Ruby 4.0: $/ and $-0 stop storing the object they were handed

internal/vm/globals.go

@@ -412,7 +412,16 @@ func (vm *VM) storeGVar(name string, v object.Value) {
 		}
 		vm.lastMatch = object.NilV
 		return
-	case "$/", "$-0", `$\`, "$,":
+	case "$/", "$-0":
+		// ⛔ $/ and $-0 take a DIFFERENT setter from $\ and $, since 4.0, and
+		// reading the four as one would carry the new freezing into two
+		// variables MRI deliberately left alone (io.c v4.0.9 Init_IO):
+		//     rb_define_hooked_variable("$,",   …, rb_deprecated_str_setter);
+		//     rb_define_hooked_variable("$/",   …, deprecated_rs_setter);
+		//     rb_define_hooked_variable("$-0",  …, deprecated_rs_setter);
+		//     rb_define_hooked_variable("$\\",  …, rb_deprecated_str_setter);
+		v = vm.checkRecordSepGvar(name, v)
+	case `$\`, "$,":
 		v = vm.checkStrGvar(name, v)
 	case "$;", "$-F":
 		v = vm.checkFieldSepGvar(name, v)
@@ -451,6 +460,73 @@ func (vm *VM) checkStrGvar(name string, v object.Value) object.Value {
 	return v
 }
 
+// checkRecordSepGvar applies io.c v4.0.9 deprecated_rs_setter to $/ and $-0,
+// which from Ruby 4.0 no longer store the object they were handed:
+//
+//	rb_deprecated_str_setter(val, id, &val);       // TypeError, and the warning
+//	if (!NIL_P(val)) {
+//	    if (rb_str_equal(val, rb_default_rs)) val = rb_default_rs;
+//	    else                                  val = rb_str_frozen_bare_string(val);
+//	}
+//
+// ⛔ THE EQUALITY BRANCH COMES FIRST, and it overrides the "a frozen bare String
+// is kept" rule for the one value most likely to be assigned. A frozen, plain
+// "\n" satisfies that rule word for word and is still NOT what gets stored:
+//
+//	$/ = "\n".dup.freeze
+//	$/.equal?(that string)   # => false
+//	$/.equal?($/ as it was)  # => true, the canonical separator
+//
+// Measured on ruby 4.0.7 before this was written. No example in ruby/spec
+// reaches it — language/predefined_spec.rb uses "abc" throughout, which can only
+// take the second branch — so the suite passes either way and the identity is
+// wrong. Proposed upstream as ruby/spec#1410.
+func (vm *VM) checkRecordSepGvar(name string, v object.Value) object.Value {
+	v = vm.checkStrGvar(name, v)
+	if object.IsNil(v) {
+		return v
+	}
+	s := v.(*object.String)
+	// rb_str_equal against a one-byte US-ASCII "\n": byte equality decides it.
+	// Any ASCII-compatible encoding holding that single byte IS == "\n", and an
+	// ASCII-incompatible one (UTF-16, UTF-32) cannot hold it in one byte at all,
+	// so there is no case where the encodings would change the answer.
+	if s.Str() == defaultRecordSeparator.Str() {
+		return defaultRecordSeparator
+	}
+	return frozenBareString(s)
+}
+
+// frozenBareString is string.c v4.0.9 rb_str_frozen_bare_string:
+//
+//	if (BARE_STRING_P(orig) && OBJ_FROZEN_RAW(orig)) return orig;
+//	return str_new_frozen(rb_cString, orig);
+//
+// where BARE_STRING_P (string.c:372) is `RBASIC_CLASS(str) == rb_cString &&
+// !rb_shape_obj_has_ivars(str)`. "Bare" is therefore BOTH halves: the exact
+// class, AND no instance variables. A String subclass carrying @ivar fails both
+// and a frozen plain String passes both, which is precisely what the three
+// examples in language/predefined_spec.rb distinguish.
+//
+// ⛔ ONLY THE IVAR HALF IS TESTED HERE, and that is not an omission. In rbgo a
+// *object.String is class String by construction (classOf maps the Go type
+// straight to vm.cString), and an instance of a String SUBCLASS never reaches
+// this function at all — checkStrGvar's type assertion rejects it first, with
+// "value of $/ must be String". MRI accepts one, because its check is T_STRING
+// and a subclass is still T_STRING. That divergence is a defect of its own,
+// filed rather than papered over here: it is why
+// predefined_spec.rb's "makes a new frozen String from the assigned String"
+// still fails, and fixing it belongs with the assertion, not with this rule.
+func frozenBareString(s *object.String) *object.String {
+	if s.Frozen && genericIvarsOf(s, false) == nil {
+		return s
+	}
+	out := object.NewString(s.Str())
+	out.Enc = s.Enc
+	out.Frozen = true
+	return out
+}
+
 // checkFieldSepGvar applies string.c v3_4_0 rb_fs_setter to $; / $-F: nil, a
 // String or a Regexp is taken as is, anything else is offered #to_str
 // (rb_fs_check's rb_check_string_type) before the TypeError, and a non-nil

internal/vm/record_separator_test.go

@@ -0,0 +1,86 @@
+package vm
+
+import "testing"
+
+// Ruby 4.0 stops storing the object assigned to $/ and $-0
+// (io.c v4.0.9 deprecated_rs_setter, language/predefined_spec.rb:698-722).
+//
+// Every expectation below was measured on a real
+// `ruby 4.0.7 (2026-09-15 revision 229531a6cf) [arm64-darwin25]` before it was
+// written here, because two of them are not in ruby/spec at all.
+func TestRecordSeparatorAssignment(t *testing.T) {
+	for _, tc := range []struct{ name, src, want string }{
+		{
+			"an unfrozen String is replaced by a frozen copy",
+			`$VERBOSE = nil; s = +"abc"; $/ = s; p [$/.equal?(s), $/.frozen?, $/ == s]`,
+			"[false, true, true]",
+		},
+		{
+			"a frozen String with no instance variables is kept",
+			`$VERBOSE = nil; s = "abc".freeze; $/ = s; p $/.equal?(s)`,
+			"true",
+		},
+
+		// ⛔ NOT IN ruby/spec. The equality branch runs BEFORE the "frozen bare
+		// string is kept" rule, so a frozen plain "\n" — which satisfies that
+		// rule word for word — is still replaced, by the canonical separator.
+		// Every example in predefined_spec.rb uses "abc" and so can only reach
+		// the other branch; proposed upstream as ruby/spec#1410.
+		{
+			"a String equal to the default becomes the default object",
+			`$VERBOSE = nil; d = $/; s = "\n".dup.freeze; $/ = s; p [$/.equal?(s), $/.equal?(d)]`,
+			"[false, true]",
+		},
+		{
+			"and so does an unfrozen one",
+			`$VERBOSE = nil; d = $/; $/ = +"\n"; p $/.equal?(d)`,
+			"true",
+		},
+
+		// $-0 shares the slot, so it takes the same setter.
+		{
+			"$-0 is replaced the same way",
+			`$VERBOSE = nil; s = +"abc"; $-0 = s; p [$-0.equal?(s), $-0.frozen?]`,
+			"[false, true]",
+		},
+		{
+			"and still reads as the same object as $/",
+			`$VERBOSE = nil; $/ = +"xyz"; p $-0.equal?($/)`,
+			"true",
+		},
+
+		// ⛔ CONTROLS: $\ and $, take rb_deprecated_str_setter, NOT this one
+		// (io.c v4.0.9 Init_IO). Carrying the freezing into them would be a
+		// silent divergence in the opposite direction, and nothing else in the
+		// suite asserts it.
+		{
+			`$\ keeps the object it is given`,
+			`$VERBOSE = nil; s = +"abc"; $\ = s; p [$\.equal?(s), $\.frozen?]`,
+			"[true, false]",
+		},
+		{
+			`$, keeps the object it is given`,
+			`$VERBOSE = nil; s = +"abc"; $, = s; p [$,.equal?(s), $,.frozen?]`,
+			"[true, false]",
+		},
+
+		// Unchanged by this: nil, the TypeError, and the returned value.
+		{"nil is accepted", `$/ = nil; p $/`, "nil"},
+		{
+			"a non-String is still refused without #to_str",
+			`begin; $/ = 1; rescue TypeError => e; p e.message; end`,
+			`"value of $/ must be String"`,
+		},
+		{
+			"assignment returns the value given, not the value stored",
+			`$VERBOSE = nil; s = +"xyz"; p((($/ = s)).equal?(s))`,
+			"true",
+		},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			if got := runSrc(t, tc.src); got != tc.want {
+				t.Errorf("%s\n got %q\nwant %q", tc.src, got, tc.want)
+			}
+		})
+	}
+}