Struct #is_a? fails for generic module included in abstract superstruct

[RespiteSage]

In 0.36.0, if you have a struct Sub which extends an abstract struct Sup, where Sup includes a generic module Mod(T) as Mod(Sup), Sub#is_a? Mod(Sup) will evaluate to false, when it was formerly true on 0.35.1.

For example:

module Mod(T)
end

abstract struct Sup
  include Mod(Sup)
end

struct Sub < Sup
end

puts Sub.new.is_a? Mod(Sup) # => false

Notably, this only seems to occur when the abstract superstruct is a part of the generic type of the module when it's included. So,

module Mod
end

abstract struct Sup
  include Mod
end

struct Sub < Sup
end

puts Sub.new.is_a? Mod # => true

and

module Mod(T)
end

abstract struct Sup
  include Mod(Int32)
end

struct Sub < Sup
end

puts Sub.new.is_a? Mod(Int32) # => true

and

module Mod(T)
end

abstract struct Sup
  include Mod(Set(Int32))
end

struct Sub < Sup
end

puts Sub.new.is_a? Mod(Set(Int32)) # => true

but

module Mod(T)
end

abstract struct Sup
  include Mod(Set(Sup))
end

struct Sub < Sup
end

puts Sub.new.is_a? Mod(Set(Sup)) # => false

[straight-shoota]

It seems #10214 broke this (and it stayed broken with #10224).

[straight-shoota]

Possibly related to #9660, but this used to work.

[HertzDevil]

It is not enough that Sub.new.is_a?(Mod(Sup)). The following still fails after #10304:

module Mod(T)
end

abstract struct Sup
  include Mod(Sup)
end

struct Sub < Sup
end

Sup < Mod(Sup)       # => false
{% Sup < Mod(Sup) %} # => false
Sub < Mod(Sup)       # => false
{% Sub < Mod(Sup) %} # => false

This only happens with abstract structs.

[HertzDevil]

Downcasts also fail in a manner unrelated to #10214 and not fixed by #10779:

module Mod(T)
end

abstract struct Sup
  include Mod(Sup)
end

struct Sub < Sup
end

{% p Mod(Sup).includers %} # => [] of ::NoReturn

Sub.new.as(Mod(Sup)).as(Sup)
# Error: can't cast Mod(Sup) to Sup
# 0.36.0 only: can't cast Sub to Mod(Sup)

It appears Sub is registered as an includer of Mod(Sup+) but not Mod(Sup). So maybe we should simply devirtualize the generic type arguments somewhere else...?