This PR aims to fix #5770.

Briefly, it adds a palette argument to discrete scales that can be used to customise the position discrete variables are placed in.
Demonstrations perhaps show off the functionality better than text descriptions.

The reprex from the issue works:

devtools::load_all("~/packages/ggplot2")
#> ℹ Loading ggplot2

p <- ggplot(diamonds, aes(clarity, price)) +
  geom_boxplot()

p + scale_x_discrete(palette = \(x) c(1, 3, 4, 6, 7, 9, 10, 12))

You can use the palette to invert the order:

p + scale_x_discrete(palette = \(x) -seq_along(x))

If two limits share a position, position adjustments treat these like a single position (invoking dodging in the example) but axis guides show duplicated labels.

p + scale_x_discrete(palette = \(x) c(1, 2, 2, 3, 3, 4, 4, 5))

Invalid palette output is checked:

p + scale_x_discrete(palette = \(x) x)
#> Error in `scale_x_discrete()`:
#> ! The `palette` function must return a <numeric> vector.

p + scale_x_discrete(palette = \(x) seq_along(x)[-1])
#> Error in `scale_x_discrete()`:
#> ! The `palette` function must return at least 8 values.

When mixing the discrete scale with continuous values in a different layer;

p2 <- p + annotate("line", x = c(1, 4.5, 8), y = c(5000, 10000, 5000))
p2

The continuous values stay as-is

p2 + scale_x_discrete(palette = \(x) c(1, 3, 4, 6, 7, 9, 10, 12))

^{Created on 2024-03-12 with reprex v2.1.0}

There are at this moment two outstanding issues:

• Discrete non-position scales take palette functions that take n as argument. In this PR, palette functions take the limits as arguments. Should I let palettes take n as an argument instead? I think the benefit of the current approach is that one could design a lookup table as palette argument.
• I did not understand one test, see comment below.