"""Tests for the Layer-3 experiment-vs-model comparison harness.

Two layers of guarantees:

1. **Structural** — the harness loads the digitisation worksheet, resolves
   soil parameters, and produces finite, physically-ordered BW predictions
   at every operating point (always exercised).
2. **Accuracy** — once measured values are digitised into the worksheet,
   BW lands inside the published Bekker-Wong model-form band against the
   real measurements. These activate automatically per source as soon as
   that source has ``meas_*`` data; they ``skip`` while it is blank, so the
   suite stays green before digitisation without ever silently passing on
   absent data.
"""

from __future__ import annotations

import math

import pytest

from roverdevkit.validation.terramechanics_experiment import (
    compare_to_experiment,
    load_experiment_points,
    summarise,
)

# Published BW single-wheel model-form error is ~15-30 %; allow a generous
# band so the accuracy check guards against gross disagreement / unit bugs
# rather than over-fitting to a particular digitisation.
BW_MODEL_FORM_BAND_PCT = 40.0

# Per-source acceptance bands. Ding and Hurrell are held to the tight model-form
# band. Wang & Han 2016 (KLS-1) is a documented stress case at the edge of the
# rigid-wheel kernel's regime: the smallest/most-lightly-loaded wheel (R=85 mm,
# 59 N) on a firm, dense, fines-rich simulant that barely sinks, so the
# force-balance sinkage solve over-predicts DP/sinkage and cannot capture the
# s~0.5 soil-disturbance DP collapse. This is a MODEL-FORM limit, not a soil
# mis-specification: KLS-1's pressure-sinkage moduli are its OWN bevameter-fit
# values (Lim et al. 2021), yet the error persists. Its loose band guards
# against unit/sign bugs only -- it is NOT a validation claim.
SOURCE_BAND_PCT = {
    "ding2011": BW_MODEL_FORM_BAND_PCT,
    "hurrell2025_rashid1": BW_MODEL_FORM_BAND_PCT,
    "wang_han_2016_kls1": 200.0,
}


@pytest.fixture(scope="module")
def comparison():
    return compare_to_experiment()


def test_worksheet_loads_with_known_sources() -> None:
    points = load_experiment_points()
    assert len(points) > 0
    sources = {p.source for p in points}
    assert "ding2011" in sources
    assert "wang_han_2016_kls1" in sources


def test_bw_predictions_finite_everywhere(comparison) -> None:
    """Every operating point yields a finite BW prediction (no full burial)."""
    assert comparison["bw_drawbar_pull_n"].notna().all()
    assert comparison["bw_sinkage_m"].notna().all()
    assert comparison["bw_torque_nm"].notna().all()
    assert (comparison["bw_sinkage_m"] > 0.0).all()


def test_bw_drawbar_pull_rises_with_slip(comparison) -> None:
    """Within each (source, grouser) family, DP increases monotonically in slip."""
    grouped = comparison.sort_values("slip").groupby(["source", "grouser_height_m"])
    for _, group in grouped:
        dp = group["bw_drawbar_pull_n"].to_numpy()
        diffs = dp[1:] - dp[:-1]
        assert (diffs >= -1e-6).all(), f"DP not monotonic in slip: {dp}"


def test_grousers_increase_drawbar_pull() -> None:
    """At matched slip/load, a grousered wheel out-pulls a smooth one."""
    df = compare_to_experiment()
    for source in ("ding2011", "wang_han_2016_kls1"):
        sub = df[df["source"] == source]
        smooth = sub[sub["grouser_height_m"] == 0.0].set_index("slip")[
            "bw_drawbar_pull_n"
        ]
        lugged = sub[sub["grouser_height_m"] > 0.0].set_index("slip")[
            "bw_drawbar_pull_n"
        ]
        shared = smooth.index.intersection(lugged.index)
        assert len(shared) > 0
        for slip in shared:
            assert lugged[slip] >= smooth[slip] - 1e-6


def test_summary_has_expected_shape(comparison) -> None:
    summary = summarise(comparison)
    assert summary["n_operating_points"] == len(comparison)
    assert summary["n_digitised"] + summary["n_pending_digitisation"] <= summary[
        "n_operating_points"
    ]


@pytest.mark.parametrize(
    "source", ["ding2011", "wang_han_2016_kls1", "hurrell2025_rashid1"]
)
def test_bw_within_band_when_digitised(comparison, source: str) -> None:
    """BW drawbar pull lands in the published band against real measurements.

    Skips until the source's ``meas_drawbar_pull_n`` column is populated, so
    the check never passes vacuously on absent data.
    """
    sub = comparison[
        (comparison["source"] == source)
        & comparison["meas_drawbar_pull_n"].notna()
        # near zero-slip DP crosses zero, so % error is ill-defined there
        & (comparison["slip"] >= 0.1)
    ]
    if sub.empty:
        pytest.skip(f"{source}: drawbar-pull measurements not yet digitised")
    errors = sub["bw_dp_abs_pct_err"].dropna()
    assert len(errors) > 0
    median_err = float(errors.median())
    band = SOURCE_BAND_PCT.get(source, BW_MODEL_FORM_BAND_PCT)
    assert median_err < band, (
        f"{source}: BW median DP error {median_err:.1f}% exceeds "
        f"{band:.0f}% model-form band"
    )
    assert not math.isnan(median_err)