I think we don't need this. The original issue is about utf-16, but another realistic case is other ascii-compatible encoding method defined by code with magic comments.

#encoding: sjis
def a【codepoint=825C letter here】; end

So we don't need to consider UTF-16 specific case.
And It looks like BOM is not preserved.

'😄'.encode(Encoding::UTF_16)
# => "\xFE\xFF\xD8\x3D\xDE\x04" (has BOM)
'😄'.encode(Encoding::UTF_16).encode(Encoding::UTF_8).chars
# => ["😄"] (BOM is removed)

Other part of this pull request looks good 👍

This is true for all tested Ruby implementations except TruffleRuby.

However, TruffleRuby adds the BOM in the beginning of the string after a conversion.

Benoit would rather see that Ruby would generally prevent such method definitions.

I see, but I don't think this should be in lib code.
Can you remove this and omit truffleruby in the added test, just like test_regexp_completor_handles_encoding_errors_gracefully does?

I see. That makes sense. 🙂

Thank you, I changed it.

It would be better to use Encoding::UTF_16LE/BE for testing, Encoding::UTF_16 is basically deprecated and a dummy encoding.
Even better would be to use something more realistic than UTF-16, since that's not even valid source encoding: #52 (comment)
So I think testing with SJIS as in @tompng's example is much more realistic and useful, using UTF-16 is basically a user mistake in the first place.

Adding a realistic test case looks good. I'd like to check for some existing test that use UTF16LE, encoding-invalid method and add/change in a followup pull request.