Stop mutating source IO state during magic-byte detection

[andreaslillebo]

Stop mutating source IO state during magic-byte detection

Problem

Marcel::MimeType.for(io) permanently mutates the source IO's encoding. Callers that read the IO after mime detection get back ASCII-8BIT bytes regardless of how they originally opened it:

require "stringio"
io = StringIO.new("hello world")
io.external_encoding   # => #<Encoding:UTF-8>
Marcel::MimeType.for(io)
io.external_encoding   # => #<Encoding:BINARY (ASCII-8BIT)>

For StringIO objects the underlying string is also re-tagged in place. When the bytes are then JSON-encoded, this trips the json gem's UTF-8 string passed as BINARY warning, which becomes an error in json 3.0.

Cause

magic_match calls io.binmode and io.set_encoding(Encoding::BINARY) before reading magic bytes:

io.binmode if io.respond_to?(:binmode)
io.set_encoding(Encoding::BINARY) if io.respond_to?(:set_encoding)
buffer = (+"").encode(Encoding::BINARY)

MAGIC.send(method) { |type, matches| magic_match_io(io, matches, buffer) }

These two lines were carried over from mimemagic in #30 and predate the buffer parameter pattern in magic_match_io. They're redundant - magic_match_io reads via IO#read(n, buffer) with a binary-encoded buffer, which already produces a binary result regardless of the IO's mode.

Solution

This PR has two commits.

Commit 1: remove the redundant mutations. binmode and set_encoding(Encoding::BINARY) are gone. magic_match_io reads through the binary buffer.

Commit 2: remove the workarounds those mutations required. With the mutations gone, the support code added to keep them safe under Ruby 3.4 is no longer needed:

• +io.to_s (the mutable-string fallback, Fix Ruby 3.4 frozen string literal warnings with StringIO #140) - only needed because set_encoding would later try to mutate the StringIO's underlying string. With no set_encoding call, a frozen string is fine.
• io.close_write (Fix Ruby 3.4 frozen string literal warnings with StringIO #140) - only needed to make the StringIO read-only before binmode/set_encoding so Ruby 3.4 wouldn't warn about modifying the underlying frozen string. With no encoding modification, there's nothing to warn about.

The "no Ruby 3.4 frozen string warnings with StringIO" regression test from #140 still passes.

Test

Adds a regression test asserting external_encoding survives a call to Marcel::MimeType.for:

test "does not mutate the source IO's encoding" do
  io = StringIO.new("hello world")
  original_encoding = io.external_encoding

  Marcel::MimeType.for(io)

  assert_equal original_encoding, io.external_encoding
end