feat: plot pass aligned to P1–P4 — lift bars, similarity heatmap, time-quality scatter

[rxavier]

Summary

P6 — plot pass aligned with P1–P4. Three new high-value diagnostic plots:

error_lift_bars()

Bar chart of error lift by target class/bin from ErrorReport.lift_by_target. Lift > 1 means the class is over-represented in errors. Dashed line at lift=1 for reference.

from poniard.error_analysis import ErrorAnalyzer
ea = ErrorAnalyzer.from_poniard(clf)
report = ea.analyze(X, y)
PoniardPlotFactory(X, y, clf).error_lift_bars(report.lift_by_target)

similarity_heatmap()

Pairwise prediction-error similarity between estimators via get_predictions_similarity(on_errors=True). Shows which models fail on the same samples (useful for diversity-aware ensembling).

PoniardPlotFactory(X, y, clf).similarity_heatmap(X, y)

time_quality_scatter()

Metric vs fit_time scatter with the Pareto-optimal front highlighted in red. Dummies shown in purple.

PoniardPlotFactory(X, y, clf).time_quality_scatter()

Kept: partial dependence, all existing plots.

Tests

• 210 tests pass, ruff clean
• 9 new tests covering all new plots (classifier, regressor, custom metric, top_n, etc.)