cnn_bigru/tests/test_50_samples.py

"""test_50_samples.py — Teste de 50 amostras para erros lógicos ou falhas.

Executa o pipeline completo:
    1. Inicializa xeon_runtime + memory_optimizer
    2. Treina BBPE tokenizer em corpus sintético
    3. Cria dataset streaming multimodal (50 amostras)
    4. Instancia modelo multimodal + generator + verifier + anti-hallucination
    5. Executa treinamento cooperativo com:
       - Synergy search (N tentativas)
       - Hypothesis controller (ativa em punições)
       - Auto-learner (ajuste dinâmico de LR + spectral norm)
    6. Executa inferência com sampling
    7. Avalia perplexidade
    8. Reporta erros lógicos/falhas encontradas

Usage:
    python -m cnn_bigru.tests.test_50_samples
"""
from __future__ import annotations

import logging
import os
import sys
import time
import traceback
from pathlib import Path
from typing import Dict, List

# Setup paths (must come before torch import for xeon_runtime)
PROJECT_ROOT = Path(__file__).resolve().parent.parent.parent
sys.path.insert(0, str(PROJECT_ROOT))

# Ativa Xeon runtime ANTES de importar torch
from cnn_bigru.utils.xeon_runtime import optimize_xeon_environment
N_CORES = optimize_xeon_environment()

import numpy as np
import torch
from torch.utils.data import DataLoader

from cnn_bigru.tokenizer.bbpe_tokenizer import BBPETokenizer
from cnn_bigru.data.streaming_dataset import (
    MultimodalStreamingDataset,
    collate_multimodal,
)
from cnn_bigru.models.multimodal_model import MultimodalCNNBiGRU
from cnn_bigru.models.generator_verifier import (
    GeneratorCNNBiGRU,
    VerifierCNNBiGRU,
    AntiHallucinationLayer,
)
from cnn_bigru.models.rope import RotaryPositionEmbedding
from cnn_bigru.models.transformer_block import (
    TransformerBlockConfig,
    CausalSelfAttention,
    TransformerBlock,
    TransformerDecoderStack,
)
from cnn_bigru.models.context_window import (
    ContextWindowConfig,
    ContextWindowManager,
    KVCache,
)
from cnn_bigru.utils.ewc import EWCConfig, EWCState
from cnn_bigru.losses.losses import LossConfig, MultiLoss
from cnn_bigru.training.trainer import TrainerConfig, CooperativeTrainer
from cnn_bigru.training.auto_learner import (
    AutoLearnConfig,
    orthogonal_init_model,
)
from cnn_bigru.training.hypothesis_controller import (
    HypothesisConfig,
    HypothesisController,
)
from cnn_bigru.inference.inference import (
    generate_with_sampling,
    evaluate_perplexity,
    make_default_context_window,
)
from cnn_bigru.utils.memory_optimizer import MemoryOptimizer

logging.basicConfig(
    level=logging.INFO,
    format="%(asctime)s | %(levelname)s | %(name)s | %(message)s",
    datefmt="%H:%M:%S",
)
logger = logging.getLogger("test_50")


# ============================================================================
# Corpus para treino do tokenizer
# ============================================================================

CORPUS_TEXTS = [
    "o modelo cooperativo cnn-bigru combina fluxos paralelos",
    "a atencao cruzada troca informacoes entre redes a e b",
    "gru bidirecional captura dependencias temporais em ambas direcoes",
    "a porta de atenuacao controla o fluxo de informacao entre celulas",
    "a camada anti-alucinacao usa logica fuzzy de lukasiewicz",
    "o verificador classifica passos com sigmoid binaria",
    "penalidades linear e exponencial reduzem o erro grave",
    "ajuste dinamico de lr estabiliza o treinamento cooperativo",
    "normalizacao espectral limita a norma dos pesos",
    "inicializacao ortogonal estabiliza matrizes recorrentes",
    "o gerador decodificador usa atencao bahdanau sobre o encoder",
    "a fusao multimodal combina texto imagem e audio",
    "hipoteses sao ativadas quando o verificador pune o passo",
    "synergy search tenta n configuracoes e escolhe a melhor",
    "a perplexidade mede a confusao do modelo na previsao",
    "top-k e top-p filtram a distribuicao de probabilidade",
    "temperature ajusta a entropia das previsoes do modelo",
    "presence penalty pune tokens ja aparecidos na geracao",
    "frequency penalty pune proporcionalmente a frequencia do token",
    "streaming dataset carrega amostras sem materializacao completa",
    "byte bpe tokeniza qualquer string utf-8 sem unk",
    "o otimizador adamw combina momentum e weight decay",
    "gradient clipping previne explosao de gradiente",
    "amp reduz vram com precisao mista bfloat16",
    "memory optimizer limpa cache entre batches",
] * 4  # ~100 amostras para treino BBPE


def step_01_init_runtime() -> Dict:
    """Inicializa runtime e reporta configurações."""
    logger.info("=" * 70)
    logger.info("PASSO 1: Inicialização do runtime Xeon + memory optimizer")
    logger.info("=" * 70)
    from cnn_bigru.utils.xeon_runtime import get_runtime_info
    info = get_runtime_info()
    for k, v in info.items():
        logger.info("  %s: %s", k, v)
    mem = MemoryOptimizer(enable_amp=False)
    mem.configure()
    mem_info = mem.get_memory_mb()
    logger.info("  Memória: %s", mem_info)
    return {"runtime": info, "memory": mem_info}


def step_02_train_tokenizer() -> BBPETokenizer:
    """Treina BBPE tokenizer em corpus sintético."""
    logger.info("=" * 70)
    logger.info("PASSO 2: Treino do BBPE tokenizer")
    logger.info("=" * 70)
    t0 = time.time()
    tok = BBPETokenizer.train_from_texts(
        CORPUS_TEXTS,
        vocab_size=2000,
        min_frequency=1,
    )
    elapsed = time.time() - t0
    logger.info("  Vocab size: %d", tok.vocab_size)
    logger.info("  BOS/PAD/EOS IDs: %d/%d/%d", tok.bos_id, tok.pad_id, tok.eos_id)
    logger.info("  Tempo: %.2fs", elapsed)

    # Validação roundtrip
    test_strs = CORPUS_TEXTS[:5]
    roundtrip = tok.validate_roundtrip(test_strs)
    logger.info("  Roundtrip accuracy: %.2f%%", roundtrip * 100)

    if roundtrip < 0.8:
        logger.warning("  Roundtrip baixo — possível problema no tokenizer")
    return tok, roundtrip


def step_03_create_dataset(tokenizer: BBPETokenizer, n_samples: int = 50) -> DataLoader:
    """Cria dataset streaming multimodal com N amostras."""
    logger.info("=" * 70)
    logger.info("PASSO 3: Criação do dataset streaming multimodal (%d amostras)", n_samples)
    logger.info("=" * 70)

    # Para teste determinístico, usamos fallback sintético
    dataset = MultimodalStreamingDataset(
        n_samples=n_samples,
        hf_datasets=[],  # skip HF forçando sintético
        use_synthetic_fallback=True,
        seed=42,
        image_size=(28, 28, 1),
        audio_shape=(32, 40),
    )

    # Coleta todas as amostras em uma lista (para DataLoader iterável)
    samples = list(dataset)
    logger.info("  Amostras coletadas: %d", len(samples))
    assert len(samples) == n_samples, f"Esperado {n_samples}, obtido {len(samples)}"

    # Cria DataLoader com collate_fn
    from torch.utils.data import DataLoader
    loader = DataLoader(
        samples,
        batch_size=8,
        shuffle=False,
        collate_fn=lambda b: collate_multimodal(b, tokenizer, max_len=32),
    )
    logger.info("  DataLoader criado: batch_size=8, max_len=32")
    return loader


def step_04_init_model(tokenizer: BBPETokenizer, device: str = "cpu") -> Dict:
    """Instancia todos os componentes do modelo."""
    logger.info("=" * 70)
    logger.info("PASSO 4: Instanciação dos modelos")
    logger.info("=" * 70)
    V = tokenizer.vocab_size

    # Modelo multimodal principal
    model = MultimodalCNNBiGRU(
        vocab_size=V,
        num_classes=3,
        embedding_dim=32,
        cnn_filters=32,
        gru_hidden=32,
        n_heads=4,
        dropout=0.1,
        pad_idx=tokenizer.pad_id,
        img_channels=1,
        img_hidden=16,
        img_out_dim=32,
        audio_freq=40,
        audio_hidden=16,
        audio_out_dim=32,
        fusion_dim=64,
        use_spectral_norm=False,
    )

    # Aplica inicialização ortogonal
    n_init = orthogonal_init_model(model)
    logger.info("  Modelo multimodal: %d params, %d camadas ortogonalizadas",
                sum(p.numel() for p in model.parameters()), n_init)

    # Generator (encoder + decoder)
    generator = GeneratorCNNBiGRU(
        vocab_size=V,
        embedding_dim=32,
        cnn_filters=32,
        gru_hidden=32,
        n_heads=4,
        dropout=0.1,
        pad_idx=tokenizer.pad_id,
        max_proof_len=16,
    )
    orthogonal_init_model(generator)

    # Verifier
    verifier = VerifierCNNBiGRU(
        vocab_size=V,
        embedding_dim=32,
        cnn_filters=32,
        gru_hidden=32,
        n_heads=4,
        dropout=0.1,
        pad_idx=tokenizer.pad_id,
    )
    orthogonal_init_model(verifier)

    # Anti-hallucination
    anti_hall = AntiHallucinationLayer(vocab_size=V, embed_dim=16)

    logger.info("  Generator params: %d", sum(p.numel() for p in generator.parameters()))
    logger.info("  Verifier params: %d", sum(p.numel() for p in verifier.parameters()))
    logger.info("  Anti-hall params: %d", sum(p.numel() for p in anti_hall.parameters()))

    return {
        "model": model,
        "generator": generator,
        "verifier": verifier,
        "anti_hall": anti_hall,
    }


def step_05_train(
    models: Dict,
    tokenizer: BBPETokenizer,
    dataloader: DataLoader,
    device: str = "cpu",
    ewc_state: EWCState = None,
) -> Dict:
    """Executa treinamento cooperativo.

    NOVO v2.0: aceita ewc_state opcional para testar EWC.
    """
    logger.info("=" * 70)
    logger.info("PASSO 5: Treinamento cooperativo (synergy + hipóteses + auto-learn)")
    logger.info("=" * 70)

    cfg = TrainerConfig(
        num_epochs=2,
        max_batches_per_epoch=6,
        batch_size=8,
        n_synergy_attempts=3,
        use_synergy_search=True,
        use_hypotheses=True,
        n_hypotheses=4,
        loss_config=LossConfig(
            alpha=1.0, beta=0.5, gamma_loss=0.3, delta=0.01,
            lambda_penal=0.1, mu_exp_penal=0.05,
            gamma_exp=1.0, threshold_err=0.5,
            l2_reg=1e-5, use_curvature=True, curvature_eps=1e-3,
        ),
        auto_config=AutoLearnConfig(
            kappa_curv=0.01, grad_clip=1.0, spectral_radius=1.0,
            lr_min=1e-6, lr_max=1e-2,
            initial_lr_G=1e-3, initial_lr_V=1e-3,
            use_spectral_norm=True, apply_after_step=True,
            l2_reg=1e-5,
        ),
        ewc_config=ewc_state.config if ewc_state else None,
        device=device,
        log_every=1,
        use_verifier_real=True,  # NOVO v2.0: usar verificador real
        use_generator=True,
        use_hypothesis_output=True,
    )

    trainer = CooperativeTrainer(
        model=models["model"],
        tokenizer=tokenizer,
        config=cfg,
        generator=models["generator"],
        verifier=models["verifier"],
        anti_hallucination=models["anti_hall"],
        ewc_state=ewc_state,
    )

    try:
        result = trainer.train(dataloader)
        logger.info("  Treinamento OK | final_loss=%.4f | final_ppl=%.2f | elapsed=%.1fs",
                    result["final_loss"], result["final_ppl"], result["elapsed_s"])
        return result
    except Exception as e:
        logger.error("  Treinamento FALHOU: %s", e)
        logger.error(traceback.format_exc())
        raise


def step_05b_test_new_modules(
    models: Dict,
    tokenizer: BBPETokenizer,
    device: str = "cpu",
) -> Dict:
    """Testa os novos módulos v2.0: EWC, Context Window, RoPE, TransformerBlock.

    NOVO v2.0: esta função exercita isoladamente cada novo módulo para garantir
    que estão funcionando e integráveis ao pipeline.
    """
    logger.info("=" * 70)
    logger.info("PASSO 5b: Teste dos novos módulos v2.0 (EWC, ContextWindow, RoPE, TransformerBlock)")
    logger.info("=" * 70)

    results = {"ewc": None, "context_window": None, "rope": None, "transformer": None}

    # ---------- EWC ----------
    try:
        logger.info("  [EWC] Testando EWCState...")
        ewc_cfg = EWCConfig(
            enabled=True,
            lambda_ewc=100.0,
            n_samples_fisher=5,  # poucas amostras para teste rápido
            online_gamma=0.9,
            device=device,
        )
        ewc_state = EWCState(ewc_cfg)
        # Antes de consolidar: penalty deve ser 0
        pen0 = ewc_state.penalty(models["model"])
        logger.info("    Penalty antes de consolidar: %.6f", float(pen0))
        assert pen0.item() == 0.0, "EWC penalty deveria ser 0 antes de consolidar"

        # Simular forward_fn para cálculo de Fisher
        # IMPORTANTE: NÃO usar torch.no_grad() aqui — a EWC precisa de gradientes
        # para calcular a matriz de Fisher (grad^2 da log-verossimilhança)
        V = tokenizer.vocab_size
        B = 2
        T = 8
        ids_a = torch.randint(0, V, (B, T), device=device)
        ids_b = torch.randint(0, V, (B, T), device=device)

        def forward_fn(_idx=None):
            # Sem torch.no_grad() — a EWC interna ativa enable_grad()
            out = models["model"](ids_a, ids_b, mode="classify")
            return out["logits"]

        # Consolidar (calcula Fisher e armazena theta_star)
        ewc_state.consolidate(models["model"], forward_fn=forward_fn)
        assert ewc_state.num_tasks() == 1, f"Esperado 1 tarefa, obtido {ewc_state.num_tasks()}"

        # Penalty deve ser > 0 agora (parâmetros não mudaram desde consolidate, mas
        # a penalidade ainda assim é computada e deve ser >= 0)
        pen1 = ewc_state.penalty(models["model"])
        logger.info("    Penalty após consolidar: %.6f", float(pen1.detach()))
        assert pen1.item() >= 0.0, "EWC penalty deve ser >= 0"

        results["ewc"] = {
            "ok": True,
            "penalty_before": float(pen0),
            "penalty_after": float(pen1),
            "num_tasks": ewc_state.num_tasks(),
        }
        logger.info("  [EWC] ✓ OK")

        # Guardar para uso posterior no trainer
        results["ewc_state"] = ewc_state
    except Exception as e:
        logger.error("  [EWC] ✗ FALHOU: %s", e)
        logger.error(traceback.format_exc())
        results["ewc"] = {"ok": False, "error": str(e)}

    # ---------- Context Window ----------
    try:
        logger.info("  [ContextWindow] Testando ContextWindowManager + KVCache...")
        cw = make_default_context_window(
            max_window=32,
            n_sink=2,
            embed_dim=64,
            n_heads=4,
            n_layers=2,
            device=device,
        )
        # Inicializar cache
        cache = cw.init_cache(batch_size=1, device=torch.device(device))
        assert cache is not None
        assert cache.get_seq_len() == 0

        # Append tokens
        tokens1 = torch.tensor([[1, 2, 3, 4, 5]], dtype=torch.long, device=device)
        window = cw.append_tokens(tokens1)
        assert window.size(1) == 5
        assert cw.kv_cache.get_seq_len() == 0  # cache só é populado quando chamamos KVCache.update

        # Simular update do cache KV
        dummy_k = torch.randn(1, 4, 5, 16, device=device)  # [B, n_heads, T, head_dim]
        dummy_v = torch.randn(1, 4, 5, 16, device=device)
        new_k, new_v = cache.update(0, dummy_k, dummy_v)
        assert cache.get_seq_len() == 5

        # Adicionar mais tokens para forçar eviction
        tokens2 = torch.tensor([[6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20]],
                               dtype=torch.long, device=device)
        cw.append_tokens(tokens2)
        # Total histórico: 5 + 15 = 20, mas max_window=32, então sem eviction ainda
        assert cw.token_history and sum(t.size(1) for t in cw.token_history) == 20

        # Forçar eviction adicionando mais tokens
        tokens3 = torch.tensor([[21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34]],
                               dtype=torch.long, device=device)
        cw.append_tokens(tokens3)  # total = 34 > 32 = max_window
        cw.evict_cache()

        info = cw.get_info()
        logger.info("    Context window info: %s", info)
        results["context_window"] = {"ok": True, "info": info}
        logger.info("  [ContextWindow] ✓ OK")
    except Exception as e:
        logger.error("  [ContextWindow] ✗ FALHOU: %s", e)
        logger.error(traceback.format_exc())
        results["context_window"] = {"ok": False, "error": str(e)}

    # ---------- RoPE ----------
    try:
        logger.info("  [RoPE] Testando RotaryPositionEmbedding...")
        head_dim = 32
        max_seq = 16
        rope = RotaryPositionEmbedding(head_dim=head_dim, max_seq_len=max_seq).to(device)
        # x: [B, n_heads, T, head_dim]
        x = torch.randn(2, 4, 8, head_dim, device=device)
        x_rot = rope(x)
        assert x_rot.shape == x.shape, f"Shape mismatch: {x_rot.shape} vs {x.shape}"
        # Verificar que RoPE preserva a norma (é uma rotação)
        norm_before = x.norm(dim=-1)
        norm_after = x_rot.norm(dim=-1)
        diff = (norm_before - norm_after).abs().max().item()
        logger.info("    Norma preservada (diff=%.6f)", diff)
        assert diff < 1e-4, f"RoPE não preservou norma (diff={diff})"

        # Testar com offset (para geração autoregressiva)
        x_rot2 = rope(x, offset=10)
        assert x_rot2.shape == x.shape

        results["rope"] = {"ok": True, "norm_diff": diff}
        logger.info("  [RoPE] ✓ OK")
    except Exception as e:
        logger.error("  [RoPE] ✗ FALHOU: %s", e)
        logger.error(traceback.format_exc())
        results["rope"] = {"ok": False, "error": str(e)}

    # ---------- TransformerBlock ----------
    try:
        logger.info("  [TransformerBlock] Testando CausalSelfAttention + TransformerBlock...")
        embed_dim = 64
        n_heads = 4
        n_layers = 2
        V = tokenizer.vocab_size

        block_config = TransformerBlockConfig(
            embed_dim=embed_dim,
            n_heads=n_heads,
            ff_dim=4 * embed_dim,
            dropout=0.1,
            use_rope=True,
            max_seq_len=32,
        )
        block = TransformerBlock(block_config).to(device)

        # Forward sem cache
        x = torch.randn(2, 8, embed_dim, device=device)
        out, k, v = block(x)
        assert out.shape == x.shape, f"Output shape: {out.shape} vs {x.shape}"
        assert k.shape == (2, n_heads, 8, embed_dim // n_heads)
        assert v.shape == (2, n_heads, 8, embed_dim // n_heads)

        # Forward com padding mask
        mask = torch.tensor([[1, 1, 1, 1, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 0]],
                            dtype=torch.float, device=device)
        out2, k2, v2 = block(x, padding_mask=mask)
        assert out2.shape == x.shape

        # Forward com cache KV
        cache_k = None
        cache_v = None
        for t in range(3):
            x_t = torch.randn(2, 1, embed_dim, device=device)
            out_t, cache_k, cache_v = block(
                x_t, kv_cache_k=cache_k, kv_cache_v=cache_v, position_offset=t
            )
            assert out_t.shape == (2, 1, embed_dim)
            assert cache_k.size(2) == t + 1, f"Cache len: {cache_k.size(2)}, esperado {t+1}"
        logger.info("    Cache KV após 3 steps: %d tokens", cache_k.size(2))

        # Testar TransformerDecoderStack completo
        decoder = TransformerDecoderStack(
            vocab_size=V,
            embed_dim=embed_dim,
            n_heads=n_heads,
            n_layers=n_layers,
            max_seq_len=32,
            use_rope=True,
            pad_id=tokenizer.pad_id,
            weight_tying=True,
        ).to(device)

        idx = torch.randint(0, V, (2, 8), device=device)
        logits, new_caches = decoder(idx)
        assert logits.shape == (2, 8, V), f"Logits shape: {logits.shape}"
        assert len(new_caches) == n_layers

        # Verificar weight tying
        assert decoder.lm_head.weight is decoder.token_embedding.weight

        results["transformer"] = {
            "ok": True,
            "block_output_shape": list(out.shape),
            "decoder_logits_shape": list(logits.shape),
            "weight_tying": True,
        }
        logger.info("  [TransformerBlock] ✓ OK")
    except Exception as e:
        logger.error("  [TransformerBlock] ✗ FALHOU: %s", e)
        logger.error(traceback.format_exc())
        results["transformer"] = {"ok": False, "error": str(e)}

    return results


def step_06_inference(
    model: MultimodalCNNBiGRU,
    tokenizer: BBPETokenizer,
    device: str = "cpu",
    generator: GeneratorCNNBiGRU = None,
    use_context_window: bool = True,
) -> Dict:
    """Executa inferência com sampling.

    NOVO v2.0: usa generator (se disponível) para geração autoregressiva REAL,
    e integra context_window para suporte a sequências longas.
    """
    logger.info("=" * 70)
    logger.info("PASSO 6: Inferência com Temperatura + Top-K + Top-P + Penalidades")
    if generator is not None:
        logger.info("  (usando GeneratorCNNBiGRU para geração autoregressiva REAL)")
    else:
        logger.info("  (sem generator — usando fallback single-step)")
    logger.info("=" * 70)

    prompts = [
        ("o modelo coopera entre", "fluxos paralelos"),
        ("atencao cruzada troca", "informacoes entre redes"),
        ("gru bidirecional captura", "dependencias temporais"),
    ]

    # Inicializar context window se solicitado
    cw = None
    if use_context_window:
        try:
            cw = make_default_context_window(
                max_window=64,
                n_sink=2,
                embed_dim=32,
                n_heads=4,
                n_layers=2,
                device=device,
            )
            logger.info("  Context window ativo (max_window=64, sink=2)")
        except Exception as e:
            logger.warning("  Context window falhou ao inicializar: %s — desativando", e)
            cw = None

    results = []
    for prompt_a, prompt_b in prompts:
        try:
            result = generate_with_sampling(
                model=model,
                tokenizer=tokenizer,
                prompt_a=prompt_a,
                prompt_b=prompt_b,
                max_new_tokens=16,
                temperature=0.7,
                top_k=20,
                top_p=0.9,
                presence_penalty=0.3,
                frequency_penalty=0.3,
                device=device,
                generator=generator,
                context_window=cw,
            )
            logger.info("  Prompt A: %s", prompt_a)
            logger.info("  Prompt B: %s", prompt_b)
            logger.info("  Gerado: %s", result["text"][:100])
            logger.info("  Metrics: %s", result["metrics"])
            results.append({"prompt_a": prompt_a, "prompt_b": prompt_b, **result})
        except Exception as e:
            logger.error("  Inferência falhou para (%s, %s): %s", prompt_a, prompt_b, e)
            logger.error(traceback.format_exc())
            results.append({"prompt_a": prompt_a, "prompt_b": prompt_b, "error": str(e)})

    return {"results": results}


def step_07_eval_ppl(
    model: MultimodalCNNBiGRU,
    tokenizer: BBPETokenizer,
    dataloader: DataLoader,
    device: str = "cpu",
    generator: GeneratorCNNBiGRU = None,
) -> Dict:
    """Avalia perplexidade.

    NOVO v2.0: usa generator (se disponível) para PPL real baseado em
    geração autoregressiva com teacher forcing.
    """
    logger.info("=" * 70)
    logger.info("PASSO 7: Avaliação de Perplexidade (PPL)")
    if generator is not None:
        logger.info("  (usando GeneratorCNNBiGRU com teacher forcing)")
    else:
        logger.info("  (sem generator — usando fallback single-step)")
    logger.info("=" * 70)
    try:
        result = evaluate_perplexity(
            model=model,
            dataloader=dataloader,
            tokenizer=tokenizer,
            device=device,
            max_batches=5,
            generator=generator,
        )
        logger.info("  Loss: %.4f | PPL: %.2f | batches: %d | used_generator: %s",
                    result["loss"], result["ppl"], result["n_batches"],
                    result.get("used_generator", False))
        return result
    except Exception as e:
        logger.error("  PPL falhou: %s", e)
        logger.error(traceback.format_exc())
        return {"error": str(e)}


def step_08_report_errors(errors: List[str], warnings: List[str]) -> None:
    """Reporta erros lógicos ou falhas encontradas."""
    logger.info("=" * 70)
    logger.info("PASSO 8: Relatório de erros lógicos ou falhas")
    logger.info("=" * 70)

    if not errors:
        logger.info("  ✓ NENHUM ERRO CRÍTICO encontrado")
    else:
        logger.error("  ✗ %d ERRO(S) CRÍTICO(S) encontrado(s):", len(errors))
        for e in errors:
            logger.error("    - %s", e)

    if not warnings:
        logger.info("  ✓ Nenhum warning relevante")
    else:
        logger.warning("  ⚠ %d warning(s):", len(warnings))
        for w in warnings:
            logger.warning("    - %s", w)

    logger.info("=" * 70)


def main():
    """Executa o teste completo de 50 amostras."""
    logger.info("INICIANDO TESTE DE 50 AMOSTRAS — CNN-BiGRU MULTIMODAL COOPERATIVO")
    logger.info("Project root: %s", PROJECT_ROOT)
    logger.info("Device: %s", "cpu")
    logger.info("Cores: %d", N_CORES)

    errors: List[str] = []
    warnings: List[str] = []

    try:
        # Step 1
        step_01_init_runtime()

        # Step 2
        tokenizer, roundtrip = step_02_train_tokenizer()
        if roundtrip < 0.8:
            warnings.append(f"BBPE roundtrip accuracy = {roundtrip:.2%} (esperado >= 80%)")

        # Step 3
        dataloader = step_03_create_dataset(tokenizer, n_samples=50)

        # Step 4
        models = step_04_init_model(tokenizer, device="cpu")

        # Verificação: forward pass simples antes do treino
        try:
            batch = next(iter(dataloader))
            with torch.no_grad():
                out = models["model"](
                    batch["input_ids_a"], batch["input_ids_b"],
                    images=batch["images"], audios=batch["audios"],
                    mode="classify",
                )
            logger.info("  Forward pass OK | logits shape: %s", out["logits"].shape)
            assert out["logits"].shape == (batch["input_ids_a"].size(0), 3), \
                f"Shape inesperado: {out['logits'].shape}"
        except Exception as e:
            errors.append(f"Forward pass inicial falhou: {e}")
            logger.error(traceback.format_exc())

        # Step 5b: Testar novos módulos v2.0 (EWC, Context Window, RoPE, TransformerBlock)
        ewc_state = None
        if not errors:
            try:
                new_modules_result = step_05b_test_new_modules(models, tokenizer, device="cpu")
                # Verificar se algum módulo falhou
                for mod_name, mod_res in new_modules_result.items():
                    if mod_name == "ewc_state":
                        continue
                    if isinstance(mod_res, dict) and not mod_res.get("ok", True):
                        errors.append(f"Módulo {mod_name} falhou: {mod_res.get('error', 'unknown')}")
                # Guardar EWC state para usar no treino
                ewc_state = new_modules_result.get("ewc_state")
                if ewc_state is None:
                    warnings.append("EWC state não foi criado — EWC não será testado no treino")
            except Exception as e:
                errors.append(f"Teste de novos módulos falhou: {e}")
                logger.error(traceback.format_exc())

        # Step 5: Treinamento (com EWC se disponível)
        if not errors:
            try:
                train_result = step_05_train(models, tokenizer, dataloader, device="cpu",
                                             ewc_state=ewc_state)
            except Exception as e:
                errors.append(f"Treinamento falhou: {e}")
                train_result = None
        else:
            train_result = None

        # Step 6: Inferência com generator + context window
        if not errors:
            try:
                step_06_inference(
                    models["model"], tokenizer, device="cpu",
                    generator=models.get("generator"),
                    use_context_window=True,
                )
            except Exception as e:
                errors.append(f"Inferência falhou: {e}")
                logger.error(traceback.format_exc())

        # Step 7: PPL com generator
        if not errors:
            try:
                step_07_eval_ppl(
                    models["model"], tokenizer, dataloader, device="cpu",
                    generator=models.get("generator"),
                )
            except Exception as e:
                warnings.append(f"PPL falhou (não crítico): {e}")

        # Step 8
        step_08_report_errors(errors, warnings)

        # Resumo final
        logger.info("=" * 70)
        logger.info("RESUMO FINAL DO TESTE DE 50 AMOSTRAS")
        logger.info("=" * 70)
        logger.info("  Amostras processadas: 50")
        logger.info("  Erros críticos: %d", len(errors))
        logger.info("  Warnings: %d", len(warnings))
        if train_result:
            logger.info("  Loss final: %.4f", train_result["final_loss"])
            logger.info("  PPL final: %.2f", train_result["final_ppl"])
            logger.info("  Tempo: %.1fs", train_result["elapsed_s"])
            logger.info("  Synergy attempts: %d", len(train_result.get("synergy_history", [])))
            logger.info("  Hypothesis activations: %d",
                        sum(h.get("hypothesis_activations", 0) for h in train_result.get("history", [])))

        if errors:
            logger.error("  STATUS: FALHA — %d erro(s)", len(errors))
            return 1
        else:
            logger.info("  STATUS: SUCESSO")
            return 0

    except Exception as e:
        logger.error("ERRO FATAL: %s", e)
        logger.error(traceback.format_exc())
        return 2


if __name__ == "__main__":
    try:
        rc = main()
    except SystemExit:
        raise
    except Exception as e:
        logger.error("Unhandled exception: %s", e)
        rc = 2
    # Evita o "Fatal Python error: PyGILState_Release" no shutdown causado
    # por threads do tokenizers/datasets que ainda estão ativas.
    import os
    os._exit(rc)