v0.17.0 — statistical graphics for regression diagnostics

Breaking changes

• Aesthetic gained the variants Xintercept, Yintercept, Slope and
  Intercept (exhaustive matches on Aesthetic need new arms).

• geom_hline / geom_vline (constant and mapped) now train the y / x
  position scale on their intercepts, as in ggplot2: a reference line outside
  the data range widens the axis instead of being drawn off-panel.

• geom_abline is now drawn in data space (y = intercept + slope · x in the
  scales' — possibly transformed — units) and clipped to the panel; it used to
  interpret slope/intercept in normalized panel units.

• StepDirection gained the variant Mid.

• PositionDodge is now a struct with width: Option<f64> and reverse: bool (a same-named const keeps .position(PositionDodge) compiling);
  PositionDodge2 gained the public fields width and reverse (construct
  it with new / default and the with_* builders). Both now dodge the
  groups present at each x (ggplot2's default preserve = "total") instead
  of reserving a slot for every group, and key groups on group + fill +
  colour together.

• Reference-line geoms no longer inherit the plot-level mapping (ggplot2's
  inherit.aes = FALSE): a plot-level color no longer leaks into
  geom_hline. New Geom::inherit_aes() (default true) controls this.

• StatQQ / StatQQLine use the exact normal quantile in every build (they
  used the Abramowitz–Stegun approximation, |error| ≈ 4.5e-4, without
  regression), and StatQQLine needs 2 sample values instead of 4.

Fixed

• geom_errorbar, geom_linerange and geom_pointrange collapsed to a
  point under coord_flip (only the x pixel of one end was used); every end
  is now transformed. geom_errorbar / geom_linerange honour a mapped
  color; an error bar is one polyline (cap–bar–cap) per row.

Fixed — dodged bars

• geom_col / geom_bar with position_dodge on a discrete x now shift
  and narrow each bar by the stored dodge offset and the number of groups at
  that x (new position::DODGE_N_COL), so grouped bars sit side by side
  instead of on top of each other.

Added — statistical geoms (reference lines, QQ, Cook's contours, KM/ECDF, dodge)

• Data-mapped reference lines (U1). Aes::{xintercept, yintercept, slope, intercept} plus GGPlot::{geom_hline_aes, geom_vline_aes, geom_abline_aes}
  (and *_aes_with(geom, aes) for custom default styling) draw one line per
  row of the layer data, per facet panel, coloured / linetyped by any mapped
  color / linetype / alpha. Constant lines are now one-row layers, so
  they appear in every facet panel. Lines carry data-series / data-value
  (the intercept) and a tooltip on the native SVG path. Non-finite intercepts
  are dropped with a warning; a numeric xintercept on a discrete axis sits
  between categories (1-based, as in ggplot2) without becoming a level, and a
  reference line added before a discrete layer no longer forces a continuous
  scale. GeomHline::mapped(), GeomVline::mapped(), GeomAbline::mapped().
• QQ plots against several distributions, with confidence bands (U2).
  QQDistribution::{Normal { mean, sd }, StudentT { df }, Exponential { rate }, HalfNormal { sd }} (ggplot2's distribution + dparams; constructors
  normal(), t(df), exponential(), half_normal()), StatQQDist,
  StatQQLineDist (line through the line_p quartiles over the range of the
  theoretical quantiles; also emits slope/intercept) and StatQQBand
  (qqplotr's stat_qq_band: QQBandType::Pointwise normal-theory envelope or
  QQBandType::Ks DKW band, level), drawn by the new GeomQQBand. Builder
  methods GGPlot::{stat_qq(dist), stat_qq_line(dist), stat_qq_band(band), geom_qq_band(), geom_qq_band_with(geom, band)}. Validated against
  ggplot2 4.0 and qqplotr 0.0.7 to 1e-9 (tests/qq_dist_r.rs,
  validation/generate_diagnostics.R). QQ points now carry data-x /
  data-value / tooltips; geom_qq_line draws one line per group.
• Cook's-distance contours (U4). GGPlot::stat_cooks_contour(p, levels)
  / geom_cooks_contour_with(GeomCooksContour) draw R's plot.lm(which = 5)
  contours ±√(level · p · (1 − h) / h) on a residuals-vs-leverage panel:
  dashed, clipped to the panel (geometrically — the native SVG has no clip
  paths), labelled with the level, in every facet panel, optionally limited
  to with_h_range(lo, hi); they train no scale. Helpers
  geom::cooks::{cooks_contour_y, cooks_distance}. Checked against R
  (tests/cooks_contour_r.rs).
• Kaplan–Meier / ECDF / horizontal-CI building blocks (U7).
  GeomStepribbon (geom_stepribbon(), geom_stepribbon_with(geom); step
  edges with StepDirection::{Hv, Vh, Mid}, ±Inf x to the panel edge) for KM
  confidence bands; censor marks geom_censor_marks(censor_col) /
  GeomCensorMarks + StatCensored (+ glyphs at rows whose column is
  > 0 / true); stat_ecdf() and stat_ecdf_band(level) / StatEcdfBand
  (simultaneous DKW band F̂ ± √(ln(2/(1−level))/(2n)), validated against R);
  geom_errorbarh() / GeomErrorbarh (xmin/xmax/y, one hoverable
  polyline per interval with data-value="xmin xmax").
• geom_step draws one step line per group (colour / group / linetype) with
  mapped linetype and data-series, and gains StepDirection::Mid
  (ggplot2's "mid"). stat_ecdf ignores non-finite input values.
• Multi-model coefficient forests (U8). position_dodge(width) /
  PositionDodge::new(width).with_reverse(bool) and
  PositionDodge2::new(padding).with_width(w).with_reverse(bool) now work on
  a discrete x axis (terms): the per-row offset is stored in
  .x_dodge_offset (position::DODGE_OFFSET_COL) and applied by
  geom_point, geom_pointrange, geom_errorbar, geom_linerange,
  geom_ribbon and censor marks, so several models' intervals per term sit
  side by side — also under coord_flip. Interval geoms map color per row
  and carry data-x (term) / data-series (model) / data-value (estimate,
  else "ymin ymax") plus a tooltip.
• examples/diagnostics.rs (runs with --no-default-features): a
  regression-diagnostic gallery on the native SVG path — QQ with pointwise +
  KS bands, QQ against t(5), residuals vs leverage with Cook's contours,
  Kaplan–Meier (survival::lung) with CI step ribbon and censor marks, ECDF
  with DKW band, dodged three-model coefficient forest, per-facet mapped ACF
  bounds — written to assets/gallery/diagnostics/*.svg.
• ggplot_rs::stat::distribution: dependency-free qnorm (AS 241), dnorm,
  pt / qt / dt (any p, any df) and ln_gamma, available in every
  feature configuration (the existing stat::dist::qt still returns the
  normal 0.975 quantile without regression).
• geom_ribbon draws one band per group (group / colour / fill) with a
  mapped fill, instead of a single polygon through every row.
• The native SVG backend draws every PointShape (square, triangle, diamond,
  +, ×) instead of falling back to circles; +/× are one stroked
  <path> per point, so they keep their data-* hover attributes.

Added — composition, label repel, GLM smooths, table brackets

• geom_smooth(method = "glm") families: SmoothFamily::Binomial(link)
  (SmoothBinomialLink::{Logit, Probit, Cloglog}, shorthand
  SmoothFamily::binomial()), SmoothFamily::Gamma(link)
  (SmoothGammaLink::{Inverse, Log}, shorthand SmoothFamily::gamma()) and
  SmoothFamily::NegativeBinomial (log link, θ by ML as MASS::glm.nb), plus
  the GeomSmooth::glm(family) builder. Bands follow ggplot2's
  predictdf.glm: linkinv(η ± qnorm(0.975)·se(η)), so they stay inside the
  response range. Validated against R glm()/predict() in
  tests/glm_smooth_r.rs.
• GGPlot::geom_bracket_table(table, BracketTable) — significance brackets
  from a precomputed test table (ggpubr stat_pvalue_manual), no test is
  recomputed. Reads the anofox test contract (group1, group2, p_adj
  falling back to p_value per row, test_id) plus optional y_position /
  label columns; label templates ("p = {p_adj}", {p}, {p.signif} /
  {stars} with configurable cutpoints, any {column}), hide_ns, automatic
  stacking above the data for rows without y_position. Rows with missing or
  unknown groups are dropped with a build warning. Pure grammar: available
  without the ggpubr feature.
• Brackets (geom_bracket*) now carry host hover metadata: a <title>
  tooltip, data-x="g1 vs g2", data-series (test_id for table brackets)
  and data-value (the p-value).
• geom_text_repel / geom_label_repel (GeomTextRepel, GeomLabelRepel,
  RepelParams, RepelDirection; ggrepel): deterministic, seeded label
  layout that avoids other labels and the labelled points and stays inside the
  panel; nudge, box_padding, point_padding, force/force_pull,
  direction, max_iter + max_time bounds, max_overlaps (dropped labels
  are reported as a warning), connecting segments beyond
  min_segment_length. Sweep-pruned collision checks; above max_labels
  (default 500) the force layout is skipped with a warning.
• DrawBackend::warn (default no-op): geoms can report draw-time warnings;
  the native SVG backend collects them (SvgBackend::take_warnings) and
  render_svg_native_with_warnings returns them after the build warnings.
• PlotGrid (ggplot_rs::compose, in the prelude): patchwork-style
  composition on the native SVG path — PlotGrid::new().add(p).ncol(2), or
  a | b (side by side) and a / b (stacked; chains flatten, mixed operators
  nest), add_spacer, nrow/byrow, relative widths/heights, spacing,
  shared title/subtitle/caption, panel tags (TagLevels::{Lower, Upper, Numeric, LowerRoman, UpperRoman, Custom} + tag_affixes), and
  collect_legends(true) (each distinct legend drawn once to the right or
  bottom, identical legends de-duplicated). Renders one SVG
  (render_svg_native[_with_size|_with_warnings|_at]); every sub-plot is a
  nested <svg> that keeps its host attributes and adds
  data-panel="<tag or index>"; the root carries data-grid="<rows> <cols>".
  Sub-plot warnings are returned prefixed panel <tag>: .
• examples/regression_diagnostics.rs: a plotters-free 2×2 diagnostics grid
  with repelled labels, table-driven brackets and a logistic GLM smooth.

Changed

• SmoothFamily::Poisson bands are now formed on the link scale and mapped
  through exp (as R/ggplot2), instead of a response-scale interval.
• ggpubr::ggarrange now delegates to PlotGrid: cells are nested SVG
  fragments without a per-cell xmlns, positioned as x="300.00", and carry
  data-panel="<index>".
• The regression feature now requires anofox-regression ^0.5.17 — the
  version the anofox-statistics DuckDB extension uses — and GLM smooths use
  that extension's IRLS settings (tolerance 1e-8, ≤ 100 iterations), so SQL
  fits and plotted smooths agree.