mu-ad5x serve: abort the presence task on shutdown — mu serve hung after stdin close

[tcovert-c137]

Fixes mu-ad5x: since the etcd-lease presence merge (949fed77), mu serve never exits after stdin closes on any box where etcd is reachable — including the serve child under mu solo. Repro: just smoke completes the ask, logs dialogue presence: lease-registered, then hangs until the recipe's 5s guard kills it; one commit earlier it exits cleanly.

Root cause

presence::spawn starts an infinite reconcile loop (lease keepalive/re-grant + peer-key reconcile every ttl/3) holding a Sessions clone, and drops the JoinHandle. mu's shutdown is reference-driven — serve's exit waits on session channels closing — so the immortal clone wedges the documented shutdown cascade. serve/mod.rs already names this exact hazard and solves it for the MCP listener ("transport::serve's shutdown cascade … deadlocks if any external clone keeps those alive. AbortOnDrop is captured by the handler closure…"); the presence task was the one spawn without the guard.

Change

• presence::spawn returns its JoinHandle, marked #[must_use = "tie this handle to shutdown (AbortOnDrop) or serve never exits (mu-ad5x)"] so the omission can't silently recur.
• serve wraps it in AbortOnDrop captured by the transport closure, exactly like mcp_guard: dropping the handler aborts the task and releases its Sessions clone, letting the cascade complete.

Aborting is crash-equivalent and safe by the lease design: the etcd lease expires at TTL and the peer keys vanish with it — the same cleanup a daemon crash gets. No graceful revoke needed.

Verification

On this box (etcd reachable, presence active): just smoke now logs lease-registered, answers, and exits cleanly (exit 0) where it previously hung to the kill guard. Full just ci-aipr green.

Gate

pre-pr-check green; consensus panel PASS — unanimous approve in 1 round, all four (focused) seats live.

🤖 Generated with Claude Code

https://claude.ai/code/session_01REB3QZvP3arA2cLMVRgYam