Summary

Implements the three daemon-lifecycle follow-ups deferred from #179 (connect-retry grace for daemon restarts).

1. Version field in DaemonHandshake

• The handshake now carries client_version (the connecting binary's CARGO_PKG_VERSION) with #[serde(default)]: a new daemon parses old handshakes (missing field → empty, never treated as skew), and old daemons ignore the extra field, so mixed-version pairs keep interoperating.
• The daemon logs a daemon_version_skew event (with a tracedecay daemon restart hint) when a client's version differs, deduped per client version in DaemonEngine — proxy clients reconnect per request, so an undeduped log would flood.

2. Stale daemon on non-systemd paths

• Client-side detection that works against already-deployed old daemons: the serve stdio proxy reads result.serverInfo.version from proxied initialize responses (daemons of every version send it) and warns on stderr: "daemon is version X but this client is Y — run tracedecay daemon restart".
• New tracedecay daemon restart subcommand restarts the installed service (reusing refresh_installed_service); when no service is installed it errors with guidance to restart the manual tracedecay daemon run process.
• tracedecay post-update now distinguishes "no daemon at all" from "an unmanaged daemon is running" and warns loudly in the latter case instead of the quiet "skipping daemon restart" line.
• Design choice: warn/log rather than daemon self-exit on serving a newer client. A self-exiting daemon is only safe under a supervisor; for a manual tracedecay daemon run it would drop the request mid-flight and leave no daemon behind, which is worse than a stale one. Plain tracedecay upgrade (which skips post-update) is covered by the skew warnings rather than by changing upgrade's semantics.

3. Serve in-process fallback race

• tracedecay serve startup no longer commits to in-process mode just because the socket file is missing. New should_proxy_serve_to_daemon: if the socket exists → proxy (unchanged fast path); if it's missing but the installed service unit claims exactly this socket (the tracedecay update restart window, where shutdown unlinks the socket before the new daemon rebinds) → probe with the same connect_with_restart_grace used for per-request reconnects; otherwise → in-process immediately, so startup stays instant when no daemon was ever installed. A stopped-but-installed service falls back to in-process after the grace instead of hard-failing every request.

Test plan

• cargo check --all-targets, cargo clippy --all-targets, cargo fmt --check — clean
• TRACEDECAY_DATA_DIR=$(mktemp -d) cargo test --lib daemon:: — 47 passed, including new tests:
  • handshake interop: old client → new daemon (missing field parses as empty), new client → old daemon (unknown-field tolerance of the serde derive), version advertised on the wire
  • client_version_skew only flags real mismatches (empty/old-client version is not skew)
  • DaemonEngine logs each skewed client version once
  • daemon_version_skew_warning fires only for mismatched initialize responses and names both versions + the restart command
  • serve transport decision: proxies when socket exists; instant in-process fallback with no socket/service (or a service on a different socket); waits out the restart window when the service owns the socket; falls back after grace when nothing rebinds
• cargo test --test mcp_cli_serve_test — 14 passed, including new end-to-end regression test: serve started with no socket but an installed unit proxies to a sentinel daemon that binds 400ms later (distinguishable from in-process fallback) and surfaces the version-skew warning on stderr
• cargo test --test tool_daemon_test --test cli_help_test --test update_plugin_test — all passed