"""TDD for spaceutil.encoder. Requires the real Qwen3 model - marked
`model` and skipped by default. Run with `pytest -m model`."""

from __future__ import annotations

import numpy as np
import pytest

pytestmark = pytest.mark.model


def test_encode_query_returns_normalized_vector(synthetic_embed_provenance):
    from spaceutil.encoder import encode_query_via_optcg

    vec = encode_query_via_optcg("red blocker with draw", synthetic_embed_provenance)
    arr = np.asarray(vec, dtype=np.float32)

    assert arr.shape == (synthetic_embed_provenance.embedding_dim,)
    norm = float(np.linalg.norm(arr))
    assert abs(norm - 1.0) < 1e-4


def test_paraphrases_are_close(synthetic_embed_provenance):
    """Loose floor - the model is small. We assert > 0.4 cosine, well
    above random."""
    from spaceutil.encoder import encode_query_via_optcg

    a = np.asarray(
        encode_query_via_optcg("red blocker", synthetic_embed_provenance),
        dtype=np.float32,
    )
    b = np.asarray(
        encode_query_via_optcg("crimson defender", synthetic_embed_provenance),
        dtype=np.float32,
    )
    cosine = float(np.dot(a, b))
    assert cosine > 0.4, f"paraphrases too distant: cosine={cosine}"


def test_model_cached_across_calls(synthetic_embed_provenance):
    """Second call must reuse the loaded model - the warmup pattern is
    load-bearing for cold-start UX."""
    import time

    from spaceutil.encoder import encode_query_via_optcg, get_encoder

    _ = get_encoder(synthetic_embed_provenance)  # warmup

    t0 = time.time()
    _ = encode_query_via_optcg("hello world", synthetic_embed_provenance)
    first = time.time() - t0

    t0 = time.time()
    _ = encode_query_via_optcg("goodbye world", synthetic_embed_provenance)
    second = time.time() - t0

    assert second < first * 5, (
        f"second call ({second:.3f}s) much slower than first ({first:.3f}s) - "
        "model likely re-loaded"
    )