Changes in version 0.2.0                        

Breaking changes

  - Row counts for multi-segment groups changed. The old per-segment
    sigmoid architecture produced n * (k-1) - (k-2) rows for k knots.
    The sigmoid method now produces n * (k-1) - (k-2) rows (identical),
    and the hermite method produces n * (k-1) + 1 rows. Code that
    depended on exact row counts may need updating.

Bug fixes

  - Fixed segment-join discontinuity in multi-period ribbons (#1). When
    a group had 3+ time points, the v0.1.0 architecture computed each
    sigmoid segment independently and concatenated them, producing two
    artefacts at every interior knot:
    
    1.  Positional gap (C0 discontinuity): The logistic sigmoid σ(t) on
        t ∈ [−s, s] never reaches exactly 0 or 1, so adjacent segments
        did not share the same y-value at their join. For smooth = 8 the
        gap was ~0.00057 data units per join — sub-pixel in most cases
        but present.
    
    2.  Derivative kink (C1 discontinuity): Each segment's slope at its
        endpoint is σ′(±s) · Δy / Δx, which depends on the segment's
        y-range. At a direction reversal (e.g. rank goes 1→5→2), the
        outgoing slope of segment k and the incoming slope of segment
        k+1 differ, producing a visible corner. This was the artefact
        reported in the screenshot.
    
    Both are now fixed. See method parameter below.

  - Duplicate x-values within a group are now silently deduplicated
    (keeping the first occurrence) instead of producing degenerate
    zero-width sigmoid segments.

  - smooth <= 0 no longer crashes the sigmoid method. Previously, smooth
    = 0 caused a division-by-zero (σ(0) − σ(−0) = 0) producing NaN
    output. Now smooth <= 0 falls back to linear interpolation via
    stats::approxfun(), following base R's pattern where degenerate
    parameters yield the limiting case rather than an error.

  - width < 0 no longer produces invisible ribbons. Negative width is
    now treated as positive (abs(width)), consistent with how ggplot2's
    own geoms handle negative size parameters. Previously, negative
    width inverted ymin/ymax, causing GeomRibbon to render nothing
    silently.

  - n < 2 now errors with an informative message via cli::cli_abort(),
    following the ggplot2 convention for parameter validation in
    compute_group. Previously, n = 0 crashed with an opaque "arguments
    imply differing number of rows" error, and n = 1 produced a
    single-point degenerate output.

  - scale_fill_rank() now defaults to limits = NULL (auto-range from
    data) instead of limits = c(1, 60). The previous default meant that
    any dataset with fewer than ~30 groups would show nearly
    monochromatic fills, because all avg_y values mapped to the green
    end of the gradient. Users who need explicit limits can still pass
    limits = c(1, 60) or any other range.

New features

  - New method parameter for geom_bump_ribbon() with two options:
    
      - "sigmoid" (default): preserves the logistic S-curve visual
        character. Each segment is a clamped sigmoid (rescaled so
        endpoints map to exact knot values), with a Hermite-basis
        derivative correction that drives the slope to zero at every
        interior knot. This eliminates both the positional gap and the
        derivative kink while keeping the sigmoid shape and honouring
        the smooth parameter.
    
      - "hermite": uses stats::splinefunH() with zero slopes at all
        knots. A single spline is evaluated across all knots in one
        pass, with C1 continuity guaranteed by construction. The smooth
        parameter is ignored. Visually similar (the cubic smoothstep 3t²
        − 2t³ closely approximates the logistic function) but not
        identical.

  - The smooth parameter now works correctly with the "sigmoid" method
    across all values including previously broken edge cases (smooth = 0
    produced constant output, negative values reversed the curve).

Internal changes

  - New internal functions smooth_path(), smooth_path_sigmoid(), and
    smooth_path_hermite() in R/sigmoid.R. The original sigmoid_path() is
    retained for backward compatibility but is no longer called by
    StatBumpRibbon.

  - Removed unused importFrom(rlang, .data) from NAMESPACE.

  - Added cli to Imports for cli::cli_abort() parameter validation,
    following the ggplot2 convention. Zero marginal dependency cost
    since cli is already a transitive dependency of ggplot2.

  - Test suite expanded from 22 to 50 tests, covering both methods,
    segment-join accuracy, duplicate-x handling, smooth parameter
    effect, input validation (n, width, smooth), and scale_fill_rank
    auto-range.

                        Changes in version 0.1.0                        

  - Initial release.
  - geom_bump_ribbon() — sigmoid-curved filled ribbons between rank
    positions.
  - geom_bump_line() — sigmoid-curved lines between rank positions.
  - StatBumpRibbon / StatBumpLine — custom Stats computing logistic
    sigmoid paths.
  - scale_fill_rank() — convenience green-yellow-red gradient scale.
  - theme_bump() — dark theme for rank comparison infographics.
  - Multi-segment support (3+ time points chain automatically).
  - 3-bend "exit–channel–enter" pattern via 4 x-values per group.
  - after_stat(avg_y) for rank-based gradient fills/colours.
  - avg_y inverse-transforms through scales$y$trans$inverse() so
    scale_y_reverse() works transparently with fill/colour mapping.