Fix a macOS call to a full cell reporting unavailable instead of capacity

[flavorjones]

Motivation

A full cell writes capacity and closes the connection without reading the request (Supervisor#refuse). The client's send then fails. Since #67, Transport::Socket#deliver ignores that failure when it is Errno::EPIPE or Errno::ECONNRESET, so receive reads the capacity answer.

On macOS the send sometimes fails with Errno::ENOTCONN instead. deliver doesn't rescue it, so the call reports unavailable. test_an_answer_that_arrives_before_the_request_is_sent_is_still_the_verdict failed on the macOS job for this reason (run):

unavailable: Errno::ENOTCONN: Socket is not connected - sendmsg(2).
Expected: "capacity"
  Actual: "unavailable"

The XNU kernel's sosend checks the socket under its lock. It then drops the lock to copy the request in, and calls uipc_send without checking again. If the cell's close lands in that window, uipc_send tests SS_ISCONNECTED before SS_CANTSENDMORE and returns ENOTCONN instead of EPIPE. In a probe on a macOS runner (run), sendmsg raised ENOTCONN in 5 of 5,000 tries, and the answer was readable after every one. On Linux it never did.

Clients on Linux are not affected.

Details

deliver also rescues Errno::ENOTCONN.

On Linux, ENOTCONN from the send means the socket never connected because the cell's backlog was full (see the comment on call). That call still reports unavailable at once, because the read that follows fails with Errno::EINVAL. Its detail changes from Errno::ENOTCONN: Transport endpoint is not connected - sendmsg(2) to Errno::EINVAL: Invalid argument. ControlTimeoutTest covers that path.

The new test stubs Connection#send_message to raise Errno::ENOTCONN once the cell's answer is on the socket. Before this change it failed with the same message as the macOS job.

Fixes #110

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.