better locale robustness

[MichaelChirico]

Closes #6140; follow-up to #6074

Looking more carefully, coerceAs() will receive a factor as-is from R unless it's doing some lazy evaluation. It's only job is to map what its input into the factor structure it receives.

So I don't think coerceAs() is doing anything wrong, rather, the test is poorly structured:

# return a vector where the first element is the first level of the factor, the second is the second level
coerceAs(1:2, factor(c("x","y")))
# coerce c("x", "y") to a factor, picking up levels automatically
factor(c("x","y"))

i.e. in the second case, 'x' will always come first regardless of which level it is, while in the first, the first level always comes first.

Therefore I think the proposed edit (to always supply levels explicitly) is appropriate -- it abstracts the base-owned logic of generating levels away.

An uglier alternative is to use coerceAs(order(c('x', 'y')), factor(c('x', 'y'))). I don't see any real advantage of this approach.

[jangorecki]

LGTM