Respect unread buffered data when checking EOF

[OskarEichler]

Summary

Return false from BufferedIO#eof? while unread bytes remain in its own buffer. Only ask the underlying IO after that buffer is exhausted.

Reproduction

require 'net/protocol'
require 'stringio'
io = Net::BufferedIO.new(StringIO.new("one\ntwo\n"))
p io.readline # "one"
p io.eof?     # before: true; after: false
p io.readline # "two"
p io.eof?     # true

A bounded local Socket.pair reproduction also shows the old eof? blocking while a complete second line is already buffered and the peer remains open. The patched call returns false without waiting. The remaining buffered line is preserved when the peer closes.

Verification

• Ruby 4.0.6 via rbenv; reviewed master 6cba9756b72eaa5a939b4cb6e032ec1da88b07bd.
• Existing bundle exec rake test: 30 tests, 66 assertions, zero failures/errors on the baseline and this branch.
• eof: 19 checks, 0 failures in an external scratch script; baseline fails the affected expectations.
• Ruby syntax and git diff --check pass. Supplemental RuboCop Lint has the same 11 pre-existing offenses; no lint-clean claim.
• No repository tests added or modified, following the consuming project's explicit no-new-tests instruction. The reproduction is included here for maintainer use.
• Existing PRs/issues were checked for overlap, including Improve BufferedIO performance #14, Fix Net::ReadTimeout errors in net-pop by handling multi-byte terminators getting split up in readuntil #30, Fix readuntil missing a terminator that spans two reads #66 and Add limit option to Net::BufferedIO#readuntil #67. This change does not replace their buffering/terminator/read-limit fixes.

Compatibility and limitations

No signature changes. eof? intentionally stops reporting EOF, or waiting on the underlying socket, while this wrapper still has unread data.

The patch also applies to installed release 0.3.0, whose runtime file matches the reviewed master. Gem version, dependencies and Ruby >= 2.6 requirement stay unchanged. Only macOS/Ruby 4.0.6 was run locally; other Ruby/OS combinations require upstream CI. No production traffic or external service was used.