Воспроизводимый рабочий процесс исследования моделей и стратегий фьючерсов CME

# research_workflow.py


```py
"""Reader-facing model and futures-strategy workflow for CME futures."""

from __future__ import annotations

import hashlib
import inspect
from collections.abc import Iterable, Mapping
from dataclasses import dataclass
from pathlib import Path
from typing import Any, Literal, cast

import polars as pl
import yaml

from case_studies.research import (
    BacktestResult,
    CandidateSet,
    DecisionArtifact,
    OfficialPopulation,
    PredictionResult,
    ResolvedModelRequest,
    Result,
    StateTransitionPolicy,
    Study,
    candidate_set_supersedes,
    plan_backtests,
    require_resolved_requests_cover_the_catalog,
    run_backtests,
    run_models,
)
from case_studies.research.contracts import ExecutionTier
from case_studies.research.execution import ModelExecution
from case_studies.research.strategy import strategy_warmup_periods
from case_studies.utils.artifact_digest import value_digest
from case_studies.utils.backtest_loaders import get_backtest_config, load_backtest_prices_for
from case_studies.utils.backtest_runner import precompute_weights
from case_studies.utils.registry import prediction_hash_from_parts
from case_studies.utils.runtime import source_commit
from case_studies.utils.sweep_config import top_n_cap
from data import load_cme_futures
from utils.modeling import load_configs
from utils.paths import REPO_ROOT

CASE_STUDY = "cme_futures"


def _declared_sweep_labels() -> tuple[str, ...]:
    """The labels ``config/setup.yaml`` declares for the sweep, primary first.

    Read from the declaration rather than restated beside it. A literal tuple here is a
    second copy of what ``setup.yaml`` says, and nothing compares the two: adding a
    variant to the sweep would leave every notebook fitting the old set and publishing a
    population one label short, silently. That is the failure a ``PRIMARY_LABEL`` default
    causes in other case studies, reached by a different route.

    Anchored at ``REPO_ROOT`` rather than ``get_case_study_dir``, which honours
    ``ML4T_OUTPUT_DIR``. The sweep declaration is committed source, not an output, so a
    preview workspace or a test redirect must not be able to change which labels are
    declared. The order is ``sweep_labels``' order - primary, then variants as written -
    because a population hashes its members as an ordered list, and re-ordering would give
    every published population a new identity.
    """
    setup = yaml.safe_load(
        (REPO_ROOT / "case_studies" / CASE_STUDY / "config" / "setup.yaml").read_text()
    )
    labels = (setup or {}).get("labels") or {}
    primary = labels.get("primary")
    if not primary:
        raise ValueError(f"{CASE_STUDY}: setup.yaml declares no labels.primary")
    ordered = [str(primary)]
    ordered += [str(name) for name in (labels.get("variants") or []) if str(name) != str(primary)]
    return tuple(ordered)


ALL_LABELS = _declared_sweep_labels()
FRONT_CONTRACT_POSITION = 0
ROLL_POLICY = "volume_rolled_multiplicative_back_adjustment"
EXPIRY_POLICY = "continuous_front_contract_rolls_before_delivery"
# The result-protocol fields that differ between the two return horizons, and the axis the
# final cross-horizon comparison therefore spans.
HORIZON_DEPENDENT_PROTOCOL_FIELDS = ("cv", "feature_artifacts", "label_artifact")
# Every name here has to be one a notebook actually registers: `official_prediction_catalog`
# resolves each with `OfficialPopulation.one`, which raises on a name that does not exist, so a
# name declared and never written stops `12_model_analysis` rather than being ignored. All six now
# read `<case study>-<what the producer publishes>-validation-v1`, which is what the other eight
# case studies do. The first four name the training family; 10a and 10b split one family between
# two notebooks, so they name the model instead.
MODEL_POPULATION_NAMES = (
    "cme_futures-linear-validation-v1",
    "cme_futures-gbm-validation-v1",
    "cme_futures-tabular_dl-validation-v1",
    "cme_futures-deep_learning-validation-v1",
    "cme_futures-pca-validation-v1",
    "cme_futures-sdf-validation-v1",
)


@dataclass(frozen=True)
class FuturesPricePath:
    """Reader-facing prices and the rows that establish their roll lineage."""

    prices: pl.DataFrame
    audit: pl.DataFrame
    roll_transitions: pl.DataFrame
    expiry_rules: pl.DataFrame


@dataclass(frozen=True)
class FuturesBacktestExecution:
    results: tuple[BacktestResult, ...]
    catalog_rows: pl.DataFrame
    # None under a preview run. A population is canonical by definition and
    # `OfficialPopulation.create` refuses a preview member outright
    # (research/population.py:101-103), so a reduced re-execution publishes no snapshot.
    population: OfficialPopulation | None


def open_study(
    *,
    execution_tier: str,
    workspace: str | Path | None = None,
    entry_point: str | None = None,
) -> Study:
    """Open canonical regeneration, canonical execution into a workspace, or a reader preview.

    Canonical with no workspace regenerates the case study's artifacts in place, which is the
    production path and needs the generated-artifact directory symlinks a maintainer worktree
    carries. Canonical *with* a workspace is the same computation writing to that root instead,
    which is the only form a checkout without those symlinks can run: CI seeds its fixture into
    ``ML4T_OUTPUT_DIR`` and has no `features`, `labels` or `run_log` symlink to regenerate over,
    so `Study.regenerate` refused there and took notebooks 13 through 17 red on every run.

    ``execution_tier`` and ``entry_point`` are stamped on the returned study rather than left to
    their defaults. Both were dropped here while the library set them, which is the hazard a
    private wrapper carries: this one restates the library's construction and had drifted from it
    on both fields. A study built with no tier reports ``canonical`` whatever it was opened for,
    and `OfficialPopulation.create` refuses a preview member by reading exactly that field, so a
    preview through this wrapper could publish a population the library would have refused.
    """
    if execution_tier == "canonical":
        if workspace is None:
            return Study.regenerate(CASE_STUDY, entry_point=entry_point)
        return Study.open(
            CASE_STUDY,
            workspace=Path(workspace).expanduser().resolve(),
            entry_point=entry_point,
        )
    if execution_tier != "preview":
        raise ValueError("execution_tier must be canonical or preview")
    if workspace is None:
        raise ValueError("preview execution requires an explicit workspace")
    workspace = Path(workspace).expanduser().resolve()
    try:
        return Study.open(
            CASE_STUDY,
            workspace=workspace,
            entry_point=entry_point,
            execution_tier=ExecutionTier.PREVIEW,
        )
    except ValueError as error:
        generated = tuple(
            REPO_ROOT / "case_studies" / CASE_STUDY / name
            for name in ("features", "labels", "run_log")
        )
        if "artifact bundle" not in str(error) or not all(path.is_symlink() for path in generated):
            raise
        workspace.mkdir(parents=True, exist_ok=True)
        shared_config = workspace / "config"
        if not shared_config.exists():
            shared_config.symlink_to(
                REPO_ROOT / "case_studies" / "config", target_is_directory=True
            )
        return Study(
            case_study=CASE_STUDY,
            root=REPO_ROOT / "case_studies" / CASE_STUDY,
            release_root=REPO_ROOT,
            output_root=workspace,
            read_only=False,
            entry_point=entry_point,
            execution_tier=ExecutionTier.PREVIEW,
            manifest={
                "schema_version": 1,
                "case_study": CASE_STUDY,
                "baseline_source_commit": source_commit(REPO_ROOT),
                "preview_only": True,
            },
        )


def model_request_catalog(
    family: str,
    *,
    labels: Iterable[str] = ALL_LABELS,
    config_names: Iterable[str] | None = None,
) -> pl.DataFrame:
    """Return the declared model population as visible Polars rows."""
    selected = set(config_names) if config_names is not None else None
    if selected is not None and not selected:
        # An empty selection is the caller's, so say so. Falling through left every row filtered
        # out and the function reported "no declared requests for <family>", blaming the family's
        # menu for a list the caller passed empty.
        raise ValueError("config_names is empty; omit it to request every declared configuration")
    rows = []
    missing_by_label = {}
    for label in labels:
        configs = load_configs(CASE_STUDY, label, family)
        available = {str(config["config_name"]) for config in configs}
        missing = sorted((selected or set()) - available)
        if missing:
            missing_by_label[label] = missing
        for config in configs:
            name = str(config["config_name"])
            if selected is None or name in selected:
                rows.append({"family": family, "label": label, "config_name": name})
    if missing_by_label:
        raise ValueError(f"unknown {family!r} configurations by label: {missing_by_label}")
    if not rows:
        raise ValueError(f"no declared requests for {family!r}")
    return pl.DataFrame(rows).unique(maintain_order=True)


def resolve_model_requests(
    study: Study,
    request_catalog: pl.DataFrame,
    *,
    execution_tier: str,
    overrides: dict[str, Any] | None = None,
    preview_reductions: dict[str, Any] | None = None,
) -> tuple[ResolvedModelRequest, ...]:
    """Resolve visible catalog rows through the shared family boundary."""
    required = {"family", "label", "config_name"}
    missing = required - set(request_catalog.columns)
    if missing:
        raise ValueError(f"model request catalog is missing {sorted(missing)}")
    return tuple(
        study.model(
            **row,
            execution_tier=execution_tier,
            overrides=dict(overrides or {}),
            preview_reductions=dict(preview_reductions or {}),
        ).resolve()
        for row in request_catalog.select(*sorted(required)).iter_rows(named=True)
    )


def run_model_catalog(
    study: Study,
    request_catalog: pl.DataFrame,
    *,
    execution_tier: str,
    overrides: dict[str, Any] | None = None,
    preview_reductions: dict[str, Any] | None = None,
) -> ModelExecution:
    """Execute every visible model request and require complete catalog output."""
    resolved = resolve_model_requests(
        study,
        request_catalog,
        execution_tier=execution_tier,
        overrides=overrides,
        preview_reductions=preview_reductions,
    )
    return run_resolved_model_requests(study, resolved)


def run_resolved_model_requests(
    study: Study,
    resolved_requests: Iterable[ResolvedModelRequest],
) -> ModelExecution:
    """Execute already-resolved requests without loading their input artifacts again."""
    resolved = tuple(resolved_requests)
    if not resolved:
        raise ValueError("resolved model requests cannot be empty")
    execution = run_models(study, requests=resolved)
    expected_rows = sum(len(run.predictions) for run in execution.runs)
    if execution.catalog_rows.height != expected_rows:
        raise RuntimeError("model execution did not return every checkpoint catalog row")
    if execution.catalog_rows.filter(~pl.col("complete")).height:
        raise RuntimeError("model execution returned incomplete prediction rows")
    return execution


def run_official_model_catalog(
    study: Study,
    request_catalog: pl.DataFrame,
    *,
    population_name: str,
    resolved_requests: Iterable[ResolvedModelRequest] | None = None,
    supersedes: str | None = None,
) -> tuple[ModelExecution, OfficialPopulation]:
    """Snapshot and execute one complete canonical model population.

    ``supersedes`` names the population hash this run replaces, and it is a value a person
    sets after reading the registry rather than one the notebook can derive. It belongs in
    the snapshot, so passing it changes the population identity: a re-run that supplies it
    is registering a new population that records what it replaced, not re-registering the
    old one. Where a population already exists under this name and the membership has
    changed, ``OfficialPopulation.create`` refuses without it and names the hash required.
    """
    resolved = tuple(resolved_requests or ())
    if not resolved:
        resolved = resolve_model_requests(study, request_catalog, execution_tier="canonical")
    if any(request.spec["execution_tier"] != "canonical" for request in resolved):
        raise ValueError("official model populations require canonical requests")
    require_resolved_requests_cover_the_catalog(request_catalog, resolved)
    expected = expected_prediction_hashes(resolved)
    population = OfficialPopulation.create(
        study,
        name=population_name,
        member_kind="prediction",
        members=expected,
        supersedes=supersedes,
    )
    execution = run_resolved_model_requests(study, resolved)
    actual = tuple(prediction.hash for run in execution.runs for prediction in run.predictions)
    if set(actual) != set(expected) or len(actual) != len(expected):
        missing = sorted(set(expected) - set(actual))
        extra = sorted(set(actual) - set(expected))
        raise RuntimeError(f"model population mismatch: missing={missing}, extra={extra}")
    population.require_complete()
    return execution, population


def resolved_model_plan(resolved_requests: Iterable[ResolvedModelRequest]) -> pl.DataFrame:
    """Show the data, folds, checkpoints, and eligibility each request will use."""
    rows = []
    for request in resolved_requests:
        computation = request.spec.get("computation", request.spec)
        expected = request._context.expected_keys
        entity = next(
            (column for column in ("product", "symbol") if column in expected.columns),
            None,
        )
        fold = next((column for column in ("fold", "fold_id") if column in expected.columns), None)
        if entity is None or fold is None:
            raise ValueError("resolved model eligibility has no entity or fold key")
        timestamps = expected.get_column("timestamp")
        rows.append(
            {
                "family": request.family,
                "label": request.spec["label"],
                "config_name": request.spec.get("config_name"),
                "task": (computation.get("task") or {}).get("type", "regression"),
                "feature_count": len(computation.get("feature_names") or []),
                "eligible_entities": expected.get_column(entity).n_unique(),
                "eligible_rows": expected.height,
                "folds": expected.get_column(fold).n_unique(),
                "validation_start": timestamps.min(),
                "validation_end": timestamps.max(),
                "checkpoints": len(computation["checkpoint_schedule"]),
                "execution_tier": request.spec["execution_tier"],
                "training_hash": request.identity,
            }
        )
    return pl.DataFrame(rows).sort("label", "family", "config_name")


def product_universe_table() -> pl.DataFrame:
    """Return the configured CME product groups with expiry references."""
    setup = yaml.safe_load(
        (REPO_ROOT / "case_studies" / CASE_STUDY / "config" / "setup.yaml").read_text()
    )
    rows = [
        {"sector": sector, "product": product}
        for sector, products in setup["universe"]["product_groups"].items()
        for product in products
    ]
    universe = pl.DataFrame(rows)
    return universe.join(
        _expiry_rules(universe.get_column("product").to_list()),
        on="product",
        how="left",
    ).sort("sector", "product")


def official_prediction_catalog(
    study: Study,
    population_names: Iterable[str],
) -> pl.DataFrame:
    """Return the exact complete catalog rows from declared official populations."""
    members = []
    for name in population_names:
        population = OfficialPopulation.one(study, name=name)
        if population.member_kind != "prediction":
            raise ValueError(f"official population {name!r} does not contain predictions")
        members.extend(population.require_complete())
    if len(members) != len(set(members)):
        raise ValueError("official prediction populations overlap")
    catalog = study.predictions.table().filter(pl.col("prediction_hash").is_in(members))
    if catalog.height != len(members) or catalog.filter(~pl.col("complete")).height:
        raise ValueError("official prediction catalog is incomplete")
    return catalog.sort("label", "family", "config_name", "checkpoint_kind", "checkpoint_value")


def expected_prediction_hashes(resolved_requests) -> tuple[str, ...]:
    """Project the declared checkpoint population to immutable prediction identities."""
    hashes = []
    for request in resolved_requests:
        computation = request.spec.get("computation", request.spec)
        for checkpoint in computation["checkpoint_schedule"]:
            hashes.append(
                prediction_hash_from_parts(
                    request.identity,
                    checkpoint["value"],
                    "validation",
                    checkpoint_kind=checkpoint["kind"],
                    identity_version=request.spec["identity_version"],
                )
            )
    if len(hashes) != len(set(hashes)):
        raise ValueError("declared request population contains duplicate prediction identities")
    return tuple(hashes)


def _expiry_rules(products: list[str]) -> pl.DataFrame:
    path = REPO_ROOT / "data" / "futures" / "market" / "futures_specs.yaml"
    configured = yaml.safe_load(path.read_text())["products"]
    missing = sorted(set(products) - set(configured))
    if missing:
        raise ValueError(f"products have no contract specification: {missing}")
    rows = []
    for product in products:
        spec = configured[product]
        if not spec.get("expiry_rule") or not spec.get("contract_months"):
            raise ValueError(f"{product} has an incomplete expiry specification")
        rows.append(
            {
                "product": product,
                "expiry_rule": str(spec["expiry_rule"]),
                "contract_months": ",".join(str(value) for value in spec["contract_months"]),
            }
        )
    return pl.DataFrame(rows).sort("product")


def load_futures_price_path(
    label: str,
    *,
    split: Literal["validation", "holdout"] = "validation",
    max_products: int = 0,
    products: Iterable[str] | None = None,
    warmup_periods: int = 0,
) -> FuturesPricePath:
    """Load front-contract prices while retaining the rows that prove each roll."""
    selected_products = sorted(set(products or ()))
    if max_products and selected_products:
        raise ValueError("select CME prices by max_products or products, not both")
    engine_prices = load_backtest_prices_for(
        CASE_STUDY,
        label,
        split=split,
        max_symbols=max_products,
        warmup_periods=warmup_periods,
    ).rename({"symbol": "product"})
    if selected_products:
        engine_prices = engine_prices.filter(pl.col("product").is_in(selected_products))
        loaded_products = set(engine_prices.get_column("product"))
        missing_products = sorted(set(selected_products) - loaded_products)
        if missing_products:
            raise ValueError(f"selected CME products have no backtest prices: {missing_products}")
    resolved_products = sorted(engine_prices.get_column("product").unique().to_list())
    if not resolved_products:
        raise ValueError(f"CME backtest prices for {label!r} resolved to no products")
    price_keys = engine_prices.select("product", "timestamp").unique()
    start = str(price_keys.get_column("timestamp").min())[:10]
    end = str(price_keys.get_column("timestamp").max())[:10]
    loaded_audit = load_cme_futures(
        products=resolved_products,
        start_date=start,
        end_date=end,
    )
    audit = (
        cast(pl.DataFrame, loaded_audit.collect())
        if isinstance(loaded_audit, pl.LazyFrame)
        else loaded_audit
    )
    audit = audit.rename({"session_date": "timestamp", "tenor": "position"})
    required = {
        "product",
        "position",
        "timestamp",
        "adj_open",
        "adj_high",
        "adj_low",
        "adj_close",
        "raw_close",
        "cum_ratio",
    }
    missing = required - set(audit.columns)
    if missing:
        raise ValueError(f"CME price source is missing columns: {sorted(missing)}")
    audit = audit.filter(pl.col("position") == FRONT_CONTRACT_POSITION).with_columns(
        pl.col("timestamp").cast(price_keys.schema["timestamp"]),
        pl.col("product").cast(price_keys.schema["product"]),
    )
    audit = audit.join(price_keys, on=["product", "timestamp"], how="semi").sort(
        "product", "timestamp"
    )
    if audit.n_unique(["product", "position", "timestamp"]) != audit.height:
        raise ValueError("front-contract roll audit contains duplicate keys")
    missing_audit = price_keys.join(
        audit.select("product", "timestamp"), on=["product", "timestamp"], how="anti"
    )
    if not missing_audit.is_empty():
        raise ValueError("backtest price keys are missing from the front-contract roll audit")
    ratio = pl.col("cum_ratio").cast(pl.Float64)
    invalid_ratio = audit.filter(
        ratio.is_null() | ratio.is_nan() | ratio.is_infinite() | (ratio <= 0)
    )
    if not invalid_ratio.is_empty():
        raise ValueError("front-contract roll ratios must be finite positive values")
    scale = pl.max_horizontal(pl.col("adj_close").abs(), pl.lit(1.0))
    inconsistent = audit.filter(
        (pl.col("adj_close") - pl.col("raw_close") * pl.col("cum_ratio")).abs() > scale * 1e-10
    )
    if not inconsistent.is_empty():
        raise ValueError("roll-adjusted prices do not equal raw prices times the roll ratio")
    previous_ratio = pl.col("cum_ratio").shift(1).over("product")
    audit = audit.with_columns(
        previous_ratio.alias("previous_cum_ratio"),
        (previous_ratio.is_not_null() & (pl.col("cum_ratio") != previous_ratio)).alias(
            "roll_transition"
        ),
    )
    transitions = audit.filter(pl.col("roll_transition")).select(
        "product",
        "timestamp",
        "raw_close",
        "adj_close",
        "previous_cum_ratio",
        "cum_ratio",
        (pl.col("cum_ratio") / pl.col("previous_cum_ratio")).alias("roll_adjustment_factor"),
    )
    return FuturesPricePath(
        prices=engine_prices.sort("timestamp", "product"),
        audit=audit,
        roll_transitions=transitions,
        expiry_rules=_expiry_rules(resolved_products),
    )


def _resolved_signal(signal: dict[str, Any]) -> dict[str, Any]:
    resolved = dict(signal)
    resolved.setdefault("long_short", get_backtest_config(CASE_STUDY).long_short)
    return resolved


def _resolved_allocation(allocation: dict[str, Any] | None) -> dict[str, Any] | None:
    if allocation is None:
        return None
    resolved = dict(allocation)
    resolved.setdefault("long_short", get_backtest_config(CASE_STUDY).long_short)
    return resolved


def resolve_product_weights(
    prediction: PredictionResult,
    *,
    prices: pl.DataFrame,
    signal: dict[str, Any],
    allocation: dict[str, Any] | None = None,
    risk: dict[str, Any] | None = None,
    costs: dict[str, Any] | None = None,
) -> pl.DataFrame:
    """Resolve built-in research decisions and return canonical product weights."""
    if "product" not in prices.columns or "symbol" in prices.columns:
        raise ValueError("reader-facing CME prices require product and cannot contain symbol")
    study = prediction.study
    signal = _resolved_signal(signal)
    allocation = _resolved_allocation(allocation)
    unresolved = study.strategy(
        prediction=prediction,
        signal=signal,
        allocation=allocation,
        risk=risk,
        costs=costs,
    )
    spec = unresolved.resolve(prices=prices)
    engine_prices = prices.rename({"product": "symbol"})
    weights = precompute_weights(
        prediction.load(),
        spec,
        engine_prices,
        label=unresolved.label,
        case_study=CASE_STUDY,
        prediction_hash=prediction.hash,
    ).rename({"symbol": "product"})
    prediction_frame = prediction.load()
    entity_columns = [
        column for column in ("symbol", "product") if column in prediction_frame.columns
    ]
    if len(entity_columns) != 1:
        raise ValueError("CME predictions require exactly one entity key: symbol or product")
    fold_columns = [column for column in ("fold", "fold_id") if column in prediction_frame.columns]
    if len(fold_columns) != 1:
        raise ValueError("CME predictions require exactly one fold key: fold or fold_id")
    entity = entity_columns[0]
    fold = fold_columns[0]
    if entity == "symbol":
        prediction_frame = prediction_frame.rename({"symbol": "product"})
    prediction_frame = prediction_frame.with_columns(
        pl.col("product").cast(weights.schema["product"]),
        pl.col("timestamp").cast(weights.schema["timestamp"]),
    )
    fold_by_time = (
        prediction_frame.select("timestamp", fold)
        .unique()
        .group_by("timestamp")
        .agg(pl.col(fold).n_unique().alias("n_folds"), pl.col(fold).first().alias("fold"))
    )
    if fold_by_time.filter(pl.col("n_folds") != 1).height:
        raise ValueError("each CME decision timestamp must belong to exactly one fold")
    eligible = prediction_frame.select("product", "timestamp").unique()
    missing = weights.select("product", "timestamp").join(
        eligible, on=["product", "timestamp"], how="anti"
    )
    if not missing.is_empty():
        raise ValueError("resolved CME decisions contain keys outside prediction eligibility")
    return (
        weights.join(fold_by_time.select("timestamp", "fold"), on="timestamp", how="left")
        .select("product", "timestamp", "weight", "fold")
        .sort("timestamp", "product")
    )


def publish_product_weights(
    prediction: PredictionResult,
    *,
    prices: pl.DataFrame,
    signal: dict[str, Any],
    allocation: dict[str, Any] | None = None,
    risk: dict[str, Any] | None = None,
    costs: dict[str, Any] | None = None,
    canonical: bool = False,
) -> DecisionArtifact:
    """Publish validated CME weights with product, roll, expiry, and fold lineage."""
    signal = _resolved_signal(signal)
    allocation = _resolved_allocation(allocation)
    weights = resolve_product_weights(
        prediction,
        prices=prices,
        signal=signal,
        allocation=allocation,
        risk=risk,
        costs=costs,
    )
    source_identity: dict[str, Any] | None = None
    if canonical:
        replay = resolve_product_weights(
            prediction,
            prices=prices,
            signal=signal,
            allocation=allocation,
            risk=risk,
            costs=costs,
        )
        if value_digest(replay) != value_digest(weights):
            raise RuntimeError("canonical CME decision replay was not deterministic")
        source_identity = {
            "module": "case_studies.cme_futures.research_workflow",
            "source_digest": hashlib.sha256(
                inspect.getsource(resolve_product_weights).encode()
            ).hexdigest(),
            "declared_inputs": {
                "prediction_hashes": [prediction.hash],
                "prices": value_digest(prices),
            },
            "determinism": {"deterministic": True},
            "clean_replay_digest": value_digest(replay),
        }
    return DecisionArtifact.publish(
        prediction.study,
        kind="target_weights",
        decisions=weights,
        prediction_hashes=[prediction.hash],
        parameters={
            "signal": signal,
            "allocation": allocation,
            "risk": risk,
            "costs": costs,
            "cadence": "7d",
            "contract_position": FRONT_CONTRACT_POSITION,
            "roll_policy": ROLL_POLICY,
            "expiry_policy": EXPIRY_POLICY,
        },
        source_identity=source_identity,
        # Positions carry across a fold boundary rather than being liquidated.
        #
        # The five expanding-window folds are adjacent in calendar time and nothing
        # happens in the market on the four dates that separate them, so flattening
        # there would be an artifact of the evaluation index rather than a property
        # of the strategy. Liquidating also cannot be executed here in any case: the
        # configured weekly_friday_close -> monday_open delay means the engine has no
        # weight row to snap the reset onto, and research/strategy.py:128-149 refuses
        # to snap it forward because that carries the old state across the boundary
        # and then pays a round trip for no change in exposure.
        #
        # What this gives up: per-fold metrics are no longer computed from a flat
        # start, so a fold inherits at most one week of exposure from the fold before
        # it. Four boundaries against roughly 260 weekly decisions, so about 1.5% of
        # them. Any notebook reporting per-fold results must say so rather than claim
        # each fold begins flat.
        state_transition_policy=StateTransitionPolicy(
            fold_boundary="continue",
            temporal_gap="continue",
        ),
        canonical=canonical,
    )


def selected_prediction(study: Study, catalog_row: dict[str, Any]) -> PredictionResult:
    """Resolve one selected catalog row without exposing its hash to notebook control flow."""
    result = Result.open(
        study,
        str(catalog_row["prediction_hash"]),
        include_preview=catalog_row.get("execution_tier") == "preview",
    )
    if not isinstance(result, PredictionResult):
        raise ValueError("selected catalog row does not identify a prediction")
    return result


def strategy_request_frame(rows: list[dict[str, Any]]) -> pl.DataFrame:
    """Build visible request rows while preserving each nested strategy dictionary."""
    if not rows:
        raise ValueError("strategy request rows cannot be empty")
    nested = ("signal", "allocation", "risk", "costs")
    scalar_rows = [{key: value for key, value in row.items() if key not in nested} for row in rows]
    frame = pl.DataFrame(scalar_rows)
    return frame.with_columns(
        *(pl.Series(name, [row.get(name) for row in rows], dtype=pl.Object) for name in nested)
    )


def supersedes_declaration(execution_tier: str, declared: str) -> str | None:
    """What `06_linear` and `07_gbm` may offer as their population's `supersedes`.

    `population_supersedes` answers a registry question - does the name's current generation
    match this declaration - and answers it by reading `study.root`. That is the right question
    and the wrong root for a preview in a maintainer worktree. `open_study` gives a preview its
    own root only when the case study's generated directories are real; where they are symlinks
    to shared artifacts, which `create_experiment` cannot copy, it reads them in place and
    redirects only the writes. `root` stays the canonical case directory, the lookup finds the
    canonical generation, and the helper correctly returns a hash that this run must not use:
    `run_model_population` refuses a preview carrying one, so the preview fails before its first
    fit. It fails only for the maintainer, because a CI checkout has no symlinks.

    The tier is therefore decided before the registry is consulted, not after. A preview
    population is discarded with its workspace and has no lineage to extend, so there is no tier
    other than canonical in which the declaration means anything.

    This lives here rather than inline in each notebook so that both pass the same expression and
    a test can exercise it. It deliberately does not wrap `population_supersedes`: what that
    function decides is the registry's business and is tested against real registries.
    """
    return declared if execution_tier == "canonical" else None


def preview_prediction_candidates(
    study: Study,
    *,
    labels: Iterable[str],
    limit: int,
) -> pl.DataFrame:
    """The preview validation predictions to backtest, capped per label.

    The cap is per label and not over the whole frame. A single head across a label-sorted
    frame spends the budget on whichever label sorts first, so a later label is left with
    fewer rows than it was allotted - or with none at all when the first label alone fills
    the budget.

    Only the second of those is caught downstream: a label reduced to zero is refused by the
    emptiness check below, while a label reduced merely *below* its budget is not, and that
    one is invisible. Both are the silent narrowing the sweep contracts exist to prevent, so
    the cap is applied per group rather than repaired afterwards.

    This lives here rather than inline in `13_backtest` so the notebook and its test call the
    same code. A test that rebuilds the expression asserts Polars' semantics and keeps passing
    when the notebook regresses.
    """
    labels = list(labels)
    if not labels:
        raise ValueError("preview prediction selection requires at least one label")
    if limit < 1:
        raise ValueError("preview prediction selection requires a positive limit")
    candidates = (
        study.predictions.table(include_preview=True)
        .filter(
            (pl.col("execution_tier") == "preview")
            & (pl.col("split") == "validation")
            & pl.col("complete")
            & pl.col("label").is_in(labels)
        )
        .sort("label", "family", "config_name", "checkpoint_kind", "checkpoint_value")
        .group_by("label", maintain_order=True)
        .head(limit)
    )
    starved = [label for label in labels if candidates.filter(pl.col("label") == label).is_empty()]
    if starved:
        raise RuntimeError(
            f"preview execution found no complete validation predictions for {starved}"
        )
    return candidates


def run_official_backtest_requests(
    study: Study,
    requests: pl.DataFrame,
    *,
    population_name: str | None,
    supersedes: str | None = None,
) -> FuturesBacktestExecution:
    """Resolve, snapshot, and execute visible futures strategy requests.

    ``population_name`` is None for a preview run, which publishes no snapshot: a backtest
    inherits its tier from the predictions it reads (research/execution.py:364), preview
    results are refused entry to an official population, and the workspace holding them is
    discarded afterwards. Everything else - the expected-identity snapshot before the engine
    runs, the order check, the per-request completeness - applies to both tiers.

    ``supersedes`` names the generation of ``population_name`` this run retires. Anything that
    moves a backtest identity - a corrected label, a changed accounting field, a re-run after a
    registry reset - produces a different member list under the same name, and
    ``OfficialPopulation.create`` refuses to write it without being told which snapshot it
    replaces. The notebooks declare it as a parameter, so the sweep can be re-run without
    editing this module.
    """
    required = {"request_name", "prediction_hash", "label", "signal"}
    missing = required - set(requests.columns)
    if missing:
        raise ValueError(f"strategy requests are missing columns: {sorted(missing)}")
    if requests.is_empty() or requests.get_column("request_name").n_unique() != requests.height:
        raise ValueError("strategy request names must be non-empty and unique")
    catalog = study.predictions.table(include_preview=True)
    price_cache: dict[tuple[str, int], FuturesPricePath] = {}
    prepared = []
    expected = []
    for row in requests.iter_rows(named=True):
        selected = catalog.filter(pl.col("prediction_hash") == row["prediction_hash"])
        if selected.height != 1 or not selected.item(0, "complete"):
            raise ValueError(
                f"prediction {row['prediction_hash']!r} is absent, ambiguous, or incomplete"
            )
        if selected.item(0, "label") != row["label"]:
            raise ValueError("strategy request label does not match its prediction catalog row")
        prediction = selected_prediction(study, selected.row(0, named=True))
        allocation = _resolved_allocation(row.get("allocation"))
        risk = row.get("risk")
        costs = row.get("costs")
        warmup = strategy_warmup_periods({"strategy": {"allocation": allocation}})
        cache_key = (str(row["label"]), warmup)
        if cache_key not in price_cache:
            price_cache[cache_key] = load_futures_price_path(
                str(row["label"]),
                split="validation",
                warmup_periods=warmup,
            )
        price_path = price_cache[cache_key]
        signal = _resolved_signal(row["signal"])
        decision = publish_product_weights(
            prediction,
            prices=price_path.prices,
            signal=signal,
            allocation=allocation,
            risk=risk,
            costs=costs,
            # A decision artifact is canonical when the prediction it decides on is, and not
            # by declaration: `DecisionArtifact.publish` resolves its lineage with
            # `include_preview=not canonical` (research/decisions.py:192), so asserting
            # canonical over a preview prediction sends that lookup to the wrong registry and
            # the publish fails on its own input with `Unknown result hash`.
            canonical=prediction.execution_tier == "canonical",
        )
        plan = plan_backtests(
            study,
            predictions=selected,
            signal=signal,
            decision=decision,
            prices=price_path.prices,
            allocation=allocation,
            risk=risk,
            costs=costs,
            chapter=row.get("chapter"),
        )
        if len(plan.members) != 1:
            raise RuntimeError("one futures strategy request must resolve to one backtest")
        expected_hash = plan.expected_hashes[0]
        expected.append(expected_hash)
        prepared.append((row, selected, price_path, signal, allocation, decision, expected_hash))
    if len(expected) != len(set(expected)):
        raise ValueError("strategy requests resolve to duplicate backtest identities")
    population = (
        OfficialPopulation.create(
            study,
            name=population_name,
            member_kind="backtest",
            members=expected,
            supersedes=supersedes,
        )
        if population_name is not None
        else None
    )
    results = []
    rows = []
    for row, selected, price_path, signal, allocation, decision, expected_hash in prepared:
        # `run_backtests` says per member whether it computed the backtest or served an
        # identical one from the registry, and that distinction is kept rather than dropped.
        # Without it a re-run of an already-registered sweep displays the same table of N
        # complete rows as the run that computed them, and a reader cannot tell a cold sweep
        # from a no-op - a wrong reading that looks exactly like a right one, and one that
        # gets more misleading every time the sweep is re-run.
        executed = run_backtests(
            study,
            predictions=selected,
            signal=signal,
            prices=price_path.prices,
            allocation=allocation,
            risk=row.get("risk"),
            costs=row.get("costs"),
            chapter=row.get("chapter"),
            decision=decision,
        )
        result = executed.results[0]
        source = "computed" if executed.diagnostics[0]["status"] == "completed" else "reused"
        if result.hash != expected_hash:
            raise RuntimeError(f"backtest identity changed: {expected_hash} -> {result.hash}")
        results.append(result)
        rows.append(
            {
                "request_name": row["request_name"],
                "label": row["label"],
                "prediction_hash": row["prediction_hash"],
                "decision_hash": decision.hash,
                "backtest_hash": result.hash,
                "complete": result.complete,
                "source": source,
            }
        )
    if tuple(result.hash for result in results) != tuple(expected):
        raise RuntimeError("backtest execution did not preserve declared request order")
    if population is not None:
        population.require_complete()
    elif not all(result.complete for result in results):
        raise RuntimeError("a preview backtest did not complete")
    return FuturesBacktestExecution(tuple(results), pl.DataFrame(rows), population)


CANDIDATE_SET_STAGES = ("signal", "allocation", "risk", "pre-overlay", "final-validation")


def candidate_set_name(stage: str, label: str) -> str:
    """Name one per-label candidate set.

    Every writer and every reader of a per-label set goes through this function. A stage the
    funnel does not define is refused here rather than minting a second namespace a reader
    would then find empty: a candidate set opened under a name nothing wrote does not fail
    loudly, it carries an absent or stale pool into the notebooks downstream.
    """
    if stage not in CANDIDATE_SET_STAGES:
        raise ValueError(
            f"unknown candidate set stage {stage!r}, expected one of {CANDIDATE_SET_STAGES}"
        )
    return f"cme_futures-{stage}-{label}-v1"


# `setup` is the case-study configuration digest, not a feature input. The latent-factor adapter
# records it alongside the feature artifacts and the other families do not, so it is dropped before
# the comparison rather than being read as a difference in what the members trained on.
NON_FEATURE_ARTIFACT_ROLES = frozenset({"setup"})
_NORMALIZED_FEATURE_ARTIFACT_CONTRACT = {"comparable_fields": ["feature_artifacts"]}


def _feature_artifact_digests(result: Result) -> dict[str, str]:
    """Return ``{role: digest}`` for the feature inputs, from any recorded shape.

    Three shapes reach the registry for the same information. The tabular families record
    ``{role: {"sha256": <hex>, "size": int}}``, the latent-factor adapter records
    ``[{"role": ..., "sha256": "sha256:<hex>"}]``, and a role may carry a bare digest string.
    Comparing the recorded field by equality therefore refuses a mixed-family set over how the
    same digests were written down. Normalizing first keeps the guarantee the protocol check
    exists to give - every member of a comparable set read the same features - instead of
    declaring the field uncomparable and giving it up.
    """
    recorded = result.protocol()["feature_artifacts"]
    if isinstance(recorded, dict):
        entries = recorded.items()
    else:
        entries = [(entry["role"], entry) for entry in recorded or ()]
    digests = {}
    for role, value in entries:
        if role in NON_FEATURE_ARTIFACT_ROLES:
            continue
        digest = value.get("sha256") if isinstance(value, dict) else value
        if not digest:
            raise ValueError(f"result {result.hash} records no digest for {role!r}")
        digests[role] = str(digest).removeprefix("sha256:")
    return digests


def _require_agreeing_feature_artifacts(members: Iterable[Result]) -> None:
    """Refuse a candidate set whose members did not read the same feature artifacts."""
    baseline = None
    for member in members:
        digests = _feature_artifact_digests(member)
        if baseline is None:
            baseline = digests
            continue
        if digests != baseline:
            differing = sorted(set(baseline) ^ set(digests)) or sorted(
                role for role in baseline if baseline[role] != digests[role]
            )
            raise ValueError(f"candidate set members read different feature artifacts: {differing}")


def _create_comparable_set(
    study: Study,
    name: str,
    members: list[Result],
    *,
    supersedes_by_set: Mapping[str, str] | None = None,
) -> CandidateSet:
    """Create one set whose members are checked to share their feature artifacts.

    ``supersedes_by_set`` maps a set name to the generation this run retires, keyed by name
    rather than passed as one value because a notebook freezes several sets and the checker in
    ``scripts/check_supersedes_literals.py`` places a declaration by the name it states. The
    declaration is resolved through :func:`case_studies.research.candidate_set_supersedes` rather
    than offered straight, so a reader's clean clone - which has no generation to replace, and
    often no ``candidate_sets`` table at all - publishes generation one instead of being refused.

    Nothing reached this argument before, and a candidate set is immutable under its name, so any
    run whose membership moved stopped here with ``a changed candidate set named '...' must
    explicitly supersedes <hash>`` and no parameter could answer it. Widening the sweep changes
    the members by construction, which made every wider run unreachable.
    """
    _require_agreeing_feature_artifacts(members)
    return CandidateSet.create(
        study,
        name,
        members,
        comparison_contract=dict(_NORMALIZED_FEATURE_ARTIFACT_CONTRACT),
        supersedes=candidate_set_supersedes(
            study, name=name, declared=(supersedes_by_set or {}).get(name)
        ),
    )


def create_label_candidate_sets(
    study: Study,
    execution: FuturesBacktestExecution,
    *,
    stage: str,
    supersedes_by_set: Mapping[str, str] | None = None,
) -> dict[str, CandidateSet]:
    """Create one immutable comparable backtest set per label.

    ``supersedes_by_set`` is keyed by the set's own name, so one declaration covers every label
    this stage freezes.
    """
    labels = execution.catalog_rows.get_column("label").unique().sort().to_list()
    output = {}
    for label in labels:
        hashes = execution.catalog_rows.filter(pl.col("label") == label).get_column("backtest_hash")
        members_by_hash = {result.hash: result for result in execution.results}
        members = [members_by_hash[value] for value in hashes]
        output[label] = _create_comparable_set(
            study,
            candidate_set_name(stage, label),
            members,
            supersedes_by_set=supersedes_by_set,
        )
    return output


# The candidate-set stage names and the registry's ``backtest_runs.stage`` values are two
# vocabularies. A set is named for the funnel step it freezes; a row is named for the
# computation that produced it. Only the three executed steps appear in both - `pre-overlay`
# and `final-validation` name unions, and no backtest row carries either.
REGISTRY_STAGE_BY_SET_STAGE = {
    "signal": "signal",
    "allocation": "allocation",
    "risk": "risk_overlay",
}


def stage_backtest_results(
    study: Study,
    *,
    stage: str,
    label: str,
    execution_tier: str,
) -> tuple[BacktestResult, ...]:
    """The eligible results of one funnel step for one label, at the tier that produced them.

    Canonical reads the immutable per-label candidate set, which is what makes a selection
    reproducible from the registry alone: the pool is fixed before anything ranks it. A preview
    has no set to read, because `CandidateSet.create` refuses a preview member outright
    (research/comparison.py:50-51), so it reads its own workspace rows for the same step. The
    guarantee therefore differs by tier and the notebooks say so: canonical selects from a
    frozen pool, preview selects from whatever its reduced run produced.
    """
    if execution_tier == "canonical":
        members = CandidateSet.one(study, name=candidate_set_name(stage, label)).members
        return tuple(Result.open(study, value) for value in members)
    if execution_tier != "preview":
        raise ValueError("execution_tier must be canonical or preview")
    registry_stage = REGISTRY_STAGE_BY_SET_STAGE.get(stage)
    if registry_stage is None:
        raise ValueError(f"{stage!r} names a union of steps, which no backtest row carries")
    rows = study.backtests.table(include_preview=True).filter(
        (pl.col("execution_tier") == "preview")
        & (pl.col("split") == "validation")
        & (pl.col("stage") == registry_stage)
        & (pl.col("label") == label)
        & pl.col("complete")
    )
    return tuple(
        Result.open(study, value, include_preview=True)
        for value in rows.get_column("backtest_hash")
    )


def rank_by_validation_sharpe(
    study: Study,
    results: Iterable[Result],
) -> tuple[BacktestResult, ...]:
    """Order backtest results by validation Sharpe, ties broken by identity.

    The same rule `CandidateSet._ranked_validation_hashes` applies, so a preview ranking and a
    canonical one differ in which rows they see and in nothing else. A member with no Sharpe is
    refused rather than sorted to an end, which is what a null would otherwise do silently.

    Unless it is bankrupt. A null Sharpe used to mean exactly one thing - the run was not
    measured - and refusing was the whole of the right answer. It now means two, because a path whose equity reaches zero stops compounding and registers
    `sharpe`, `sortino`, `calmar`, `omega`, `stability` and `tail_ratio` as null on purpose:
    ranking a bankrupt path is the thing that issue exists to prevent. The `ruin` column is what
    separates the two, so this reads it rather than testing the Sharpe alone:

    * ``ruin = 1.0`` - bankrupt. It sorts last, ahead of nothing, and is never selected. It does
      not disqualify the set, because a sweep that produced one bankrupt member and eleven
      solvent ones has a perfectly good ranking of the eleven.
    * null Sharpe, no ruin flag - not measured. Refused, as before.

    Sorting last rather than dropping is what `strategy_analysis.rank_returns_on_common_support`
    already does for the same condition, so a reader comparing the two sees one rule.

    A registry written before #920 has no `ruin` column at all, which is not the same as no
    bankrupt member; there the check falls back to the null test that was the whole rule then.
    """
    members = {result.hash: result for result in results}
    if not members:
        raise ValueError("ranking by validation Sharpe requires at least one result")
    table = study.backtests.table(include_preview=True)
    ruined = (
        (pl.col("ruin") == 1.0).fill_null(False) if "ruin" in table.columns else pl.lit(False)  # noqa: FBT003
    )
    rows = (
        table.filter(
            pl.col("backtest_hash").is_in(list(members)) & (pl.col("sharpe").is_not_null() | ruined)
        )
        .with_columns(ruined.alias("_ruined"))
        .sort(
            "_ruined", "sharpe", "backtest_hash", descending=[False, True, False], nulls_last=True
        )
    )
    if rows.height != len(members):
        raise ValueError("a result being ranked has no validation Sharpe recorded")
    return tuple(members[value] for value in rows.get_column("backtest_hash"))


def pre_overlay_results(
    study: Study,
    *,
    label: str,
    execution_tier: str,
    supersedes_by_set: Mapping[str, str] | None = None,
) -> tuple[BacktestResult, ...]:
    """The signal and allocation pool for one label, at the tier that produced it.

    ``supersedes_by_set`` is forwarded to the set this builds. Without it a caller could declare
    a generation for `cme_futures-pre-overlay-<label>-v1` and still be refused, because the
    canonical branch creates that set here and the declaration never arrived: measured
    2026-09-19, when the weekend risk lane passed all six set names as ``live`` and stopped on
    the one name this function owns.
    """
    if execution_tier == "canonical":
        pool = pre_overlay_candidate_set(study, label=label, supersedes_by_set=supersedes_by_set)
        return tuple(Result.open(study, value) for value in pool.members)
    return (
        *stage_backtest_results(study, stage="signal", label=label, execution_tier=execution_tier),
        *stage_backtest_results(
            study, stage="allocation", label=label, execution_tier=execution_tier
        ),
    )


def final_validation_results(
    study: Study,
    *,
    label: str,
    execution_tier: str,
    supersedes_by_set: Mapping[str, str] | None = None,
) -> tuple[BacktestResult, ...]:
    """The full selection pool for one label: signal, allocation and the risk overlay."""
    if execution_tier == "canonical":
        pool = final_validation_candidate_set(
            study, label=label, supersedes_by_set=supersedes_by_set
        )
        return tuple(Result.open(study, value) for value in pool.members)
    return (
        *pre_overlay_results(
            study,
            label=label,
            execution_tier=execution_tier,
            supersedes_by_set=supersedes_by_set,
        ),
        *stage_backtest_results(study, stage="risk", label=label, execution_tier=execution_tier),
    )


def shortlist_signal_configurations(
    study: Study,
    *,
    label: str,
    limit: int,
    execution_tier: str = "canonical",
) -> tuple[BacktestResult, ...]:
    """Select the strongest signal result for each distinct model configuration.

    ``limit`` is a promise, not a cap: a positive one that the population cannot fill raises,
    because a caller that asked for 20 and got 8 is looking at a degenerate population and
    silently ranking the 8 would hide it. ``limit=0`` asks for every distinct configuration,
    the spelling ``top_n_predictions.signal`` already uses in every ``setup.yaml``, and is the
    only way to ask for all of them without first knowing how many there are - which is the
    one thing this function exists to compute. Asking with a number large enough to be sure,
    999 against a population of 50, is indistinguishable from the degenerate case and raises.
    """
    cap = top_n_cap(limit)
    pool = stage_backtest_results(study, stage="signal", label=label, execution_tier=execution_tier)
    selected = []
    configurations = set()
    for result in rank_by_validation_sharpe(study, pool):
        assert isinstance(result, BacktestResult)
        training = result.lineage()["training_spec"]
        key = (training["family"], training.get("config_name"))
        if key in configurations:
            continue
        configurations.add(key)
        selected.append(result)
        if len(selected) == cap:
            break
    if cap is None:
        if not selected:
            raise ValueError("signal population holds no distinct configurations")
    elif len(selected) != cap:
        raise ValueError(
            f"signal population has {len(selected)} distinct configurations, expected {limit}"
        )
    return tuple(selected)


def _union_members(study: Study, *pools: Iterable[str]) -> list[Result]:
    """The distinct results across several stage pools, in first-seen order.

    A backtest is identified by its prediction and its strategy spec, and the funnel stage is
    not part of either. So two stages register one row whenever the later stage changed nothing
    about a configuration - an allocation that resolves to the same spec the signal stage
    already ran is exactly that - and the stages' pools then overlap. `CandidateSet.create`
    refuses a repeated member, so concatenating the pools produced a union that raised precisely
    when two stages agreed, which is the case the union exists to describe.
    """
    seen: dict[str, None] = {}
    for pool in pools:
        for value in pool:
            seen.setdefault(value, None)
    return [Result.open(study, value) for value in seen]


def pre_overlay_candidate_set(
    study: Study, *, label: str, supersedes_by_set: Mapping[str, str] | None = None
) -> CandidateSet:
    """Return the immutable union of signal and allocation validation results."""
    signal = CandidateSet.one(study, name=candidate_set_name("signal", label))
    allocation = CandidateSet.one(study, name=candidate_set_name("allocation", label))
    members = _union_members(study, signal.members, allocation.members)
    return _create_comparable_set(
        study,
        candidate_set_name("pre-overlay", label),
        members,
        supersedes_by_set=supersedes_by_set,
    )


def final_validation_candidate_set(
    study: Study, *, label: str, supersedes_by_set: Mapping[str, str] | None = None
) -> CandidateSet:
    """Return the selection pool across signal, allocation, and risk-overlay stages."""
    pre_overlay = pre_overlay_candidate_set(study, label=label, supersedes_by_set=supersedes_by_set)
    risk = CandidateSet.one(study, name=candidate_set_name("risk", label))
    members = _union_members(study, pre_overlay.members, risk.members)
    return _create_comparable_set(
        study,
        candidate_set_name("final-validation", label),
        members,
        supersedes_by_set=supersedes_by_set,
    )


def final_selection_candidate_set(
    study: Study, *, supersedes_by_set: Mapping[str, str] | None = None
) -> CandidateSet:
    """Return the one immutable pool the case-study configuration is selected from.

    The funnel runs per label through the risk overlay. The object of selection is one
    case-study configuration, so the per-label pools are compared as a single set and the
    label comes from the row that wins rather than being fixed in advance.

    The return horizon is the axis the comparison spans, and naming it is what makes the
    comparison legible: the two pools carry different label artifacts, different derived
    features and a different purge interval. Everything else must match, and the candidate
    set refuses a member whose split or execution tier disagrees.
    """
    members = []
    for label in ALL_LABELS:
        pool = final_validation_candidate_set(
            study, label=label, supersedes_by_set=supersedes_by_set
        )
        members.extend(Result.open(study, value) for value in pool.members)
    name = "cme_futures-final-selection-v1"
    return CandidateSet.create(
        study,
        name,
        members,
        comparison_contract={"comparable_fields": list(HORIZON_DEPENDENT_PROTOCOL_FIELDS)},
        supersedes=candidate_set_supersedes(
            study, name=name, declared=(supersedes_by_set or {}).get(name)
        ),
    )


def selection_catalog(study: Study, members: Iterable[str]) -> pl.DataFrame:
    """Return the registered catalog rows for one selection pool, given its member hashes.

    Takes the hashes rather than a `CandidateSet` because only a canonical pool is a set: a
    preview pool is the rows its own reduced run produced, and there is no immutable object to
    pass. The check that every member is described by the catalog applies to both.
    """
    members = list(members)
    table = study.backtests.table(include_preview=True).filter(
        pl.col("backtest_hash").is_in(members)
    )
    if table.height != len(members):
        raise ValueError("the backtest catalog does not describe every candidate")
    return table.select(
        "label",
        "family",
        "config_name",
        "checkpoint_kind",
        "checkpoint_value",
        "stage",
        "signal_method",
        "allocation_method",
        "risk_method",
        "sharpe",
        "max_drawdown",
        "num_trades",
        "avg_turn