fix(server): auto-reap stale sockets and add `jcode server stop` (#291)

After an `install.sh`/auto-update upgrade, the old daemon can survive and
keep answering on the runtime socket (or leave a dead socket behind), so a
fresh launch either talks to stale code or wedges in a connect-retry loop
(issues #277/#291).

Two complementary fixes:

1. `reap_stale_socket_if_dead`: on the client launch path, before spawning a
   server, unlink a provably-dead socket pair. Safety: only reaps when there
   is no live listener AND we can acquire the exclusive daemon `flock` (a live
   daemon holds that lock for its whole lifetime), with a post-lock re-check to
   close the startup race. This can never disturb a running server.

2. `jcode server stop`: graceful retirement for installers/users. Looks up the
   daemon owning the active socket in the registry, sends SIGTERM (the daemon
   has a graceful handler), waits for exit (escalating to SIGKILL only if it
   refuses), then reaps any leftover socket. Supports --json.

Tested live: starts a real daemon in an isolated runtime dir, `server stop`
signals it, the process exits, and the socket/lock/registry entry are cleared.
Added 3 unit tests for the reaper (dead socket+lock removed; live listener
spared; socket spared while lock held).