[tannevaled]

Closes #772. Takes the SystemExit half of #774.

What it was

rbgo raised SocketError for everything that went wrong on the way to a peer,
with getaddrinfo: in front of a Go error — a label that is simply false
for a refused connection, where no name was being resolved.

The cost is that the ordinary idioms stop working in both directions: code
rescuing Errno::ECONNREFUSED to back off and retry never matches, and code
rescuing SocketError to report a bad hostname matches a perfectly resolvable
host whose port is shut.

Measured

MRI 4.0.7 has three answers where rbgo had one:

	MRI 4.0.7	rbgo before	rbgo after
TCPSocket.new("127.0.0.1", <closed>)	Errno::ECONNREFUSED	SocketError	Errno::ECONNREFUSED
TCPSocket.new("no-such-host.invalid", 80)	Socket::ResolutionError	SocketError	Socket::ResolutionError
Socket.tcp("192.0.2.1", 80, connect_timeout:)	Errno::ETIMEDOUT	—	— (needs Socket.tcp, absent)

And the message is MRI's, byte for byte:

Connection refused - connect(2) for "127.0.0.1" port 1

Socket::ResolutionError is new and is a subclass of SocketError, so
every rescue SocketError written before this still catches a resolution
failure. A test asserts that rather than leaving it to be discovered.

Both raisers had to change, which is why the classifier is shared: TCPSocket.new
goes through socket.go, and Net::HTTP — the path #772 was filed on — through
raiseTransportErr. The Net::HTTP side passes no address, because its phase is
not always a connect and naming connect(2) there would describe a call that
did not happen.

Three tests asserted the defect

nethttp_bind_test.go  TestNetHTTPErrors
nethttp_bind_test.go  TestNetHTTPDialXferError  (three arms)
socket_test.go        TestSocketErrors

All rescued SocketError for a refused connection — green exactly while every
way of failing to reach a peer arrived as one class, so fixing the tool turned
them red. They now assert MRI's class rather than being loosened, and
TestNetHTTPErrors asserts both classes separately: Errno::ECONNREFUSED
alone would not notice a name failure being given the same one.

The first of them was reported as green by a filtered grep over the suite
output. They were found by running the whole suite and counting the
failures.

Also here: fork { exit n } stops the child, not the program (#774)

pid = fork { exit 7 }
_, st = Process.wait2(pid)
puts "status=#{st.exitstatus}"      # rbgo before: never reached, process exits 7

rbgo runs the block in-process by design, so the block is the child, and
the child's exit is runForkBlock returning the status rather than a panic
escaping to the top. exit, exit! and abort all travel that path and all
three now match MRI; a block that returns normally is still 0, as the control.

This does not close #774. An uncaught exception in the block still takes
the parent down, where MRI prints it in the child and exits 1. That needs an
exception renderer inside the VM — the one that exists lives in the CLI — and
is left deliberately rather than done halfway.

The tests get their closed port by binding a listener and closing it. "Port 1
is surely shut" is a statement about the host a test happens to run on.

🤖 Generated with Claude Code