foxhunt/ml/src/lib.rs

//! Machine Learning Models for Foxhunt
//!
//! This crate provides comprehensive machine learning models and algorithms
//! for the Foxhunt high-frequency trading system. All ML operations use
//! enterprise-grade safety controls to prevent system failures.
//!
//! ## Safety Features
//!
//! - **Comprehensive mathematical safety**: All operations handle NaN/Infinity gracefully
//! - **Tensor bounds checking**: Prevents buffer overflows and memory issues
//! - **Model drift detection**: Automatic monitoring of model performance degradation
//! - **Financial validation**: Ensures all predictions use unified financial types
//! - **Memory management**: Prevents OOM conditions and memory leaks
//! - **Timeout handling**: Prevents hanging operations
//!
//! ## Usage
//!
//! ```rust
//! use ml_models::safety::{get_global_safety_manager, MLSafetyConfig};
//!
//! // Initialize safety with custom configuration
//! let config = MLSafetyConfig::default();
//! let safety_manager = get_global_safety_manager();
//!
//! // All ML operations should go through the safety manager
//! let result = safety_manager.safe_math_operation("prediction", || {
//!     // Your ML computation here
//!     Ok(42.0)
//! }).await?;
//! ```

#![warn(missing_docs)]
#![warn(missing_debug_implementations)]
#![warn(rust_2018_idioms)]
#![warn(missing_docs)]
#![warn(missing_debug_implementations)]
#![warn(rust_2018_idioms)]
#![deny(
    clippy::unwrap_used,
    clippy::expect_used,
    clippy::panic,
    clippy::unimplemented,
    clippy::unreachable,
    clippy::indexing_slicing
)]

// Import common types properly - NO ALIASES THAT CONFLICT!
use serde::{Deserialize, Serialize};
use candle_core::Tensor;
use candle_nn::Optimizer; // For Adam optimizer support
use candle_core::Var; // For tensor variables


// Note: Optimizer trait not available in candle_optimisers v0.9
// Files using optimizers may need to be updated or removed

// Note: For candle_nn types like Linear and Dropout, they implement Module trait
// Use Module::forward(&self, input) instead of self.forward(input)

/// Wrapper for Adam optimizer to provide required methods
pub struct Adam {
    optimizer: candle_optimisers::adam::Adam,
    learning_rate: f64,
}

impl Adam {
    pub fn new(vars: Vec<Var>, params: candle_optimisers::adam::ParamsAdam) -> Result<Self, MLError> {
        let learning_rate = params.lr;
        let optimizer = candle_optimisers::adam::Adam::new(vars, params)
            .map_err(|e| MLError::TrainingError(format!("Failed to create Adam optimizer: {}", e)))?;

        Ok(Self {
            optimizer,
            learning_rate,
        })
    }

    pub fn backward_step(&mut self, loss: &Tensor) -> Result<(), MLError> {
        // Calculate gradients
        let grads = loss.backward()
            .map_err(|e| MLError::TrainingError(format!("Backward pass failed: {}", e)))?;

        // Apply optimizer step using trait method
        Optimizer::step(&mut self.optimizer, &grads)
            .map_err(|e| MLError::TrainingError(format!("Optimizer step failed: {}", e)))?;

        Ok(())
    }

    pub fn learning_rate(&self) -> f64 {
        self.learning_rate
    }
}

// Direct type imports - no compatibility aliases
use rust_decimal::Decimal;


#[derive(Debug, Clone, thiserror::Error, serde::Serialize, serde::Deserialize)]
pub enum CommonTypeError {
    #[error("Type error: {0}")]
    Error(String),
}

#[derive(Debug, Clone, Copy, PartialEq, Eq, serde::Serialize, serde::Deserialize)]
pub enum MarketRegime {
    Normal,
    Trending,
    Sideways,
    Bull,
    Bear,
    Crisis,
}

#[derive(Debug, Clone, thiserror::Error, serde::Serialize, serde::Deserialize)]
pub enum CommonError {
    #[error("Error: {0}")]
    General(String),
}

impl CommonError {
    pub fn validation(msg: impl Into<String>) -> Self {
        Self::General(msg.into())
    }
    pub fn config(msg: impl Into<String>) -> Self {
        Self::General(msg.into())
    }
    pub fn service(category: ErrorCategory, msg: impl Into<String>) -> Self {
        Self::General(format!("{:?}: {}", category, msg.into()))
    }
}

#[derive(Debug, Clone, Copy, PartialEq, Eq, serde::Serialize, serde::Deserialize)]
pub enum ErrorCategory {
    System,
}

// Now using real types from common crate

// Core ML types
#[derive(Debug, Clone, Serialize, Deserialize)]
#[cfg_attr(feature = "database", derive(sqlx::FromRow))]
pub struct Trade {
    pub symbol: String,
    pub price: Decimal,
    pub quantity: Decimal,
    pub timestamp: u64,
    pub side: String,
}

// Health status for ensemble models
#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
pub enum HealthStatus {
    Healthy,
    Degraded,
    Unhealthy,
}

// Import specific types from trading_engine that we need
// (removed wildcard prelude to avoid conflicts)

// Using Decimal for financial types

// Placeholder types for compilation - should be imported from appropriate crates in production
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct MarketDataSnapshot {
    pub timestamp: chrono::DateTime<chrono::Utc>,
    pub symbol: String,
    pub price: Decimal,
    pub volume: Decimal,
}

#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct FeatureVector(pub Vec<f64>);

#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct IntegerTensor(pub Vec<i64>);

#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct UpdateSummary {
    pub updated_models: usize,
    pub total_models: usize,
}

use thiserror::Error;

/// Machine Learning specific errors
#[derive(Debug, Clone, Error, Serialize, Deserialize)]
pub enum MLError {
    /// Configuration error
    #[error("Configuration error: {reason}")]
    ConfigError { reason: String },

    /// Configuration error (alternative naming)
    #[error("Configuration error: {0}")]
    ConfigurationError(String),

    /// Dimension mismatch error
    #[error("Dimension mismatch: expected {expected}, got {actual}")]
    DimensionMismatch { expected: usize, actual: usize },

    /// Graph-related error
    #[error("Graph error: {message}")]
    GraphError { message: String },

    /// Resource limit exceeded
    #[error("Resource limit exceeded: {resource} limit {limit}")]
    ResourceLimit { resource: String, limit: usize },

    /// Serialization error
    #[error("Serialization error: {reason}")]
    SerializationError { reason: String },

    /// Validation error
    #[error("Validation error: {message}")]
    ValidationError { message: String },

    /// Concurrency error
    #[error("Concurrency error in operation: {operation}")]
    ConcurrencyError { operation: String },

    /// Invalid input error
    #[error("Invalid input: {0}")]
    InvalidInput(String),

    /// Training error
    #[error("Training error: {0}")]
    TrainingError(String),

    /// Inference error
    #[error("Inference error: {0}")]
    InferenceError(String),

    /// Model error
    #[error("Model error: {0}")]
    ModelError(String),

    /// Model not trained error
    #[error("Model not trained: {0}")]
    NotTrained(String),

    /// Anyhow error wrapping
    #[error("General error: {0}")]
    AnyhowError(String),

    /// Tensor creation error
    #[error("Tensor creation error in {operation}: {reason}")]
    TensorCreationError { operation: String, reason: String },

    /// Lock error
    #[error("Lock error: {0}")]
    LockError(String),

    /// Model not found error
    #[error("Model not found: {0}")]
    ModelNotFound(String),

    /// Insufficient data error
    #[error("Insufficient data: {0}")]
    InsufficientData(String),
}

// Implement From trait for candle_core::Error
impl From<candle_core::Error> for MLError {
    fn from(err: candle_core::Error) -> Self {
        MLError::ModelError(format!("Candle error: {}", err))
    }
}

// NOTE: Commented out workspace dependency - will be re-enabled when workspace is available
// impl From<error_handling::TradingError> for MLError {
//     fn from(err: error_handling::TradingError) -> Self {
//         match err {
//             error_handling::TradingError::InvalidPrice { value, reason } => {
//                 MLError::ValidationError {
//                     message: format!("Invalid price {}: {}", value, reason),
//                 }
//             }
//             error_handling::TradingError::InvalidQuantity { value, reason } => {
//                 MLError::ValidationError {
//                     message: format!("Invalid quantity {}: {}", value, reason),
//                 }
//             }
//             error_handling::TradingError::FinancialSafety { message, .. } => {
//                 MLError::ValidationError {
//                     message: format!("Financial safety error: {}", message),
//                 }
//             }
//             error_handling::TradingError::DivisionByZero { operation } => {
//                 MLError::ValidationError {
//                     message: format!("Division by zero in {}", operation),
//                 }
//             }
//             error_handling::TradingError::ModelInference { reason, model } => {
//                 MLError::InferenceError(format!("Model inference error for {}: {}", model, reason))
//             }
//             error_handling::TradingError::GpuComputation { reason, operation } => {
//                 let msg = match operation {
//                     Some(op) => format!("GPU computation error ({}): {}", op, reason),
//                     None => format!("GPU computation error: {}", reason),
//                 };
//                 MLError::ModelError(msg)
//             }
//             other => MLError::ModelError(format!("Trading error: {}", other)),
//         }
//     }
// }
// Implement From trait for anyhow::Error
impl From<anyhow::Error> for MLError {
    fn from(err: anyhow::Error) -> Self {
        MLError::AnyhowError(err.to_string())
    }
}

// UNIFIED ERROR HANDLING: Convert all ML errors to CommonError for workspace consistency
impl From<MLError> for CommonError {
    fn from(err: MLError) -> Self {
        match err {
            MLError::ConfigError { reason } => CommonError::config(
                format!("ML configuration error: {}", reason)
            ),
            MLError::ConfigurationError(msg) => CommonError::config(
                format!("ML configuration error: {}", msg)
            ),
            MLError::DimensionMismatch { expected, actual } => CommonError::validation(
                format!("ML dimension mismatch: expected {}, got {}", expected, actual)
            ),
            MLError::GraphError { message } => CommonError::service(
                ErrorCategory::System,
                format!("ML graph error: {}", message)
            ),
            MLError::ResourceLimit { resource, limit } => CommonError::service(
                ErrorCategory::System,
                format!("ML resource limit exceeded: {} limit {}", resource, limit)
            ),
            MLError::SerializationError { reason } => CommonError::service(
                ErrorCategory::System,
                format!("ML serialization error: {}", reason)
            ),
            MLError::ValidationError { message } => CommonError::validation(
                format!("ML validation error: {}", message)
            ),
            MLError::ConcurrencyError { operation } => CommonError::service(
                ErrorCategory::System,
                format!("ML concurrency error in operation: {}", operation)
            ),
            MLError::InvalidInput(msg) => CommonError::validation(
                format!("ML invalid input: {}", msg)
            ),
            MLError::TrainingError(msg) => CommonError::service(
                ErrorCategory::System,
                format!("ML training error: {}", msg)
            ),
            MLError::InferenceError(msg) => CommonError::service(
                ErrorCategory::System,
                format!("ML inference error: {}", msg)
            ),
            MLError::ModelError(msg) => CommonError::service(
                ErrorCategory::System,
                format!("ML model error: {}", msg)
            ),
            MLError::NotTrained(msg) => CommonError::service(
                ErrorCategory::System,
                format!("ML model not trained: {}", msg)
            ),
            MLError::AnyhowError(msg) => CommonError::service(
                ErrorCategory::System,
                format!("ML error: {}", msg)
            ),
            MLError::TensorCreationError { operation, reason } => CommonError::service(
                ErrorCategory::System,
                format!("ML tensor creation error in {}: {}", operation, reason)
            ),
            MLError::LockError(msg) => CommonError::service(
                ErrorCategory::System,
                format!("ML lock error: {}", msg)
            ),
            MLError::ModelNotFound(msg) => CommonError::service(
                ErrorCategory::System,
                format!("ML model not found: {}", msg)
            ),
            MLError::InsufficientData(msg) => CommonError::validation(
                format!("ML insufficient data: {}", msg)
            ),
        }
    }
}

// Convert common type errors to MLError
impl From<CommonTypeError> for MLError {
    fn from(err: CommonTypeError) -> Self {
        MLError::ModelError(format!("Common type error: {}", err))
    }
}

impl From<serde_json::Error> for MLError {
    fn from(err: serde_json::Error) -> Self {
        MLError::SerializationError {
            reason: err.to_string(),
        }
    }
}

impl From<inference::RealInferenceError> for MLError {
    fn from(err: inference::RealInferenceError) -> Self {
        match err {
            inference::RealInferenceError::GpuRequired { reason } => {
                MLError::ModelError(format!("GPU required: {}", reason))
            }
            inference::RealInferenceError::ComputationFailed { reason } => {
                MLError::InferenceError(reason)
            }
            inference::RealInferenceError::FeatureMismatch { expected, actual } => {
                MLError::DimensionMismatch { expected, actual }
            }
            inference::RealInferenceError::PredictionValidation { reason } => {
                MLError::ValidationError { message: reason }
            }
            inference::RealInferenceError::HardwareError { reason } => {
                MLError::ModelError(format!("Hardware error: {}", reason))
            }
            other => MLError::InferenceError(other.to_string()),
        }
    }
}

// Implement From<ProductionTrainingError> for MLError
impl From<training_pipeline::ProductionTrainingError> for MLError {
    fn from(err: training_pipeline::ProductionTrainingError) -> Self {
        match err {
            training_pipeline::ProductionTrainingError::ConfigError { reason } => {
                MLError::ConfigError { reason }
            }
            training_pipeline::ProductionTrainingError::ArchitectureError { reason } => {
                MLError::ModelError(format!("Architecture error: {}", reason))
            }
            training_pipeline::ProductionTrainingError::DataError { reason } => {
                MLError::ValidationError {
                    message: format!("Data error: {}", reason),
                }
            }
            training_pipeline::ProductionTrainingError::OptimizationError { reason } => {
                MLError::TrainingError(format!("Optimization error: {}", reason))
            }
            training_pipeline::ProductionTrainingError::FinancialError { reason } => {
                MLError::ValidationError {
                    message: format!("Financial error: {}", reason),
                }
            }
            training_pipeline::ProductionTrainingError::SafetyViolation { reason } => {
                MLError::ValidationError {
                    message: format!("Safety violation: {}", reason),
                }
            }
            training_pipeline::ProductionTrainingError::ConvergenceError { reason } => {
                MLError::TrainingError(format!("Convergence error: {}", reason))
            }
            training_pipeline::ProductionTrainingError::ResourceError { reason } => {
                MLError::ModelError(format!("Resource error: {}", reason))
            }
            training_pipeline::ProductionTrainingError::GpuRequired { reason } => {
                MLError::ModelError(format!("GPU required: {}", reason))
            }
        }
    }
}

// Note: From trait for liquid::LiquidError is implemented in the liquid module to avoid conflicts

/// Result type for ML operations
pub type MLResult<T> = Result<T, MLError>;

/// New unified result type using CommonError for better integration
pub type UnifiedMLResult<T> = Result<T, CommonError>;

/// Precision factor for fixed-point arithmetic
pub const PRECISION_FACTOR: i64 = 100_000_000;

/// Maximum inference latency target in microseconds
pub const MAX_INFERENCE_LATENCY_US: u64 = 100;

// ========== CORE ML MODULES ==========
// Core ML modules
pub mod checkpoint;
pub mod dqn;
pub mod ensemble;
pub mod flash_attention;
pub mod integration;
pub mod labeling;
pub mod liquid;
pub mod mamba;
pub mod microstructure;
pub mod ppo;
pub mod risk;
pub mod safety;
pub mod tft;
pub mod tgnn;
pub mod tlob;
pub mod transformers;
pub mod universe;

// ========== INFRASTRUCTURE MODULES ==========
// Infrastructure
pub mod benchmarks;
pub mod common;
pub mod training;

// ========== CORE EXPORTS ==========
// Core exports
pub mod error;
pub mod error_consolidated;
pub mod features;
pub mod inference;
pub mod model;
pub mod operations;
pub mod performance;
pub mod production;
pub mod validation;

// ========== ADDITIONAL MODULES ==========
// Additional ML processing modules
pub mod batch_processing; // Batch processing for ML operations
pub mod bridge; // Type system bridge for ML-Financial integration
pub mod operations_safe; // Safe operations module
pub mod ops_production; // Production ML operations
pub mod portfolio_transformer; // Portfolio-specific transformer
pub mod regime_detection; // Market regime detection
pub mod tensor_ops;
// TLOB transformer implementation moved to tlob module
pub mod examples;
// Removed examples_stubs module - contained only placeholder implementations
pub mod integration_test;
// TODO: Re-enable when model_loader types are available
// pub mod model_loader_integration;
pub mod models_demo;
pub mod observability;
pub mod stress_testing; // Stress testing framework
pub mod training_pipeline; // Complete training pipeline system
pub mod traits; // Common traits for ML models // Production observability and monitoring // Integration with model_loader crate





// ========== MISSING TYPES STUBS ==========

/// Application result wrapper for ML operations
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct MLAppResult<T> {
    pub data: T,
    pub success: bool,
    pub message: Option<String>,
    pub execution_time_ms: u64,
    pub metadata: HashMap<String, String>,
}

impl<T> MLAppResult<T> {
    /// Create a successful result
    pub fn success(data: T) -> Self {
        Self {
            data,
            success: true,
            message: None,
            execution_time_ms: 0,
            metadata: HashMap::new(),
        }
    }

    /// Create a failed result with message
    pub fn error(data: T, message: String) -> Self {
        Self {
            data,
            success: false,
            message: Some(message),
            execution_time_ms: 0,
            metadata: HashMap::new(),
        }
    }

    /// Set execution time
    pub fn with_timing(mut self, execution_time_ms: u64) -> Self {
        self.execution_time_ms = execution_time_ms;
        self
    }

    /// Add metadata
    pub fn with_metadata(mut self, key: String, value: String) -> Self {
        self.metadata.insert(key, value);
        self
    }
}

/// Performance profile configuration for HFT models
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct HFTPerformanceProfile {
    pub max_latency_us: u64,
    pub target_throughput: u32,
    pub memory_limit_mb: u64,
    pub cpu_affinity: Option<Vec<usize>>,
    pub gpu_enabled: bool,
    pub batch_size: u32,
    pub optimization_level: OptimizationLevel,
}

/// Optimization levels for HFT performance
#[derive(Debug, Clone, Copy, Serialize, Deserialize)]
pub enum OptimizationLevel {
    /// Maximum speed, minimal safety checks
    UltraLow,
    /// Balanced speed and safety
    Low,
    /// Standard optimization
    Medium,
    /// Conservative with full validation
    High,
}

impl Default for HFTPerformanceProfile {
    fn default() -> Self {
        Self {
            max_latency_us: 100,      // 100 microseconds target
            target_throughput: 10000, // 10k operations per second
            memory_limit_mb: 1024,    // 1GB memory limit
            cpu_affinity: None,
            gpu_enabled: false,
            batch_size: 1,
            optimization_level: OptimizationLevel::Medium,
        }
    }
}

/// Create HFT performance profile with default settings
pub fn create_hft_performance_profile() -> HFTPerformanceProfile {
    HFTPerformanceProfile::default()
}

/// Create HFT performance profile with custom latency target
pub fn create_hft_performance_profile_with_latency(max_latency_us: u64) -> HFTPerformanceProfile {
    HFTPerformanceProfile {
        max_latency_us,
        ..Default::default()
    }
}

/// Create HFT performance profile optimized for ultra-low latency
pub fn create_ultra_low_latency_profile() -> HFTPerformanceProfile {
    HFTPerformanceProfile {
        max_latency_us: 10,       // 10 microseconds target
        target_throughput: 50000, // 50k operations per second
        memory_limit_mb: 512,     // Reduced memory for cache efficiency
        gpu_enabled: true,        // Enable GPU acceleration
        batch_size: 1,            // No batching for minimal latency
        optimization_level: OptimizationLevel::UltraLow,
        ..Default::default()
    }
}

// ========== UNIFIED ML MODEL INTERFACE ==========

use async_trait::async_trait;
use futures::future::join_all;
use std::collections::HashMap;
use std::sync::Arc;
use tokio::sync::RwLock;

/// Features vector for ML model input
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct Features {
    /// Raw feature values
    pub values: Vec<f64>,
    /// Feature names for debugging
    pub names: Vec<String>,
    /// Timestamp of features
    pub timestamp: u64,
    /// Symbol these features are for
    pub symbol: Option<String>,
}

impl Features {
    pub fn new(values: Vec<f64>, names: Vec<String>) -> Self {
        Self {
            values,
            names,
            timestamp: std::time::SystemTime::now()
                .duration_since(std::time::UNIX_EPOCH)
                .unwrap_or_default()
                .as_micros() as u64,
            symbol: None,
        }
    }

    pub fn with_symbol(mut self, symbol: String) -> Self {
        self.symbol = Some(symbol);
        self
    }
}

/// Model prediction result
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct ModelPrediction {
    /// Predicted value (price direction, probability, etc.)
    pub value: f64,
    /// Model confidence (0.0 to 1.0)
    pub confidence: f64,
    /// Additional model-specific metadata
    pub metadata: HashMap<String, serde_json::Value>,
    /// Prediction timestamp
    pub timestamp: u64,
    /// Model identifier
    pub model_id: String,
}

impl ModelPrediction {
    pub fn new(model_id: String, value: f64, confidence: f64) -> Self {
        Self {
            value,
            confidence,
            metadata: HashMap::new(),
            timestamp: std::time::SystemTime::now()
                .duration_since(std::time::UNIX_EPOCH)
                .unwrap_or_default()
                .as_micros() as u64,
            model_id,
        }
    }

    pub fn with_metadata(mut self, key: String, value: serde_json::Value) -> Self {
        self.metadata.insert(key, value);
        self
    }
}

/// Feedback for model weight updates
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct Feedback {
    /// Actual outcome (for supervised learning)
    pub actual_value: Option<f64>,
    /// Reward signal (for reinforcement learning)
    pub reward: Option<f64>,
    /// Trading performance metrics
    pub performance_metrics: HashMap<String, f64>,
    /// Timestamp of feedback
    pub timestamp: u64,
}

impl Feedback {
    pub fn new() -> Self {
        Self {
            actual_value: None,
            reward: None,
            performance_metrics: HashMap::new(),
            timestamp: std::time::SystemTime::now()
                .duration_since(std::time::UNIX_EPOCH)
                .unwrap_or_default()
                .as_micros() as u64,
        }
    }

    pub fn with_actual(mut self, actual: f64) -> Self {
        self.actual_value = Some(actual);
        self
    }

    pub fn with_reward(mut self, reward: f64) -> Self {
        self.reward = Some(reward);
        self
    }
}

/// Unified interface for all ML models in the system
#[async_trait]
pub trait MLModel: Send + Sync {
    /// Get unique model identifier
    fn name(&self) -> &str;

    /// Get model type
    fn model_type(&self) -> ModelType;

    /// Make prediction based on features
    async fn predict(&self, features: &Features) -> MLResult<ModelPrediction>;

    /// Get current model confidence score (0.0 to 1.0)
    fn get_confidence(&self) -> f64;

    /// Update model weights based on feedback (optional - not all models support online learning)
    async fn update_weights(&mut self, _feedback: &Feedback) -> MLResult<()> {
        // Default implementation does nothing (for immutable models)
        Ok(())
    }

    /// Check if model is ready for predictions
    fn is_ready(&self) -> bool {
        true // Default to ready
    }

    /// Get model metadata
    fn get_metadata(&self) -> ModelMetadata;

    /// Validate input features
    fn validate_features(&self, features: &Features) -> MLResult<()> {
        // Default validation - check for empty features
        if features.values.is_empty() {
            return Err(MLError::ValidationError {
                message: "Empty feature vector".to_string(),
            });
        }
        Ok(())
    }
}

/// Thread-safe model registry using DashMap for high-performance concurrent access
pub struct ModelRegistry {
    /// Models stored by name
    models: dashmap::DashMap<String, Arc<dyn MLModel>>,
    /// Registry metadata
    metadata: Arc<RwLock<RegistryMetadata>>,
}

#[derive(Debug, Clone)]
struct RegistryMetadata {
    created_at: std::time::SystemTime,
    total_registrations: u64,
    last_access: std::time::SystemTime,
}

impl ModelRegistry {
    /// Create new model registry
    pub fn new() -> Self {
        Self {
            models: dashmap::DashMap::new(),
            metadata: Arc::new(RwLock::new(RegistryMetadata {
                created_at: std::time::SystemTime::now(),
                total_registrations: 0,
                last_access: std::time::SystemTime::now(),
            })),
        }
    }

    /// Register a model in the registry
    pub async fn register(&self, model: Arc<dyn MLModel>) -> MLResult<()> {
        let name = model.name().to_string();

        // Check if model is ready
        if !model.is_ready() {
            return Err(MLError::ModelError(format!("Model {} is not ready", name)));
        }

        self.models.insert(name.clone(), model);

        // Update metadata
        {
            let mut meta = self.metadata.write().await;
            meta.total_registrations += 1;
            meta.last_access = std::time::SystemTime::now();
        }

        tracing::info!("Registered ML model: {}", name);
        Ok(())
    }

    /// Get model by name
    pub async fn get(&self, name: &str) -> Option<Arc<dyn MLModel>> {
        // Update last access time
        {
            let mut meta = self.metadata.write().await;
            meta.last_access = std::time::SystemTime::now();
        }

        self.models.get(name).map(|entry| entry.value().clone())
    }

    /// Get all registered models
    pub fn get_all(&self) -> Vec<Arc<dyn MLModel>> {
        self.models
            .iter()
            .map(|entry| entry.value().clone())
            .collect()
    }

    /// Get model names
    pub fn get_model_names(&self) -> Vec<String> {
        self.models
            .iter()
            .map(|entry| entry.key().clone())
            .collect()
    }

    /// Remove model from registry
    pub async fn remove(&self, name: &str) -> Option<Arc<dyn MLModel>> {
        let result = self.models.remove(name).map(|(_, model)| model);

        if result.is_some() {
            tracing::info!("Removed ML model: {}", name);
        }

        result
    }

    /// Get registry statistics
    pub async fn get_stats(&self) -> RegistryStats {
        let meta = self.metadata.read().await;
        RegistryStats {
            total_models: self.models.len(),
            total_registrations: meta.total_registrations,
            created_at: meta.created_at,
            last_access: meta.last_access,
        }
    }

    /// Parallel prediction across all models
    pub async fn predict_all(&self, features: &Features) -> Vec<MLResult<ModelPrediction>> {
        let models = self.get_all();
        let futures = models.iter().map(|model| {
            let features = features.clone();
            async move { model.predict(&features).await }
        });

        join_all(futures).await
    }

    /// Parallel prediction across specific models
    pub async fn predict_selected(
        &self,
        model_names: &[String],
        features: &Features,
    ) -> Vec<MLResult<ModelPrediction>> {
        let futures = model_names.iter().map(|name| {
            let name = name.clone();
            let features = features.clone();
            async move {
                if let Some(model) = self.get(&name).await {
                    model.predict(&features).await
                } else {
                    Err(MLError::ModelNotFound(name))
                }
            }
        });

        join_all(futures).await
    }
}

impl Default for ModelRegistry {
    fn default() -> Self {
        Self::new()
    }
}

/// Registry statistics
#[derive(Debug, Clone)]
pub struct RegistryStats {
    pub total_models: usize,
    pub total_registrations: u64,
    pub created_at: std::time::SystemTime,
    pub last_access: std::time::SystemTime,
}

/// Global model registry instance (singleton pattern)
static GLOBAL_REGISTRY: once_cell::sync::Lazy<Arc<ModelRegistry>> =
    once_cell::sync::Lazy::new(|| Arc::new(ModelRegistry::new()));

/// Get global model registry
pub fn get_global_registry() -> Arc<ModelRegistry> {
    GLOBAL_REGISTRY.clone()
}

// ========== PARALLEL EXECUTION OPTIMIZATIONS ==========

/// High-performance parallel executor for ML models optimized for sub-50μs latency
pub struct ParallelExecutor {
    /// Performance profile
    profile: HFTPerformanceProfile,
    /// CPU affinity settings
    cpu_affinity: Option<Vec<usize>>,
    /// Thread pool for CPU-bound operations
    cpu_pool: Arc<rayon::ThreadPool>,
    /// Async runtime handle
    runtime_handle: tokio::runtime::Handle,
}

impl ParallelExecutor {
    /// Create new parallel executor with HFT performance profile
    pub fn new(profile: HFTPerformanceProfile) -> Result<Self, MLError> {
        // Create dedicated thread pool based on profile
        let cpu_pool = rayon::ThreadPoolBuilder::new()
            .num_threads(
                profile
                    .cpu_affinity
                    .as_ref()
                    .map(|v| v.len())
                    .unwrap_or(num_cpus::get()),
            )
            .thread_name(|i| format!("ml-cpu-{}", i))
            .build()
            .map_err(|e| MLError::ModelError(format!("Failed to create thread pool: {}", e)))?;

        let runtime_handle = tokio::runtime::Handle::try_current()
            .map_err(|e| MLError::ModelError(format!("No tokio runtime available: {}", e)))?;

        let cpu_affinity = profile.cpu_affinity.clone();

        Ok(Self {
            profile,
            cpu_affinity,
            cpu_pool: Arc::new(cpu_pool),
            runtime_handle,
        })
    }

    /// Execute parallel predictions with latency optimization
    pub async fn execute_parallel_predictions(
        &self,
        models: Vec<Arc<dyn MLModel>>,
        features: Features,
    ) -> Vec<MLResult<ModelPrediction>> {
        let start_time = std::time::Instant::now();

        // Determine execution strategy based on performance profile
        let results = match self.profile.optimization_level {
            OptimizationLevel::UltraLow => {
                // Ultra-low latency: parallel execution with minimal overhead
                self.execute_ultra_low_latency(models, features).await
            }
            OptimizationLevel::Low => {
                // Low latency: parallel with basic batching
                self.execute_low_latency(models, features).await
            }
            OptimizationLevel::Medium => {
                // Medium: balanced parallel execution
                self.execute_balanced(models, features).await
            }
            OptimizationLevel::High => {
                // High: conservative with full validation
                self.execute_conservative(models, features).await
            }
        };

        let execution_time = start_time.elapsed();

        // Log performance if exceeding target latency
        if execution_time.as_micros() > self.profile.max_latency_us as u128 {
            tracing::warn!(
                "Parallel execution exceeded target latency: {}μs > {}μs",
                execution_time.as_micros(),
                self.profile.max_latency_us
            );
        }

        results
    }

    /// Ultra-low latency execution (<10μs target)
    async fn execute_ultra_low_latency(
        &self,
        models: Vec<Arc<dyn MLModel>>,
        features: Features,
    ) -> Vec<MLResult<ModelPrediction>> {
        // Use futures::future::join_all for minimal overhead
        let futures = models.into_iter().map(|model| {
            let features = features.clone();
            async move { model.predict(&features).await }
        });

        join_all(futures).await
    }

    /// Low latency execution with basic optimizations
    async fn execute_low_latency(
        &self,
        models: Vec<Arc<dyn MLModel>>,
        features: Features,
    ) -> Vec<MLResult<ModelPrediction>> {
        // Group models by type for potential batching
        let mut model_groups: HashMap<ModelType, Vec<Arc<dyn MLModel>>> = HashMap::new();

        for model in models {
            let model_type = model.model_type();
            model_groups.entry(model_type).or_default().push(model);
        }

        let mut all_futures = Vec::new();

        for (_, group_models) in model_groups {
            for model in group_models {
                let features = features.clone();
                all_futures.push(async move { model.predict(&features).await });
            }
        }

        join_all(all_futures).await
    }

    /// Balanced execution with moderate optimizations
    async fn execute_balanced(
        &self,
        models: Vec<Arc<dyn MLModel>>,
        features: Features,
    ) -> Vec<MLResult<ModelPrediction>> {
        // Validate features once for all models
        for model in &models {
            if let Err(e) = model.validate_features(&features) {
                tracing::debug!(
                    "Feature validation failed for model {}: {}",
                    model.name(),
                    e
                );
            }
        }

        let futures = models.into_iter().map(|model| {
            let features = features.clone();
            async move {
                if model.is_ready() {
                    model.predict(&features).await
                } else {
                    Err(MLError::ModelError(format!(
                        "Model {} not ready",
                        model.name()
                    )))
                }
            }
        });

        join_all(futures).await
    }

    /// Conservative execution with full validation
    async fn execute_conservative(
        &self,
        models: Vec<Arc<dyn MLModel>>,
        features: Features,
    ) -> Vec<MLResult<ModelPrediction>> {
        let mut results = Vec::new();

        for model in models {
            // Comprehensive validation
            if !model.is_ready() {
                results.push(Err(MLError::ModelError(format!(
                    "Model {} not ready",
                    model.name()
                ))));
                continue;
            }

            if let Err(e) = model.validate_features(&features) {
                results.push(Err(e));
                continue;
            }

            // Execute with timeout
            let prediction_future = model.predict(&features);
            let timeout_duration = std::time::Duration::from_micros(self.profile.max_latency_us);

            match tokio::time::timeout(timeout_duration, prediction_future).await {
                Ok(result) => results.push(result),
                Err(_) => results.push(Err(MLError::ModelError(format!(
                    "Model {} prediction timed out after {}μs",
                    model.name(),
                    self.profile.max_latency_us
                )))),
            }
        }

        results
    }

    /// Get execution statistics
    pub fn get_stats(&self) -> ExecutorStats {
        ExecutorStats {
            optimization_level: self.profile.optimization_level,
            target_latency_us: self.profile.max_latency_us,
            cpu_threads: self.cpu_pool.current_num_threads(),
            cpu_affinity: self.cpu_affinity.clone(),
        }
    }
}

/// Executor performance statistics
#[derive(Debug, Clone)]
pub struct ExecutorStats {
    pub optimization_level: OptimizationLevel,
    pub target_latency_us: u64,
    pub cpu_threads: usize,
    pub cpu_affinity: Option<Vec<usize>>,
}

/// Latency optimizer for ML inference pipelines
pub struct LatencyOptimizer {
    /// Target latency in microseconds
    target_latency_us: u64,
    /// Performance history
    performance_history: Arc<RwLock<Vec<PerformancePoint>>>,
    /// Optimization parameters
    optimization_params: OptimizationParams,
}

#[derive(Debug, Clone)]
struct PerformancePoint {
    timestamp: std::time::Instant,
    latency_us: u64,
    model_count: usize,
    batch_size: u32,
    success: bool,
}

#[derive(Debug, Clone)]
struct OptimizationParams {
    max_batch_size: u32,
    adaptive_batching: bool,
    prefetch_enabled: bool,
    cache_predictions: bool,
}

impl Default for OptimizationParams {
    fn default() -> Self {
        Self {
            max_batch_size: 8,
            adaptive_batching: true,
            prefetch_enabled: true,
            cache_predictions: false, // Disabled for real-time trading
        }
    }
}

impl LatencyOptimizer {
    /// Create new latency optimizer
    pub fn new(target_latency_us: u64) -> Self {
        Self {
            target_latency_us,
            performance_history: Arc::new(RwLock::new(Vec::new())),
            optimization_params: OptimizationParams::default(),
        }
    }

    /// Record performance measurement
    pub async fn record_performance(
        &self,
        latency_us: u64,
        model_count: usize,
        batch_size: u32,
        success: bool,
    ) {
        let point = PerformancePoint {
            timestamp: std::time::Instant::now(),
            latency_us,
            model_count,
            batch_size,
            success,
        };

        {
            let mut history = self.performance_history.write().await;
            history.push(point);

            // Keep only recent history (last 1000 measurements)
            if history.len() > 1000 {
                let excess = history.len() - 1000;
                history.drain(0..excess);
            }
        }
    }

    /// Get optimization recommendations
    pub async fn get_recommendations(&self) -> OptimizationRecommendations {
        let history = self.performance_history.read().await;

        if history.is_empty() {
            return OptimizationRecommendations::default();
        }

        let recent_points: Vec<&PerformancePoint> = history.iter().rev().take(100).collect();

        let avg_latency =
            recent_points.iter().map(|p| p.latency_us).sum::<u64>() / recent_points.len() as u64;

        let success_rate =
            recent_points.iter().filter(|p| p.success).count() as f64 / recent_points.len() as f64;

        OptimizationRecommendations {
            current_avg_latency_us: avg_latency,
            target_latency_us: self.target_latency_us,
            success_rate,
            meets_target: avg_latency <= self.target_latency_us,
            recommended_batch_size: self.calculate_optimal_batch_size(&recent_points),
            recommended_model_limit: self.calculate_optimal_model_limit(&recent_points),
        }
    }

    fn calculate_optimal_batch_size(&self, points: &[&PerformancePoint]) -> u32 {
        // Simple heuristic: find batch size with best latency/success ratio
        let mut batch_performance: HashMap<u32, (u64, f64)> = HashMap::new();

        for point in points {
            let entry = batch_performance
                .entry(point.batch_size)
                .or_insert((0, 0.0));
            entry.0 += point.latency_us;
            entry.1 += if point.success { 1.0 } else { 0.0 };
        }

        batch_performance
            .into_iter()
            .filter(|(_, (_, success_count))| *success_count > 0.0)
            .min_by_key(|(_, (latency, success_count))| {
                // Optimize for latency with success rate weighting
                ((*latency as f64) / success_count) as u64
            })
            .map(|(batch_size, _)| batch_size)
            .unwrap_or(1)
    }

    fn calculate_optimal_model_limit(&self, points: &[&PerformancePoint]) -> usize {
        // Find the sweet spot where adding more models doesn't improve latency
        let mut model_performance: HashMap<usize, u64> = HashMap::new();

        for point in points {
            if point.success {
                let entry = model_performance.entry(point.model_count).or_insert(0);
                *entry += point.latency_us;
            }
        }

        model_performance
            .into_iter()
            .filter(|(_, avg_latency)| *avg_latency <= self.target_latency_us)
            .max_by_key(|(model_count, _)| *model_count)
            .map(|(model_count, _)| model_count)
            .unwrap_or(1)
    }
}

/// Optimization recommendations from latency analysis
#[derive(Debug, Clone)]
pub struct OptimizationRecommendations {
    pub current_avg_latency_us: u64,
    pub target_latency_us: u64,
    pub success_rate: f64,
    pub meets_target: bool,
    pub recommended_batch_size: u32,
    pub recommended_model_limit: usize,
}

impl Default for OptimizationRecommendations {
    fn default() -> Self {
        Self {
            current_avg_latency_us: 0,
            target_latency_us: 50,
            success_rate: 0.0,
            meets_target: false,
            recommended_batch_size: 1,
            recommended_model_limit: 1,
        }
    }
}

/// Create optimized parallel executor for HFT scenarios
pub fn create_hft_parallel_executor() -> Result<ParallelExecutor, MLError> {
    let profile = create_ultra_low_latency_profile();
    ParallelExecutor::new(profile)
}

/// Create latency optimizer with HFT targets
pub fn create_hft_latency_optimizer() -> LatencyOptimizer {
    LatencyOptimizer::new(50) // 50 microsecond target
}

// ========== CANONICAL TRAINING AND VALIDATION METRICS ==========
// These are the unified types that all ML modules must use to prevent type conflicts

/// Canonical training metrics used throughout ML module
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct TrainingMetrics {
    /// Training loss value
    pub loss: f64,
    /// Training accuracy (0.0 to 1.0)
    pub accuracy: f64,
    /// Training precision (0.0 to 1.0)
    pub precision: f64,
    /// Training recall (0.0 to 1.0)
    pub recall: f64,
    /// Training F1 score (0.0 to 1.0)
    pub f1_score: f64,
    /// Total training time in seconds
    pub training_time_seconds: f64,
    /// Number of epochs trained
    pub epochs_trained: u32,
    /// Whether convergence was achieved
    pub convergence_achieved: bool,
    /// Additional model-specific metrics
    pub additional_metrics: HashMap<String, f64>,
}

impl TrainingMetrics {
    /// Create new training metrics
    pub fn new() -> Self {
        Self {
            loss: 0.0,
            accuracy: 0.0,
            precision: 0.0,
            recall: 0.0,
            f1_score: 0.0,
            training_time_seconds: 0.0,
            epochs_trained: 0,
            convergence_achieved: false,
            additional_metrics: HashMap::new(),
        }
    }

    /// Add an additional metric
    pub fn add_metric(&mut self, name: &str, value: f64) {
        self.additional_metrics.insert(name.to_string(), value);
    }

    /// Check if training was successful
    pub fn is_successful(&self) -> bool {
        self.convergence_achieved && self.accuracy > 0.5
    }
}

impl Default for TrainingMetrics {
    fn default() -> Self {
        Self::new()
    }
}

/// Canonical validation metrics used throughout ML module
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct ValidationMetrics {
    /// Validation loss value
    pub validation_loss: f64,
    /// Validation accuracy (0.0 to 1.0)
    pub validation_accuracy: f64,
    /// Validation precision (0.0 to 1.0)
    pub validation_precision: f64,
    /// Validation recall (0.0 to 1.0)
    pub validation_recall: f64,
    /// Validation F1 score (0.0 to 1.0)
    pub validation_f1_score: f64,
    /// Number of samples validated
    pub samples_validated: usize,
    /// Additional model-specific validation metrics
    pub additional_metrics: HashMap<String, f64>,
}

impl ValidationMetrics {
    /// Create new validation metrics
    pub fn new() -> Self {
        Self {
            validation_loss: 0.0,
            validation_accuracy: 0.0,
            validation_precision: 0.0,
            validation_recall: 0.0,
            validation_f1_score: 0.0,
            samples_validated: 0,
            additional_metrics: HashMap::new(),
        }
    }

    /// Add an additional validation metric
    pub fn add_metric(&mut self, name: &str, value: f64) {
        self.additional_metrics.insert(name.to_string(), value);
    }

    /// Check if validation was successful
    pub fn is_successful(&self) -> bool {
        self.validation_accuracy > 0.5 && self.samples_validated > 0
    }
}

impl Default for ValidationMetrics {
    fn default() -> Self {
        Self::new()
    }
}

// Public exports moved to end of file after all type definitions

// ========== CANONICAL ML TYPES ==========
// These are the unified types that all ML modules must use to prevent type conflicts

/// Canonical inference result used throughout ML module
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct InferenceResult {
    /// Model identifier
    pub model_id: String,
    /// Prediction value (primary prediction)
    pub prediction_value: f64,
    /// Confidence score (0.0 to 1.0)
    pub confidence: f64,
    /// Latency in microseconds
    pub latency_us: u64,
    /// Timestamp in microseconds since UNIX epoch
    pub timestamp: u64,
    /// Model metadata
    pub metadata: ModelMetadata,
}

impl InferenceResult {
    /// Create new inference result
    pub fn new(
        model_id: String,
        prediction_value: f64,
        confidence: f64,
        latency_us: u64,
        timestamp: u64,
        metadata: ModelMetadata,
    ) -> Self {
        Self {
            model_id,
            prediction_value,
            confidence,
            latency_us,
            timestamp,
            metadata,
        }
    }

    /// Extract prediction as float value
    pub fn prediction_as_float(&self) -> f64 {
        self.prediction_value
    }

    /// Get the model identifier
    pub fn model_id(&self) -> &str {
        &self.model_id
    }
}

/// Canonical model metadata used throughout ML module
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct ModelMetadata {
    /// Type of the model
    pub model_type: ModelType,
    /// Model version
    pub version: String,
    /// Number of features used for inference
    pub features_used: usize,
    /// Memory usage in megabytes
    pub memory_usage_mb: f64,
    /// Additional metadata key-value pairs
    pub additional_metadata: HashMap<String, String>,
}

impl ModelMetadata {
    /// Create new model metadata
    pub fn new(
        model_type: ModelType,
        version: String,
        features_used: usize,
        memory_usage_mb: f64,
    ) -> Self {
        Self {
            model_type,
            version,
            features_used,
            memory_usage_mb,
            additional_metadata: HashMap::new(),
        }
    }

    /// Add additional metadata
    pub fn add_metadata(&mut self, key: &str, value: String) {
        self.additional_metadata.insert(key.to_string(), value);
    }

    /// Mark the model as trained
    pub fn mark_trained(&mut self) {
        self.add_metadata("training_status", "trained".to_string());
        self.add_metadata(
            "training_timestamp",
            std::time::SystemTime::now()
                .duration_since(std::time::UNIX_EPOCH)
                .unwrap_or_default()
                .as_secs()
                .to_string(),
        );
    }
}

/// Canonical model type enum used throughout ML module
#[derive(Debug, Clone, Copy, Serialize, Deserialize, PartialEq, Eq, Hash)]
pub enum ModelType {
    /// Compact Deep Q-Network
    CompactDQN,
    /// Distilled micro network for ultra-low latency
    DistilledMicroNet,
    /// Standard Deep Q-Network
    DQN,
    /// Rainbow DQN with all enhancements
    RainbowDQN,
    /// MAMBA model (SSM)
    MAMBA,
    /// Temporal Fusion Transformer
    TFT,
    /// Temporal Graph Neural Network
    TGGN,
    /// Liquid Neural Network
    LNN,
    /// Temporal Limit Order Book transformer
    TLOB,
    /// Proximal Policy Optimization
    PPO,
    /// Transformer for sequence modeling
    Transformer,
    /// Mamba state space model (alias for MAMBA)
    Mamba,
    /// Liquid time constant networks (alias for LNN)
    LiquidNet,
    /// Temporal Graph Neural Network (alias for TGGN)
    TGNN,
    /// Ensemble methods
    Ensemble,
}

impl ModelType {
    /// Get file extension for model type
    pub fn file_extension(&self) -> &'static str {
        match self {
            ModelType::DQN => "dqn",
            ModelType::MAMBA | ModelType::Mamba => "mamba",
            ModelType::TFT => "tft",
            ModelType::TGGN | ModelType::TGNN => "tggn",
            ModelType::LNN | ModelType::LiquidNet => "lnn",
            ModelType::CompactDQN => "compact_dqn",
            ModelType::DistilledMicroNet => "distilled",
            ModelType::RainbowDQN => "rainbow_dqn",
            ModelType::TLOB => "tlob",
            ModelType::PPO => "ppo",
            ModelType::Transformer => "transformer",
            ModelType::Ensemble => "ensemble",
        }
    }

    /// Get model type from string
    pub fn from_str(s: &str) -> Option<Self> {
        match s.to_lowercase().as_str() {
            "dqn" => Some(ModelType::DQN),
            "mamba" => Some(ModelType::MAMBA),
            "tft" => Some(ModelType::TFT),
            "tggn" | "tgnn" => Some(ModelType::TGGN),
            "lnn" | "liquidnet" => Some(ModelType::LNN),
            "compact_dqn" | "compactdqn" => Some(ModelType::CompactDQN),
            "distilled" | "distilledmicronet" => Some(ModelType::DistilledMicroNet),
            "rainbow_dqn" | "rainbowdqn" => Some(ModelType::RainbowDQN),
            "tlob" => Some(ModelType::TLOB),
            "ppo" => Some(ModelType::PPO),
            "transformer" => Some(ModelType::Transformer),
            "ensemble" => Some(ModelType::Ensemble),
            _ => None,
        }
    }
}

// TEMPORARILY COMMENTED OUT - These modules need to be checked for availability
// Re-export training pipeline system (from existing training_pipeline module)
// pub use training_pipeline::{
//     ProductionMLTrainingSystem, ProductionTrainingConfig, ProductionTrainingMetrics,
//     FinancialFeatures, MicrostructureFeatures, RiskFeatures, TrainingResult,
// };

// Note: All types in this module are already public and available
// External crates can import them directly as: use ml::{Features, ModelPrediction, etc.}

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn test_ml_error_creation() -> Result<(), Box<dyn std::error::Error>> {
        let error = MLError::ConfigError {
            reason: "test".to_string(),
        };
        assert!(error.to_string().contains("Configuration error"));
        Ok(())
    }
}