vellumplot 0.4.0

New features

• Categorical datashading in one call. mark_datashade() now accepts a mapped
  discrete color/fill aesthetic and shades categorically (datashader's
  count_cat): each category is aggregated separately and every cell is coloured by
  the count-weighted average of the category hues it holds, opacity by density — so
  you see which category dominates where, and where they mix, with a colour
  legend. Category hues come from the usual discrete colour scale
  (scale_color_*() apply). This replaces the old idiom of stacking one datashade
  layer per category with blend = "screen". New span/clip arguments clamp the
  density range (absolute limits or percentiles) so a few extreme cells don't
  flatten the rest. Requires the accompanying vellum dev version.

• mark_image() draws images at data points. A new mark places a bitmap
  image (e.g. a country flag or company logo) at each (x, y), replacing the
  usual point marker. src is a column of local file paths or one constant
  path, and size sets the height in millimetres (the width follows each
  image's native aspect ratio). Requires the suggested magick package (#33).

• coord_trans() — nonlinear display transforms. Warps the display of one
  or both position axes after the scale has trained, so gridlines bunch up and
  straight lines curve while the axis keeps its original data-value labels (unlike
  scale_*(trans=), which rescales the data and relabels in transformed space).
  Each of x/y takes a transform name ("log10", "sqrt", "identity") or a
  scales::transform_*() object, e.g. coord_trans(y = "log10"). Supports the
  common marks (points, lines, areas/ribbons, bars, tiles, smooths, text, …);
  segment/interval, boxplot, edges, and raster/datashade marks are not warped yet
  and raise a clear error under coord_trans().

• mark_violin() and mark_ridgeline() no longer clip at the edges. The
  position scales are now trained to include each mark's drawn footprint -- the
  full kernel-density support for a violin, and the ridge height (scale) above
  the top category for a ridgeline -- so the tallest ridge and the density tails
  are no longer cropped by the panel. Explicit axis limits still take precedence.

• Marginal distributions with add_marginal(). A new plot modifier draws a
  density or histogram of the panel's x variable along the top edge and of its
  y variable along the right edge, each sharing the scatter's axis so they line
  up (the vellumplot analogue of ggExtra::ggMarginal()). Pick the distribution
  with type = "density" / "histogram", the edges with sides ("t", "r",
  "tr"), and the extent with size; group = TRUE splits each marginal by a
  discrete color/fill mapping in the matching palette. It reads x/y from
  the first plain (identity-stat) layer and reuses the existing density/bin
  stats, so no axes or legends are duplicated. This version supports a single
  panel only (an error is raised with facets, a non-Cartesian coordinate system,
  or a locked aspect ratio).

• Positional constants now render. A bare literal on a coordinate channel
  (e.g. mark_segment(x = i, y = 0, xend = i, yend = value) for a lollipop
  baseline) is now treated as a constant-valued coordinate rather than a style
  param, so it populates the layer's values and trains the position scale.
  Previously y = 0 was dropped from scale training, collapsing each segment's
  start onto its end (zero-length, invisible), and a segment-only plot errored
  with "Every layer needs an x and y encoding". Segment/edge endpoints (xend /
  yend) now also extend the axis domain, so a segment-only plot derives its
  range from both ends. Affects the positional channels x, y, xend,
  yend, xmin, xmax, ymin, ymax.

• Faster mark_sf() on large maps. Non-interactive sf layers now batch all
  features that share a resolved fill/colour into a single grob per style group
  (polygons into one evenodd path_grob, lines into one NA-separated
  lines_grob), instead of one grob per feature, and the layer's bounding box is
  computed in a single linear pass rather than by growing coordinate vectors. On
  a 40k-polygon grid this cut scene compilation from ~24s to ~0.8s. Interactive
  layers (those declaring data_id / tooltip / hover_group) are unchanged —
  they still emit one keyed grob per feature so every feature stays addressable.
  Note: because a batched polygon group is filled with one evenodd rule, two
  same-fill features that overlap would cancel in the overlap region (the same
  property a self-overlapping MULTIPOLYGON already has); disjoint features —
  every real choropleth and coverage map — are unaffected.

Bug fixes

• linetype now works on stroked marks. mark_segment(), mark_linerange(),
  mark_rule(), mark_errorbar(), mark_rug(), mark_edges(), and
  mark_contour() (and annotate("segment", ...)) previously ignored
  linetype and always drew solid lines; they now honour it, consistent with
  mark_line() (#30).

• annotate("rect", ...) honours infinite bounds. A bound of -Inf/Inf
  (e.g. xmin = -Inf, xmax = Inf for a full-width band) now extends the
  rectangle to the panel edge, matching ggplot2, instead of silently rendering
  nothing (#29).

• Continuous labels no longer group thousands by default. A default
  continuous axis or legend now renders 4-digit-plus values without a grouping
  separator, so years and IDs read 2010 instead of 2 010 (#27). Pass
  explicit labels to restore grouping.

• More inputs fail fast. A design layout area must be a solid rectangle (and
  area() requires t <= b, l <= r); a per-row interactivity value
  (tooltip/data_id/…) must be length 1 or the data's row count; and
  add_marginal() requires a cleanly numeric column. Each now errors clearly
  instead of silently mis-rendering.

• Clearer alt text for grid facets. A facet_grid() plot's automatic alt text
  now distinguishes the row and column variables (e.g. "Faceted by rows cyl,
  columns am") instead of running them together.

• Compositions can be themed. theme() now accepts a composition (from
  [concat()] / [hconcat()] / [vconcat()]) and styles its figure-level chrome —
  the title/subtitle/caption bands, the collected legend, panel spacing, and
  tags. Previously a composition's figure chrome always used the default theme.
  Each sub-plot still carries its own theme.

• Gradient fills fail clearly on marks that don't support them. A
  fill = linear_gradient(...) is painted by mark_area(), mark_ribbon(), and
  mark_bar(); on any other mark it now raises a clear error instead of leaking
  an undefined paint into the renderer.

• Inputs are validated up front. theme() now checks panel.spacing,
  plot.margin, aspect.ratio, and axis.ticks.length (and rejects NA/Inf
  in any numeric setting); vplot() checks width/height the way it already
  checked dpi; a discrete shape/linetype scale with more categories than
  its palette errors instead of silently reusing a glyph; and a binned colour
  scale given fewer than two breaks errors instead of crashing in
  colorRampPalette(). Valid input is unaffected.

• Legends read consistently. A horizontal (top/bottom) continuous colour
  legend now draws the NA key that the vertical one always did.
  guide_legend(reverse = TRUE) now actually flips a continuous colourbar
  (previously a silent no-op) and no longer desyncs a binned colour scale's
  boundaries from its swatches. A merged colour+shape legend's swatches are now
  tagged for interactive highlight/select. Minor gridlines past the outermost
  breaks use the local break spacing at each end, so they sit correctly when
  breaks are unevenly spaced.

• Numeric facets order numerically. Faceting on a numeric variable now orders
  panels 1, 2, 10 rather than the lexicographic 1, 10, 2; multi-variable
  facets order by each variable's own type.

• Log/reversed axes. Bars and areas on a scale_*(trans = "log10") (or
  "sqrt") axis draw from the axis floor instead of a degenerate shape (the zero
  baseline that maps to -Inf is clamped into the trained range). A descending
  limit such as ylim(hi, lo) on a bar/area is no longer silently un-reversed by
  the zero baseline.

• Grouped histogram density. after_stat(density) / after_stat(prop) on a
  grouped histogram or bar now normalizes per group (each group integrates/sums
  to 1), matching ggplot2, rather than dividing by the grand total.

• Clearer errors instead of silent surprises or crashes. coord_trans()
  rejects a "reverse" transform (a no-op there — use
  scale_*(trans = "reverse")) and rejects bar/area marks on a nonlinearly
  transformed axis (a zero baseline has no place on a log display). Histogram /
  2-D bin / hexbin stats abort cleanly on all-NA input instead of a cryptic
  seq() error. mark_raster() and z-surface contours reject a grid with a
  duplicated cell (previously a silent transparent/NA hole). Composition
  auto-tags continue past 26 sub-plots (Z, AA, AB, …) instead of NA.

• mark_area() now stacks. mark_area(position = ...) was silently ignored;
  it gains a position argument ("stack" default, plus "fill" and
  "identity"), so areas sharing an x with a mapped fill/color combine into
  a band instead of overlapping from the zero baseline. An area with no fill
  mapping is unchanged.

• Mapped aesthetics that were silently dropped now take effect. A mapped
  fill on mark_label() colours the label background (previously it always fell
  back to white); a mapped or constant alpha is honoured per category/tick by
  mark_violin(), mark_ridgeline(), mark_contour(), mark_contour_filled()
  and mark_rug() (previously collapsed to one value or ignored), and mark_rug()
  also honours a mapped color.

• British spelling works everywhere. hover_colour/selected_colour are now
  recognised as interactivity arguments (previously only the American spelling
  was), and mark_bin2d()/mark_hex() no longer add a default count fill when a
  British colour = is supplied.

• vgraph() warns on a layout-column clash. A vertex/edge attribute named
  x/y (or xend/yend) is overwritten by the layout coordinates; that used
  to happen silently and now emits a warning naming the attribute.