* MCP bridge: reconnect before idle sockets die, retry reads, actionable timeouts (#2487)
The app closes idle control connections after 30s (IO_TIMEOUT) while the
bridge kept one persistent connection, so the first call after a pause
failed with "operation may have completed". The bridge now drops and
reconnects before sending once the connection has been idle 25s, so a
request is never written to a socket the app may have dropped.
The default call timeout drops from 60s to 55s, strictly below the app's
60s reply deadline, so the bridge's own error — naming the method and the
likely causes — is the one callers see instead of a race between the
app's deadline reply and the caller's own cutoff.
Idempotent read methods (ui.inspect and friends) get one automatic retry
on a transport failure; edits keep #1007 semantics and are never resent.
A timed-out call is never retried (it would only double the wait).
* MCP bridge: caller-vouched read retries, per-failure wording, docs contract (#2487 review)
The retry decision moves from a hand-maintained method list — most of whose
entries were engine command ids that never travel as control-protocol
methods, so they could never match — to the call sites that know a call is
a read: call never resends (#1007), call_read retries once on a transport
failure, and the MCP server opts its pure-read tools in. Failures are
worded by what they prove: a connection that could never be established is
"the request was not sent", not "may have completed"; a read timeout says
a read may be retried safely instead of demanding a state inspection. The
55s client cap, the read-retry exception, and the idle reconnect are now
documented in the control protocol and the security book.
3e3e580217