The `:day`, `:hour` and `:minute` clauses of `Time.diff/3`,
`DateTime.diff/3` and `NaiveDateTime.diff/3` compute the difference in
seconds first and then divide again, so the sub-second remainder is
rounded twice. The intermediate conversion floors while the final
`div/2` truncates, so the result follows neither rule:
# elapsed is 23:59:59.999998, less than a full day
Time.diff(~T[00:00:00.000001], ~T[23:59:59.999999], :hour)
#=> -24, while the opposite direction returns 23
# elapsed is 86399.5s, and the docs promise incomplete days round to zero
NaiveDateTime.diff(~N[2000-01-01 00:00:00.7], ~N[2000-01-02 00:00:00.2], :day)
#=> -1
Compute the difference in microseconds and divide once instead. This
keeps the documented truncation ("fractional results are not supported
and are truncated", "it also rounds incomplete days to zero") in both
directions. Positive results are unchanged, so the existing
`DateTime.diff/3` "almost 7 days" test now gains its negative
counterpart.
Assisted-by: Claude Code:claude-opus-4-8
nseaprotector
002e42d574