src/common/leaderboard.py

"""Shared leaderboard renderer for the Calibration Benchmark dashboard.

Extracts the scored-table pipeline from the home page so both the Optimization
leaderboard (home) and the UQ leaderboard (pages/UQLeaderboard.py) can reuse
it without duplicating code.

Usage::

    from common.leaderboard import render_leaderboard

    render_leaderboard(
        metric_store,
        target_col="rmse_target",
        target_label="RMSE Target Level",
        title="Optimization Leaderboard",
        state_prefix="opt",
        raw_page="pages/RawData.py",
    )
"""
from __future__ import annotations

import altair as alt
import pandas as pd
import streamlit as st

# Stable categorical palette (Vega tableau10). Methods are assigned colors by their
# sorted position in the full available-method list, so toggling visibility never
# reassigns colors to other methods.
_METHOD_PALETTE = [
    "#4c78a8", "#f58518", "#e45756", "#72b7b2", "#54a24b",
    "#eeca3b", "#b279a2", "#ff9da6", "#9d755d", "#bab0ac",
]


def render_leaderboard(
    metric_store: pd.DataFrame,
    *,
    target_col: str,
    target_label: str,
    title: str,
    state_prefix: str,
    raw_page: str | None = None,
    show_failure_panel: bool = False,
    show_scoring_modes: bool = True,
    canonical_target_levels: list[float] | None = None,
    budget_store: pd.DataFrame | None = None,
) -> None:
    """Render a scored leaderboard backed by *metric_store*.

    Parameters
    ----------
    metric_store:
        DataFrame produced by ``load_metric_store()`` or ``load_uq_store()``.
        Must contain at least the columns ``benchmark``, ``algorithm_type``,
        ``abbreviation``, ``Method``, ``family``, ``ensemble_size``, ``metric``,
        ``failure_rate``, and *target_col*.
    target_col:
        Name of the target-coordinate column, e.g. ``"rmse_target"`` or
        ``"uq_target"``.
    target_label:
        Human-readable label for the target-level radio control,
        e.g. ``"RMSE Target Level"`` or ``"UQ Target Level"``.
    title:
        Leaderboard section header text.
    state_prefix:
        Short string used to namespace ``st.session_state`` keys so multiple
        leaderboard pages keep independent control state.  Use ``"opt"`` for
        the Optimization leaderboard and ``"uq"`` for the UQ leaderboard.
    raw_page:
        Optional Streamlit page path for an "Open Raw Data" link shown at the
        bottom.  Pass ``None`` to suppress the link.
    show_failure_panel:
        If ``True``, render a grouped-bar failure-rate chart below the main
        performance chart.
    canonical_target_levels:
        If provided, the target-level selector always offers exactly these
        values (as strings) regardless of what is present in the data.  Use
        this to pin the UQ leaderboard to its fixed set of target-scaling
        levels even when some have 100 % failure.
    budget_store:
        Optional DataFrame produced by ``load_uq_budget_store()``.  When
        provided, an additional "Budget and Iterations for Coverage" section is
        rendered below the main performance chart, showing mean budget
        (N_ens·k_iter, solid lines) and mean iterations (k_iter, dashed lines)
        vs ensemble size, followed by a coverage-failure-rate bar chart.
    """
    st.header(title)

    if metric_store.empty:
        st.warning("No metric data found. Expected NetCDF files in `data/` with a `metric` variable.")
        return

    # Derived column name for the string version of the target coordinate
    target_str_col = f"{target_col}_str"

    benchmark_values = sorted(metric_store["benchmark"].unique().tolist())
    benchmark_options = ["All"] + benchmark_values
    selected_benchmark = st.selectbox("Benchmark", options=benchmark_options, index=0)

    filtered = (
        metric_store.copy()
        if selected_benchmark == "All"
        else metric_store[metric_store["benchmark"] == selected_benchmark].copy()
    )
    filtered[target_str_col] = filtered[target_col].astype(str)

    if canonical_target_levels is not None:
        target_options = ["All targets"] + [str(float(t)) for t in canonical_target_levels]
    else:
        target_options = ["All targets"] + sorted(
            metric_store[target_col].astype(str).unique().tolist()
        )

    scoring_options = [
        "Mean Forward Model Runs",
        "Minimum Forward Model Runs",
        "Smallest Optimal Ensemble Size",
        "Custom Blend",
    ]

    # Session-state keys namespaced by state_prefix so two leaderboard pages
    # don't share control state.
    k_target = f"{state_prefix}_selected_target"
    k_scoring = f"{state_prefix}_scoring_mode"
    k_weight = f"{state_prefix}_fwdruns_weight_percent"
    k_methods = f"{state_prefix}_selected_methods"

    current_target = st.session_state.get(k_target, "All targets")
    if current_target not in target_options:
        current_target = target_options[0]

    current_scoring_mode = st.session_state.get(k_scoring, "Mean Forward Model Runs")
    if current_scoring_mode not in scoring_options:
        current_scoring_mode = scoring_options[0]

    current_fwdruns_weight_percent = int(st.session_state.get(k_weight, 80))
    current_fwdruns_weight_percent = max(0, min(100, current_fwdruns_weight_percent))

    selected_target = current_target
    scoring_mode = current_scoring_mode
    fwdruns_weight = current_fwdruns_weight_percent / 100.0
    ensemble_weight = 1.0 - fwdruns_weight

    # Available methods for the current benchmark selection; used to populate the
    # multiselect and to prune any stale saved selections when the benchmark changes.
    available_methods = sorted(filtered["abbreviation"].dropna().unique().tolist())
    saved_methods = st.session_state.get(k_methods, available_methods)
    valid_saved = [m for m in saved_methods if m in available_methods]
    st.session_state[k_methods] = valid_saved if valid_saved else available_methods

    # Stable color scale: domain covers ALL methods so colors don't shift when a
    # subset is displayed.
    color_domain = available_methods
    color_range = [_METHOD_PALETTE[i % len(_METHOD_PALETTE)] for i in range(len(available_methods))]
    method_color = alt.Color(
        "abbreviation:N",
        title="Method",
        scale=alt.Scale(domain=color_domain, range=color_range),
    )

    def build_scored_table(input_df: pd.DataFrame, add_rank: bool = True) -> pd.DataFrame:
        ranking_source = (
            input_df
            if selected_target == "All targets"
            else input_df[input_df[target_str_col] == selected_target]
        )
        if ranking_source.empty:
            return ranking_source

        # Failure rate from every row (NaN metric rows carry failure_rate=100)
        failure_agg = ranking_source.groupby(
            ["algorithm_type", "abbreviation", "Method", "family"], as_index=False
        ).agg(**{"Mean Failure Rate (%)": ("failure_rate", "mean")})

        # Metric stats only from runs that reached the target (non-NaN metric)
        valid_rows = ranking_source.dropna(subset=["metric"])
        if valid_rows.empty:
            return pd.DataFrame()

        scored_df = valid_rows.groupby(
            ["algorithm_type", "abbreviation", "Method", "family"], as_index=False
        ).agg(
            **{"Mean Forward Model Runs": ("metric", "mean")},
            **{"Minimum Forward Model Runs": ("metric", "min")},
            **{"Targets Used": (target_str_col, "nunique")},
            **{"Ensemble Sizes Used": ("ensemble_size", "nunique")},
        )

        best_per_target = (
            valid_rows.sort_values(["algorithm_type", target_str_col, "metric", "ensemble_size"])
            .groupby(["algorithm_type", "abbreviation", "Method", "family", target_str_col], as_index=False)
            .first()[
                [
                    "algorithm_type",
                    "abbreviation",
                    "Method",
                    "family",
                    target_str_col,
                    "ensemble_size",
                ]
            ]
        )

        optimal_ensemble = best_per_target.groupby(
            ["algorithm_type", "abbreviation", "Method", "family"], as_index=False
        ).agg(**{"Optimal Ensemble Size": ("ensemble_size", "mean")})

        scored_df = scored_df.merge(
            optimal_ensemble,
            on=["algorithm_type", "abbreviation", "Method", "family"],
            how="left",
        )
        scored_df = scored_df.merge(
            failure_agg,
            on=["algorithm_type", "abbreviation", "Method", "family"],
            how="left",
        )

        scored_df["Optimal Ensemble Size"] = scored_df["Optimal Ensemble Size"].round(2)
        scored_df["Mean Forward Model Runs"] = scored_df["Mean Forward Model Runs"].round(4)
        scored_df["Minimum Forward Model Runs"] = scored_df["Minimum Forward Model Runs"].round(4)
        scored_df["Mean Failure Rate (%)"] = scored_df["Mean Failure Rate (%)"].round(1)

        mean_runs_min = scored_df["Mean Forward Model Runs"].min()
        mean_runs_max = scored_df["Mean Forward Model Runs"].max()
        if mean_runs_max > mean_runs_min:
            scored_df["mean_runs_score"] = (
                100.0 * (mean_runs_max - scored_df["Mean Forward Model Runs"]) / (mean_runs_max - mean_runs_min)
            )
        else:
            scored_df["mean_runs_score"] = 100.0

        minimum_runs_min = scored_df["Minimum Forward Model Runs"].min()
        minimum_runs_max = scored_df["Minimum Forward Model Runs"].max()
        if minimum_runs_max > minimum_runs_min:
            scored_df["minimum_runs_score"] = (
                100.0 * (minimum_runs_max - scored_df["Minimum Forward Model Runs"]) / (minimum_runs_max - minimum_runs_min)
            )
        else:
            scored_df["minimum_runs_score"] = 100.0

        ens_min = scored_df["Optimal Ensemble Size"].min()
        ens_max = scored_df["Optimal Ensemble Size"].max()
        if ens_max > ens_min:
            scored_df["ensemble_score"] = (
                100.0 * (ens_max - scored_df["Optimal Ensemble Size"]) / (ens_max - ens_min)
            )
        else:
            scored_df["ensemble_score"] = 100.0

        if scoring_mode == "Mean Forward Model Runs":
            scored_df["Score"] = scored_df["mean_runs_score"]
            sort_columns = ["Mean Forward Model Runs", "Optimal Ensemble Size", "abbreviation"]
            ascending = [True, True, True]
        elif scoring_mode == "Minimum Forward Model Runs":
            scored_df["Score"] = scored_df["minimum_runs_score"]
            sort_columns = ["Minimum Forward Model Runs", "Optimal Ensemble Size", "abbreviation"]
            ascending = [True, True, True]
        elif scoring_mode == "Smallest Optimal Ensemble Size":
            scored_df["Score"] = scored_df["ensemble_score"]
            sort_columns = ["Optimal Ensemble Size", "Mean Forward Model Runs", "abbreviation"]
            ascending = [True, True, True]
        else:
            scored_df["Score"] = (
                fwdruns_weight * scored_df["mean_runs_score"]
                + ensemble_weight * scored_df["ensemble_score"]
            )
            sort_columns = ["Score", "Mean Forward Model Runs", "Optimal Ensemble Size", "abbreviation"]
            ascending = [False, True, True, True]

        scored_df = scored_df.sort_values(sort_columns, ascending=ascending).reset_index(drop=True)

        if add_rank:
            scored_df["Rank"] = scored_df.index + 1
            scored_df["Placement"] = scored_df["Rank"].apply(
                lambda rank: f"{ {1: '🥇', 2: '🥈', 3: '🥉'}.get(rank, '')} #{rank}".strip()
            )

        return scored_df

    if selected_benchmark == "All":
        benchmark_scores = []
        for benchmark_name in benchmark_values:
            benchmark_df = metric_store[metric_store["benchmark"] == benchmark_name].copy()
            benchmark_df[target_str_col] = benchmark_df[target_col].astype(str)
            scored = build_scored_table(benchmark_df, add_rank=False)
            if scored.empty:
                continue
            scored["benchmark"] = benchmark_name
            benchmark_scores.append(scored)

        if benchmark_scores:
            combined_scores = (
                benchmark_scores[0].copy()
                if len(benchmark_scores) == 1
                else pd.concat(benchmark_scores, ignore_index=True)
            )
            leaderboard_df = combined_scores.groupby(
                ["algorithm_type", "abbreviation", "Method", "family"], as_index=False
            ).agg(
                Score=("Score", "mean"),
                **{"Mean Forward Model Runs": ("Mean Forward Model Runs", "mean")},
                **{"Minimum Forward Model Runs": ("Minimum Forward Model Runs", "mean")},
                **{"Optimal Ensemble Size": ("Optimal Ensemble Size", "mean")},
                **{"Targets Used": ("Targets Used", "mean")},
                **{"Ensemble Sizes Used": ("Ensemble Sizes Used", "mean")},
                **{"Benchmarks Used": ("benchmark", "nunique")},
                **{"Mean Failure Rate (%)": ("Mean Failure Rate (%)", "mean")},
            )

            leaderboard_df["Mean Forward Model Runs"] = leaderboard_df["Mean Forward Model Runs"].round(4)
            leaderboard_df["Minimum Forward Model Runs"] = leaderboard_df["Minimum Forward Model Runs"].round(4)
            leaderboard_df["Optimal Ensemble Size"] = leaderboard_df["Optimal Ensemble Size"].round(2)
            leaderboard_df["Targets Used"] = leaderboard_df["Targets Used"].round().astype(int)
            leaderboard_df["Ensemble Sizes Used"] = leaderboard_df["Ensemble Sizes Used"].round().astype(int)
            leaderboard_df["Mean Failure Rate (%)"] = leaderboard_df["Mean Failure Rate (%)"].round(1)
            leaderboard_df = leaderboard_df.sort_values(
                ["Score", "Mean Forward Model Runs", "abbreviation"], ascending=[False, True, True]
            ).reset_index(drop=True)
            leaderboard_df["Rank"] = leaderboard_df.index + 1
            leaderboard_df["Placement"] = leaderboard_df["Rank"].apply(
                lambda rank: f"{ {1: '🥇', 2: '🥈', 3: '🥉'}.get(rank, '')} #{rank}".strip()
            )
        else:
            leaderboard_df = pd.DataFrame(
                columns=[
                    "Placement",
                    "abbreviation",
                    "Method",
                    "family",
                    "Mean Forward Model Runs",
                    "Minimum Forward Model Runs",
                    "Score",
                    "Optimal Ensemble Size",
                    "Targets Used",
                    "Ensemble Sizes Used",
                    "Benchmarks Used",
                    "Mean Failure Rate (%)",
                ]
            )
    else:
        leaderboard_df = build_scored_table(filtered, add_rank=True)
        leaderboard_df["Benchmarks Used"] = 1

    if scoring_mode == "Mean Forward Model Runs":
        score_basis = "normalized mean of best forward model runs over selected target levels (lower is better)"
    elif scoring_mode == "Minimum Forward Model Runs":
        score_basis = "normalized minimum of forward model runs over selected targets and ensemble sizes (lower is better)"
    elif scoring_mode == "Smallest Optimal Ensemble Size":
        score_basis = "normalized mean optimal ensemble size over selected target levels (lower is better)"
    else:
        score_basis = (
            "weighted blend of normalized forward-model-runs score and normalized ensemble-size score "
            f"(forward-runs weight {fwdruns_weight:.0%}, ensemble-size weight {ensemble_weight:.0%})"
        )

    if selected_benchmark == "All":
        score_basis = f"{score_basis}; in All mode, each method's final score is the mean of its per-benchmark scores"

    # Controls expander — always shown so users can change target even when the
    # current selection yields all failures.
    with st.expander("Scoring & Target Controls", expanded=False):
        st.radio(
            target_label,
            options=target_options,
            horizontal=True,
            key=k_target,
        )

        if show_scoring_modes:
            st.radio(
                "Scoring Method",
                options=scoring_options,
                horizontal=True,
                key=k_scoring,
            )

            if st.session_state.get(k_scoring, "Mean Forward Model Runs") == "Custom Blend":
                st.slider(
                    "Blend Weight: Forward Runs vs Ensemble Size",
                    min_value=0,
                    max_value=100,
                    step=5,
                    key=k_weight,
                    help=(
                        "Higher forward-runs weight prioritizes fewer model evaluations; "
                        "higher ensemble-size weight prioritizes smaller ensembles."
                    ),
                )

        st.multiselect(
            "Methods to display in charts",
            options=available_methods,
            key=k_methods,
        )

    selected_methods = st.session_state.get(k_methods, available_methods)
    if not selected_methods:
        selected_methods = available_methods

    if leaderboard_df.empty:
        st.warning(
            "All runs failed to reach the target at this selection. "
            "See the failure rate chart below."
            if show_failure_panel
            else "No rows available for the current benchmark/target selection."
        )
    else:
        if selected_benchmark == "All":
            table_column_order = [
                "Placement",
                "abbreviation",
                "Method",
                "family",
                "Score",
                "Mean Failure Rate (%)",
                "Targets Used",
                "Ensemble Sizes Used",
                "Benchmarks Used",
            ]
        else:
            table_column_order = [
                "Placement",
                "abbreviation",
                "Method",
                "family",
                "Score",
                "Mean Forward Model Runs",
                "Minimum Forward Model Runs",
                "Mean Failure Rate (%)",
                "Optimal Ensemble Size",
                "Targets Used",
                "Ensemble Sizes Used",
                "Benchmarks Used",
            ]

        st.subheader(f"Ranked Leaderboard — {selected_benchmark}")
        st.dataframe(
            leaderboard_df,
            hide_index=True,
            use_container_width=True,
            column_config={
                "Placement": st.column_config.TextColumn("Placement"),
                "family": st.column_config.TextColumn("Family"),
                "Method": st.column_config.TextColumn("Method"),
                "abbreviation": st.column_config.TextColumn("Abbrev."),
                "Mean Forward Model Runs": st.column_config.NumberColumn("Mean Forward Model Runs", format="%.4f"),
                "Minimum Forward Model Runs": st.column_config.NumberColumn("Minimum Forward Model Runs", format="%.4f"),
                "Score": st.column_config.ProgressColumn("Score (0-100)", min_value=0.0, max_value=100.0, format="%.1f"),
                "Optimal Ensemble Size": st.column_config.NumberColumn("Mean Optimal Ensemble Size", format="%.2f"),
                "Mean Failure Rate (%)": st.column_config.NumberColumn("Mean Failure Rate (%)", format="%.1f"),
                "Targets Used": st.column_config.NumberColumn("Targets Used", format="%d"),
                "Ensemble Sizes Used": st.column_config.NumberColumn("Ensemble Sizes Used", format="%d"),
                "Benchmarks Used": st.column_config.NumberColumn("Benchmarks Used", format="%d"),
            },
            column_order=table_column_order,
        )

        st.info(
            f"Score is a normalized 0–100 ranking based on **{score_basis}**. "
            "For Mean/Minimum forward-runs scoring, values are computed from all selected metric target levels "
            "and all ensemble sizes after averaging over random seeds."
        )

        if selected_benchmark != "All":
            st.subheader("Mean Forward Model Runs vs Ensemble Size")
            chart_source = (
                filtered
                if selected_target == "All targets"
                else filtered[filtered[target_str_col] == selected_target]
            )
            chart_source = chart_source[chart_source["abbreviation"].isin(selected_methods)]
            chart_df = chart_source.dropna(subset=["metric"]).groupby(
                ["abbreviation", "ensemble_size"], as_index=False
            ).agg(mean_forward_runs=("metric", "mean"))

            all_ens_combos = chart_source[["abbreviation", "ensemble_size"]].drop_duplicates()
            ens_ticks = sorted(all_ens_combos["ensemble_size"].unique().tolist()) if not all_ens_combos.empty else []

            if not chart_df.empty:
                _ok = chart_df[["abbreviation", "ensemble_size"]].assign(_ok=True)
                fail_df = all_ens_combos.merge(_ok, on=["abbreviation", "ensemble_size"], how="left")
                fail_df = fail_df[fail_df["_ok"].isna()].drop(columns="_ok").assign(mean_forward_runs=0.0)
            else:
                fail_df = all_ens_combos.assign(mean_forward_runs=0.0)

            all_failed = chart_df.empty
            chart_layers = []
            if not chart_df.empty:
                chart_layers.append(
                    alt.Chart(chart_df)
                    .mark_line(point=True)
                    .encode(
                        x=alt.X(
                            "ensemble_size:Q",
                            title="Ensemble Size",
                            axis=alt.Axis(values=ens_ticks, format="d"),
                        ),
                        y=alt.Y("mean_forward_runs:Q", title="Mean Forward Model Runs"),
                        color=method_color,
                        tooltip=["abbreviation", "ensemble_size", alt.Tooltip("mean_forward_runs:Q", format=".4f")],
                    )
                )
            if not fail_df.empty:
                y_fwd = (
                    alt.Y("mean_forward_runs:Q", title="Mean Forward Model Runs", scale=alt.Scale(domain=[0, 1]))
                    if all_failed
                    else alt.Y("mean_forward_runs:Q", title="Mean Forward Model Runs")
                )
                chart_layers.append(
                    alt.Chart(fail_df)
                    .mark_point(shape="cross", angle=45, size=200, filled=True, opacity=1.0)
                    .encode(
                        x=alt.X(
                            "ensemble_size:Q",
                            title="Ensemble Size",
                            axis=alt.Axis(values=ens_ticks, format="d"),
                        ),
                        y=y_fwd,
                        color=method_color,
                        tooltip=[
                            alt.Tooltip("abbreviation:N", title="Method"),
                            alt.Tooltip("ensemble_size:Q", title="Ensemble Size"),
                            alt.Tooltip("mean_forward_runs:Q", title="Value (all failed)"),
                        ],
                    )
                )
            if chart_layers:
                st.altair_chart(alt.layer(*chart_layers), use_container_width=True)

    # Mean-iterations-for-coverage section (UQ only, when budget_store provided)
    if budget_store is not None and not budget_store.empty and selected_benchmark != "All":
        bf = budget_store[budget_store["benchmark"] == selected_benchmark].copy()
        bf[target_str_col] = bf[target_col].astype(str)
        if selected_target != "All targets":
            bf = bf[bf[target_str_col] == selected_target]
        bf = bf[bf["abbreviation"].isin(selected_methods)]

        iters_df = (
            bf[["abbreviation", "ensemble_size", "mean_iters"]]
            .dropna(subset=["mean_iters"])
            .groupby(["abbreviation", "ensemble_size"], as_index=False)
            .agg(mean_iters=("mean_iters", "mean"))
        )
        all_ens_combos_iters = bf[["abbreviation", "ensemble_size"]].drop_duplicates()
        if not all_ens_combos_iters.empty:
            iters_ticks = sorted(all_ens_combos_iters["ensemble_size"].unique().tolist())

            if not iters_df.empty:
                _ok_iters = iters_df[["abbreviation", "ensemble_size"]].assign(_ok=True)
                fail_df_iters = all_ens_combos_iters.merge(_ok_iters, on=["abbreviation", "ensemble_size"], how="left")
                fail_df_iters = fail_df_iters[fail_df_iters["_ok"].isna()].drop(columns="_ok").assign(mean_iters=0.0)
            else:
                fail_df_iters = all_ens_combos_iters.assign(mean_iters=0.0)

            all_failed_iters = iters_df.empty
            st.subheader("Mean Iterations for Coverage vs Ensemble Size")
            iters_layers = []
            if not iters_df.empty:
                iters_layers.append(
                    alt.Chart(iters_df)
                    .mark_line(point=True)
                    .encode(
                        x=alt.X(
                            "ensemble_size:Q",
                            title="Ensemble Size",
                            axis=alt.Axis(values=iters_ticks, format="d"),
                        ),
                        y=alt.Y("mean_iters:Q", title="Mean Iterations"),
                        color=method_color,
                        tooltip=[
                            alt.Tooltip("abbreviation:N", title="Method"),
                            alt.Tooltip("ensemble_size:Q", title="Ensemble Size"),
                            alt.Tooltip("mean_iters:Q", format=".2f", title="Mean Iterations"),
                        ],
                    )
                )
            if not fail_df_iters.empty:
                y_iters = (
                    alt.Y("mean_iters:Q", title="Mean Iterations", scale=alt.Scale(domain=[0, 1]))
                    if all_failed_iters
                    else alt.Y("mean_iters:Q", title="Mean Iterations")
                )
                iters_layers.append(
                    alt.Chart(fail_df_iters)
                    .mark_point(shape="cross", angle=45, size=200, filled=True, opacity=1.0)
                    .encode(
                        x=alt.X(
                            "ensemble_size:Q",
                            title="Ensemble Size",
                            axis=alt.Axis(values=iters_ticks, format="d"),
                        ),
                        y=y_iters,
                        color=method_color,
                        tooltip=[
                            alt.Tooltip("abbreviation:N", title="Method"),
                            alt.Tooltip("ensemble_size:Q", title="Ensemble Size"),
                            alt.Tooltip("mean_iters:Q", title="Value (all failed)"),
                        ],
                    )
                )
            if iters_layers:
                st.altair_chart(alt.layer(*iters_layers), use_container_width=True)

    # Failure panel — rendered regardless of whether the scored table has rows
    if show_failure_panel and selected_benchmark != "All":
        failure_source = (
            filtered
            if selected_target == "All targets"
            else filtered[filtered[target_str_col] == selected_target]
        )
        failure_source = failure_source[failure_source["abbreviation"].isin(selected_methods)]
        if not failure_source.empty:
            failure_df = failure_source.groupby(
                ["abbreviation", "ensemble_size"], as_index=False
            ).agg(mean_failure_rate=("failure_rate", "mean"))
            failure_df = failure_df.sort_values("ensemble_size")

            target_str = (
                "All Targets"
                if selected_target == "All targets"
                else f"Target {selected_target}"
            )
            st.subheader(f"Failure Rate of Hitting {target_str}")
            ens_ticks_fail = sorted(failure_df["ensemble_size"].unique().tolist())
            failure_chart = (
                alt.Chart(failure_df)
                .mark_bar()
                .encode(
                    x=alt.X(
                        "ensemble_size:O",
                        title="Ensemble Size",
                        sort=[str(e) for e in ens_ticks_fail],
                        axis=alt.Axis(labelAngle=0),
                    ),
                    xOffset=alt.XOffset("abbreviation:N"),
                    y=alt.Y(
                        "mean_failure_rate:Q",
                        title="Failure Rate (%)",
                        scale=alt.Scale(domain=[0, 100]),
                    ),
                    color=method_color,
                    tooltip=[
                        alt.Tooltip("abbreviation:N", title="Method"),
                        alt.Tooltip("ensemble_size:O", title="Ensemble Size"),
                        alt.Tooltip("mean_failure_rate:Q", format=".1f", title="Failure Rate (%)"),
                    ],
                )
            )
            ceiling_line = (
                alt.Chart(alt.Data(values=[{}]))
                .mark_rule(color="grey", strokeDash=[4, 4])
                .encode(y=alt.datum(100))
            )
            st.altair_chart(failure_chart + ceiling_line, use_container_width=True)

    st.caption("Top 3 are shown as podium spots; remaining methods are directly comparable via normalized score.")

    if raw_page is not None:
        st.page_link(raw_page, label="Open Raw Data & CSV Export", icon="🧾")