feat(voz): legenda por ênfase, forced align, IA local e correções de zoom/revisão

Trabalho da branch feat/revisao-enfases: pipeline de edição por voz ganha alinhamento forçado (whisperx), roteirização por LLM local (Ollama), e a etapa 5 (revisão de frases) passa a refletir de verdade o que é aplicado. - generate_subtitles_by_emphasis: legenda comum cobre o clipe inteiro, legenda dinâmica só nas frases de ênfase, e a comum é desativada (enabled="0") onde a dinâmica cobre, em vez de nunca ser gerada ali. - validate_subtitle_layout ignora títulos com enabled="0" — corrige falso positivo de colisão contra o que está desativado no lugar dele. - Corrige zoom/marcador sendo descartado quando a borda encosta exatamente no início de um corte. - Etapa 5 do Assistente: recarrega quando as decisões da IA mudam (com fresh=true, ignorando a revisão salva antiga) — resolve a dessincronia entre "ativa" na tela e o que já foi cortado no FCPXML. - Etapa "Processar" reaplica as decisões da revisão (_phrase_actions.json) antes da cadeia de remoção de silêncio/legendas — antes, desativar uma frase na etapa 5 não tinha efeito nenhum no vídeo final. - Etapa "Concluído" fundida em "Processar" — abrir no Final Cut/Finder aparece assim que termina, sem slide extra. - Palavra clicável na etapa 5 agora funciona como toggle (clique de novo desfaz) e mostra a própria ênfase (sublinhado colorido + peso da fonte). - fcpxml/forced_align.py, fcpxml/llm_local.py, ai_edit.py: alinhamento fonético via whisperx e roteirização local via Ollama/Gemma. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
2026-08-21 18:26:04 -04:00
parent 711c397dfe
commit 7b5aed79ee
36 changed files with 2922 additions and 624 deletions
@@ -0,0 +1,259 @@
+"""Tests for the local LLM integration (Ollama): prompt building, JSON
+extraction, and turning a model response into validated voice actions.
+
+The model itself is mocked — these tests cover the client contract
+(extraction, validation, message shape) without a running Ollama. A real
+end-to-end run lives in the manual test harness (see ENGINE notes) because it
+needs the local model server.
+"""
+
+import json
+
+import pytest
+
+from fcpxml import llm_local
+from fcpxml.llm_local import (
+    _extract_json,
+    build_edit_messages,
+    generate_voice_actions,
+)
+from fcpxml.voice_actions import parse_actions
+
+
+def _fake_timeline() -> dict:
+    """A minimal but well-formed voice timeline (matches build_voice_timeline)."""
+    return {
+        "version": "1.0",
+        "source": "demo.mp4",
+        "rotation": 0.0,
+        "language": "pt",
+        "layers": {
+            "transcript": True,
+            "acoustics": True,
+            "speakers": False,
+            "emotion": False,
+            "alignment": False,
+        },
+        "scales": {},
+        "summary": {
+            "duration": 12.0,
+            "speaker_count": 1,
+            "segment_count": 2,
+            "word_count": 4,
+            "avg_emphasis": 0.4,
+            "peak_selection": "top 2%",
+            "peak_count": 1,
+            "peak_moments": [],
+        },
+        "speakers": [
+            {
+                "id": "SPEAKER_00",
+                "name": "Speaker 1",
+                "speaking_seconds": 12.0,
+                "share": 1.0,
+                "segment_count": 2,
+                "avg_segment": 6.0,
+                "word_count": 4,
+                "samples": [],
+            }
+        ],
+        "segments": [
+            {
+                "start": 0.0,
+                "end": 6.0,
+                "speaker": "SPEAKER_00",
+                "text": "Hoje vamos falar de mastopexia.",
+                "gap_before": 0.0,
+                "take_boundary": False,
+                "avg_energy": 0.5,
+                "peak_emphasis": 0.6,
+                "emotion": "neutral",
+                "emotion_confidence": 0.0,
+                "arousal": 0.5,
+                "valence": 0.6,
+                "words": [
+                    {"text": "Hoje", "start": 0.1, "end": 0.5, "speaker": "SPEAKER_00",
+                     "energy": 0.3, "pitch_delta": 0.2, "rate_delta": 0.1, "pause_before": 0.1,
+                     "emphasis": 0.3, "emotion": "neutral", "emotion_confidence": 0.0,
+                     "arousal": 0.4, "valence": 0.6, "energy_raw": 0.3, "pitch_hz": 120.0},
+                    {"text": "vamos", "start": 0.6, "end": 0.9, "speaker": "SPEAKER_00",
+                     "energy": 0.4, "pitch_delta": 0.3, "rate_delta": 0.2, "pause_before": 0.0,
+                     "emphasis": 0.5, "emotion": "neutral", "emotion_confidence": 0.0,
+                     "arousal": 0.5, "valence": 0.6, "energy_raw": 0.4, "pitch_hz": 125.0},
+                ],
+            },
+            {
+                "start": 6.0,
+                "end": 12.0,
+                "speaker": "SPEAKER_00",
+                "text": "E aí cara, tá gravando?",
+                "gap_before": 0.2,
+                "take_boundary": False,
+                "avg_energy": 0.4,
+                "peak_emphasis": 0.3,
+                "emotion": "neutral",
+                "emotion_confidence": 0.0,
+                "arousal": 0.4,
+                "valence": 0.5,
+                "words": [
+                    {"text": "E", "start": 6.1, "end": 6.3, "speaker": "SPEAKER_00",
+                     "energy": 0.3, "pitch_delta": 0.2, "rate_delta": 0.1, "pause_before": 0.1,
+                     "emphasis": 0.3, "emotion": "neutral", "emotion_confidence": 0.0,
+                     "arousal": 0.4, "valence": 0.5, "energy_raw": 0.3, "pitch_hz": 120.0},
+                    {"text": "cara", "start": 6.4, "end": 6.7, "speaker": "SPEAKER_00",
+                     "energy": 0.4, "pitch_delta": 0.3, "rate_delta": 0.2, "pause_before": 0.0,
+                     "emphasis": 0.4, "emotion": "neutral", "emotion_confidence": 0.0,
+                     "arousal": 0.4, "valence": 0.5, "energy_raw": 0.4, "pitch_hz": 122.0},
+                ],
+            },
+        ],
+    }
+
+
+def test_extract_json_unfenced():
+    text = '{"source": "x", "actions": []}'
+    assert _extract_json(text) == {"source": "x", "actions": []}
+
+
+def test_extract_json_with_fence():
+    text = '```json\n{"source": "x", "actions": [{"kind": "cut", "start": 1, "end": 2}]}\n```'
+    data = _extract_json(text)
+    assert data["source"] == "x"
+    assert data["actions"][0]["kind"] == "cut"
+
+
+def test_extract_json_with_prose_around():
+    text = 'Aqui está:\n{"source": "x", "actions": []}\nfim.'
+    assert _extract_json(text) == {"source": "x", "actions": []}
+
+
+def test_extract_json_invalid_returns_none():
+    assert _extract_json("no json here") is None
+    assert _extract_json("") is None
+
+
+def test_extract_json_unwraps_wrapped_list():
+    # Some models wrap the expected {"source","actions"} object in a
+    # single-element list; the actions must still be found.
+    text = '```json\n[{"source": "x", "actions": [{"kind": "cut", "start": 1, "end": 2}]}]\n```'
+    data = _extract_json(text)
+    assert isinstance(data, dict)
+    assert data["source"] == "x"
+    assert data["actions"][0]["kind"] == "cut"
+
+
+def test_build_edit_messages_shape():
+    system, user = build_edit_messages(_fake_timeline())
+    assert "editor" in system.lower()
+    assert "demo.mp4" in user
+    assert "segments" in user
+    assert "SOMENTE" in user
+
+
+def test_generate_voice_actions_parses_model_json(monkeypatch):
+    canned = json.dumps({
+        "source": "demo.mp4",
+        "actions": [
+            {"kind": "cut", "start": 6.0, "end": 12.0,
+             "reason": "papo casual com a equipe"},
+            {"kind": "zoom", "start": 0.1, "end": 0.9, "params": {"scale": 1.3},
+             "reason": "ênfase em 'vamos'"},
+        ],
+    })
+    monkeypatch.setattr(llm_local, "ollama_chat", lambda **kwargs: canned)
+
+    result = generate_voice_actions(_fake_timeline(), model="test")
+    actions, errors = parse_actions([a.as_dict() for a in result["actions"]])
+    assert errors == []
+    assert len(actions) == 2
+    assert {a.kind for a in actions} == {"cut", "zoom"}
+    assert result["errors"] == []
+
+
+def test_generate_voice_actions_reports_bad_rows(monkeypatch):
+    canned = json.dumps({
+        "source": "demo.mp4",
+        "actions": [
+            {"kind": "cut", "start": 6.0, "end": 12.0, "reason": "ok"},
+            {"kind": "bogus", "start": 1, "end": 2},
+            {"kind": "zoom", "start": 0.1, "end": 0.05},
+        ],
+    })
+    monkeypatch.setattr(llm_local, "ollama_chat", lambda **kwargs: canned)
+
+    result = generate_voice_actions(_fake_timeline(), model="test")
+    assert len(result["actions"]) == 1
+    assert result["actions"][0].kind == "cut"
+    assert len(result["errors"]) >= 2
+
+
+def test_generate_voice_actions_handles_transport_error(monkeypatch):
+    def _boom(**kwargs):
+        raise RuntimeError("ollama down")
+    monkeypatch.setattr(llm_local, "ollama_chat", _boom)
+
+    result = generate_voice_actions(_fake_timeline(), model="test")
+    assert result["actions"] == []
+    assert any("ollama" in e.lower() for e in result["errors"])
+
+
+def test_build_edit_messages_is_compact():
+    """The prompt must drop the heavy per-word audio features so a real
+    recording fits in the model context (the 47k-token dump made Ollama drop
+    the connection)."""
+    import json as _json
+
+    timeline = _fake_timeline()
+    # pad with the kind of audio detail a real timeline carries
+    timeline["segments"][0]["words"][0]["samples"] = [0.1, 0.2, 0.3]
+    timeline["speakers"][0]["samples"] = [1, 2, 3]
+    raw = _json.dumps(timeline, ensure_ascii=False)
+    system, user = build_edit_messages(timeline)
+    assert "demo.mp4" in user and "segments" in user
+    # projected prompt is materially smaller than the raw timeline
+    assert len(user) < len(raw)
+    # heavy fields we deliberately drop never reach the model
+    assert '"samples"' not in user
+    assert '"energy_raw"' not in user
+    assert '"pitch_hz"' not in user
+
+
+def test_ollama_chat_wraps_empty_response(monkeypatch):
+    """A dropped connection that yields an empty body must surface as a clear
+    RuntimeError (not an unhandled JSONDecodeError crashing the pipeline)."""
+
+    class _Resp:
+        def raise_for_status(self):
+            pass
+
+        def json(self):
+            raise ValueError("Expecting value: line 1 column 1")
+
+    monkeypatch.setattr(llm_local.httpx, "post", lambda *a, **k: _Resp())
+    try:
+        llm_local.ollama_chat(model="test", messages=[{"role": "user", "content": "x"}])
+    except RuntimeError as exc:
+        assert "Falha ao falar com o modelo local" in str(exc)
+    else:
+        raise AssertionError("expected RuntimeError")
+
+
+def test_list_ollama_models_parses_tags(monkeypatch):
+    class _Resp:
+        def raise_for_status(self):
+            pass
+
+        def json(self):
+            return {"models": [{"name": "gemma3:4b"}, {"name": "gemma3:12b"}]}
+
+    monkeypatch.setattr(llm_local.httpx, "get", lambda *a, **k: _Resp())
+    assert llm_local.list_ollama_models() == ["gemma3:12b", "gemma3:4b"]
+
+
+def test_list_ollama_models_returns_empty_on_error(monkeypatch):
+    def _boom(*a, **k):
+        raise RuntimeError("ollama down")
+
+    monkeypatch.setattr(llm_local.httpx, "get", _boom)
+    assert llm_local.list_ollama_models() == []
+