fix(#295): client-side detection of stale pre-self-heal servers

A server old enough to predate the self-reported staleness machinery sends
`server_has_update: None`, so it can never tell the client it is stale. The
client then attaches and a newer-protocol request (e.g. set_route) explodes
against the ancient daemon.

Add an independent client-side check: when the server cannot self-report
(server_has_update: None), compare the server's reported release version
against the client's own and defer + reload when the server is strictly older.

Selfdev-safe: gated on clean release semvers on BOTH sides. Any -dev/dirty/
unparseable version (self-dev or branched daemons, which share a base semver
and cannot be ordered) is left to the existing mtime-directional path, so we
never force-downgrade or loop a dev daemon (#277/#291).

Reuses the existing pending_server_reload machinery (force reload + #277 loop
backoff). When the older server cannot heal via reload, the user is pointed at
`jcode server stop` + relaunch.

Tests: unit coverage for the gate precedence and selfdev-safe semver parsing,
plus e2e handle_server_event tests for the ancient (None) defer case and the
current-release no-defer case.