
     elif ocr_engine == "auto":
-        use_rapid = RAPIDOCR_AVAILABLE and img_bgr is not None
+        engine_name = "tesseract"
-        engine_name = "rapid" if use_rapid else "tesseract"
     elif ocr_engine == "rapid":
         if not RAPIDOCR_AVAILABLE:
             logger.warning("RapidOCR requested but not available, falling back to Tesseract")

klausur-service/backend/ocr_pipeline_api.py

@@ -1416,9 +1416,11 @@ async def _word_batch_stream_generator(
     # 2. Send preparing event (keepalive for proxy)
     yield f"data: {json.dumps({'type': 'preparing', 'message': 'Cell-First OCR laeuft parallel...'})}\n\n"
 
-    # 3. Run batch OCR in thread pool (CPU-bound, don't block event loop)
+    # 3. Run batch OCR in thread pool with periodic keepalive events.
+    # The OCR takes 30-60s and proxy servers (Nginx) may drop idle SSE
+    # connections after 30-60s. Send keepalive every 5s to prevent this.
     loop = asyncio.get_event_loop()
-    cells, columns_meta = await loop.run_in_executor(
+    ocr_future = loop.run_in_executor(
         None,
         lambda: build_cell_grid_v2(
             ocr_img, col_regions, row_geoms, img_w, img_h,
@@ -1426,6 +1428,25 @@ async def _word_batch_stream_generator(
         ),
     )
 
+    # Send keepalive events every 5 seconds while OCR runs
+    keepalive_count = 0
+    while not ocr_future.done():
+        try:
+            cells, columns_meta = await asyncio.wait_for(
+                asyncio.shield(ocr_future), timeout=5.0,
+            )
+            break  # OCR finished
+        except asyncio.TimeoutError:
+            keepalive_count += 1
+            elapsed = int(time.time() - t0)
+            yield f"data: {json.dumps({'type': 'keepalive', 'elapsed': elapsed, 'message': f'OCR laeuft... ({elapsed}s)'})}\n\n"
+            if await request.is_disconnected():
+                logger.info(f"SSE batch: client disconnected during OCR for {session_id}")
+                ocr_future.cancel()
+                return
+    else:
+        cells, columns_meta = ocr_future.result()
+
     if await request.is_disconnected():
         logger.info(f"SSE batch: client disconnected after OCR for {session_id}")
         return