foxhunt/ml/tests/gpu_benchmark_integration_tests.rs

//! GPU Benchmark Integration Tests
//!
//! Comprehensive end-to-end tests for the GPU training benchmark system.
//! Tests full workflow from data loading through model benchmarking to result generation.
//!
//! ## Test Categories
//!
//! 1. **Setup & Fixtures**: Test helpers and data loading
//! 2. **Module Integration**: Cross-module functionality
//! 3. **Model Benchmarks**: Full model training runs (GPU-only)
//! 4. **End-to-End**: Complete benchmark coordinator
//! 5. **Error Handling**: Failure modes and recovery
//! 6. **Performance**: Overhead and efficiency validation
//!
//! ## GPU Test Convention
//!
//! Tests marked with `#[ignore]` require GPU and are slow (5-60 minutes).
//! Run with: `cargo test -- --ignored --nocapture`

use ml::benchmark::*;
use std::path::PathBuf;
use std::sync::Arc;
use tokio;

// ============================================================================
// Section 1: Setup and Fixtures (Test Helpers)
// ============================================================================

/// Test helper: Create GPU manager with graceful CPU fallback
fn setup_gpu_manager() -> Arc<GpuHardwareManager> {
    Arc::new(
        GpuHardwareManager::new()
            .expect("Failed to create GPU manager (check CUDA installation)"),
    )
}

/// Test helper: Load small subset of DBN data for testing
/// Uses 6E.FUT (Euro FX futures) with 1000 bars maximum
fn load_test_data() -> Vec<MarketDataPoint> {
    let data_dir = PathBuf::from("test_data/real/databento/ml_training");

    if !data_dir.exists() {
        eprintln!(
            "WARNING: Test data directory not found: {:?}",
            data_dir
        );
        return vec![];
    }

    let mut loader = DbnDataLoader::new(data_dir.clone());

    // Try to load 6E.FUT (Euro FX) - one of the smaller datasets
    match loader.load_symbol_data("6E.FUT") {
        Ok(data) => {
            let subset: Vec<_> = data.into_iter().take(1000).collect();
            println!(
                "[Test Helper] Loaded {} market data points from 6E.FUT",
                subset.len()
            );
            subset
        }
        Err(e) => {
            eprintln!("WARNING: Failed to load test data: {}", e);
            vec![]
        }
    }
}

/// Test helper: Create minimal batch size config for testing
fn create_test_batch_config() -> BatchSizeConfig {
    BatchSizeConfig::new(32, 1) // batch_size=32, gradient_accumulation_steps=1
}

// ============================================================================
// Section 2: Module Integration Tests (6 tests)
// ============================================================================

#[tokio::test]
#[ignore] // GPU-only test (requires CUDA)
async fn test_gpu_warmup_reduces_variance() {
    // Test that warmup protocol reduces first-epoch timing variance
    // Expected: >50% reduction in standard deviation

    let gpu_manager = setup_gpu_manager();

    // Measure cold start variance (no warmup)
    let mut cold_times = Vec::new();
    for _ in 0..5 {
        let start = std::time::Instant::now();
        gpu_manager
            .device()
            .synchronize()
            .expect("Device sync failed");
        cold_times.push(start.elapsed().as_micros() as f64);
    }

    let cold_mean = cold_times.iter().sum::<f64>() / cold_times.len() as f64;
    let cold_variance = cold_times
        .iter()
        .map(|x| (x - cold_mean).powi(2))
        .sum::<f64>()
        / cold_times.len() as f64;
    let cold_std_dev = cold_variance.sqrt();

    // Warmup GPU
    gpu_manager.warmup().expect("GPU warmup failed");

    // Measure warm start variance (after warmup)
    let mut warm_times = Vec::new();
    for _ in 0..5 {
        let start = std::time::Instant::now();
        gpu_manager
            .device()
            .synchronize()
            .expect("Device sync failed");
        warm_times.push(start.elapsed().as_micros() as f64);
    }

    let warm_mean = warm_times.iter().sum::<f64>() / warm_times.len() as f64;
    let warm_variance = warm_times
        .iter()
        .map(|x| (x - warm_mean).powi(2))
        .sum::<f64>()
        / warm_times.len() as f64;
    let warm_std_dev = warm_variance.sqrt();

    println!("[GPU Warmup Test]");
    println!("  Cold std dev: {:.2}μs", cold_std_dev);
    println!("  Warm std dev: {:.2}μs", warm_std_dev);
    println!(
        "  Reduction: {:.1}%",
        (1.0 - warm_std_dev / cold_std_dev) * 100.0
    );

    // Assert warmup reduces variance by >50%
    assert!(
        warm_std_dev < cold_std_dev * 0.5,
        "Warmup should reduce timing variance by >50% (cold: {:.2}, warm: {:.2})",
        cold_std_dev,
        warm_std_dev
    );
}

#[tokio::test]
async fn test_statistical_sampler_with_real_timings() {
    // Test statistical sampler with realistic epoch timings
    // Verify 95% CI calculation, outlier removal

    let mut sampler = StatisticalSampler::new(2); // 2 warmup epochs

    // Simulate 10 epoch timings with some variance (in seconds, not milliseconds)
    let epoch_times_s = vec![
        0.150, 0.145, // Warmup (ignored)
        0.100, 0.102, 0.098, 0.101, 0.099, // Normal epochs
        0.097, 0.103, 0.1005, 0.1015, // More normal epochs
        0.200, // Outlier (spike)
    ];

    for &time_s in epoch_times_s.iter() {
        sampler.add_sample(time_s);
    }

    let stats = sampler.compute_statistics().expect("Failed to compute statistics");

    println!("[Statistical Sampler Test]");
    println!("  Mean: {:.2}ms", stats.mean_seconds * 1000.0);
    println!("  Median (P50): {:.2}ms", stats.p50_median * 1000.0);
    println!("  Std Dev: {:.2}ms", stats.std_dev * 1000.0);
    println!(
        "  95% CI: [{:.2}, {:.2}]ms",
        stats.confidence_interval_95.0 * 1000.0,
        stats.confidence_interval_95.1 * 1000.0
    );
    println!("  Number of samples: {}", stats.num_samples);
    println!("  Outliers removed: {}", stats.outliers_removed);

    // Verify samples were recorded (after warmup)
    assert!(stats.num_samples > 0, "Should have recorded samples");

    // Verify mean is reasonable (should be ~100ms, excluding outlier)
    let mean_ms = stats.mean_seconds * 1000.0;
    assert!(
        mean_ms > 90.0 && mean_ms < 150.0,
        "Mean should be ~100ms, got {:.2}ms",
        mean_ms
    );

    // Verify median is close to mean (data should be roughly normal)
    let median_ms = stats.p50_median * 1000.0;
    assert!(
        (median_ms - mean_ms).abs() < 50.0,
        "Median should be reasonably close to mean"
    );

    // Verify 95% CI includes the mean
    assert!(
        stats.confidence_interval_95.0 < stats.mean_seconds
            && stats.mean_seconds < stats.confidence_interval_95.1,
        "Mean should be within 95% CI"
    );
}

#[tokio::test]
#[ignore] // GPU-only test
async fn test_batch_size_finder_gpu() {
    // Test batch size finder on real GPU
    // Verify OOM detection, binary search convergence

    let gpu_manager = setup_gpu_manager();
    let device = gpu_manager.device().clone();

    let finder = BatchSizeFinder::new(device);

    // Simple test function that fails for very large batch sizes
    let test_fn = |batch_size: usize| -> Result<bool, ml::MLError> {
        if batch_size > 8192 {
            Err(ml::MLError::ModelError("Out of memory".to_string())) // Simulate OOM
        } else {
            Ok(true)
        }
    };

    let config = finder
        .find_optimal_batch_size(test_fn)
        .expect("Batch size finder failed");

    println!("[Batch Size Finder Test]");
    println!("  Found batch size: {}", config.batch_size);
    println!("  Gradient accum steps: {}", config.gradient_accumulation_steps);
    println!("  Effective batch size: {}", config.effective_batch_size);

    // Verify batch size is reasonable
    assert!(
        config.batch_size >= 16 && config.batch_size <= 8192,
        "Batch size should be in range [16, 8192], got {}",
        config.batch_size
    );

    // Verify effective batch size >= batch size
    assert!(
        config.effective_batch_size >= config.batch_size,
        "Effective batch size should be >= batch size"
    );
}

#[tokio::test]
#[ignore] // GPU-only test
async fn test_memory_profiler_accuracy() {
    // Test memory profiler against real GPU allocations
    // Verify peak/avg calculations accurate

    use candle_core::{Device, Tensor};

    let gpu_manager = setup_gpu_manager();

    // Check if CUDA is available
    let device = gpu_manager.device();
    if !matches!(device, Device::Cuda(_)) {
        println!("[Memory Profiler Test] Skipping - CUDA not available");
        return;
    }

    let mut profiler = MemoryProfiler::new(0); // GPU 0

    // Take baseline snapshot
    profiler.take_snapshot().expect("Failed to take snapshot");

    // Allocate some memory (simulate model tensors)
    let _tensor1 = Tensor::zeros(&[1000, 1000], candle_core::DType::F32, device)
        .expect("Failed to allocate tensor1");
    profiler.take_snapshot().expect("Failed to take snapshot");

    let _tensor2 = Tensor::zeros(&[2000, 2000], candle_core::DType::F32, device)
        .expect("Failed to allocate tensor2");
    profiler.take_snapshot().expect("Failed to take snapshot");

    // Drop first tensor (memory should decrease)
    drop(_tensor1);
    profiler.take_snapshot().expect("Failed to take snapshot");

    let peak_mb = profiler.peak_usage_mb();
    let avg_mb = profiler.avg_usage_mb();
    let snapshot_count = profiler.snapshot_count();

    println!("[Memory Profiler Test]");
    println!("  Peak memory: {:.2}MB", peak_mb);
    println!("  Average memory: {:.2}MB", avg_mb);
    println!("  Total snapshots: {}", snapshot_count);

    // Verify peak >= average (peak should be at least as much as average)
    assert!(
        peak_mb >= avg_mb,
        "Peak memory should be >= average memory"
    );

    // Verify we took 4 snapshots
    assert_eq!(snapshot_count, 4);

    // Verify peak is reasonable (should be >0 since we allocated tensors)
    assert!(
        peak_mb > 0.0,
        "Peak memory should be >0 after allocations"
    );
}

#[tokio::test]
async fn test_stability_validator_detects_divergence() {
    // Test stability validator with diverging loss curve
    // Verify early detection (<5 epochs)

    let mut validator = StabilityValidator::new();

    // Simulate diverging training (exponentially increasing loss)
    let losses = vec![1.0, 1.5, 2.5, 5.0, 10.0, 25.0];
    let gradient_norms = vec![0.1, 0.5, 1.0, 5.0, 20.0, 100.0];

    for (epoch, (&loss, &grad_norm)) in losses.iter().zip(gradient_norms.iter()).enumerate() {
        validator.record_loss(loss);
        validator.record_gradient_norm(grad_norm);

        let metrics = validator.validate();

        println!(
            "[Stability Validator Test] Epoch {}: loss={:.2}, grad_norm={:.2}, trend={:?}",
            epoch, loss, grad_norm, metrics.loss_trend
        );

        // Check if divergence detected early
        if epoch >= 4 {
            // Should detect divergence by epoch 4
            assert!(
                matches!(metrics.loss_trend, LossTrend::Diverging),
                "Should detect diverging loss trend by epoch 4"
            );
            assert!(
                matches!(metrics.gradient_health, GradientHealth::Exploding),
                "Should detect exploding gradients by epoch 4"
            );
            break;
        }
    }
}

#[tokio::test]
async fn test_data_loader_loads_all_symbols() {
    // Test DBN data loader with all available symbols
    // Verify data validation catches errors

    let data_dir = PathBuf::from("test_data/real/databento/ml_training");

    if !data_dir.exists() {
        println!("[Data Loader Test] Skipping - test data not found");
        return;
    }

    let mut loader = DbnDataLoader::new(data_dir);

    // Load all data
    let data = match loader.load_all_data() {
        Ok(d) => d,
        Err(e) => {
            println!("[Data Loader Test] Failed to load data: {}", e);
            return;
        }
    };

    let stats = loader.data_statistics();

    println!("[Data Loader Test]");
    println!("  Total files: {}", stats.total_files);
    println!("  Total bars: {}", stats.total_bars);
    println!("  Symbols: {:?}", stats.symbols);
    println!("  Date range: {:?}", stats.date_range);
    println!("  Invalid bars: {}", stats.invalid_bars);

    // Verify we found some data
    assert!(stats.total_files > 0, "Should find DBN files");
    assert!(stats.total_bars > 0, "Should load some data bars");
    assert!(!stats.symbols.is_empty(), "Should have symbols");

    // Verify data validation (use data we already loaded)
    if data.is_empty() {
        println!("[Data Loader Test] Warning: No data loaded for validation");
        return;
    }

    // Check first few bars for validity
    for (i, bar) in data.iter().take(10).enumerate() {
        assert!(
            bar.is_valid(),
            "Bar {} should be valid",
            i
        );
    }
}

// ============================================================================
// Section 3: Model Benchmark Tests (4 tests, GPU-only)
// ============================================================================

#[tokio::test]
#[ignore] // GPU-only, slow (5-10 min)
async fn test_dqn_benchmark_full_run() {
    // Full DQN benchmark: 10 epochs, real data
    // Verify: Memory <150MB, training stable, statistics valid

    let test_data = load_test_data();
    if test_data.is_empty() {
        println!("[DQN Benchmark Test] Skipping - no test data available");
        return;
    }

    let gpu_manager = setup_gpu_manager();
    let mut runner = DqnBenchmarkRunner::new(gpu_manager);

    let result = runner
        .run_benchmark(10)
        .await
        .expect("DQN benchmark failed");

    println!("[DQN Benchmark Test]");
    println!("  Model: {}", result.model_name);
    println!("  Total epochs: {}", result.total_epochs);
    println!(
        "  Mean epoch time: {:.2}ms",
        result.statistics.mean_seconds * 1000.0
    );
    println!("  Peak memory: {:.2}MB", result.memory_peak_mb);
    println!("  Average loss: {:.4}", result.avg_loss);
    println!("  Stability: {:?}", result.stability);

    // Verify memory usage is reasonable (<150MB target)
    assert!(
        result.memory_peak_mb < 150.0,
        "DQN should use <150MB VRAM, got {:.2}MB",
        result.memory_peak_mb
    );

    // Verify training completed all epochs
    assert_eq!(result.total_epochs, 10);

    // Verify statistics are valid
    assert!(result.statistics.mean_seconds > 0.0);
    assert!(result.statistics.std_dev >= 0.0);

    // Verify losses recorded
    assert_eq!(result.training_losses.len(), 10);
    assert!(result.avg_loss > 0.0);

    // Verify stability metrics computed
    // LossTrend should be one of: Converging, Diverging, or Stagnant
    // GradientHealth should be one of: Healthy, Exploding, or Vanishing
    assert!(
        matches!(result.stability.loss_trend, LossTrend::Converging | LossTrend::Diverging | LossTrend::Stagnant),
        "Loss trend should be computed"
    );
    assert!(
        matches!(result.stability.gradient_health, GradientHealth::Healthy | GradientHealth::Exploding | GradientHealth::Vanishing),
        "Gradient health should be computed"
    );
}

#[tokio::test]
#[ignore] // GPU-only, slow (5-10 min)
async fn test_ppo_benchmark_full_run() {
    // Full PPO benchmark: 10 epochs, real data
    // Verify: Memory <200MB, training stable, statistics valid

    let test_data = load_test_data();
    if test_data.is_empty() {
        println!("[PPO Benchmark Test] Skipping - no test data available");
        return;
    }

    let gpu_manager = setup_gpu_manager();
    let mut runner = PpoBenchmarkRunner::new(gpu_manager);

    let result = runner
        .run_benchmark(10)
        .await
        .expect("PPO benchmark failed");

    println!("[PPO Benchmark Test]");
    println!("  Model: {}", result.model_name);
    println!("  Total epochs: {}", result.total_epochs);
    println!(
        "  Mean epoch time: {:.2}ms",
        result.statistics.mean_seconds * 1000.0
    );
    println!("  Peak memory: {:.2}MB", result.memory_peak_mb);
    println!("  Avg policy loss: {:.4}", result.avg_policy_loss);
    println!("  Avg value loss: {:.4}", result.avg_value_loss);

    // Verify memory usage is reasonable (<200MB target)
    assert!(
        result.memory_peak_mb < 200.0,
        "PPO should use <200MB VRAM, got {:.2}MB",
        result.memory_peak_mb
    );

    // Verify training completed all epochs
    assert_eq!(result.total_epochs, 10);

    // Verify statistics are valid
    assert!(result.statistics.mean_seconds > 0.0);
    assert!(result.statistics.std_dev >= 0.0);

    // Verify epoch times recorded
    assert_eq!(result.epoch_times_ms.len(), 10);

    // Verify losses are reasonable
    assert!(result.avg_policy_loss.is_finite());
    assert!(result.avg_value_loss.is_finite());
}

#[tokio::test]
#[ignore] // GPU-only, slow (10-20 min) - PLACEHOLDER
async fn test_mamba2_benchmark_full_run() {
    // Full MAMBA-2 benchmark: 10 epochs, real data
    // Verify: Memory 150-500MB, gradient accumulation working

    // TODO: Implement when MAMBA-2 benchmark module is ready
    // Expected to be implemented by Module 7a

    println!("[MAMBA-2 Benchmark Test] PLACEHOLDER - waiting for Module 7a");
    println!("  Expected memory: 150-500MB VRAM");
    println!("  Expected features: Gradient accumulation for large model");

    // This test will be implemented after Module 7a is complete
}

#[tokio::test]
#[ignore] // GPU-only, slow (10-20 min) - PLACEHOLDER
async fn test_tft_benchmark_full_run() {
    // Full TFT benchmark: 10 epochs, real data (batch_size ≤4)
    // Verify: Memory 1.5-2.5GB, doesn't OOM

    // TODO: Implement when TFT benchmark module is ready
    // Expected to be implemented by Module 7b

    println!("[TFT Benchmark Test] PLACEHOLDER - waiting for Module 7b");
    println!("  Expected memory: 1.5-2.5GB VRAM");
    println!("  Expected constraint: batch_size ≤ 4 (very memory intensive)");

    // This test will be implemented after Module 7b is complete
}

// ============================================================================
// Section 4: End-to-End Coordinator Test (1 test)
// ============================================================================

#[tokio::test]
#[ignore] // GPU-only, very slow (30-60 min) - PLACEHOLDER
async fn test_full_benchmark_coordinator() {
    // Run complete benchmark: DQN + PPO + MAMBA-2 + TFT
    // Verify:
    // - JSON output generated
    // - Decision framework applied correctly
    // - All statistics valid
    // - No crashes or panics

    // TODO: Implement when coordinator module (Module 10) is ready

    println!("[Full Coordinator Test] PLACEHOLDER - waiting for Module 10");
    println!("  Expected duration: 30-60 minutes");
    println!("  Expected output: JSON report with all model benchmarks");
    println!("  Expected decision: local_gpu or cloud_gpu recommendation");

    // This test will run:
    // 1. Load CLI args with reduced epochs (5 instead of 30)
    // 2. Run all benchmarks sequentially
    // 3. Apply decision framework
    // 4. Generate JSON report
    // 5. Verify all outputs valid
}

// ============================================================================
// Section 5: Error Handling Tests (4 tests)
// ============================================================================

#[tokio::test]
async fn test_graceful_cpu_fallback() {
    // Test CPU fallback when GPU unavailable

    let gpu_manager = setup_gpu_manager();
    let device = gpu_manager.device();

    println!("[CPU Fallback Test]");
    println!("  Device: {:?}", device);

    // GPU manager should always succeed (falls back to CPU if no CUDA)
    match device {
        candle_core::Device::Cpu => {
            println!("  Status: Running on CPU (CUDA not available)");
        }
        candle_core::Device::Cuda(_) => {
            println!("  Status: Running on GPU (CUDA available)");
        }
        _ => {
            panic!("Unexpected device type");
        }
    }

    // Verify device is usable
    let result = gpu_manager.warmup();
    assert!(result.is_ok(), "Warmup should succeed on any device");
}

#[tokio::test]
#[ignore] // GPU-only test
async fn test_thermal_throttling_detection() {
    // Test thermal warning/error detection (mock high temp)

    let gpu_manager = setup_gpu_manager();

    // Check thermal throttling
    match gpu_manager.check_thermal_throttling() {
        Ok(is_throttling) => {
            println!("[Thermal Test]");
            println!("  Thermal throttling: {}", is_throttling);

            // Verify not throttling under normal conditions
            assert!(
                !is_throttling,
                "GPU should not be thermal throttling under normal test conditions"
            );
        }
        Err(e) => {
            println!("[Thermal Test] Could not check thermal throttling: {}", e);
            // Not a failure - some systems don't expose thermal info
        }
    }
}

#[tokio::test]
#[ignore] // GPU-only test
async fn test_oom_handling() {
    // Test OOM detection and recovery

    use candle_core::{Device, Tensor};

    let gpu_manager = setup_gpu_manager();
    let device = gpu_manager.device();

    if !matches!(device, Device::Cuda(_)) {
        println!("[OOM Test] Skipping - CUDA not available");
        return;
    }

    // Try to allocate unreasonably large tensor (should fail)
    let result = Tensor::zeros(
        &[100_000, 100_000], // 10 billion floats = 40GB
        candle_core::DType::F32,
        device,
    );

    println!("[OOM Test]");
    match result {
        Ok(_) => {
            println!("  WARNING: Large allocation succeeded (unexpected)");
        }
        Err(e) => {
            println!("  Successfully caught OOM error: {}", e);
            // This is the expected path
        }
    }

    // Verify GPU is still functional after OOM
    let recovery_result = Tensor::zeros(&[100, 100], candle_core::DType::F32, device);

    assert!(
        recovery_result.is_ok(),
        "GPU should recover after OOM error"
    );
}

#[tokio::test]
async fn test_invalid_data_handling() {
    // Test data loader with corrupted/invalid data

    // Create invalid market data point
    let invalid_bar = MarketDataPoint {
        timestamp: 0,
        symbol: "TEST".to_string(),
        open: 100.0,
        high: 50.0,  // Invalid: high < low
        low: 120.0,  // Invalid: low > high
        close: 110.0,
        volume: -10.0, // Invalid: negative volume
    };

    // Verify validation catches errors
    assert!(
        !invalid_bar.is_valid(),
        "Invalid data should be detected by validation"
    );

    // Test with NaN values
    let nan_bar = MarketDataPoint {
        timestamp: 0,
        symbol: "TEST".to_string(),
        open: f64::NAN,
        high: 100.0,
        low: 90.0,
        close: 95.0,
        volume: 1000.0,
    };

    assert!(
        !nan_bar.is_valid(),
        "NaN values should be detected by validation"
    );

    // Test with inconsistent OHLC
    let inconsistent_bar = MarketDataPoint {
        timestamp: 0,
        symbol: "TEST".to_string(),
        open: 100.0,
        high: 105.0,
        low: 95.0,
        close: 110.0, // Invalid: close > high
        volume: 1000.0,
    };

    assert!(
        !inconsistent_bar.is_valid(),
        "Inconsistent OHLC should be detected"
    );

    println!("[Invalid Data Test]");
    println!("  ✓ High < Low detected");
    println!("  ✓ Negative volume detected");
    println!("  ✓ NaN values detected");
    println!("  ✓ Close > High detected");
}

// ============================================================================
// Section 6: Performance Tests (2 tests)
// ============================================================================

#[tokio::test]
#[ignore] // Performance validation
async fn test_memory_profiler_overhead() {
    // Verify memory profiler adds <10ms overhead per snapshot

    use std::time::Instant;

    let mut profiler = MemoryProfiler::new(0);

    // Measure snapshot time
    let iterations = 100;
    let start = Instant::now();

    for _ in 0..iterations {
        profiler.take_snapshot().ok(); // Ignore errors for benchmark
    }

    let elapsed = start.elapsed();
    let avg_per_snapshot = elapsed.as_micros() as f64 / iterations as f64;

    println!("[Memory Profiler Overhead Test]");
    println!("  Total snapshots: {}", iterations);
    println!("  Total time: {:.2}ms", elapsed.as_millis());
    println!("  Average per snapshot: {:.2}μs", avg_per_snapshot);

    // Verify overhead is <10ms (10,000μs) per snapshot
    assert!(
        avg_per_snapshot < 10_000.0,
        "Memory profiler overhead should be <10ms per snapshot, got {:.2}μs",
        avg_per_snapshot
    );
}

#[tokio::test]
#[ignore] // Performance validation
async fn test_statistical_sampler_performance() {
    // Verify statistics computation <1ms for 20 samples

    use std::time::Instant;

    let mut sampler = StatisticalSampler::new(0); // No warmup for this test

    // Record 20 samples (in seconds)
    for i in 0..20 {
        sampler.add_sample(0.100 + (i as f64 * 0.0005)); // ~100ms with small increments
    }

    // Measure statistics computation time
    let start = Instant::now();
    let _stats = sampler.compute_statistics();
    let elapsed = start.elapsed();

    println!("[Statistical Sampler Performance Test]");
    println!("  Computation time: {:.2}μs", elapsed.as_micros());

    // Verify computation is <1ms (1000μs)
    assert!(
        elapsed.as_micros() < 1000,
        "Statistics computation should be <1ms, got {}μs",
        elapsed.as_micros()
    );
}