Implemented a second LLM stage for grammatical review

tests/test_prompts.py

@@ -1,7 +1,7 @@
 import json
 
 from audita.chunking import chunk_transcript
-from audita.prompts import build_glossary_correction_messages
+from audita.prompts import build_glossary_correction_messages, build_grammar_correction_messages
 from audita.schemas import parse_glossary_yaml, parse_transcript_json
 
 
@@ -65,3 +65,62 @@ def test_prompt_uses_simplified_segment_payload():
     assert "speaker" not in prompt_segments[0]
     assert "start" not in prompt_segments[0]
     assert "end" not in prompt_segments[0]
+
+
+def test_grammar_prompt_limits_readability_corrections_and_protects_glossary():
+    transcript = parse_transcript_json(
+        """
+        [
+          {"id": 1, "speaker": "Eric", "start": 0.0, "end": 1.0, "text": "then lyra went their"}
+        ]
+        """
+    )
+    glossary = parse_glossary_yaml(
+        """
+        glossary:
+          - name: "Lyra"
+            category: npc
+            summary: "Lyra is a hostile NPC."
+        """
+    )
+    section = chunk_transcript(transcript, max_section_tokens=16000)[0]
+
+    messages = build_grammar_correction_messages(section, glossary)
+    prompt_text = "\n".join(message["content"] for message in messages)
+
+    assert "capitalization" in prompt_text
+    assert "commas, periods, em dashes, and ellipses" in prompt_text
+    assert "homophone fixes" in prompt_text
+    assert "spelling fixes" in prompt_text
+    assert "Do not paraphrase" in prompt_text
+    assert "protected vocabulary" in prompt_text
+    assert "appears exactly once" in prompt_text
+    assert "Do not return speaker, start, or end fields" in prompt_text
+
+
+def test_grammar_prompt_uses_simplified_segment_payload():
+    transcript = parse_transcript_json(
+        """
+        [
+          {"id": 1, "speaker": "Eric", "start": 0.0, "end": 1.0, "text": "then lyra went their"}
+        ]
+        """
+    )
+    glossary = parse_glossary_yaml(
+        """
+        glossary:
+          - name: "Lyra"
+            category: npc
+            summary: "Lyra is a hostile NPC."
+        """
+    )
+    section = chunk_transcript(transcript, max_section_tokens=16000)[0]
+
+    messages = build_grammar_correction_messages(section, glossary)
+    transcript_json = messages[1]["content"].split("Transcript section:\n", maxsplit=1)[1]
+    prompt_segments = json.loads(transcript_json)
+
+    assert prompt_segments == [{"id": 1, "original_text": "then lyra went their"}]
+    assert "speaker" not in prompt_segments[0]
+    assert "start" not in prompt_segments[0]
+    assert "end" not in prompt_segments[0]