While investigating the test failures reported in #6171 I ran across a CGI test that fails:

req = Net::HTTP::Get.new("/webrick.cgi/%A4%DB%A4%B2/%A4%DB%A4%B2")
http.request(req){|res|
  assert_equal("/\xA4\xDB\xA4\xB2/\xA4\xDB\xA4\xB2", res.body, log.call)
}

This URL gets parsed out and passed to a CGI subprocess via an environment variable, but because it does not decode as valid UTF-8 it's not possible for JRuby to either set or get that variable without mangling the original bytes.

Unfortunately the root cause is Java's own System.getenv, which promises a map from String to String and therefore must be able to decode environment variables to characters.

The only way we can support this use case, where differently-encoded text or undecodable bytes are stored in an environment variable, would be to bypass System.getenv and go straight to the native getenv function. Since this is a much larger fix than I wanted to attempt for #6171, I'm filing this issue and will expand WEBrick's own test exclusion to also exclude the "java" platform.