feat(diarization): добавлено разделение транскрипта по говорящим

Зачем:
- локальным транскриптам нужна структура реплик для конспектов и протоколов.

Что:
- добавлены пословные таймкоды для всех ASR-бэкендов и сведение с Sherpa-ONNX.
- реализованы CLI-флаги, деградация без потери ASR и speaker Markdown.
- добавлены проверяемый кеш моделей, тесты и документация.

Проверка:
- `pytest` — 283 passed, 1 skipped.
- `pyright` — 0 errors.
- Ruff и `git diff --check` — без ошибок.
- выполнены три контрольных прогона на реальных записях.

tests/test_diarization.py

@@ -0,0 +1,120 @@
+from local_transcriber.diarization import build_speaker_transcript
+from local_transcriber.types import (
+    SmallSpeakerCluster,
+    SpeakerInterval,
+    SpeakerTurn,
+    Word,
+)
+
+
+def test_build_speaker_transcript_assigns_and_groups_words():
+    words = [
+        Word(start=0.0, end=0.8, text="Добрый"),
+        Word(start=0.8, end=1.4, text="день."),
+        Word(start=1.5, end=2.1, text="Привет!"),
+    ]
+    intervals = [
+        SpeakerInterval(start=0.0, end=1.4, cluster=7),
+        SpeakerInterval(start=1.4, end=2.3, cluster=3),
+    ]
+
+    transcript = build_speaker_transcript(words, intervals, recording_duration=30.0)
+
+    assert [
+        (turn.speaker, turn.start, turn.end, turn.text) for turn in transcript.turns
+    ] == [
+        (1, 0.0, 1.4, "Добрый день."),
+        (2, 1.5, 2.1, "Привет!"),
+    ]
+    assert transcript.cluster_count == 2
+    assert transcript.unassigned_word_count == 0
+
+
+def test_build_speaker_transcript_reports_small_cluster_without_filtering_it():
+    words = [
+        Word(start=0.0, end=1.0, text="Редкая реплика."),
+        Word(start=5.0, end=6.0, text="Основная реплика."),
+    ]
+    intervals = [
+        SpeakerInterval(start=0.0, end=4.9, cluster=4),
+        SpeakerInterval(start=5.0, end=10.0, cluster=9),
+    ]
+
+    transcript = build_speaker_transcript(words, intervals, recording_duration=100.0)
+
+    assert [turn.speaker for turn in transcript.turns] == [1, 2]
+    assert transcript.small_clusters == [SmallSpeakerCluster(speaker=1, duration=4.9)]
+
+
+def test_build_speaker_transcript_keeps_equal_overlap_unassigned():
+    words = [Word(start=0.0, end=1.0, text="Спорное слово")]
+    intervals = [
+        SpeakerInterval(start=0.0, end=0.1, cluster=8),
+        SpeakerInterval(start=0.3, end=0.5, cluster=8),
+        SpeakerInterval(start=0.0, end=0.3, cluster=2),
+    ]
+
+    transcript = build_speaker_transcript(words, intervals, recording_duration=10.0)
+
+    assert transcript.turns[0].speaker is None
+    assert transcript.unassigned_word_count == 1
+
+
+def test_build_speaker_transcript_keeps_word_without_overlap_unknown():
+    transcript = build_speaker_transcript(
+        [Word(start=5.0, end=6.0, text="Вне разметки")],
+        [SpeakerInterval(start=0.0, end=1.0, cluster=1)],
+        recording_duration=10.0,
+    )
+
+    assert transcript.turns == [SpeakerTurn(5.0, 6.0, "Вне разметки", None)]
+    assert transcript.unassigned_word_count == 1
+
+
+def test_build_speaker_transcript_splits_at_two_second_pause():
+    transcript = build_speaker_transcript(
+        [
+            Word(0.0, 1.0, "До паузы."),
+            Word(3.0, 4.0, "После паузы."),
+        ],
+        [SpeakerInterval(0.0, 4.0, 1)],
+        recording_duration=10.0,
+    )
+
+    assert [turn.text for turn in transcript.turns] == [
+        "До паузы.",
+        "После паузы.",
+    ]
+
+
+def test_build_speaker_transcript_does_not_exceed_sixty_seconds():
+    transcript = build_speaker_transcript(
+        [
+            Word(0.0, 30.0, "Начало."),
+            Word(30.0, 60.0, "Продолжение."),
+            Word(60.0, 61.0, "Новая реплика."),
+        ],
+        [SpeakerInterval(0.0, 61.0, 1)],
+        recording_duration=70.0,
+    )
+
+    assert [turn.text for turn in transcript.turns] == [
+        "Начало. Продолжение.",
+        "Новая реплика.",
+    ]
+
+
+def test_build_speaker_transcript_preserves_punctuation_without_leading_space():
+    transcript = build_speaker_transcript(
+        [
+            Word(0.0, 0.4, "Тарадата"),
+            Word(0.4, 0.5, "+"),
+            Word(0.5, 0.7, "Click"),
+            Word(0.7, 0.8, "—"),
+            Word(0.8, 1.0, "это"),
+        ],
+        [SpeakerInterval(0.0, 1.0, 1)],
+        recording_duration=10.0,
+    )
+
+    assert transcript.turns[0].text == "Тарадата+ Click— это"