gunnar_server/

serve.rs

//! `gunnar serve` starts its listeners together and stops them together. They
//! share nothing but the control plane behind them.
//!
//! ```text
//!   :2222            SSH   git-upload-pack / git-receive-pack, publickey auth
//!   :8080            HTTP  the same two verbs over smart HTTP, access-token
//!                          auth — only with `--features http` AND `--http-bind`
//!   127.0.0.1:50051  gRPC  the control plane; authenticated (sshsig + the
//!                          ACL model) unless the development escape hatch is on
//! ```
//!
//! **The two git listeners are two doors onto one room**, and the wiring below
//! is what keeps that true: the same `Dispatch`, the same `WireTransport`, the
//! same `ServerMetrics` and the one `allow_anonymous` field are handed to both.
//! A second control plane or a second transport here is how SSH and HTTP would
//! come to disagree about who may push.
//!
//! Git traffic never passes through gRPC. There is no bridge and none is
//! wanted: it is that separation which makes the v1 authentication hole
//! survivable, because an open gRPC port cannot push code.
//!
//! # What this module is responsible for saying out loud
//!
//! 1. **Whether the unauthenticated enrolment door is open**, at every start,
//!    on stdout *and* in the log. `gunnar_core::AuthTable::prepare_to_serve`
//!    returns the state as a value; [`announce_readiness`] is the caller that
//!    has to branch on it, and `tests/bootstrap.rs` asserts every branch.
//!
//!    Until 2026-08-05 an empty identity table with no `--admin-key` was a
//!    **refusal to start** (the secret-store seam). That is why gunnar had no first boot: the
//!    box could not be reached to be given its first key, and the only way in
//!    was `gunnar key add` with the server down, which needs a shell the
//!    appliance does not have. It now starts and serves `Bootstrap.Enroll`
//!    until somebody uses it — see `gunnar_grpc::service::bootstrap`.
//! 2. **gRPC binds loopback unless told otherwise**, and says so at startup
//!    whatever it binds.
//!
//! # Two listeners, and now two runtimes
//!
//! Separate listeners separate the *ports*. They did not separate the *schedulers*, and that
//! turned out to be most of what the separation was for.
//!
//! `ControlPlane` is a synchronous trait — `gunnar-grpc`'s `seam.rs` says so,
//! and says the backend belongs behind `spawn_blocking` "at the call site" —
//! and no call site does it. `Admin.StorageInfo` runs an unbounded recursive
//! `std::fs::read_dir` that `stat`s every loose object in the repository,
//! straight from its `async fn`, on a worker thread. `Repos.Create` opens a git
//! directory and replays an arrow log. While both listeners shared one runtime,
//! a handful of concurrent control-plane calls occupied every worker and the SSH
//! listener's accept, handshake and exec dispatch simply did not run. The loopback default leaves
//! the control port unauthenticated, so that was reachable on purpose as well as
//! by accident.
//!
//! **The control plane therefore gets its own runtime**, built here, with its
//! own worker threads and its own blocking pool. Two properties follow that
//! thirty `spawn_blocking` wrappers would not have given:
//!
//! * it holds for *every* control-plane call, including ones written later by
//!   somebody who never read `seam.rs`. The fix is by construction rather than a
//!   convention thirty call sites have to keep (one implementation);
//! * a control plane that stalls completely degrades the control plane only. Git
//!   keeps being served.
//!
//! The gRPC listener is bound **on** that runtime rather than handed to it: a
//! tokio socket registers with the I/O driver of the runtime that created it,
//! and one whose driver belongs to a runtime that is shutting down is a stall
//! waiting to happen. `tests/control_plane_starvation.rs` measures the whole
//! thing from an OS thread outside both runtimes.

use std::net::SocketAddr;
use std::sync::Arc;

use anyhow::{bail, Context, Result};
use gunnar_core::observe::{MetricsSnapshot, RequestId, ServerMetrics};
use gunnar_core::{ensure_host_key, Readiness};
use gunnar_grpc::{EventBus, ServerConfig, ServiceContext};
use gunnar_ssh::{BoundServer, GunnarSsh, SshConfig};
use tokio::sync::oneshot;
use tokio::task::JoinHandle;

use gunnar_common::bindspec::BindSpec;

use crate::control_plane::CoreControlPlane;
use crate::stores::ServerStores;
use crate::transport::{WireConfig, WireTransport};

/// What `gunnar serve` was asked for.
#[derive(Debug, Clone)]
pub struct ServeConfig {
    /// Where the SSH listener binds.
    pub ssh_bind: SocketAddr,
    /// Where the gRPC control plane binds. Loopback by default.
    ///
    /// A [`BindSpec`] rather than a `SocketAddr` because an appliance's private
    /// address is DHCP-learned: the declaration names `eth1:50051` and
    /// [`run`] resolves it, ONCE, after `GUNNAR_NET` has brought the link up
    /// and before anything binds. Downstream — the readiness line, the
    /// dashboard row, `Admin.Health` — then reports the address the listener
    /// actually took, not the spelling it was asked for.
    pub grpc_bind: BindSpec,
    /// The DEVICE the control plane listens on, when `grpc_bind` was spelled as
    /// an interface (`eth1:50051`) — set by [`run`]/[`start`] from that spelling
    /// and honoured as `SO_BINDTODEVICE` on the listener socket, so the
    /// interface name means the interface and not merely the address it held
    /// at boot. MEASURED 2026-09-05 (the KVM bench rig, two slirp NICs on the
    /// same 10.0.2.0/24): an address-only bind of eth1's 10.0.2.15 answered
    /// through eth0, because both NICs held that address and the route decided.
    /// On a cloud whose utility NIC is `eth1` on its own subnet the two agree;
    /// the device bind makes them agree everywhere. `None` for an address spec.
    pub grpc_bind_device: Option<String>,
    /// An OpenSSH public key line to install as the bootstrap admin when the
    /// identity table is empty (the secret-store seam).
    pub admin_key: Option<String>,
    /// **The keys `Bootstrap.Restore` is taken from** (`--restore-pubkey`, h-v1-6a): OpenSSH
    /// public key lines or `@<path>`, resolved and checked at start. An installer-fed box adds
    /// the fed admin keys itself; a box with neither refuses every restore by name.
    pub restore_pubkeys: Vec<String>,
    /// Start the SSH listener only.
    pub no_grpc: bool,
    /// **`--wg-only`: every box-to-box connection runs over the estate's
    /// WireGuard overlay** (owner, 2026-10-04). The address the control plane
    /// is bound on (`eth1:50051`, the utility network) serves NOTHING: the
    /// control plane and the replication listener answer only on the tunnel's
    /// address ([`crate::tunnel_addr`]), and SSH — bound for the public's git
    /// clients — drops a connection that arrives at the utility address before
    /// a byte is spoken. A plain-IP connection between two boxes is refused.
    pub wg_only: bool,
    /// `--role twin`: no forge listener from the first boot (loopback only),
    /// whether or not `Bootstrap.BecomeTwin` has run yet.
    pub twin_role: bool,
    /// Whether a caller with **no key** may clone a repository marked
    /// [`Visibility::Public`](gunnar_core::Visibility::Public). `true` by
    /// default; nothing is anonymously readable until a repository is
    /// deliberately made public, because `Private` is the default for a new
    /// one.
    ///
    /// Reaches [`SshConfig::allow_anonymous`](gunnar_ssh::SshConfig::allow_anonymous)
    /// **and `HttpConfig::allow_anonymous`**, from this one field. Two fields
    /// is how one of them ends up stubbed: an operator who passed
    /// `--no-anonymous` and got it honoured on one transport has a server that
    /// serves strangers and reports that it does not. The gRPC control plane
    /// has no anonymous mode to switch, at any setting.
    pub allow_anonymous: bool,
    /// **`--anonymous allow|refuse`** (the browse replica): whether an
    /// anonymous read is SERVED here or sent to the browse replica the
    /// refusal names. Orthogonal to `allow_anonymous`, which is whether a
    /// stranger authenticates at all; this is what the dispatcher tells one
    /// who did. Both front doors read it through the one `Control`.
    pub anonymous: gunnar_ssh::AnonymousPolicy,
    /// Where the smart-HTTP listener binds. `None` does not start
    /// it, which is the default.
    ///
    /// Off unless asked for, because a second listening port is a decision an
    /// operator makes and not one a default makes for them. The feature it
    /// needs is off by default too, so an appliance image compiles no HTTP at
    /// all.
    #[cfg(feature = "http")]
    pub http_bind: Option<SocketAddr>,
    /// Whether the HTTP listener will read a credential off a plain connection
    /// Default: only when a proxy attests TLS.
    #[cfg(feature = "http")]
    pub http_cleartext: gunnar_http::Cleartext,
    /// Terminate TLS on the smart-HTTP listener itself. `None` — the default —
    /// is the proxy-terminated shape; `SelfMinted` is the single-executable
    /// shape (leaf minted from this server's own issuer at start, chain
    /// closed for clients that pin the issuer); `Identity` is
    /// bring-your-own-PEM.
    #[cfg(feature = "http")]
    pub http_tls: Option<HttpTls>,
    // (HttpTls is defined right below ServeConfig.)
    /// **Build every push**, in this process, hermetically (roadmap phase 3).
    ///
    /// `None` — the default — is a server that stores and serves and does not
    /// build. `Some` starts the runner on the same event bus every push
    /// publishes on, which is what makes the golden path's fourth step
    /// *nothing at all*.
    ///
    /// It is in the serving process on purpose: a second process cannot
    /// subscribe to an in-process bus, and a runner that polled the attestation
    /// log instead would build a push some time *after* it landed rather than
    /// *because* it landed.
    pub runner: Option<crate::runner::WatchConfig>,
    /// Require a **client certificate** on the gRPC control plane, minted by
    /// this server's own issuer (the mTLS plan).
    ///
    /// `None` — the default — is the loopback shape the unauthenticated v1 assumes: the listener
    /// binds `127.0.0.1` and a credential in metadata is the identity.
    /// `Some(GrpcTls::SelfMinted)` is the shape a control plane reachable off
    /// the box needs, and it changes two things at once: the transport refuses
    /// a peer whose certificate this issuer did not sign, and the gate maps the
    /// certificate that got through to an ACL principal.
    ///
    /// Deliberately **not** on by default. Turning it on rewrites how every
    /// existing client must connect, and a security default that silently
    /// breaks the working configuration gets switched back off in a hurry —
    /// which leaves it off everywhere. It is a flag an operator turns on when
    /// the listener leaves loopback, and `--grpc-bind` off loopback without it
    /// is what the startup warning names.
    pub grpc_tls: Option<GrpcTls>,
    /// `--web-relying-party`, unparsed: the WebAuthn relying parties the gRPC
    /// plane answers for (`gunnar_grpc::webauthn::RelyingParty`).
    pub web_relying_parties: Vec<String>,
    /// `--web-require-user-verification` (A34-12).
    pub web_require_user_verification: bool,
    /// **The mutual replication listener** (`--grpc-replication-bind`, lane T9):
    /// the door a twin follows this box through. Served only with a client
    /// certificate this box's issuer signed, TLS 1.3 only, its leaf minted from
    /// the issuer at every start under the fixed name
    /// `gunnar_grpc::client_signer::REPLICATION_SERVER_NAME`.
    ///
    /// `None` DERIVES it, and the derivation is what lets the ONE medium carry
    /// no flag: a control plane bound off loopback (the appliance,
    /// `0.0.0.0:50051`) gets this listener on the same address, port + 1
    /// (`50052`); a loopback one (a laptop, a test) gets none. Its main
    /// listener stays exactly what it was, because turning `--grpc-tls-self` on
    /// for the whole plane would lock out every client that has no certificate
    /// — and the bootstrap door with them.
    pub grpc_replication_bind: Option<SocketAddr>,
    // (HttpTls is defined right below ServeConfig.)
    /// How many git transfers may be in flight across the whole server.
    /// `None` takes
    /// [`DEFAULT_MAX_CONCURRENT_TRANSFERS`](gunnar_ssh::config::DEFAULT_MAX_CONCURRENT_TRANSFERS).
    ///
    /// The fix note on the admission gate says every one of these numbers needs to be
    /// settable, because the right one is deployment-specific; until now
    /// neither was reachable from `gunnar serve` and the two defaults were the
    /// only values any deployment could have. That is also why the pressure
    /// harness could not reach contention without 320 client processes.
    pub max_concurrent_transfers: Option<usize>,
    /// The **memory budget** the admission ledger may give out to git
    /// operations in flight, in bytes (gate 2). `None` takes
    /// [`crate::budget::budgets`]'s derived figure — `MemAvailable` x
    /// [`crate::budget::TRANSFER_MEMORY_PERCENT`] — rather than
    /// `gunnar-ssh`'s compiled-in 8 GiB, which was a number about the cgroup
    /// one bakeoff ran in and not about this box.
    pub memory_budget_bytes: Option<u64>,
    /// How long a transfer may **wait** for a slot before it is refused.
    /// `None` takes [`gunnar_ssh::governor::DEFAULT_ADMISSION_WAIT`];
    /// `Some(Duration::ZERO)` restores the instant refusal.
    pub admission_wait: Option<std::time::Duration>,
    /// A principal's guaranteed floor and its share of a contended server.
    /// `None` takes
    /// [`DEFAULT_MAX_TRANSFERS_PER_PRINCIPAL`](gunnar_ssh::config::DEFAULT_MAX_TRANSFERS_PER_PRINCIPAL).
    pub max_transfers_per_principal: Option<usize>,
    /// How many git reads may be served **without a key** at once.
    /// `None` takes
    /// [`DEFAULT_MAX_ANONYMOUS_TRANSFERS`](gunnar_ssh::admission::DEFAULT_MAX_ANONYMOUS_TRANSFERS).
    ///
    /// Note the asymmetry with the field above, which is deliberate: that one
    /// is a floor a principal is guaranteed and may exceed on a quiet box, this
    /// one is a wall that binds however empty the server is. See
    /// [`gunnar_ssh::admission`] for why work-conserving is the right instinct
    /// for a tenant and the wrong one for a stranger.
    pub max_anonymous_transfers: Option<usize>,
    /// **The usage share** (`admit-share`): `--admit-limit` and
    /// `--admit-window-days`, parsed. See [`gunnar_ssh::SshConfig::admit`].
    #[cfg(feature = "admit-share")]
    pub admit: gunnar_ssh::edda_admit::Settings,
    /// Make the per-principal number an unconditional wall again rather than a
    /// max-min share. See
    /// [`SshConfig::per_principal_limit_is_hard`](gunnar_ssh::SshConfig::per_principal_limit_is_hard).
    pub per_principal_limit_is_hard: bool,
    /// How many worker threads the control plane's **own** runtime gets.
    /// `None` derives it from [`control_plane_workers`].
    ///
    /// A knob rather than a constant, because the right number is a deployment
    /// question: a box whose operators live in the UI wants more, an appliance
    /// serving one CI fleet wants fewer. A test also has to be able to make the
    /// pool small enough to saturate it deliberately.
    pub control_plane_workers: Option<usize>,
    /// The largest object `Refs.ReadObject` will inflate. `None`
    /// takes
    /// [`CoreControlPlane`](crate::CoreControlPlane)'s default of 64
    /// MiB; `Some(0)` means no ceiling.
    ///
    /// A knob for the same reason `control_plane_workers` is one: the number is
    /// `memory budget / concurrent inflates`, and the budget is a deployment
    /// question. It is also the only one of these that turns a request which
    /// works today into a refusal, which is exactly why it must be reachable
    /// rather than compiled in.
    pub max_object_bytes: Option<u64>,
    /// How often the counter heartbeat is written. `None` turns it
    /// off.
    ///
    /// The heartbeat answers *"what is this server doing right now"* from a log
    /// the operator already has open, with nothing to scrape. It stays quiet
    /// while nothing changes, so an idle appliance does not fill its journal.
    pub metrics_interval: Option<std::time::Duration>,
    /// Where the **scrape endpoint** binds. `None` does not start it,
    /// which is the default.
    ///
    /// The heartbeat above and this are the same nineteen numbers reached two
    /// ways, and both are wanted: the heartbeat needs nothing set up and is
    /// already in the journal when an incident is reconstructed afterwards; the
    /// endpoint can be *asked*, at whatever rate the person asking chooses,
    /// which is the only way `transfers_in_flight` distinguishes a slow
    /// transfer from a stuck one while it is still happening.
    ///
    /// Off unless asked for, for the same reason [`ServeConfig::http_bind`] is:
    /// a second listening port is an operator's decision. See
    /// [`crate::scrape`] for why it carries no authentication and what
    /// that does and does not expose.
    pub metrics_bind: Option<SocketAddr>,
    /// Replication to a second gunnar (replication). `None` reads
    /// `<data-root>/replication.json`, which is what an operator configures and
    /// what a restart picks up; `Some` overrides it in-process and is how a
    /// test points a primary at a standby it just started on an ephemeral port.
    ///
    /// A configuration with no targets starts nothing at all: no task, no
    /// subscriber. A server that replicates nowhere pays nothing for the
    /// feature existing.
    #[cfg(feature = "replication")]
    pub replication: Option<crate::mirror::ReplicationConfig>,
    /// **I11's passive heartbeat**: probe this primary and take over if it stops
    /// answering. `None` — the default — starts no task at all.
    ///
    /// Off unless asked for, and that is the design rather than caution:
    /// `control_plane.rs::replication_abdicate` states the rule, *"naming a new
    /// primary is the operator's decision, never a default this box picks"*, and
    /// a box that promoted itself because nobody configured it not to would
    /// make that decision for every deployment that never heard of the feature.
    ///
    /// It does nothing on a box whose role is `Primary`: the task re-reads the
    /// role every tick and leaves when it is not a standby, so setting it on
    /// both boxes of a pair is harmless and is what a symmetric unit file wants.
    /// See [`crate::standby_watch`] for what the probe does and does not prove.
    #[cfg(feature = "replication")]
    pub standby_watch: Option<crate::standby_watch::StandbyWatch>,
    /// **The estate repairs itself** (5.4): the boot check after the control plane
    /// listens (it RECHARGES only with `auto`, the estate's opt-in, off by default)
    /// and the pull channel to monetize. `None` (the default): neither runs.
    /// See [`crate::repair`].
    #[cfg(feature = "replication")]
    pub repair: Option<crate::repair::RepairConfig>,
    /// **Where replication's counters are published to**, for the console band, the
    /// staleness alarm and `Admin.Stats`.
    ///
    /// It is on the config rather than returned from [`Server::start`] because
    /// the console band starts BEFORE the server does — [`run`] hands the same
    /// view to both, and `start` drops the replicator into it once there is
    /// one. Left at its default it reports "nothing configured", which is what
    /// a server with no standby is.
    pub replication_view: crate::replication_view::ReplicationView,
    /// **Where the band's `tenants` row is read from**, the same shape and
    /// for the same reason as `replication_view`: a metered `Server::start`
    /// fills it from its entitlement slot; a community build never does, and
    /// the band draws no row.
    pub tenants_view: crate::dashboard::TenantView,
    /// **Where the band's `flow` row is read from.** Filled on every build —
    /// counting what moved is not the metered appliance's capability. See
    /// [`crate::dashboard::FlowView`].
    pub flow_view: crate::dashboard::FlowView,
    /// What the two git state machines serve under (`gunnar-wire`'s own
    /// `Config` types, held once — see [`WireConfig`]).
    ///
    /// Defaults to `gunnar-wire`'s defaults, which is what every deployment
    /// silently ran on while `with_config` had no callers. The knobs it carries
    /// were fully implemented in both protocol versions and reachable from
    /// nowhere; `--no-filter` in particular is the difference between an
    /// appliance that declines partial clone and one that only says it might.
    pub wire: WireConfig,
    /// **The metered appliance's key and pay page** (`metered` only): what
    /// `--monetize-pubkey` and `--billing-url` resolve to. `None` on a metered
    /// build starts it with every namespace Free and `Entitlement.Set` refusing
    /// by name — never with a key it did not have.
    #[cfg(feature = "metered")]
    pub metered: Option<crate::entitlement::MeteredConfig>,
    /// **The disk-headroom floor** (`--disk-floor`): the community plan gate,
    /// first link of every front door's chain and the guard on `Repos.Create`.
    /// See `gunnar_ssh::headroom`. Always present — the default is 5 % or
    /// 2 GiB, whichever is larger — because a box with no floor is a box that
    /// finds out it is full from a torn pack.
    pub disk_floor: gunnar_ssh::DiskFloor,
    /// **What measures the data set** for the floor and the space ledger.
    /// `None` — every deployment — is one `statvfs` on the data root. A test
    /// that must fill a data set of a stated size without a filesystem of that
    /// size hands its own probe here (`tests/push_space_admission.rs`'s
    /// virtual disk); nothing else about the server changes.
    pub space_probe: Option<Arc<dyn gunnar_ssh::SpaceProbe>>,
}

/// The `--anonymous` policy type, re-exported so a caller that builds a
/// [`ServeConfig`] (the test harness, `gunnar-cli`) names it from here.
pub use gunnar_ssh::AnonymousPolicy;

impl Default for ServeConfig {
    fn default() -> Self {
        ServeConfig {
            ssh_bind: SocketAddr::from(([0, 0, 0, 0], gunnar_ssh::config::DEFAULT_PORT)),
            grpc_bind: BindSpec::Addr(gunnar_grpc::DEFAULT_BIND),
            grpc_bind_device: None,
            admin_key: None,
            restore_pubkeys: Vec::new(),
            no_grpc: false,
            wg_only: false,
            twin_role: false,
            allow_anonymous: true,
            anonymous: gunnar_ssh::AnonymousPolicy::Allow,
            #[cfg(feature = "http")]
            http_bind: None,
            #[cfg(feature = "http")]
            http_cleartext: gunnar_http::Cleartext::default(),
            #[cfg(feature = "http")]
            http_tls: None,
            grpc_tls: None,
            web_relying_parties: Vec::new(),
            web_require_user_verification: false,
            grpc_replication_bind: None,
            runner: None,
            max_concurrent_transfers: None,
            memory_budget_bytes: None,
            admission_wait: None,
            max_transfers_per_principal: None,
            max_anonymous_transfers: None,
            #[cfg(feature = "admit-share")]
            admit: gunnar_ssh::edda_admit::Settings::default(),
            per_principal_limit_is_hard: false,
            control_plane_workers: None,
            max_object_bytes: None,
            metrics_interval: Some(METRICS_INTERVAL),
            metrics_bind: None,
            #[cfg(feature = "replication")]
            replication: None,
            #[cfg(feature = "replication")]
            standby_watch: None,
            #[cfg(feature = "replication")]
            repair: None,
            replication_view: crate::replication_view::ReplicationView::new(),
            tenants_view: crate::dashboard::TenantView::new(),
            flow_view: crate::dashboard::FlowView::new(),
            wire: WireConfig::default(),
            #[cfg(feature = "metered")]
            metered: None,
            disk_floor: gunnar_ssh::DiskFloor::default(),
            space_probe: None,
        }
    }
}

impl ServeConfig {
    /// A configuration whose listeners take kernel-assigned loopback ports.
    /// A constructor for tests, not a mode: nothing about what is authenticated
    /// or dispatched changes.
    pub fn ephemeral() -> Self {
        ServeConfig {
            ssh_bind: SocketAddr::from(([127, 0, 0, 1], 0)),
            grpc_bind: BindSpec::Addr(SocketAddr::from(([127, 0, 0, 1], 0))),
            grpc_bind_device: None,
            ..ServeConfig::default()
        }
    }
}

/// How long [`Server::shutdown`] lets the gRPC listener finish gracefully
/// before it is aborted.
///
/// Long enough for an in-flight unary call to complete on a loaded box, short
/// enough that a Ctrl-C is answered while the operator is still watching. See
/// [`Server::shutdown`] for why any bound is needed at all.
pub const GRPC_SHUTDOWN_GRACE: std::time::Duration = std::time::Duration::from_secs(2);

/// How often `gunnar serve` writes its counters to the log.
///
/// Sixty seconds: often enough that an operator watching an incident sees the
/// numbers move while they are still watching, rare enough that a year of
/// journal on an idle appliance is not made of heartbeats. The line is skipped
/// when nothing has changed since the last one, so "quiet" and "not running"
/// stay distinguishable. A *busy* server is never quiet.
pub const METRICS_INTERVAL: std::time::Duration = std::time::Duration::from_secs(60);

/// The size of the blocking pool that carries git transfers.
///
/// **A decision, not a default.** Every in-flight transfer occupies one blocking
/// thread for its whole duration (`WireTransport::serve` bridges the async
/// duplex into `gunnar-wire`'s blocking `BufRead`/`Write` through
/// `SyncIoBridge`), so this number *was* the server's real concurrency ceiling,
/// and it was 512 because that is what `tokio` picks when nobody says. Past it,
/// transfers queued with no log, no counter and no error.
///
/// The ceiling that matters is now
/// [`SshConfig::max_concurrent_transfers`](gunnar_ssh::SshConfig::max_concurrent_transfers),
/// which refuses visibly. This pool has to stay comfortably **above** it: it
/// also carries every other blocking thing the server does — pack indexing, ref
/// transactions, the arrow log — and a pool sized exactly to the transfer limit
/// would deadlock the moment a transfer needed a second blocking task.
/// `the_blocking_pool_is_larger_than_the_transfer_limit` asserts the ordering
/// rather than leaving it to this paragraph.
pub const GIT_BLOCKING_THREADS: usize = 512;

/// Worker threads that RUN the control plane's calls, derived from the
/// machine.
///
/// A quarter of the cores, at least four and at most eight. The control plane is
/// an administrative surface — a handful of operators and a UI — not the
/// serving path, and the whole point of giving it a separate pool is that stalls
/// stay inside a bounded number of threads. Giving it one core's worth would
/// make a single slow `StorageInfo` serialise the UI; giving it the machine
/// would recreate the problem with extra steps. The floor was two until
/// R5.1-G (2026-10-07): two slow synchronous calls held both, and every other
/// call — the console's pages, monetize's `Entitlement.Set` — queued behind
/// them.
///
/// These are the HANDLER workers ([`ServiceContext::with_handler_runtime`]).
/// The HTTP/2 connections run on [`CONTROL_CONNECTION_WORKERS`] workers of
/// their own, which no handler can hold, so a client's keepalive PING is
/// answered however long the calls take.
pub fn control_plane_workers() -> usize {
    let cores = std::thread::available_parallelism()
        .map(|n| n.get())
        .unwrap_or(4);
    (cores / 4).clamp(4, 8)
}

/// Worker threads that drive the control plane's HTTP/2 connections (R5.1-G).
///
/// They accept, read and write frames — and answer a client's keepalive PING —
/// and nothing else: every call runs on the handler runtime
/// ([`control_plane_workers`] workers). They used to be one runtime, and when
/// the slow synchronous calls held all of its workers nothing answered the
/// PING; the console and monetize then killed the connection (their keepalive:
/// a PING every 10 s, dead 5 s later) and every call in flight on it failed.
pub const CONTROL_CONNECTION_WORKERS: usize = 2;

/// How many CPUs this process may **run on**, from its affinity mask.
///
/// `/proc/self/status`'s `Cpus_allowed_list`, because that is the kernel's own
/// answer and not a request anybody made. `Some(n)` or `None` when the file is
/// not there to be read; a missing answer is never guessed at, because the whole
/// point of [`report_cpu_budget`] is a comparison and half a comparison is
/// worse than none.
fn cpus_allowed() -> Option<usize> {
    let status = std::fs::read_to_string("/proc/self/status").ok()?;
    let list = status
        .lines()
        .find_map(|line| line.strip_prefix("Cpus_allowed_list:"))?;
    count_cpu_list(list)
}

/// How many CPUs a kernel CPU list names: `0-31`, `3`, `0-3,8,10-11`.
///
/// Split out of [`cpus_allowed`] so it can be exercised on lists this host does
/// not have. It matters: the reference desktop reports a plain `0-31`, so a version of this that
/// counted ranges and dropped the singletons passed every test on this box
/// while answering **0 allowed CPUs on a server pinned to one core** — the
/// condition the caller exists to detect, read exactly backwards.
fn count_cpu_list(list: &str) -> Option<usize> {
    let mut total = 0usize;
    for range in list.trim().split(',') {
        let range = range.trim();
        if range.is_empty() {
            continue;
        }
        total += match range.split_once('-') {
            Some((lo, hi)) => {
                let (lo, hi) = (
                    lo.trim().parse::<usize>().ok()?,
                    hi.trim().parse::<usize>().ok()?,
                );
                hi.checked_sub(lo)?.checked_add(1)?
            }
            None => {
                range.parse::<usize>().ok()?;
                1
            }
        };
    }
    (total > 0).then_some(total)
}

/// Whether the affinity is wide enough against the quota to be worth a line,
/// and the two numbers to say it with.
///
/// Pure, and given both readings rather than taking them, so the condition can
/// be seen in both states from a test on any machine. Either reading missing is
/// `None`: half a comparison is worse than none.
///
/// **Twice, not any excess.** The two numbers are rarely exactly equal and a
/// line that fires on nine-versus-eight is a line operators learn to skip.
/// Twice is where the measurement above already showed a real cost.
fn spread_budget(allowed: Option<usize>, quota: Option<f64>) -> Option<(usize, usize)> {
    let (allowed, quota) = (allowed?, quota?);
    let budget = quota.ceil().max(1.0) as usize;
    (allowed >= budget.saturating_mul(2)).then_some((allowed, budget))
}

/// How many CPUs' worth of time this process may **use**, from cgroup v2
/// `cpu.max`. `None` when there is no quota, which is the ordinary case.
fn cpu_quota() -> Option<f64> {
    let own = std::fs::read_to_string("/proc/self/cgroup").ok()?;
    let rel = own
        .lines()
        .find_map(|line| line.strip_prefix("0::"))?
        .trim()
        .trim_start_matches('/');
    let path = std::path::Path::new("/sys/fs/cgroup")
        .join(rel)
        .join("cpu.max");
    let raw = std::fs::read_to_string(path).ok()?;
    let mut parts = raw.split_whitespace();
    let quota = parts.next()?;
    if quota == "max" {
        return None;
    }
    let quota: f64 = quota.parse().ok()?;
    let period: f64 = parts.next()?.parse().ok()?;
    (period > 0.0).then(|| quota / period)
}

/// **Say when the box will let this server run far wider than it will let it
/// work**, because that costs more than anything in the request path.
///
/// A container started with `--cpus 8` on a 32-core host gets eight CPUs of
/// *quota* spread over thirty-two CPUs of *affinity*. gunnar serves one transfer
/// on three threads that hand work to each other — a hyper task on a runtime
/// worker, the `spawn_blocking` thread that runs the git state machine, and the
/// pack-entry worker `iter_from_counts` spawns — so at N concurrent transfers
/// the scheduler has ~3N runnable threads and thirty-two places to put them,
/// while only eight of them may run at a time. Every hand-off then crosses a
/// core, and a hand-off that crosses a core is a futex.
///
/// Measured on the reference desktop, 2026-08-04, one binary, one 8-CPU budget, the `year`
/// corpus (28 893 objects) over smart HTTP, ramp 1→32 clients:
///
/// | budget applied as | CPU-seconds per request, 1 client → 32 |
/// |---|---|
/// | `cpu.max` quota only, 32 CPUs allowed | 0.42 → 0.94 |
/// | the same quota, affinity pinned to 8  | 0.34 → 0.35 |
///
/// The quota-only arm spent **40% of the server's CPU** in the kernel's futex
/// hash-bucket spinlock (`native_queued_spin_lock_slowpath`, `perf record` on
/// the serving process). The pinned arm is flat: gunnar's per-request cost does
/// not in fact grow with concurrency, and the ramp that said it did was
/// measuring where the threads were allowed to run.
///
/// So this is a line in the log and not a change of behaviour: gunnar must not
/// pin itself. It does not know which CPUs are *its* to take, and choosing wrong
/// would be worse than the condition being reported. The operator does know, and
/// `--cpuset-cpus` is the whole fix.
///
/// The decision is [`spread_budget`], separately, because a condition read off
/// the host it is running on can only ever be seen in the state that host
/// happens to be in — and this one has to be seen in both.
fn report_cpu_budget() {
    let Some((allowed, budget)) = spread_budget(cpus_allowed(), cpu_quota()) else {
        return;
    };
    tracing::warn!(
        cpus_allowed = allowed,
        cpu_quota = budget,
        "this server may run on {allowed} CPUs but may only use {budget} CPUs' worth of \
         them. Every hand-off between the three threads that serve one transfer then \
         crosses a core, and in a 2026-08-04 measurement that cost 2.7x the CPU per \
         request at 32 concurrent clients against the identical budget applied as a \
         cpuset. Start the \
         container with `--cpuset-cpus` (or the unit with `AllowedCPUs=`) naming {budget} \
         CPUs, not only `--cpus`/`CPUQuota=`."
    );
}

/// The runtime `gunnar serve` runs the **git** listener on.
///
/// Built here rather than with `#[tokio::main]` so a client-only build links no
/// async runtime at all, and built with an explicit
/// [`GIT_BLOCKING_THREADS`] so the number that used to be the server's
/// accidental concurrency ceiling is now written down somewhere.
pub fn runtime() -> Result<tokio::runtime::Runtime> {
    tokio::runtime::Builder::new_multi_thread()
        .max_blocking_threads(GIT_BLOCKING_THREADS)
        .thread_name("gunnar-git")
        .enable_all()
        .build()
        .context("building the git runtime")
}

/// Both listeners, bound and running.
/// How the smart-HTTP listener gets its TLS identity (see
/// [`ServeConfig::http_tls`]).
#[cfg(feature = "http")]
#[derive(Debug, Clone)]
pub enum HttpTls {
    /// Mint a leaf from this server's own issuer at start — the
    /// single-executable deployment. Refuses to start if it cannot mint;
    /// silent cleartext fallback is the defect this enum exists to prevent.
    SelfMinted,
    /// Operator-supplied chain and key.
    Identity(gunnar_http::TlsIdentity),
}

/// How the gRPC control plane gets its **mutual** TLS material.
///
/// The mirror of [`HttpTls`], with one difference that is the whole point: the
/// smart-HTTP listener terminates TLS to encrypt, and this one terminates TLS
/// to *identify*. So there is no bring-your-own-CA arm here yet. A certificate
/// the gate can turn into a principal has to have been registered in the ACL
/// table, and the only thing that mints and registers in one act is this
/// server's own issuer (`Bootstrap.Enroll`, `gunnar control bootstrap`).
/// Trusting a foreign CA would mean a certificate that passes the handshake and
/// can never pass the gate, which is a confusing way to be refused.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum GrpcTls {
    /// Mint the listener's own leaf from this server's issuer at start, and
    /// pin that issuer as the CA every client certificate must come from.
    ///
    /// Refuses to start if it cannot mint. A control plane that asked for
    /// client certificates and came up without them would be the same silent
    /// downgrade [`HttpTls`] refuses.
    SelfMinted,
}

/// **A tokio `Runtime` that is always torn down with `shutdown_background`,
/// never by a blocking `Drop`.**
///
/// Dropping a `Runtime` blocks the current thread on its worker and blocking
/// pools' shutdown, and tokio **panics** when that blocking drop happens inside
/// another runtime's context: *"Cannot drop a runtime in a context where
/// blocking is not allowed."*
///
/// [`Server::start`] is an `async fn` that builds the control plane's own
/// runtime and then runs several more fallible steps before it moves that
/// runtime into [`Server`]. A `?` anywhere in that window used to drop a bare
/// `Runtime` on the async stack — and on the appliance gunnar is **PID 1**, so
/// that panic is the kernel's `Attempted to kill init`. Measured 2026-08-31:
/// a power cut left a 0-byte `replication.json`, its parse in `start`
/// failed, the `?` unwound past the live runtime, and the box wedged on every
/// reboot. Wrapping the runtime in this guard makes the early return safe **by
/// construction** — one drop path, and it never blocks — instead of by
/// remembering to hand-`take()` it at each of the several exits.
///
/// Since R5.1-G it holds the control plane's TWO runtimes — the connections'
/// and the calls' — under the one guard, in that order.
#[derive(Default)]
struct ControlRuntime(Vec<tokio::runtime::Runtime>);

impl ControlRuntime {
    /// Hold a runtime whose drop must be non-blocking.
    fn new(runtime: tokio::runtime::Runtime) -> Self {
        ControlRuntime(vec![runtime])
    }

    /// Take over every runtime `other` holds, so one guard drops them all.
    fn hold(&mut self, other: &mut ControlRuntime) {
        self.0.append(&mut other.0);
    }

    /// A handle on the first runtime held, if any.
    fn handle(&self) -> Option<tokio::runtime::Handle> {
        self.0.first().map(|runtime| runtime.handle().clone())
    }

    /// Tear the runtimes down now, without blocking, if any are still held.
    /// Idempotent, so an explicit call and the `Drop` cannot double-free.
    fn shutdown(&mut self) {
        for runtime in self.0.drain(..) {
            runtime.shutdown_background();
        }
    }
}

impl Drop for ControlRuntime {
    fn drop(&mut self) {
        self.shutdown();
    }
}

pub struct Server {
    /// An `Option` only because [`Server`] now has a [`Drop`] impl, which makes
    /// moving a field out of `self` in [`Server::shutdown`] illegal.
    ssh: Option<BoundServer>,
    grpc: Option<JoinHandle<Result<()>>>,
    grpc_shutdown: Option<oneshot::Sender<()>>,
    /// The bound gRPC address and whether that listener authenticates, as
    /// ONE field. They are printed together and must not be able to disagree.
    grpc_endpoint: Option<(SocketAddr, bool)>,
    /// The control plane's **own** runtime, kept alive here because
    /// dropping it would stop the listener it carries.
    ///
    /// A [`ControlRuntime`] guard, not a bare `Runtime`: dropping a `Runtime`
    /// blocks, and doing so from inside another runtime's context panics
    /// ("Cannot drop a runtime in a context where blocking is not allowed"),
    /// and both [`Server::start`] and [`Server::shutdown`] run in async
    /// context. The guard routes *every* drop through `shutdown_background`, so
    /// no `?`, panic or forgotten `take()` can turn a teardown into an abort —
    /// which as PID 1 is a kernel panic.
    control_runtime: ControlRuntime,
    stores: Arc<ServerStores>,
    /// Whether the unauthenticated `Bootstrap.Enroll` window was open when this
    /// server started, so the banner on stdout says the same thing the log does.
    ///
    /// A field rather than a re-read: the banner and the warning must not be
    /// able to disagree about the single most important fact a fresh server has
    /// (one implementation).
    bootstrap_window_open: bool,
    /// **The SSH host key this server is serving, `SHA256:…`.**
    ///
    /// Kept so the startup banner can print it. It is logged at start through
    /// `tracing`, which on the appliance means stderr, which means
    /// `/dev/console`, which is the LAST `console=` on the command line, which
    /// is the serial line — and the UpCloud shape has no serial line. So the
    /// one emission this box makes of the one fact a client has to pin
    /// reached nobody in production, and `known_hosts` got filled by scanning
    /// instead. A field, not a re-read, for the same reason
    /// `bootstrap_window_open` is one: the banner and the log must not be able
    /// to disagree about it.
    host_key_fingerprint: String,
    /// The counters both listeners write to.
    metrics: Arc<ServerMetrics>,
    /// The heartbeat task, aborted at shutdown so a stopped server stops
    /// narrating.
    heartbeat: Option<JoinHandle<()>>,
    /// I11's standby watch, held so it is aborted with the server. A watchdog
    /// that outlived the box it watches over could promote a server that is
    /// shutting down.
    #[cfg(feature = "replication")]
    standby_watch: Option<JoinHandle<()>>,
    /// The boot check and the pull channel ([`crate::repair`]), aborted with the
    /// server (a recharge it was running fails its restore stream, never ends it).
    #[cfg(feature = "replication")]
    repair: Option<JoinHandle<()>>,
    /// The event bus both the transport and the control plane publish to.
    bus: Arc<EventBus>,
    /// The smart-HTTP listener's task and the address it bound (smart HTTP).
    /// Dropped with the `Server`, and therefore aborted.
    #[cfg(feature = "http")]
    http: Option<(JoinHandle<std::io::Result<()>>, SocketAddr, bool)>,
    /// The scrape listener's task and the address it bound. Dropped
    /// with the `Server`, and therefore aborted: a stopped server must stop
    /// answering, or a scraper keeps drawing the last numbers it saw.
    scrape: Option<(JoinHandle<std::io::Result<()>>, SocketAddr)>,
    /// The replicator, when any standby is configured.
    /// the replicator and its workers, in the LIVE SLOT the control plane
    /// reconfigures through.
    ///
    /// It was an `Option<Arc<Replicator>>` plus an `Option<ReplicationTasks>`
    /// until 2026-08-30 — decided once at startup, from a file. That is
    /// unusable on the shape this ships to: an OS-less appliance with no shell,
    /// where the only way to name a standby is over the control plane. See
    /// [`ReplicationSlot`](crate::mirror::ReplicationSlot).
    ///
    /// Dropped with the `Server`, and therefore aborted, so a stopped server
    /// stops replicating.
    #[cfg(feature = "replication")]
    replication: Arc<crate::mirror::ReplicationSlot>,
    /// The outbound webhook delivery slot, reconfigured through the control plane.
    /// Dropped with the `Server`, and therefore aborted, so a stopped server
    /// stops delivering.
    #[cfg(feature = "webhooks")]
    webhook: Arc<crate::webhook::WebhookSlot>,
}

/// One line every `interval` naming what the server has done, but only when
/// something moved.
///
/// **Not a scrape endpoint.** The load target is a forge, and the first
/// question at an incident gets answered from the log the operator already has
/// open. A Prometheus surface is a separate decision with its own cost: another
/// listener, and another thing to authenticate. This makes the counters
/// readable today, in four lines.
async fn heartbeat(metrics: Arc<ServerMetrics>, interval: std::time::Duration) {
    let mut ticker = tokio::time::interval(interval);
    // The first tick fires immediately; skip it so start-up does not emit a row
    // of zeroes that reads like traffic.
    ticker.tick().await;
    let mut previous = metrics.snapshot();
    loop {
        ticker.tick().await;
        let now = metrics.snapshot();
        // An idle server says nothing. A server that has served ANYTHING since
        // the last beat says what, including one whose only change is that a
        // transfer finished; that is how `transfers_in_flight` coming back to
        // zero becomes visible.
        if now == previous {
            continue;
        }
        emit_metrics(&now);
        previous = now;
    }
}

/// The gRPC control plane's request ids, minted from the **same** counter the
/// SSH listener and the smart-HTTP listener mint theirs from.
///
/// This one line is why `gunnar_grpc::router` takes a source instead of owning
/// one: `gunnar-grpc` has no `gunnar-core` edge and must not grow one (it would
/// put `aws-lc-rs`, C and a build script into the `--features client,control`
/// graph), and a private counter over there would mean a control-plane call and
/// a git transfer could print the same `req=` value in one log.
///
/// `gunnar_grpc::trace`'s rendering is asserted equal to
/// [`RequestId`]'s by `the_grpc_request_id_renders_exactly_as_gunnar_core_
/// renders_one` in `tests/observability.rs`, because a shared counter rendered
/// two ways is still two identifiers.
fn grpc_request_id() -> u64 {
    RequestId::next().get()
}

/// The heartbeat line.
///
/// **This function used to BE the line**, and it listed eighteen of the
/// nineteen counters. `connections_anonymous` — written by both transports
/// (`gunnar-ssh/src/server.rs`, `gunnar-http/src/service.rs`), and the answer to
/// *how much of my traffic is strangers*, which the field's own documentation
/// calls the first question an operator asks after making anything public —
/// appeared in no line gunnar had ever printed. Nothing was wrong with the
/// counter; the hand-written list next to it had simply never been updated when
/// the field was added.
///
/// So it is not a hand-written list any more. The line expands from the same
/// macro arm that declares the fields, in `gunnar-core`, beside the scrape
/// endpoint's renderer, and neither can fall behind the struct (one implementation). This
/// wrapper stays only so the heartbeat task and a test call the same one thing.
fn emit_metrics(now: &MetricsSnapshot) {
    now.emit_heartbeat();
}

impl Server {
    /// Bind everything and start accepting.
    ///
    /// Returns once both sockets are bound, so a caller knows the real
    /// addresses before the first client connects, which is what makes port 0
    /// usable in a test.
    pub async fn start(stores: Arc<ServerStores>, config: ServeConfig) -> Result<Server> {
        report_cpu_budget();
        // The bind spelling becomes an address HERE, before anything else in
        // this function can report one. `run` has already resolved it and
        // replaced the spec with the answer, so on the appliance path this is
        // the identity case; a caller that reaches `start` directly (every
        // test, and `gunnar serve` under an init system) gets the resolution
        // it would otherwise have had to do itself.
        let mut config = config;
        // A bind by ROLE (`utility:50051`, the lean estate) becomes this box's NIC first.
        config.grpc_bind = crate::boot::role_bound_here(config.grpc_bind)?;
        // The interface spelling names a DEVICE as well as an address; keep the
        // device before the spec becomes an address, so the listener can bind it.
        if let BindSpec::Interface { interface, .. } = &config.grpc_bind {
            config.grpc_bind_device = Some(interface.clone());
        }
        // **An interface whose lease PID 1 still asks for** (its boot DHCP failed —
        // estate-box3 D4, a box rebooted while its network was down) is WAITED for: the
        // control plane binds once the address arrives instead of never.
        if let BindSpec::Interface { interface, .. } = &config.grpc_bind {
            let has_address = |i: &str| matches!(appliance_pid1::net::interface_ipv4(i), Ok(Some(_)));
            if !has_address(interface) && appliance_pid1::net::dhcp_pending(interface) {
                tracing::info!(interface = %interface, "the control plane waits for {interface}'s address (PID 1 asks DHCP again in the background)");
                while !has_address(interface) && appliance_pid1::net::dhcp_pending(interface) {
                    tokio::time::sleep(std::time::Duration::from_secs(1)).await;
                }
            }
        }
        // ★ **An estate box is WireGuard-only** (owner 2026-10-04: every
        //   box-to-box connection over the tunnel), whatever its medium says: a
        //   pairing seed with a WireGuard tunnel IS an estate (the front, the
        //   relay, the overlay). A standalone appliance — no seed — is not.
        #[cfg(feature = "wg-client")]
        if !config.wg_only {
            if let Ok(Some(seed)) = crate::pairing::seed(stores.root()) {
                if crate::tunnel_addr::estate_is_wg_only(Some(seed.tunnel == nordisk_pairing::seed::TunnelKind::Wireguard)) {
                    config.wg_only = true;
                    tracing::warn!("serve: this box belongs to an estate (a WireGuard pairing seed) — WireGuard-only: its utility address serves nothing, boxes speak over the tunnel");
                }
            }
        }
        // The posture the pairing reads (which address of the primary a twin joins).
        crate::tunnel_addr::set_wg_only(config.wg_only);
        let resolved_grpc_bind = match resolve_grpc_bind(&config.grpc_bind)? {
            Some(addr) => addr,
            None => {
                // The interface is there and holds no address. There is nothing
                // to bind, so there is no control plane — the same shape
                // `--no-grpc` produces, reached for the same honest reason and
                // reported through the same field, so the dashboard row, the
                // readiness line and `Admin.Health` all agree.
                config.no_grpc = true;
                gunnar_grpc::DEFAULT_BIND
            }
        };
        // ── the bootstrap: refuse to serve rather than start unreachable ───
        let admin_key = match &config.admin_key {
            Some(spec) => Some(read_key_argument(spec)?),
            None => None,
        };
        // The keys a restore is taken from, resolved and CHECKED now: a line that does not
        // parse would otherwise be a key that silently lets nobody in, found out on the day
        // of the restore.
        let restore_keys = config
            .restore_pubkeys
            .iter()
            .map(|spec| {
                let line = read_key_argument(spec)?;
                gunnar_core::auth::parse_public_key(line.trim())
                    .map_err(|err| anyhow::anyhow!("--restore-pubkey {spec:?} is not an OpenSSH public key line: {err:#}"))?;
                Ok(line.trim().to_owned())
            })
            .collect::<Result<Vec<String>>>()?;
        let readiness = stores
            .auth()
            .prepare_to_serve(admin_key.as_deref())
            .with_context(|| {
                format!(
                    "reading the identity table at {}. This is NOT the same as finding it \
                     empty: a data path that cannot be read must never be served as a server \
                     nobody has enrolled yet, because that would reopen the bootstrap window on \
                     a disk fault.",
                    stores.root().display()
                )
            })?;
        announce_readiness(&readiness, stores.root(), resolved_grpc_bind, config.no_grpc);
        let bootstrap_window_open = readiness.bootstrap_window_open();

        // ── the SSH host key, minted in-process and persisted ──────────────
        // A twin frames nothing: its host key is read (the primary's, replayed)
        // or minted through the UNLOGGED store, and its follower starts here.
        let host_key_store = if stores.is_twin() { stores.secrets_unframed()? } else { stores.secrets()? };
        let host_key =
            ensure_host_key(&host_key_store).context("loading or minting the SSH host key")?;
        if stores.is_twin() {
            crate::twin::start_follower(Arc::clone(&stores));
        }
        // The pairing loop, once the control plane serves: join the front,
        // take the role, become the twin if elected (pairing.rs).
        #[cfg(feature = "wg-client")]
        match crate::pairing::seed(stores.root()) {
            Ok(Some(seed)) => crate::pairing::start(Arc::clone(&stores), seed),
            Ok(None) => {}
            Err(e) => tracing::error!(error = %format!("{e:#}"), "pairing: the seed is present and unusable — this box will not pair"),
        }
        let host_key_fingerprint = host_key.fingerprint().to_string();
        tracing::info!(
            fingerprint = %host_key_fingerprint,
            "ssh host key ready (persisted; it survives a restart)"
        );
        // The fingerprint crosses the ssh-key version split as BYTES: gunnar-core
        // links ssh-key 0.6 and russh pins 0.7.0-rc.10, so no key type can be
        // handed from one to the other.
        let russh_key = russh_private_key(host_key.private_openssh())
            .context("handing the host key to russh")?;

        // ── the certificate issuer, minted at first start ──────────────────
        //
        // The same discipline as the host key above: a fresh appliance mints
        // its own, and a restart finds the one it minted. `Pki::create`'s
        // refusal lives under the document's compare-and-swap, so two racing
        // starts cannot mint two issuers. Failure is a warning, not a refusal
        // to serve: SSH-key auth carries the server whether or not certificates
        // exist, and an operator who imports a CA later gets exactly the
        // `issuer import` refusal they should.
        match stores.pki() {
            Ok(pki) => match pki.issuer() {
                Ok(Some(live)) => tracing::info!(
                    subject = %live.subject,
                    fingerprint = %live.fingerprint,
                    "issuer ready (persisted; it survives a restart)"
                ),
                Ok(None) => match pki.create("gunnar issuer", gunnar_core::DEFAULT_ISSUER_DAYS) {
                    Ok(created) => tracing::warn!(
                        subject = %created.subject,
                        fingerprint = %created.fingerprint,
                        "FIRST START: this server minted its own certificate issuer. Its \
                         private key is sealed in the secret store. `gunnar issuer show --pem` \
                         prints the certificate clients pin."
                    ),
                    Err(err) => tracing::warn!(
                        %err,
                        "the issuer could not be minted at first start; certificate issuance \
                         is unavailable until `gunnar issuer create` or `issuer import` succeeds"
                    ),
                },
                Err(err) => tracing::warn!(%err, "the PKI document could not be read"),
            },
            Err(err) => tracing::warn!(
                %err,
                "no secret store, so no issuer: certificate issuance is unavailable"
            ),
        }

        // ── can this server be moved onto mutual TLS? ──────────────────────
        //
        // Asked HERE, and the position is two decisions rather than one.
        //
        // **After the issuer block above**, because that block MINTS one on a
        // first start: asking before it would refuse a server whose issuer is
        // four lines from existing, and the refusal would name the one thing
        // the operator did not have to do.
        //
        // **Before anything binds, and before the control plane's runtime is
        // built.** Two reasons:
        //
        // 1. `--grpc-tls-self` is a one-way door for everybody who cannot
        //    complete the handshake afterwards, and exactly one state makes
        //    it a door with nobody on either side. A server asked for client
        //    certificates that cannot honour the ask must be a startup error
        //    naming the cause, never a listener that came up — the same
        //    discipline `--http-tls-*` already follows.
        // 2. **A `?` past the control runtime is an appliance that will not
        //    boot.** `ControlRuntime`'s own docs record the measurement: that
        //    `Runtime` is built bare inside this function and dropping one on
        //    an async stack panics, and on the appliance gunnar is PID 1, so
        //    the panic is `Attempted to kill init`. The first draft of this
        //    check sat beside the trust root, inside that window, and
        //    `tests/tls_transition.rs` failed with exactly that panic instead
        //    of the refusal it was asserting. That test is the guard that it
        //    does not move back.
        //
        // `crate::tls_preflight` holds the argument and the verdicts;
        // `gunnar issuer readiness` is this same function asked in advance,
        // so the answer an operator plans against is the answer acted on here
        // (one implementation).
        if !config.no_grpc && config.grpc_tls.is_some() {
            let preflight = crate::tls_preflight::preflight(&stores)
                .context("checking whether this server can move to mutual TLS")?;
            if let Some(refusal) = preflight.refusal() {
                anyhow::bail!("{refusal}");
            }
            preflight.warn();
        }

        // ── the SSH listener ───────────────────────────────────────────────
        //
        // ONE set of counters, minted here and shared by the listener
        // and the transport. Two `ServerMetrics` would produce two partial
        // truths, the listener's refusals in one and the transport's pushes in
        // the other, and neither answers "what is this server doing".
        // The stores' set (A28 M17): the flow flush's reconcile pass, started before this,
        // counts its misses on it too.
        let metrics = Arc::clone(stores.metrics());
        let bus = Arc::new(EventBus::new(gunnar_grpc::DEFAULT_EVENT_BUFFER));
        let transport = Arc::new(
            WireTransport::new(Arc::clone(&stores))
                .with_bus(Arc::clone(&bus))
                .with_metrics(Arc::clone(&metrics))
                .with_wire_config(config.wire.clone()),
        );
        let defaults = SshConfig::default();
        // The DECLARED budget (`crate::budget`): cores x TRANSFERS_PER_CORE,
        // capped by the descriptors the handle cache has not already spoken
        // for, with `MemAvailable` sizing the ledger. An operator's
        // `--max-concurrent-transfers` still wins — `.or()` below — and so does
        // `GUNNAR_MAX_TRANSFERS`, which is how the appliance's kernel command
        // line reaches it. Before this the ceiling was the shipped 256, which
        // the governor then walked down to 19 under load while the
        // ledger's flat 320 MiB-per-push refused everything past 25.
        let mut budgets = crate::budget::budgets();
        // **Recorded, so the box can say later whether its socket is where a
        // twin's belongs** (2026-09-21, lane T4). A box told it is a twin AFTER
        // this line keeps whatever was bound here until it restarts;
        // `twin::git_listener_needs_restart` is how that is reported instead of
        // being left to `ss`.
        let twin_shaped = stores.is_twin() || config.twin_role;
        crate::twin::note_git_bind_is_twin_shaped(twin_shaped);
        let ssh_config = SshConfig {
            bind: if twin_shaped {
                // **No forge listener on a twin.** It serves no git to anyone: a
                // read would grow the caches the twin must not hold. The listener
                // the rest of `start` is wired around stays, on loopback, port 0.
                tracing::warn!(target: gunnar_core::observe::TARGET_TWIN, asked = %config.ssh_bind, "this box is a TWIN: the git listener is bound to loopback only");
                SocketAddr::from(([127, 0, 0, 1], 0))
            } else {
                config.ssh_bind
            },
            allow_anonymous: config.allow_anonymous,
            // The banner names the BUILD, not just the crate version. On the
            // OS-less appliance this listener is the only thing that can be
            // asked what is running — there is no shell to run `gunnar
            // --version` in. See `gunnar_common::ssh_server_id`.
            server_id: gunnar_common::ssh_server_id(),
            // `or` rather than `unwrap_or`: `None` here means "the operator did
            // not say", which takes the default. Turning a limit *off* is
            // `Some(0)`-free and stays a code path, not a flag, because an
            // appliance with no transfer ceiling is the unbounded listener
            // exactly as it was.
            max_concurrent_transfers: config.max_concurrent_transfers.or(Some(budgets.transfers)),
            memory_budget_bytes: config.memory_budget_bytes.or(Some(budgets.memory_bytes)),
            admission_wait: config.admission_wait.or(defaults.admission_wait),
            max_transfers_per_principal: config
                .max_transfers_per_principal
                .or(defaults.max_transfers_per_principal),
            // `or` for the same reason as the two above: `None` means the
            // operator did not say, which takes `SshConfig::default`'s
            // ceiling. Switching it OFF is deliberately not reachable from a
            // flag — `Some(0)` refuses everything anonymous, and letting
            // anonymous inherit a keyed identity's share is the hole anonymous read opened
            // and this closed, not a configuration.
            max_anonymous_transfers: config
                .max_anonymous_transfers
                .or(defaults.max_anonymous_transfers),
            per_principal_limit_is_hard: config.per_principal_limit_is_hard,
            #[cfg(feature = "admit-share")]
            admit: config.admit,
            ..defaults
        };
        // The row says what is IN FORCE, after the operator's flags: a line
        // that printed the derivation while the server ran on something else
        // would be worse than no line, because the first thing anybody does
        // with it is compare it to the refusals underneath it.
        if let Some(n) = config.max_concurrent_transfers {
            budgets.transfers = n;
            budgets.transfers_from = crate::budget::Source::Operator;
        }
        if let Some(bytes) = config.memory_budget_bytes {
            budgets.memory_bytes = bytes;
        }
        if let Some(wait) = config.admission_wait {
            budgets.wait_ms = wait.as_millis() as u64;
        }
        crate::budget::announce(&budgets);

        #[cfg(feature = "http")]
        let http_transport: Arc<dyn gunnar_ssh::seam::GitTransport> = Arc::clone(&transport) as _;
        // The SAME `WireTransport`, reached through a second trait rather than
        // a second object: `open_lfs` hands out the store under the repository's
        // own git dir and the `WireConfig::lfs` limits the SSH session is built
        // with, so an operator's `--max-lfs-object-bytes` binds on both
        // transports or on neither.
        // ── the metered appliance's entitlement slot, before any front door ─
        //
        // Opened here so BOTH front doors and the control plane share ONE
        // slot (one implementation): the same arc-swap the SSH dispatch reads is the one
        // `Entitlement.Set` swaps. Loading it re-verifies every stored fact
        // (see `crate::entitlement`), and a key that does not parse is a
        // startup error — a metered box that silently ran Free would be the
        // provider's revenue leaking with no line to say so.
        #[cfg(feature = "metered")]
        let entitlement = match &config.metered {
            Some(metered) => Some(
                crate::entitlement::EntitlementSlot::open(Arc::clone(&stores), metered)
                    .context("opening the entitlement slot")?,
            ),
            None => {
                tracing::warn!(
                    "metered build started without --monetize-pubkey: every namespace is Free \
                     and Entitlement.Set refuses every fact"
                );
                None
            }
        };
        // The band's tenants row reads the SAME slot `Admin.Stats` counts from
        // (one implementation). Filled here, before any door and outside the gRPC branch, so
        // a `--no-grpc` metered box still shows its tenants.
        // **This estate SELLS room** (the feed's `sells_room`, LIVE; gunnar-grpc `selling`):
        // ONE flag, read by the entitlement slot's push gate and the gRPC context alike.
        #[cfg(feature = "metered")]
        let sells_room = entitlement.as_ref().map(|slot| slot.sells_room_flag()).unwrap_or_else(|| Arc::new(std::sync::atomic::AtomicBool::new(false)));
        #[cfg(not(feature = "metered"))]
        let sells_room = Arc::new(std::sync::atomic::AtomicBool::new(false));
        // **The front's egress door** (the lean estate, A19): a box laid with no public IPv4 is
        // fed it, and every client of this box that reaches the internet takes it from then on.
        nordisk_pairing::feedbus::on_published("the front's egress door", crate::feed::take_egress);
        {
            let flag = Arc::clone(&sells_room);
            nordisk_pairing::feedbus::on_published("shop room: a person's account with no fact", move |feed| {
                let sells_room = crate::feed::Fed::from_feed(feed).sells_room;
                if flag.swap(sells_room, std::sync::atomic::Ordering::Relaxed) != sells_room {
                    tracing::info!(target: "gunnar::audit", event = "shop_room", sells_room, "{}", if sells_room { "this estate's room comes from the shop: a person's own account with no fact holds no room until its fact lands" } else { "this estate's room does not come from a shop: an account with no fact is uncapped" });
                }
            });
        }
        #[cfg(feature = "metered")]
        if let Some(slot) = &entitlement {
            // A key the installer feeds later is applied to THIS slot in place
            // (owner 2026-09-25: avoid restarts; `feed::monetize_arrived`).
            let taker = Arc::clone(slot);
            crate::feed::on_monetize_key("entitlement slot", Box::new(move |k| taker.set_key(k).map(|_| ())));
            // ★ UW-6: the SITE's pay page, fed per estate (`refusal_link`), replaces the
            // constant medium's — now and on every feed after (LIVE).
            let pay = Arc::clone(slot);
            nordisk_pairing::feedbus::on_published("entitlement: the site's pay page", move |feed| {
                if let Some(url) = crate::feed::Fed::from_feed(feed).billing_url.as_deref() {
                    if let Err(why) = pay.set_billing_url(url) {
                        tracing::warn!(target: "gunnar::audit", event = "entitlement.billing_url_refused", reason = %why, "the fed pay page was refused; the one in force stays");
                    }
                }
            });
            let slot = Arc::clone(slot);
            config.tenants_view.fill(move || slot.counts());
        }
        // ── spools a crash left behind (cov-13) ──
        //
        // A pushed pack of 64 MiB or more is spooled to `incoming-*.pack` while
        // it arrives and removed by `Drop` when it lands or fails. A process
        // that died mid-receive ran no `Drop`, so the file is still here, full
        // size, under a name nothing reads — room the next push is refused for.
        // No door is open yet, so no live push owns one.
        let (mut swept, mut swept_bytes) = crate::push_space::sweep_spool_leftovers(stores.repos_root());
        // …and an LFS upload's `<txlog>/spool/*.part`, the same shape one door over.
        let (lfs, lfs_bytes) = crate::push_space::sweep_lfs_spool(&stores.txlog().dir().join(gunnar_core::txlog::SPOOL_DIR));
        swept += lfs;
        swept_bytes += lfs_bytes;
        if swept > 0 {
            tracing::warn!(
                files = swept,
                bytes = swept_bytes,
                "deleted {swept} pushed-pack spool file(s) a crash left behind ({})",
                gunnar_ssh::headroom::human(swept_bytes)
            );
        }
        // ── the disk-headroom gate: the community plan gate, every build ──
        //
        // Measured once here so the boot log states the floor IN FORCE beside
        // what is free right now; refused per request in the chain below and
        // in `Repos.Create`. A root that cannot be measured is logged as the
        // refusal it will be, not hidden until the first push finds it.
        //
        // ★ The filesystem's preallocation, measured once here (round 5.2): what a growing
        // spool-sized file occupies beyond its bytes on THIS data set (XFS doubles it; tmpfs and
        // ext4 add nothing), so every push reserves what it will really hold while it lands.
        let (fs_slack, unmeasured) = gunnar_ssh::space::FsSlack::of(stores.root());
        match &unmeasured {
            None => tracing::info!(root = %stores.root().display(), measured_permille = ?fs_slack.measured_permille, slack_permille = fs_slack.slack_permille, "{}", fs_slack.words(stores.root(), None)),
            Some(why) => tracing::warn!(root = %stores.root().display(), slack_permille = fs_slack.slack_permille, "{}", fs_slack.words(stores.root(), Some(why))),
        }
        let ledger = match &config.space_probe {
            None => gunnar_ssh::SpaceLedger::new(stores.root(), config.disk_floor),
            Some(probe) => gunnar_ssh::SpaceLedger::with_probe(stores.root(), config.disk_floor, Arc::clone(probe)),
        };
        let disk_gate = Arc::new(gunnar_ssh::DiskHeadroomGate::over(Arc::new(ledger.with_fs_slack(fs_slack))));
        match disk_gate.headroom() {
            Ok(room) => tracing::info!(
                root = %stores.root().display(),
                free = %gunnar_ssh::headroom::human(room.free),
                floor = %gunnar_ssh::headroom::human(room.floor),
                disk_floor = %config.disk_floor,
                has_room = room.has_room(),
                "disk headroom: pushes and Repos.Create are refused under the floor"
            ),
            Err(err) => tracing::error!(
                root = %stores.root().display(),
                %err,
                "disk headroom could not be measured; every push and Repos.Create will be \
                 refused until it can be"
            ),
        }
        // The front doors' control: the gate CHAIN, disk first, then the plan.
        let front_control = || {
            front_control(
                &stores,
                config.anonymous.clone(),
                Arc::clone(&disk_gate),
                #[cfg(feature = "metered")]
                entitlement.as_ref(),
            )
        };
        #[cfg(feature = "http")]
        let http_lfs: Arc<dyn gunnar_ssh::seam::LfsTransport> = Arc::clone(&transport) as _;
        // **Repository settings over the forge's door** (owner, 2026-10-14):
        // `gunnar-repo …` on this listener, answered by the control plane's own
        // `Repos` handlers once the control plane exists (set further down);
        // until then — or on `--no-grpc` — the verb says it is not answered here.
        let repo_door_ctx: Arc<std::sync::OnceLock<gunnar_grpc::service::ServiceContext>> =
            Arc::new(std::sync::OnceLock::new());
        let ssh = GunnarSsh::new(ssh_config, front_control(), transport)
            // The boot window (`crate::readiness`): while the start-up disk
            // test is saturating /data, a git transfer is refused with
            // `starting` and EX_TEMPFAIL rather than served an open that
            // fails and reads as a missing repository.
            .with_readiness(crate::readiness::flag())
            // The admission gate is the ONE door all three git fronts ask, so
            // it is where a refusal is charged to the tenant that met it. See
            // `GunnarSsh::with_flow`.
            .with_flow(Arc::clone(stores.flow()))
            .with_metrics(Arc::clone(&metrics));
        // **The sealed box's tunnel-address door** (T17): `gunnar-tunnel-address`
        // over this listener, server-wide Admin, answered by the SAME function
        // `Bootstrap.AssignTunnelAddress` and `Admin.SetTunnelAddress` call. The
        // listener already held the caller to server Admin, so `may_write`.
        #[cfg(feature = "wg-client")]
        let ssh = {
            let root = stores.root().to_path_buf();
            ssh.with_tunnel_address_door(Arc::new(move |address: &str| {
                crate::appliance::tunnel_address_answer(&root, address, true)
                    .map(|answer| crate::appliance::render_tunnel_answer(&answer))
                    .map_err(|err| format!("{err}"))
            }))
        };
        let ssh = {
            let cell = Arc::clone(&repo_door_ctx);
            ssh.with_repo_door(Arc::new(move |ask: gunnar_ssh::RepoAsk| {
                let cell = Arc::clone(&cell);
                Box::pin(async move {
                    let Some(ctx) = cell.get() else {
                        return Err("this server answers repository settings only on its control plane".to_owned());
                    };
                    let setting = match ask.action {
                        gunnar_ssh::RepoAction::Show => gunnar_grpc::repo_door::RepoSetting::Show,
                        gunnar_ssh::RepoAction::Visibility(public) => gunnar_grpc::repo_door::RepoSetting::Visibility(public),
                        gunnar_ssh::RepoAction::DefaultBranch(b) => gunnar_grpc::repo_door::RepoSetting::DefaultBranch(b),
                        gunnar_ssh::RepoAction::Create { sha256 } => gunnar_grpc::repo_door::RepoSetting::Create(sha256),
                    };
                    gunnar_grpc::repo_door::ask(ctx, &ask.principal, &ask.fingerprint, ask.repo.namespace(), ask.repo.name(), setting)
                        .await
                        .map(|repo| gunnar_grpc::repo_door::render(&repo))
                        .map_err(|status| status.message().to_owned())
                })
            }))
        };
        // **--wg-only** (owner 2026-10-04): a connection at the utility address
        // — where the control plane is bound — is refused before a byte; the
        // public's git clients and the tunnel are served as ever.
        let ssh = if config.wg_only && !config.no_grpc {
            let utility = resolved_grpc_bind.ip();
            tracing::warn!(%utility, "ssh: --wg-only — a connection at the utility address is refused; boxes speak over the tunnel");
            ssh.with_refused_local(Arc::new(move |local: std::net::IpAddr| local == utility))
        } else {
            ssh
        };
        let ssh = ssh
            .bind(russh_key)
            .await
            .with_context(|| format!("binding the SSH listener on {}", config.ssh_bind))?;
        // The anonymous setting is on the *bind* line rather than in a separate
        // one, so an operator reading a boot log sees what this listener will
        // serve strangers in the same place they see where it is listening.
        tracing::info!(
            addr = %ssh.local_addr(),
            anonymous_read = config.allow_anonymous,
            "gunnar ssh listener: git-upload-pack, git-receive-pack; publickey, plus \
             anonymous read of public repositories when enabled"
        );

        // ── the smart-HTTP listener ────────────────────────────────────────
        //
        // The SAME `Dispatch` the SSH listener authorises with, the SAME
        // `WireTransport`, the SAME counters. That is the whole point of the
        // crate: a second transport, not a second server. If this ever needed
        // its own control plane or its own transport, HTTP and SSH would be
        // able to disagree about who may push.
        #[cfg(feature = "http")]
        let http = match config.http_bind {
            None => None,
            Some(bind) => {
                // The TLS identity, resolved BEFORE binding: a `--http-tls-*`
                // that cannot produce one must be "the server refused to
                // start", never a listener that quietly came up cleartext —
                // that silent fallback would be the whole defect the flag
                // exists to prevent.
                let tls = match &config.http_tls {
                    None => None,
                    Some(HttpTls::Identity(identity)) => Some(identity.clone()),
                    Some(HttpTls::SelfMinted) => {
                        let pki = stores.pki().context(
                            "--http-tls-self needs the secret store; no PKI is available",
                        )?;
                        let issuer = pki
                            .issuer()
                            .context("reading the issuer for --http-tls-self")?
                            .context(
                                "--http-tls-self needs an issuer and none exists; the first-start \
                                 mint above must have failed — its warning says why",
                            )?;
                        // The names a client may dial this listener by. rcgen
                        // types each SAN as IP or DNS on its own. Loopback is
                        // always present so `git clone https://127.0.0.1/…`
                        // works out of the box; a concrete bind address joins
                        // it. An operator with real DNS names brings their own
                        // certificate — that is what `--http-tls-cert` is for.
                        let mut hosts = vec!["localhost".to_owned(), "127.0.0.1".to_owned()];
                        let ip = bind.ip();
                        if !ip.is_unspecified() && !ip.is_loopback() {
                            hosts.push(ip.to_string());
                        }
                        let issued = pki
                            .issue_server_identity(&hosts, gunnar_core::DEFAULT_LEAF_DAYS)
                            .context("minting the smart-HTTP server certificate")?;
                        let key_pem = issued.key_pem.clone().context(
                            "the issuer minted a server certificate but returned no key",
                        )?;
                        tracing::info!(
                            fingerprint = %issued.fingerprint,
                            sans = ?hosts,
                            "smart-HTTP TLS: server certificate minted from this server's own \
                             issuer; clients pin the issuer (`gunnar issuer show --pem`)"
                        );
                        Some(gunnar_http::TlsIdentity {
                            // Leaf first, issuer after, so a pinning client can
                            // build the chain from the handshake alone.
                            cert_chain_pem: format!("{}{}", issued.cert_pem, issuer.cert_pem),
                            key_pem,
                        })
                    }
                };
                let http_config = gunnar_http::HttpConfig {
                    listen: bind,
                    // One field, both listeners. See `ServeConfig::allow_anonymous`.
                    allow_anonymous: config.allow_anonymous,
                    cleartext: config.http_cleartext,
                    tls,
                    ..gunnar_http::HttpConfig::default()
                };
                let bound = gunnar_http::GunnarHttp::new(
                    http_config,
                    gunnar_ssh::server::Dispatch::new(front_control()),
                    Arc::clone(&http_transport),
                )
                .with_metrics(Arc::clone(&metrics))
                // ── ONE admission gate for BOTH front doors ─────────────────
                //
                // The SSH listener's own gate, handed in rather than rebuilt,
                // so a git transfer over smart HTTP counts against the same
                // server-wide limit, the same per-principal share and the same
                // memory budget as one over SSH. Before this the HTTP front had
                // no gate at all (measured: 64 clients, 64 slots, nothing
                // refused); two independent gates would have been the next
                // worst thing, doubling the appliance's real ceiling in
                // silence. This is the same-writer fix the crate's own comment
                // above demands: "a second transport, not a second server."
                .with_admission(ssh.admission_gate())
                .with_lfs(Arc::clone(&http_lfs))
                // The registry rides the same listener as git, on the same
                // credentials: one port, one identity table, one place an
                // operator opens in a firewall.
                .with_registry(Arc::new(crate::registry::StoreRegistry::new(Arc::clone(
                    &stores,
                ))))
                .bind()
                .await
                .with_context(|| format!("binding the HTTP listener on {bind}"))?;
                let addr = bound.local_addr();
                let tls_on = bound.tls();
                tracing::info!(
                    %addr,
                    anonymous_read = config.allow_anonymous,
                    cleartext_credentials = ?config.http_cleartext,
                    tls = tls_on,
                    "gunnar smart-http listener: info/refs, git-upload-pack, git-receive-pack, \
                     info/lfs/objects (batch, get, put, verify)"
                );
                // A32-14, stays a task: a listener; its handle is kept and it ends at the server's stop.
                Some((tokio::spawn(bound.serve()), addr, tls_on))
            }
        };

        // ── the scrape endpoint ────────────────────────────────────────────
        //
        // The SAME `Arc<ServerMetrics>` the heartbeat reads and both transports
        // write. Two readings of one set of counters, never two sets: a scrape
        // that disagreed with the heartbeat would leave an operator with no way
        // to decide which one to believe during the one hour they need either.
        //
        // On the GIT runtime and not the control plane's: the whole point of
        // this endpoint is that it still answers when the control plane is what
        // is wedged, and putting it on the pool that starves would have made it
        // useless in exactly that case.
        let scrape = match config.metrics_bind {
            None => None,
            Some(bind) => {
                let bound = crate::scrape::bind(bind, Arc::clone(&metrics))
                    .await
                    .with_context(|| format!("binding the metrics listener on {bind}"))?;
                let addr = bound.local_addr();
                // A32-14, stays a task: a listener; its handle is kept and it ends at the server's stop.
                Some((tokio::spawn(bound.serve()), addr))
            }
        };

        // ── the gRPC control plane, on a runtime of its own ────────────────
        // ── replication slot: built EMPTY, and built EARLY ─────────────────
        //
        // Before the control plane, because the control plane takes a clone of
        // it: `Mirror.SetReplication` has to work on a box that has never had a
        // target and holds no repositories yet, which is every box on its first
        // boot. And it captures THIS runtime — the git one — because a blocking
        // send-pack belongs here and not on the control plane an operator is
        // using while the box is in trouble.
        #[cfg(feature = "replication")]
        let replication = crate::mirror::ReplicationSlot::new(
            Arc::clone(&stores),
            Arc::clone(&bus),
            tokio::runtime::Handle::current(),
        );

        // ── the outbound webhook slot: same shape, same reason ─────────────
        //
        // Built empty and early, for the same reason the replication slot is:
        // the control plane takes a clone of it so `Webhooks.AddWebhook` works
        // on a box that has never had an endpoint. Its workers do async POSTs,
        // not blocking send-packs, but sharing the git runtime keeps them off
        // the control plane an operator uses while the box is in trouble.
        #[cfg(feature = "webhooks")]
        let webhook = crate::webhook::WebhookSlot::new(
            Arc::clone(&stores),
            Arc::clone(&bus),
            tokio::runtime::Handle::current(),
        );

        // The control plane's copy of the replication view, taken while the
        // plane is built and filled below once the replicator exists.
        let mut control_view: Option<crate::replication_view::ReplicationView> = None;
        #[cfg(feature = "replication")]
        let mut control_for_repair: Option<Arc<dyn gunnar_grpc::ControlPlane>> = None;
        // The SAME grow machine `Admin.GrowStart` drives: a pulled `grow_go_ahead` (G16-d).
        #[cfg(feature = "replication")]
        let mut grow_for_repair: Option<Arc<crate::grow::GrowMachine>> = None;
        let (grpc, grpc_shutdown, grpc_endpoint, control_runtime) = if config.no_grpc {
            tracing::warn!(
                "the gRPC control plane is disabled; repositories, keys and grants can only be \
                 administered from this host's command line"
            );
            (None, None, None, ControlRuntime::default())
        } else {
            let workers = config
                .control_plane_workers
                .unwrap_or_else(control_plane_workers);
            // **The calls on one runtime, the connections on another**
            // (R5.1-G). The handler runtime is held by a guard from the moment
            // it exists, so the `?` below cannot drop a bare runtime here (see
            // `ControlRuntime`).
            let mut handler_runtime = ControlRuntime::new(
                tokio::runtime::Builder::new_multi_thread()
                    .worker_threads(workers)
                    .thread_name("gunnar-control-call")
                    .enable_all()
                    .build()
                    .context("building the control plane's handler runtime")?,
            );
            let handlers = handler_runtime
                .handle()
                .context("the control plane's handler runtime was gone as soon as it was built")?;
            let control_runtime = tokio::runtime::Builder::new_multi_thread()
                .worker_threads(CONTROL_CONNECTION_WORKERS)
                .thread_name("gunnar-control")
                .enable_all()
                .build()
                .context("building the control plane's runtime")?;

            let grpc_bind = resolved_grpc_bind;
            let grpc_bind_device = config.grpc_bind_device.clone();
            let web_relying_parties = config
                .web_relying_parties
                .iter()
                .map(|party| party.parse::<gunnar_grpc::webauthn::RelyingParty>().map_err(|why| anyhow::anyhow!("--web-relying-party: {why}")))
                .collect::<Result<Vec<_>>>()?;
            let ssh_hint = ssh.local_addr().to_string();
            // **ONE twin heartbeat registry, shared by the gRPC plane that
            // RECORDS a twin's poll and the drain that READS what it applied.**
            //
            // It used to be created inside `ServiceContext` and owned only
            // there, so the grow world could not see it — and the drain proved
            // itself instead by pushing to the replication roster, which for a
            // twin can never succeed (a twin serves no git). Built here so both
            // halves hold the same `Arc`; two registries would be two answers
            // to "what has the twin applied", and the one in front of the
            // unmount would be the empty one.
            let heartbeats = Arc::new(gunnar_grpc::heartbeat::Registry::default());
            // ── the runtime data-set growth (`crate::grow`) ────────────────
            //
            // Over the ONE admission gate the fronts share (the drain pauses
            // it) and these stores (the identity handle, the close/reopen).
            // With `grow-runtime` the world is the box; without it the machine
            // still answers `GrowStatus` honestly and refuses `GrowStart` by
            // name.
            #[cfg(feature = "grow-runtime")]
            let grow_world: Arc<dyn crate::grow::World> = Arc::new(
                crate::grow::world::BoxWorld::new(ssh.admission_gate(), Arc::clone(&stores))
                    // So the drain can prove the standbys are caught up once
                    // the gate is shut and nothing is in flight — the only
                    // moment at which "everything is drained to the twin" is
                    // a true sentence. Without this the world falls back to
                    // `NoStandby` and a growth would proceed having proved
                    // nothing about a twin that exists.
                    .with_replication(Arc::clone(&replication))
                    // The twin's OWN reported position — the only thing on
                    // this box that says whether it holds the set.
                    .with_heartbeats(Arc::clone(&heartbeats)),
            );
            #[cfg(not(feature = "grow-runtime"))]
            let grow_world: Arc<dyn crate::grow::World> = Arc::new(crate::grow::world::NoEngine);
            // **The go-ahead policy is the metered half's, and only the metered
            // half's.** AUDITED 2026-09-18: this line read
            // `GrowMachine::new(grow_world)` on every build, which installs
            // `OperatorPolicy` — and that policy refuses every signer but
            // `operator` by name. monetize signs its go-ahead `signer:
            // "monetize"`, so its orchestrator reached `phase::STARTING`,
            // called `Admin.GrowStart` and was refused AFTER it had reserved
            // the spend: front-driven growth was dead at step 3 and had never
            // worked. `MonetizePolicy` verifies that signature under the key
            // the box already holds from `--monetize-pubkey` (the same one
            // `EntitlementSlot::open` above was given), and it takes an
            // `operator` go-ahead exactly as before. `cfg`, not a runtime
            // branch: a community binary has no monetize verifier in it at all.
            #[cfg(feature = "metered")]
            let grow = match &config.metered {
                Some(metered) => {
                    // Keyless when the key is fed and has not arrived: it is then
                    // applied here in place too (`feed::on_monetize_key`).
                    let policy = Arc::new(match &metered.monetize_pubkey {
                        Some(key) => crate::grow::MonetizePolicy::new(key).map_err(|e| anyhow::anyhow!("the growth go-ahead policy: {e}"))?,
                        None => crate::grow::MonetizePolicy::awaiting(),
                    });
                    let taker = Arc::clone(&policy);
                    crate::feed::on_monetize_key("growth go-ahead policy", Box::new(move |k| taker.set_key(k)));
                    crate::grow::GrowMachine::with_policy(grow_world, policy)
                }
                // A metered binary started without `--monetize-pubkey` serves
                // Free (the warning three hundred lines up); it has no key to
                // verify a `monetize` go-ahead with either, so it keeps the
                // operator door and refuses monetize by name rather than
                // pretending to a check it cannot make.
                None => crate::grow::GrowMachine::new(grow_world),
            };
            #[cfg(not(feature = "metered"))]
            let grow = crate::grow::GrowMachine::new(grow_world);
            #[cfg(feature = "replication")]
            {
                grow_for_repair = Some(Arc::clone(&grow));
            }
            let mut plane = CoreControlPlane::new(Arc::clone(&stores))
                // `Bootstrap.Restore`'s keys from the medium; the fed ones are read per ask.
                .with_restore_keys(restore_keys.clone())
                // `Repos.Create` asks the SAME gate the front doors ask (one implementation):
                // a disk that refuses a push refuses a new repository with the
                // same line.
                .with_headroom(Arc::clone(&disk_gate))
                .with_grow(grow)
                // The in-flight count `Mirror.FlushReplication` reports (T11).
                .with_admission(ssh.admission_gate())
                // The SAME bus the SSH front door publishes a push onto, so a
                // `MirrorSynced` and a `RefUpdated` reach one `Events.Watch`
                // stream. The plane publishes it from the mirror queue's worker
                // when a sync completes — `Mirror.Sync` itself only enqueues,
                // and an event announcing a fetch that has not happened would
                // be worse than no event.
                .with_bus(Arc::clone(&bus))
                // `Bootstrap.Status` names the role before the join (a twin has no banner).
                .with_twin_role(config.twin_role)
                // **`serving_git` is a reading of THIS socket.** Both addresses,
                // so the status door can say what was asked for and what was
                // taken: on a twin those differ, because the branch above binds
                // loopback:0 whatever `--ssh-bind` named, and a box that reports
                // `git serving` while its :2222 sends an RST is a health line
                // that cannot fail.
                .with_git_listener(crate::readiness::GitListener {
                    asked: config.ssh_bind,
                    bound: ssh.local_addr(),
                });
            // The plane publishes into the SAME view the band reads, so
            // `Admin.Stats` and the console can never disagree about whether a
            // standby is taking bytes (one implementation).
            // The WRITE side. `replication_view` just below is the READ side
            // one slot, two halves, and neither is the other's twin —
            // a `SetReplication` changes what is running, a `Stats` reads what
            // it did.
            #[cfg(feature = "replication")]
            {
                plane = plane.with_replication(Arc::clone(&replication));
            }
            #[cfg(feature = "webhooks")]
            {
                plane = plane.with_webhooks(Arc::clone(&webhook));
            }
            #[cfg(feature = "metered")]
            if let Some(slot) = &entitlement {
                plane = plane.with_entitlement(Arc::clone(slot));
            }
            let plane_view = plane.replication_view();
            if let Some(bytes) = config.max_object_bytes {
                plane = plane.with_max_object_bytes(bytes);
            }
            // **An import the last process did not live to finish is settled and
            // resumed** (import-big, 2026-10-26), on the control runtime the
            // mirror queue's lane lives on. A twin originates no sync.
            if !config.twin_role {
                let _on = control_runtime.enter();
                let resumed = plane.resume_interrupted_mirrors();
                if resumed > 0 {
                    tracing::info!(target: "gunnar::git", resumed, "mirror syncs interrupted by the last stop were resumed");
                }
            }
            // The trust root, resolved BEFORE the listener is spawned, for the
            // reason `--http-tls-*` is resolved before its bind: a control
            // plane that was asked for client certificates and could not mint
            // its own identity must be a startup error naming the cause, never
            // a listener that came up accepting anyone.
            let grpc_trust = match config.grpc_tls {
                None => None,
                Some(GrpcTls::SelfMinted) => {
                    let pki = stores
                        .pki()
                        .context("--grpc-tls-self needs the secret store; no PKI is available")?;
                    let issuer = pki
                        .issuer()
                        .context("reading the issuer for --grpc-tls-self")?
                        .context(
                            "--grpc-tls-self needs an issuer and none exists; the first-start \
                             mint above must have failed — its warning says why",
                        )?;
                    // The same SAN reasoning as the HTTP listener's: loopback
                    // always, plus a concrete bind address when there is one.
                    let mut hosts = vec!["localhost".to_owned(), "127.0.0.1".to_owned()];
                    let ip = grpc_bind.ip();
                    if !ip.is_unspecified() && !ip.is_loopback() {
                        hosts.push(ip.to_string());
                    }
                    let issued = pki
                        .issue_server_identity(&hosts, gunnar_core::DEFAULT_LEAF_DAYS)
                        .context("minting the gRPC server certificate")?;
                    let key_pem = issued
                        .key_pem
                        .clone()
                        .context("the issuer minted a server certificate but returned no key")?;
                    tracing::info!(
                        fingerprint = %issued.fingerprint,
                        sans = ?hosts,
                        "gRPC control plane: mutual TLS on. Clients must present a certificate \
                         this server's issuer signed AND registered (`gunnar control bootstrap`, \
                         `gunnar key add`)"
                    );
                    Some(gunnar_grpc::Trust {
                        // The issuer is both halves here: the CA every client
                        // certificate must chain to, and the second link of
                        // the chain this listener presents. Beside it, the
                        // CENTRAL monetize's attach CA when this estate
                        // attached from another cloud (f-grpc-tls): the
                        // central dials in with a client identity under it.
                        #[cfg(feature = "wg-client")]
                        client_ca: crate::fed_principals::grpc_client_cas(&issuer.cert_pem, stores.root()).into_bytes(),
                        // No fed attach CA without the installer's feed (wg-client).
                        #[cfg(not(feature = "wg-client"))]
                        client_ca: issuer.cert_pem.clone().into_bytes(),
                        cert: format!("{}{}", issued.cert_pem, issuer.cert_pem).into_bytes(),
                        key: key_pem.into_bytes(),
                    })
                }
            };
            // Taken from the trust root itself rather than from
            // `config.grpc_tls`, so the field `Admin.Health` reports is the
            // one the listener is actually built with.
            let grpc_mutual_tls = grpc_trust.is_some();

            // ── the MUTUAL REPLICATION LISTENER (lane T9) ──────────────────────
            //
            // The door a twin follows this box through. Resolved HERE, before
            // anything is spawned, for the reason the main trust root is: a box
            // that should serve its twin and cannot mint the identity to do it
            // must fail to start NAMING the cause, not come up with a twin that
            // retries into nothing for ever. `grpc_replication_bind` unset derives
            // it from the main bind (off loopback ⇒ port + 1), which is what lets
            // the one medium carry no flag.
            let replication_root = stores.root().to_path_buf();
            let replication = {
                // The ONE derivation (`client_signer::replication_address`), so the
                // operator side that dials it cannot drift from what is bound here.
                let derived = (!grpc_bind.ip().is_loopback() && grpc_bind.port() != 0)
                    .then(|| gunnar_grpc::client_signer::replication_address(&grpc_bind.to_string()).ok())
                    .flatten()
                    .and_then(|at| at.parse::<SocketAddr>().ok());
                // **Where it binds** (owner 2026-09-22, amended 2026-09-24): on the
                // derived (eth1) address for the process lifetime, and ALSO on
                // `wg0`'s address while the tunnel is up (`replication_door`); an
                // explicit `--grpc-replication-bind` is never widened.
                let explicit = config.grpc_replication_bind.is_some();
                match config.grpc_replication_bind.or(derived) {
                    None => None,
                    Some(base) => {
                        let bind = base;
                        let pki = stores.pki().with_context(|| {
                            format!("the mutual replication listener on {bind} needs the secret store; no PKI is available")
                        })?;
                        let issuer = pki.issuer().context("reading the issuer for the replication listener")?.with_context(|| {
                            format!(
                                "the mutual replication listener on {bind} needs this box's issuer and there is \
                                 none — the first-start mint above must have failed, and its warning says why. \
                                 Without it no twin of this box can ever follow it."
                            )
                        })?;
                        let mut hosts = vec![
                            gunnar_grpc::client_signer::REPLICATION_SERVER_NAME.to_owned(),
                            "localhost".to_owned(),
                            "127.0.0.1".to_owned(),
                        ];
                        for ip in [bind.ip(), base.ip()] {
                            if !ip.is_unspecified() && !ip.is_loopback() && !hosts.contains(&ip.to_string()) {
                                hosts.push(ip.to_string());
                            }
                        }
                        // Minted at EVERY start, so a restart is a renewal and a
                        // lapse needs a year of uninterrupted uptime.
                        let issued = pki
                            .issue_server_identity(&hosts, gunnar_core::DEFAULT_LEAF_DAYS)
                            .context("minting the replication listener's certificate")?;
                        let key_pem = issued
                            .key_pem
                            .clone()
                            .context("the issuer minted the replication certificate but returned no key")?;
                        Some((
                            base,
                            explicit,
                            bind,
                            gunnar_grpc::Trust {
                                client_ca: issuer.cert_pem.clone().into_bytes(),
                                cert: format!("{}{}", issued.cert_pem, issuer.cert_pem).into_bytes(),
                                key: key_pem.into_bytes(),
                            },
                        ))
                    }
                }
            };

            let control = Arc::new(plane);
            control_view = Some(plane_view);
            #[cfg(feature = "replication")]
            {
                control_for_repair = Some(control.clone());
            }
            let bus_for_grpc = Arc::clone(&bus);
            // ── ONE admission gate on ALL THREE front doors ────────────────
            //
            // `Repos.ArchiveOut` streams a whole repository's stored bytes down
            // the control-plane channel — a transfer in everything but name — so
            // it takes a permit from the SAME `Arc<Admission>` the SSH and HTTP
            // git fronts hold (`ssh.admission_gate()`), not a second gate (one
            // implementation). The *cheap* control RPCs get no gate on purpose: the control
            // plane runs on its own runtime, so their walks starve
            // nothing but themselves. Captured here and moved into the control
            // runtime's spawn below.
            let grpc_transfer_gate =
                crate::grpc_gate::control_transfer_gate(ssh.admission_gate());
            let browse_flag = stores.browse();
            // ── the durable audit trail (GUN-044) ──────────────────────────
            //
            // Opened HERE, on the startup path, because `AuditChain::open`
            // verifies the existing chain and an operator has to be told about
            // a broken one at the moment the server starts — not the first time
            // somebody runs a verification months later.
            //
            // A chain that will not open does NOT stop the server. The control
            // plane still authenticates, still authorises and still writes the
            // `tracing` record it always did; what it loses is the durable
            // copy, and it says so at `error!` with the reason. Refusing to
            // serve would turn a full disk — or a chain somebody has already
            // tampered with — into an outage, which hands whoever tampered
            // with it exactly what they wanted.
            let audit_ledger = {
                // **Off the stripe.** The trail records who changed what on
                // this box, which is exactly the question asked the morning
                // after the four striped data disks have gone — so a copy that
                // dies with them answers nobody. `ServerStores::audit_path`
                // adopts an existing chain from the control directory once and
                // falls back to it unchanged if the slow disk will not take it.
                let path = stores.audit_path();
                // ── and it must SURVIVE A RESTORE (lane T11, 2026-09-21) ───
                //
                // D6: `log/` was a per-box path, so the twin never held the
                // trail and `twin replay` brought a box back with no audit
                // history at all — silently, reporting success. A system whose
                // audit trail is destroyed by its own recovery procedure
                // cannot answer the only question that will be asked.
                //
                // So the primary's chain is opened FRAMED: every entry is an
                // `Append` transaction, the twin holds it, and a restored box
                // continues the dead primary's chain from its own tip.
                //
                // A twin — and a box mid-restore — originates nothing, so it
                // gets an ordinary chain at its OWN per-box path
                // (`audit_path` decides the name), and this branch does not
                // even try to frame it.
                let originates = !stores.is_twin() && !crate::twin::is_restoring(stores.root());
                let opened = if originates {
                    match gunnar_core::AuditChain::open_framed(&path) {
                        Ok(chain) => Ok(chain),
                        // **Never silent.** An unframable chain is the D6 bug
                        // coming back, so it is said in the words an operator
                        // needs, and the server still writes a local trail.
                        Err(err @ gunnar_core::AuditError::Unframed { .. }) => {
                            tracing::error!(
                                target: gunnar_grpc::AUDIT_TARGET,
                                path = %path.display(),
                                error = %err,
                                "this box's audit trail is NOT replicated: it will be written \
                                 here and it will NOT survive a restore from the twin. Every \
                                 other class of state comes back; this one will not."
                            );
                            gunnar_core::AuditChain::open(&path)
                        }
                        Err(err) => Err(err),
                    }
                } else {
                    gunnar_core::AuditChain::open(&path)
                };
                match opened {
                    Ok(chain) => Some(crate::grpc_gate::control_audit_ledger(Arc::new(chain))),
                    Err(err) => {
                        tracing::error!(
                            target: gunnar_grpc::AUDIT_TARGET,
                            path = %path.display(),
                            error = %err,
                            "the control-plane audit chain could not be opened; this server \
                             will log its decisions and will NOT write them to a trail. If the \
                             chain is reported BROKEN, the file has been edited or truncated \
                             since gunnar wrote it."
                        );
                        None
                    }
                }
            };
            let (tx, rx) = oneshot::channel();
            // The listener is bound INSIDE the control runtime. A tokio socket
            // registers with the I/O driver of the runtime that created it, so
            // one bound out here would keep the git runtime's driver in the
            // control plane's critical path, which is the exact coupling this
            // change exists to remove.
            let (bound_tx, bound_rx) = oneshot::channel::<Result<(SocketAddr, bool)>>();
            // The box's jera ledger (RA3-5): the job queue's jobs and the long duties are rows in
            // `<data>/jobs`, read by `nornir jobs`; one that cannot open is said, and they run unrecorded.
            if let Err(e) = gunnar_common::jobs::set_ledger(&stores.root().join("jobs")) {
                tracing::error!("ALERT jobs.ledger {e}: the jobs run, unrecorded");
            }
            let sweep_stores = Arc::clone(&stores);
            let activity_stores = Arc::clone(&stores);
            let wg_only = config.wg_only;
            // A32-14, stays a task: the control plane on its own runtime; its handle is kept and joined at the stop.
            let join = control_runtime.spawn(async move {
                // Bound WITH PATIENCE: a control plane on a tailnet or VPN
                // address is asked for before that interface exists on a cold
                // boot (builder, every boot until 2026-09-10: six `EADDRNOTAVAIL`
                // exits and six blind systemd restarts before the seventh try
                // found the address). The wait is bounded and only for that
                // one errno; everything else still fails at once.
                let listener = match bind_when_the_address_arrives(
                    || bind_grpc_listener(grpc_bind, grpc_bind_device.as_deref()),
                    GRPC_BIND_PATIENCE,
                    GRPC_BIND_RETRY_EVERY,
                )
                .await
                .0
                .with_context(|| match &grpc_bind_device {
                    Some(dev) => format!("binding the gRPC listener on {grpc_bind} (device {dev})"),
                    None => format!("binding the gRPC listener on {grpc_bind}"),
                }) {
                    Ok(listener) => listener,
                    Err(err) => {
                        // Report the bind failure to `start`, which is the only
                        // place that can turn it into a startup error. Swallowing
                        // it here would leave `gunnar serve` announcing a control
                        // plane that never bound.
                        let _ = bound_tx.send(Err(err));
                        return Ok(());
                    }
                };
                let addr = match listener
                    .local_addr()
                    .context("reading the gRPC local address")
                {
                    Ok(addr) => addr,
                    Err(err) => {
                        let _ = bound_tx.send(Err(err));
                        return Ok(());
                    }
                };
                let server_config = ServerConfig::default()
                    .with_bind(addr)
                    .with_ssh_bind_hint(ssh_hint)
                    .with_web_relying_parties(web_relying_parties)
                    .with_web_user_verification(if config.web_require_user_verification {
                        gunnar_grpc::config::UserVerification::Required
                    } else {
                        gunnar_grpc::config::UserVerification::Preferred
                    })
                    // Reported from the SAME value that decides whether the
                    // listener is built with a trust root, three lines below
                    // (`router_with_trust(…, grpc_trust.as_ref())`). Derived
                    // from anything else it would be a second opinion about
                    // whether the door is locked, and `Admin.Health`'s job
                    // here is to be the answer rather than to remember one.
                    .with_mutual_tls(grpc_mutual_tls);
                // Every start, on purpose: a hole nobody is reminded of is a hole
                // nobody fixes.
                gunnar_grpc::warn_authentication_disabled(&server_config);
                // Reported out rather than re-derived by the caller. `serve`
                // prints this on stdout, and a second reader of "is auth on"
                // is a second thing that can be wrong: the line it replaced
                // said `UNAUTHENTICATED in v1` unconditionally, and had
                // gone stale against a control plane that authenticates.
                let authenticated = server_config.authentication_enabled();
                let ctx = ServiceContext::new(control, bus_for_grpc, server_config)
                    // The one "this estate sells room" flag the push gate reads too.
                    .with_sells_room(Arc::clone(&sells_room))
                    // Every call on the handler runtime, the connections on
                    // this one (R5.1-G): no slow call can hold the worker that
                    // answers a client's keepalive PING.
                    .with_handler_runtime(handlers)
                    // The transfer-shaped RPC's permit, from the git
                    // fronts' own gate.
                    .with_transfer_gate(grpc_transfer_gate)
                    // The browse replica: the browse flag the mutator refusal reads —
                    // the same `Arc` the replication slot flips.
                    .with_browse_flag(browse_flag)
                    // The SAME registry the drain reads (see where it is built).
                    .with_heartbeats(Arc::clone(&heartbeats));
                // The hash-chained trail, when there is one; `tracing` alone
                // when there is not. Either way the gate's decisions also
                // stamp who last changed something (2026-10-21, the admin's
                // user list): `grpc_gate::activity_ledger`.
                let ctx = ctx.with_audit_ledger(crate::grpc_gate::activity_ledger(
                    audit_ledger,
                    Arc::clone(activity_stores.activity()),
                ));
                // The forge door's `gunnar-repo` verb answers from here on.
                let _ = repo_door_ctx.set(ctx.clone());
                // **Company SSO's network**: the one way gunnar reaches an
                // account's identity provider. `GUNNAR_SSO_ALLOW_LOOPBACK_HTTP=1`
                // is the operator's test door (an IdP on this box's loopback,
                // over http) and nothing else.
                let sso_loopback = std::env::var("GUNNAR_SSO_ALLOW_LOOPBACK_HTTP").is_ok_and(|v| v.trim() == "1");
                if sso_loopback {
                    tracing::warn!("GUNNAR_SSO_ALLOW_LOOPBACK_HTTP=1: an identity provider on this box's loopback is reachable over http (a test door)");
                }
                ctx.sso.set_allow_loopback_http(sso_loopback);
                // A32-18: the egress is `sso-egress`'s (every appliance row names it); a build
                // without it sets none, so an identity provider's call is refused by name.
                #[cfg(feature = "sso-egress")]
                ctx.sso.set_egress(std::sync::Arc::new(crate::sso_egress::UreqEgress::with_nets(
                    sso_loopback,
                    crate::sso_egress::system_lookup(),
                    ctx.sso.private_nets_handle(),
                )));
                #[cfg(not(feature = "sso-egress"))]
                tracing::info!("company SSO has no way out in this build (gunnar-server's `sso-egress` is off): an identity provider's call is refused by name");
                // **The operator's internal networks a company IdP may live
                // on** (owner 2026-10-06): the installer's feed, applied now and
                // on every feed after (LIVE); a list gunnar refuses is said and
                // applied as none — today's guard.
                {
                    let sso = std::sync::Arc::clone(&ctx.sso);
                    nordisk_pairing::feedbus::on_published(
                        "sso: the operator's identity-provider networks",
                        move |feed| {
                            let nets = match crate::feed::sso_private_nets(Some(&crate::feed::Fed::from_feed(feed))) {
                                Ok(nets) => nets,
                                Err(why) => {
                                    tracing::warn!(target: "gunnar::audit", event = "sso.private_nets_refused", reason = %why, "the fed identity-provider networks were refused; none are admitted");
                                    gunnar_grpc::sso::PrivateNets::default()
                                }
                            };
                            if nets != sso.private_nets() {
                                tracing::info!(target: "gunnar::audit", event = "sso.private_nets", nets = %nets, "identity providers on the operator's internal networks: {}", if nets.is_empty() { "none (public addresses only)".to_owned() } else { nets.to_string() });
                                sso.set_private_nets(nets);
                            }
                        },
                    );
                }
                // **The transaction log's default retention** (h-txlog-2x): a PRIMARY
                // retains its log behind its slowest REPORTING twin, minus a window, every
                // ten minutes — never with no twin, never past a silent or unheard one, and
                // never on a twin (its log is its only copy; the pass says so and does
                // nothing). `GUNNAR_TXLOG_RETAIN=off` turns it off.
                // A32-14, stays a task: a supervised async loop (a panicked pass is said) on the server's runtime; it ends with it.
                tokio::spawn(gunnar_grpc::retention::run(ctx.clone()));
                // **The account audit logs' retention** (2026-10-05): a year by
                // default, swept every six hours on a primary.
                // A jera job (A32-14) that ends with the stores; one that cannot start is said.
                if let Err(e) = crate::account_audit_sweep::start(&sweep_stores, crate::account_audit_sweep::EVERY) {
                    tracing::error!(target: "gunnar::audit", "ALERT {} no sweep ({e}): the account audit logs are not pruned on this run", crate::account_audit_sweep::SWEEP_JOB);
                }
                // The replication listener: bound BEFORE the address is reported,
                // so a port that will not bind is a startup error and not a twin
                // that retries into a closed door. The same context — the same
                // gate, grants and audit trail — on a door that demands a
                // certificate at the handshake.
                // One stop for the replication door, whichever address it is on
                // (it may rebind when the tunnel comes up at runtime).
                let (replication_stop_tx, replication_stop_rx) = tokio::sync::watch::channel(false);
                match replication {
                    Some((base, explicit, bind, trust)) => {
                        let listener = match tokio::net::TcpListener::bind(bind).await {
                            Ok(listener) => listener,
                            Err(err) => {
                                let _ = bound_tx.send(Err(anyhow::Error::new(err).context(format!(
                                    "binding the mutual replication listener on {bind} (the door this box's twin follows it through)"
                                ))));
                                return Ok(());
                            }
                        };
                        let bound = listener.local_addr().unwrap_or(base);
                        let recorded = if wg_only { record_replication_door_wg_only(&replication_root, bound.port(), None) } else { record_replication_door(&replication_root, bound, None) };
                        if let Err(err) = recorded {
                            let _ = bound_tx.send(Err(err));
                            return Ok(());
                        }
                        // Built once up front so a broken trust is a STARTUP error;
                        // every rebind builds its own from the same trust.
                        if let Err(err) = gunnar_grpc::router_with_trust(ctx.clone(), grpc_request_id, Some(&trust)) {
                            let _ = bound_tx.send(Err(err.context("configuring the mutual replication listener")));
                            return Ok(());
                        }
                        tracing::info!(
                            %bind, suite = gunnar_grpc::preferred_bulk_suite(),
                            "mutual replication listener up: TLS 1.3, a client certificate from this box's issuer required"
                        );
                        // A32-14, stays a task: the replication listener's accept loop, ended by the server's stop.
                        tokio::spawn(replication_door(
                            listener,
                            base,
                            explicit,
                            wg_only,
                            ctx.clone(),
                            grpc_request_id,
                            trust,
                            replication_root.clone(),
                            replication_stop_rx,
                        ));
                    }
                    None => {
                        if let Err(err) = crate::twin_tls::record_replication_port(&replication_root, None) {
                            let _ = bound_tx.send(Err(err));
                            return Ok(());
                        }
                    }
                }
                if bound_tx.send(Ok((addr, authenticated))).is_err() {
                    // `start` gave up on us; do not leave a listener running.
                    return Ok(());
                }
                if wg_only {
                    // **--wg-only** (owner 2026-10-04): the bound — utility —
                    // listener serves nothing; the control plane answers on the
                    // tunnel's address only, following it (`control_doors`).
                    gunnar_grpc::router_with_trust(ctx.clone(), grpc_request_id, grpc_trust.as_ref())
                        .context("configuring mutual TLS on the gRPC control plane")?;
                    let (control_stop_tx, control_stop_rx) = tokio::sync::watch::channel(false);
                    let serve_control = move |l: tokio::net::TcpListener, mut stop: tokio::sync::watch::Receiver<bool>| {
                        let router = gunnar_grpc::router_with_trust(ctx.clone(), grpc_request_id, grpc_trust.as_ref());
                        async move {
                            let router = match router {
                                Ok(r) => r,
                                Err(err) => {
                                    tracing::error!(%err, "the control plane's tunnel door could not be configured");
                                    return;
                                }
                            };
                            let at = l.local_addr().map(|a| a.to_string()).unwrap_or_default();
                            if let Err(err) = router
                                .serve_with_incoming_shutdown(tokio_stream::wrappers::TcpListenerStream::new(l), async move {
                                    let _ = stop.wait_for(|s| *s).await;
                                })
                                .await
                            {
                                tracing::error!(%err, %at, "a control plane door stopped");
                            }
                        }
                    };
                    // A32-14, stays a task: the control doors' accept loop; its handle is kept and it stops at the stop signal.
                    let doors = tokio::spawn(control_doors(listener, addr, true, crate::tunnel_addr::subscribe(), control_stop_rx, serve_control));
                    let _ = rx.await;
                    // One stop for every door.
                    let _ = replication_stop_tx.send(true);
                    let _ = control_stop_tx.send(true);
                    let _ = doors.await;
                    Ok(())
                } else {
                    gunnar_grpc::router_with_trust(ctx, grpc_request_id, grpc_trust.as_ref())
                        .context("configuring mutual TLS on the gRPC control plane")?
                        .serve_with_incoming_shutdown(
                            tokio_stream::wrappers::TcpListenerStream::new(listener),
                            async {
                                let _ = rx.await;
                                // One stop for both doors.
                                let _ = replication_stop_tx.send(true);
                            },
                        )
                        .await
                        .context("the gRPC control plane stopped")
                }
            });

            let (addr, grpc_authenticated) = match bound_rx.await {
                Ok(Ok(pair)) => pair,
                Ok(Err(err)) => {
                    control_runtime.shutdown_background();
                    return Err(err);
                }
                Err(_) => {
                    control_runtime.shutdown_background();
                    bail!(
                        "the gRPC control plane task ended before it reported an address; the \
                         listener on {grpc_bind} never came up"
                    );
                }
            };
            tracing::info!(
                %addr,
                control_plane_workers = workers,
                control_connection_workers = CONTROL_CONNECTION_WORKERS,
                "gunnar grpc control plane listening on its own runtime (a control-plane \
                 stall cannot reach the git listener, and its calls run on a handler runtime \
                 that cannot hold the connections)"
            );
            let mut held = ControlRuntime::new(control_runtime);
            held.hold(&mut handler_runtime);
            (
                Some(join),
                Some(tx),
                Some((addr, grpc_authenticated)),
                held,
            )
        };

        // ── the log level, changeable from here on ─────────────────────────
        //
        // Started here rather than in `main` because it needs a runtime, and
        // deliberately NOT held on `Server`: the task owns two signal streams
        // and nothing else, and a level-control task that stopped with the
        // server would be one more thing to get wrong at shutdown.
        //
        // It also has to exist for a NEGATIVE reason. The default disposition
        // of both `SIGUSR1` and `SIGUSR2` is to TERMINATE the process, so
        // before this an operator who reached for the usual signal killed the
        // server they were debugging.
        gunnar_common::logging::spawn_signal_control();

        // ── the counter heartbeat ──────────────────────────────────────────
        let heartbeat = config.metrics_interval.map(|interval| {
            let metrics = Arc::clone(&metrics);
            // A32-14, stays a task: its handle is kept and aborted at the server's stop.
            tokio::spawn(async move { heartbeat(metrics, interval).await })
        });

        // ── the runner, on the same bus every push publishes on ────────────
        //
        // Subscribed **before** the listeners bind, so a push that lands in the
        // first millisecond is seen. Subscribing after would lose it silently,
        // which is the shape of bug that shows up once a month and is blamed on
        // the network.
        //
        // Its own OS thread rather than a tokio task: a build is a blocking
        // child process that can run for minutes, and putting it on a runtime
        // worker would starve whatever else that runtime is for.
        let runner = match config.runner.clone() {
            None => None,
            Some(runner_config) => {
                if let Err(err) = crate::sandbox::available() {
                    if !runner_config.allow_network {
                        bail!(
                            "the runner was asked for hermetic builds and this kernel cannot \
                             give it a sandbox: {}\n\n\
                             Refusing to start rather than building with the network up and \
                             signing records that say otherwise. Pass --runner-allow-network to \
                             build anyway; every record will say the network was up.",
                            err.explain()
                        );
                    }
                }
                let stores = Arc::clone(&stores);
                let bus = Arc::clone(&bus);
                tracing::info!(
                    workspace = %runner_config.workspace.display(),
                    builder = %runner_config.builder,
                    hermetic = !runner_config.allow_network,
                    "the build runner is watching this server's pushes"
                );
                // A28 M18, stays a thread: the build runner's watch for the server's life, its handle
                // held by the server (`_runner`). A watch that ends is said (`runner::said_end`).
                Some(std::thread::spawn(move || {
                    let end = crate::runner::watch(&stores, &bus, &runner_config, |outcome| {
                        match outcome {
                            crate::runner::Watched::Built(built) => tracing::info!(
                                target: "gunnar::runner",
                                event = "runner.built",
                                repo = %built.record.source_repo,
                                commit = %built.record.source_commit,
                                recipe = %built.record.recipe,
                                artifacts = built.record.subjects.len(),
                                "a push was built and attested"
                            ),
                            crate::runner::Watched::Failed { repo, commit, why } => {
                                // At `error`, and the watch continues: one
                                // repository's broken recipe must not stop
                                // every other repository's builds.
                                tracing::error!(
                                    target: "gunnar::runner",
                                    event = "runner.failed",
                                    %repo, %commit, %why,
                                    "a push could not be built; the runner is still watching"
                                )
                            }
                            crate::runner::Watched::Skipped { repo, r#ref, why } => {
                                tracing::debug!(
                                    target: "gunnar::runner",
                                    event = "runner.skipped",
                                    %repo, r#ref = %r#ref, %why,
                                    "a ref moved and was not built"
                                )
                            }
                        }
                    });
                    match crate::runner::said_end(&end) {
                        (tracing::Level::ERROR, words) => tracing::error!(target: "gunnar::runner", event = "runner.stopped", "{words}"),
                        (_, words) => tracing::warn!(target: "gunnar::runner", event = "runner.stopped", "{words}"),
                    }
                }))
            }
        };
        // Held so the thread is not detached from the server's lifetime; the
        // watch ends when the bus's last sender goes.
        let _runner = runner;

        // ── replication, a SUBSCRIBER and nothing else ─────────────────────
        //
        // Started last, and started on the git runtime deliberately: the work
        // it does is a git push over SSH, and the blocking part of it goes to
        // the blocking pool. Nothing above this line knows it exists, which is
        // the property that makes "replication can never fail a push" true by
        // construction rather than by care.
        #[cfg(feature = "replication")]
        {
            use crate::mirror::ReplicationConfig;
            let replication_config = match config.replication.clone() {
                Some(explicit) => explicit,
                None => ReplicationConfig::load(stores.root())
                    .context("reading the replication configuration")?,
            };
            // An empty target list leaves the slot empty, which is still replication's
            // "a server that replicates nowhere pays nothing": no replicator, no
            // subscriber, no task. What it no longer means is that the operator
            // has no way to add one.
            replication
                .apply(replication_config)
                .context("starting replication")?;
        }

        // ── I11's passive heartbeat, if this box was asked for one ─────────
        //
        // Started AFTER the slot has its configuration, because the task's
        // first act is to read the role out of it: a watch spawned before
        // `apply` would read the pre-boot file and could leave immediately on a
        // box that is about to become a standby.
        //
        // On the GIT runtime, with the replicator, and for the same reason: the
        // probe is a blocking-shaped network wait, and the control plane is the
        // surface an operator uses while a box is in trouble.
        #[cfg(feature = "replication")]
        let standby_watch = config.standby_watch.map(|watch| {
            tracing::info!(
                target: "gunnar::replication",
                primary = %watch.primary,
                probe_interval_ms = watch.probe_interval.as_millis() as u64,
                promote_after = watch.promote_after,
                earliest_takeover_ms = watch.earliest_takeover().as_millis() as u64,
                "watching the primary; this box takes over by itself if it stops answering"
            );
            let slot = Arc::clone(&replication);
            // A32-14, stays a task: its handle is kept and aborted at the server's stop.
            tokio::spawn(async move { crate::standby_watch::watch(slot, watch).await })
        });

        // ── 5.4: the boot check, then the pull channel ─────────────────────
        //
        // After the control plane listens: the boot check reads this box's own
        // `Bootstrap.Status`, and a recharge into this box goes through it.
        #[cfg(feature = "replication")]
        let repair = match (config.repair, grpc_endpoint) {
            (Some(cfg), Some((bound, _))) => {
                let own = crate::repair::own_addr(bound);
                let mut entitlements = crate::repair::Entitlements::of(Arc::clone(&stores));
                if let Some(control) = control_for_repair.clone() {
                    entitlements = entitlements.with_control(control);
                }
                #[cfg(feature = "metered")]
                let mut entitlements = entitlements.with_slot(entitlement.clone());
                if let Some(grow) = grow_for_repair.clone() {
                    entitlements = entitlements.with_grow(crate::paid_growth::PaidGrowth::new(grow));
                }
                // The environment's relay; else (a SOLD estate's main, P1) the one the shop FED it,
                // read at each mail — not ready until the feed names it, so a resend is refused
                // by name before anything is minted.
                #[cfg(feature = "fleet-pull")]
                let postman: Option<Arc<dyn crate::postman::Postman>> = match crate::postman::Smtp::from_env() {
                    Ok(Some(smtp)) => Some(Arc::new(smtp)),
                    Ok(None) => Some(Arc::new(crate::postman::FedRelay { fed: crate::feed::Fed::current })),
                    Err(why) => {
                        tracing::warn!(target: crate::repair::TARGET, "no mail relay: a sign_in_again order is refused by name ({why})");
                        None
                    }
                };
                let entitlements = entitlements.with_twin(cfg.twin.clone());
                #[cfg(feature = "fleet-pull")]
                let entitlements = entitlements.with_postman(postman);
                let stores = Arc::clone(&stores);
                crate::repair::arm(cfg.clone(), own.clone(), tokio::runtime::Handle::current());
                // A32-14, stays a task: the repair boot and the fleet pump, async on the server's runtime; its handle is kept and aborted at the stop.
                Some(tokio::spawn(async move {
                    crate::repair::boot(Arc::clone(&stores), cfg.clone(), own.clone()).await;
                    // AIRGAP B5 + M12: only a `fleet-pull` build carries the pull channel (and
                    // the shared pump it runs on); any other build refuses a named one by name
                    // (`repair::resolve_fleet`) and pulls nothing.
                    #[cfg(not(feature = "fleet-pull"))]
                    let _ = (entitlements, own);
                    #[cfg(feature = "fleet-pull")]
                    if let Some(fleet) = crate::repair::fleet_when_fed(cfg.fleet.clone(), crate::repair::FED_EVERY).await {
                        let estate = crate::repair::ServerEstate::new(stores.root(), &cfg, own.clone())
                            .map_err(|why| tracing::warn!(target: crate::repair::TARGET, "the pull channel runs without a repair key: every order is refused by name ({why})"))
                            .ok();
                        let (_keep, stop) = tokio::sync::watch::channel(false);
                        crate::repair::pull_loop(stores, fleet, estate, entitlements, own, stop).await;
                    }
                }))
            }
            (Some(_), None) => {
                tracing::warn!(target: "gunnar::repair", "the boot check and the pull channel need the control plane, which is off");
                None
            }
            (None, _) => None,
        };

        // Outbound webhooks: load webhooks.json and start delivering. An empty file
        // (or none) leaves the slot empty — no subscriber, no task — the "a
        // server with no webhooks pays nothing" property; a malformed file is a
        // hard error here, so a typo is caught at boot and not per delivery.
        #[cfg(feature = "webhooks")]
        {
            let webhook_config = crate::webhook::WebhookConfig::load(stores.root())
                .context("reading the webhook configuration")?;
            webhook
                .apply(webhook_config)
                .context("starting outbound webhook delivery")?;
        }

        // ── publish the counters, so something finally READS them ──────────
        //
        // One closure, two readers: `config.replication_view` is the console
        // band's and `control_view` is `Admin.Stats`'s. Before this the
        // counters existed and nothing looked at them, which is how every
        // configured standby receiving NOTHING went unnoticed for weeks.
        //
        // The closure reads the slot on EVERY call rather than closing over one
        // `Arc<Replicator>`. That is what keeps the band and `Admin.Stats`
        // honest after a live `SetReplication`: a captured replicator would go
        // on reporting the standby that was just removed, and report nothing at
        // all about the one that was just added. With no target configured the
        // slot is empty and the answer is `ReplicationReport::default()` —
        // `configured: false`, which is a fact rather than an absence.
        #[cfg(feature = "replication")]
        for view in [Some(config.replication_view.clone()), control_view.clone()]
            .into_iter()
            .flatten()
        {
            let slot = Arc::clone(&replication);
            view.fill(move || {
                slot.replicator()
                    .as_deref()
                    .map(|replicator| {
                        // The cheap report, plus the ref number PEEKED from the
                        // last on-demand walk — never a walk of its own. This
                        // closure runs on every `Admin.Stats` and every band
                        // frame, so a reader that walked would be the cold-status
                        // "re-open every store every 60 s from a polling
                        // caller". `ref_coverage_cached` opens nothing; the
                        // number is whatever the most recent `--refs` warmed,
                        // and zeros until the first one.
                        let mut report = crate::replication_view::report(replicator);
                        report.refs = replicator.ref_coverage_cached();
                        report
                    })
                    .unwrap_or_default()
            });
        }
        let _ = &control_view;

        Ok(Server {
            ssh: Some(ssh),
            grpc,
            grpc_shutdown,
            grpc_endpoint,
            control_runtime,
            stores,
            bootstrap_window_open,
            host_key_fingerprint,
            metrics,
            heartbeat,
            #[cfg(feature = "replication")]
            standby_watch,
            #[cfg(feature = "replication")]
            repair,
            bus,
            #[cfg(feature = "http")]
            http,
            scrape,
            #[cfg(feature = "replication")]
            replication,
            #[cfg(feature = "webhooks")]
            webhook,
        })
    }

    /// The event bus.
    ///
    /// The same object `Events.Watch` streams from, exposed here so an
    /// in-process caller (the replicator, and the tests that assert what it
    /// publishes) can subscribe without a gRPC listener and a credential in
    /// between.
    pub fn bus(&self) -> &Arc<EventBus> {
        &self.bus
    }

    /// The replicator, when this server has any standby configured.
    ///
    /// Exposed so an operator surface, and `tests/replication.rs`, can read the
    /// counters and the per-standby status without reaching into the tasks.
    #[cfg(feature = "replication")]
    pub fn replication(&self) -> Option<Arc<crate::mirror::Replicator>> {
        self.replication.replicator()
    }

    /// The live slot, for a caller that RECONFIGURES rather than only observes:
    /// `Mirror.SetReplication`'s in-process half and `tests/replication.rs`.
    #[cfg(feature = "replication")]
    pub fn replication_slot(&self) -> &Arc<crate::mirror::ReplicationSlot> {
        &self.replication
    }

    /// The live webhook slot, for a caller that reconfigures outbound delivery:
    /// `Webhooks.AddWebhook`'s in-process half and `tests/webhooks.rs`.
    #[cfg(feature = "webhooks")]
    pub fn webhook_slot(&self) -> &Arc<crate::webhook::WebhookSlot> {
        &self.webhook
    }

    /// **The SSH host key this server serves**, `SHA256:…` — the fingerprint a
    /// client pins, read off the key that is actually loaded rather than off a
    /// file somebody remembers.
    pub fn host_key_fingerprint(&self) -> &str {
        &self.host_key_fingerprint
    }

    /// The address the SSH listener actually bound.
    pub fn ssh_addr(&self) -> SocketAddr {
        self.ssh().local_addr()
    }

    /// The address the scrape endpoint bound, if it was started.
    ///
    /// `None` is the ordinary answer: nothing binds unless `--metrics-bind`
    /// asked for it. A test that asked for port 0 learns its port here.
    pub fn metrics_addr(&self) -> Option<SocketAddr> {
        self.scrape.as_ref().map(|(_, addr)| *addr)
    }

    /// The address the smart-HTTP listener bound, if it was started.
    #[cfg(feature = "http")]
    pub fn http_addr(&self) -> Option<SocketAddr> {
        self.http.as_ref().map(|(_, addr, _)| *addr)
    }

    /// Whether the smart-HTTP listener terminates TLS itself.
    #[cfg(feature = "http")]
    pub fn http_tls(&self) -> bool {
        self.http.as_ref().is_some_and(|(_, _, tls)| *tls)
    }

    /// The SSH listener. Present for the whole life of a `Server`; the `Option`
    /// exists only so [`Server::shutdown`] can move it out past the [`Drop`]
    /// impl.
    fn ssh(&self) -> &BoundServer {
        self.ssh
            .as_ref()
            .expect("the SSH listener is taken only by shutdown, which consumes the Server")
    }

    /// The address the gRPC listener actually bound, if it was started.
    pub fn grpc_addr(&self) -> Option<SocketAddr> {
        self.grpc_endpoint.map(|(addr, _)| addr)
    }

    /// The bound gRPC address together with whether that listener
    /// authenticates, which is what `serve` prints.
    ///
    /// One accessor rather than two, because the banner used to carry a
    /// hardcoded `UNAUTHENTICATED in v1` beside a control plane that
    /// authenticates. A caller that cannot get the address without the verdict
    /// cannot print a stale one.
    pub fn grpc_endpoint(&self) -> Option<(SocketAddr, bool)> {
        self.grpc_endpoint
    }

    /// Was the unauthenticated enrolment window open when this server started?
    ///
    /// What `serve` prints beside the gRPC address. `false` is the ordinary
    /// answer, which is why the banner says nothing at all in that case.
    pub fn bootstrap_window_open(&self) -> bool {
        self.bootstrap_window_open
    }

    /// The stores behind both listeners.
    pub fn stores(&self) -> &Arc<ServerStores> {
        &self.stores
    }

    /// Serve until a stop is requested: a signal, or on the appliance the ACPI
    /// power button ([`appliance_pid1::runtime::stop_request`]).
    ///
    /// **Not `ctrl_c()`.** That is `SIGINT` alone, and nothing in an appliance's
    /// life sends `SIGINT`: `podman stop` sends `SIGTERM` and every service
    /// manager sends `SIGTERM`. A cloud's soft stop (UpCloud `stop`, QEMU
    /// `system_powerdown`) is not a signal at all: it is an ACPI power-button
    /// press the kernel reports on the `acpi_event` generic-netlink group
    /// (corrected 2026-09-28; this said it became `SIGTERM`, and a soft stop
    /// then waited out the cloud's timeout and was hard-stopped). PID 1 also
    /// gets **no default disposition** for a signal it has not handled — the
    /// kernel discards it. See edda's `appliance_pid1::net`.
    pub async fn wait(self) -> Result<()> {
        let signal = appliance_pid1::runtime::stop_request().await;
        tracing::info!(signal, "shutting both listeners down");
        self.shutdown().await
    }

    /// Ask both listeners to stop, and wait for them.
    ///
    /// # Why the gRPC wait is bounded (measured 2026-08-03)
    ///
    /// `serve_with_incoming_shutdown` is *graceful*: it stops accepting and
    /// then waits for every in-flight request to finish. `Events.Watch` (the event stream)
    /// is server-streaming and by design **never finishes**; it ends when the
    /// subscriber goes away. So one connected UI made this wait forever, and
    /// `gunnar serve` could not be stopped with Ctrl-C at all. Measured: a test
    /// binary wedged for eleven minutes with the watch connection still
    /// `ESTAB`, no CPU, no progress.
    ///
    /// Graceful is still tried first — an in-flight `Repos.Create` should be
    /// allowed to land — but after [`GRPC_SHUTDOWN_GRACE`] the task is aborted
    /// and the process is free to exit. A watcher losing its stream at shutdown
    /// is what a watcher is built to handle. A server that cannot be stopped
    /// is not.
    pub async fn shutdown(mut self) -> Result<()> {
        // One last line before the listeners go, so the final numbers are in
        // the log rather than only in a process that is about to exit.
        if let Some(beat) = self.heartbeat.take() {
            beat.abort();
        }
        // I11's watchdog goes with the listeners. A watch that outlived the
        // server it belongs to could promote a box that is shutting down —
        // writing `role: primary` to disk as its last act, so the NEXT boot
        // comes up writable. Aborting it here is what keeps a shutdown a
        // shutdown.
        #[cfg(feature = "replication")]
        if let Some(watch) = self.standby_watch.take() {
            watch.abort();
        }
        #[cfg(feature = "replication")]
        if let Some(repair) = self.repair.take() {
            repair.abort();
        }
        // Replication stops with the server. An in-flight attempt is abandoned
        // rather than waited for: the job it was draining is still in the
        // queue's shape ("converge this repository") and the next start
        // re-derives it from the repository itself.
        #[cfg(feature = "replication")]
        self.replication.stop();
        #[cfg(feature = "webhooks")]
        self.webhook.stop();
        self.log_metrics();
        if let Some(tx) = self.grpc_shutdown.take() {
            let _ = tx.send(());
        }
        if let Some(join) = self.grpc.take() {
            let abort = join.abort_handle();
            match tokio::time::timeout(GRPC_SHUTDOWN_GRACE, join).await {
                Ok(Ok(result)) => result?,
                Ok(Err(err)) if err.is_cancelled() => {}
                Ok(Err(err)) => return Err(anyhow::anyhow!("the gRPC task panicked: {err}")),
                Err(_elapsed) => {
                    tracing::warn!(
                        grace = ?GRPC_SHUTDOWN_GRACE,
                        "the gRPC control plane did not stop gracefully; almost certainly an \
                         open Events.Watch stream, which by design never ends. Aborting it so \
                         the process can exit."
                    );
                    abort.abort();
                }
            }
        }
        // `shutdown_background`, not a drop: `shutdown` is an `async fn` and may
        // well be running on a runtime thread, where dropping a `Runtime`
        // panics. The listener is already stopped or aborted above, so there is
        // nothing left worth blocking for. The `ControlRuntime` guard's own
        // `Drop` would do this too; doing it explicitly here just tears the
        // control plane down before the SSH listener rather than after.
        self.control_runtime.shutdown();
        if let Some(ssh) = self.ssh.take() {
            ssh.shutdown("gunnar is shutting down")
                .await
                .context("stopping the SSH listener")?;
        }
        Ok(())
    }

    /// What the SSH admission gate has done since start-up.
    ///
    /// Here so an operator surface (and `tests/`) can read the saturation
    /// counters without reaching into `gunnar-ssh`.
    pub fn admission(&self) -> gunnar_ssh::AdmissionStats {
        self.ssh().admission()
    }

    /// Every counter this server keeps, read now.
    ///
    /// One reading of one set: the SSH listener and the git transport write to
    /// the same block, so `connections_refused` and `pushes_rejected` are facts
    /// about the same server rather than two crates' private tallies.
    pub fn metrics(&self) -> MetricsSnapshot {
        self.metrics.snapshot()
    }

    /// The counter block itself, for a caller that wants to keep a handle to it.
    pub fn metrics_handle(&self) -> &Arc<ServerMetrics> {
        &self.metrics
    }

    /// Write the counter line now, without waiting for the next heartbeat.
    pub fn log_metrics(&self) {
        emit_metrics(&self.metrics.snapshot());
    }
}

impl Drop for Server {
    fn drop(&mut self) {
        if let Some(beat) = self.heartbeat.take() {
            beat.abort();
        }
        // Same reason as in `shutdown`: a watchdog that survives its server
        // could promote a box on its way out.
        #[cfg(feature = "replication")]
        if let Some(watch) = self.standby_watch.take() {
            watch.abort();
        }
        #[cfg(feature = "replication")]
        if let Some(repair) = self.repair.take() {
            repair.abort();
        }
        #[cfg(feature = "replication")]
        self.replication.stop();
        #[cfg(feature = "webhooks")]
        self.webhook.stop();
        // A `Server` dropped without `shutdown` (a test that panicked, say)
        // would otherwise take its control runtime down by a blocking `Drop`,
        // which panics on a runtime thread and would bury the real failure. The
        // `ControlRuntime` field's own `Drop` already routes through
        // `shutdown_background`; this is that same call, made explicit.
        self.control_runtime.shutdown();
    }
}

/// **Say, at every start, whether the unauthenticated enrolment door is open.**
///
/// Silence here is the disease this repository keeps finding. A server standing
/// with an open door is a server anybody who can reach the control port can
/// take; an operator must never have to guess which state their box is in, and
/// on an unenrolled server there is by definition no credential with which to
/// ask `Admin.Health`.
///
/// Split out of [`Server::start`] and given both readings rather than taking
/// them, so every branch can be driven from a test on any box. A warning that
/// reads its own host can only ever be seen in the state that host happens to
/// be in.
///
/// Returns the line it emitted, so a test asserts on the **text an operator
/// sees** rather than on a boolean that proves nothing about the log.
pub fn announce_readiness(
    readiness: &Readiness,
    data_root: &std::path::Path,
    grpc_bind: SocketAddr,
    no_grpc: bool,
) -> String {
    match readiness {
        Readiness::Ready => {
            let line = format!("identity table ready at {}", data_root.display());
            tracing::info!(data_root = %data_root.display(), "{line}");
            line
        }
        Readiness::AwaitingBootstrap if no_grpc => {
            let line = format!(
                "NO IDENTITY HAS EVER BEEN ENROLLED at {}, and the gRPC control plane is \
                 DISABLED, so there is no way to enrol one. Nobody can reach this server. Start \
                 without --no-grpc and call Bootstrap.Enroll, or pass --admin-key.",
                data_root.display()
            );
            tracing::error!(data_root = %data_root.display(), "{line}");
            line
        }
        Readiness::AwaitingBootstrap => {
            let line = format!(
                "NO IDENTITY HAS EVER BEEN ENROLLED at {}. Bootstrap.Enroll on {grpc_bind} is \
                 open to ANY caller that can reach it, with NO credential, and it creates a \
                 gunnar-admin. It closes permanently on the first success and deleting \
                 identities does not reopen it. Enrol now, or bind the control plane where \
                 nobody else can reach it.",
                data_root.display()
            );
            tracing::warn!(
                data_root = %data_root.display(),
                %grpc_bind,
                bootstrap_window = "OPEN",
                "{line}"
            );
            line
        }
        Readiness::SealedAndEmpty { record } => {
            let who = match record {
                Some(record) => format!(
                    "was bootstrapped by \"{}\" at {}",
                    record.principal_id, record.at
                ),
                None => "holds principals but no keys and carries no bootstrap record, so it \
                         was written before that record existed, and"
                    .to_owned(),
            };
            let line = format!(
                "this server {who} its identity table at {} now holds no keys. NOBODY CAN \
                 AUTHENTICATE against it, and the enrolment window stays shut — deleting \
                 identities does not reopen it. Add a key with `gunnar key add` while the \
                 server is down.",
                data_root.display()
            );
            tracing::error!(
                data_root = %data_root.display(),
                bootstrap_window = "SHUT",
                "{line}"
            );
            line
        }
    }
}

/// **Turn `--grpc-bind` into the address the listener binds.**
///
/// The one caller of [`BindSpec::resolve_with`] in the server, and the one
/// place `appliance_pid1::net::interface_ipv4` is wired to it. An address resolves to itself
/// and never touches netlink; an interface is asked, once, and a refusal stops
/// the start rather than widening the bind — the whole point of the spelling.
/// Returns the address to bind, or `None` when the named interface exists and
/// holds no IPv4 — see [`BindSpec::resolve_with`] for why that is `Ok` and not
/// an error, and what the caller owes it.
/// **Bind the control plane's listener — on a DEVICE when the spelling named
/// one.** `eth1:50051` resolves to the address eth1 holds, and an address alone
/// is not the interface: MEASURED 2026-09-05 on the KVM bench rig (two slirp
/// NICs, both 10.0.2.15), the listener bound to eth1's address answered on
/// eth0's forward and refused eth1's, because the kernel routes by address and
/// both interfaces held it. `SO_BINDTODEVICE` makes the socket the interface's
/// — the same call the DHCP client makes for its broadcast socket — so the
/// declaration `--grpc-bind eth1:50051` means eth1 on every machine, not just on
/// one whose subnets happen to differ. An address spelling (`None`) binds as
/// before. The socket is made non-blocking and handed to the control runtime's
/// I/O driver by `from_std`, on the runtime that calls this.
fn bind_grpc_listener(addr: SocketAddr, device: Option<&str>) -> Result<tokio::net::TcpListener> {
    use socket2::{Domain, Protocol, Socket, Type};
    let domain = if addr.is_ipv6() { Domain::IPV6 } else { Domain::IPV4 };
    let sock = Socket::new(domain, Type::STREAM, Some(Protocol::TCP))
        .context("opening the gRPC listener socket")?;
    sock.set_reuse_address(true).context("SO_REUSEADDR on the gRPC listener")?;
    if let Some(dev) = device {
        sock.bind_device(Some(dev.as_bytes()))
            .with_context(|| format!("SO_BINDTODEVICE {dev} on the gRPC listener"))?;
    }
    sock.bind(&addr.into()).with_context(|| format!("bind {addr}"))?;
    sock.listen(1024).context("listen on the gRPC listener")?;
    sock.set_nonblocking(true).context("non-blocking gRPC listener")?;
    tokio::net::TcpListener::from_std(sock.into()).context("registering the gRPC listener with the runtime")
}

/// How long `serve` waits for the gRPC bind address to turn up on an interface
/// before it gives up. Thirty seconds covers a tailnet or WireGuard interface
/// that comes up after this service on a cold boot; anything slower is a box
/// with a real problem, and the bounded exit hands it to the supervisor.
pub const GRPC_BIND_PATIENCE: std::time::Duration = std::time::Duration::from_secs(30);

/// The pause between two bind attempts while waiting for the address.
pub const GRPC_BIND_RETRY_EVERY: std::time::Duration = std::time::Duration::from_millis(500);

/// `EADDRNOTAVAIL` and nothing else. That errno means "no interface on this
/// box has that address", which on a cold boot means *yet*. `EADDRINUSE`, a
/// bad device, a permission refusal: none of those get better by waiting.
fn bind_failed_because_the_address_is_not_here_yet(err: &anyhow::Error) -> bool {
    err.chain()
        .filter_map(|cause| cause.downcast_ref::<std::io::Error>())
        .any(|io| io.kind() == std::io::ErrorKind::AddrNotAvailable)
}

/// Run `attempt` until it binds, retrying only while the failure is the
/// address-not-here-yet errno, every `every`, for at most `patience`.
///
/// Returns the final result AND the number of retries taken, so a caller (and
/// the test) can tell a first-try bind from a waited-for one.
async fn bind_when_the_address_arrives<T, F>(
    mut attempt: F,
    patience: std::time::Duration,
    every: std::time::Duration,
) -> (Result<T>, u32)
where
    F: FnMut() -> Result<T>,
{
    let started = std::time::Instant::now();
    let mut retries = 0u32;
    loop {
        match attempt() {
            Ok(bound) => {
                if retries > 0 {
                    tracing::info!(
                        retries,
                        waited_ms = started.elapsed().as_millis() as u64,
                        "the gRPC bind address arrived; the control plane is bound"
                    );
                }
                return (Ok(bound), retries);
            }
            Err(err) if bind_failed_because_the_address_is_not_here_yet(&err) => {
                if started.elapsed() >= patience {
                    return (
                        Err(err.context(format!(
                            "the address was on no interface for {} s ({retries} retries)",
                            patience.as_secs()
                        ))),
                        retries,
                    );
                }
                if retries == 0 {
                    tracing::warn!(
                        patience_s = patience.as_secs(),
                        "the gRPC bind address is on no interface yet (EADDRNOTAVAIL); waiting for \
                         it. A tailnet or VPN interface that comes up after this service is the \
                         usual cause"
                    );
                }
                retries += 1;
                tokio::time::sleep(every).await;
            }
            Err(err) => return (Err(err), retries),
        }
    }
}

#[cfg(test)]
mod grpc_bind_patience_tests {
    use super::*;
    use std::time::Duration;

    fn not_here_yet() -> anyhow::Error {
        anyhow::Error::from(std::io::Error::from(std::io::ErrorKind::AddrNotAvailable))
            .context("bind 100.64.0.1:50061")
    }

    /// The address arrives on the third try: bound, and the retry count says
    /// the wait was real (a wrapper that never retried would return 0 here).
    #[test]
    fn an_address_that_arrives_later_is_bound_and_the_retries_are_counted() {
        let rt = tokio::runtime::Builder::new_current_thread().enable_all().build().unwrap();
        let mut left = 2;
        let (bound, retries) = rt.block_on(bind_when_the_address_arrives(
            || {
                if left > 0 {
                    left -= 1;
                    Err(not_here_yet())
                } else {
                    Ok(7u8)
                }
            },
            Duration::from_secs(5),
            Duration::from_millis(1),
        ));
        assert_eq!(bound.unwrap(), 7);
        assert_eq!(retries, 2);
    }

    /// Any other errno fails on the first try, with zero retries: waiting on
    /// `EADDRINUSE` would hide a second server on the port for thirty seconds.
    #[test]
    fn an_address_in_use_is_not_waited_for() {
        let rt = tokio::runtime::Builder::new_current_thread().enable_all().build().unwrap();
        let mut calls = 0;
        let (bound, retries) = rt.block_on(bind_when_the_address_arrives(
            || -> Result<u8> {
                calls += 1;
                Err(anyhow::Error::from(std::io::Error::from(std::io::ErrorKind::AddrInUse)))
            },
            Duration::from_secs(5),
            Duration::from_millis(1),
        ));
        assert!(bound.is_err());
        assert_eq!(retries, 0);
        assert_eq!(calls, 1);
    }

    /// The patience is a bound: an address that never comes gives up with the
    /// wait named in the error, not a hang.
    #[test]
    fn an_address_that_never_comes_gives_up_after_the_patience() {
        let rt = tokio::runtime::Builder::new_current_thread().enable_all().build().unwrap();
        let (bound, retries) = rt.block_on(bind_when_the_address_arrives(
            || -> Result<u8> { Err(not_here_yet()) },
            Duration::from_millis(30),
            Duration::from_millis(5),
        ));
        let msg = format!("{:#}", bound.unwrap_err());
        assert!(msg.contains("on no interface for 0 s"), "{msg}");
        assert!(retries >= 3, "retries={retries}");
    }

    /// The real socket path: TEST-NET-1 is on no interface anywhere, and the
    /// kernel's answer is the errno the wait is keyed on. This is the red that
    /// proves the classifier reads the errno that `bind_grpc_listener` actually
    /// produces, not one the test invented.
    #[test]
    fn the_real_bind_error_for_an_absent_address_is_classified_as_not_here_yet() {
        let rt = tokio::runtime::Builder::new_current_thread().enable_io().build().unwrap();
        let _guard = rt.enter();
        let err = bind_grpc_listener("192.0.2.1:0".parse().unwrap(), None)
            .expect_err("TEST-NET-1 is on no interface");
        assert!(bind_failed_because_the_address_is_not_here_yet(&err), "{err:#}");
        let in_use = anyhow::Error::from(std::io::Error::from(std::io::ErrorKind::AddrInUse));
        assert!(!bind_failed_because_the_address_is_not_here_yet(&in_use));
    }
}

#[cfg(test)]
mod grpc_bind_device_tests {
    use super::*;

    /// A device that does not exist is refused BY NAME before anything binds —
    /// the red that proves the device half is wired at all.
    #[test]
    fn a_device_the_machine_does_not_have_is_refused_by_name() {
        let rt = tokio::runtime::Builder::new_current_thread().enable_io().build().unwrap();
        let _guard = rt.enter();
        let err = bind_grpc_listener("127.0.0.1:0".parse().unwrap(), Some("nosuchdev0"))
            .expect_err("a nonexistent device cannot be bound");
        let msg = format!("{err:#}");
        assert!(msg.contains("SO_BINDTODEVICE nosuchdev0"), "{msg}");
    }

    /// The address-only spelling binds exactly as before, and on the loopback
    /// device by name too (the one device every box has).
    #[test]
    fn an_address_binds_and_so_does_the_loopback_device() {
        let rt = tokio::runtime::Builder::new_current_thread().enable_io().build().unwrap();
        let _guard = rt.enter();
        let plain = bind_grpc_listener("127.0.0.1:0".parse().unwrap(), None).unwrap();
        assert!(plain.local_addr().unwrap().port() != 0);
        match bind_grpc_listener("127.0.0.1:0".parse().unwrap(), Some("lo")) {
            Ok(l) => assert!(l.local_addr().unwrap().port() != 0),
            // Kernels before 5.7 gate SO_BINDTODEVICE behind CAP_NET_RAW; a
            // permission refusal is that gate, not a wiring defect.
            Err(e) => assert!(format!("{e:#}").contains("SO_BINDTODEVICE lo"), "{e:#}"),
        }
    }
}

fn resolve_grpc_bind(spec: &BindSpec) -> Result<Option<SocketAddr>> {
    match spec
        .resolve_with(|name| appliance_pid1::net::interface_ipv4(name).map_err(|e| format!("{e:#}")))
        .map_err(|why| anyhow::anyhow!("{why}"))?
    {
        gunnar_common::bindspec::Resolved::Bound(addr) => Ok(Some(addr)),
        gunnar_common::bindspec::Resolved::NoAddress { interface, why } => {
            // ERROR and not WARN: an operator who asked for a control plane
            // has not got one, and nothing else on the box will say so.
            tracing::error!(
                interface = %interface,
                "{why} The gRPC control plane will NOT start. The git listener is unaffected \
                 and this server still serves and authenticates pushes and fetches over SSH."
            );
            Ok(None)
        }
    }
}

/// `--admin-key` is either the key line itself or `@<path>`.
fn read_key_argument(spec: &str) -> Result<String> {
    match spec.strip_prefix('@') {
        Some(path) => std::fs::read_to_string(path)
            .with_context(|| format!("reading the public key from {path}"))
            .map(|text| text.trim().to_owned()),
        None => Ok(spec.trim().to_owned()),
    }
}

/// **The startup banner's host-key line.**
///
/// The fingerprint IN FULL, never shortened: the dashboard's `short_fp` is for
/// recognising a key you already know, and this line exists to be the first
/// place anybody learns it. It is read off a VNC screenshot by a human and
/// typed into a `known_hosts` line, and twelve characters cannot be typed into
/// anything.
///
/// The word `sshkey` leads it so a screenshot can be searched for one token.
///
/// Why this is worth a line of banner at all: `known_hosts` records no
/// provenance. Once written, a key taken off the wire by `ssh-keyscan` and a
/// key read off a console are the same bytes, for ever after, and nothing in
/// the file can tell the two apart. The only thing that makes a forge push
/// mean anything is WHERE the fingerprint came from — a record written by the
/// install (`deploy/upcloud`'s receipt) or a reading off the box's own console.
/// Never from whatever answered on `:2222`.
pub(crate) fn host_key_banner_line(fingerprint: &str) -> String {
    format!("sshkey {fingerprint}   (pin THIS; a key scanned off the port proves nothing)")
}

/// The host key, as russh's `ssh-key` sees it.
fn russh_private_key(openssh: &[u8]) -> Result<gunnar_ssh::russh::keys::PrivateKey> {
    gunnar_ssh::russh::keys::PrivateKey::from_openssh(openssh)
        .map_err(|err| anyhow::anyhow!("the stored host key did not parse: {err}"))
}

/// Run `gunnar serve` to completion.
pub async fn run(stores: Arc<ServerStores>, config: ServeConfig, check_only: bool) -> Result<()> {
    // The descriptor limit, raised once. `ServerStores::open` already did this
    // (its handle cache is sized from the answer); calling it here is the
    // guarantee that the `limits.nofile` line — a WARN row on the dashboard
    // when the limit is low — exists before a listener binds, whichever path
    // built `stores`.
    let _ = crate::limits::nofile();
    // ── THE NARRATOR, BEFORE ANYTHING ELSE ──────────────────────────────────
    //
    // MEASURED 2026-09-23: `boot::ship_boot_narration` ends at `end of
    // hand-off` — the moment the DHCP lease makes the wire usable — and the
    // estate's twin died in the silence after it, having read its config,
    // bound its listeners, taken its role, held its certificate and started
    // following. None of that was said. From here on every boundary is.
    //
    // The sink is `korp.log_udp=` off the kernel command line and nothing
    // else. Absent ⇒ nothing is armed, no thread is started, no datagram is
    // sent, and the two file readers (the ESP's BOOTLOG.TXT and /data) carry
    // the story as they always did. That is the off switch.
    use crate::narrate::Stage;
    crate::narrate::install();
    crate::narrate::data_home(stores.root());
    crate::narrate::at(
        Stage::Config,
        &format!(
            "data root {}, engine {:?}, ssh {}, grpc {:?}, twin_role={}, is_twin={}",
            stores.root().display(),
            stores.engine(),
            config.ssh_bind,
            config.grpc_bind,
            config.twin_role,
            stores.is_twin(),
        ),
    );
    // ── PID 1's own duties, before anything binds ───────────────────────────
    //
    // The network first and by hand, because the interface has to exist before
    // a listener can be reached over it and on this appliance nothing else
    // brings it up: measured on the reference desktop 2026-08-03, the kernel is built with
    // `CONFIG_IP_PNP` off, so `ip=` on the command line configures nothing.
    // A refused static plan or malformed GUNNAR_NET is fatal on purpose: the
    // alternative is a server that starts, logs "ssh 0.0.0.0:2222" and can
    // never be reached, which is the same failure `Readiness::RefuseToServe`
    // exists to prevent. A DHCP timeout is NOT fatal — bring_up_network warns
    // and serves unreachable, the missing-GUNNAR_NET posture — because there
    // the world declined to answer; the declaration did not lie.
    crate::narrate::at(Stage::Network, "bringing the declared interfaces up over netlink");
    let dhcp_session = crate::boot::bring_up_network()?;
    // The lease. The UDP narration wire is retired (2026-09-24): the installer
    // collects this box's whole boot log over its own mutual TLS.
    crate::narrate::at(Stage::Lease, &format!("{} DHCP lease(s) held", dhcp_session.len()));
    crate::narrate::at(Stage::Sink, "the boot log is kept for the installer (its bootstrap listener serves it)");
    // **The clock, and here rather than anywhere later.** It is after the
    // network because it is fetched over the network, and before every listener
    // because `gunnar_grpc::auth::MAX_CLOCK_SKEW_MS` is ±5 minutes: an appliance
    // that serves a single request on a clock two hours out refuses a genuine
    // credential and a genuine Stripe webhook signature, which is precisely what
    // gunnar.rs measured on 2026-09-14. Never fatal, never a `?`: see
    // `crate::clock`'s note on why a failed clock fetch must not cost a forge.
    crate::clock::synchronise_at_boot().await;
    // And again when a credential is refused for skew, and every six hours: a
    // boot-time step fixes the clock the box boots with, never the drift of the
    // months until the next re-image. Background threads; see `crate::clock`.
    crate::clock::arm_triggers();
    // **The bind spelling becomes an address here, and not one line earlier.**
    // `--grpc-bind eth1:50051` names an interface whose address arrives with
    // the DHCP lease above; asking before `bring_up_network` would find an
    // interface with no address every time. Asking AFTER is also what makes
    // the refusal actionable: if it fires, the lease genuinely did not land,
    // and the warning `bring_up_network` already logged says why.
    let mut config = config;
    // **A bind by ROLE** (`utility:50051`, the lean estate, A19 row 17) becomes this box's
    // utility NIC here, after the network is up: `eth1` on a box with two NICs, `eth0` on a
    // box laid behind its front with its utility NIC alone.
    config.grpc_bind = crate::boot::role_bound_here(config.grpc_bind).context("resolving the --grpc-bind role to this box's NIC")?;
    if let BindSpec::Interface { interface, .. } = &config.grpc_bind {
        tracing::info!(interface = %interface, "the control plane's utility door is {interface} on this box");
    }
    // **An interface whose lease PID 1 still asks for** (its boot DHCP failed — estate-box3
    // D4, a box rebooted while its network was down) is WAITED for, here on the appliance's
    // path as in `Server::start`: the control plane binds once the address arrives.
    if let BindSpec::Interface { interface, .. } = &config.grpc_bind {
        let has_address = |i: &str| matches!(appliance_pid1::net::interface_ipv4(i), Ok(Some(_)));
        if !has_address(interface) && appliance_pid1::net::dhcp_pending(interface) {
            tracing::info!(interface = %interface, "the control plane waits for {interface}'s address (PID 1 asks DHCP again in the background)");
            while !has_address(interface) && appliance_pid1::net::dhcp_pending(interface) {
                tokio::time::sleep(std::time::Duration::from_secs(1)).await;
            }
        }
    }
    let resolved = resolve_grpc_bind(&config.grpc_bind).with_context(|| {
        "resolving --grpc-bind after the network came up. A medium that names a NIC this \
         machine does not have is a declaration that lies, and gunnar stops rather than \
         falling back to a wildcard bind: an unauthenticated control plane on every \
         interface is the defect the interface spelling exists to close."
    })?;
    let resolved_grpc_bind = match resolved {
        Some(addr) => addr,
        None => {
            config.no_grpc = true;
            gunnar_grpc::DEFAULT_BIND
        }
    };
    // Pin the answer into the config so `Server::start`, the dashboard and the
    // readiness line all report the address the listener took rather than the
    // spelling it was asked for — and so the netlink question is asked once.
    if let BindSpec::Interface { interface, .. } = &config.grpc_bind {
        config.grpc_bind_device = Some(interface.clone());
    }
    config.grpc_bind = BindSpec::Addr(resolved_grpc_bind);
    if !config.no_grpc && !resolved_grpc_bind.ip().is_loopback() {
        tracing::info!(
            grpc_bind = %resolved_grpc_bind,
            "gRPC control plane will bind a NON-loopback address"
        );
    }
    // Whether /data survived the last install (the ESP record korp-installer
    // left) — one event, one dashboard row, before the band starts.
    crate::boot::announce_install_mode(stores.root());
    // The start-up disk test, appliance only (same gate as the dashboard, or
    // `GUNNAR_DISKTEST=1`): one thread, one `disktest` row, never waited for.
    crate::disktest::spawn(stores.root().to_path_buf());
    if appliance_pid1::net::is_init() {
        tracing::info!(
            "running as pid 1: no init system underneath, so this process handles SIGTERM and \
             reaps re-parented orphans itself"
        );
    }
    // Installed unconditionally. A reaper that only exists at pid 1 is a reaper
    // no test on a developer machine ever runs.
    let reaper = tokio::spawn(appliance_pid1::runtime::reaper());
    // **The soft-delete sweep**: the half of `repo delete` that gives the bytes
    // back. Hourly, and once now. Without it a retention window would be a
    // promise to keep repositories forever, and `/data` would fill with
    // repositories nobody can list.
    let tombstone_sweep = crate::stores::spawn_tombstone_sweep(Arc::clone(&stores));
    // **The flow flush**: the half of the activity ledger that survives a
    // restart. The counters a transfer bumps live in atomics in this process;
    // this copies them into the control store every minute, and rolls the UTC
    // day on a box that is idle at midnight. See `stores::spawn_flow_flush`.
    // The band's fold, refreshed by the flush task once a minute. `None` until
    // the first flush: a row of zeroes before anything has been folded would
    // claim the box served nobody today.
    let flow_band: Arc<std::sync::Mutex<Option<gunnar_core::Fleet>>> =
        Arc::new(std::sync::Mutex::new(None));
    let flow_flush = crate::stores::spawn_flow_flush(Arc::clone(&stores), Arc::clone(&flow_band));
    config.flow_view.fill(move || *flow_band.lock().expect("the flow band"));
    // Kept out of the move into `Server::start` below, for the shutdown flush.
    let flow_at_exit = Arc::clone(stores.flow());
    // The VGA status screen, appliance only (PID 1 booted by korp-start) or
    // `GUNNAR_DASHBOARD=<device>`. One thread, fed by the log tap; the lease
    // line above is already in the tap's buffer and is its first row. Never
    // fatal, never waited for — see `crate::dashboard`.
    // The band starts before the server does, so it is handed a VIEW and the
    // replicator is dropped into it below, once `Server::start` has one. A box
    // with no standby leaves it unfilled and the band draws no `standby` row.
    let replication_view = config.replication_view.clone();
    crate::dashboard::spawn(crate::dashboard::Facts {
        ssh_bind: config.ssh_bind,
        grpc_bind: (!config.no_grpc).then_some(resolved_grpc_bind),
        data_root: stores.root().to_path_buf(),
        build: gunnar_common::BUILD_REF.to_owned(),
        engine: Some(stores.engine()),
        replication: replication_view.clone(),
        tenants: config.tenants_view.clone(),
        flow: config.flow_view.clone(),
    });
    // Exactly when the address is DHCP-learned, a task keeps the lease alive
    // (RFC 2131 §4.4.5 — appliance_pid1::runtime::dhcp_renewal). Static and unconfigured paths
    // return an EMPTY list above and gain no task: renewal exists only where a
    // lease does — and a two-NIC appliance holds TWO leases, each with its own
    // T1, so there is one task per lease rather than one task guessing which
    // interface it is keeping alive.
    #[cfg(feature = "dhcp-client")]
    let renewals: Vec<tokio::task::JoinHandle<()>> = dhcp_session
        .into_iter()
        .map(|session| tokio::spawn(crate::boot::dhcp_renewal(session)))
        .collect();
    // Built without `dhcp-client`: `DhcpSession` is UNINHABITED, so this is the
    // compiler proving there is no lease rather than a runtime check claiming
    // it. The `Some` arm is `match session {}` — unreachable by construction —
    // and the abort sites below need no cfg of their own.
    #[cfg(not(feature = "dhcp-client"))]
    let renewals: Vec<tokio::task::JoinHandle<()>> = dhcp_session
        .into_iter()
        .map(|session| match session {})
        .collect();

    // ── the tunnel, AFTER the outer address ─────────────────────────────────
    //
    // Order is load-bearing and not alphabetical: a WireGuard peer is reached
    // over the outer address, so a handshake attempted before DHCP has
    // finished has no source address to leave from. It is strictly additive —
    // no `wg0.conf` in the data root means one INFO line and nothing else
    // changes — and it is deliberately NOT fatal when the kernel refuses,
    // because an appliance that will not serve git because a tunnel failed is
    // worse than one that serves on its outer address and says so loudly.
    //
    // Nothing below has to be told about it: both listeners bind `0.0.0.0`
    // (`ServeConfig::default`, ssh 0.0.0.0:2222), so `wg0` is served the
    // moment it has an address.
    // **This binding is the tunnel's lifetime.** The data plane is a thread in
    // THIS process now, not a device the kernel owns: `Tunnel` holds the pump
    // and the tun fd, and dropping the last handle stops the pump and deletes
    // `wg0`. The clone below goes to a watcher task that RETURNS after the
    // first handshake, so if this binding were consumed instead of kept the
    // tunnel would come up, announce itself and vanish a few seconds later.
    // It stays in scope for the whole of `run`, which is the process.
    // A pairing box's `wg0.conf` is written by the pairing loop once the
    // installer's door has assigned the address (pairing.rs); on a reboot it
    // is already here and this lane raises it like any other.
    #[cfg(feature = "wg-client")]
    let tunnel = crate::wg::bring_up(stores.root())?;
    #[cfg(feature = "wg-client")]
    let handshake = tunnel.clone().map(|t| {
        tokio::spawn(crate::wg::handshake_watch(
            t,
            20,
            std::time::Duration::from_secs(3),
        ))
    });
    // Built without `wg-client`: there is no tunnel type, no boringtun and no
    // `wg_appliance_core` in this binary at all, so there is nothing to bring
    // up, nothing to watch and no banner line to write. No `tunnel` binding
    // exists in this build — only the watcher handle the two abort paths below
    // reach for, and it is permanently `None`.
    #[cfg(not(feature = "wg-client"))]
    let handshake: Option<tokio::task::JoinHandle<()>> = None;

    // Read off both BEFORE the move below, so the role line can be narrated
    // after the listeners are up: `Server::start` consumes `stores` and
    // `config`, and the role is a property of the boot rather than of the
    // server object.
    let config_twin_role = config.twin_role;
    let stores_is_twin = stores.is_twin();
    // `Server::start` fills `config.replication_view` — the same object the
    // band above was handed — once the replicator exists.
    let server = Server::start(stores, config).await?;
    // The staleness alarm: one WARN into the log and one event-stream event the first
    // time a standby stops taking bytes — with "last successful replication:
    // NEVER" as its own message. Harmless when nothing is configured: the view
    // reports no standbys and the task never fires.
    let staleness_watch = crate::replication_view::spawn_staleness_watch(
        replication_view.clone(),
        Arc::clone(server.bus()),
    );
    // The banner is BUILT once and written twice (one implementation, one writer): to stdout,
    // which on the appliance is `/dev/console` = the serial line that
    // `installtest/src/iso.rs` reads its boot proof off, and — appliance only —
    // mirrored to `/dev/tty0` so a KVM GRAPHICS console ends on gunnar's own
    // status instead of the kernel's last printk. Measured 2026-08-12: with the
    // banner on serial alone, a VGA console showed a healthy boot log that never
    // once said "gunnar".
    let mut banner = String::new();
    {
        use std::fmt::Write as _;
        writeln!(banner, "ssh   {}", server.ssh_addr()).ok();
        // **THE HOST KEY, ON A CHANNEL PRODUCTION HAS.** Directly under the
        // address it belongs to, because the two together are the whole of
        // what a client needs to trust this box.
        //
        // MEASURED 2026-09-23. The fingerprint was logged and nowhere else:
        // `tracing` → stderr → `/dev/console` → the LAST `console=` on the
        // command line → the serial line, and the UpCloud shape has no serial
        // line (a panic there prints PID 1's first `print!` to nothing; the
        // ISO's own gate VMs are the only boxes that ever had it, in their
        // `serial.log`). So in production the one fact a client must pin
        // reached NOBODY, and this laptop's `known_hosts` has eight
        // `ssh-keyscan`-shaped lines for `[git.gunnar.rs]:2222` to prove what
        // fills the gap.
        //
        // This banner is the channel that answers: it is printed on stdout AND
        // mirrored to `/dev/tty0` (`appliance_pid1::console::mirror_to_gfx_console`, a few
        // lines below), which is the VNC screen of an UpCloud VM. The
        // WireGuard public key is already here for exactly this argument — on
        // a shell-less appliance the banner is the only channel that can
        // answer "is the key on this box the one I enrolled?".
        //
        // The narration wire (`korp.log_udp=`) was considered FIRST and is
        // not usable: it is armed and flushed long before this point, but the
        // one door onto it (`appliance_pid1::narration::narration_door`) classifies any
        // line carrying `sha256:` as `pubkey` and replaces it with a withheld
        // notice, the collector refuses it identically
        // (`<ops>/xtask/src/bootnarr.rs::secret_class`), and the 40-character
        // opaque-run rule catches the fingerprint's body with the label
        // stripped off — so no spelling escapes. That door is an OWNER RULE
        // ("a PUBLIC verifying key … is withheld anyway, because it is not
        // progress"), spelled in three repositories. Putting the fingerprint
        // on that wire needs the rule reversed and all three vocabularies
        // changed together; it is not something this line may do quietly.
        writeln!(banner, "{}", host_key_banner_line(server.host_key_fingerprint())).ok();
        match server.grpc_endpoint() {
            // The open enrolment window comes FIRST on this line, before the
            // reassuring "authenticated": a server that has never been enrolled is
            // one anybody who can reach the port can take, and an operator reading
            // three lines of banner must not have to know that "authenticated"
            // does not cover `Bootstrap.Enroll`.
            Some((addr, _)) if server.bootstrap_window_open() => writeln!(
                banner,
                "grpc  {addr}  (NOT ENROLLED — Bootstrap.Enroll is open to anyone who can reach \
                 this port. `gunnar control bootstrap <name>` claims it, once.)"
            )
            .ok(),
            Some((addr, true)) => {
                writeln!(banner, "grpc  {addr}  (authenticated: sshsig + the ACL model)").ok()
            }
            Some((addr, false)) => writeln!(
                banner,
                "grpc  {addr}  (UNAUTHENTICATED — the escape hatch is on)"
            )
            .ok(),
            None => writeln!(banner, "grpc  disabled").ok(),
        };
        // Only when it is on, like metrics below. Without this line a server asked
        // to bind HTTP on port 0 has bound a listener nobody can find: ssh, grpc,
        // metrics and the log level were all announced, and the one listener a
        // stock `git` client talks to was not.
        #[cfg(feature = "http")]
        if let Some(addr) = server.http_addr() {
            if server.http_tls() {
                writeln!(
                    banner,
                    "https {addr}  (TLS terminated here; clients pin the issuer)"
                )
                .ok();
            } else {
                writeln!(banner, "http  {addr}").ok();
            }
        }
        // Only when it is on: a line reading "metrics disabled" on every
        // server that never asked for the feature is banner an operator learns to
        // skip, and the banner's whole value is that it is short enough to read.
        if let Some(addr) = server.metrics_addr() {
            writeln!(
                banner,
                "metrics http://{addr}{}   (no authentication{})",
                crate::scrape::SCRAPE_PATH,
                if addr.ip().is_loopback() {
                    ""
                } else {
                    "; NOT on loopback"
                }
            )
            .ok();
        }
        // The tunnel, and ONLY when there is one — same rule as metrics above.
        //
        // It carries the appliance's WireGuard PUBLIC KEY, which is not
        // decoration: peering is mutual allowlisting by public key, and a key
        // the hub does not have in its `[Peer]` stanza gets **silence** rather
        // than a rejection. So "is the key on this box the one I enrolled?" is
        // the first question of every failed tunnel, and on a shell-less
        // appliance the banner — serial AND the mirrored graphics console — is
        // the only channel that can answer it.
        #[cfg(feature = "wg-client")]
        if let Some(tunnel) = &tunnel {
            writeln!(banner, "{}", tunnel.banner_line()).ok();
        }
        // On stdout beside the addresses rather than in the log: this is
        // the one fact an operator needs BEFORE the incident, and by the time they
        // want it the log is the thing they cannot read enough of. It also warns
        // them off the trap — both signals default to terminating a process, and
        // that is what they did to this one until today.
        writeln!(
            banner,
            "log   {}   (SIGUSR1 → one rung more verbose, SIGUSR2 → back; pid {})",
            gunnar_common::logging::current().unwrap_or_else(|| "?".to_owned()),
            std::process::id()
        )
        .ok();
    }
    print!("{banner}");
    appliance_pid1::console::mirror_to_gfx_console(&banner);
    // **ONE LINE PER LISTENER, WITH THE ADDRESS IT TOOK.** Not the spelling it
    // was asked for: `--grpc-bind eth1:50051` and `10.13.9.24:50051` are two
    // different facts and only the second one can be dialled. A listener the
    // build does not carry says nothing rather than saying `disabled`, so a
    // reader counting listeners counts the ones that exist.
    crate::narrate::at(Stage::Listener, &format!("ssh BOUND {}", server.ssh_addr()));
    match server.grpc_endpoint() {
        Some((addr, _)) => crate::narrate::at(
            Stage::Listener,
            &format!(
                "grpc BOUND {addr}{}",
                if server.bootstrap_window_open() { " (NOT ENROLLED — Bootstrap.Enroll is open)" } else { "" }
            ),
        ),
        None => crate::narrate::at(Stage::Listener, "grpc NOT bound — the control plane is off"),
    }
    #[cfg(feature = "http")]
    if let Some(addr) = server.http_addr() {
        crate::narrate::at(
            Stage::Listener,
            &format!("{} BOUND {addr}", if server.http_tls() { "https" } else { "http" }),
        );
    }
    if let Some(addr) = server.metrics_addr() {
        crate::narrate::at(Stage::Listener, &format!("metrics BOUND {addr}"));
    }
    // **THE ROLE**, from the two places that decide it: what the medium
    // declared and what the data home already is. They are narrated together
    // because a disagreement between them is the interesting case — a box
    // installed from main's medium onto a disk that already holds a twin's
    // certificate is exactly the shape the estate re-images into.
    crate::narrate::at(
        Stage::Role,
        &format!(
            "{} (declared twin_role={}, data home says is_twin={})",
            if stores_is_twin || config_twin_role { "TWIN" } else { "PRIMARY" },
            config_twin_role,
            stores_is_twin,
        ),
    );
    if check_only {
        reaper.abort();
        tombstone_sweep.abort();
        flow_flush.abort();
        staleness_watch.abort();
        for renewal in &renewals {
            renewal.abort();
        }
        if let Some(handshake) = &handshake {
            handshake.abort();
        }
        server.shutdown().await?;
        return Ok(());
    }
    // **SERVING.** The end of the boot story and the beginning of the ticks:
    // from here the only thing the box is expected to say is one short line a
    // minute, and the ear's threshold is measured against that. A box that
    // reached here and then went quiet is a DEATH AT A NAMED PLACE — which is
    // the reading the estate did not have on 2026-09-23.
    crate::narrate::at(
        Stage::Serving,
        &format!(
            "up at {} s",
            crate::narrate::uptime_s(),
        ),
    );
    let result = server.wait().await;
    reaper.abort();
    tombstone_sweep.abort();
    // One last flush before the process goes. The flush COPIES rather than
    // adds, so doing it here as well as on the tick cannot double a count, and
    // it is what stops an orderly shutdown losing the last minute of today.
    if let Err(err) = flow_at_exit.flush(gunnar_core::now_ms()) {
        tracing::warn!(
            target: "gunnar::store",
            event = "flow.flush_failed",
            %err,
            "the flow ledger could not be written on shutdown; up to one flush interval of \
             today's activity counters is lost"
        );
    }
    flow_flush.abort();
    staleness_watch.abort();
    for renewal in &renewals {
        renewal.abort();
    }
    if let Some(handshake) = &handshake {
        handshake.abort();
    }
    result
}

/// The public key argument reader, exposed for the wiring tests.
#[cfg(test)]
mod tests {
    use super::read_key_argument;
    use super::*;

    /// **The banner line an operator reads off the VNC screen.**
    ///
    /// The fingerprint must be there IN FULL and unmangled: this line is the
    /// only production channel that carries it (the log goes to a serial
    /// console the UpCloud shape does not have, and the narration wire's one
    /// door withholds anything containing `sha256:`), and a fingerprint that
    /// has been shortened for display cannot be pinned by anybody.
    ///
    /// Red-first: shortening the fingerprint the way the dashboard's
    /// `short_fp` does, or dropping it from the line, fails here.
    #[test]
    fn the_banner_names_the_host_key_in_full() {
        const FP: &str = "SHA256:z/fSv0Z0RZS+Lccbc6ZoOWwt/fbj1VFJGSBKQbo4icE";
        let line = host_key_banner_line(FP);
        assert!(line.contains(FP), "{line}");
        assert!(line.starts_with("sshkey "), "one searchable token: {line}");
        assert!(!line.contains('\n'), "one line, or the banner is not a banner: {line}");
        // And it warns off the thing that fills known_hosts when this line is
        // missing, because a fingerprint with no provenance is not evidence.
        assert!(line.contains("scanned off the port proves nothing"), "{line}");
    }

    /// The visible limit has to sit **inside** the pool that carries
    /// it, or the invisible ceiling is back, at whichever of the two numbers is
    /// smaller, which is the one nobody is looking at.
    #[test]
    fn the_blocking_pool_is_larger_than_the_transfer_limit() {
        let transfers = gunnar_ssh::SshConfig::default()
            .max_concurrent_transfers
            .expect("the appliance default must have a transfer limit");
        assert!(
            transfers < GIT_BLOCKING_THREADS,
            "max_concurrent_transfers ({transfers}) is not below GIT_BLOCKING_THREADS \
             ({GIT_BLOCKING_THREADS}), so the blocking pool becomes the real ceiling again and \
             the admission counters would show a server that is never full while every transfer \
             queues"
        );
        assert!(
            GIT_BLOCKING_THREADS >= transfers * 2,
            "the blocking pool ({GIT_BLOCKING_THREADS}) leaves no headroom over \
             {transfers} concurrent transfers for the other blocking work a transfer needs: \
             pack indexing, ref transactions, the arrow log"
        );
    }

    /// The control plane must get a pool that is bounded on both sides — and
    /// at least FOUR workers (R5.1-G): two slow synchronous calls (a snapshot,
    /// a disk walk) held both workers of the old floor, and every other call —
    /// the console's page loads, monetize's `Entitlement.Set` — waited behind
    /// them.
    #[test]
    fn the_control_plane_gets_its_own_bounded_pool() {
        let workers = control_plane_workers();
        assert!(
            (4..=8).contains(&workers),
            "control_plane_workers() returned {workers}; below four, two slow synchronous calls \
             hold every worker and the whole administrative surface queues behind them, and an \
             unbounded pool recreates the starved listener with extra steps"
        );
    }

    #[test]
    fn a_literal_key_and_an_at_path_read_the_same() {
        let dir = tempfile::tempdir().expect("a temp dir");
        let path = dir.path().join("id.pub");
        std::fs::write(&path, "ssh-ed25519 AAAAC3Nz literal\n").expect("write");
        assert_eq!(
            read_key_argument("ssh-ed25519 AAAAC3Nz literal").expect("literal"),
            read_key_argument(&format!("@{}", path.display())).expect("from file"),
            "`--admin-key @file` and `--admin-key '<line>'` must produce the same key, or a \
             bootstrap that works one way fails the other"
        );
    }

    /// The 32-CPUs-of-affinity-on-8-CPUs-of-quota condition, seen **both** ways.
    ///
    /// A warning that reads its own host can only ever be observed in the state
    /// that host is in, and a build box that happens to be unconstrained would
    /// have exercised the silent branch for ever while the loud one rotted.
    #[test]
    fn a_quota_far_narrower_than_the_affinity_is_reported_and_a_matching_one_is_not() {
        // The measured bad case: `podman run --cpus 8` on a 32-core host.
        assert_eq!(
            spread_budget(Some(32), Some(8.0)),
            Some((32, 8)),
            "eight CPUs of quota spread over thirty-two is the condition that cost 2.7x the \
             CPU per request at 32 clients; it must be said out loud"
        );
        // The same budget as a cpuset: nothing to report.
        assert_eq!(spread_budget(Some(8), Some(8.0)), None);
        // Exactly at the line, and one either side of it, because `>=` and `>`
        // are one character apart and both compile.
        assert_eq!(spread_budget(Some(16), Some(8.0)), Some((16, 8)));
        assert_eq!(spread_budget(Some(15), Some(8.0)), None);
        // A fractional quota rounds UP to the CPUs an operator would name.
        assert_eq!(spread_budget(Some(32), Some(1.5)), Some((32, 2)));
        // No quota at all is the ordinary case and is not a finding. Neither is
        // an unreadable affinity: half a comparison would be a number invented
        // to fill the other half.
        assert_eq!(spread_budget(Some(32), None), None);
        assert_eq!(spread_budget(None, Some(8.0)), None);
    }

    /// `Cpus_allowed_list` is ranges *and* singletons, and the two mixed.
    ///
    /// Against literals rather than against this host: the reference desktop reports a plain
    /// `0-31`, and a parser that counted ranges and silently dropped singletons
    /// passed on this box while answering 0 for a server pinned to one CPU.
    #[test]
    fn a_cpu_list_is_counted_whether_it_is_ranges_singletons_or_both() {
        assert_eq!(count_cpu_list("0-31"), Some(32));
        assert_eq!(count_cpu_list("0-7"), Some(8));
        assert_eq!(
            count_cpu_list("3"),
            Some(1),
            "one pinned CPU is one CPU, not none"
        );
        assert_eq!(count_cpu_list("0,2,4,6"), Some(4));
        assert_eq!(count_cpu_list("0-3,8,10-11"), Some(7));
        assert_eq!(
            count_cpu_list("  0-3 , 8 \n"),
            Some(5),
            "the kernel line is padded"
        );
        // Unreadable is None, never a number: a guessed count is the one thing
        // that turns this whole check into a lie.
        assert_eq!(count_cpu_list(""), None);
        assert_eq!(count_cpu_list("banana"), None);
        assert_eq!(
            count_cpu_list("7-3"),
            None,
            "a reversed range is not 5 CPUs"
        );
    }

    #[test]
    fn the_affinity_of_this_process_is_read_as_a_count_of_cpus() {
        let n = cpus_allowed().expect("Linux always publishes Cpus_allowed_list");
        assert!(
            (1..=1 << 20).contains(&n),
            "this process reports {n} allowed CPUs, which is not a count of CPUs"
        );
        // The whole comparison rests on this being the *affinity* and not the
        // machine, so it must never exceed what the machine has.
        let online = std::thread::available_parallelism()
            .map(|n| n.get())
            .unwrap_or(1);
        assert!(
            n >= online,
            "the affinity ({n}) is narrower than what std reports as available ({online}), \
             which would make every comparison against it too quiet"
        );
    }

    #[test]
    fn a_missing_key_file_is_an_error_not_an_empty_key() {
        let err = read_key_argument("@/nonexistent/gunnar/key.pub")
            .expect_err("a missing key file must fail");
        // An empty string here would reach `bootstrap_admin` and fail there
        // with a parse error, but it would also be indistinguishable from "no
        // key was asked for" in any caller that checked for emptiness.
        assert!(format!("{err:#}").contains("reading the public key"));
    }
}

/// **The front doors' gate chain, in the one order it is ever built** (one implementation):
/// the disk-headroom gate first — every build — then, on a `metered` build
/// with a monetize key, the entitlement gate. SSH and smart-HTTP take the same
/// `Control`, so a push is refused for the same reason on both; and
/// `tests/entitlement.rs` builds the chain through this function rather than
/// by hand, so what it asserts is the wiring `serve` runs, not a copy of it.
pub fn front_control(
    stores: &crate::stores::ServerStores,
    anonymous: gunnar_ssh::AnonymousPolicy,
    disk: Arc<gunnar_ssh::DiskHeadroomGate>,
    #[cfg(feature = "metered")] entitlement: Option<&Arc<crate::entitlement::EntitlementSlot>>,
) -> gunnar_ssh::Control {
    #[allow(unused_mut)]
    // **The data set's space ledger** (cov-13): the disk gate's own, so the door
    // before the first byte and every receive loop after it count one set of
    // promises against one floor. Installed here because this is the one place
    // the chain is built.
    stores.install_space_ledger(Arc::clone(disk.ledger()));
    let control = stores
        .ssh_control()
        .with_anonymous_policy(anonymous)
        .with_plan_gate(disk as Arc<dyn gunnar_ssh::PlanGate>);
    #[cfg(feature = "metered")]
    let control = match entitlement {
        Some(slot) => control.with_plan_gate(Arc::clone(slot) as Arc<dyn gunnar_ssh::PlanGate>),
        None => control,
    };
    // **And the same chain for a mirror's worker** (2026-10-02): a mirror sync
    // grows a namespace exactly as a push does, so it is held to the same
    // room by the same gates — `mirror_fetch::room_for` reads what is
    // installed here. One chain, built in this one place.
    stores.install_plan_gates(control.plan_gates().to_vec());
    control
}

/// **The replication door's address, written where `twin certify` reads it**:
/// the bound port always; the HOSTS only while a second door is open on the
/// tunnel — then `<tunnel>,<utility>`, the twin's order of preference (it
/// dials the first that answers and falls back to the next). With no tunnel
/// door nothing is recorded and the twin dials the host it was joined to.
pub fn record_replication_door(root: &std::path::Path, utility: SocketAddr, tunnel: Option<SocketAddr>) -> Result<()> {
    crate::twin_tls::record_replication_port(root, Some(utility.port()))?;
    let hosts: Vec<std::net::IpAddr> = match tunnel {
        Some(t) => [t.ip(), utility.ip()].into_iter().filter(|ip| !ip.is_unspecified()).collect(),
        None => vec![],
    };
    crate::twin_tls::record_replication_host(root, &hosts)
}

/// **The mutual replication listener** (owner 2026-09-22, amended 2026-09-24):
/// serves on `listener` — the UTILITY address — until the box stops, and opens
/// a SECOND door on the tunnel address while there is one
/// ([`crate::tunnel_addr::replication_extra`]), closing it when the tunnel
/// goes away or moves. The utility door is never taken away: raising wg0
/// happens before any handshake, and a tunnel that never answers must not
/// cost the twin the path that works.
#[allow(clippy::too_many_arguments)]
pub async fn replication_door(
    listener: tokio::net::TcpListener,
    base: SocketAddr,
    explicit: bool,
    wg_only: bool,
    ctx: gunnar_grpc::ServiceContext,
    request_ids: gunnar_grpc::RequestIdSource,
    trust: gunnar_grpc::Trust,
    root: std::path::PathBuf,
    stop: tokio::sync::watch::Receiver<bool>,
) {
    // The door's certificate is the twin's PAIRING: one a principal no longer
    // holds (a nuked twin's, superseded when its replacement was certified —
    // twin reborn 2026-09-30) is refused here even when a header rides along.
    let mut ctx = ctx;
    ctx.config = ctx.config.clone().with_registered_client_certificate_required(true);
    let serve = move |listener: tokio::net::TcpListener, mut stop: tokio::sync::watch::Receiver<bool>| {
        let router = gunnar_grpc::router_with_trust(ctx.clone(), request_ids, Some(&trust));
        async move {
            let router = match router {
                Ok(r) => r,
                Err(err) => {
                    tracing::error!(%err, "the mutual replication listener could not be configured; no twin can follow this box");
                    return;
                }
            };
            let at = listener.local_addr().map(|a| a.to_string()).unwrap_or_default();
            let served = router
                .serve_with_incoming_shutdown(tokio_stream::wrappers::TcpListenerStream::new(listener), async move {
                    let _ = stop.wait_for(|s| *s).await;
                })
                .await;
            if let Err(err) = served {
                tracing::error!(%err, %at, "a mutual replication door stopped");
            }
        }
    };
    replication_doors_wg(listener, base, explicit, wg_only, root, crate::tunnel_addr::subscribe(), stop, serve).await
}

/// [`replication_door`]'s rule with the serving and the tunnel watch handed in
/// — the tests drive it with a plain accept loop and their own tunnel channel.
pub async fn replication_doors<S, F>(
    listener: tokio::net::TcpListener,
    base: SocketAddr,
    explicit: bool,
    root: std::path::PathBuf,
    tunnel: tokio::sync::watch::Receiver<Option<std::net::Ipv4Addr>>,
    stop: tokio::sync::watch::Receiver<bool>,
    serve: S,
) where
    S: Fn(tokio::net::TcpListener, tokio::sync::watch::Receiver<bool>) -> F,
    F: std::future::Future<Output = ()> + Send + 'static,
{
    replication_doors_wg(listener, base, explicit, false, root, tunnel, stop, serve).await
}

/// [`replication_doors`] under a posture: `wg_only` closes the utility door
/// and serves the tunnel's alone ([`doors`]); the twin is told the tunnel host
/// only.
#[allow(clippy::too_many_arguments)]
pub async fn replication_doors_wg<S, F>(
    listener: tokio::net::TcpListener,
    base: SocketAddr,
    explicit: bool,
    wg_only: bool,
    root: std::path::PathBuf,
    tunnel: tokio::sync::watch::Receiver<Option<std::net::Ipv4Addr>>,
    stop: tokio::sync::watch::Receiver<bool>,
    serve: S,
) where
    S: Fn(tokio::net::TcpListener, tokio::sync::watch::Receiver<bool>) -> F,
    F: std::future::Future<Output = ()> + Send + 'static,
{
    let record = move |utility: SocketAddr, door: Option<SocketAddr>| {
        let at = if wg_only { record_replication_door_wg_only(&root, utility.port(), door) } else { record_replication_door(&root, utility, door) };
        if let Err(err) = at {
            tracing::error!(%err, "the replication doors changed but their addresses could not be recorded");
        }
    };
    doors(listener, base, wg_only, tunnel, stop, serve, move |b, t| crate::tunnel_addr::replication_extra(b, explicit, t), record, "mutual replication listener").await
}

/// **The replication door under `--wg-only`, recorded**: the port, and the
/// tunnel's address as the ONE host (none while there is no tunnel: the twin
/// dials the host it was joined to, the primary's overlay address).
pub fn record_replication_door_wg_only(root: &std::path::Path, port: u16, tunnel: Option<SocketAddr>) -> Result<()> {
    crate::twin_tls::record_replication_port(root, Some(port))?;
    let hosts: Vec<std::net::IpAddr> = tunnel.map(|t| t.ip()).into_iter().collect();
    crate::twin_tls::record_replication_host(root, &hosts)
}

/// **The control plane's doors** (`:50051`). Not `wg_only`: the listener as
/// bound serves for the process lifetime, as it always has. `wg_only` (owner,
/// 2026-10-04): the bound — utility — listener is CLOSED, so a plain-IP
/// connection is refused by the kernel, and a door on the tunnel's address,
/// same port, serves while there is a tunnel ([`doors`]).
pub async fn control_doors<S, F>(
    listener: tokio::net::TcpListener,
    base: SocketAddr,
    wg_only: bool,
    tunnel: tokio::sync::watch::Receiver<Option<std::net::Ipv4Addr>>,
    stop: tokio::sync::watch::Receiver<bool>,
    serve: S,
) where
    S: Fn(tokio::net::TcpListener, tokio::sync::watch::Receiver<bool>) -> F,
    F: std::future::Future<Output = ()> + Send + 'static,
{
    doors(listener, base, wg_only, tunnel, stop, serve, |_, _| None, |_, _| {}, "control plane").await
}

/// **One rule for a listener and the tunnel**, shared by the replication
/// listener and the control plane.
///
/// * not `wg_only`: `listener` (the utility address) serves until `stop`;
///   `extra` names a SECOND door on the tunnel while there is one (the
///   replication rule, [`crate::tunnel_addr::replication_extra`]) or none;
/// * `wg_only`: `listener` is CLOSED at once — nothing serves on the utility
///   address, and the kernel refuses a plain-IP connection — and a door on the
///   tunnel's address, same port, serves while there is a tunnel, moving with
///   it.
///
/// `record` hears every change: `(the utility address, the tunnel door)`.
#[allow(clippy::too_many_arguments)]
pub async fn doors<S, F, X, R>(
    listener: tokio::net::TcpListener,
    base: SocketAddr,
    wg_only: bool,
    mut tunnel: tokio::sync::watch::Receiver<Option<std::net::Ipv4Addr>>,
    mut stop: tokio::sync::watch::Receiver<bool>,
    serve: S,
    extra: X,
    mut record: R,
    what: &'static str,
) where
    S: Fn(tokio::net::TcpListener, tokio::sync::watch::Receiver<bool>) -> F,
    F: std::future::Future<Output = ()> + Send + 'static,
    X: Fn(SocketAddr, Option<std::net::Ipv4Addr>) -> Option<SocketAddr>,
    R: FnMut(SocketAddr, Option<SocketAddr>),
{
    // The port actually bound (a `:0` bind in a test picks one).
    let port = listener.local_addr().map(|a| a.port()).unwrap_or(base.port());
    let base = SocketAddr::new(base.ip(), port);
    let utility = if wg_only {
        drop(listener);
        tracing::warn!(%base, "{what}: --wg-only — the utility door is CLOSED; it answers only on the WireGuard tunnel's address");
        None
    } else {
        Some(tokio::spawn(serve(listener, stop.clone())))
    };
    // The open tunnel door: its address and the switch that closes it.
    let mut door: Option<(SocketAddr, tokio::sync::watch::Sender<bool>)> = None;
    let mut first = true;
    loop {
        let t = *tunnel.borrow_and_update();
        let want = if wg_only { t.map(|t| SocketAddr::new(std::net::IpAddr::V4(t), port)) } else { extra(base, t) };
        if first || door.as_ref().map(|(a, _)| *a) != want {
            first = false;
            if let Some((at, close)) = door.take() {
                let _ = close.send(true);
                if wg_only {
                    tracing::warn!(%at, "{what}: the tunnel went away or moved: its door is closed, and with --wg-only nothing serves until it is back");
                } else {
                    tracing::warn!(%at, "{what}: the tunnel went away or moved: the door on it is closed; the utility door {base} keeps serving");
                }
            }
            if let Some(at) = want {
                match tokio::net::TcpListener::bind(at).await {
                    Ok(l) => {
                        let (close, closed) = tokio::sync::watch::channel(false);
                        tokio::spawn(serve(l, closed));
                        if wg_only {
                            tracing::info!(%at, "{what}: up on the WireGuard tunnel's address, and only there (--wg-only)");
                        } else {
                            tracing::warn!(%at, %base, "{what} ALSO on the tunnel's address (no restart); the utility door stays");
                        }
                        door = Some((at, close));
                    }
                    Err(err) => tracing::error!(%err, %at, "{what}: the tunnel's door could not bind"),
                }
            }
            record(base, door.as_ref().map(|(a, _)| *a));
        }
        tokio::select! {
            _ = async { let _ = stop.wait_for(|s| *s).await; } => break,
            changed = tunnel.changed() => {
                if changed.is_err() {
                    let _ = stop.wait_for(|s| *s).await;
                    break;
                }
            }
        }
    }
    if let Some((_, close)) = door.take() {
        let _ = close.send(true);
    }
    if let Some(u) = utility {
        let _ = u.await;
    }
}

#[cfg(test)]
mod replication_door_tests {
    use super::*;
    use std::net::Ipv4Addr;
    use tokio::io::AsyncWriteExt as _;

    /// A door that answers every connection with one byte, until closed.
    async fn answering(listener: tokio::net::TcpListener, mut stop: tokio::sync::watch::Receiver<bool>) {
        loop {
            tokio::select! {
                _ = async { let _ = stop.wait_for(|s| *s).await; } => return,
                got = listener.accept() => if let Ok((mut c, _)) = got { let _ = c.write_all(b"R").await; },
            }
        }
    }

    /// Does a door answer at `at`? A closed one refuses at once; the 10 s bounds are hang guards
    /// for a door that accepted and never wrote, long enough that a loaded box's slow answer is
    /// still an answer (2 s was not, at load 16 under `nice`).
    async fn answers(at: SocketAddr) -> bool {
        use tokio::io::AsyncReadExt as _;
        let Ok(Ok(mut c)) = tokio::time::timeout(std::time::Duration::from_secs(10), tokio::net::TcpStream::connect(at)).await else { return false };
        let mut b = [0u8; 1];
        matches!(tokio::time::timeout(std::time::Duration::from_secs(10), c.read_exact(&mut b)).await, Ok(Ok(_)))
    }

    async fn settle() {
        tokio::time::sleep(std::time::Duration::from_millis(150)).await;
    }

    /// **Wait for the doors to act**, not for a fixed time (R5.1, 2026-10-07). The doors act on a
    /// change when their task next runs, and on a loaded box (load 32 on dwarf) that was not
    /// within the 150 ms `settle` these asserts used to trust: "the tunnel door answers" failed
    /// with the door about to open. `cond` is asked every 20 ms until it holds; the two minutes
    /// are a hang guard (a test that takes under a second alone took 30 s and more at load 16
    /// under `nice`), so only a door that never acts fails.
    async fn eventually<F, Fut>(mut cond: F) -> bool
    where
        F: FnMut() -> Fut,
        Fut: std::future::Future<Output = bool>,
    {
        let deadline = std::time::Instant::now() + std::time::Duration::from_secs(120);
        loop {
            if cond().await {
                return true;
            }
            if std::time::Instant::now() >= deadline {
                return false;
            }
            tokio::time::sleep(std::time::Duration::from_millis(20)).await;
        }
    }

    /// **WireGuard-only (owner 2026-10-04)**: the utility door is CLOSED — a
    /// plain-IP connection to the utility address is refused — and only the
    /// tunnel door answers, recorded as the one host the twin is told.
    #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
    async fn under_wg_only_the_utility_door_is_closed_and_only_the_tunnel_door_answers() {
        let dir = tempfile::tempdir().unwrap();
        let root = dir.path().to_path_buf();
        let utility = tokio::net::TcpListener::bind("127.0.0.1:0").await.unwrap();
        let base = utility.local_addr().unwrap();
        let (tunnel_tx, tunnel_rx) = tokio::sync::watch::channel(None::<Ipv4Addr>);
        let (stop_tx, stop_rx) = tokio::sync::watch::channel(false);
        let doors = tokio::spawn(replication_doors_wg(utility, base, false, true, root.clone(), tunnel_rx, stop_rx, answering));
        let refused = |at: SocketAddr| async move { matches!(tokio::net::TcpStream::connect(at).await, Err(e) if e.kind() == std::io::ErrorKind::ConnectionRefused) };

        assert!(eventually(|| refused(base)).await, "no tunnel: the utility door is closed, not served");
        assert_eq!(crate::twin_tls::replication_host(&root), None);

        let t = Ipv4Addr::new(127, 0, 0, 2);
        tunnel_tx.send_replace(Some(t));
        let on_tunnel = SocketAddr::new(t.into(), base.port());
        assert!(eventually(|| answers(on_tunnel)).await, "the tunnel door answers");
        assert!(eventually(|| refused(base)).await, "and a plain-IP connection is refused");
        assert!(
            eventually(|| std::future::ready(crate::twin_tls::replication_host(&root).as_deref() == Some("127.0.0.2"))).await,
            "the twin is told the tunnel alone: {:?}",
            crate::twin_tls::replication_host(&root)
        );

        // The CONTROL PLANE follows the same rule on its own port.
        let control = tokio::net::TcpListener::bind("127.0.0.1:0").await.unwrap();
        let cbase = control.local_addr().unwrap();
        let (_ctx, crx) = tokio::sync::watch::channel(Some(t));
        let (cstop_tx, cstop_rx) = tokio::sync::watch::channel(false);
        let cdoors = tokio::spawn(control_doors(control, cbase, true, crx, cstop_rx, answering));
        assert!(eventually(|| answers(SocketAddr::new(t.into(), cbase.port()))).await, "the control plane answers on the tunnel");
        assert!(eventually(|| refused(cbase)).await, "and refuses the plain IP");

        stop_tx.send_replace(true);
        cstop_tx.send_replace(true);
        tokio::time::timeout(std::time::Duration::from_secs(5), doors).await.expect("the doors stop").unwrap();
        tokio::time::timeout(std::time::Duration::from_secs(5), cdoors).await.expect("the control doors stop").unwrap();
    }

    /// Fix 2026-09-24: raising the tunnel no longer MOVES the replication
    /// listener off the utility address. With the tunnel down, not bindable, up,
    /// and gone again, the utility door answers throughout; the tunnel door is
    /// an addition, recorded `<tunnel>,<utility>` only while it is open.
    #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
    async fn the_utility_door_answers_whatever_the_tunnel_does() {
        let dir = tempfile::tempdir().unwrap();
        let root = dir.path().to_path_buf();
        let utility = tokio::net::TcpListener::bind("127.0.0.1:0").await.unwrap();
        let base = utility.local_addr().unwrap();
        let (tunnel_tx, tunnel_rx) = tokio::sync::watch::channel(None::<Ipv4Addr>);
        let (stop_tx, stop_rx) = tokio::sync::watch::channel(false);
        let doors = tokio::spawn(replication_doors(utility, base, false, root.clone(), tunnel_rx, stop_rx, answering));

        // Tunnel down: the utility door answers, no host recorded.
        assert!(eventually(|| answers(base)).await);
        assert!(
            eventually(|| std::future::ready(crate::twin_tls::replication_port(&root) == Some(base.port()))).await,
            "the twin is told the utility port: {:?}",
            crate::twin_tls::replication_port(&root)
        );
        assert_eq!(crate::twin_tls::replication_host(&root), None);

        // A tunnel address this box cannot bind (raised elsewhere / not yet on
        // the interface): the utility door still answers.
        tunnel_tx.send_replace(Some(Ipv4Addr::new(10, 254, 254, 254)));
        settle().await;
        assert!(eventually(|| answers(base)).await, "a tunnel door that cannot bind must not cost the utility door");
        assert_eq!(crate::twin_tls::replication_host(&root), None);

        // Tunnel up: BOTH answer; the twin is told tunnel first, utility next.
        let t = Ipv4Addr::new(127, 0, 0, 2);
        tunnel_tx.send_replace(Some(t));
        let on_tunnel = SocketAddr::new(t.into(), base.port());
        assert!(eventually(|| answers(on_tunnel)).await, "the tunnel door answers");
        assert!(answers(base).await, "and the utility door still does");
        assert!(
            eventually(|| std::future::ready(crate::twin_tls::replication_host(&root).as_deref() == Some("127.0.0.2,127.0.0.1"))).await,
            "the twin is told tunnel first, utility next: {:?}",
            crate::twin_tls::replication_host(&root)
        );

        // Tunnel gone: its door closes, the utility door stays.
        tunnel_tx.send_replace(None);
        assert!(eventually(|| async move { !answers(on_tunnel).await }).await, "the tunnel door is closed");
        assert!(answers(base).await);
        assert!(
            eventually(|| std::future::ready(crate::twin_tls::replication_host(&root).is_none())).await,
            "no host once the tunnel is gone: {:?}",
            crate::twin_tls::replication_host(&root)
        );

        stop_tx.send_replace(true);
        tokio::time::timeout(std::time::Duration::from_secs(5), doors).await.expect("the doors stop").unwrap();
    }
}