v0.16.0 — hardened SVG output, slim dependency tree, host-integration features

Breaking changes

• The library is now an rlib only (was ["cdylib", "rlib"]), which
  unblocks static / emscripten side-module consumers. The wasm-bindgen browser
  exports (ggplot_rs::wasm::*: render_geo, render_plot, render_bar,
  render_hist, geo_bounds, scatter_frame, render_scatter_rgba,
  render_scatter_xy) moved to the new, unpublished workspace crate
  crates/ggplot-rs-wasm (cdylib). Build the browser bundle with
  wasm-pack build crates/ggplot-rs-wasm --target web --out-dir ../../web/pkg --out-name ggplot_rs.
  The library no longer depends on wasm-bindgen.
• The wasm feature is now a deprecated alias for sf (kept so
  features = ["wasm"] dependents keep compiling); it no longer adds
  wasm-bindgen/serde_json.
• New default-on feature plotters (implies the new png feature) gates
  the plotters/image-backed APIs: GGPlot::{render_svg, render_svg_with_size, render_png, render_png_with_size, save, save_with_size, ggsave},
  ggpubr::{ggarrange_png, ggarrange_save_png} and
  render::plotters_backend. Crates using default-features = false that call
  these must add features = ["plotters"]; the self-contained
  render_svg_native* path is always available. canvas now implies png.
  cli implies plotters.
• DataFrame::add_column no longer panics on a length mismatch: it pads the
  shorter side with Value::Na and records an issue
  (DataFrame::issues() / validate()); GGPlot::try_build and every
  render_*/save* then return GGError::ValidationError. This is reachable
  from GGPlot::new(Vec<(String, Vec<Value>)>) with ragged columns, which used
  to panic. New DataFrame::try_add_column rejects the column instead.
• DataFrame::group_by / unique_values keep Value::Na distinct from the
  literal string "NA" (they used to merge into one group).
• label_number(…, prefix, …) is now sign-aware: -5 with prefix "€"
  renders "-€5" (was "€-5"), and a value rounding to zero drops its sign.
• PositionJitter gained a seed: u64 field, and width/height are now
  Option<f64> (None = 0.4 × the data's resolution, as in ggplot2). Build it
  with PositionJitter::new(w, h), PositionJitter::default(), .with_width(),
  .with_height() and .with_seed() instead of a struct literal.
  GeomJitter::{width, height} are likewise Option<f64> (default None).
• Jitter and mean_cl_boot now draw from an internal SplitMix64 stream instead
  of rand::StdRng: output is still deterministic, but the exact jittered
  positions / bootstrap intervals differ from 0.15.

Added

• ggplot_rs::format::format_value(&Value) -> String — the public tooltip /
  data-label value formatter (previously the crate-private geom::tip_value);
  ggplot_rs::format also re-exports the label_* formatters.

• label_currency(prefix, suffix, big_mark, accuracy) — sign-aware currency
  labels with a configurable thousands mark ("-€1.234.567"). label_dollar
  is now implemented on top of it (output unchanged).

• tests/native_svg.rs: smoke coverage of the plotters-free SVG path across
  geoms/coords/facets/themes, run in every feature configuration.

• Value::group_key() -> GroupKey (injective w.r.t. missingness, borrows
  strings) and Value::key_str() -> Cow<str> (allocation-free form of
  to_group_key).

• DataFrame::try_add_column, DataFrame::issues, DataFrame::validate.

Changed

• Minimal dependency tree (default-features = false, features = ["sf"]):
  38 crates → 3 (indexmap, equivalent, hashbrown). No plotters, image,
  ab_glyph, chrono, rand or getrandom.
• Removed the rand dependency (and the wasm32-only getrandom dependency)
  from the library; randomness comes from a small internal seeded PRNG.
• Default jitter amount is resolution-based (ggplot2's 0.4 * resolution(x)):
  integer-spaced data keeps ±0.4, finely spaced continuous data is no longer
  smeared by a fixed ±0.4 data units.

Performance

• stat_density switches to a binned estimator (linear binning onto a grid
  with spacing ≤ bw/50 + a kernel table truncated at ±8 bw) for groups with
  ≥ 2048 values: O(n + 512·window) instead of O(n·512) kernel evaluations,
  within ~2e-5 of the exact estimate relative to the peak. Smaller groups use
  the exact sum, unchanged.
• Layers without their own data borrow the plot data instead of cloning the
  whole frame per layer.
• Discrete scale training/lookups, stat_count, position_stack and
  position_fill use borrowed keys and hash lookups instead of allocating a
  String per value and scanning the distinct x values linearly per row
  (O(n·k) → O(n)).
• DataFrame::group_by builds borrowed keys (no per-row String clones).

Fixed

• Empty manual scales (scale_color_manual(vec![]), scale_fill_manual,
  scale_shape_manual, scale_linetype_manual) no longer panic with a
  division by zero; the geom falls back to its default colour/shape/linetype.

• format_epoch_secs is total over i64 (it overflowed for i64::MIN).

• stat_binhex / geom_hex output order is deterministic (row-major) instead
  of following HashMap iteration order, so identical data renders identical
  SVG.

• label_number / label_currency with an accuracy that keeps decimals no
  longer inserts thousands marks into the fractional part
  (1234.5 @ 0.01 → "1,234.50", was "1,23,4.50"-style garbage).

Rendering robustness

• An empty plot (zero-row data, no columns, or no layers) renders an empty
  panel with its title/axes instead of failing with
  ValidationError("geom_… requires aesthetic …").
• A layer whose stat produces no rows (a density/violin of one value, loess on
  two points, …) draws nothing; the other layers still render.
• Rows with non-finite position values (NaN, ±Inf, or a transform that is
  undefined such as log10(0)) are dropped before stats, like ggplot2's
  "Removed n rows containing non-finite values". Stat output is filtered the
  same way.
• Build warnings are exposed: BuiltPlot::warnings() and
  GGPlot::render_svg_native_with_warnings(w, h) -> (String, Vec<String>).
• The native SVG backend never writes NaN/inf into an attribute: marks with
  non-finite geometry are dropped and non-finite style values fall back to
  defaults.
• Density/violin bandwidth follows R's bw.nrd0, including its fallbacks for
  zero-spread samples (previously a constant sample produced NaN).
• A position scale's transform now applies to its whole column family
  (ymin/ymax/yend, …), not only x/y.
• Faceted panels keep per-mark hover metadata (tooltips / data-x).

SVG host integration

• Hoverable marks carry data-series (colour/fill/group level) and
  data-value (the raw measured value; for stacked/filled bars the segment's
  own value) next to data-x. Stacked-bar tooltips now report the segment
  value rather than the cumulative stack top.
• The root <svg> carries the trained, expanded position domains:
  data-domain="x0 x1 y0 y1" (both axes continuous), data-xdomain,
  data-ydomain, data-xlevels/data-ylevels (JSON arrays for discrete axes)
  and data-flip="true" under coord_flip.
• GGPlot::render_svg_native_at(x, y, w, h) renders a nested, positioned
  <svg x y width height viewBox> fragment without xmlns, for composing
  dashboards without string surgery (SvgBackend::finish_fragment).

Rect bounds

• -Inf/Inf (f64::NEG_INFINITY/f64::INFINITY) in geom_rect /
  annotate("rect") xmin/xmax/ymin/ymax extend the rect to the panel
  edge (ggplot2) and do not train scales — also on discrete axes. New
  Geom::allows_infinite() hook; other geoms still drop infinite positions.

Theme / colour

• Theme presets (theme_minimal(), theme_void(), …) and theme(t) keep a
  previously set primary_color (unless the new theme sets its own), so call
  order no longer matters.
• primary_color no longer overrides colours of geoms added with
  geom_*_with(...) (Layer::explicit_style); default-configured geoms still
  take the brand colour.
• ScaleColorDiscrete::sorted(), GGPlot::scale_fill_discrete_sorted() and
  scale_color_discrete_sorted(): discrete colour levels in sorted order, so a
  level keeps its palette colour across charts regardless of data order.

Tiles

• geom_tile on a continuous axis is sized by the data resolution (smallest
  gap between distinct values, like ggplot2's resolution()); width/height
  are now multiples of it (default 1 = abutting tiles). The tile extents
  (xmin/xmax/ymin/ymax) train the scales, so edge tiles no longer spill
  past the axes and no with_expand(…) workaround is needed. New
  Geom::setup_data() hook (ggplot2's setup_data). Free facet scales also
  train on extent columns.

New chart types

• geom_candlestick() / geom_ohlc() (+ _with): financial charts from the
  new open/high/low/close aesthetics (Aes::open() …,
  Aesthetic::{Open, High, Low, Close}); configurable up/down colours;
  the y scale trains on high–low; hover carries data-value = close and
  data-series = up/down (or the mapped group).
• coord_radar() / CoordRadar: radar/spider charts — discrete x → spokes
  (first level at 12 o'clock), y → radius from 0, straight segments, rings at
  the y breaks, spokes and outside labels. geom_polygon now groups by
  group, else color/fill (no longer requires group), strokes with the
  mapped colour, and orders radar vertices by spoke; geom_line closes its
  path under radar and orders points by mapped x position. New
  Coord::is_radar() / Coord::train_scales() hooks.
• Gauges: CoordPolar::with_span(start, end) sweeps the theta range over a
  partial arc (e.g. -PI/2..PI/2 for a half donut) and fits the arc to the
  panel; geom_rect draws annulus sectors under polar coords, so a gauge is
  geom_rect bands + a geom_segment needle.
• Calendar heatmaps: geom_calendar() / geom_calendar_with(tile, monday_first)
  and StatCalendar lay a date x (DateTime, epoch seconds or
  YYYY-MM-DD) out as week × weekday tiles with month labels, Mon/Wed/Fri
  labels, month boundary outlines and square cells; spans over
  MAX_CALENDAR_YEARS (50) are clipped to the most recent years with a
  warning. Tiles' hover key (data-x, tooltip) is the ISO date.
• geom_step accepts ±Inf (stat_ecdf's padding reaches the panel edges
  again).
• Gallery: candlestick, radar, gauge, calendar examples.