The hub and the document server swallowed every send failure as a counted
loss, so a 70,000-byte reliable frame returned Ok and was resent forever.
Failures now split in two. Permanent means the datagram can never be sent as
built (EMSGSIZE, EINVAL, EAFNOSUPPORT and their Winsock equivalents,
is_permanent_send_error in cultnet-rs); everything else stays a counted loss.
A caller-directed hub send returns the error and queues nothing: the packets
it created are withdrawn and their sequences reissued, so the peer never sees
a gap (a permanent failure after an earlier fragment left ends the session).
Inside a poll the failure ends that peer's session with a goodbye whose reason
names the error, reported to the hub caller as Disconnected, and never fails
the poll. The document server ends the session the same way and reports a
snapshot response it can never send as an application rejection. Tests pin
each path, the classification per OS, BytesSent after a healthy send and a zero
failure count for healthy sends.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
64ed64a894