fix(#667): CLI no longer hangs on the receipt socket; resolve captains by socket path

[tu11aa]

Two more defects from the slice 4 live smoke. Neither was reachable from unit tests.

1. squadrant ping delivered, then hung forever

The message arrived and the captain answered — and the CLI process never exited:

$ pgrep -fl 'index.js ping'
63345 node .../dist/index.js ping s4lab SLICE4-ALPHA...     # still resident, work long done

A listening server keeps Node's event loop alive. Any short-lived CLI that built a captain channel bound a receipt socket and could never terminate. Anything scripting squadrant ping would block indefinitely.

• server.unref() after listen — the process can now exit once its own awaits settle. Safe for the daemon, which stays alive on its own control socket rather than this one.
• stop() now also unlinks the socket file; close() does not, so every process that bound one was leaving a squadrantd-<pid>.sock behind in the shared directory.

2. Every captain ping reported accepted-unconfirmed

For a captain that was alive and answering:

captain-channel s4lab: accepted via claude-peer (unconfirmed — no turn observed)

The captain channel resolved the registry entry by cwd, and cwd is not canonical:

{"pid":44257,"cwd":"/private/tmp/s4lab","status":"idle",
 "messagingSocketPath":"/tmp/cc-socks/squadrant-captain-s4lab.sock"}

The project path is /tmp/s4lab; the session registers /private/tmp/s4lab because /tmp is a symlink on macOS. The raw string compare silently missed, statusFor returned undefined, and the confirmation branch never ran.

Added readClaudeStatusBySocketPath and switched the captain channel to it — the same identifier #689 adopted for crews: squadrant picks it at launch and the session echoes it back verbatim.

This removes the last cwd-based match from the delivery path. Slice 1 had already dropped cwd for crews because the operator's own Claude window shares the project root; non-canonical paths are a second, independent reason. Worth treating "never match a session by cwd" as settled.

Verification — slice 4 smoke now passes all four cases

Mode	Result
unset/off	unchanged, and the no-auto-boot refusal still fires
on	✔ Delivered to 's4lab' (accepted via claude-peer), returns in 8s, captain answered
shadow	shadow probe — reachable, message delivered exactly once
gone (captain killed)	session gone — falling back to pane → ⚠ not reachable, one fallback, both logged

6 new tests (unref, optional-unref, file cleanup, plus the earlier reject/unlink cases). pnpm build clean, 2660 tests green.

Related: #667 slice 4 (#685), #689, #691, #695.