is should imply truthiness for non-boolean non-nil types

[euclidianAce]

(Possibly related to/duplicate of #776, or maybe a special case of it?)

Consider the following:

local record R end

local function convert(_: string): R end

local u: string | R
local _r: R = u is R
	and u
	or convert(u)

$ tl check example.tl
example.tl:8:13: argument 1: got string | R, expected string

The only case where the argument to convert will not be a string is when u itself is nil. But since u is R and u (-> type(u) == "table" and u) will never produce a nil, convert's argument is a string (well, string | nil).

In fact this error can be avoided by using and assert(u) instead of just and u.

I haven't dug around much in the fact inference code so I'm not sure how feasible it is to say a given variable should be truthy in a given context. Also with custom is implementations it may not be as simple as this, but it would be nice to not need a runtime assert for this.

[hishamhm]

local _r: R = u is R
	and u
	or convert(u)

To make this work (i.e, given a , we would have to special-case the var is T and var or __ construct.

If the and part contains anything other than the plain variable tested with is, that part may be nil, and then the or part would still run when var is T (as shown in #776), so if var was of the type T | U, we can't assume U in the or part if the and part is an arbitrary expression. But we could for this special-cased construct you described.