//! End-to-end Qwen3.6 / Qwen3-Next (`qwen35moe`) GGUF decode oracle. Ignored by
//! default — needs the ~22 GB Q5_K_M 35B-A3B checkpoint. Validates the
//! Gated-DeltaNet + gated-attention hybrid MoE path (lazy per-expert dequant) on
//! a *sharp* prompt.
//!
//! Reference (llama.cpp / llama-simple, no BOS): "The quick brown fox jumps over
//! the lazy" -> " dog" (5388).
//!
//! Run: `cargo test -p jouleclaw-loader-gguf --release --features rayon \
//!   --test real_qwen35_oracle -- --ignored --nocapture`

use jouleclaw_loader_gguf::qwen35moe::Qwen35Moe;

const GG: &str =
    "/Volumes/macos_4TB_external_1/neuronexus_ai/model_library/qwen3.6-35b-a3b-gguf/Qwen3.6-35B-A3B-UD-Q5_K_M.gguf";

const FOX: [u32; 8] = [760, 3841, 13477, 37550, 33075, 888, 279, 15217];
const DOG: u32 = 5388;

#[test]
#[ignore]
fn real_qwen35_predicts_dog() {
    if !std::path::Path::new(GG).exists() {
        eprintln!("skip: qwen35moe GGUF not present");
        return;
    }
    let m = Qwen35Moe::load(GG).expect("load qwen35moe gguf");
    eprintln!(
        "qwen35moe: {} layers (fa every {}), d={}, attn {}q/{}kv hd={} nrot={}, deltanet {}k/{}v heads hd={}, experts {}/{} +shared",
        m.cfg.n_layers, m.cfg.full_attn_interval, m.cfg.d_model, m.cfg.n_head, m.cfg.n_kv,
        m.cfg.head_dim, m.cfg.n_rot, m.cfg.n_k_heads, m.cfg.n_v_heads, m.cfg.head_v,
        m.cfg.n_used, m.cfg.n_experts
    );
    let logits = m.forward(&FOX);
    let mut idx: Vec<usize> = (0..logits.len()).collect();
    idx.sort_by(|&a, &b| logits[b].partial_cmp(&logits[a]).unwrap());
    eprintln!("top5 = {:?} (expected {DOG} = ' dog')", &idx[..5]);
    assert_eq!(idx[0], DOG as usize, "greedy next token diverges from llama.cpp");
}