fix: stream archives instead of holding them in memory, and close a leaked socket

[tas50]

1. The build context was assembled entirely in memory

Tar.pack_directory wrote the whole archive into a StringIO before sending a byte. A build context is whatever the caller points at — a Rails app with its assets, a monorepo subtree, a directory holding a model checkpoint — so its full size was charged to the heap on top of what the daemon was about to receive.

It now packs to a Tempfile and streams. The connection layer already sends a readable body chunked, so the archive never needs to be resident.

pack_dockerfile still uses StringIO — a short generated Dockerfile has no reason to touch disk.

2. archive_in slurped the IO it was given

body: archive.respond_to?(:read) ? archive.read : archive,

That defeats the point of accepting an IO — a container filesystem is exactly the kind of archive nobody wants in memory. It is handed over as it stands now.

3. The fix both of those depend on

attach_payload chose between body_stream and body with a class check:

when IO, StringIO

which looks exhaustive and is not. Tempfile is a delegator around File, so Tempfile.new.is_a?(IO) is false. A packed context would have fallen through to body = payload and gone out as the delegator's to_s — a string like #<Tempfile:0x...> where the daemon expected a tar.

What Net::HTTP actually needs from a body stream is read, so that is what is asked for. There is a test asserting the request contains real archive bytes and the word Tempfile appears nowhere in it.

4. A failed TLS handshake leaked the socket

sync_close only ties the SSLSocket to the socket underneath once the SSLSocket exists and owns it. An expired certificate, a hostname mismatch or an untrusted CA raised before that, and dial converted it to a ConnectionError with nothing closing the descriptor — so a retry loop waiting for a daemon to come up exhausts file descriptors rather than failing cleanly.

The test stands up a TCP server that never speaks TLS and asserts the socket the transport opened is closed after the failure.

Note on the allowlist

spec/zero_dependency_spec.rb (added in #19) caught require "tempfile" immediately, which is the guard doing its job. tempfile is a default gem, not a bundled one — confirmed by loading it with GEM_HOME/GEM_PATH empty — so it is safe for a gem that declares no runtime dependencies, and it is allowlisted explicitly rather than by loosening the rule.

Verification

• bundle exec rake — 333 runs, 1128 assertions, 0 failures
• steep check clean, chefstyle clean
• DOCKER_API_NG_INTEGRATION=1 rake integration — 19 runs, 49 assertions, 0 failures against a real daemon, which is the run that matters here: it puts a genuinely streamed tar through /build and /containers/{id}/archive.
• spec/archive_streaming_spec.rb — 2 failures and 2 errors on main.