fix: remove EnsureSSHServer call and fix authorized_keys permissions

- Remove EnsureSSHServer() call in manager.go that was overriding setupSSHInContainer() config
- Fix authorized_keys ownership with chown root:root
- Attempt to work around ld.so.preload GPU library logging issue
- Remove unused klog import from manager.go

The old EnsureSSHServer() was being called after setupSSHInContainer(),
overriding the secure SSH configuration (PasswordAuthentication no,
PermitRootLogin prohibit-password) with insecure defaults. This caused
containers to still have password authentication enabled despite
setupSSHInContainer() setting it to disabled.

Known issue: TensorFusion GPU libraries loaded via /etc/ld.so.preload
interfere with SSH protocol. The preload file is read-only mounted and
cannot be modified at runtime.  This requires a fix in the TensorFusion
GPU library to suppress logging during SSH sessions.

Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>

Author: andystime

internal/studio/manager.go

@@ -17,7 +17,6 @@ import (
 
 	"github.com/NexusGPU/gpu-go/internal/errors"
 	"github.com/NexusGPU/gpu-go/internal/platform"
-	"k8s.io/klog/v2"
 )
 
 var (
@@ -223,12 +222,15 @@ func (m *Manager) Create(ctx context.Context, opts *CreateOptions) (*Environment
 		return nil, err
 	}
 
-	// Setup SSH server if backend supports it
-	if sshBackend, ok := backend.(SSHServerBackend); ok {
-		if err := sshBackend.EnsureSSHServer(ctx, env.ID); err != nil {
-			klog.Warningf("Failed to setup SSH server: %v (continuing anyway)", err)
-		}
-	}
+	// SSH server is now set up by setupSSHInContainer() during container creation
+	// No need to call EnsureSSHServer which would override the secure configuration
+	//
+	// // Setup SSH server if backend supports it
+	// if sshBackend, ok := backend.(SSHServerBackend); ok {
+	// 	if err := sshBackend.EnsureSSHServer(ctx, env.ID); err != nil {
+	// 		klog.Warningf("Failed to setup SSH server: %v (continuing anyway)", err)
+	// 	}
+	// }
 
 	m.clearUnreachableSSH(ctx, env)
 

internal/studio/ssh_setup.go

@@ -111,15 +111,31 @@ echo "SSH setup completed successfully"
 	if sshPublicKey != "" {
 		klog.V(2).Infof("Adding SSH public key to container")
 		addKeyCmd := execCmd("exec", containerID, "sh", "-c",
-			fmt.Sprintf("echo '%s' >> /root/.ssh/authorized_keys && chmod 600 /root/.ssh/authorized_keys", sshPublicKey))
+			fmt.Sprintf("echo '%s' >> /root/.ssh/authorized_keys && chmod 600 /root/.ssh/authorized_keys && chown root:root /root/.ssh/authorized_keys", sshPublicKey))
 		if output, err := addKeyCmd.CombinedOutput(); err != nil {
 			klog.Warningf("Failed to add SSH key (non-fatal): %v, output: %s", err, string(output))
 		}
 	}
 
 	// Start SSH daemon in background
+	// Temporarily disable ld.so.preload to prevent GPU library logs from breaking SSH protocol
 	klog.V(2).Infof("Starting SSH daemon in container")
-	startSSHCmd := execCmd("exec", "-d", containerID, "/usr/sbin/sshd", "-D")
+	startSSHScript := `
+# Temporarily disable ld.so.preload to prevent library logs from breaking SSH
+if [ -f /etc/ld.so.preload ]; then
+    mv /etc/ld.so.preload /etc/ld.so.preload.disabled
+fi
+
+# Start SSH daemon
+/usr/sbin/sshd -D &
+
+# Re-enable ld.so.preload after sshd has forked
+sleep 1
+if [ -f /etc/ld.so.preload.disabled ]; then
+    mv /etc/ld.so.preload.disabled /etc/ld.so.preload
+fi
+`
+	startSSHCmd := execCmd("exec", "-d", containerID, "sh", "-c", startSSHScript)
 	if output, err := startSSHCmd.CombinedOutput(); err != nil {
 		klog.Errorf("Failed to start SSH daemon: %v, output: %s", err, string(output))
 		return fmt.Errorf("failed to start SSH daemon: %w\nOutput: %s", err, string(output))