Refactor to request matching/replacement substrings from LLMs, rather than requesting a complete replacement for the full original text

tests/test_pipeline.py

@@ -49,8 +49,8 @@ def _transcript():
 def test_pipeline_processes_with_fake_llm_and_cleans_work_dir(tmp_path):
     correction = CorrectionCandidate(
         segment_id=0,
-        original_text="I ask Chontia.",
-        corrected_text="I ask Chauntea.",
+        original_text="Chontia",
+        corrected_text="Chauntea",
         confidence=0.95,
     )
     fake_client = FakeLLMClient([CorrectionSet(corrections=[correction])])
@@ -71,7 +71,7 @@ def test_pipeline_skips_bad_correction_and_preserves_diagnostics(tmp_path):
     correction = CorrectionCandidate(
         segment_id=0,
         original_text="Different text.",
-        corrected_text="I ask Chauntea.",
+        corrected_text="Chauntea",
         confidence=0.95,
     )
     fake_client = FakeLLMClient([CorrectionSet(corrections=[correction])])
@@ -93,4 +93,4 @@ def test_pipeline_skips_bad_correction_and_preserves_diagnostics(tmp_path):
     assert skipped_path.exists()
     diagnostics = json.loads(skipped_path.read_text(encoding="utf-8"))
     assert diagnostics["skipped_corrections"][0]["segment_id"] == 0
-    assert "original_text" in diagnostics["skipped_corrections"][0]["reason"]
+    assert "does not match any substring" in diagnostics["skipped_corrections"][0]["reason"]