unikernel: bake volume config into images + per-service user-mode net

Enables migrating real services (config/cert dirs, stateless) to
unikernels. Volume host dirs are copied into a per-VM staging tree
mirroring guest paths; the ops config goes in the staging root with the
top-level dirs in Dirs, so ops bakes them at the right absolute paths.
(Staging is required — an absolute /srv MapDirs source makes ops descend
into the agent's podman overlay storage and fail.) A component may set
network = "user" to use QEMU user-mode NAT instead of the isolated
bridge (Phase-1 networking for first migrations, before a gateway proxy).

Verified: mcat (the MCIAS policy tester) deployed as a Nanos unikernel
via 'mcp deploy', booting with its baked /srv/mcat config+certs, serving
HTTPS verified against the platform CA, configured against MCIAS.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

internal/runtime/qemu.go

@@ -204,6 +204,7 @@ func (q *QEMU) Pull(ctx context.Context, image string) error {
 type opsConfig struct {
 	Args      []string          `json:"Args,omitempty"`
 	Env       map[string]string `json:"Env,omitempty"`
+	Dirs      []string          `json:"Dirs,omitempty"` // dirs (relative to the config file) baked into the image
 	RunConfig opsRunConfig      `json:"RunConfig"`
 }
 
@@ -263,9 +264,15 @@ func (q *QEMU) Run(ctx context.Context, spec ContainerSpec) error {
 		cpus = 1
 	}
 
+	// A component can opt out of the isolated bridge with network = "user"
+	// (e.g. a stateless service that needs NAT egress to its dependencies
+	// before the gateway proxy exists). Otherwise unikernels use the bridge
+	// when one is configured.
+	useBridge := q.bridgeMode() && spec.Network != "user"
+
 	// In bridge mode, allocate a static IP on the isolated bridge for this VM.
 	var vmIP string
-	if q.bridgeMode() {
+	if useBridge {
 		ip, allocErr := q.allocateIP(spec.Name)
 		if allocErr != nil {
 			return fmt.Errorf("allocate VM IP: %w", allocErr)
@@ -282,11 +289,52 @@ func (q *QEMU) Run(ctx context.Context, spec ContainerSpec) error {
 			CPUs:   cpus,
 		},
 	}
-	if q.bridgeMode() {
+	if useBridge {
 		cfg.RunConfig.IPAddress = vmIP
 		cfg.RunConfig.NetMask = "255.255.255.0"
 		cfg.RunConfig.Gateway = q.Gateway
 	}
+	// Bake each volume's host directory into the image (read-only). Unikernels
+	// have no writable host mount yet, so this suits config/cert directories;
+	// stateful services need a different storage story (9p/virtio-blk).
+	//
+	// Volume contents are copied into a per-VM staging tree that mirrors the
+	// guest paths (e.g. /srv/mcat -> <bake>/srv/mcat). We then put the ops
+	// config in the staging root and list the top-level dirs in `Dirs`: `ops`
+	// resolves Dirs relative to the config file and bakes the tree at the
+	// matching absolute path. (Staging is required: an absolute /srv source
+	// would make `ops` descend into the agent's podman overlay storage and
+	// fail.)
+	bakeDir := ""
+	if len(spec.Volumes) > 0 {
+		bakeDir = filepath.Join(q.vmDir(spec.Name), "bake")
+		_ = os.RemoveAll(bakeDir)
+		topLevel := map[string]bool{}
+		for _, v := range spec.Volumes {
+			parts := strings.SplitN(v, ":", 2)
+			host := parts[0]
+			guest := host
+			if len(parts) == 2 {
+				guest = parts[1]
+			}
+			rel := strings.TrimPrefix(guest, "/")
+			stage := filepath.Join(bakeDir, rel)
+			if err := os.MkdirAll(filepath.Dir(stage), 0o750); err != nil {
+				return fmt.Errorf("stage volume dir: %w", err)
+			}
+			if err := os.CopyFS(stage, os.DirFS(host)); err != nil {
+				return fmt.Errorf("stage volume %q: %w", host, err)
+			}
+			if i := strings.Index(rel, "/"); i >= 0 {
+				topLevel[rel[:i]] = true
+			} else {
+				topLevel[rel] = true
+			}
+		}
+		for d := range topLevel {
+			cfg.Dirs = append(cfg.Dirs, d)
+		}
+	}
 	if len(spec.Env) > 0 {
 		cfg.Env = map[string]string{}
 		for _, e := range spec.Env {
@@ -295,7 +343,13 @@ func (q *QEMU) Run(ctx context.Context, spec ContainerSpec) error {
 			}
 		}
 	}
+	// Place the ops config in the staging root when baking volumes, so its
+	// relative Dirs resolve to the staged tree; otherwise keep it in the VM
+	// state dir.
 	cfgPath := filepath.Join(q.vmDir(spec.Name), "ops.json")
+	if bakeDir != "" {
+		cfgPath = filepath.Join(bakeDir, "ops.json")
+	}
 	cfgBytes, err := json.MarshalIndent(cfg, "", "  ")
 	if err != nil {
 		return fmt.Errorf("marshal ops config: %w", err)
@@ -311,6 +365,9 @@ func (q *QEMU) Run(ctx context.Context, spec ContainerSpec) error {
 	}
 	build := exec.CommandContext(ctx, q.opsPath(), "build", bin, "-c", cfgPath, "-i", spec.Name) //nolint:gosec
 	build.Env = q.opsEnv()
+	if bakeDir != "" {
+		build.Dir = bakeDir
+	}
 	if out, err := build.CombinedOutput(); err != nil {
 		return fmt.Errorf("ops build %q: %w: %s", spec.Name, err, out)
 	}
@@ -339,7 +396,7 @@ func (q *QEMU) Run(ctx context.Context, spec ContainerSpec) error {
 	//     reaching only mc-proxy on the gateway. This makes mediation
 	//     mandatory rather than cooperative.
 	var netDevice, netBackend string
-	if q.bridgeMode() {
+	if useBridge {
 		tap, tapErr := q.createTAP(spec.Name)
 		if tapErr != nil {
 			return fmt.Errorf("create TAP: %w", tapErr)
@@ -368,17 +425,6 @@ func (q *QEMU) Run(ctx context.Context, spec ContainerSpec) error {
 		"-device", netDevice,
 		"-netdev", netBackend,
 	}
-	// 9p passthrough for host /srv/<service> volumes (best-effort; Nanos must
-	// support the 9p client for the guest to mount it).
-	for i, v := range spec.Volumes {
-		parts := strings.SplitN(v, ":", 2)
-		host := parts[0]
-		tag := fmt.Sprintf("srv%d", i)
-		args = append(args,
-			"-virtfs", fmt.Sprintf("local,path=%s,mount_tag=%s,security_model=none,id=%s", host, tag, tag),
-		)
-	}
-
 	cmd := exec.CommandContext(ctx, q.qemuPath(), args...) //nolint:gosec
 	if out, err := cmd.CombinedOutput(); err != nil {
 		return fmt.Errorf("qemu launch %q: %w: %s", spec.Name, err, out)