feat(agent): v0.3.5 Windows inbound forwarding, UI actions, lifecycle

- Configure tailscale serve automatically for each instance on Windows userspace networking.
- Add local UI buttons: start/stop/reset/delete instances (stop/start preserve volumes).
- Clean shutdown: stop tailscaled and instances, notify server with instance_stopped.
- Restart tailscaled on agent boot using persisted state when pre-auth key is absent.
- Sync instance stopped/deleted status to dashboard (server/lib/websocket.ts).
- Security: include prior authz/scoping changes across API routes, ephemeral pre-auth keys, ACL policy, internal API key.
- Update SUIVI_VPN_ONDEMAND.md and docs/ONBOARDING_CLIENT.md.
- Bump agent version to 0.3.5.

agent/tailscale.go

@@ -9,6 +9,7 @@ import (
 	"os/exec"
 	"path/filepath"
 	"runtime"
+	"strconv"
 	"sync"
 	"syscall"
 	"time"
@@ -61,6 +62,9 @@ func startTailscale(dataDir, nodeID, headscaleURL, authKey string) (string, erro
 	if err := os.MkdirAll(tsDataDir, 0700); err != nil {
 		return "", fmt.Errorf("create tailscale dir: %w", err)
 	}
+	// Make sure a previous tailscaled (e.g. left behind after a crash or
+	// force-kill) does not block the new daemon on the same socket/state.
+	killStaleTailscaled(tsDataDir)
 	if runtime.GOOS == "windows" {
 		// Windows uses named pipes for tailscaled IPC, not Unix sockets.
 		tsSocket = `\\.\pipe\studioe5-tailscaled`
@@ -88,6 +92,9 @@ func startTailscale(dataDir, nodeID, headscaleURL, authKey string) (string, erro
 		tsCmd = nil
 		return "", fmt.Errorf("start tailscaled: %w", err)
 	}
+	if err := os.WriteFile(filepath.Join(tsDataDir, "tailscaled.pid"), []byte(strconv.Itoa(tsCmd.Process.Pid)), 0644); err != nil {
+		log.Printf("Cannot write tailscaled pid file: %v", err)
+	}
 
 	// Give tailscaled a moment to start listening.
 	time.Sleep(1 * time.Second)
@@ -99,11 +106,14 @@ func startTailscale(dataDir, nodeID, headscaleURL, authKey string) (string, erro
 	upArgs := []string{
 		"--socket=" + tsSocket,
 		"up",
-		"--authkey=" + authKey,
 		"--login-server=" + headscaleURL,
 		"--hostname=" + nodeID,
 		"--accept-dns=false",
 	}
+	// The auth key is omitted on reconnect: Tailscale reuses the existing state.
+	if authKey != "" {
+		upArgs = append(upArgs, "--authkey="+authKey)
+	}
 	if runtime.GOOS == "windows" {
 		// On Windows, keep the VPN up even after the tailscale.exe CLI client disconnects.
 		upArgs = append(upArgs, "--unattended")
@@ -181,6 +191,9 @@ func stopTailscaleLocked() {
 	_ = tsCmd.Wait()
 	tsCmd = nil
 	tsIP = ""
+	if tsDataDir != "" {
+		_ = os.Remove(filepath.Join(tsDataDir, "tailscaled.pid"))
+	}
 	log.Printf("Tailscale stopped")
 }
 
@@ -200,4 +213,70 @@ func getTailscaleIP() string {
 	return tsIP
 }
 
+// setupTailscaleServe configures Tailscale to proxy inbound Tailnet traffic
+// on the given TCP port to localhost:<port>. This is required on Windows
+// because userspace networking does not forward incoming connections to
+// loopback by default.
+func setupTailscaleServe(port int) error {
+	tsCmdMu.Lock()
+	defer tsCmdMu.Unlock()
+	if tsSocket == "" {
+		return fmt.Errorf("tailscale socket not initialized")
+	}
+
+	portStr := strconv.Itoa(port)
+	// Clean up any stale config for this port first.
+	offCmd := exec.Command(tailscaleBin("tailscale"), "--socket="+tsSocket, "serve", "--bg", "--tcp="+portStr, "off")
+	hideWindow(offCmd)
+	_ = offCmd.Run()
+
+	serveCmd := exec.Command(tailscaleBin("tailscale"), "--socket="+tsSocket, "serve", "--bg", "--tcp="+portStr, "tcp://localhost:"+portStr)
+	hideWindow(serveCmd)
+	out, err := serveCmd.CombinedOutput()
+	if err != nil {
+		return fmt.Errorf("tailscale serve: %w: %s", err, string(out))
+	}
+	log.Printf("Tailscale serve configured for port %s", portStr)
+	return nil
+}
+
+// removeTailscaleServe removes the Tailscale serve proxy for a port when an
+// instance is stopped or deleted.
+func removeTailscaleServe(port int) {
+	tsCmdMu.Lock()
+	defer tsCmdMu.Unlock()
+	if tsSocket == "" {
+		return
+	}
+	portStr := strconv.Itoa(port)
+	offCmd := exec.Command(tailscaleBin("tailscale"), "--socket="+tsSocket, "serve", "--bg", "--tcp="+portStr, "off")
+	hideWindow(offCmd)
+	_ = offCmd.Run()
+	log.Printf("Tailscale serve removed for port %s", portStr)
+}
+
+// killStaleTailscaled terminates a previously started tailscaled process that
+// may have been left running after the agent was force-killed.
+func killStaleTailscaled(tsDataDir string) {
+	pidFile := filepath.Join(tsDataDir, "tailscaled.pid")
+	data, err := os.ReadFile(pidFile)
+	if err != nil {
+		return
+	}
+	var pid int
+	if _, err := fmt.Sscanf(string(data), "%d", &pid); err != nil {
+		return
+	}
+	proc, err := os.FindProcess(pid)
+	if err != nil {
+		return
+	}
+	if err := proc.Signal(syscall.Signal(0)); err == nil {
+		log.Printf("Killing stale tailscaled process %d", pid)
+		_ = proc.Kill()
+		_, _ = proc.Wait()
+	}
+	_ = os.Remove(pidFile)
+}
+