chore(skore): Align default metrics with is_computable and discouraged - #3195

Summary

• Replace the task-branching default_metrics factory with a flat DEFAULT_METRICS catalog.
• Rename Metric.available → Metric.is_computable (technical: can we compute it?).
• Add Metric.discouraged(report) -> str | None (methodological: should we register it?). Defaults skip discouraged metrics silently; metrics.add(..., force=False) raises with the reason unless force=True.
• Move the has_default_score display filter onto Score.discouraged, so sklearn's default score is no longer registered for LogisticRegression / RandomForest / etc.
• Allow metrics.add(Score(), force=True) to put a discouraged score back into summarize / get; metrics.score() remains the dedicated escape hatch.

Behavior change

Defaults only register metrics that are both computable and not discouraged. User .add never checks is_computable; only discouraged can block it (unless force=True).

For estimators that use scikit-learn's default ClassifierMixin.score / RegressorMixin.score, "score" is skipped from the registry for the same reason (use report.metrics.score(), or add(Score(), force=True) to put it back).

Example

The point of discouraged is to refuse registering a metric that can be computed but should not be reported for the problem at hand. Override it on your metric class; .add raises with that reason unless the caller passes force=True:

from sklearn.datasets import load_breast_cancer
from sklearn.linear_model import LogisticRegression
from sklearn.metrics import accuracy_score
from skore import evaluate
from skore._sklearn.metrics import FunctionKind, Metric

class AccuracyOnImbalanced(Metric):
    """Accuracy that refuses to be registered on binary classification reports."""

    name = "accuracy_on_imbalanced"
    verbose_name = "Accuracy (imbalanced)"
    function = staticmethod(accuracy_score)
    response_method = "predict"
    greater_is_better = True
    function_kind = FunctionKind.METRIC

    @staticmethod
    def discouraged(report):
        if report._ml_task == "binary-classification":
            return (
                "accuracy is often misleading for binary classification; "
                "prefer precision, recall, or ROC AUC"
            )
        return None

X, y = load_breast_cancer(return_X_y=True)
report = evaluate(LogisticRegression(max_iter=10_000), X, y, splitter=0.2)

# Refused with the reason from discouraged().
report.metrics.add(AccuracyOnImbalanced())
# ValueError: Cannot add 'accuracy_on_imbalanced': accuracy is often misleading
# for binary classification; prefer precision, recall, or ROC AUC. Pass
# force=True to register it anyway.

# Opt in explicitly when you still want it.
report.metrics.add(AccuracyOnImbalanced(), force=True)
report.metrics.accuracy_on_imbalanced()
# 0.94...

Custom metrics without a discouraged override keep today's behavior: .add registers them unconditionally (still not filtered by is_computable).

Out of scope (flagged, not fixed here)

• The neg_ alias claim in summarize / get docstrings that lookup does not implement
• Metric.new silently dropping some arguments
• MetricRow / MetricsSummaryRow duplication
• Stale 'score' listing in accepted design record docs/design/0007-show-neg-when-used.md (left as history)

Test plan

• Registry / discouraged / force-add unit tests
• CV and comparison metrics.score() still work when score is not registered
• Doctests on metrics accessors
• No circular registry ↔ report reference (state size test)
• Merged with main (#3196 registry-key row names kept via resolved.name)