feat: diagnostics-first API — ErrorReport, compare, pareto, per-sample times

[rxavier]

Summary

Roadmap phases P0, P1, P3, P4 — diagnostics-first API overhaul.

P1 — ErrorAnalyzer report type

• ErrorReport dataclass replaces loose dict from analyze()
  • universal_failures — samples every model got wrong
  • disagreement_set — samples where models split (ensembling candidates)
  • lift_by_target / lift_by_feature — error rate vs global rate
• from_poniard defaults to all non-dummy estimators when estimator_names=None

P3 — Statistical comparison

• compare() — paired fold t-tests with mean_diff, wins_a, wins_b, ties, p_value
• Excludes dummies by default, supports metric/estimator subsets

P4 — Time-quality surface

• pareto() — Pareto-optimal estimators (metric vs time)
• best_under(seconds=...) — best metric within time budget
• Both accept time_col with options: fit_time, score_time, fit_time_per_sample, score_time_per_sample

Per-sample times

• fit() now stores fold sizes and computes fit_time_per_sample / score_time_per_sample
• Both columns appear in get_results() automatically

P0 — README rewrite

• Lead with diagnostics: multi-model diagnostics positioning
• Error analysis is the hero section
• Statistical comparison and time-quality promoted to top-level sections

Tests

• 196 tests pass, ruff clean
• 22 new tests for compare, pareto, best_under, and new ErrorReport fields