feat(chunk): paragraph mode giu xuong dong - translator.py

src/translator.py

@@ -22,6 +22,7 @@ from hardware import (
     detect_hardware_profile,
     resolve_gpu_indices,
 )
+from line_restore import assemble_paragraph_output
 from token_chunker import source_token_ids, split_for_translation
 
 import ctranslate2
@@ -1200,11 +1201,17 @@ class HachimiTranslator:
             yield end, total, f"Đã xong {end}/{total} chunk", None, None
 
         assemble_start = time.perf_counter()
-        rows = [
-            (index, chunk, translated)
-            for index, (chunk, translated) in enumerate(zip(chunks, translations), start=1)
-        ]
-        full_text = "\n".join(translations)
+        if chunk_mode == "paragraph":
+            # Chunk gom nhiều dòng → model nuốt xuống dòng. Khôi phục bố cục dòng
+            # gốc (cho full_text) + ghép rows 1:1 dòng-gốc/dòng-dịch để bảng đối
+            # chiếu khớp bản tải về. Xem line_restore.assemble_paragraph_output.
+            rows, full_text = assemble_paragraph_output(text, chunks, translations)
+        else:
+            rows = [
+                (index, chunk, translated)
+                for index, (chunk, translated) in enumerate(zip(chunks, translations), start=1)
+            ]
+            full_text = "\n".join(translations)
         self._profile_add("assemble_s", time.perf_counter() - assemble_start)
         cache_stats = getattr(self._tokenizer, "cache_stats", None)
         if callable(cache_stats):