Implemented a second-stage validator for grammar corrections that allows homophone-related changes

tests/test_prompts.py

@@ -4,6 +4,7 @@ from audita.chunking import chunk_transcript
 from audita.prompts import (
     build_glossary_correction_messages,
     build_grammar_correction_messages,
+    build_grammar_spoken_form_validation_messages,
     build_grammar_validation_messages,
 )
 from audita.schemas import parse_glossary_yaml, parse_transcript_json
@@ -194,6 +195,28 @@ def test_grammar_validation_prompt_rejects_semantic_changes():
     assert "became visible" in prompt_text
     assert "became invisible" in prompt_text
     assert "reverses the meaning" in prompt_text
-    assert "spelling and homophone fixes only when" in prompt_text
+    assert "do not try to rescue likely homophone or transcription fixes" in prompt_text
+    assert "handled in a separate spoken-form validation step" in prompt_text
     assert "correction_index" in prompt_text
     assert "is_meaning_preserving" in prompt_text
+
+
+def test_grammar_spoken_form_validation_prompt_allows_homophone_rescue():
+    messages = build_grammar_spoken_form_validation_messages(
+        [
+            {
+                "correction_index": 0,
+                "id": 1,
+                "original_segment_text": "ChatGPT still can't really do that with a dam.",
+                "corrected_segment_text": "ChatGPT still can't really do that with a damn.",
+                "original_text": "dam",
+                "corrected_text": "damn",
+            }
+        ]
+    )
+    prompt_text = "\n".join(message["content"] for message in messages)
+
+    assert "likely homophone, spoken-form, or transcription fix" in prompt_text
+    assert '"dam" to "damn"' in prompt_text
+    assert '"became visible" to "became invisible"' in prompt_text
+    assert "is_likely_spoken_form_correction" in prompt_text