feat: kurze Absätze nach Inhalt mit dem passenden Nachbarn verbinden

Passt ein Absatz unter MIN_PARAGRAPH_SENTENCES an beide Nachbarn,
entscheidet das Modell, ob er den vorherigen Gedanken abschließt oder den
folgenden einleitet (JSON {"anschluss": "vorher"|"nachher"}); passt nur
einer, wird ohne Modell verbunden. Antworten werden in .parts/
zwischengespeichert. Der Cache und der JSON-Aufruf sind dafür als
cached_json() und chat_json() herausgezogen, vorhandene Caches gelten weiter.

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

tests/test_abschrift_korrigieren.py

@@ -14,6 +14,14 @@ REPO = Path(__file__).resolve().parent.parent
 SAMPLE = REPO / "testdaten" / "ramstein.de.txt"
 GLOSSARY = REPO / "ramstein.glossar.txt"
 SHARED_GLOSSARY = REPO / "krone-schmalz.glossar.txt"
+# The autouse fixture below replaces it; tests of the real function use this
+REAL_FIND_MERGE_SIDE = ak.find_merge_side
+
+
+@pytest.fixture(autouse=True)
+def merge_to_previous(monkeypatch):
+    """No model call when joining short paragraphs: always the previous one."""
+    monkeypatch.setattr(ak, "find_merge_side", lambda *args: "vorher")
 
 
 def flat(text):
@@ -197,6 +205,40 @@ def test_find_paragraph_starts_bad_json(monkeypatch):
         ak.find_paragraph_starts(["s"] * 3, 1, ak.PARAGRAPH_SYSTEM)
 
 
+def merge_body(content):
+    return json.dumps({"message": {"content": content}}).encode("utf-8")
+
+
+@pytest.mark.parametrize("side", ["vorher", "nachher"])
+def test_find_merge_side(monkeypatch, side):
+    requests = []
+
+    def urlopen(request, timeout=None):
+        requests.append(json.loads(request.data))
+        return FakeResponse(merge_body(json.dumps({"anschluss": side})))
+
+    monkeypatch.setattr(ak.urllib.request, "urlopen", urlopen)
+    assert REAL_FIND_MERGE_SIDE("A.", "B.", "C.", ak.MERGE_SYSTEM) == side
+    payload = requests[0]
+    assert payload["format"] == ak.MERGE_FORMAT
+    assert payload["messages"][1]["content"] == "[vorher]\nA.\n\n[kurz]\nB.\n\n[nachher]\nC."
+
+
+@pytest.mark.parametrize("content", ["kein JSON", '{"anschluss": "egal"}', "[]"])
+def test_find_merge_side_bad_answer(monkeypatch, content):
+    monkeypatch.setattr(ak.urllib.request, "urlopen", fake_urlopen(merge_body(content)))
+    with pytest.raises(RuntimeError, match="unbrauchbare Antwort"):
+        REAL_FIND_MERGE_SIDE("A.", "B.", "C.", ak.MERGE_SYSTEM)
+
+
+def test_cached_merge_side(tmp_path, monkeypatch):
+    monkeypatch.setattr(ak, "find_merge_side", lambda *args: "nachher")
+    prefix = tmp_path / "verbinden0005"
+    assert ak.cached_merge_side("A.", "B.", "C.", "sys", prefix) == ("nachher", False)
+    monkeypatch.setattr(ak, "find_merge_side", no_model_calls)
+    assert ak.cached_merge_side("A.", "B.", "C.", "sys", prefix) == ("nachher", True)
+
+
 # === Sentences and paragraphs ===
 
 def test_split_sentences_keeps_closing_quotes():
@@ -242,7 +284,34 @@ def test_group_paragraphs():
     ([(1, 10), (11, 11)], [(1, 10), (11, 11)]),            # no room anywhere
 ])
 def test_merge_short_paragraphs(paragraphs, expected):
-    assert ak.merge_short_paragraphs(paragraphs) == expected
+    merged, sides = ak.merge_short_paragraphs(paragraphs)
+    assert merged == expected
+    assert sum(sides.values()) == len(paragraphs) - len(expected)
+
+
+def test_merge_short_paragraphs_asks_only_with_two_options():
+    calls = []
+
+    def choose(previous, short, following):
+        calls.append((previous, short, following))
+        return "nachher"
+
+    paragraphs = [(1, 2), (3, 6), (7, 7), (8, 11), (12, 21), (22, 22), (23, 26)]
+    merged, sides = ak.merge_short_paragraphs(paragraphs, choose)
+    # (1, 2) can only join its follower; (22, 22) only its follower, the
+    # previous one is full; (7, 7) has room on both sides
+    assert calls == [((1, 6), (7, 7), (8, 11))]
+    assert merged == [(1, 6), (7, 11), (12, 21), (22, 26)]
+    assert sides == {"nachher": 3}
+
+
+def test_merge_short_paragraphs_rechecks_joined_short_paragraph():
+    # Two single sentences in a row: the first joins the second, the pair is
+    # still too short and joins its neighbour
+    merged, _ = ak.merge_short_paragraphs(
+        [(1, 10), (11, 11), (12, 12), (13, 16)], lambda *args: "nachher"
+    )
+    assert merged == [(1, 10), (11, 16)]
 
 
 def paragraph_lengths(text):
@@ -287,9 +356,34 @@ def test_build_paragraphs_uses_cache(tmp_path, monkeypatch):
     first = ak.build_paragraphs(text, tmp_path)
 
     monkeypatch.setattr(ak, "find_paragraph_starts", no_model_calls)
+    monkeypatch.setattr(ak, "find_merge_side", no_model_calls)
     assert ak.build_paragraphs(text, tmp_path) == first
 
 
+def test_build_paragraphs_asks_where_short_paragraphs_belong(tmp_path, monkeypatch, capsys):
+    text = flat(SAMPLE.read_text(encoding="utf-8"))
+    # Paragraphs of 5 and 1 sentences alternate: every single one has room on both sides
+    monkeypatch.setattr(
+        ak, "find_paragraph_starts",
+        lambda sentences, first, system: {
+            n for n in range(max(first, 2), first + len(sentences)) if n % 6 in (1, 0)
+        },
+    )
+    asked = []
+
+    def find(previous, short, following, system):
+        asked.append(system)
+        return "nachher"
+
+    monkeypatch.setattr(ak, "find_merge_side", find)
+    result = ak.build_paragraphs(text, tmp_path)
+
+    assert asked and set(asked) == {ak.MERGE_SYSTEM}
+    assert flat(result) == text
+    assert "mit dem folgenden" in capsys.readouterr().out
+    assert min(paragraph_lengths(result)) >= ak.MIN_PARAGRAPH_SENTENCES
+
+
 # === Command line and end-to-end ===
 
 @pytest.fixture