Killing a stale daemon deletes the live daemon's socket

A daemon unlinks daemon.sock on shutdown without checking whether the node at that path is still its own. So killing a stranded daemon removes the live daemon's socket: that daemon keeps running but is unreachable, and the next client spawns a replacement. kill / pkill cleanup raises the process count instead of lowering it.

S=~/.cocoindex_code/daemon.sock

ccc daemon status                               # a daemon is live (not passive — self-starts)
A=$(cat ~/.cocoindex_code/daemon.pid)
ls -i $S                                        # 179763394  ...daemon.sock   <- A's node

nohup ccc run-daemon >/dev/null 2>&1 & sleep 5
B=$(cat ~/.cocoindex_code/daemon.pid)
ls -i $S                                        # 179763419  ...daemon.sock   <- B unlinked A's node, bound its own

Both daemons still hold a socket, and both report the same path — but they are different objects, and only B's is the one at the path:

$ lsof -p $A -p $B | grep daemon.sock
python3.1 40288 ... 12u unix 0x31379fddf5fa3701 ... /Users/…/daemon.sock   <- A, unlinked
python3.1 40341 ... 14u unix 0x72f981a44d8e296f ... /Users/…/daemon.sock   <- B, at the path

Now SIGTERM the displaced daemon:

kill $A; sleep 2
ls -i $S                                        # No such file or directory — A deleted B's node
kill -0 $B && echo "B alive, no socket"

The shutdown block guards one artifact and not the other (daemon.py, 0.2.32):

if sys.platform != "win32":
    try:
        Path(sock_path).unlink(missing_ok=True)   # unconditional
    except Exception:
        pass
try:
    stored = pid_path.read_text().strip()
    if stored == str(os.getpid()):                # guarded
        pid_path.unlink(missing_ok=True)

The PID file has the "am I still the owner?" test; the socket doesn't. Recording the socket's inode at bind time and unlinking only on a match would close it.

Separate from #243, which serializes client startup but doesn't touch this path.

Start in daemon.py at the shutdown block for ccc run-daemon and trace where the Unix socket is bound. Reproduce the two-daemon scenario from the issue, then verify that stopping a displaced daemon leaves the live daemon's socket available while its own cleanup still works.