Grammar corrections now run through a deterministic guard that blocks protected glossary names/aliases from being changed

tests/test_pipeline.py

@@ -305,6 +305,59 @@ def test_grammar_stage_runs_after_glossary_and_sees_corrected_text(tmp_path):
     assert revised[0].text == "I ask Chauntea."
 
 
+def test_grammar_stage_cannot_reverse_glossary_protected_term(tmp_path):
+    transcript = parse_source_transcript_json(
+        """
+        [
+          {"speaker": "Eric", "start": 0.0, "end": 1.0, "text": "Frank moves."}
+        ]
+        """
+    )
+    glossary = parse_glossary_yaml(
+        """
+        glossary:
+          - name: "Hrank"
+            category: pc
+            summary: "Hrank is a player character."
+        """
+    )
+    glossary_correction = CorrectionCandidate(
+        id=1,
+        original_text="Frank",
+        corrected_text="Hrank",
+        confidence=0.95,
+    )
+    grammar_reversal = CorrectionCandidate(
+        id=1,
+        original_text="Hrank",
+        corrected_text="Frank",
+        confidence=0.95,
+    )
+    fake_client = FakeLLMClient(
+        [
+            CorrectionSet(corrections=[glossary_correction]),
+            CorrectionSet(corrections=[grammar_reversal]),
+        ]
+    )
+    progress = []
+
+    revised = process_transcript(
+        transcript,
+        glossary,
+        _config(tmp_path, grammar_max_llm_passes=1),
+        llm_client=fake_client,
+        progress=progress.append,
+    )
+
+    assert revised[0].text == "Hrank moves."
+    assert any("Skipping grammar correction for id 1" in message for message in progress)
+    run_dirs = list((tmp_path / "work").iterdir())
+    assert len(run_dirs) == 1
+    diagnostics = json.loads((run_dirs[0] / "skipped-corrections.json").read_text(encoding="utf-8"))
+    assert diagnostics["skipped_corrections"][0]["stage"] == "grammar"
+    assert "protected glossary term" in diagnostics["skipped_corrections"][0]["reason"]
+
+
 def test_grammar_stage_retries_repeated_span_and_applies_unique_retry(tmp_path):
     transcript = parse_source_transcript_json(
         """