aprender-core 0.30.0

// CONTRACT: tensor-inventory-v1.yaml
// HASH: sha256:b4c5d6e7f8901234
// Generated by: pv probar --binding
// DO NOT EDIT — regenerate with `pv probar --binding`

use proptest::prelude::*;

/// Model config for tensor counting.
#[derive(Debug, Clone)]
struct TensorConfig {
    layers: u32,
    has_bias: bool,
    has_qk_norm: bool,
    num_norms: u32, // typically 2 (pre-attn, pre-ffn)
    tied_embeddings: bool,
}

/// Count tensors per layer.
fn per_layer_tensors(cfg: &TensorConfig) -> u32 {
    let base = 7; // qkv_proj, o_proj, gate_proj, up_proj, down_proj, + 2 norm weights
    let bias = if cfg.has_bias { 3 } else { 0 }; // q,k,v bias or qkv bias
    let qk_norm = if cfg.has_qk_norm { 2 } else { 0 }; // q_norm_weight, k_norm_weight
    let extra_norms = cfg.num_norms.saturating_sub(2) * 1; // extra norm weights beyond 2
    base + bias + qk_norm + extra_norms
}

/// Count total tensors in model.
fn total_tensors(cfg: &TensorConfig) -> u32 {
    let base = if cfg.tied_embeddings { 2 } else { 3 }; // embed + lm_head + final_norm (tied: -1)
    base + cfg.layers * per_layer_tensors(cfg)
}

/// Bytes for given quantization scheme.
fn quant_bytes(params: u64, scheme: &str) -> u64 {
    match scheme {
        "Q4K" => params * 18 / 32,  // 18 bytes per 32-element superblock
        "Q6K" => params * 26 / 32,  // 26 bytes per 32-element block
        "Q8_0" => params * 34 / 32, // 34 bytes per 32-element block
        "F16" => params * 2,
        "F32" => params * 4,
        _ => 0,
    }
}

fn config_strategy() -> impl Strategy<Value = TensorConfig> {
    (
        1u32..100,
        proptest::bool::ANY,
        proptest::bool::ANY,
        2u32..4,
        proptest::bool::ANY,
    )
        .prop_map(
            |(layers, has_bias, has_qk_norm, num_norms, tied)| TensorConfig {
                layers,
                has_bias,
                has_qk_norm,
                num_norms,
                tied_embeddings: tied,
            },
        )
}

proptest! {
    /// Obligation: Tensor count formula (invariant)
    /// Formal: total = base + L * per_layer for valid configs
    #[test]
    fn prop_tensor_count_formula(
        cfg in config_strategy()
    ) {
        let total = total_tensors(&cfg);
        let base = if cfg.tied_embeddings { 2u32 } else { 3u32 };
        let expected = base + cfg.layers * per_layer_tensors(&cfg);
        prop_assert_eq!(
            total, expected,
            "tensor count mismatch: total={}, expected base({}) + {}*{}={}",
            total, base, cfg.layers, per_layer_tensors(&cfg), expected
        );
        // Must be positive
        prop_assert!(total > 0, "tensor count must be > 0, got {}", total);
    }

    /// Obligation: Architecture delta linear (invariant)
    /// Formal: delta(A,B) = L * (per_layer_B - per_layer_A)
    #[test]
    fn prop_architecture_delta(
        layers in 1u32..100,
        a_bias in proptest::bool::ANY,
        b_bias in proptest::bool::ANY,
        a_qk in proptest::bool::ANY,
        b_qk in proptest::bool::ANY
    ) {
        let cfg_a = TensorConfig {
            layers,
            has_bias: a_bias,
            has_qk_norm: a_qk,
            num_norms: 2,
            tied_embeddings: false,
        };
        let cfg_b = TensorConfig {
            layers,
            has_bias: b_bias,
            has_qk_norm: b_qk,
            num_norms: 2,
            tied_embeddings: false,
        };

        let total_a = total_tensors(&cfg_a);
        let total_b = total_tensors(&cfg_b);
        let per_layer_a = per_layer_tensors(&cfg_a) as i64;
        let per_layer_b = per_layer_tensors(&cfg_b) as i64;

        let actual_delta = total_b as i64 - total_a as i64;
        let expected_delta = layers as i64 * (per_layer_b - per_layer_a);
        prop_assert_eq!(
            actual_delta, expected_delta,
            "delta mismatch: actual={}, expected L({}) * ({} - {}) = {}",
            actual_delta, layers, per_layer_b, per_layer_a, expected_delta
        );
    }

    /// Obligation: Tied embedding count (invariant)
    /// Formal: tied => count(untied) - count(tied) = 1
    #[test]
    fn prop_tied_embedding_count(
        layers in 1u32..100,
        has_bias in proptest::bool::ANY,
        has_qk_norm in proptest::bool::ANY
    ) {
        let untied = TensorConfig {
            layers,
            has_bias,
            has_qk_norm,
            num_norms: 2,
            tied_embeddings: false,
        };
        let tied = TensorConfig {
            layers,
            has_bias,
            has_qk_norm,
            num_norms: 2,
            tied_embeddings: true,
        };

        let diff = total_tensors(&untied) - total_tensors(&tied);
        prop_assert_eq!(
            diff, 1,
            "tied vs untied diff={}, expected 1", diff
        );
    }

    /// Obligation: Quantization byte ordering (monotonicity)
    /// Formal: Q4K < Q6K < Q8 < F16 < F32 bytes for same params
    #[test]
    fn prop_quant_byte_ordering(
        // Use multiples of 32 to avoid rounding issues
        params_base in 1u64..1_000_000
    ) {
        let params = params_base * 32;
        let q4k = quant_bytes(params, "Q4K");
        let q6k = quant_bytes(params, "Q6K");
        let q8 = quant_bytes(params, "Q8_0");
        let f16 = quant_bytes(params, "F16");
        let f32_bytes = quant_bytes(params, "F32");

        prop_assert!(
            q4k < q6k,
            "Q4K({}) >= Q6K({}) for params={}", q4k, q6k, params
        );
        prop_assert!(
            q6k < q8,
            "Q6K({}) >= Q8_0({}) for params={}", q6k, q8, params
        );
        prop_assert!(
            q8 < f16,
            "Q8_0({}) >= F16({}) for params={}", q8, f16, params
        );
        prop_assert!(
            f16 < f32_bytes,
            "F16({}) >= F32({}) for params={}", f16, f32_bytes, params
        );
    }

    /// Obligation: SIMD inventory equivalence (equivalence)
    #[test]
    #[ignore = "SIMD equivalence — trueno domain"]
    fn prop_simd_equivalence(
        _x in proptest::collection::vec(0u32..100, 1..32usize)
    ) {
        // Tensor inventory is pure math — no SIMD variant
    }
}