feat: overhaul error analysis (roadmap 1.2, 3.1-3.3)

[rxavier]

Summary

Overhauls ErrorAnalyzer to make it a first-class feature (roadmap items 1.2 — fix bugs in differentiated modules, 3.1 — test coverage, 3.2 — data passing, 3.3 — documentation).

Changes

Bug fixes

• analyze_features called _train_test_split_from_cv() with no X, y (would NameError); now passes them explicitly.
• Positional-vs-name index mismatch in analyze_features: feature selection is now unified on positional indices; a positional int is never compared to a column name.
• _compute_predictions/rank_errors/analyze_target reached into self._poniard.X/y which were never stored on the estimator; they now require X/y explicitly (ErrorAnalyzer stays stateless).
• _rank_errors_continuous_multioutput broadcast bug (np.std(y, axis=0) returned a Series → pandas alignment failure).
• Multiclass ranking: pandas upcast row['y'] to float → KeyError: 'proba_2.0'.
• Standalone analyze_features iterated .feature_types dict without .items().

Redesign

• Return contract: rank_errors / merge_errors return plain DataFrames (dropped the {"values", "idx"} wrapper). merge_errors accepts a dict of DataFrames or a single DataFrame.
• Regression error definition: samples flagged by |residual| > (1 - error_quantile) quantile threshold (default worst 10%), replacing the global ±1 std band. Per-target thresholds for multioutput.
• Fold-consistent importances: permutation importances averaged over the Poniard estimator's CV folds instead of a random 80/20 split.
• Interpretation: analyze_target returns error_count / target_count / error_rate per class/bin (always a DataFrame; as_ratio/wrt_target removed). Categorical features in analyze_features get an error-rate-per-category table.
• Orchestrator: new analyze(X, y, ...) returns a single report with ranked_errors, merged_errors, summary, by_target, by_feature.

Tests & docs

• New tests/test_error_analysis.py with 33 tests (all target types, error paths, report structure, quantile threshold).
• ErrorAnalyzer section added to README.

Breaking changes

• rank_errors/merge_errors return DataFrames instead of {"values", "idx"} dicts.
• analyze_target drops as_ratio / wrt_target; now always returns a DataFrame with error_rate.
• analyze_features requires X (and y when estimator_name is used).
• rank_errors requires X and y when built via from_poniard.

Tests

pytest — 128 passed (95 existing + 33 new).