//! End-to-end Nemotron-H-MoE (`nemotron_h_moe`) GGUF decode oracle. Ignored by
//! default — needs the ~17 GB Q4_K_M 30B checkpoint on disk. Validates the
//! 30B-class MoE path (lazy per-expert dequant) on a *sharp* prompt so the
//! argmax can't be flipped by f32-vs-quant matmul noise.
//!
//! Reference (llama.cpp / llama-simple, no BOS): "The quick brown fox jumps over
//! the lazy" -> " dog" (10575).
//!
//! Run: `cargo test -p jouleclaw-loader-gguf --release --features rayon \
//!   --test real_nemotron_oracle -- --ignored --nocapture`

use jouleclaw_loader_gguf::nemotron_gguf::NemotronMoe;

const GG: &str =
    "/Volumes/macos_4TB_external_1/neuronexus_ai/model_library/nemotron-3-nano-Q4/nvidia_Nemotron-3-Nano-30B-A3B-Q4_K_M.gguf";

const FOX: [u32; 8] = [1784, 7586, 22980, 94137, 72993, 2136, 1278, 42757];
const DOG: u32 = 10575;

#[test]
#[ignore]
fn real_nemotron_predicts_dog() {
    if !std::path::Path::new(GG).exists() {
        eprintln!("skip: nemotron GGUF not present");
        return;
    }
    let m = NemotronMoe::load(GG).expect("load nemotron gguf");
    eprintln!(
        "nemotron: {} layers, d={}, m_heads={} inner={} groups={} state={}, experts={}/{} +shared, attn {}q/{}kv hd={}",
        m.cfg.n_layers, m.cfg.d_model, m.cfg.m_heads, m.cfg.m_inner, m.cfg.n_groups, m.cfg.state,
        m.cfg.n_used, m.cfg.n_experts, m.cfg.n_heads, m.cfg.n_kv, m.cfg.head_dim
    );
    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");
}