feat: Änderungsbericht listet eigenmächtige Modelländerungen „Zu prüfen“

Verglichen wird der Text nach den Glossar-Ersetzungen mit dem Ergebnis des
Modells. Gemeldet werden ersetzte gültige Wörter (Systemwortliste
/usr/share/dict/ngerman bzw. american-english), stark veränderte, eingefügte
und gestrichene Wörter; Hörfehler-Korrekturen, reine Zusammen-/Getrennt-
schreibung, Füllwörter und Glossarwörter nicht. Eigener Abschnitt im
HTML-Bericht, Rahmen im Text, Liste in der Konsole. aenderungen.py versteht
dafür -g und -l.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

tests/conftest.py

@@ -18,6 +18,22 @@ def ollama_available(request, monkeypatch):
         monkeypatch.setattr(module, "check_ollama", lambda model: None)
 
 
+@pytest.fixture(autouse=True)
+def small_wordlist(tmp_path_factory, monkeypatch):
+    """Reports use a small word list instead of the system one."""
+    import aenderungen
+    folder = tmp_path_factory.mktemp("wordlist")
+    de = folder / "de.txt"
+    de.write_text(
+        "\n".join(["Professor", "das", "dass", "ist", "Haus", "unterschreiben", "Bar"]),
+        encoding="utf-8",
+    )
+    monkeypatch.setattr(aenderungen, "WORDLISTS", {"de": de, "en": folder / "fehlt.txt"})
+    aenderungen.load_wordlist.cache_clear()
+    yield
+    aenderungen.load_wordlist.cache_clear()
+
+
 @pytest.fixture
 def fake_yt_dlp():
     """Factory for a subprocess.run stand-in.

tests/test_abschrift_korrigieren.py

@@ -336,6 +336,26 @@ def test_main_writes_change_report(sample_copy, monkeypatch, capsys):
     assert f"Bericht: {report}" in out
 
 
+def test_main_report_suspects_only_model_changes(sample_copy, monkeypatch, capsys):
+    glossary = sample_copy.with_name("g.txt")
+    glossary.write_text("Rammstein => Ramstein\n", encoding="utf-8")
+
+    def correct(chunk, index, system):
+        return chunk.strip().replace(" das ", " dies ", 1)
+
+    monkeypatch.setattr(ak, "correct_chunk", correct)
+    monkeypatch.setattr(ak, "find_paragraph_starts", every_nth_start(5))
+    run_main(monkeypatch, "-i", sample_copy, "-g", glossary)
+
+    page = sample_copy.with_name("ramstein.de.aenderungen.html").read_text(encoding="utf-8")
+    section = page.split("<h2>Zu prüfen")[1].split("<h2>Alle Änderungen")[0]
+    assert "Rammstein" in page and "Rammstein" not in section
+    assert "dies" in section
+    out = capsys.readouterr().out
+    assert "gültiges Wort ersetzt: „das“ → „dies“" in out
+    assert "Stelle(n) zu prüfen, siehe" in out
+
+
 def test_main_without_change_report(sample_copy, monkeypatch):
     monkeypatch.setattr(ak, "correct_chunk", identity_correct)
     monkeypatch.setattr(ak, "find_paragraph_starts", every_nth_start(5))

tests/test_aenderungen.py

@@ -117,3 +117,121 @@ def test_cli_missing_file(tmp_path, monkeypatch, capsys):
     with pytest.raises(SystemExit):
         ae.main()
     assert "Datei nicht gefunden" in capsys.readouterr().err
+
+
+# === Suspect changes ===
+
+WORDS = {"professor", "das", "dass", "ist", "haus", "unterschreiben", "bar"}
+
+
+def suspects(model_input, corrected, wordlist=WORDS, known=()):
+    found = ae.review_candidates(ae.compare(model_input, corrected), wordlist, set(known))
+    return [(s.reason, s.old, s.new) for s in found]
+
+
+@pytest.mark.parametrize("model_input, corrected, expected", [
+    ("Der Professor kam", "Der Professorin kam", [(ae.VALID_WORD, "Professor", "Professorin")]),
+    ("wir unterschreiben das", "wir unterscheiden das",
+     [(ae.VALID_WORD, "unterschreiben", "unterscheiden")]),
+    ("im CT1 heute", "im ZDF heute", [(ae.CHANGED_A_LOT, "CT1", "ZDF")]),
+    ("er kam heute", "er ist kam heute", [(ae.INSERTED, "", "ist")]),
+    ("er kam heute", "er kam", [(ae.DELETED, "heute", "")]),
+])
+def test_review_candidates_reports(model_input, corrected, expected):
+    assert suspects(model_input, corrected) == expected
+
+
+@pytest.mark.parametrize("model_input, corrected", [
+    ("das Kremmel heute", "das Kreml heute"),                 # recognition error
+    ("die Gradwanderung", "die Gratwanderung"),
+    ("die Clan Kriminalität", "die Clankriminalität"),        # spacing only
+    ("KI generierte Videos", "KI-generierte Videos"),         # hyphen only
+    ("er äh kam ähm heute", "er kam heute"),                  # fillers
+    ("das ist gut", "Das ist gut."),                          # case, punctuation
+])
+def test_review_candidates_ignores(model_input, corrected):
+    assert suspects(model_input, corrected) == []
+
+
+def test_review_candidates_glossary_words_are_known():
+    known = ae.glossary_words(["Egon Bahr"], [("Bars", "Bahrs")])
+    assert known == {"egon", "bahr", "bahrs"}
+    assert suspects("Bar sagte", "Bahr sagte", known=known) == []
+    assert suspects("Bar sagte", "Bahr sagte") == [(ae.VALID_WORD, "Bar", "Bahr")]
+
+
+def test_review_candidates_without_wordlist():
+    # Similar replacements of real words go unnoticed, the rest still counts
+    assert suspects("Der Professor kam", "Der Professorin kam", wordlist=None) == []
+    assert suspects("im CT1 heute", "im ZDF heute", wordlist=None) == [
+        (ae.CHANGED_A_LOT, "CT1", "ZDF")
+    ]
+
+
+def test_review_candidates_indices_point_into_corrected_words():
+    comparison = ae.compare("a b Professor c d", "A b Professorin c d.")
+    [suspect] = ae.review_candidates(comparison, WORDS, set())
+    assert comparison.words[suspect.j1:suspect.j2] == ["Professorin"]
+
+
+def test_load_wordlist(capsys):
+    assert "professor" in ae.load_wordlist("de")
+    assert ae.load_wordlist("en") is None
+    assert "keine Wortliste für „en“ (Paket wamerican)" in capsys.readouterr().out
+
+
+def suspect_section(page):
+    return page.split("<h2>Zu prüfen")[1].split("<h2>Alle Änderungen")[0]
+
+
+def test_report_suspects_only_against_model_input(tmp_path, capsys):
+    target = tmp_path / "bericht.html"
+    # Glossary turned "Kremmel" into "Kreml" before the model; the model
+    # replaced "Professor" and deleted "heute" on its own
+    comparison = ae.write_report(
+        "der Professor im Kremmel heute dort", "Der Professorin im Kreml dort.", target,
+        "rede", model_input="der Professor im Kreml heute dort",
+    )
+    page = target.read_text(encoding="utf-8")
+    section = suspect_section(page)
+
+    assert [s.reason for s in comparison.suspects] == [ae.VALID_WORD, ae.DELETED]
+    assert "Zu prüfen (2)" in page
+    assert "Professorin" in section and "Kremmel" not in section
+    assert "<span class=warn title='gültiges Wort ersetzt: „Professor“ → „Professorin“'>" in page
+    assert "<span class=warn title='gestrichen: „heute“ → „∅“'>⚠</span>" in page
+    # Framing changes nothing about the text itself
+    assert report_text(page) == "Der Professorin im Kreml ⚠ dort.".split()
+
+    ae.print_summary(comparison)
+    out = capsys.readouterr().out
+    assert "Zu prüfen: 2 Stellen" in out
+    assert "gültiges Wort ersetzt: „Professor“ → „Professorin“" in out
+
+
+def test_report_without_suspects(tmp_path):
+    target = tmp_path / "bericht.html"
+    comparison = ae.write_report("das Kremmel", "Das Kreml.", target, "rede")
+    assert comparison.suspects == []
+    assert "Keine auffälligen Änderungen" in suspect_section(target.read_text(encoding="utf-8"))
+
+
+def test_cli_glossary(tmp_path, monkeypatch, capsys):
+    raw = tmp_path / "rede.txt"
+    raw.write_text("das Kremmel und\nBar heute\n", encoding="utf-8")
+    corrected = tmp_path / "rede.korrigiert.txt"
+    corrected.write_text("Das Kreml und Bahr heute.\n", encoding="utf-8")
+    glossary = tmp_path / "g.txt"
+    glossary.write_text("Kremmel => Kreml\nEgon Bahr\n", encoding="utf-8")
+
+    argv = ["aenderungen.py", "-a", str(raw), "-b", str(corrected)]
+    monkeypatch.setattr(sys, "argv", argv)
+    ae.main()
+    assert "gültiges Wort ersetzt: „Bar“ → „Bahr“" in capsys.readouterr().out
+
+    monkeypatch.setattr(sys, "argv", [*argv, "-g", str(glossary)])
+    ae.main()
+    out = capsys.readouterr().out
+    assert "Zu prüfen: 0 Stellen" in out
+    # The report still compares against the raw transcript
+    assert "1× „Kremmel“ → „Kreml“" in out