mcpx: CLI never notices a running daemon from an older binary (no daemon-version check) #265

Problem

The mcpx CLI in this repo never notices when the daemon it connected to was started by a different (older) binary. After an upgrade, a daemon started by the old binary keeps running, the new CLI talks to it over the socket/url ladder rungs, and a command or field the old daemon lacks fails confusingly instead of saying why.

This is the same defect as dezren39/darwin-nix#344 (fixed there by dezren39/nix PR #386, merged 2026-10-18); this repo's pkgs/mcpx copy does not have that fix.

Evidence (verified on main 2d56eb2, 2026-10-18)

• The daemon already reports its version: pkgs/mcpx/internal/daemon/server.go handleHealth returns "version": s.version in the GET /v1/health answer.
• The CLI never reads that field:
  • No DaemonVersion() method in pkgs/mcpx/internal/cli/client.go — the neighbouring Endpoint() parses only endpoint from the same /v1/health answer.
  • No pkgs/mcpx/internal/cli/versioncheck.go exists (no daemonVersionNotice, daemonVersionWarning, or checkDaemonVersion anywhere in pkgs/mcpx/internal/cli/).
  • App.ensure() in pkgs/mcpx/internal/cli/commands.go:81 caches the client/ladder and returns without any version comparison.
• Therefore a version mismatch is silent on every command that reuses an already-running daemon via the socket or url rungs (Connect in pkgs/mcpx/internal/cli/connect.go).

Why it matters

Upgrades are the normal path (mcpx upgrade, package updates). Without the notice, the first symptom of a stale daemon is a confusing per-command failure, with no hint that mcpx restart --daemon is the fix.

Suggested fix (mirrors dezren39/nix PR #386)

1. Add Client.DaemonVersion(ctx) in client.go: GET /v1/health, parse version; an older daemon reporting none yields "" with no error (absence is not a mismatch).
2. Add versioncheck.go: daemonVersionNotice(cli, daemon, remote) — empty when equal or either side empty; otherwise one line naming both versions, with mcpx restart --daemon advice locally and "restart it where it runs" for a remote daemon. checkDaemonVersion only runs for socket/url ladder rungs (spawn/inline daemons run this binary by construction) and is best-effort (a failed health read stays silent).
3. Call a.checkDaemonVersion(ctx, c, ladder) in ensure() after a successful connect.
4. Port the regression tests from dezren39/nix (versioncheck_test.go): notice table cases, health-server warning cases (mismatch / matching / no-version / unreachable), and the spawned-daemon skip.

Acceptance criteria

• With a daemon reporting a different version reached via socket/url, a CLI command prints exactly one stderr notice naming both versions and the restart command.
• Equal versions, a daemon reporting no version, and spawn/inline daemons produce no notice and no extra health call failure mode.
• New versioncheck tests pass; go test ./internal/cli/ in pkgs/mcpx is green; ablation (removing the ensure() call) makes the wiring test fail.