katgpt-types 0.2.0

use super::*;

#[test]
fn test_with_overrides_none_unchanged() {
    let config = Config::draft();
    let overrides = InferenceOverrides::default();
    let original_tb = config.tree_budget;
    let original_temp = config.temperature;
    let original_dl = config.draft_lookahead;
    let result = config.with_overrides(&overrides);
    assert_eq!(result.tree_budget, original_tb);
    assert_eq!(result.temperature, original_temp);
    assert_eq!(result.draft_lookahead, original_dl);
}

#[test]
fn test_with_overrides_some_applied() {
    let config = Config::draft();
    let overrides = InferenceOverrides {
        tree_budget: Some(99),
        temperature: Some(0.123),
        ..Default::default()
    };
    let original_dl = config.draft_lookahead;
    let result = config.with_overrides(&overrides);
    assert_eq!(result.tree_budget, 99);
    assert!((result.temperature - 0.123).abs() < 1e-6);
    // Non-overridden fields stay the same
    assert_eq!(result.draft_lookahead, original_dl);
}

#[test]
fn test_with_overrides_all_fields() {
    let config = Config::draft();
    let overrides = InferenceOverrides {
        tree_budget: Some(1),
        draft_lookahead: Some(2),
        parallel_threshold: Some(3),
        screening_threshold: Some(0.1),
        temperature: Some(0.2),
        sparse_threshold: Some(0.3),
        early_exit_patience: Some(4),
        early_exit_gap: Some(0.4),
        mtp_activation_threshold: Some(5),
        mtp_cluster_vocab_threshold: Some(6),
        mtp_shared_kv_prompt_threshold: Some(7),
        mtp_cluster_size: Some(8),
        mtp_min_output_tokens: Some(10),
        mtp_cluster_topk: Some(2),
        sp_kv_threshold: Some(0.5),
        width_rollouts: Some(9),
        early_stop_threshold: Some(0.6),
        convergence_selector: Some(ConvergenceSelector::Top1Converged),
        mls_layers: Some(3),
        // drafter_lora_path is consumed by the caller, not applied to Config
        drafter_lora_path: None,
        max_plan_horizon: Some(5),
        #[cfg(feature = "hydra_budget")]
        hydra_skip_threshold: None,
        #[cfg(feature = "hydra_budget")]
        hydra_skip_erasure_draft: None,
        depth_tier: None,
    };
    let result = config.with_overrides(&overrides);
    assert_eq!(result.tree_budget, 1);
    assert_eq!(result.draft_lookahead, 2);
    assert_eq!(result.parallel_threshold, 3);
    assert!((result.screening_threshold - 0.1).abs() < 1e-6);
    assert!((result.temperature - 0.2).abs() < 1e-6);
    assert!((result.sparse_threshold - 0.3).abs() < 1e-6);
    assert_eq!(result.early_exit_patience, 4);
    assert!((result.early_exit_gap - 0.4).abs() < 1e-6);
    assert_eq!(result.mtp_activation_threshold, 5);
    assert_eq!(result.mtp_cluster_vocab_threshold, 6);
    assert_eq!(result.mtp_shared_kv_prompt_threshold, 7);
    assert_eq!(result.mtp_cluster_size, 8);
    assert_eq!(result.mtp_min_output_tokens, 10);
    assert_eq!(result.mtp_cluster_topk, 2);
    assert!((result.sp_kv_threshold - 0.5).abs() < 1e-6);
    assert_eq!(result.width_rollouts, 9);
    assert!((result.early_stop_threshold - 0.6).abs() < 1e-6);
    assert_eq!(
        result.convergence_selector,
        ConvergenceSelector::Top1Converged
    );
    assert_eq!(result.mls_layers, 3);
    // max_plan_horizon caps draft_lookahead when lower (Plan 112 T11)
    assert_eq!(result.draft_lookahead, 2); // 2 (from override) < 5 (horizon cap), stays 2
}

#[test]
fn test_early_exit_defaults_disabled() {
    let config = Config::micro();
    assert_eq!(config.early_exit_patience, 0);
    assert!((config.early_exit_gap).abs() < 1e-6);
}

#[test]
#[cfg(feature = "domain_latent")]
fn test_domain_latent_save_load_roundtrip() {
    let tmp = std::env::temp_dir().join("katgpt_core_test_domain_latent.bin");
    let original = DomainLatent::from_vec(vec![1.0, 2.0, 3.0, 4.0]);
    original.save(&tmp).unwrap();
    let loaded = DomainLatent::load(&tmp).unwrap();
    assert_eq!(original.embedding, loaded.embedding);
    drop(std::fs::remove_file(&tmp));
}

#[test]
#[cfg(feature = "domain_latent")]
fn test_domain_latent_zeros() {
    let dl = DomainLatent::zeros(8);
    assert_eq!(dl.embedding.len(), 8);
    assert!(dl.embedding.iter().all(|&v| v == 0.0));
}

#[test]
#[cfg(feature = "domain_latent")]
fn test_domain_latent_invalid_magic() {
    let tmp = std::env::temp_dir().join("katgpt_core_test_bad_magic.bin");
    let mut buf = b"XXXX".to_vec();
    buf.push(1); // version
    buf.extend_from_slice(&4u32.to_le_bytes()); // kv_dim
    buf.extend_from_slice(
        &[
            0.0f32.to_le_bytes(),
            0.0f32.to_le_bytes(),
            0.0f32.to_le_bytes(),
            0.0f32.to_le_bytes(),
        ]
        .concat(),
    );
    let hash = blake3::hash(&buf);
    buf.extend_from_slice(hash.as_bytes());
    std::fs::write(&tmp, &buf).unwrap();
    assert!(DomainLatent::load(&tmp).is_err());
    drop(std::fs::remove_file(&tmp));
}

#[test]
#[cfg(feature = "domain_latent")]
fn test_domain_latent_checksum_mismatch() {
    let tmp = std::env::temp_dir().join("katgpt_core_test_bad_checksum.bin");
    let mut buf = b"DLAT".to_vec();
    buf.push(1); // version
    buf.extend_from_slice(&4u32.to_le_bytes()); // kv_dim
    buf.extend_from_slice(&[0.0f32.to_le_bytes(); 4].concat());
    buf.extend_from_slice(&[0u8; 32]); // wrong checksum
    std::fs::write(&tmp, &buf).unwrap();
    assert!(DomainLatent::load(&tmp).is_err());
    drop(std::fs::remove_file(&tmp));
}

#[test]
#[cfg(feature = "domain_latent")]
fn test_domain_latent_file_too_small() {
    let tmp = std::env::temp_dir().join("katgpt_core_test_too_small.bin");
    std::fs::write(&tmp, b"DLAT").unwrap();
    assert!(DomainLatent::load(&tmp).is_err());
    drop(std::fs::remove_file(&tmp));
}

#[test]
fn test_config_game() {
    let config = Config::game();
    assert_eq!(config.vocab_size, 10);
    assert_eq!(config.block_size, 170);
    assert_eq!(config.n_embd, 32);
    assert_eq!(config.n_head, 4);
    assert_eq!(config.head_dim, 8);
    assert_eq!(config.mlp_hidden, 128);
    assert!(config.validate().is_ok());
}

#[test]
#[cfg(feature = "gpart_adapter")]
fn test_gpart_isometry() {
    let adapter = GpartAdapter {
        d: 4,
        seed: 42,
        theta: vec![1.0, 2.0, 3.0, 4.0],
    };
    assert!(adapter.check_isometry(256));
}

#[test]
#[cfg(feature = "gpart_adapter")]
fn test_gpart_apply_correctness() {
    let adapter = GpartAdapter {
        d: 2,
        seed: 123,
        theta: vec![1.0, -1.0],
    };
    let mut weights = vec![0.0f32; 8];
    adapter.apply(&mut weights);
    // Every weight should have changed (each element assigned to group 0 or 1)
    assert!(weights.iter().all(|&w| w != 0.0));
    // Verify deterministic: apply again with same adapter should produce same result
    let mut weights2 = vec![0.0f32; 8];
    adapter.apply(&mut weights2);
    assert_eq!(weights, weights2);
}

#[test]
#[cfg(feature = "gpart_adapter")]
fn test_gpart_commitment_roundtrip() {
    let adapter = GpartAdapter {
        d: 4,
        seed: 99,
        theta: vec![0.5, -0.5, 1.0, -1.0],
    };
    let commit = adapter.commitment();
    assert!(adapter.verify(&commit));
}

#[test]
#[cfg(feature = "gpart_adapter")]
fn test_gpart_tamper_detection() {
    let adapter = GpartAdapter {
        d: 4,
        seed: 99,
        theta: vec![0.5, -0.5, 1.0, -1.0],
    };
    let commit = adapter.commitment();
    // Tampered commitment
    let mut tampered = commit;
    tampered[0] ^= 0xFF;
    assert!(!adapter.verify(&tampered));
    // Different theta produces different commitment
    let tampered_adapter = GpartAdapter {
        d: 4,
        seed: 99,
        theta: vec![0.5, -0.5, 1.0, 0.0], // last element changed
    };
    assert_ne!(adapter.commitment(), tampered_adapter.commitment());
    // Different seed produces different commitment
    let seed_adapter = GpartAdapter {
        d: 4,
        seed: 100,
        theta: vec![0.5, -0.5, 1.0, -1.0],
    };
    assert_ne!(adapter.commitment(), seed_adapter.commitment());
}

#[test]
#[cfg(feature = "gpart_adapter")]
fn test_gpart_save_load_roundtrip() {
    let tmp = std::env::temp_dir().join("katgpt_core_test_gpart.bin");
    let adapter = GpartAdapter {
        d: 8,
        seed: 42,
        theta: vec![1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0],
    };
    adapter.save(&tmp).unwrap();
    let loaded = GpartAdapter::load(&tmp).unwrap();
    assert_eq!(loaded.d, adapter.d);
    assert_eq!(loaded.seed, adapter.seed);
    assert_eq!(loaded.theta, adapter.theta);
    drop(std::fs::remove_file(&tmp));
}

#[test]
#[cfg(feature = "gpart_adapter")]
fn test_gpart_determinism_same_seed() {
    let adapter = GpartAdapter {
        d: 4,
        seed: 42,
        theta: vec![1.0, 2.0, 3.0, 4.0],
    };
    let mut w1 = vec![0.0f32; 64];
    let mut w2 = vec![0.0f32; 64];
    adapter.apply(&mut w1);
    adapter.apply(&mut w2);
    assert_eq!(w1, w2);
}

#[test]
#[cfg(feature = "gpart_pruning")]
fn test_gpart_masked_all_true_matches_unmasked() {
    // When group_mask is all-true, masked apply must equal unmasked apply.
    let adapter = GpartAdapter {
        d: 4,
        seed: 42,
        theta: vec![1.0, 2.0, 3.0, 4.0],
    };
    let n = 256;
    let mut w_ref = vec![0.0f32; n];
    let mut assignments = vec![0usize; n];
    let mut group_sizes = vec![0usize; adapter.d];
    adapter.apply_with_scratch(&mut w_ref, &mut assignments, &mut group_sizes);

    let mut w_masked = vec![0.0f32; n];
    let mask = vec![true; adapter.d];
    adapter.apply_with_scratch_masked(&mut w_masked, &mut assignments, &mut group_sizes, &mask);

    for (i, (a, b)) in w_ref.iter().zip(w_masked.iter()).enumerate() {
        assert_eq!(a, b, "mismatch at index {i}");
    }
}

#[test]
#[cfg(feature = "gpart_pruning")]
fn test_gpart_masked_zeroes_masked_groups() {
    // If group g is masked off, NO base_weights element assigned to g
    // should receive any delta from theta[g]. Concretely: the masked
    // delta for elements of masked groups must be exactly zero.
    let adapter = GpartAdapter {
        d: 4,
        seed: 7,
        theta: vec![1.0, 2.0, 3.0, 4.0],
    };
    let n = 512;
    let mut w_zero = vec![0.0f32; n]; // baseline: all groups masked off
    let mut assignments = vec![0usize; n];
    let mut group_sizes = vec![0usize; adapter.d];
    let all_false = vec![false; adapter.d];
    adapter.apply_with_scratch_masked(&mut w_zero, &mut assignments, &mut group_sizes, &all_false);
    // With every group masked, weights should remain exactly zero.
    assert!(w_zero.iter().all(|&w| w == 0.0));
}

#[test]
#[cfg(feature = "gpart_pruning")]
fn test_gpart_masked_single_group_preserves_others() {
    // Masking group 0 only: every element in groups 1..d should still receive
    // its full delta, while elements in group 0 should stay at baseline.
    let adapter = GpartAdapter {
        d: 4,
        seed: 7,
        theta: vec![1.0, 2.0, 3.0, 4.0],
    };
    let n = 512;

    // Full (unmasked) reference.
    let mut w_full = vec![0.0f32; n];
    let mut assignments_full = vec![0usize; n];
    let mut group_sizes = vec![0usize; adapter.d];
    adapter.apply_with_scratch(&mut w_full, &mut assignments_full, &mut group_sizes);

    // Masked: drop group 0.
    let mut w_masked = vec![0.0f32; n];
    let mut assignments = vec![0usize; n];
    let mut mask = vec![true; adapter.d];
    mask[0] = false;
    adapter.apply_with_scratch_masked(&mut w_masked, &mut assignments, &mut group_sizes, &mask);

    // assignments[] is regenerated deterministically from seed, so it matches
    // between the two calls — compare per-element using the masked call's
    // assignment vector.
    for i in 0..n {
        let g = assignments[i];
        if g == 0 {
            // Group 0 masked off → masked output should equal baseline (0.0).
            assert_eq!(w_masked[i], 0.0, "group-0 element {i} should be zero");
        } else {
            // Other groups untouched → should equal full reference.
            assert_eq!(
                w_masked[i], w_full[i],
                "group-{g} element {i} should match unmasked"
            );
        }
    }
}

#[test]
#[cfg(feature = "gpart_pruning")]
fn test_gpart_topk_mask_edge_cases() {
    let adapter = GpartAdapter {
        d: 4,
        seed: 0,
        theta: vec![0.1, 0.5, 0.3, 0.9],
    };
    // k >= d → all active
    assert!(adapter.topk_mask(adapter.d).iter().all(|&b| b));
    assert!(adapter.topk_mask(adapter.d + 10).iter().all(|&b| b));
    // k == 0 → all inactive
    assert!(adapter.topk_mask(0).iter().all(|&b| !b));
}

#[test]
#[cfg(feature = "gpart_pruning")]
fn test_gpart_topk_mask_selects_largest_magnitudes() {
    // θ = [0.1, 0.5, -0.3, 0.9] → |θ| = [0.1, 0.5, 0.3, 0.9]
    // top-2 by magnitude should keep groups {3 (0.9), 1 (0.5)}.
    let adapter = GpartAdapter {
        d: 4,
        seed: 0,
        theta: vec![0.1, 0.5, -0.3, 0.9],
    };
    let mask = adapter.topk_mask(2);
    assert_eq!(mask.len(), 4);
    assert!(mask[1] && mask[3], "groups 1 and 3 must be in top-2");
    assert!(!mask[0] && !mask[2], "groups 0 and 2 must be pruned");
    // Exactly 2 active.
    assert_eq!(mask.iter().filter(|&&b| b).count(), 2);
}

// ── Issue 299: multi-adapter LoRA loading ──────────────────────

/// Build a Plan 008 LoRA binary file with `n` adapters for testing.
/// Each adapter's A/B matrices are filled with a distinct sentinel pattern
/// so we can verify the loader returns the right data per adapter.
fn build_test_lora_file(path: &std::path::Path, n_adapters: usize, rank: usize) {
    let in_dim = 4usize;
    let out_dim = 4usize;
    let alpha = 8.0f32;

    // Payload: [n_adapters(4)][rank(4)][alpha(4)] + per-adapter [in_dim(4)][out_dim(4)][A][B]
    let header = 4 + 4 + 4;
    let per_adapter = 4 + 4 + (rank * in_dim + out_dim * rank) * std::mem::size_of::<f32>();
    let mut payload = Vec::with_capacity(header + n_adapters * per_adapter);

    payload.extend_from_slice(&(n_adapters as u32).to_le_bytes());
    payload.extend_from_slice(&(rank as u32).to_le_bytes());
    payload.extend_from_slice(&alpha.to_le_bytes());

    for i in 0..n_adapters {
        payload.extend_from_slice(&(in_dim as u32).to_le_bytes());
        payload.extend_from_slice(&(out_dim as u32).to_le_bytes());
        // A: [rank × in_dim], sentinel = (i + 1) * 0.1
        let a_sentinel = (i as f32 + 1.0) * 0.1;
        for _ in 0..(rank * in_dim) {
            payload.extend_from_slice(&a_sentinel.to_le_bytes());
        }
        // B: [out_dim × rank], sentinel = (i + 1) * -0.2
        let b_sentinel = (i as f32 + 1.0) * -0.2;
        for _ in 0..(out_dim * rank) {
            payload.extend_from_slice(&b_sentinel.to_le_bytes());
        }
    }

    let checksum = blake3::hash(&payload);
    let mut file_data = Vec::with_capacity(4 + 4 + 32 + payload.len());
    file_data.extend_from_slice(b"LORA");
    file_data.extend_from_slice(&1u32.to_le_bytes());
    file_data.extend_from_slice(checksum.as_bytes());
    file_data.extend_from_slice(&payload);
    std::fs::write(path, &file_data).unwrap();
}

#[test]
fn test_lora_load_single_adapter_returns_one_element_vec() {
    let tmp = std::env::temp_dir().join("katgpt_core_test_lora_single.bin");
    build_test_lora_file(&tmp, 1, 4);
    let adapters = LoraAdapter::load(&tmp).expect("single-adapter load should succeed");
    assert_eq!(
        adapters.len(),
        1,
        "single-adapter file must yield 1-element Vec"
    );
    assert_eq!(adapters[0].rank, 4);
    assert_eq!(adapters[0].in_dim, 4);
    assert_eq!(adapters[0].out_dim, 4);
    assert!((adapters[0].alpha - 8.0).abs() < 1e-6);
    drop(std::fs::remove_file(&tmp));
}

/// Issue 299 regression: a 12-adapter file (L2 model, 6 adapters/layer × 2 layers)
/// must load ALL 12 adapters. Previously `load()` returned only the first.
#[test]
fn test_lora_load_multi_adapter_returns_all() {
    let tmp = std::env::temp_dir().join("katgpt_core_test_lora_multi.bin");
    build_test_lora_file(&tmp, 12, 4);
    let adapters = LoraAdapter::load(&tmp).expect("12-adapter load should succeed");
    assert_eq!(
        adapters.len(),
        12,
        "12-adapter file must yield 12-element Vec (Issue 299 regression)"
    );

    // Verify each adapter has its own sentinel data — proves layers 1..11 aren't dropped
    for (i, a) in adapters.iter().enumerate() {
        let expected_a = (i as f32 + 1.0) * 0.1;
        let expected_b = (i as f32 + 1.0) * -0.2;
        assert_eq!(a.rank, 4, "adapter {i} rank");
        assert_eq!(a.in_dim, 4, "adapter {i} in_dim");
        assert_eq!(a.out_dim, 4, "adapter {i} out_dim");
        assert_eq!(a.a.len(), 4 * 4, "adapter {i} A matrix size");
        assert_eq!(a.b.len(), 4 * 4, "adapter {i} B matrix size");
        assert!(
            a.a.iter().all(|&v| (v - expected_a).abs() < 1e-6),
            "adapter {i} A data mismatch"
        );
        assert!(
            a.b.iter().all(|&v| (v - expected_b).abs() < 1e-6),
            "adapter {i} B data mismatch"
        );
    }
    drop(std::fs::remove_file(&tmp));
}

#[test]
fn test_lora_load_first_returns_only_first_adapter() {
    let tmp = std::env::temp_dir().join("katgpt_core_test_lora_first.bin");
    build_test_lora_file(&tmp, 6, 4);
    let first = LoraAdapter::load_first(&tmp).expect("load_first should succeed");
    assert_eq!(first.rank, 4);
    // Adapter 0 sentinel is 1.0 * 0.1 = 0.1
    assert!(first.a.iter().all(|&v| (v - 0.1).abs() < 1e-6));
    assert!(first.b.iter().all(|&v| (v - (-0.2)).abs() < 1e-6));
    drop(std::fs::remove_file(&tmp));
}

#[test]
fn test_lora_load_rejects_truncated_multi_adapter() {
    // Truncate a 12-adapter file so only 6 adapters' bytes fit — loader must error
    let tmp = std::env::temp_dir().join("katgpt_core_test_lora_truncated.bin");
    build_test_lora_file(&tmp, 12, 4);
    let full = std::fs::read(&tmp).unwrap();
    // Cut the payload in half (keep header + checksum + ~half the adapters)
    let cut = full.len() / 2 + 20; // keep enough to pass header check
    std::fs::write(&tmp, &full[..cut]).unwrap();
    // Re-generate checksum won't match — but if it did, the truncation check fires.
    // Either way, load must fail (not silently return wrong number of adapters).
    let result = LoraAdapter::load(&tmp);
    assert!(
        result.is_err(),
        "truncated multi-adapter file must error, not silently drop adapters"
    );
    drop(std::fs::remove_file(&tmp));
}

#[test]
fn test_lora_load_rejects_bad_magic() {
    let tmp = std::env::temp_dir().join("katgpt_core_test_lora_bad_magic.bin");
    std::fs::write(&tmp, b"NOPE\0\0\0\0").unwrap();
    assert!(LoraAdapter::load(&tmp).is_err());
    drop(std::fs::remove_file(&tmp));
}

#[test]
fn test_lora_load_rejects_zero_adapters() {
    let tmp = std::env::temp_dir().join("katgpt_core_test_lora_zero.bin");
    // Build a valid-checksum file that declares 0 adapters
    let mut payload = Vec::new();
    payload.extend_from_slice(&0u32.to_le_bytes()); // n_adapters = 0
    payload.extend_from_slice(&4u32.to_le_bytes()); // rank
    payload.extend_from_slice(&8.0f32.to_le_bytes()); // alpha
    let checksum = blake3::hash(&payload);
    let mut buf = Vec::new();
    buf.extend_from_slice(b"LORA");
    buf.extend_from_slice(&1u32.to_le_bytes());
    buf.extend_from_slice(checksum.as_bytes());
    buf.extend_from_slice(&payload);
    std::fs::write(&tmp, &buf).unwrap();
    assert!(
        LoraAdapter::load(&tmp).is_err(),
        "zero-adapter declaration must be rejected"
    );
    drop(std::fs::remove_file(&tmp));
}