parse_inline_*_annotation fails on multibyte content (since #2863)

[pjocprac]

Repro

require "rbs"
require "rbs/inline_parser"

src = <<~RUBY
class C
  # 日本語
  #: () -> void
  def m; end
end
RUBY

buffer = RBS::Buffer.new(name: Pathname("t.rb"), content: src)
prism = Prism.parse(src)
result = RBS::InlineParser.parse(buffer, prism)
result.diagnostics.each { |d| puts "#{d.class.name.split('::').last}: #{d.message}" }

Expected

No AnnotationSyntaxError for the #: () -> void annotation.

Actual

AnnotationSyntaxError: Syntax error: unexpected token for inline leading annotation

Replacing # 日本語 with an ASCII-only comment (e.g. # comment) parses cleanly without diagnostics.

Environment

• rbs: 4.0.2 (also reproduces on master, 4.1.0.pre.1)
• ruby: 3.4.7
• platform: arm64-darwin24

Hypothesis

Parser.parse_inline_leading_annotation and parse_inline_trailing_annotation in lib/rbs/parser_aux.rb#L122-L130 pass the range argument directly to the C parser:

def self.parse_inline_leading_annotation(source, range, variables: [])
  buf = buffer(source)
  _parse_inline_leading_annotation(buf, range.begin || 0, range.end || buf.last_position, variables)
end

RBS::AST::Ruby::CommentBlock builds comment_buffer.ranges using character offsets. After #2863, the C parser expects byte offsets. Other parse_* methods in the same file use byte_range(range, buf.content) to convert character offsets to byte offsets, but parse_inline_*_annotation were not updated. With ASCII-only input the two coincide, hiding the bug; multibyte content makes the C parser start at an invalid byte position.

I'd like to confirm whether this was intentional or an oversight. I've prepared a proposed fix in #2945.