Extend glossary term protection to the initial glossary LLM stage

tests/test_pipeline.py

@@ -79,6 +79,136 @@ def test_pipeline_processes_with_fake_llm_and_cleans_work_dir(tmp_path):
     assert list((tmp_path / "work").iterdir()) == []
 
 
+def test_glossary_stage_can_correct_toward_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,
+    )
+    fake_client = FakeLLMClient(
+        [
+            CorrectionSet(corrections=[glossary_correction]),
+            CorrectionSet(corrections=[]),
+        ]
+    )
+
+    revised = process_transcript(
+        transcript,
+        glossary,
+        _config(tmp_path),
+        llm_client=fake_client,
+    )
+
+    assert revised[0].text == "Hrank moves."
+    assert list((tmp_path / "work").iterdir()) == []
+
+
+def test_glossary_stage_cannot_change_away_from_protected_term(tmp_path):
+    transcript = parse_source_transcript_json(
+        """
+        [
+          {"speaker": "Eric", "start": 0.0, "end": 1.0, "text": "Hrank moves."}
+        ]
+        """
+    )
+    glossary = parse_glossary_yaml(
+        """
+        glossary:
+          - name: "Hrank"
+            category: pc
+            summary: "Hrank is a player character."
+        """
+    )
+    glossary_reversal = CorrectionCandidate(
+        id=1,
+        original_text="Hrank",
+        corrected_text="Frank",
+        confidence=0.95,
+    )
+    fake_client = FakeLLMClient(
+        [
+            CorrectionSet(corrections=[glossary_reversal]),
+            CorrectionSet(corrections=[]),
+        ]
+    )
+
+    revised = process_transcript(
+        transcript,
+        glossary,
+        _config(tmp_path, glossary_max_llm_passes=1),
+        llm_client=fake_client,
+    )
+
+    assert revised[0].text == "Hrank moves."
+    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"))
+    skipped = diagnostics["skipped_corrections"][0]
+    assert skipped["stage"] == "glossary"
+    assert skipped["reason"] == "correction changes protected glossary term usage"
+
+
+def test_glossary_stage_cannot_decanonicalize_protected_term(tmp_path):
+    transcript = parse_source_transcript_json(
+        """
+        [
+          {"speaker": "Eric", "start": 0.0, "end": 1.0, "text": "Hrank moves."}
+        ]
+        """
+    )
+    glossary = parse_glossary_yaml(
+        """
+        glossary:
+          - name: "Hrank"
+            category: pc
+            summary: "Hrank is a player character."
+        """
+    )
+    decapitalization = CorrectionCandidate(
+        id=1,
+        original_text="Hrank",
+        corrected_text="hrank",
+        confidence=0.95,
+    )
+    fake_client = FakeLLMClient(
+        [
+            CorrectionSet(corrections=[decapitalization]),
+            CorrectionSet(corrections=[]),
+        ]
+    )
+
+    revised = process_transcript(
+        transcript,
+        glossary,
+        _config(tmp_path, glossary_max_llm_passes=1),
+        llm_client=fake_client,
+    )
+
+    assert revised[0].text == "Hrank moves."
+    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"))
+    skipped = diagnostics["skipped_corrections"][0]
+    assert skipped["stage"] == "glossary"
+    assert skipped["reason"] == "correction changes protected glossary term capitalization"
+
+
 def test_pipeline_normalizes_before_llm_prompts(tmp_path):
     transcript = parse_source_transcript_json(
         """
@@ -355,7 +485,7 @@ def test_grammar_stage_cannot_reverse_glossary_protected_term(tmp_path):
     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"]
+    assert diagnostics["skipped_corrections"][0]["reason"] == "correction changes protected glossary term usage"
 
 
 def test_grammar_stage_can_correct_toward_protected_term(tmp_path):
@@ -439,7 +569,7 @@ def test_grammar_stage_cannot_change_away_from_protected_term(tmp_path):
     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"]
+    assert diagnostics["skipped_corrections"][0]["reason"] == "correction changes protected glossary term usage"
 
 
 def test_grammar_stage_retries_repeated_span_and_applies_unique_retry(tmp_path):