feat: Ausschnitte an Satzgrenzen erweitern (--exakt schneidet genau)

Bei -s/-e werden angeschnittene Sätze an beiden Enden vollständig
aufgenommen (höchstens 30 s über den Zeitraum hinaus). rede_korrigieren.py
weist das Modell bei Ausschnitten zusätzlich an, nichts zu ergänzen.

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

tests/test_extract_text.py

@@ -147,3 +147,21 @@ def test_time_range_order_error(monkeypatch, capsys):
         run_main(monkeypatch, "-u", URL, "-s", "2:00", "-e", "1:00")
     assert exc.value.code == 2
     assert "muss vor -e/--ende liegen" in capsys.readouterr().err
+
+
+CUT_VTT = (
+    "WEBVTT\n\n00:00:01.000 --> 00:00:02.000\nGuten Abend und\n\n"
+    "00:00:02.000 --> 00:00:03.000\nwillkommen. Schön, dass\n\n"
+    "00:00:03.000 --> 00:00:04.000\nSie da sind.\n"
+)
+
+
+@pytest.mark.parametrize("exact, text", [
+    ([], "Guten Abend und\nwillkommen. Schön, dass\nSie da sind.\n"),
+    (["--exakt"], "willkommen. Schön, dass\n"),
+])
+def test_time_range_sentence_bounds(fake_yt_dlp, tmp_path, monkeypatch, exact, text):
+    monkeypatch.setattr(yt.subprocess, "run", fake_yt_dlp(GERMAN_VIDEO, vtt=CUT_VTT))
+    target = tmp_path / "out.txt"
+    run_main(monkeypatch, "-u", URL, "-s", "2", "-e", "3", "-o", target, *exact)
+    assert target.read_text(encoding="utf-8") == text

tests/test_rede_korrigieren.py

@@ -615,3 +615,36 @@ def test_ollama_open_passes_http_errors(monkeypatch):
     monkeypatch.setattr(rk.urllib.request, "urlopen", not_found)
     with pytest.raises(rk.urllib.error.HTTPError):
         rk.ollama_open(rk.OLLAMA_URL)
+
+
+# === Excerpts: sentence bounds and prompt note ===
+
+@pytest.mark.parametrize("lang", ["de", "en"])
+def test_build_system_excerpt_note(lang):
+    prompts = rk.PROMPTS[lang]
+    assert prompts["excerpt"] not in rk.build_system(["X"], lang)
+    system = rk.build_system(["X"], lang, excerpt=True)
+    assert system.startswith(prompts["system"])
+    assert system.index(prompts["excerpt"]) < system.index(prompts["glossary"])
+    assert rk.build_system([], lang, excerpt=True).endswith(prompts["excerpt"] + "\n")
+
+
+CUT_VTT = (
+    "WEBVTT\n\n00:00:01.000 --> 00:00:02.000\nDer Kremmel ist\n\n"
+    "00:00:02.000 --> 00:00:03.000\nin Moskau. Das ist\n\n"
+    "00:00:03.000 --> 00:00:04.000\nbekannt.\n"
+)
+
+
+@pytest.mark.parametrize("exact, raw", [
+    ([], "Der Kremmel ist\nin Moskau. Das ist\nbekannt.\n"),
+    (["--exakt"], "in Moskau. Das ist\n"),
+])
+def test_main_url_excerpt(fake_yt_dlp, yt_dlp_installed, tmp_path, monkeypatch, exact, raw):
+    monkeypatch.setattr(subprocess, "run", fake_yt_dlp(VIDEO, vtt=CUT_VTT))
+    systems = recording_model(monkeypatch)
+    target = tmp_path / "rede.korrigiert.txt"
+    run_main(monkeypatch, "-u", URL, "-s", "2", "-e", "3", "-o", target, *exact)
+
+    assert (tmp_path / "rede.txt").read_text(encoding="utf-8") == raw
+    assert rk.PROMPTS["de"]["excerpt"] in systems[0]

tests/test_vtt2txt.py

@@ -188,17 +188,17 @@ def test_time_suffix(start, end, suffix):
     assert vtt2txt.time_suffix(start, end) == suffix
 
 
-def test_vtt_to_lines_time_range():
+def test_vtt_to_lines_exact_time_range():
     vtt = "WEBVTT\n\n" + cue(1, "A") + cue(2, "A", "B") + cue(3, "B", "C") + cue(4, "C", "D")
-    assert vtt2txt.vtt_to_lines(vtt, start=2, end=4) == ["B", "C"]
-    assert vtt2txt.vtt_to_lines(vtt, start=3) == ["C", "D"]
-    assert vtt2txt.vtt_to_lines(vtt, end=2) == ["A"]
+    assert vtt2txt.vtt_to_lines(vtt, start=2, end=4, exact=True) == ["B", "C"]
+    assert vtt2txt.vtt_to_lines(vtt, start=3, exact=True) == ["C", "D"]
+    assert vtt2txt.vtt_to_lines(vtt, end=2, exact=True) == ["A"]
 
 
 def test_vtt_to_lines_time_range_skips_window_from_before_start():
     # "B" was said before the range; its repetition in the first cue must not count
     vtt = "WEBVTT\n\n" + cue(1, "A", "B") + cue(2, "B", "C")
-    assert vtt2txt.vtt_to_lines(vtt, start=2) == ["C"]
+    assert vtt2txt.vtt_to_lines(vtt, start=2, exact=True) == ["C"]
 
 
 def test_main_time_range_default_name(sample_copy, monkeypatch):
@@ -225,3 +225,70 @@ def test_cli_time_errors(sample_copy, monkeypatch, capsys, args, message):
         run_main(monkeypatch, "-i", sample_copy, *args)
     assert exc.value.code == 2
     assert message in capsys.readouterr().err
+
+
+def timed_cue(seconds, *lines):
+    return f"{vtt2txt.format_time(seconds)}.000 --> {vtt2txt.format_time(seconds + 1)}.000\n" \
+        + "\n".join(lines) + "\n\n"
+
+
+def captions(*cues):
+    """VTT with one new line per cue: cues are (seconds, text)."""
+    return "WEBVTT\n\n" + "".join(timed_cue(t, text) for t, text in cues)
+
+
+SPEECH = captions(
+    (100, "Erster Satz ist hier."), (102, "Der zweite Satz"), (104, "beginnt vorher. Dann"),
+    (106, "kommt der dritte"), (108, "und endet. Der vierte"), (110, "läuft weiter"),
+    (112, "und endet auch. Schluss"), (114, "folgt."),
+)
+
+
+def test_snap_extends_both_sides_to_sentences():
+    # Range holds only "kommt der dritte" and "und endet. Der vierte"
+    assert vtt2txt.vtt_to_lines(SPEECH, start=106, end=110) == [
+        "Dann", "kommt der dritte", "und endet. Der vierte", "läuft weiter",
+        "und endet auch.",
+    ]
+
+
+def test_snap_keeps_range_already_at_sentence_bounds():
+    assert vtt2txt.vtt_to_lines(SPEECH, start=102, end=104) == [
+        "Der zweite Satz", "beginnt vorher.",
+    ]
+    assert vtt2txt.vtt_to_lines(SPEECH, start=100, end=102) == ["Erster Satz ist hier."]
+
+
+def test_snap_reaches_start_and_end_of_captions():
+    assert vtt2txt.vtt_to_lines(SPEECH, start=114) == ["Schluss", "folgt."]
+    speech = captions((0, "ohne Satzzeichen"), (2, "am Anfang. Rest"))
+    assert vtt2txt.vtt_to_lines(speech, start=2) == ["ohne Satzzeichen", "am Anfang. Rest"]
+
+
+def test_snap_cuts_exactly_without_sentence_end_in_reach():
+    far = vtt2txt.MAX_SNAP_SECONDS + 5
+    speech = captions(
+        (0, "Satz."), (2, "ein endloser"), (far, "Satz ohne"), (2 * far, "Ende"), (3 * far, "jetzt."),
+    )
+    assert vtt2txt.vtt_to_lines(speech, start=far, end=far + 1) == ["Satz ohne"]
+
+
+def test_snap_exact_keeps_cut():
+    assert vtt2txt.vtt_to_lines(SPEECH, start=106, end=110, exact=True) == [
+        "kommt der dritte", "und endet. Der vierte",
+    ]
+
+
+def test_snap_closing_quote_belongs_to_sentence():
+    speech = captions((0, "Er sagte „Nein.“ Und"), (2, "dann"), (4, "ging er."))
+    assert vtt2txt.vtt_to_lines(speech, start=2, end=4) == ["Und", "dann", "ging er."]
+
+
+def test_main_exact_option(tmp_path, monkeypatch):
+    source = tmp_path / "rede.vtt"
+    source.write_text(SPEECH, encoding="utf-8")
+    run_main(monkeypatch, "-i", source, "-s", "106", "-e", "110", "--exakt")
+    target = tmp_path / "rede_00-01-46_bis_00-01-50.txt"
+    assert target.read_text(encoding="utf-8") == "kommt der dritte\nund endet. Der vierte\n"
+    run_main(monkeypatch, "-i", source, "-s", "106", "-e", "110")
+    assert target.read_text(encoding="utf-8").endswith("und endet auch.\n")