Blame for generator/code/tests/test_demography_semantics.py

from __future__ import annotations

import dataclasses
import json
import sys
from types import SimpleNamespace

import msprime
import numpy as np
import pytest

from simulator.config import (
    DEMOGRAPHY_PROB_FIELDS,
    DEMOGRAPHY_TYPE_IDS,
    GENERATOR_RECIPE_IDS,
    GENERATOR_RECIPE_TO_DEMOGRAPHY_TYPE,
    ID_TO_DEMOGRAPHY_TYPE,
    ID_TO_GENERATOR_RECIPE,
    SOURCE_PROB_FIELDS,
    Config,
)
from simulator.demography import (
    DEMO_PROB_FIELDS,
    DEMO_SAMPLERS,
    _canonical_smcpp_sawtooth,
    _canonical_zigzag_1s14,
    _clean_demography,
    _native_exponential_transition,
    _sawtooth,
    build_demography,
    choose_generator_recipe,
    resolve_demography_type,
    sample_custom_demography,
    summarize_demography,
)
from simulator.dataset_state import SAMPLE_GENERATOR_SCHEMA_VERSION
from simulator.sample import simulate_one
from simulator.stdpopsim_support import (
    StdpopsimSimulationError,
    _extract_single_population_trajectory,
    try_stdpopsim_sample,
)
from simulator.utils import bin_average_log10_ne, time_grid


def _only_demography(name: str, *, time_bins: int = 48) -> Config:
    weights = {field: 0.0 for field in DEMOGRAPHY_PROB_FIELDS}
    field_by_name = dict(DEMO_PROB_FIELDS)
    weights[field_by_name[name]] = 1.0
    return dataclasses.replace(Config(), time_bins=time_bins, **weights)


EXPECTED_STATIC_RECIPE_TYPES = {
    "constant": "constant",
    "spectral_walk": "random_walk",
    "bottleneck": "bottleneck",
    "ooa_chain": "bottleneck",
    "sawtooth": "zigzag",
    "literature_sawtooth": "zigzag",
    "empirical_human_ne_template": "empirical_human_ne_template",
    "stdpopsim_single_pop": "stdpopsim",
}


def _log_time_average_oracle(
    edges: np.ndarray,
    breaks: list[float],
    values: list[float],
    growth_rates: list[float],
) -> np.ndarray:
    """Independent closed-form oracle for mean(log10 Ne) in log-time bins."""
    starts = [0.0, *breaks]
    result: list[float] = []
    for left, right in zip(edges[:-1], edges[1:]):
        cuts = [float(left), *(t for t in breaks if left < t < right), float(right)]
        integral = 0.0
        for a, b in zip(cuts[:-1], cuts[1:]):
            epoch = int(np.searchsorted(breaks, np.sqrt(a * b), side="right"))
            log_width = np.log(b / a)
            integral += np.log10(values[epoch]) * log_width
            integral -= growth_rates[epoch] / np.log(10.0) * (
                (b - a) - starts[epoch] * log_width
            )
        result.append(float(integral / np.log(right / left)))
    return np.asarray(result)


def _shape_path_oracle(
    edges: np.ndarray,
    breaks: list[float],
    values: list[float],
    growth_rates: list[float],
) -> np.ndarray:
    starts = [0.0, *breaks]

    def evaluate(epoch: int, time: float) -> float:
        return float(
            values[epoch]
            * np.exp(-growth_rates[epoch] * (time - starts[epoch]))
        )

    path = [float(values[0])]
    for epoch, boundary in enumerate(breaks):
        if 0.0 < boundary < float(edges[-1]):
            path.extend([evaluate(epoch, boundary), float(values[epoch + 1])])
    right_epoch = int(np.searchsorted(breaks, edges[-1], side="left"))
    path.append(evaluate(right_epoch, float(edges[-1])))
    return np.asarray(path, dtype=np.float64)


@pytest.mark.parametrize(
    ("present_ne", "older_ne", "expected_sign"),
    [
        (100_000.0, 10_000.0, 1),
        (10_000.0, 100_000.0, -1),
    ],
)
def test_native_exponential_matches_msprime_trajectory(
    present_ne: float,
    older_ne: float,
    expected_sign: int,
) -> None:
    end = 1_000.0
    breaks, values, rates = _native_exponential_transition(present_ne, older_ne, end)
    demography = build_demography(breaks, values, rates)
    demography.validate()

    times = np.array([0.0, 100.0, 500.0, end - 1e-6, end, 2 * end])
    observed = demography.debug().population_size_trajectory(times)[:, 0]
    expected = np.where(
        times < end,
        present_ne * np.exp(-rates[0] * times),
        older_ne,
    )

    assert np.sign(rates[0]) == expected_sign
    np.testing.assert_allclose(observed, expected, rtol=2e-12, atol=1e-8)


def test_continuous_target_uses_exact_log_time_integral() -> None:
    breaks, values, rates = _native_exponential_transition(100_000.0, 10_000.0, 500.0)
    edges = np.array([50.0, 200.0, 1_000.0, 5_000.0])

    observed = bin_average_log10_ne(edges, breaks, values, rates)
    expected = _log_time_average_oracle(edges, breaks, values, rates)

    np.testing.assert_allclose(observed, expected, rtol=0.0, atol=2e-6)


def test_zero_growth_rates_preserve_piecewise_constant_targets() -> None:
    edges = np.geomspace(50.0, 500_000.0, 49)
    breaks = [200.0, 3_000.0, 70_000.0]
    values = [20_000.0, 2_000.0, 80_000.0, 10_000.0]

    previous_api = bin_average_log10_ne(edges, breaks, values)
    explicit_zero_rates = bin_average_log10_ne(edges, breaks, values, [0.0] * len(values))

    np.testing.assert_array_equal(previous_api, explicit_zero_rates)


@pytest.mark.parametrize(
    "scenario",
    [
        "exponential",
        "sawtooth",
        "ooa_chain",
        "stdpopsim_single_pop",
    ],
)
def test_native_model_target_matches_msprime_debugger(scenario: str) -> None:
    cfg = _only_demography(scenario)
    demography, target, meta = sample_custom_demography(np.random.default_rng(20260723), cfg)
    edges, _ = time_grid(cfg)

    numeric: list[float] = []
    debugger = demography.debug()
    boundaries = np.asarray(meta["demography_breaks"], dtype=np.float64)
    for left, right in zip(edges[:-1], edges[1:]):
        # Integrate piece by piece between epoch boundaries.  A size reset is a
        # jump in log Ne, and a trapezoid rule laid across one converges at
        # O(1/n): on a bin holding a boundary, 513 points leave an error near
        # 6e-4, which is thirty times this test's tolerance and entirely the
        # oracle's.  Splitting at the boundary makes every piece smooth, so the
        # oracle is limited by the trapezoid rule on a smooth function instead.
        inside = boundaries[(boundaries > left) & (boundaries < right)]
        cuts = np.unique(np.concatenate(([left], inside, [right])))
        total = 0.0
        for piece_left, piece_right in zip(cuts[:-1], cuts[1:]):
            log_times = np.linspace(np.log(piece_left), np.log(piece_right), 4_097)
            # Step in from both ends so the samples stay strictly inside the
            # epoch and never pick up the value from across the jump.
            log_times[0] += 1e-12
            log_times[-1] -= 1e-12
            ne = debugger.population_size_trajectory(np.exp(log_times))[:, 0]
            log_ne = np.log10(ne)
            total += float(
                np.sum((log_ne[:-1] + log_ne[1:]) * np.diff(log_times) / 2.0)
            )
        numeric.append(total / float(np.log(right) - np.log(left)))

    # Not every recipe in this list grows exponentially on every draw: eight of
    # the eighteen stdpopsim models are purely piecewise constant, so the flag
    # has to follow the drawn trajectory rather than the sampler's name.  What
    # the test is really pinning is that the two agree.
    rates = meta["demography_growth_rates"]
    assert meta["native_continuous_growth"] == any(float(r) != 0.0 for r in rates)
    assert len(rates) == len(meta["demography_values"])
    np.testing.assert_allclose(target, numeric, rtol=0.0, atol=2e-5)


@pytest.mark.parametrize("seed", range(30))
def test_exponential_metadata_records_realized_model(seed: int) -> None:
    """One sampler with a signed rate, in place of three named directions.

    ``continuous_exponential_growth``, ``rapid_recent_growth_extreme`` and
    ``continuous_exponential_decline`` recorded the same quantity under three
    metadata keys and two conventions -- present-over-old for two of them,
    old-over-present for the third -- so nothing could group them.  There is one
    signed field now, and it has to agree with the trajectory it describes.
    """

    cfg = _only_demography("exponential")
    _, _, meta = sample_custom_demography(np.random.default_rng(seed), cfg)
    present_ne, older_ne = meta["demography_values"]
    end = meta["demography_breaks"][0]
    rate = meta["demography_growth_rates"][0]

    assert meta["exponential_log10_factor"] == pytest.approx(
        np.log10(present_ne / older_ne)
    )
    assert meta["exponential_direction"] == (
        "growth" if present_ne > older_ne else "decline"
    )
    assert rate == pytest.approx(np.log(present_ne / older_ne) / end)
    assert meta["native_continuous_growth"] is True


def test_canonical_zigzag_1s14_matches_published_event_table() -> None:
    cfg = Config()
    edges, mids = time_grid(cfg)
    breaks, values, meta = _canonical_zigzag_1s14(
        np.random.default_rng(1),
        cfg,
        edges,
        mids,
    )

    expected_breaks = [
        33.333334976,
        133.3334544,
        533.33324512,
        2_133.3352704,
        8_533.329632,
        34_133.318528,
    ]
    expected_values = [
        71_560.0,
        71_560.0,
        7_156.029247798124,
        71_560.21678822454,
        7_155.999712600628,
        71_559.65925132985,
        7_156.0,
    ]
    expected_rates = [
        0.0,
        0.02302578256,
        -0.0057564631,
        0.00143911578,
        -0.0003597785,
        8.9944801565e-5,
        0.0,
    ]

    assert breaks == expected_breaks
    np.testing.assert_allclose(values, expected_values, rtol=2e-14, atol=1e-9)
    assert meta["demography_growth_rates"] == expected_rates
    assert meta["literature_model_id"] == "Zigzag_1S14"

    demography = build_demography(breaks, values, expected_rates)
    ancient = demography.debug().population_size_trajectory([breaks[-1], 100_000.0])[:, 0]
    np.testing.assert_array_equal(ancient, np.array([7_156.0, 7_156.0]))


def test_smcpp_sawtooth_is_kept_distinct_from_zigzag_1s14() -> None:
    cfg = Config()
    edges, mids = time_grid(cfg)
    s_breaks, s_values, s_meta = _canonical_smcpp_sawtooth(
        np.random.default_rng(1),
        cfg,
        edges,
        mids,
    )
    z_breaks, z_values, _ = _canonical_zigzag_1s14(
        np.random.default_rng(1),
        cfg,
        edges,
        mids,
    )

    assert s_meta["literature_model_id"] == "Terhorst2017_sawtooth"
    assert s_meta["literature_demography_anchor"] is True
    assert s_breaks[2] == 533.50498912
    assert z_breaks[2] == 533.33324512
    assert s_values[-1] == pytest.approx(7_164.809184944985)
    assert z_values[-1] == 7_156.0
    expected_rates = [
        0.0,
        *(
            rate / (4 * 14_312)
            for rate in [1_318.18, -329.546, 82.3865, -20.5966, 5.14916]
        ),
        0.0,
    ]
    assert s_meta["demography_growth_rates"] == expected_rates


def test_continuous_metadata_round_trip_reconstructs_ground_truth() -> None:
    cfg = _only_demography("sawtooth")
    demography, _, meta = sample_custom_demography(np.random.default_rng(81), cfg)
    restored = json.loads(json.dumps(meta))
    reconstructed = build_demography(
        restored["demography_breaks"],
        restored["demography_values"],
        restored["demography_growth_rates"],
    )
    times = np.geomspace(1.0, cfg.max_time * 2, 1_000)

    expected = demography.debug().population_size_trajectory(times)
    observed = reconstructed.debug().population_size_trajectory(times)
    np.testing.assert_allclose(observed, expected, rtol=1e-13, atol=1e-9)


def test_stdpopsim_proxy_extraction_preserves_first_population_growth() -> None:
    source = msprime.Demography()
    source.add_population(name="target", initial_size=100_000.0, growth_rate=0.001)
    source.add_population(name="other", initial_size=20_000.0)
    source.add_population_parameters_change(time=100.0, growth_rate=-0.002, population="target")
    source.add_population_parameters_change(time=200.0, initial_size=500.0, population="other")
    source.add_population_parameters_change(
        time=300.0,
        initial_size=40_000.0,
        growth_rate=0.0,
        population="target",
    )
    source.sort_events()

    breaks, values, rates = _extract_single_population_trajectory(source)
    reconstructed = build_demography(breaks, values, rates)
    times = np.array([0.0, 50.0, 100.0, 150.0, 299.0, 300.0, 1_000.0])

    expected = source.debug().population_size_trajectory(times)[:, 0]
    observed = reconstructed.debug().population_size_trajectory(times)[:, 0]
    assert breaks == [100.0, 300.0]
    assert rates == [0.001, -0.002, 0.0]
    np.testing.assert_allclose(observed, expected, rtol=2e-12, atol=1e-8)


def test_stdpopsim_anchor_fails_instead_of_inventing_constant_target(
    monkeypatch: pytest.MonkeyPatch,
) -> None:
    class FakeTreeSequence:
        sequence_length = 1_000

        def trim(self):
            return self

    class FakeEngine:
        def simulate(self, model, contig, samples, seed):
            return FakeTreeSequence()

    model = SimpleNamespace(
        id="BrokenModel",
        populations=[SimpleNamespace(id="pop")],
        model=None,
        get_samples=lambda *args: {"pop": int(args[0])},
    )
    species = SimpleNamespace(
        demographic_models=[model],
        get_demographic_model=lambda model_id: model,
        get_contig=lambda **kwargs: SimpleNamespace(**kwargs),
    )
    fake_stdpopsim = SimpleNamespace(
        get_species=lambda species_id: species,
        get_engine=lambda engine_id: FakeEngine(),
    )
    monkeypatch.setitem(sys.modules, "stdpopsim", fake_stdpopsim)
    cfg = dataclasses.replace(
        Config(),
        enable_stdpopsim=True,
        stdpopsim_models="BrokenModel",
        n_haplotypes=4,
        seq_len=1_000,
    )

    with pytest.raises(
        StdpopsimSimulationError,
        match="exposes no msprime demography",
    ):
        try_stdpopsim_sample(np.random.default_rng(123), cfg)


def test_native_literature_model_survives_full_sample_generation() -> None:
    cfg = _only_demography("stdpopsim_single_pop")
    source_weights = {field: 0.0 for field in SOURCE_PROB_FIELDS}
    source_weights["p_clean_constant"] = 1.0
    cfg = dataclasses.replace(
        cfg,
        seq_len=5_000,
        n_haplotypes=4,
        **source_weights,
    )

    sample = simulate_one((0, dataclasses.asdict(cfg), {"rec_maps": {}, "mut_maps": {}}, 721))
    meta = sample["metadata"]

    assert sample["demography_type_id"] == DEMOGRAPHY_TYPE_IDS["stdpopsim"]
    assert sample["generator_recipe_id"] == GENERATOR_RECIPE_IDS["stdpopsim_single_pop"]
    assert sample["target_log10_ne"].shape == (cfg.time_bins,)
    assert np.isfinite(sample["target_log10_ne"]).all()
    assert meta["demography_type"] == "stdpopsim"
    assert meta["generator_recipe"] == "stdpopsim_single_pop"
    assert meta["scenario"] == "stdpopsim_single_pop"
    # The row has to say it is held out, or the split cannot route it.
    assert meta["held_out_evaluation"] is True
    assert meta["literature_demography_anchor"] is True
    assert meta["demography_growth_rates"][-1] == 0.0
    assert meta["n_variants"] == sample["positions"].size


@pytest.mark.parametrize("seed", range(20))
def test_random_zigzag_is_continuous_alternating_and_terminally_constant(seed: int) -> None:
    cfg = Config()
    edges, mids = time_grid(cfg)
    breaks, values, meta = _sawtooth(np.random.default_rng(seed), cfg, edges, mids)
    rates = meta["demography_growth_rates"]

    starts = [0.0, *breaks]
    for epoch, boundary in enumerate(breaks):
        endpoint = values[epoch] * np.exp(-rates[epoch] * (boundary - starts[epoch]))
        assert endpoint == pytest.approx(values[epoch + 1], rel=2e-12)

    signs = np.sign(rates[1:-1])
    assert np.all(signs != 0)
    assert np.all(signs[1:] == -signs[:-1])
    assert rates[-1] == 0.0

    demography = build_demography(breaks, values, rates)
    terminal = demography.debug().population_size_trajectory(
        [breaks[-1], max(cfg.max_time * 2, breaks[-1] * 2)]
    )[:, 0]
    np.testing.assert_allclose(terminal[0], terminal[1], rtol=0.0, atol=0.0)


@pytest.mark.parametrize("sampler", [_sawtooth])
def test_continuous_sawtooth_ground_truth_does_not_depend_on_label_bins(sampler) -> None:
    cfg32 = Config(time_bins=32)
    cfg64 = Config(time_bins=64)
    edges32, mids32 = time_grid(cfg32)
    edges64, mids64 = time_grid(cfg64)

    result32 = sampler(np.random.default_rng(991), cfg32, edges32, mids32)
    result64 = sampler(np.random.default_rng(991), cfg64, edges64, mids64)

    assert result32 == result64


def test_all_registered_demographies_have_probability_fields() -> None:
    assert {name for name, _ in DEMO_PROB_FIELDS} == set(DEMO_SAMPLERS)
    assert {field for _, field in DEMO_PROB_FIELDS} == set(DEMOGRAPHY_PROB_FIELDS)
    # Every id names a sampler this generator still has.  Nothing is kept for
    # a dataset written before a change: the schema version refuses those long
    # before an id is ever read.
    assert set(GENERATOR_RECIPE_IDS) == set(DEMO_SAMPLERS) | {"stdpopsim"}
    assert set(GENERATOR_RECIPE_TO_DEMOGRAPHY_TYPE) == set(GENERATOR_RECIPE_IDS)
    assert len(set(DEMOGRAPHY_TYPE_IDS.values())) == len(DEMOGRAPHY_TYPE_IDS)
    assert len(set(GENERATOR_RECIPE_IDS.values())) == len(GENERATOR_RECIPE_IDS)
    assert {
        value: key for key, value in DEMOGRAPHY_TYPE_IDS.items()
    } == ID_TO_DEMOGRAPHY_TYPE
    assert {
        value: key for key, value in GENERATOR_RECIPE_IDS.items()
    } == ID_TO_GENERATOR_RECIPE


def test_public_demography_types_are_compact_and_explicit() -> None:
    assert set(DEMOGRAPHY_TYPE_IDS) == {
        "constant",
        "random_walk",
        "bottleneck",
        "expansion",
        "contraction",
        "zigzag",
        "empirical_human_ne_template",
        "stdpopsim",
    }
    assert {
        recipe: demography_type
        for recipe, demography_type in GENERATOR_RECIPE_TO_DEMOGRAPHY_TYPE.items()
        if demography_type is not None and recipe != "stdpopsim"
    } == EXPECTED_STATIC_RECIPE_TYPES
    assert {
        recipe
        for recipe, demography_type in GENERATOR_RECIPE_TO_DEMOGRAPHY_TYPE.items()
        if demography_type is None
    } == {
        # Direction is a property of the drawn trajectory, not the sampler.
        "size_event",
        "event_on_background",
        "multi_event",
        "exponential",
    }


def test_direction_is_read_from_the_trajectory_not_the_sampler() -> None:
    """Whether a size change is an expansion or a contraction is a property of
    what was drawn, so it cannot be a fixed attribute of the recipe."""

    assert resolve_demography_type("size_event", [20_000.0, 10_000.0], {}) == "expansion"
    assert resolve_demography_type("size_event", [10_000.0, 20_000.0], {}) == "contraction"
    assert resolve_demography_type("size_event", [10_000.0, 10_000.0], {}) == "constant"


@pytest.mark.parametrize("generator_recipe", sorted(DEMO_SAMPLERS))
def test_taxonomy_preserves_every_recipe_trajectory_and_adds_uniform_metrics(
    generator_recipe: str,
) -> None:
    cfg = _only_demography(generator_recipe)
    expected_rng = np.random.default_rng(830_001)
    assert choose_generator_recipe(expected_rng, cfg) == generator_recipe
    edges, mids = time_grid(cfg)
    raw_breaks, raw_values, raw_meta = DEMO_SAMPLERS[generator_recipe](
        expected_rng,
        cfg,
        edges,
        mids,
    )
    expected_breaks, expected_values, expected_rates = _clean_demography(
        raw_breaks,
        raw_values,
        raw_meta.get("demography_growth_rates"),
        # Mirror the production path: which band a recipe is clipped to is part
        # of what this test claims to preserve, so omitting it would compare a
        # stress recipe against the regular band.
        stress="stress" in generator_recipe,
    )
    expected_target = bin_average_log10_ne(
        edges,
        expected_breaks,
        expected_values,
        expected_rates,
    )

    _, observed_target, meta = sample_custom_demography(
        np.random.default_rng(830_001),
        cfg,
    )

    assert meta["generator_recipe"] == generator_recipe
    assert meta["generator_recipe_id"] == GENERATOR_RECIPE_IDS[generator_recipe]
    assert meta["scenario"] == generator_recipe
    expected_type = EXPECTED_STATIC_RECIPE_TYPES.get(
        generator_recipe,
        resolve_demography_type(generator_recipe, expected_values, raw_meta),
    )
    assert meta["demography_type"] == expected_type
    assert meta["demography_type_id"] == DEMOGRAPHY_TYPE_IDS[expected_type]
    if generator_recipe == "near_constant":
        assert meta["history_family"] == "stable"
    np.testing.assert_array_equal(meta["demography_breaks"], expected_breaks)
    np.testing.assert_array_equal(meta["demography_values"], expected_values)
    np.testing.assert_array_equal(meta["demography_growth_rates"], expected_rates)
    np.testing.assert_array_equal(observed_target, expected_target)

    shape_values = _shape_path_oracle(
        edges,
        expected_breaks,
        expected_values,
        expected_rates,
    )
    log_values = np.log10(shape_values)
    differences = np.diff(log_values)
    nonzero_signs = np.sign(differences[np.abs(differences) > 1e-12])
    expected_turns = (
        int(np.count_nonzero(nonzero_signs[1:] != nonzero_signs[:-1]))
        if nonzero_signs.size > 1
        else 0
    )
    assert meta["event_severity"] == pytest.approx(
        float(shape_values.min() / shape_values.max())
    )
    assert meta["log10_ne_range"] == pytest.approx(float(np.ptp(log_values)))
    assert meta["total_variation_log10"] == pytest.approx(float(np.abs(differences).sum()))
    assert meta["turn_count"] == expected_turns


def test_shape_metrics_include_growth_limits_and_instantaneous_resets() -> None:
    edges = np.array([1.0, 10.0, 100.0, 200.0])
    breaks = [100.0]
    values = [100.0, 100.0]
    growth_rates = [-np.log(10.0) / 100.0, 0.0]

    summary = summarize_demography(
        edges,
        breaks,
        values,
        growth_rates,
        {},
        generator_recipe="stdpopsim",
        demography_type="stdpopsim",
    )

    assert summary["min_Ne"] == pytest.approx(100.0)
    assert summary["max_Ne"] == pytest.approx(1_000.0)
    assert summary["event_severity"] == pytest.approx(0.1)
    assert summary["total_variation_log10"] == pytest.approx(2.0)
    assert summary["turn_count"] == 1
    assert summary["direction"] == "mixed"


def test_the_oscillation_strength_is_continuous_rather_than_a_boolean() -> None:
    """``oscillating_mild`` and ``zigzag_strong`` were one function and a flag.

    ``zigzag`` was worse: it returned the strong branch unchanged, so it was
    bit-identical to ``zigzag_strong`` on every seed and a classifier told them
    apart at 0.283, below the 0.500 chance rate.  What replaces all three has
    to actually fill the range between mild and strong, not sit on its ends.
    """

    cfg = Config()
    assert sum(float(getattr(cfg, field)) for field in DEMOGRAPHY_PROB_FIELDS) == pytest.approx(1.0)
    edges, mids = time_grid(cfg)
    strengths = [
        _sawtooth(np.random.default_rng(seed), cfg, edges, mids)[2][
            "oscillation_strength_fraction"
        ]
        for seed in range(400)
    ]
    assert min(strengths) < 0.05 and max(strengths) > 0.95
    # Neither end may carry a lump: a boolean in disguise would pile up there.
    interior = [value for value in strengths if 0.2 < value < 0.8]
    assert len(interior) > 0.5 * len(strengths)


def test_the_declared_versions_match_what_the_generator_writes() -> None:
    """Two versions travel with a dataset and they say different things.

    ``demography_mixture_version`` names the mixture a label was drawn from;
    ``SAMPLE_GENERATOR_SCHEMA_VERSION`` names the on-disk contract a reader has
    to understand.  Both are refused rather than adapted to when they do not
    match, because labels from two mixtures cannot be pooled and arrays from
    two schemas do not have the same dtypes.
    """

    cfg = Config()
    assert cfg.demography_mixture_version == "v1"
    # The held-out catalogue must carry weight.  At zero it is never drawn,
    # which leaves nothing to evaluate on -- an evaluation set still has to be
    # generated before it can be held out.
    assert cfg.p_stdpopsim_single_pop > 0.0
    assert cfg.p_spectral_walk == pytest.approx(0.20)
    assert SAMPLE_GENERATOR_SCHEMA_VERSION == 1