v0.24.0

Minor bump: geom_col() and geom_area() now stack by default, like
ggplot2 (an output change; see Migration below), plus legend-layout
and missing-value stacking fixes from the #167 review, each guarded by a
test.

Changed

• geom_col()'s default position is now position_stack(), matching
  geom_bar()'s (#169) — ggplot2 defaults geom_col to position_stack
  too, and 0.22/0.23 left it at PositionIdentity, a parity gap called
  out in review #167 finding 3 and tracked as #169. Migration: a plot
  that calls geom_col() with more than one group (e.g. a mapped
  fill/colour) at the same x and relies on the old identity behaviour
  (bars drawn on top of each other from y = 0, not stacked) must now
  add an explicit .position(PositionIdentity) to keep that output; a
  plot that already set .position(...) explicitly, or that never has
  more than one group at a shared x, is unaffected. The ggplot CLI's
  --geom col changes the same way (src/bin/ggplot.rs, same
  GGPlot::geom_col() call). tests/default_output_023.rs's
  all_positive_stack case (geom_col()'s default, two groups at one
  x) exercises this — see its updated expected hash and data-fn comment;
  tests/issue169_geom_col_default_stack.rs checks the default matches
  an explicit PositionStack and no longer matches PositionIdentity.
  tests/integration.rs's test_layer_position_override_build switched
  from geom_col() to geom_point() to keep testing an override of an
  identity default.
• geom_area()'s default position is now position_stack() too,
  matching ggplot2 (review #167 finding 5 on #171) — the same parity gap
  as #169, landed in this same minor so every default-position break
  ships in one release instead of trickling in across several.
  Migration: a plot that calls geom_area() with more than one
  series (a mapped fill/colour/group) at the same x and relies on
  the old identity behaviour (bands drawn on top of each other from
  y = 0, not stacked) must now add an explicit
  .position(PositionIdentity); a plot with only one series, or that
  already set .position(...) explicitly, is unaffected. The ggplot
  CLI's --geom area changes the same way. tests/default_output_022.rs's
  area case changes too, for an unrelated reason worth calling out: its
  fixture (band_data()) happens to carry its own pre-existing ymin
  column (there for geom_ribbon's case, not this one), which
  PositionIdentity always left untouched but PositionStack now
  overwrites with the (single-series, all-zero) baseline it computes —
  see that file's updated expected hash and docstring note.
  Known limitation: ggplot2 pairs geom_area with stat_align
  (each series interpolated onto the union of x before stacking);
  ggplot-rs stacks by exact x, so a series missing an x makes the series
  above it dip at that x (#172).
  tests/geom_area_default_stack.rs checks the default matches an
  explicit PositionStack and no longer matches PositionIdentity.

Fixed

• A vertical legend's explicit ncol grid that auto-reduces its column
  count to fit (#159) now reclaims the panel width that reduction no
  longer needs, instead of keeping the wider, pre-reduction column
  count's reservation forever (#168, review #167 finding 5 on #171).
  Reclaiming it is a single bounded retry, not an unconditional
  recompute: the tighter reservation is re-checked against the same fit
  test the reduction loop uses, against every currently-planned guide
  — not only the guide(s) the reduction loop itself reduced — and is
  kept only if every one of them still fits under it. A follow-up review
  round caught that first cut checking only the reduced guides: legend_x
  moves for every guide sharing the row once the reservation tightens,
  not just the one(s) reduction touched, so a second guide that already
  fit at the old, wider legend_x (its base width estimate is a fixed
  6·text_size, unrelated to the reduced guide's own labels), or whose
  labels were truncated against that old legend_x, shifted right by
  the same amount with nothing re-checking it — up to 40px off-canvas
  with no warning in a 7,992-plot sweep, 7.3% of cases with two
  simultaneous guides. Checking every plan removed all of those. A grid
  that would overflow at the tighter reservation even after that (the
  pre-existing #159 limitation where an explicit ncol(1) grid has no
  fit check of its own) keeps the wider, already-safe reservation
  instead, so this can only shrink the reservation when doing so is
  provably safe, never regress a plot that already rendered correctly.
  tests/legend_wrap.rs's legend_panel_width_is_reclaimed_after_column_reduction
  and legend_panel_width_reservation_is_not_loosened_when_unsafe cover
  the single-guide case (both sides); two_guides_ncol_retighten_repro_a_stays_on_canvas
  and ..._repro_b_stays_on_canvas cover the two-guide regression
  (untouched guide and truncated guide respectively).
• A bottom/top legend mixing a colourbar (a continuous colour/fill
  scale) with a discrete guide now leaves the discrete guide only the
  row width left over after the colourbar's own fixed share, instead of
  splitting the whole row width among the discrete guides alone — the
  wrapped discrete grid could still run past the canvas once the
  colourbar's share was added back on top (review #167 finding 7).
  tests/legend_wrap.rs's bottom_legend_colourbar_leaves_room_for_the_discrete_guide.
• A PlotGrid's collected-legend path (collect_legends(true)) now
  propagates the warnings its own legend planning produces (e.g. an
  invalid nrow(0)/ncol(0) on the collected legend), instead of
  silently discarding them — unlike the same setting on a single plot's
  own legend, which already warned (review #167 finding 6).
  tests/legend_wrap.rs's plotgrid_collected_legend_propagates_its_own_warnings.
• The legend bbox pre-pass used to measure a PlotGrid's collected
  legend (compose::capture_legend's internal Measure backend) now
  forwards a drawn label's <title> tooltip to its inner backend
  instead of silently dropping it (review #167 finding 4) — latent
  today (a collected legend's labels never get 0.23's truncate-with-
  <title> treatment yet), but would otherwise have resurfaced as soon
  as they did. src/compose.rs's
  measure_forwards_the_title_to_its_inner_backend unit test.
• A missing (NaN/SQL NULL) y is now skipped in position_stack
  and position_fill (it contributes 0 to its bucket's running
  total/denominator) and its own y/ymin/ymax are left NaN — not
  drawn — matching ggplot2's pos_stack/position_fill exactly (review
  #167 finding 2/3 on #171). ±Inf is treated as missing the same way.
  This only matters for geom_line/geom_step/
  geom_path, which keep a missing row as a line break instead of
  dropping it (Geom::keeps_na_as_break); every other geom already drops
  non-finite rows before Stat/Position ever see them. Also fixes a
  pre-existing, unrelated bug the same code touches: a Value::Na y
  (SQL NULL, as opposed to a float NaN) used to fall back to 0.0
  via as_f64().unwrap_or(0.0) and get stacked as a real point at the
  top of its bucket — a line drawn straight through a row the "removed 1
  row" warning said was gone. src/position/stack.rs's
  missing_y_contributes_zero_and_does_not_poison_its_bucket and
  value_na_is_treated_like_float_nan_not_zero, src/position/fill.rs's
  sibling test, and tests/position_stack_r.rs's
  stacked_line_with_a_missing_y_leaves_other_series_intact_at_that_x
  (an end-to-end geom_line regression, since keeps_na_as_break is
  what makes this reachable at all).