A publisher of variable-size rows has to size them against the server's
limit before it publishes, because exceeding it is not a truncation: the
server answers `Maximum Payload Violation` and closes the connection, so
the row is lost and the client reconnects.
The limit belongs to the queue's NixOS module, which is the only place
that chooses it. Reading it off the connection rather than restating it
as a constant keeps it spelled once — a client-side copy would be a
second opinion about a number it does not own, and would go stale the
day an operator raised the real one.
No test: the value comes from the server's INFO line, so anything
asserted about it without a running server would only be re-asserting
async-nats's own pre-connect default.
Refs #3805
da7fbd20b8