def _build_views(self) -> list[str]:
    # Declarative virtual views mounted directly over zero-copy Parquet files
    static_views = {
        self.visual_analytics_dir / "va_comm_threads.parquet":
            "CREATE OR REPLACE VIEW va_comm_threads AS SELECT * FROM read_parquet('{basePath}/visual_analytics/va_comm_threads.parquet');",
        self.visual_analytics_dir / "va_cluster_hierarchy.parquet":
            "CREATE OR REPLACE VIEW va_cluster_hierarchy AS SELECT * FROM read_parquet('{basePath}/visual_analytics/va_cluster_hierarchy.parquet');",
        self.visual_analytics_dir / "va_document_registry.parquet":
            "CREATE OR REPLACE VIEW va_document_registry AS SELECT * FROM read_parquet('{basePath}/visual_analytics/va_document_registry.parquet');",
    }
    for file_path, view_sql in static_views.items():
        if file_path.exists():
            views.append(view_sql)
    return views

def to_arrow_fixed_size_list(embeddings: np.ndarray, dim: int = 256) -> pa.FixedSizeListArray:
    """Zero-copy contiguous C-buffer wrap into Arrow FixedSizeListArray without .tolist() heap unpacking."""
    contiguous = np.ascontiguousarray(embeddings, dtype=np.float32)
    flat_buffer = pa.py_buffer(contiguous.ravel())
    value_array = pa.Array.from_buffers(pa.float32(), contiguous.size, [None, flat_buffer])
    return pa.FixedSizeListArray.from_arrays(value_array, dim=dim)

def compute_mps_auc_roc(pos_scores: torch.Tensor, neg_scores: torch.Tensor) -> float:
    """Compute AUC-ROC natively on the MPS device with zero CPU synchronization."""
    scores = torch.cat([pos_scores, neg_scores])
    labels = torch.cat([torch.ones_like(pos_scores), torch.zeros_like(neg_scores)])

    # Vector sorting and cumulative trapezoidal summation run edge-to-edge on Apple Silicon
    sorted_indices = torch.argsort(scores, descending=True)
    sorted_labels = labels[sorted_indices]

    tps = torch.cumsum(sorted_labels, dim=0)
    fps = torch.cumsum(1.0 - sorted_labels, dim=0)

    tpr = tps / pos_scores.size(0)
    fpr = fps / neg_scores.size(0)

    return float(torch.sum((fpr[1:] - fpr[:-1]) * tpr[:-1]).item())

def elevate_process_qos() -> bool:
    """Elevate worker QoS from background throttling to QOS_CLASS_USER_INTERACTIVE (0x21)."""
    if platform.system() == "Darwin":
        libc = ctypes.CDLL("libSystem.B.dylib")
        # QOS_CLASS_USER_INTERACTIVE = 0x21 pins execution to Apple Silicon Performance Cores
        ret = libc.pthread_set_qos_class_self_np(0x21, 0)
        os.setpriority(os.PRIO_PROCESS, os.getpid(), 0)
        return ret == 0
    return False