Breaks should be computed on pre-expanded range

[hadley]

Otherwise when you adjust the number of breaks, you can end up with breaks right on the margin of the plot

library(ggplot2)

may01 <- as.POSIXct("2020-05-01")
df <- data.frame(x = c(may01, may01 + 60 * 86400))

ggplot(df, aes(x, 1)) + 
  geom_blank() + 
  scale_x_datetime(NULL, date_breaks = "7 days", date_labels = "%d %b") + 
  scale_y_continuous(NULL, breaks = NULL)

^{Created on 2019-10-27 by the reprex package (v0.3.0)}

[paleolimbot]

This would also solve #980, although having implemented this at least once, I seem to remember that most of the test case plots looked better when the breaks were computed on the expanded range.

[thomasp85]

Closing this for now as it is not clear it will be a universally better solution