Skip to main content

boatramp_node/
node.rs

1//! The node-graph assembly: given a built store (blobs + KV), a configured
2//! [`Auth`](boatramp_server::Auth), and resolved
3//! [`ServerOptions`](boatramp_server::ServerOptions), wire the deploy store,
4//! handler runtime, compute reconcile loop, and domain-verify reconcile loop
5//! into a [`RunningNode`] ready to hand to a transport (`serve_with` & friends).
6//!
7//! This is the headline extraction of `PLAN-node-library`: the binary's
8//! `serve::run` used to inline this wiring, so no embedder or in-process test
9//! could exercise the same graph the `boatramp serve` binary runs. `run` now
10//! resolves the *environment* (args -> backends -> store, signal handlers,
11//! migration, auth) and calls [`assemble`]; the cluster path keeps its own inline
12//! copy until a later step converges it here.
13
14use std::path::Path;
15use std::sync::Arc;
16
17use boatramp_core::Storage;
18use boatramp_core::deploy::DeployStore;
19use boatramp_core::kv::KvStore;
20
21use crate::config::ServerConfig;
22use crate::error::{Error, Result};
23
24/// How often the compute reconcile loop converges desired vs actual workloads.
25/// Defaults to 30s; override with `BOATRAMP_COMPUTE_RECONCILE_TICK_MS` (milliseconds)
26/// so compute-backed tests can converge in a fraction of a second instead of
27/// waiting a full tick for the launch/scale reconcile.
28pub fn compute_reconcile_tick() -> std::time::Duration {
29    std::env::var("BOATRAMP_COMPUTE_RECONCILE_TICK_MS")
30        .ok()
31        .and_then(|s| s.parse::<u64>().ok())
32        .filter(|&ms| ms > 0)
33        .map(std::time::Duration::from_millis)
34        .unwrap_or(std::time::Duration::from_secs(30))
35}
36/// How often the domain-verify reconcile loop re-checks pending challenges.
37pub const DOMAIN_VERIFY_RECONCILE_TICK: std::time::Duration = std::time::Duration::from_secs(60);
38/// How long a compute workload may be idle before scale-to-zero sleeps it.
39pub const COMPUTE_IDLE_TIMEOUT: std::time::Duration = std::time::Duration::from_secs(300);
40
41/// The built store + resolved config handed to [`assemble`]. Owns the blob/KV
42/// backends and the auth/options the caller already resolved; borrows the parsed
43/// config and data directory.
44pub struct NodeInput<'a> {
45    /// The full parsed server config (the handler + compute sections are read here).
46    pub config: &'a ServerConfig,
47    /// The node data directory (per-site SQL, handler state).
48    pub data_dir: &'a Path,
49    /// The object store built by [`crate::blobs::build_blobs`].
50    pub storage: Arc<dyn Storage>,
51    /// The metadata KV, already cache-fronted, built by [`crate::backends::build_kv`].
52    pub kv: Arc<dyn KvStore>,
53    /// The control-plane auth built by [`crate::auth::configure_auth`].
54    pub auth: boatramp_server::Auth,
55    /// Server options, already carrying the resolved posture, daemon runtime, and
56    /// (post-`configure_auth`/`configure_oidc`) issuer / OIDC verifier.
57    pub options: boatramp_server::ServerOptions,
58    /// The public HTTP serve bind address, if known — used (under
59    /// `allow_guest_self_egress`) to let a handler guest's `wasi:http` reach this
60    /// instance's own front door over loopback. `None` (an in-process embedder with no
61    /// listener) disables self-egress.
62    pub serve_addr: Option<std::net::SocketAddr>,
63    /// The cloud blob-change watch provider (FA-5b2), if the backend is a cloud one.
64    pub watch_provider: Option<Arc<dyn boatramp_core::blob_provision::WatchProvider>>,
65    /// The provisioning tier for the watch provider.
66    pub provision_tier: boatramp_core::blob_notify::ProvisionTier,
67    /// The `wasi:messaging` substrate override for the handler runtime. `None` uses
68    /// the single-node default (`LogMessaging` over the same backends); the cluster
69    /// path passes its Raft-backed coordinator.
70    pub messaging: Option<Arc<dyn boatramp_core::messaging::Messaging>>,
71    /// The single leader gate for cron firing + the compute / domain-verify reconcile
72    /// loops. Single-node passes an always-true gate (there is one node); the cluster
73    /// passes its Raft `is_leader` check so a single node drives each sweep.
74    pub is_leader: boatramp_server::CronLeaderGate,
75    /// This node's compute scheduler id (`0` single-node; the cluster node id in a
76    /// fleet, so replicas are tagged to the right node).
77    pub node_id: u64,
78    /// The binary the re-exec'd compute workers run as — the container backend's
79    /// `__sandbox` jailer and the microVM backends' `__vmm-run`/`__vz-run` VM hosts.
80    /// `None` uses this process's own executable (`current_exe`), which is what
81    /// `boatramp serve` wants (the child *is* boatramp). An **embedding harness**
82    /// whose own binary doesn't implement those subcommands should point this at a
83    /// built `boatramp` binary, so it can drive the real container/microVM backends
84    /// in-process (only the per-workload worker re-execs; the serving plane stays
85    /// embedded). The docker backend needs neither — it talks to a daemon.
86    pub worker_exe: Option<std::path::PathBuf>,
87}
88
89/// A fully wired node: the deploy store, handler runtime, auth, and options a
90/// transport consumes, plus the detached reconcile loops kept alive for the
91/// node's serving life. Destructure it and hold `reconcile` across the serve
92/// await so the loops outlive assembly.
93pub struct RunningNode {
94    /// The deploy store (blob + KV) the router serves from.
95    pub deploy: DeployStore,
96    /// The handler runtime for wasm handlers (a disabled build ⇒ a no-op runtime).
97    pub handlers: boatramp_server::HandlerRuntime,
98    /// The control-plane auth.
99    pub auth: boatramp_server::Auth,
100    /// The resolved server options.
101    pub options: boatramp_server::ServerOptions,
102    /// The detached reconcile loops (compute + domain-verify). Tokio `JoinHandle`s
103    /// do not abort on drop, so the loops run for the process life regardless; the
104    /// handles are retained so an embedder can join/abort them on shutdown.
105    pub reconcile: Vec<tokio::task::JoinHandle<()>>,
106}
107
108/// The instance's own serve socket(s) a guest self-call may reach, given the bind `addr` and
109/// whether the posture (`allow_guest_self_egress`) permits it. A wildcard bind
110/// (`0.0.0.0`/`::`) is reachable over loopback, so it normalizes to `127.0.0.1` **and** `::1`
111/// on the serve port; a specific bind is reachable at itself. Empty when disabled or no
112/// listener.
113fn self_egress_addrs(
114    addr: Option<std::net::SocketAddr>,
115    enabled: bool,
116) -> Vec<std::net::SocketAddr> {
117    use std::net::{IpAddr, Ipv4Addr, Ipv6Addr, SocketAddr};
118    let Some(addr) = addr.filter(|_| enabled) else {
119        return Vec::new();
120    };
121    if addr.ip().is_unspecified() {
122        let port = addr.port();
123        vec![
124            SocketAddr::new(IpAddr::V4(Ipv4Addr::LOCALHOST), port),
125            SocketAddr::new(IpAddr::V6(Ipv6Addr::LOCALHOST), port),
126        ]
127    } else {
128        vec![addr]
129    }
130}
131
132/// Whether a node with **no configured control-plane issuer** should auto-provision an ephemeral
133/// in-memory fleet signer (for host session cookies + delegable capabilities) from its serve bind.
134///
135/// The fleet signer is a distinct trust domain from control-plane admin auth, so a DEV / loopback
136/// node with auth disabled should still be able to sign/verify session cookies and capabilities. It
137/// is gated **strictly to a loopback bind** (`127.0.0.1`/`::1`) or an in-process embedder with no
138/// bind address (`None`): NEVER a public or wildcard (`0.0.0.0`/`::`, reachable off-host) bind,
139/// where an ephemeral key would silently invalidate live capabilities across a restart — a
140/// public node that wants guest capabilities without control-plane auth must supply a persistent
141/// key. (`is_loopback()` is already `false` for a wildcard/unspecified address, so a `0.0.0.0` bind
142/// correctly does NOT auto-provision.)
143#[cfg(feature = "handlers")]
144fn should_autoprovision_fleet_signer(serve_addr: Option<std::net::SocketAddr>) -> bool {
145    serve_addr.is_none_or(|a| a.ip().is_loopback())
146}
147
148/// Wire [`NodeInput`] into a [`RunningNode`]: build the handler runtime, the
149/// deploy store (materializing the reserved `default` project), the compute
150/// backends + reconcile loop, and the domain-verify reconcile loop.
151///
152/// The caller has already built the store and configured auth/OIDC on `options`;
153/// this is the pure node-graph wiring, identical to what `boatramp serve` runs.
154pub async fn assemble(input: NodeInput<'_>) -> Result<RunningNode> {
155    let NodeInput {
156        config,
157        data_dir,
158        storage,
159        kv,
160        auth,
161        options,
162        serve_addr,
163        watch_provider,
164        provision_tier,
165        messaging,
166        is_leader,
167        node_id,
168        worker_exe,
169    } = input;
170    // Copy out the posture scalars up front so `options` can be moved into the
171    // returned `RunningNode` without a lingering borrow.
172    let max_handler_blob_bytes = options.posture.max_handler_blob_bytes;
173    let max_component_bytes = options.posture.max_component_bytes;
174    let allow_guest_private_egress = options.posture.allow_guest_private_egress;
175    let allow_env_secret_refs = options.posture.allow_env_secret_refs;
176    let allow_guest_email = options.posture.allow_guest_email;
177    // The instance's own serve socket(s) a guest self-call may reach, when the posture allows
178    // it: a wildcard bind (`0.0.0.0`/`::`) is reachable on loopback, so normalize to
179    // `127.0.0.1`/`::1`; a specific bind is itself.
180    let self_egress_addrs = self_egress_addrs(serve_addr, options.posture.allow_guest_self_egress);
181    let allow_shared_kernel = options.posture.allow_shared_kernel_compute;
182    let domain_verify_allow_private = options.posture.domain_verify_allow_private;
183
184    // The deploy store the router serves from — built up front so the handler
185    // runtime's managed compute-backed `sql` binding can resolve DB endpoints from
186    // the same store the reconcile writes.
187    let compute_storage = storage.clone();
188    let deploy = DeployStore::new(storage, kv.clone());
189    // **Crown-jewel frontier-sync (v0.9.0 KV-recovery, C2).** The sealed-secret / sealed-credential
190    // stores below hold IRREPLACEABLE data (a dropped write = a lost secret with no client copy), so
191    // rather than rely on per-call-site discipline we route them ALL through a `CheckpointKv`: every
192    // write advances the durable frontier before it acks, so a crash's self-heal-on-open trailing-
193    // tail quarantine can never drop one. Wrapping ONCE here is the can't-miss guarantee. (On the
194    // cluster `RaftKv` / `MemoryKv` the added checkpoint is a no-op, so this is inert there.) The
195    // mixed `DeployStore` above keeps the RAW `kv` and selects the frontier-sync per method instead
196    // (its `current/*` pointers, history, invocation records and metering stay async).
197    let secrets_kv: Arc<dyn KvStore> = Arc::new(boatramp_core::kv::CheckpointKv::new(kv.clone()));
198    // The `[secrets]` envelope (local KEK / Vault) that seals a managed SQL
199    // credential at rest. `None` ⇒ no wrapping (a managed DB then fails closed).
200    let secrets_envelope = build_secrets_envelope(config.secrets.as_ref(), data_dir)?;
201    // The project-scoped internal secret store, built from the same KV + `[secrets]`
202    // envelope that seal managed-DB credentials. Backs both the `boatramp:<name>`
203    // resolver (wired into the handler runtime below, when that feature is present)
204    // and the admin secrets API (threaded into `ServerOptions` unconditionally, so it
205    // works on a lean node too). `None` when no envelope is configured — the admin
206    // endpoints then fail closed with a clear 501, never a panic.
207    let secret_store = secrets_envelope.clone().map(|envelope| {
208        Arc::new(boatramp_core::secret_store::SecretStore::new(
209            secrets_kv.clone(),
210            envelope,
211        ))
212    });
213    // The per-TENANT sealed secret store (task #493), sealed with the SAME `[secrets]` envelope.
214    // Backs both the control-plane CRUD (threaded into `ServerOptions` below) and — when the
215    // `tenant-secrets` feature is compiled — the runtime guest binding (the SAME `Arc` handed to
216    // `set_tenant_secret_store`). `None` when no envelope is configured, so the endpoints fail
217    // closed with a clear 501 and the guest binding is not built.
218    let tenant_secret_store = secrets_envelope.clone().map(|envelope| {
219        Arc::new(boatramp_core::secret_store::TenantSecretStore::new(
220            secrets_kv.clone(),
221            envelope,
222        ))
223    });
224    // The project-scoped SMTP email-profile store, built from the same KV + envelope
225    // (the password is sealed at rest). Backs the admin API (`options` below,
226    // unconditionally, so it works on a lean node) and — when the `email` feature +
227    // `allow_guest_email` posture permit — the runtime's host-side profile
228    // resolution (wired inside `build_handler_runtime`). `None` with no envelope, so
229    // the admin email endpoints fail closed with a clear 501.
230    let email_profile_store = secrets_envelope.clone().map(|envelope| {
231        Arc::new(boatramp_core::email_config::EmailProfileStore::new(
232            secrets_kv.clone(),
233            envelope,
234        ))
235    });
236
237    // Dev-posture guest-egress extra CA(s): when the posture permits (off/refused under
238    // multi-tenant) AND the operator pointed `BOATRAMP_GUEST_EGRESS_EXTRA_CA_FILE` at a PEM, parse
239    // it into trust anchors the guest's outbound `wasi:http` TLS client trusts on TOP of the webpki
240    // roots (for a hermetic HTTPS test double). Empty otherwise. A configured-but-unreadable/
241    // unparsable file is a hard config error (fail closed), never a silent no-trust.
242    let guest_egress_extra_roots =
243        load_guest_egress_extra_roots(options.posture.allow_guest_egress_extra_ca)?;
244
245    // The handler runtime reuses the same blob/KV backends (per-site prefixed)
246    // for its wasi:blobstore/keyvalue bindings; the sql binding is selected by
247    // `[handlers.bindings.sql]` (default: per-site libsql files under <data-dir>).
248    let handlers = crate::handlers::build_handler_runtime(
249        kv.clone(),
250        compute_storage.clone(),
251        data_dir,
252        config.handlers.as_ref(),
253        messaging,
254        max_handler_blob_bytes,
255        max_component_bytes,
256        allow_guest_private_egress,
257        self_egress_addrs,
258        guest_egress_extra_roots,
259        allow_env_secret_refs,
260        allow_guest_email,
261        options.posture.require_tenancy_declaration,
262        options.posture.allow_cross_tenant_db,
263        &deploy,
264        secrets_envelope.clone(),
265    )
266    .await?;
267    // Hand the runtime the SAME per-tenant secret store `Arc` the control-plane routes hold (task
268    // #493), so a guest `tenant-secrets` `get` and a control-plane `PUT` seal/unseal against ONE
269    // store. Unset when no `[secrets]` envelope, so the guest binding is not built (fail-closed).
270    // Handlers-gated: `set_tenant_secret_store` lives on the handler runtime, so a lean (no-handlers)
271    // node has no guest binding to wire (the control-plane routes still work via `ServerOptions`).
272    #[cfg(feature = "handlers")]
273    if let Some(store) = tenant_secret_store.clone() {
274        handlers.set_tenant_secret_store(store);
275    }
276    // Wire the fleet session-cookie signer (R3, PLAN-tenancy-principal): the same issuer that mints
277    // control-plane tokens signs + verifies the host-issued anonymous session cookie AND the
278    // delegable capabilities (PLAN-delegable-capabilities). Handlers-gated: the session-cookie
279    // machinery lives on the handler runtime, so a lean (no-handlers) build has nothing to wire.
280    //
281    // The fleet signer is a DIFFERENT trust domain from control-plane admin auth (signing a
282    // customer's session cookie / an embed capability is not the authority to admit an operator to
283    // the control plane), but production derives it from the control-plane issuer for convenience.
284    // For a DEV / loopback node with control-plane auth disabled (`options.issuer` is `None`),
285    // auto-provision an EPHEMERAL in-memory Ed25519 fleet key so the guest-facing signer just works
286    // — session cookies + capability mint/verify — WITHOUT turning on control-plane auth. Strictly
287    // gated to a loopback bind (or an in-process embedder with no bind address): never on a public
288    // bind, where an ephemeral key would silently invalidate live capabilities across a restart (a
289    // public node that wants guest capabilities without control-plane auth must supply a persistent
290    // key). Ephemeral = issue + verify within one process run; nothing persisted, no cross-process
291    // or cross-deploy trust. Production is byte-identical: a real deploy supplies a control-plane key
292    // ⇒ `issuer` is `Some` ⇒ this fallback is never taken.
293    #[cfg(feature = "handlers")]
294    {
295        let fleet_signer = options.issuer.clone().or_else(|| {
296            should_autoprovision_fleet_signer(serve_addr).then(|| {
297                tracing::warn!(
298                    "control-plane auth is disabled and no signer is configured; auto-provisioning \
299                     an EPHEMERAL in-memory fleet signer (Ed25519) for host session cookies + \
300                     delegable capabilities on this loopback/dev node — regenerated each start, \
301                     never persisted. Configure a control-plane key (or a dedicated signer) for \
302                     production."
303                );
304                Arc::new(boatramp_core::cose::LocalSigner::generate(
305                    boatramp_core::cose::TokenAlg::Ed25519,
306                )) as Arc<dyn boatramp_core::cose::Signer>
307            })
308        });
309        if let Some(issuer) = fleet_signer {
310            handlers.set_session_signer(issuer);
311        }
312    }
313    // Enable guest capability minting (`boatramp:handlers/capability`, PLAN-delegable-capabilities)
314    // when the operator posture allows it. A minted capability is verified against the same fleet
315    // signer as the session cookie (wired just above), so this only enables the mint path + the TTL
316    // ceiling; posture-off (or a zero ceiling) ⇒ not offered (a guest `mint` is access-denied).
317    #[cfg(feature = "capability")]
318    if options.posture.allow_guest_mint_capability {
319        handlers.set_capability_minting(options.posture.max_guest_capability_ttl_secs);
320    }
321    // Per-project tenancy/capability posture overrides (Gap 4a): resolve each
322    // `[security.projects.<p>]` override against the fleet base so one serve process can run a
323    // strict-isolation project beside a looser one on a shared, multi-project machine. Empty ⇒
324    // every project uses the node base wired just above. Only these four in-project knobs are
325    // per-project; cross-project isolation stays structural (project = database).
326    #[cfg(feature = "handlers")]
327    {
328        let base = &options.posture;
329        let overrides: std::collections::BTreeMap<
330            String,
331            boatramp_core::security::ResolvedProjectTenancy,
332        > = config
333            .security
334            .as_ref()
335            .map(|s| {
336                s.projects
337                    .iter()
338                    .map(|(project, ovr)| (project.clone(), base.project_tenancy(ovr)))
339                    .collect()
340            })
341            .unwrap_or_default();
342        handlers.set_project_tenancy_overrides(overrides);
343    }
344    // Wire the guest project self-config capability (`boatramp:handlers/admin`) when the
345    // operator posture enables at least one surface. The controller reuses the same in-process
346    // domain-verify / email-profile / secret / site-config subsystems + the real domain probe;
347    // it's project-scoped per grant and rate-limited + audited. Posture-off ⇒ not offered.
348    #[cfg(feature = "admin")]
349    {
350        use boatramp_handlers::AdminSurface;
351        let p = &options.posture;
352        let mut surfaces = std::collections::BTreeSet::new();
353        if p.allow_guest_admin_domains {
354            surfaces.insert(AdminSurface::Domains);
355        }
356        if p.allow_guest_admin_email {
357            surfaces.insert(AdminSurface::Email);
358        }
359        if p.allow_guest_admin_site {
360            surfaces.insert(AdminSurface::Site);
361        }
362        if p.allow_guest_admin_secrets {
363            surfaces.insert(AdminSurface::Secrets);
364        }
365        if !surfaces.is_empty() {
366            let controller = Arc::new(boatramp_server::ServerAdminController::with_server_probe(
367                deploy.clone(),
368                email_profile_store.clone(),
369                secret_store.clone(),
370                p.domain_verify_allow_private,
371            ));
372            handlers.set_admin(controller, surfaces);
373        }
374    }
375    // Leader-gate cron firing (cluster: only the Raft leader fires; single-node: an
376    // always-true gate, equivalent to the unset default). The same gate drives the
377    // reconcile loops below, so all three converge on one leader per fleet. Only the
378    // handler runtime has a scheduler, so this is a no-op without the `handlers` feature.
379    #[cfg(feature = "handlers")]
380    handlers.set_cron_leader_gate(is_leader.clone());
381    // FA-5b2: on a cloud backend, wire the blob-change notification provisioner +
382    // its tier so adding a `blob` trigger provisions (and removing it retracts).
383    #[cfg(feature = "handlers")]
384    if let Some(provider) = watch_provider {
385        handlers.set_watch_provider(provider);
386        handlers.set_provision_tier(provision_tier);
387    }
388    #[cfg(not(feature = "handlers"))]
389    let _ = (watch_provider, provision_tier);
390
391    // Materialize the reserved `default` project so `project ls` / `project show
392    // default` reflect it on a fresh install, not only after a migration. Best
393    // effort: the reader backstop keeps listings correct even if this write can't
394    // land, so a transient failure must never block serving.
395    match deploy.ensure_default_project().await {
396        Ok(true) => tracing::info!("materialized the reserved `default` project record"),
397        Ok(false) => {}
398        Err(e) => tracing::warn!(
399            error = %e,
400            "could not materialize the `default` project record; readers use the synthesized default"
401        ),
402    }
403    // Wire the function-to-function invoke resolver now the deploy store exists,
404    // so a function granted `invoke` can call a sibling in-process (FI).
405    #[cfg(feature = "handlers")]
406    handlers.set_invoker(deploy.clone());
407
408    // Compute reconcile loop. Single-node is always the "leader". Backends are
409    // built from the `[compute]` config + capability detection; a no-op when none
410    // are registered. Detached for the server's life.
411    let (compute_backends, compute_node) = crate::compute::build_compute(
412        config.compute.as_ref(),
413        compute_storage,
414        data_dir,
415        node_id,
416        !allow_shared_kernel,
417        options.daemon_runtime.clone(),
418        worker_exe.as_deref(),
419    )
420    .await;
421    // Adopt the IPs of already-running replicas into each backend's fresh-on-boot
422    // IP pool BEFORE the reconcile loop starts allocating. A backend with a per-node
423    // pool (the native container backend) rebuilds it empty each process start; without
424    // this the boot reconcile could re-hand a live address to a different workload —
425    // the container-IP collision — or move a replica's endpoint on relaunch. Feeds
426    // every persisted replica's `(workload, replica, endpoint-ip)`; each backend keeps
427    // only the IPs in its own subnet (a cheap no-op for docker/cloudflare/VMM).
428    crate::compute::adopt_running_replica_ips(&deploy, &compute_backends).await;
429    // Per-project internal DNS (service discovery): start the resolver on the bridge
430    // gateway so a guest resolves peers by name within its project. On by default;
431    // starts only when the container backend + bridge are up (Linux). Detached for
432    // the node's serving life (pushed into `reconcile` below). Started before the
433    // reconcile loop consumes `compute_backends` — it borrows the registry to check
434    // the container backend is present.
435    let internal_dns =
436        crate::compute::spawn_internal_dns(config.compute.as_ref(), &compute_backends, &deploy);
437    // Activate the compute sql-shim (PLAN-compute-bindings): bind its listener +
438    // build the resolver when a sql provider and `compute.sql_shim_url` are both present.
439    #[cfg(feature = "handlers")]
440    let sql_resolver = boatramp_server::sql_shim::spawn_sql_shim(
441        handlers.sql_backends(),
442        config.compute.as_ref().and_then(|c| c.sql_shim_url.clone()),
443    )
444    .await;
445    #[cfg(not(feature = "handlers"))]
446    let sql_resolver: Option<Arc<dyn boatramp_core::compute::ComputeBindingResolver>> = None;
447
448    // Managed compute-backed SQL (PLAN-managed-compute-sql P2-b): if the handler
449    // `sql` config declares any managed database, inject its `POSTGRES_*`/`MYSQL_*`
450    // server-init env into the DB workload at launch from the sealed credential.
451    // Reaching here with a managed DB implies an envelope (build_handler_runtime
452    // fails closed otherwise), so the credential store always has one to seal with.
453    // Keep a clone of the secrets envelope for the operator-SQL capability below
454    // (the managed_db_resolver match moves the original).
455    #[cfg(any(feature = "sql-postgres", feature = "sql-mysql"))]
456    let operator_envelope = secrets_envelope.clone();
457    // …and a second clone for the tenant-deprovision capability (drops a deleted
458    // tenant's managed DB/role/credential on project/site delete). It needs a real
459    // envelope to seal/unseal + delete per-tenant credentials, so it is wired only
460    // when one is present (same fail-closed gating as the managed-DB paths).
461    #[cfg(any(feature = "sql-postgres", feature = "sql-mysql"))]
462    let deprovision_envelope = secrets_envelope.clone();
463    // …and a clone for the provisioning drift-repair capability (owner-model retrofit /
464    // reconcile). Like the migrate path it connects as the sealed owner role and re-seals
465    // credentials, so it needs a real envelope; wired only when one is present.
466    #[cfg(any(feature = "sql-postgres", feature = "sql-mysql"))]
467    let repair_envelope = secrets_envelope.clone();
468    // …and a third clone for the soft-delete tombstone reaper (the leader-gated task
469    // that hard-drops a Shared-Postgres tenant once its grace window elapses). It, too,
470    // needs a real envelope to unseal the superuser credential + delete the per-tenant
471    // one on hard-drop.
472    #[cfg(any(feature = "sql-postgres", feature = "sql-mysql"))]
473    let reaper_envelope = secrets_envelope.clone();
474    // …and a clone for the declarative managed-database capability (#501 Stage B): it
475    // provisions the declared DB + seals its credential server-side, so it needs a real
476    // envelope; wired only when one is present (fail-closed like every managed-DB path).
477    #[cfg(any(feature = "sql-postgres", feature = "sql-mysql"))]
478    let declare_envelope = secrets_envelope.clone();
479    #[cfg(any(feature = "sql-postgres", feature = "sql-mysql"))]
480    let managed_db_resolver: Option<Arc<dyn boatramp_core::compute::ManagedDbEnvResolver>> = match (
481        config
482            .handlers
483            .as_ref()
484            .and_then(|h| h.bindings.sql.as_ref()),
485        secrets_envelope,
486    ) {
487        (Some(sql), Some(envelope)) if !sql.databases.is_empty() => {
488            // Crown-jewel (C2): sealed managed-DB credentials go through the frontier-syncing
489            // `secrets_kv`, so a freshly-minted DB credential is past the durable frontier before
490            // the workload sees it — never droppable by the self-heal-on-open trailing-tail quarantine.
491            let creds =
492                crate::managed_sql::ManagedSqlCredentials::new(secrets_kv.clone(), envelope);
493            let privilege = config
494                .compute
495                .as_ref()
496                .map(|c| c.managed_db_privilege)
497                .unwrap_or_default();
498            let env =
499                crate::managed_sql::ManagedDbEnv::from_config(&sql.databases, creds, privilege);
500            (!env.is_empty()).then(|| Arc::new(env) as Arc<_>)
501        }
502        _ => None,
503    };
504    #[cfg(not(any(feature = "sql-postgres", feature = "sql-mysql")))]
505    let managed_db_resolver: Option<Arc<dyn boatramp_core::compute::ManagedDbEnvResolver>> = None;
506
507    // Turnkey managed DB: auto-register the compute workload(s) backing each managed
508    // co-located database that has none yet, so declaring the `databases` binding is
509    // enough to boot the DB (no separate `compute set` / apply). Tenant-aware — a
510    // `Shared` binding registers its one shared server; a `Single` binding registers
511    // nothing at boot (its per-tenant `<compute>-<ident>` is created durably by the lazy
512    // resolve on first `sql` use and relaunched by the reconcile, so a project that never
513    // uses `sql` — e.g. a static-only site — never gets a spurious DB). Non-clobbering +
514    // idempotent; runs before the reconcile loop so its first tick can launch what it
515    // registered.
516    #[cfg(any(feature = "sql-postgres", feature = "sql-mysql"))]
517    if let Some(sql) = config
518        .handlers
519        .as_ref()
520        .and_then(|h| h.bindings.sql.as_ref())
521        .filter(|sql| !sql.databases.is_empty())
522    {
523        crate::managed_sql::auto_register_managed_db_workloads(&deploy, &sql.databases).await;
524    }
525
526    // Operator SQL capability (managed-DB migrations/queries via the sealed credential, resolved
527    // server-side) — backs `POST /api/sql/{db}/{exec,query}`. The SAME concrete NodeOperatorSql
528    // also backs the owner-gated schema-migration runner (which reuses its owner + superuser
529    // backends), so build it ONCE and share it.
530    // Build the SAME concrete NodeOperatorSql once (when a managed DB is configured) and share it:
531    // it backs both `operator_sql` (the sql exec/query cap) and the migration runner (which reuses
532    // its owner + superuser backends). Two separate bindings so neither annotation is a complex type.
533    // The migration substrate is a single dispatcher (crate::managed_sql::DispatchMigrationRunner)
534    // routing each `(project, db)` to its engine's substrate: the sqlx NodeMigrationRunner for a
535    // Postgres/MySQL binding, the LibsqlMigrationRunner for a `libsql` binding. It compiles whenever a
536    // sqlx engine OR `migrate` (⇒ libsql) is on, so the embedded-libsql default can migrate even on a
537    // node with no external sqlx engine. `operator_sql` (the `POST /api/sql/{db}/{exec,query}` cap)
538    // stays sqlx-only — a libsql file has no operator-SQL/credential seam.
539    #[cfg(any(feature = "sql-postgres", feature = "sql-mysql"))]
540    let operator_sql: Option<Arc<dyn boatramp_core::sql::OperatorSql>>;
541    // Late-init: the match arms assign it (and, under sqlx, `operator_sql` in the same block), and the
542    // arms differ by feature-cfg — so a direct `let … = match {…}` would need cfg'd arm bodies. The
543    // late-init keeps that readable; the value is always assigned before use.
544    #[cfg(any(feature = "sql-postgres", feature = "sql-mysql", feature = "migrate"))]
545    #[allow(clippy::needless_late_init)]
546    let migration_substrate: Option<Arc<dyn boatramp_core::sql::MigrationSubstrate>>;
547    #[cfg(any(feature = "sql-postgres", feature = "sql-mysql", feature = "migrate"))]
548    match config
549        .handlers
550        .as_ref()
551        .and_then(|h| h.bindings.sql.as_ref())
552        .filter(|sql| !sql.databases.is_empty())
553    {
554        Some(sql) => {
555            // The sqlx (Postgres/MySQL) arm — the shared NodeOperatorSql backs both the operator-SQL
556            // cap and the sqlx migration runner. Only built when a sqlx engine is compiled in.
557            #[cfg(any(feature = "sql-postgres", feature = "sql-mysql"))]
558            let node_op = Arc::new(crate::managed_sql::NodeOperatorSql::new(
559                sql.databases.clone(),
560                kv.clone(),
561                operator_envelope,
562                deploy.clone(),
563            ));
564            // The operator's trusted-extension allowlist — the only extensions a migration may
565            // enable (empty ⇒ none). See ExternalSqlConfig::migrate_trusted_extensions.
566            #[cfg(any(feature = "sql-postgres", feature = "sql-mysql"))]
567            let trusted: std::collections::BTreeSet<String> = sql
568                .migrate_trusted_extensions
569                .clone()
570                .unwrap_or_default()
571                .into_iter()
572                .collect();
573            migration_substrate = Some(Arc::new(crate::managed_sql::DispatchMigrationRunner::new(
574                #[cfg(any(feature = "sql-postgres", feature = "sql-mysql"))]
575                node_op.clone(),
576                #[cfg(any(feature = "sql-postgres", feature = "sql-mysql"))]
577                trusted,
578                #[cfg(feature = "migrate")]
579                sql.databases.clone(),
580            ))
581                as Arc<dyn boatramp_core::sql::MigrationSubstrate>);
582            #[cfg(any(feature = "sql-postgres", feature = "sql-mysql"))]
583            {
584                operator_sql = Some(node_op as Arc<dyn boatramp_core::sql::OperatorSql>);
585            }
586        }
587        None => {
588            #[cfg(any(feature = "sql-postgres", feature = "sql-mysql"))]
589            {
590                operator_sql = None;
591            }
592            migration_substrate = None;
593        }
594    }
595    // When only `migrate` (no sqlx) is compiled, the operator-SQL cap does not exist.
596    #[cfg(all(
597        not(any(feature = "sql-postgres", feature = "sql-mysql")),
598        feature = "migrate"
599    ))]
600    let operator_sql: Option<Arc<dyn boatramp_core::sql::OperatorSql>> = None;
601    #[cfg(not(any(feature = "sql-postgres", feature = "sql-mysql", feature = "migrate")))]
602    let migration_substrate: Option<Arc<dyn boatramp_core::sql::MigrationSubstrate>> = None;
603    #[cfg(not(any(feature = "sql-postgres", feature = "sql-mysql", feature = "migrate")))]
604    let operator_sql: Option<Arc<dyn boatramp_core::sql::OperatorSql>> = None;
605
606    // Tenant-deprovision capability (drop a deleted tenant's managed DB/role/sealed
607    // credential on project/site delete). Wired only when a compute-backed managed
608    // database + a secrets envelope are both present — same gating as operator_sql,
609    // plus the envelope requirement (it must seal/unseal per-tenant credentials).
610    // The soft-delete grace window for a Shared-Postgres managed tenant
611    // (`handlers.bindings.sql.deprovision_grace_secs`, env-settable). Default 7 days;
612    // `0` disables the soft path (immediate hard drop). Threaded to the deprovisioner
613    // (which soft-deletes) and implicitly honored by the reaper (which only ever finds
614    // tombstones a >0 grace produced).
615    #[cfg(any(feature = "sql-postgres", feature = "sql-mysql"))]
616    let deprovision_grace_secs = config
617        .handlers
618        .as_ref()
619        .and_then(|h| h.bindings.sql.as_ref())
620        .and_then(|sql| sql.deprovision_grace_secs)
621        .unwrap_or(crate::tenant_sql::DEFAULT_DEPROVISION_GRACE_SECS);
622    #[cfg(any(feature = "sql-postgres", feature = "sql-mysql"))]
623    let tenant_deprovisioner: Option<Arc<dyn boatramp_core::sql::TenantDeprovisioner>> = config
624        .handlers
625        .as_ref()
626        .and_then(|h| h.bindings.sql.as_ref())
627        .filter(|sql| !sql.databases.is_empty())
628        .zip(deprovision_envelope)
629        .map(|(sql, envelope)| {
630            Arc::new(crate::tenant_sql::NodeTenantDeprovisioner::new(
631                deploy.clone(),
632                kv.clone(),
633                envelope,
634                sql.databases.clone(),
635                deprovision_grace_secs,
636            )) as Arc<_>
637        });
638    #[cfg(not(any(feature = "sql-postgres", feature = "sql-mysql")))]
639    let tenant_deprovisioner: Option<Arc<dyn boatramp_core::sql::TenantDeprovisioner>> = None;
640
641    // Provisioning drift-repair capability (owner-model retrofit / reconcile) — backs the
642    // `Project·Admin`-gated `/api/repair/{db}` + `/dry-run`. Same gating as operator_sql plus
643    // the envelope requirement (it re-seals the owner credential + connects as it). The
644    // envelope is threaded as `Some(_)` so a lean-but-managed node still gets a clear
645    // per-check error rather than a panic if none is configured.
646    #[cfg(any(feature = "sql-postgres", feature = "sql-mysql"))]
647    let tenant_repair: Option<Arc<dyn boatramp_core::sql::TenantRepair>> = config
648        .handlers
649        .as_ref()
650        .and_then(|h| h.bindings.sql.as_ref())
651        .filter(|sql| !sql.databases.is_empty())
652        .map(|sql| {
653            Arc::new(crate::repair::NodeTenantRepair::new(
654                sql.databases.clone(),
655                deploy.clone(),
656                kv.clone(),
657                repair_envelope.clone(),
658            )) as Arc<_>
659        });
660    #[cfg(not(any(feature = "sql-postgres", feature = "sql-mysql")))]
661    let tenant_repair: Option<Arc<dyn boatramp_core::sql::TenantRepair>> = None;
662
663    // Declarative managed-database capability (#501 Stage B) — backs the `Project·Admin`-gated
664    // `PUT /api/projects/{proj}/databases/{name}` + `POST …/ensure`. It persists a manifest
665    // `databases:` entry to `project/{project}/database/{name}`, enforces daemon-config-wins at
666    // the merge point (against the node-static `sql.databases`), refuses an identity change,
667    // binds provisioning to the caller's project, and eagerly provisions via `provision_tenant`.
668    // Requires a sqlx engine (it provisions a Postgres/MySQL server) AND a `[secrets]` envelope
669    // (a managed DB seals its credential). UNLIKE repair/operator_sql it is NOT gated on
670    // `!sql.databases.is_empty()` — a node with NO node-static databases can still accept a
671    // project declaration (the whole point of the declarative front door). The node-static map
672    // (possibly empty) is threaded so the daemon-wins conflict check has both sources.
673    #[cfg(any(feature = "sql-postgres", feature = "sql-mysql"))]
674    let managed_db_declare: Option<Arc<dyn boatramp_core::sql::ManagedDbDeclare>> =
675        declare_envelope.map(|envelope| {
676            let sql_cfg = config
677                .handlers
678                .as_ref()
679                .and_then(|h| h.bindings.sql.as_ref());
680            let static_dbs = sql_cfg.map(|sql| sql.databases.clone()).unwrap_or_default();
681            // Per-project declaration ceilings (#501 Stage B MEDIUM-1 — the disk-exhaustion
682            // guard): the operator's `max_declared_databases`/`max_declared_volume_mib`, or
683            // the built-in `apply_db_caps` defaults (16 DBs / 512 GiB) when unset.
684            let quota = crate::managed_db_declare::DeclareQuota::from_config(
685                sql_cfg.and_then(|sql| sql.max_declared_databases),
686                sql_cfg.and_then(|sql| sql.max_declared_volume_mib),
687            );
688            Arc::new(crate::managed_db_declare::NodeManagedDbDeclare::new(
689                static_dbs,
690                deploy.clone(),
691                kv.clone(),
692                envelope,
693                quota,
694            )) as Arc<_>
695        });
696    #[cfg(not(any(feature = "sql-postgres", feature = "sql-mysql")))]
697    let managed_db_declare: Option<Arc<dyn boatramp_core::sql::ManagedDbDeclare>> = None;
698
699    // Operator compute-exec capability (run a command inside a running workload) —
700    // backs `POST /api/compute/{name}/exec`, gated by the `allow_compute_exec`
701    // posture. Clone the backend registry before the reconcile loop consumes it.
702    let compute_exec: Option<Arc<dyn boatramp_core::compute::ComputeExec>> = Some(Arc::new(
703        crate::compute::NodeComputeExec::new(compute_backends.clone(), deploy.clone()),
704    ) as Arc<_>);
705
706    // Operator volume-reclamation capability (list + remove persistent volumes) —
707    // backs `GET /api/compute/volumes` + `DELETE /api/compute/volumes/{name}`.
708    // Same admin-scoped `/api/compute/*` gate; clone the registry before the
709    // reconcile loop consumes the original below.
710    let compute_volumes: Option<Arc<dyn boatramp_core::compute::ComputeVolumes>> = Some(Arc::new(
711        crate::compute::NodeComputeVolumes::new(compute_backends.clone(), deploy.clone()),
712    )
713        as Arc<_>);
714
715    // Operator reconcile-plane control capability (restart a replica) — backs
716    // `POST /api/compute/maintenance/restart` (admin-scoped). Clone the registry
717    // before the reconcile loop consumes the original below.
718    let compute_control: Option<Arc<dyn boatramp_core::compute::ComputeControl>> = Some(Arc::new(
719        crate::compute::NodeComputeControl::new(compute_backends.clone(), deploy.clone()),
720    )
721        as Arc<_>);
722
723    let compute_reconcile = boatramp_server::spawn_compute_reconcile(
724        deploy.clone(),
725        compute_backends,
726        vec![compute_node],
727        boatramp_core::compute::BackendPolicy::from_shared_kernel_allowed(allow_shared_kernel),
728        is_leader.clone(),
729        compute_reconcile_tick(),
730        COMPUTE_IDLE_TIMEOUT,
731        sql_resolver,
732        managed_db_resolver,
733    );
734
735    // Tenant tombstone reaper: leader-gated hard-drop of soft-deleted Shared-Postgres
736    // tenants past their grace window (safe deprovision — see `tenant_sql`). Wired only
737    // when a compute-backed managed database + a secrets envelope are both present
738    // (same gating as the deprovisioner); each tombstone carries its own server +
739    // superuser, so the reaper needs no per-binding config. A `0` grace never writes a
740    // tombstone, so the sweep is simply inert then.
741    #[cfg(any(feature = "sql-postgres", feature = "sql-mysql"))]
742    let tombstone_reaper: Option<tokio::task::JoinHandle<()>> = config
743        .handlers
744        .as_ref()
745        .and_then(|h| h.bindings.sql.as_ref())
746        .filter(|sql| !sql.databases.is_empty())
747        .zip(reaper_envelope)
748        .map(|(_sql, envelope)| {
749            crate::tenant_sql::spawn_tenant_tombstone_reaper(
750                deploy.clone(),
751                kv.clone(),
752                envelope,
753                is_leader.clone(),
754                crate::tenant_sql::TOMBSTONE_REAPER_TICK,
755            )
756        });
757    #[cfg(not(any(feature = "sql-postgres", feature = "sql-mysql")))]
758    let tombstone_reaper: Option<tokio::task::JoinHandle<()>> = None;
759
760    // Domain-verify auto-complete: periodically re-check every site's pending
761    // ownership challenges and attach any that now pass — a published token (e.g.
762    // via `domain add --provider`) converges without a manual `domain verify`.
763    let dv_reconcile = boatramp_server::spawn_domain_verify_reconcile(
764        deploy.clone(),
765        domain_verify_allow_private,
766        is_leader,
767        DOMAIN_VERIFY_RECONCILE_TICK,
768    );
769
770    // Wire the operator capabilities onto the options the router is built from.
771    let mut options = options;
772    options.operator_sql = operator_sql;
773    options.migration_substrate = migration_substrate;
774    options.tenant_repair = tenant_repair;
775    options.managed_db_declare = managed_db_declare;
776    options.tenant_deprovisioner = tenant_deprovisioner;
777    options.compute_exec = compute_exec;
778    options.compute_volumes = compute_volumes;
779    options.compute_control = compute_control;
780    // The internal secret store backs the admin secrets API (set/list/delete). Not
781    // handlers-gated — it must be reachable even on a lean node.
782    options.secret_store = secret_store;
783    // The per-tenant sealed secret store backs the control-plane CRUD (task #493). Like the secret
784    // store it is not handlers-gated, so the endpoints work on a lean node.
785    options.tenant_secret_store = tenant_secret_store;
786    // The email-profile store backs the admin API (`/api/email/profiles`); like the
787    // secret store it is not handlers-gated, so it works on a lean node.
788    options.email_profile_store = email_profile_store;
789
790    // The detached reconcile loops: the always-present compute + domain-verify ones,
791    // plus the optional tenant-tombstone reaper (only when a managed DB is configured).
792    let mut reconcile = vec![compute_reconcile, dv_reconcile];
793    if let Some(reaper) = tombstone_reaper {
794        reconcile.push(reaper);
795    }
796    if let Some(dns) = internal_dns {
797        reconcile.push(dns);
798    }
799
800    Ok(RunningNode {
801        deploy,
802        handlers,
803        auth,
804        options,
805        reconcile,
806    })
807}
808
809/// Build the `[secrets]` envelope (secrets-at-rest wrapping) from `boatramp.cfg`'s
810/// `[secrets]` section: `local` (a machine-local AES-256-GCM KEK) or `vault` (Vault
811/// Env var an operator points at a PEM file of extra CA(s) the guest's outbound `wasi:http` TLS
812/// client should trust on top of the webpki roots — honored only under the
813/// `allow_guest_egress_extra_ca` posture (a hermetic HTTPS test double lever).
814const GUEST_EGRESS_EXTRA_CA_ENV: &str = "BOATRAMP_GUEST_EGRESS_EXTRA_CA_FILE";
815
816/// Parse the operator's guest-egress extra-CA PEM ([`GUEST_EGRESS_EXTRA_CA_ENV`]) into rustls trust
817/// anchors, gated by the `allow_guest_egress_extra_ca` posture (`allow`). No env set ⇒ empty (the
818/// default). `allow == false` (e.g. multi-tenant) with a file set ⇒ empty + a warning (the posture
819/// refuses it). Set + readable + ≥1 cert ⇒ those certs. Set-but-unreadable / no valid cert ⇒ a hard
820/// error (fail closed — a configured-but-broken CA must not silently degrade to no-trust).
821fn load_guest_egress_extra_roots(
822    allow: bool,
823) -> Result<Vec<rustls::pki_types::CertificateDer<'static>>> {
824    let path = match std::env::var(GUEST_EGRESS_EXTRA_CA_ENV) {
825        Ok(p) if !p.is_empty() => p,
826        _ => return Ok(Vec::new()),
827    };
828    if !allow {
829        tracing::warn!(
830            env = GUEST_EGRESS_EXTRA_CA_ENV,
831            "ignoring a guest-egress extra CA: the security posture forbids it \
832             (allow_guest_egress_extra_ca is off — e.g. under multi-tenant)"
833        );
834        return Ok(Vec::new());
835    }
836    let pem =
837        std::fs::read(&path).map_err(|e| Error::GuestEgressCa(format!("reading {path:?}: {e}")))?;
838    let certs = rustls_pemfile::certs(&mut std::io::BufReader::new(&pem[..]))
839        .collect::<std::result::Result<Vec<_>, _>>()
840        .map_err(|e| Error::GuestEgressCa(format!("parsing {path:?}: {e}")))?;
841    if certs.is_empty() {
842        return Err(Error::GuestEgressCa(format!(
843            "{path:?} contained no PEM certificate"
844        )));
845    }
846    tracing::info!(
847        env = GUEST_EGRESS_EXTRA_CA_ENV,
848        count = certs.len(),
849        path = %path,
850        "guest egress trusts operator extra CA(s) (dev-posture; webpki roots still apply)"
851    );
852    Ok(certs)
853}
854
855/// Transit). `None`/empty ⇒ no wrapping. The Vault token is read from the
856/// environment (`token_env`), never a file. This seals a managed SQL credential at
857/// rest; a managed database fails closed without it.
858fn build_secrets_envelope(
859    secrets: Option<&crate::config::SecretsConfig>,
860    data_dir: &Path,
861) -> Result<Option<Arc<dyn boatramp_core::envelope::KeyEnvelope>>> {
862    use boatramp_server::envelope::{EnvelopeSpec, build_envelope};
863    let Some(cfg) = secrets else {
864        return Ok(None);
865    };
866    let spec = match cfg.envelope.as_str() {
867        "" => EnvelopeSpec::None,
868        "local" => EnvelopeSpec::Local {
869            kek_file: cfg
870                .kek_file
871                .clone()
872                .unwrap_or_else(|| data_dir.join("secrets/kek")),
873        },
874        "vault" => {
875            let v = cfg.vault.as_ref().ok_or_else(|| {
876                Error::Envelope(
877                    "secrets.envelope = \"vault\" needs a [secrets.vault] section".into(),
878                )
879            })?;
880            let token = std::env::var(&v.token_env).map_err(|_| {
881                Error::Envelope(format!("Vault token env `{}` is not set", v.token_env))
882            })?;
883            EnvelopeSpec::Vault {
884                addr: v.addr.clone(),
885                key: v.key.clone(),
886                token,
887            }
888        }
889        other => {
890            return Err(Error::Envelope(format!(
891                "unknown secrets.envelope {other:?} (want \"local\" or \"vault\")"
892            )));
893        }
894    };
895    build_envelope(spec).map_err(|e| Error::Envelope(e.to_string()))
896}
897
898#[cfg(all(test, feature = "fs"))]
899mod tests {
900    use super::*;
901    use boatramp_core::kv::MemoryKv;
902    use boatramp_core::security::SecurityProfile;
903
904    /// The dev/loopback ephemeral fleet-signer auto-provision is gated STRICTLY to a loopback bind
905    /// (or an in-process embedder with no bind): a public / wildcard / private-network bind must
906    /// NOT silently provision an ephemeral signing key (it would invalidate live capabilities on a
907    /// restart — such a node must supply a persistent key). This is the security-critical boundary.
908    #[cfg(feature = "handlers")]
909    #[test]
910    fn ephemeral_fleet_signer_auto_provisions_only_on_loopback_or_in_process() {
911        use std::net::SocketAddr;
912        let sa = |s: &str| s.parse::<SocketAddr>().unwrap();
913        // In-process (no listener) and loopback → auto-provision the dev fleet signer.
914        assert!(should_autoprovision_fleet_signer(None));
915        assert!(should_autoprovision_fleet_signer(Some(sa(
916            "127.0.0.1:8080"
917        ))));
918        assert!(should_autoprovision_fleet_signer(Some(sa("[::1]:8080"))));
919        // Off-host-reachable binds MUST NOT auto-provision an ephemeral key: a public IP, a
920        // private-network IP, and a wildcard bind (`0.0.0.0`/`::`, reachable off-host).
921        assert!(!should_autoprovision_fleet_signer(Some(sa(
922            "203.0.113.5:8080"
923        ))));
924        assert!(!should_autoprovision_fleet_signer(Some(sa(
925            "10.0.0.4:8080"
926        ))));
927        assert!(!should_autoprovision_fleet_signer(Some(sa("0.0.0.0:8080"))));
928        assert!(!should_autoprovision_fleet_signer(Some(sa("[::]:8080"))));
929    }
930
931    /// The headline in-process fidelity check (PLAN-node-library N2b.3): `assemble`
932    /// over a temp `FsStorage` + `MemoryKv` produces a `RunningNode` whose deploy
933    /// store is live (the reserved `default` project was materialized during
934    /// assembly) and whose router — the exact one `boatramp serve` builds — answers
935    /// `/healthz`. No listener is bound: the request is driven through the router
936    /// via `tower::oneshot`, so the whole assembly runs in-process.
937    #[tokio::test]
938    async fn assemble_produces_a_serving_node_over_a_temp_store() {
939        use axum::body::Body;
940        use axum::http::{Request, StatusCode};
941        use tower::ServiceExt;
942
943        let tmp = tempfile::tempdir().unwrap();
944        let storage: Arc<dyn Storage> = Arc::new(boatramp_storage::FsStorage::new(tmp.path()));
945        let kv: Arc<dyn KvStore> = Arc::new(MemoryKv::new());
946        let config = ServerConfig::default();
947        let options = boatramp_server::ServerOptions {
948            // The strict `multi-tenant` posture, as an unconfigured `serve` resolves.
949            posture: SecurityProfile::MultiTenant.preset(),
950            ..Default::default()
951        };
952
953        let node = assemble(NodeInput {
954            config: &config,
955            data_dir: tmp.path(),
956            storage,
957            kv,
958            auth: boatramp_server::Auth::disabled(),
959            options,
960            serve_addr: None,
961            watch_provider: None,
962            provision_tier: boatramp_core::blob_notify::ProvisionTier::default(),
963            messaging: None,
964            is_leader: Arc::new(|| true),
965            node_id: 0,
966            worker_exe: None,
967        })
968        .await
969        .expect("assemble a node over a temp store");
970
971        // The deploy store is live: `assemble` already materialized the reserved
972        // `default` project, so a second ensure reports "already present" (`false`).
973        assert!(
974            !node
975                .deploy
976                .ensure_default_project()
977                .await
978                .expect("read the default project"),
979            "assemble should have materialized the default project"
980        );
981
982        // The assembled router (the same wiring `serve` binds) answers /healthz.
983        let router =
984            boatramp_server::router_with(node.deploy, node.auth, node.handlers, node.options);
985        let response = router
986            .oneshot(
987                Request::builder()
988                    .uri("/healthz")
989                    .body(Body::empty())
990                    .unwrap(),
991            )
992            .await
993            .expect("route /healthz");
994        assert_eq!(response.status(), StatusCode::OK);
995    }
996}