RandomizedSearchCV was hard-capped to min(n_iter, 5), so 'tuned' mode
ignored the user's n_iter (default 20) and the documented n_iter=30
example. RandomizedSearchCV already self-limits when the parameter grid
has fewer combinations than n_iter, so the cap was unnecessary and only
reduced tuning quality.

The catalog defines 38 models (19 classifiers + 19 regressors), not the
36 (19 + 17) stated in the docs. The SVC/SVR entries are each split into
linear + rbf variants and HuberRegressor was added; the planned
HistGradientBoosting regressor was never included. Updated README, PLAN,
CHANGELOG, and the docs site to match the implementation.

Last leftover of the old InsightML name in shipped code. The public
symbol is renamed to DISSECTML_TEMPLATE; INSIGHTML_TEMPLATE is kept as a
deprecated alias (same object) and remains exported so existing imports
keep working. The Plotly template registration key was already
'dissectml'.

The build_*_section functions in report/sections/ were never imported by
the live renderer (html_renderer.render_html_report builds its own
_section_* helpers); only tests/report/test_sections.py referenced them.
This dead duplicate was also the sole source of two reported
inconsistencies: a VIF>5 threshold (live code uses >=10 everywhere) and
attribute-style leakage access (live code correctly uses dict access on
the LeakageWarning TypedDict). Removed the module and its test.

- Remove the unused _MNAR_IMPUTER constant in preprocessing.py (imputer
  selection is decided by separate logic; the constant was never read).
- Correct InsightPipeline docstrings/error: the class docstring claimed it
  is 'used internally by dml.analyze()', but analyze() calls the stages
  directly and never touches this class. run() only wires Stage 1 (EDA)
  and is a documented stub; the error now points users to dml.analyze()
  for the full pipeline instead of implying it is unavailable.

Previously _tune_one overwrote only the primary metric (accuracy/r2) with
the RandomizedSearchCV score and kept the rest of the metrics, the std
values, and the OOF predictions from the untuned model. The leaderboard
then mixed metrics from two different CV schemes, and Stage-4 comparisons
(error analysis, ROC, significance) ran on stale, un-tuned OOF predictions.

Extract the runner's CV evaluation into a shared _evaluate_pipeline() and
re-score the tuned estimator through it on the battle's CV folds, so every
metric, std, and the OOF predictions come from one consistent scheme and
stay comparable with the untuned models in the leaderboard.

The market-research page still used the pre-rename 'import dissectml as
iml' alias in its API examples. The alias name is arbitrary so the
snippets ran fine, but they contradicted the documented 'dml' convention.
Updated all four occurrences for consistency. (File is excluded from the
package build.)

Charts inherited the theme's fixed left margin (l=60), so long y-axis
tick labels — model names on the metric-comparison bar chart, and feature/
category names elsewhere — were clipped at the left edge. Enable Plotly
automargin on both axes in the theme template so the margin auto-expands
to fit labels, and set it explicitly on the metric_bar_chart y-axis.