A String subclass is refused by $/, $-0, $\ and $, where MRI accepts it #884

Assigning an instance of a String subclass to $/, $-0, $\ or $, raises TypeError in rbgo. MRI accepts it.

Measured

$ rbgo -e 'sub = Class.new(String); $VERBOSE=nil; $/ = sub.new("abc")'
-e:1:in '<main>': value of $/ must be String (TypeError)

$ ruby -v
ruby 4.0.7 (2026-09-15 revision 229531a6cf) +PRISM [arm64-darwin25]
$ ruby -e 'sub = Class.new(String); $VERBOSE=nil; $/ = sub.new("abc"); p $/.class'
String

Cause

internal/vm/globals.go checkStrGvar decides with a Go type assertion:

if _, ok := v.(*object.String); !ok {
    raise("TypeError", "value of %s must be String", name)
}

A *object.String is class String by construction in rbgo (classOf maps the Go type straight to vm.cString), so an instance of a subclass is a different Go value and fails the assertion. MRI's check is T_STRING, which a subclass still is — rb_str_setter calls StringValue-style validation on the type, not the class.

Why it matters now

It is why language/predefined_spec.rb's

it "makes a new frozen String from the assigned String" do
  string_subclass = Class.new(String)
  str = string_subclass.new("abc")
  str.instance_variable_set(:@ivar, 1)
  $/ = str
  ...

still fails after the Ruby 4.0 record-separator work: the example never reaches the freezing rule, because the assignment is refused one step earlier. I found this while implementing that rule and deliberately did not widen the assertion inside it — the two are separate changes, and this one touches every $x must be String global, not just the two that 4.0 changed.

Scope

The same assertion guards $/, $-0, $\ and $, (and checkFieldSepGvar has the same shape for $;/$-F, which additionally accepts a Regexp). A fix should establish how rbgo answers "is this value a String, subclasses included" once, rather than per-setter.

Refs #799