foxhunt/ml/tests/ppo_hyperopt_backward_compat_test.rs

//! PPO Hyperopt Backward Compatibility Tests
//!
//! This test suite ensures that the addition of `minibatch_size` as the 6th parameter
//! to PPOParams does not break existing functionality. It covers:
//!
//! 1. Default value validation (minibatch_size = 128)
//! 2. Full serialization/deserialization roundtrip
//! 3. Graceful handling of missing fields (Serde Default trait)
//! 4. Old 5-parameter continuous arrays rejected with clear errors
//! 5. New 6-parameter continuous arrays accepted and validated
//!
//! ## Migration Path
//!
//! - **Old checkpoints**: Use Default trait fallback (minibatch_size = 128)
//! - **Old hyperopt results**: Must be re-run with 6 parameters
//! - **Production configs**: Update TOML/JSON to include minibatch_size

use ml::hyperopt::adapters::ppo::PPOParams;
use ml::hyperopt::traits::ParameterSpace;

#[test]
fn test_default_ppo_params_has_reasonable_minibatch_size() {
    let params = PPOParams::default();

    // Verify default minibatch_size is production-tested value
    assert_eq!(
        params.minibatch_size, 128,
        "Default minibatch_size should be 128 (production-tested)"
    );

    // Verify it's within VRAM bounds (64-230 for RTX 3050 Ti)
    assert!(
        params.minibatch_size >= 64,
        "minibatch_size {} should be >= 64 (VRAM lower bound)",
        params.minibatch_size
    );
    assert!(
        params.minibatch_size <= 230,
        "minibatch_size {} should be <= 230 (VRAM upper bound)",
        params.minibatch_size
    );
}

#[test]
fn test_ppo_params_serialization_includes_all_fields() {
    let params = PPOParams {
        policy_learning_rate: 1e-6,
        value_learning_rate: 0.001,
        clip_epsilon: 0.2,
        value_loss_coeff: 1.0,
        entropy_coeff: 0.05,
        minibatch_size: 128,
    };

    // Serialize to JSON
    let json = serde_json::to_string(&params).expect("Serialization should succeed");

    // Verify JSON contains minibatch_size field
    assert!(
        json.contains("minibatch_size"),
        "Serialized JSON should contain 'minibatch_size' field. Got: {}",
        json
    );
    assert!(
        json.contains("128"),
        "Serialized JSON should contain minibatch_size value (128). Got: {}",
        json
    );

    // Verify all 6 fields are present
    let field_count = json.matches("\":").count();
    assert_eq!(
        field_count, 6,
        "Serialized JSON should have 6 fields (5 old + 1 new). Got: {}",
        field_count
    );
}

#[test]
fn test_ppo_params_deserialization_roundtrip() {
    let original = PPOParams {
        policy_learning_rate: 1e-6,
        value_learning_rate: 0.001,
        clip_epsilon: 0.1126,
        value_loss_coeff: 0.5,
        entropy_coeff: 0.006142,
        minibatch_size: 192,
    };

    // Serialize to JSON
    let json = serde_json::to_string(&original).expect("Serialization should succeed");

    // Deserialize back
    let deserialized: PPOParams =
        serde_json::from_str(&json).expect("Deserialization should succeed");

    // Verify all fields match (within floating-point tolerance)
    assert!(
        (deserialized.policy_learning_rate - original.policy_learning_rate).abs() < 1e-10,
        "policy_learning_rate mismatch"
    );
    assert!(
        (deserialized.value_learning_rate - original.value_learning_rate).abs() < 1e-10,
        "value_learning_rate mismatch"
    );
    assert!(
        (deserialized.clip_epsilon - original.clip_epsilon).abs() < 1e-10,
        "clip_epsilon mismatch"
    );
    assert!(
        (deserialized.value_loss_coeff - original.value_loss_coeff).abs() < 1e-10,
        "value_loss_coeff mismatch"
    );
    assert!(
        (deserialized.entropy_coeff - original.entropy_coeff).abs() < 1e-10,
        "entropy_coeff mismatch"
    );
    assert_eq!(
        deserialized.minibatch_size, original.minibatch_size,
        "minibatch_size mismatch"
    );
}

#[test]
fn test_ppo_params_handles_missing_minibatch_size_gracefully() {
    // Simulate old JSON format (5 fields, missing minibatch_size)
    let old_json = r#"{
        "policy_learning_rate": 0.000001,
        "value_learning_rate": 0.001,
        "clip_epsilon": 0.2,
        "value_loss_coeff": 1.0,
        "entropy_coeff": 0.05
    }"#;

    // Should deserialize successfully using Default trait
    let params: PPOParams = serde_json::from_str(old_json)
        .expect("Deserialization should succeed with missing minibatch_size field");

    // Verify missing field uses default value (128)
    assert_eq!(
        params.minibatch_size, 128,
        "Missing minibatch_size should default to 128"
    );

    // Verify other fields are correct
    assert!((params.policy_learning_rate - 1e-6).abs() < 1e-10);
    assert!((params.value_learning_rate - 0.001).abs() < 1e-10);
    assert!((params.clip_epsilon - 0.2).abs() < 1e-10);
    assert!((params.value_loss_coeff - 1.0).abs() < 1e-10);
    assert!((params.entropy_coeff - 0.05).abs() < 1e-10);
}

#[test]
fn test_old_5_parameter_continuous_array_rejected() {
    // Old hyperopt results with 5 parameters (pre-minibatch_size)
    let old_continuous = vec![
        1e-6_f64.ln(),  // policy_learning_rate (log scale)
        0.001_f64.ln(), // value_learning_rate (log scale)
        0.2,            // clip_epsilon
        1.0,            // value_loss_coeff
        0.05_f64.ln(),  // entropy_coeff (log scale)
    ];

    // Should fail with clear error message
    let result = PPOParams::from_continuous(&old_continuous);

    assert!(
        result.is_err(),
        "Old 5-parameter array should be rejected, but got: {:?}",
        result
    );

    // Verify error message is clear
    let error_msg = result.unwrap_err().to_string();
    assert!(
        error_msg.contains("Expected 6 parameters") || error_msg.contains("got 5"),
        "Error message should clearly state version mismatch. Got: {}",
        error_msg
    );
}

#[test]
fn test_new_6_parameter_continuous_array_accepted() {
    // New hyperopt results with 6 parameters (includes minibatch_size)
    let new_continuous = vec![
        1e-6_f64.ln(),  // policy_learning_rate (log scale)
        0.001_f64.ln(), // value_learning_rate (log scale)
        0.2,            // clip_epsilon
        1.0,            // value_loss_coeff
        0.05_f64.ln(),  // entropy_coeff (log scale)
        128.0,          // minibatch_size (linear scale)
    ];

    // Should succeed
    let result = PPOParams::from_continuous(&new_continuous);

    assert!(
        result.is_ok(),
        "New 6-parameter array should be accepted, but got error: {:?}",
        result.err()
    );

    let params = result.unwrap();

    // Verify all parameters are correctly extracted
    assert!(
        (params.policy_learning_rate - 1e-6).abs() < 1e-9,
        "policy_learning_rate mismatch: expected 1e-6, got {}",
        params.policy_learning_rate
    );
    assert!(
        (params.value_learning_rate - 0.001).abs() < 1e-9,
        "value_learning_rate mismatch: expected 0.001, got {}",
        params.value_learning_rate
    );
    assert!(
        (params.clip_epsilon - 0.2).abs() < 1e-9,
        "clip_epsilon mismatch: expected 0.2, got {}",
        params.clip_epsilon
    );
    assert!(
        (params.value_loss_coeff - 1.0).abs() < 1e-9,
        "value_loss_coeff mismatch: expected 1.0, got {}",
        params.value_loss_coeff
    );
    assert!(
        (params.entropy_coeff - 0.05).abs() < 1e-9,
        "entropy_coeff mismatch: expected 0.05, got {}",
        params.entropy_coeff
    );
    assert_eq!(
        params.minibatch_size, 128,
        "minibatch_size mismatch: expected 128, got {}",
        params.minibatch_size
    );
}

#[test]
fn test_minibatch_size_boundary_values() {
    // Test lower bound (64)
    let lower_bound = vec![
        1e-6_f64.ln(),  // policy_learning_rate
        0.001_f64.ln(), // value_learning_rate
        0.2,            // clip_epsilon
        1.0,            // value_loss_coeff
        0.05_f64.ln(),  // entropy_coeff
        64.0,           // minibatch_size (lower bound)
    ];

    let params_lower = PPOParams::from_continuous(&lower_bound).unwrap();
    assert_eq!(
        params_lower.minibatch_size, 64,
        "Lower bound (64) should be preserved"
    );

    // Test upper bound (230)
    let upper_bound = vec![
        1e-6_f64.ln(),  // policy_learning_rate
        0.001_f64.ln(), // value_learning_rate
        0.2,            // clip_epsilon
        1.0,            // value_loss_coeff
        0.05_f64.ln(),  // entropy_coeff
        230.0,          // minibatch_size (upper bound)
    ];

    let params_upper = PPOParams::from_continuous(&upper_bound).unwrap();
    assert_eq!(
        params_upper.minibatch_size, 230,
        "Upper bound (230) should be preserved"
    );
}

#[test]
fn test_minibatch_size_clamping() {
    // Test below lower bound (should clamp to 64)
    let below_bound = vec![
        1e-6_f64.ln(),  // policy_learning_rate
        0.001_f64.ln(), // value_learning_rate
        0.2,            // clip_epsilon
        1.0,            // value_loss_coeff
        0.05_f64.ln(),  // entropy_coeff
        32.0,           // minibatch_size (below lower bound)
    ];

    let params_below = PPOParams::from_continuous(&below_bound).unwrap();
    assert_eq!(
        params_below.minibatch_size, 64,
        "Values below 64 should clamp to 64"
    );

    // Test above upper bound (should clamp to 230)
    let above_bound = vec![
        1e-6_f64.ln(),  // policy_learning_rate
        0.001_f64.ln(), // value_learning_rate
        0.2,            // clip_epsilon
        1.0,            // value_loss_coeff
        0.05_f64.ln(),  // entropy_coeff
        512.0,          // minibatch_size (above upper bound)
    ];

    let params_above = PPOParams::from_continuous(&above_bound).unwrap();
    assert_eq!(
        params_above.minibatch_size, 230,
        "Values above 230 should clamp to 230"
    );
}

#[test]
fn test_to_continuous_roundtrip() {
    let original = PPOParams {
        policy_learning_rate: 1e-6,
        value_learning_rate: 0.001,
        clip_epsilon: 0.1126,
        value_loss_coeff: 0.5,
        entropy_coeff: 0.006142,
        minibatch_size: 192,
    };

    // Convert to continuous array
    let continuous = original.to_continuous();

    // Verify length is 6
    assert_eq!(
        continuous.len(),
        6,
        "to_continuous() should return 6 values"
    );

    // Convert back
    let roundtrip = PPOParams::from_continuous(&continuous).unwrap();

    // Verify all fields match (within tolerance)
    assert!(
        (roundtrip.policy_learning_rate - original.policy_learning_rate).abs() < 1e-9,
        "policy_learning_rate roundtrip failed"
    );
    assert!(
        (roundtrip.value_learning_rate - original.value_learning_rate).abs() < 1e-9,
        "value_learning_rate roundtrip failed"
    );
    assert!(
        (roundtrip.clip_epsilon - original.clip_epsilon).abs() < 1e-9,
        "clip_epsilon roundtrip failed"
    );
    assert!(
        (roundtrip.value_loss_coeff - original.value_loss_coeff).abs() < 1e-9,
        "value_loss_coeff roundtrip failed"
    );
    assert!(
        (roundtrip.entropy_coeff - original.entropy_coeff).abs() < 1e-9,
        "entropy_coeff roundtrip failed"
    );
    assert_eq!(
        roundtrip.minibatch_size, original.minibatch_size,
        "minibatch_size roundtrip failed"
    );
}

#[test]
fn test_continuous_array_wrong_length_errors() {
    // Test empty array
    let empty: Vec<f64> = vec![];
    let result_empty = PPOParams::from_continuous(&empty);
    assert!(result_empty.is_err(), "Empty array should be rejected");

    // Test 7-parameter array
    let too_long = vec![
        1e-6_f64.ln(),  // policy_learning_rate
        0.001_f64.ln(), // value_learning_rate
        0.2,            // clip_epsilon
        1.0,            // value_loss_coeff
        0.05_f64.ln(),  // entropy_coeff
        128.0,          // minibatch_size
        999.0,          // extra parameter
    ];

    let result_long = PPOParams::from_continuous(&too_long);
    assert!(result_long.is_err(), "7-parameter array should be rejected");
}