tollgate_server/

lib.rs

//! The control-plane HTTP service.
//!
//! Latency-tolerant by design — its contract is correctness (no double-spend,
//! fencing, idempotent ingest), enforced by whatever [`tollgate_store`] backend
//! it is built over. The service is generic over that backend: anything
//! satisfying [`Backend`] serves identically, which is the
//! pluggability seam the Postgres backend drops into.
//!
//! Surface (`/v1`):
//! - `POST /v1/leases/acquire`, `POST /v1/leases/release`,
//!   `POST /v1/leases/consolidate`, `POST /v1/leases/reclaim` — fenced lease
//!   lifecycle.
//! - `GET /v1/keys` — revisioned active credential pages (instance authority only).
//! - `GET /v1/snapshots` — the principal catalogue (GL-48).
//! - `GET /v1/snapshots/{principal}` — compiled snapshot fetch.
//! - `POST /v1/usage/ingest` — idempotent usage batches.
//! - `POST /v1/admin/accounts`, `POST /v1/admin/accounts/{id}/deposit`,
//!   `POST /v1/admin/accounts/{id}/status`,
//!   `POST /v1/admin/accounts/{id}/capacity-class`,
//!   `POST/GET /v1/admin/accounts/{id}/keys`,
//!   `DELETE /v1/admin/accounts/{id}/keys/{key}`,
//!   `PUT/DELETE /v1/admin/accounts/{id}/keys/{key}/snapshot` (GL-143),
//!   `PUT/DELETE /v1/admin/snapshots/{principal}` — administration.
//! - `GET /livez`, `GET /readyz` — probes.
//!
//! All protected handlers require verified instance or operator evidence.
//! [`serve`] owns TLS and refuses exposed plaintext; [`security::ServerSecurity`]
//! atomically rotates roles, credentials and TLS. Administrative audit events
//! carry receipts captured by the backend at mutation. See
//! `docs/CONTROL_PLANE_SECURITY.md` for deployment and audit collection.

pub mod config;
pub mod error;
pub mod google;
mod maintenance;
pub mod security;
pub mod transport;

use std::sync::Arc;

use axum::extract::{DefaultBodyLimit, State};
use axum::http::StatusCode;
use axum::routing::{get, post, put};
use axum::{Json, Router};
use tracing::Instrument as _;

use tollgate_auth::CredentialIssuer;
use tollgate_core::{AccountId, CapacityClass, CostUnits, KeyId, Principal, PublishableSnapshot};
use tollgate_store::wire::{
    API_PREFIX, AccountKeyResponse, AccountKeysResponse, AccountResponse, AcquireRequest,
    AcquireResponse, ConsolidateRequest, ConsolidateResponse, CreateAccountRequest, DepositRequest,
    IngestRequest, IssueKeyRequest, IssuedKeyResponse, MAX_INGEST_BODY_BYTES,
    MAX_SNAPSHOT_BODY_BYTES, PrincipalsResponse, PublishSnapshotRequest, ReleaseRequest,
    RevokeKeyResponse, SetBudgetRequest, SetBudgetResponse, SetCapacityClassRequest,
    SetStatusRequest, SetStatusResponse,
};
use tollgate_store::{
    AccountConfig, AdminState, AdminStore, Clock, DEFAULT_KEY_PAGE_LIMIT,
    DEFAULT_RECLAIM_BATCH_LIMIT, DEFAULT_ROLLOVER_BATCH_LIMIT, IngestReport, KeyDirectory,
    KeyRecord, KeySource, LeaseAllocator, ReclaimBatch, ReclaimedLease, SnapshotResolution,
    SnapshotSource, StoreError, StoreHealth, UsageSink,
};

use crate::error::{ApiError, ApiJson, ApiPath, ApiQuery};
use crate::security::{Authorization, InstanceIdentity, OperatorIdentity, Role, ServerSecurity};

/// Everything the handlers need. `S` is the storage backend; the clock is the
/// single place wall time enters the server.
pub struct ServerState<S> {
    pub store: Arc<S>,
    pub clock: Arc<dyn Clock>,
    pub security: Arc<ServerSecurity>,
    /// Who may mint credentials, if this deployment issues them at all (GL-121).
    ///
    /// `None` is a deliberate answer, not an omission: an instance that only
    /// verifies has no business holding the capability to create credentials,
    /// and the issuance routes answer `501` rather than pretending. That is the
    /// shape `list_principals` already uses for a backend that cannot
    /// enumerate — an unsupported capability is reported, never faked.
    pub issuer: Option<Arc<dyn CredentialIssuer + Send + Sync>>,
}

impl<S> Clone for ServerState<S> {
    fn clone(&self) -> Self {
        ServerState {
            store: Arc::clone(&self.store),
            clock: Arc::clone(&self.clock),
            security: Arc::clone(&self.security),
            issuer: self.issuer.clone(),
        }
    }
}

/// The full store bound the server needs from a backend.
pub trait Backend:
    LeaseAllocator
    + SnapshotSource
    + UsageSink
    + AdminStore
    + KeySource
    // A server administers credentials as well as projecting them (GL-121), so
    // its backend must be the credential *directory* and not only a source.
    // `HttpStore` implements `KeySource` but not this — correctly: it is a
    // client of a server, never the authority behind one.
    + KeyDirectory
    + StoreHealth
    + Send
    + Sync
    + 'static
{
}
impl<T> Backend for T where
    T: LeaseAllocator
        + SnapshotSource
        + UsageSink
        + AdminStore
        + KeySource
        + KeyDirectory
        + StoreHealth
        + Send
        + Sync
        + 'static
{
}

/// In-process handler router. Without an owned maintenance task `/readyz`
/// returns 503 even when the store answers; use [`serve`] for a ready service.
pub fn router<S: Backend>(state: ServerState<S>) -> Router {
    router_with_maintenance(state, maintenance::Monitor::unmanaged())
}

fn router_with_maintenance<S: Backend>(
    state: ServerState<S>,
    maintenance: maintenance::Monitor,
) -> Router {
    let authorization = |role| Authorization {
        security: Arc::clone(&state.security),
        clock: Arc::clone(&state.clock),
        role,
    };
    let instance = Router::new()
        .route("/leases/acquire", post(acquire::<S>))
        .route("/leases/release", post(release::<S>))
        .route("/leases/consolidate", post(consolidate::<S>))
        .route("/leases/reclaim", post(reclaim::<S>))
        .route("/snapshots", get(list_principals::<S>))
        .route("/keys", get(active_keys::<S>))
        .route("/snapshots/{principal}", get(fetch_snapshot::<S>))
        .route(
            "/usage/ingest",
            post(ingest::<S>).layer(DefaultBodyLimit::max(MAX_INGEST_BODY_BYTES)),
        )
        .route_layer(axum::middleware::from_fn_with_state(
            authorization(Role::Instance),
            security::authorize,
        ));
    let operator = Router::new()
        .route("/accounts", post(create_account::<S>))
        .route("/accounts/{account}", get(account::<S>))
        .route("/accounts/{account}/budget", put(set_budget::<S>))
        .route(
            "/accounts/{account}/keys",
            post(issue_key::<S>).get(list_account_keys::<S>),
        )
        .route(
            "/accounts/{account}/keys/{key}",
            axum::routing::delete(revoke_account_key::<S>),
        )
        .route(
            "/accounts/{account}/keys/{key}/snapshot",
            put(publish_key_snapshot::<S>)
                .delete(remove_key_snapshot::<S>)
                .layer(DefaultBodyLimit::max(MAX_SNAPSHOT_BODY_BYTES)),
        )
        .route("/accounts/{account}/deposit", post(deposit::<S>))
        .route("/accounts/{account}/status", post(set_status::<S>))
        .route(
            "/accounts/{account}/capacity-class",
            post(set_capacity_class::<S>),
        )
        .route(
            "/snapshots/{principal}",
            put(publish_snapshot::<S>)
                .delete(remove_snapshot::<S>)
                .layer(DefaultBodyLimit::max(MAX_SNAPSHOT_BODY_BYTES)),
        )
        .route_layer(axum::middleware::from_fn_with_state(
            authorization(Role::Operator),
            security::authorize,
        ));
    Router::new()
        .route("/livez", get(async || StatusCode::OK))
        .route(
            "/readyz",
            get(readyz::<S>).layer(axum::Extension(maintenance)),
        )
        .nest(API_PREFIX, instance.nest("/admin", operator))
        .with_state(state)
        .layer(axum::middleware::from_fn(error::report_http_failure))
}

/// Serve until `shutdown` resolves, running the maintenance sweep every
/// `reclaim_interval` in the background (INVARIANTS.md GL-9's server half).
pub async fn serve<S: Backend>(
    listener: tokio::net::TcpListener,
    state: ServerState<S>,
    reclaim_interval: std::time::Duration,
    shutdown: impl Future<Output = ()> + Send + 'static,
) -> std::io::Result<()> {
    if reclaim_interval.is_zero()
        || tokio::time::Instant::now()
            .checked_add(reclaim_interval)
            .is_none()
    {
        return Err(std::io::Error::new(
            std::io::ErrorKind::InvalidInput,
            "reclaim_interval must be positive and fit the monotonic clock",
        ));
    }
    let listener = transport::SecureListener::new(listener, Arc::clone(&state.security))?;
    let sweep_store = Arc::clone(&state.store);
    let sweep_clock = Arc::clone(&state.clock);
    let mut sweeper = maintenance::Task::spawn(move |health| {
        maintenance_sweep(sweep_store, sweep_clock, reclaim_interval, health).instrument(
            tracing::info_span!(
                "maintenance_sweep",
                interval_ms = reclaim_interval.as_millis()
            ),
        )
    });
    // Axum spawns its signal waiter. Hand it only a receiver; retaining the
    // sender and the caller's future here makes every serve exit close the
    // signal, including cancellation before the caller requests shutdown.
    let (signal, signalled) = tokio::sync::oneshot::channel::<()>();
    let mut signal = Some(signal);
    let server = axum::serve(
        listener,
        router_with_maintenance(state, sweeper.monitor.clone())
            .into_make_service_with_connect_info::<transport::PeerIdentity>(),
    )
    .with_graceful_shutdown(async move {
        // Sender drop is also a shutdown signal: it means serve exited.
        let Ok(()) = signalled.await else { return };
    });
    let server = std::future::IntoFuture::into_future(server);
    tokio::pin!(server, shutdown);
    loop {
        tokio::select! {
            biased;
            result = &mut sweeper.join => {
                if result.as_ref().is_err_and(|error| error.is_cancelled())
                    && sweeper.stop.requested()
                {
                    // Expected cancellation after the graceful-shutdown signal.
                    // Readiness was withdrawn before the abort; HTTP may drain.
                    return server.await;
                }
                sweeper.stop.stop();
                drop(signal.take());
                let reason = match result {
                    Err(error) if error.is_panic() => "panic",
                    Err(_) => "cancelled",
                    Ok(()) => "unexpected-return",
                };
                tracing::error!(
                    operation = "maintenance",
                    reason,
                    "server maintenance task stopped; stopping the listener"
                );
                return Err(std::io::Error::other("server maintenance task stopped unexpectedly"));
            }
            result = &mut server => return result,
            () = &mut shutdown, if signal.is_some() => {
                sweeper.stop.stop();
                drop(signal.take());
            }
        }
    }
}

/// Reclaim expired leases and roll due budget periods forever, on the
/// configured interval.
///
/// This is INVARIANTS.md GL-9's server half, and it is the only thing that
/// returns units stranded by a crashed holder. A store can serve `ping` and
/// fail maintenance, so each outcome is published to readiness and reported
/// independently. Silence about success would
/// leave "the sweep is running but finding nothing" and "the sweep stopped"
/// indistinguishable.
///
/// The rollover pass (GL-97) shares this tick rather than owning a timer,
/// because it needs the same three things and gets them here already: a frozen
/// cutoff, a bounded drain, and a failure that is reported rather than
/// swallowed. It is also the only trigger — without it a schedule is a stored
/// intention nothing ever acts on — and a boundary the pass misses is not lost
/// but late, since the store crosses it on the next tick that sees it. The two
/// passes are independent: one failing must not stop the other, because a
/// stuck rollover stranding quota would be a strictly worse outcome than a
/// late allowance.
async fn maintenance_sweep<S: Backend>(
    store: Arc<S>,
    clock: Arc<dyn Clock>,
    interval: std::time::Duration,
    mut health: maintenance::Publisher,
) {
    #[derive(Default)]
    struct Progress {
        leases: u128,
        units: u128,
        batches: u64,
    }

    impl Progress {
        fn record(&mut self, batch: &ReclaimBatch) -> Result<(), StoreError> {
            let batch_leases =
                u128::from(u64::try_from(batch.len()).map_err(|_| {
                    StoreError("reclaim batch lease count exceeds u64 range".into())
                })?);
            let batch_units = batch.reclaimed().iter().try_fold(0u128, |total, lease| {
                total
                    .checked_add(u128::from(lease.forfeited.get()))
                    .ok_or_else(|| StoreError("reclaim batch unit total overflow".into()))
            })?;
            let leases = self
                .leases
                .checked_add(batch_leases)
                .ok_or_else(|| StoreError("reclaim cycle lease count overflow".into()))?;
            let units = self
                .units
                .checked_add(batch_units)
                .ok_or_else(|| StoreError("reclaim cycle unit total overflow".into()))?;
            let batches = self
                .batches
                .checked_add(1)
                .ok_or_else(|| StoreError("reclaim cycle batch count overflow".into()))?;
            *self = Progress {
                leases,
                units,
                batches,
            };
            Ok(())
        }
    }

    let mut tick = tokio::time::interval(interval);
    tick.set_missed_tick_behavior(tokio::time::MissedTickBehavior::Delay);
    loop {
        tick.tick().await;
        // Freeze the cutoff for this cycle. A continuously advancing cutoff
        // could make a busy allocator feed the drain forever; a fixed one is
        // a finite backlog and still returns all quota expired at this tick.
        let now = clock.now();
        let mut progress = Progress::default();
        let outcome = loop {
            match store
                .reclaim_expired_batch(now, DEFAULT_RECLAIM_BATCH_LIMIT)
                .await
            {
                Ok(batch) => {
                    if progress.record(&batch).is_err() {
                        tracing::error!(
                            operation = "reclaim",
                            "reclaim progress counter overflowed; stopping maintenance"
                        );
                        return;
                    }
                    if !batch.is_saturated() {
                        break Ok(());
                    }
                    // MemoryStore can finish a batch without an .await that
                    // yields. Let request handlers run between backlog chunks;
                    // PostgreSQL benefits from the same explicit fairness.
                    tokio::task::yield_now().await;
                }
                Err(error) => break Err(error),
            }
        };

        let Ok(report) = health.record(maintenance::Operation::Reclaim, outcome.is_ok()) else {
            tracing::error!(
                operation = "reclaim",
                "maintenance failure counter exhausted"
            );
            return;
        };
        match outcome {
            Ok(()) => {
                if report.recovered_after > 0 {
                    tracing::info!(
                        operation = "reclaim",
                        after_failures = report.recovered_after,
                        "reclaim sweep recovered"
                    );
                }
                // A swept lease is one its holder never released, so its
                // remainder was forfeited rather than returned (GL-136). Units
                // forfeited are an operator signal: a crash, or a shutdown
                // whose release deadline lapsed.
                if progress.units > 0 {
                    tracing::warn!(
                        leases = %progress.leases,
                        forfeited_units = %progress.units,
                        batches = progress.batches,
                        "swept unreleased leases; their remainders are forfeited as provisional loss"
                    );
                } else if progress.leases > 0 {
                    tracing::info!(
                        leases = %progress.leases,
                        forfeited_units = %progress.units,
                        batches = progress.batches,
                        "swept unreleased leases with nothing left to forfeit"
                    );
                }
            }
            Err(_) => {
                macro_rules! report_failure {
                    ($level:ident) => {
                        tracing::$level!(
                            operation = "reclaim",
                            code = "storage",
                            consecutive_failures = report.failures,
                            reclaimed_leases = %progress.leases,
                            forfeited_units = %progress.units,
                            completed_batches = progress.batches,
                            "reclaim sweep failed; completed batches stay committed and remaining expired leases stay stranded until it recovers"
                        );
                    };
                }
                if report.failures >= maintenance::PERSISTENT_FAILURES {
                    report_failure!(error);
                } else {
                    report_failure!(warn);
                }
            }
        }
        if roll_due_periods(store.as_ref(), now, &mut health)
            .await
            .is_break()
        {
            return;
        }
    }
}

/// Drain the rollover pass for one tick.
///
/// Bounded per transaction and drained to a partial batch, exactly as the
/// reclaim loop above is, and for a sharper reason: every scheduled account is
/// due at the same instant, so the first tick after midnight on the 1st has
/// the whole scheduled population to cross. A single unbounded statement there
/// would hold locks across the entire account table.
///
/// Failures are reported and the tick ends. Retrying inside the tick would
/// spin against a store that is down; the next tick is the retry, and until it
/// succeeds the affected accounts keep spending last period's allowance, which
/// is late rather than wrong. Only counter exhaustion breaks maintenance;
/// a reported backend failure continues to the next scheduled retry.
async fn roll_due_periods<S: Backend>(
    store: &S,
    now: jiff::Timestamp,
    health: &mut maintenance::Publisher,
) -> std::ops::ControlFlow<()> {
    let mut accounts: u64 = 0;
    let mut batches: u64 = 0;
    loop {
        match store
            .roll_due_periods(now, DEFAULT_ROLLOVER_BATCH_LIMIT)
            .await
        {
            Ok(batch) => {
                let Some(total) = u64::try_from(batch.len())
                    .ok()
                    .and_then(|rolled| accounts.checked_add(rolled))
                    .zip(batches.checked_add(1))
                else {
                    // Unreachable short of a backend ignoring the batch limit
                    // forever, and still not a place to wrap: a counter that
                    // silently restarts would under-report a boundary that
                    // rolled more accounts than it claimed.
                    tracing::error!(
                        accounts,
                        batches,
                        "budget rollover progress counter overflowed; stopping maintenance"
                    );
                    return std::ops::ControlFlow::Break(());
                };
                (accounts, batches) = (total.0, total.1);
                if !batch.is_saturated() {
                    break;
                }
                // MemoryStore can finish a batch without an .await that
                // yields; let request handlers run between chunks.
                tokio::task::yield_now().await;
            }
            Err(_) => {
                let Ok(report) = health.record(maintenance::Operation::Rollover, false) else {
                    tracing::error!(
                        operation = "budget-rollover",
                        "maintenance failure counter exhausted"
                    );
                    return std::ops::ControlFlow::Break(());
                };
                macro_rules! report_failure {
                    ($level:ident) => {
                        tracing::$level!(
                            operation = "budget-rollover",
                            code = "storage",
                            consecutive_failures = report.failures,
                            rolled_accounts = accounts,
                            completed_batches = batches,
                            "budget rollover failed; accounts past their boundary keep last period's allowance until it recovers"
                        );
                    };
                }
                if report.failures >= maintenance::PERSISTENT_FAILURES {
                    report_failure!(error);
                } else {
                    report_failure!(warn);
                }
                return std::ops::ControlFlow::Continue(());
            }
        }
    }
    let report = health
        .record(maintenance::Operation::Rollover, true)
        .expect("a successful pass clears its failure counter without arithmetic");
    if report.recovered_after > 0 {
        tracing::info!(
            operation = "budget-rollover",
            after_failures = report.recovered_after,
            "budget rollover recovered"
        );
    }
    if accounts > 0 {
        tracing::info!(accounts, batches, "rolled budget periods");
    }
    std::ops::ControlFlow::Continue(())
}

/// Ready only when the backing store answers (review finding GL-11): a server
/// whose source of truth or maintenance is unavailable must not attract traffic.
async fn readyz<S: Backend>(
    State(state): State<ServerState<S>>,
    axum::Extension(maintenance): axum::Extension<maintenance::Monitor>,
) -> StatusCode {
    match state.store.ping().await {
        Ok(()) if maintenance.healthy() => StatusCode::OK,
        Ok(()) => StatusCode::SERVICE_UNAVAILABLE,
        Err(_) => {
            tracing::warn!(
                operation = "readiness",
                code = "storage",
                "readiness probe failed: the store did not answer"
            );
            StatusCode::SERVICE_UNAVAILABLE
        }
    }
}

async fn acquire<S: Backend>(
    _identity: InstanceIdentity,
    State(state): State<ServerState<S>>,
    ApiJson(request): ApiJson<AcquireRequest>,
) -> Result<Json<AcquireResponse>, ApiError> {
    let grant = state
        .store
        .acquire(
            request.account_id,
            request.requested,
            request.ttl.duration()?,
            state.clock.now(),
        )
        .await?;
    Ok(Json(grant))
}

async fn release<S: Backend>(
    _identity: InstanceIdentity,
    State(state): State<ServerState<S>>,
    ApiJson(request): ApiJson<ReleaseRequest>,
) -> Result<StatusCode, ApiError> {
    state
        .store
        .release(
            request.lease_id,
            request.fencing_token,
            request.unspent,
            state.clock.now(),
        )
        .await?;
    Ok(StatusCode::NO_CONTENT)
}

/// Return a lease's unspent units and re-grant against the restored balance,
/// in one backend transaction.
///
/// It is a route rather than two calls the instance could make itself for the
/// reason the trait method exists: split across the wire, another instance can
/// take the returned units in the gap, and the grant policy can hand back less
/// than was returned (`LeaseAllocator::consolidate`).
async fn consolidate<S: Backend>(
    _identity: InstanceIdentity,
    State(state): State<ServerState<S>>,
    ApiJson(request): ApiJson<ConsolidateRequest>,
) -> Result<Json<ConsolidateResponse>, ApiError> {
    let grant = state
        .store
        .consolidate(
            request.lease_id,
            request.fencing_token,
            request.unspent,
            request.requested,
            request.needed,
            request.ttl.duration()?,
            state.clock.now(),
        )
        .await?;
    Ok(Json(grant))
}

async fn reclaim<S: Backend>(
    _identity: InstanceIdentity,
    State(state): State<ServerState<S>>,
) -> Result<Json<Vec<ReclaimedLease>>, ApiError> {
    Ok(Json(state.store.reclaim_expired(state.clock.now()).await?))
}

async fn fetch_snapshot<S: Backend>(
    _identity: InstanceIdentity,
    State(state): State<ServerState<S>>,
    ApiPath(principal): ApiPath<Principal>,
) -> Result<Json<Arc<tollgate_core::AccountSnapshot>>, ApiError> {
    match state.store.snapshot(principal).await? {
        SnapshotResolution::Present(snapshot) => Ok(Json(snapshot.into_inner())),
        SnapshotResolution::Revoked { generation } => Err(ApiError::revoked(generation)),
        SnapshotResolution::Unknown => Err(ApiError::not_found("unknown-principal", "no snapshot")),
    }
}

/// The principal catalogue, for instances that track every customer rather
/// than a configured slice (GL-48).
///
/// A backend that cannot enumerate answers 501 rather than an empty list: an
/// empty catalogue and an unsupported one lead an instance to opposite
/// conclusions — forget everything, or keep what you were configured with —
/// and conflating them would silently strand it on a stale set.
async fn list_principals<S: Backend>(
    _identity: InstanceIdentity,
    State(state): State<ServerState<S>>,
) -> Result<Json<PrincipalsResponse>, ApiError> {
    match state.store.principals().await? {
        Some(principals) => Ok(Json(PrincipalsResponse { principals })),
        None => Err(ApiError::not_implemented(
            "enumeration-unsupported",
            "this backend cannot list principals",
        )),
    }
}

async fn ingest<S: Backend>(
    _identity: InstanceIdentity,
    State(state): State<ServerState<S>>,
    ApiJson(request): ApiJson<IngestRequest>,
) -> Result<Json<IngestReport>, ApiError> {
    Ok(Json(
        state
            .store
            .ingest(&request.events, state.clock.now())
            .await?,
    ))
}

#[derive(serde::Deserialize)]
#[serde(deny_unknown_fields)]
struct KeyQuery {
    after: Option<tollgate_core::KeyId>,
    limit: Option<usize>,
}

async fn active_keys<S: Backend>(
    _instance: InstanceIdentity,
    State(state): State<ServerState<S>>,
    ApiQuery(query): ApiQuery<KeyQuery>,
) -> Result<impl axum::response::IntoResponse, ApiError> {
    let limit = credential_page_limit(query.limit)?;
    let page = state
        .store
        .active_keys_page(state.clock.now(), query.after, limit)
        .await
        .and_then(|page| {
            page.validate_request(query.after, limit)?;
            Ok(page)
        })
        .map_err(|_| ApiError {
            status: StatusCode::SERVICE_UNAVAILABLE,
            code: "credential-source-unavailable",
            title: "credential source is unavailable".into(),
            generation: None,
            balance_exhaustion: None,
            balance_shortfall: None,
        })?;
    let body = tollgate_store::wire::KeysResponse {
        revision: page.revision(),
        as_of: page.as_of(),
        next_after: page.next_after(),
        keys: page.into_records(),
    };
    Ok((
        [(axum::http::header::CACHE_CONTROL, "no-store")],
        Json(body),
    ))
}

async fn create_account<S: Backend>(
    operator: OperatorIdentity,
    State(state): State<ServerState<S>>,
    ApiJson(request): ApiJson<CreateAccountRequest>,
) -> Result<StatusCode, ApiError> {
    operator
        .run(
            "create_account",
            request.account_id,
            state.clock.as_ref(),
            state.store.create_account(AccountConfig {
                account_id: request.account_id,
                initial_balance: request.initial_balance,
                status: request.status,
                capacity_class: CapacityClass::Assured,
            }),
        )
        .await?;
    Ok(StatusCode::CREATED)
}

async fn deposit<S: Backend>(
    operator: OperatorIdentity,
    State(state): State<ServerState<S>>,
    ApiPath(account): ApiPath<AccountId>,
    ApiJson(request): ApiJson<DepositRequest>,
) -> Result<StatusCode, ApiError> {
    if request.units == CostUnits::ZERO {
        return Err(ApiError::bad_request(
            "zero-deposit",
            "deposit must be positive",
        ));
    }
    operator
        .run(
            "deposit",
            account,
            state.clock.as_ref(),
            state.store.deposit(account, request.units),
        )
        .await?;
    Ok(StatusCode::NO_CONTENT)
}

async fn set_status<S: Backend>(
    operator: OperatorIdentity,
    State(state): State<ServerState<S>>,
    ApiPath(account): ApiPath<AccountId>,
    ApiJson(request): ApiJson<SetStatusRequest>,
) -> Result<Json<SetStatusResponse>, ApiError> {
    // 200 with the blast radius, not 204: the ledger half is one row, but the
    // snapshot half is however many credentials the account has, and an
    // operator has no other way to learn which (GL-51).
    let change = operator
        .run(
            "set_account_status",
            account,
            state.clock.as_ref(),
            state.store.set_account_status(account, request.status),
        )
        .await?;
    Ok(Json(SetStatusResponse {
        republished: change.republished,
        unreadable: change.unreadable,
    }))
}

/// One account's administrative state, for an operator or an application
/// backend acting for its customer (GL-121).
///
/// A read, so it takes no audit receipt: the receipt convention describes
/// committed *outcomes*, and this commits nothing. It is still operator-only —
/// an instance credential cannot reach it, because it exposes an account's
/// funding position.
///
/// 404 for an unknown account rather than a zeroed body, so a caller cannot
/// read "does not exist" as "exists with no funding".
async fn account<S: Backend>(
    _operator: OperatorIdentity,
    State(state): State<ServerState<S>>,
    ApiPath(account): ApiPath<AccountId>,
) -> Result<Json<AccountResponse>, ApiError> {
    let as_of = state.clock.now();
    let view = state
        .store
        .account_view(account)
        .await?
        .ok_or_else(|| ApiError::not_found("unknown-account", "no such account"))?;
    let funding = view.conservation;
    Ok(Json(AccountResponse {
        account_id: view.account_id,
        as_of,
        status: view.status,
        capacity_class: view.capacity_class,
        budget: view.schedule,
        period_start: view.period_start,
        balance: funding.balance,
        outstanding_lease_grants: funding.active_lease_grants,
        settled_usage: funding.settled_usage,
        expired_allowance: funding.expired,
        settlement_loss: funding.settlement_loss,
        deposited: funding.deposited,
        overage_recorded: funding.overage_recorded,
    }))
}

/// Set or clear an account's periodic allowance (GL-121).
///
/// Audited through the receipt convention, so the response reports what the
/// call actually replaced rather than echoing the request back. A repeat is a
/// success with equal `previous` and `current`: the caller's intent is
/// satisfied, and saying so is more useful than a conflict.
async fn set_budget<S: Backend>(
    operator: OperatorIdentity,
    State(state): State<ServerState<S>>,
    ApiPath(account): ApiPath<AccountId>,
    ApiJson(request): ApiJson<SetBudgetRequest>,
) -> Result<Json<SetBudgetResponse>, ApiError> {
    // The receipt carries what this call actually replaced, so the outcome is
    // mapped to it rather than re-read. Reporting `request.budget` as
    // `previous` would echo the caller's own input back as history, and a
    // separate read afterwards would be a different snapshot — either way the
    // response would stop describing what committed here.
    let previous = operator
        .run(
            "set_budget_schedule",
            account,
            state.clock.as_ref(),
            async {
                state
                    .store
                    .set_budget_schedule(account, request.budget)
                    .await
                    .map(|receipt| {
                        let replaced = match receipt.before {
                            AdminState::Budget { schedule } => schedule,
                            _ => None,
                        };
                        tollgate_store::AdminReceipt::new(replaced, receipt.before, receipt.after)
                    })
            },
        )
        .await?;
    Ok(Json(SetBudgetResponse {
        previous,
        current: request.budget,
    }))
}

/// Issue one credential for an account, disclosing its secret exactly once.
///
/// **Order matters and is the contract.** The credential is minted, then
/// stored, and only then returned. A crash between minting and storing loses a
/// secret nobody has — harmless. The reverse order would hand out a credential
/// the server has never heard of, and no later reconciliation could repair it,
/// because what persists is a digest and the secret cannot be derived from it.
///
/// A lost *response* is recoverable without reissuing: the caller chose the
/// `key_id`, so resending the same request answers `409` — the credential
/// exists, and its secret is gone. The remedy is to revoke and issue a new
/// one, never to ask for the same secret again.
async fn issue_key<S: Backend>(
    operator: OperatorIdentity,
    State(state): State<ServerState<S>>,
    ApiPath(account): ApiPath<AccountId>,
    ApiJson(request): ApiJson<IssueKeyRequest>,
) -> Result<(StatusCode, Json<IssuedKeyResponse>), ApiError> {
    let issuer = state.issuer.as_ref().ok_or_else(|| {
        ApiError::not_implemented(
            "issuance-unsupported",
            "this deployment does not issue credentials",
        )
    })?;
    let minted = issuer.mint(request.key_id).map_err(|_| {
        // 503, not 500: entropy exhaustion is an operational condition the
        // caller should retry, not a defect in the request.
        ApiError {
            status: axum::http::StatusCode::SERVICE_UNAVAILABLE,
            code: "entropy-unavailable",
            title: "credential entropy unavailable".into(),
            generation: None,
            balance_exhaustion: None,
            balance_shortfall: None,
        }
    })?;
    // Disclosed exactly as minted: the text is what the digest covers, so
    // the credential the owner presents is the credential that verifies.
    // An embedder's issuer that breaks that contract is refused before the
    // record is stored, so no unpresentable credential ever exists.
    let secret = std::str::from_utf8(&minted.secret)
        .ok()
        .filter(|text| !text.is_empty() && text.bytes().all(|b| b.is_ascii_graphic()))
        .ok_or_else(|| ApiError {
            status: axum::http::StatusCode::INTERNAL_SERVER_ERROR,
            code: "issuer-misconfigured",
            title: "the configured issuer minted a credential that cannot be presented".into(),
            generation: None,
            balance_exhaustion: None,
            balance_shortfall: None,
        })?
        .to_owned();

    let record = KeyRecord {
        key_id: minted.key_id,
        account_id: account,
        principal: minted.principal,
        digest: minted.digest,
        not_after: request.not_after,
    };
    operator
        .run(
            "issue_key",
            format!("{account}/keys/{}", minted.key_id),
            state.clock.as_ref(),
            async {
                state
                    .store
                    .insert_key_within_audited(record, request.max_active_keys, state.clock.now())
                    .await
            },
        )
        .await?;

    // 201, as `create_account` answers: this made a credential that did not
    // exist. It is also what distinguishes the successful call from the 409 a
    // resent request gets, which is the signal a retrying caller reads.
    Ok((
        StatusCode::CREATED,
        Json(IssuedKeyResponse {
            key_id: minted.key_id,
            secret,
            not_after: request.not_after,
        }),
    ))
}

/// One page of an account's credentials — metadata only.
///
/// Never the secret, never the digest, and never the principal: the principal
/// is the digest's leading 128 bits, so a listing carrying it would leak half
/// of what the verifier compares against.
async fn list_account_keys<S: Backend>(
    _operator: OperatorIdentity,
    State(state): State<ServerState<S>>,
    ApiPath(account): ApiPath<AccountId>,
    ApiQuery(page): ApiQuery<KeyQuery>,
) -> Result<Json<AccountKeysResponse>, ApiError> {
    let limit = credential_page_limit(page.limit)?;
    let as_of = state.clock.now();
    let summaries = state.store.account_keys(account, page.after, limit).await?;
    // A full page may or may not be the last; a short one certainly is. Saying
    // "there may be more" costs the caller one extra request and never skips a
    // credential, which is the safe direction for a cursor.
    let next_after = (summaries.len() == limit.get())
        .then(|| summaries.last().map(|summary| summary.key_id))
        .flatten();
    Ok(Json(AccountKeysResponse {
        as_of,
        keys: summaries
            .into_iter()
            .map(|summary| AccountKeyResponse {
                key_id: summary.key_id,
                not_after: summary.not_after,
                revoked_at: summary.revoked_at,
                live: summary.is_live(as_of),
            })
            .collect(),
        next_after,
    }))
}

/// Revoke one of an account's credentials.
///
/// **Bound to the account in the path.** A `key_id` belonging to someone else
/// is answered 404, not revoked: an application backend administering its own
/// customer must not be able to retire another customer's credential by
/// guessing or mistyping an id. The check is a read of the account's own
/// listing, so it cannot be satisfied by a credential the account does not own.
async fn revoke_account_key<S: Backend>(
    operator: OperatorIdentity,
    State(state): State<ServerState<S>>,
    ApiPath((account, key)): ApiPath<(AccountId, KeyId)>,
) -> Result<Json<RevokeKeyResponse>, ApiError> {
    let owned = state
        .store
        .account_keys(account, key.0.checked_sub(1).map(KeyId), one())
        .await?
        .into_iter()
        .any(|summary| summary.key_id == key);
    if !owned {
        return Err(ApiError::not_found(
            "unknown-credential",
            "no such credential for this account",
        ));
    }
    let now = state.clock.now();
    let outcome = operator
        .run(
            "revoke_key",
            format!("{account}/keys/{key}"),
            state.clock.as_ref(),
            state.store.revoke_key_audited(key, now),
        )
        .await?;
    Ok(Json(RevokeKeyResponse {
        key_id: key,
        retired: matches!(outcome, tollgate_store::Revocation::Retired),
    }))
}

/// Publish the policy for one of an account's credentials (GL-143).
///
/// The operator names the credential by the handles it already holds; the
/// store resolves its principal, which never appears in a request, response
/// or audit. The snapshot is bound to the key in the path: an unstated
/// `key_id` is filled in. One naming a different credential is left as
/// stated, and the store refuses it rather than this handler rewriting it.
async fn publish_key_snapshot<S: Backend>(
    operator: OperatorIdentity,
    State(state): State<ServerState<S>>,
    ApiPath((account, key)): ApiPath<(AccountId, KeyId)>,
    ApiJson(request): ApiJson<PublishSnapshotRequest>,
) -> Result<StatusCode, ApiError> {
    let mut snapshot = request.snapshot;
    if snapshot.key_id.is_none() {
        Arc::make_mut(&mut snapshot).key_id = Some(key);
    }
    let snapshot = PublishableSnapshot::try_new(snapshot)?;
    operator
        .run(
            "publish_key_snapshot",
            format!("{account}/keys/{key}/snapshot"),
            state.clock.as_ref(),
            state.store.publish_key_snapshot(account, key, snapshot),
        )
        .await?;
    Ok(StatusCode::NO_CONTENT)
}

/// Withdraw the policy bound to one of an account's credentials (GL-143).
///
/// Revocation does not withdraw it, so this is part of revoking a key; it is
/// allowed for a revoked credential for exactly that reason.
async fn remove_key_snapshot<S: Backend>(
    operator: OperatorIdentity,
    State(state): State<ServerState<S>>,
    ApiPath((account, key)): ApiPath<(AccountId, KeyId)>,
) -> Result<StatusCode, ApiError> {
    operator
        .run(
            "remove_key_snapshot",
            format!("{account}/keys/{key}/snapshot"),
            state.clock.as_ref(),
            state.store.remove_key_snapshot(account, key),
        )
        .await?;
    Ok(StatusCode::NO_CONTENT)
}

/// Both credential feeds validate input before any backend read.
fn credential_page_limit(limit: Option<usize>) -> Result<std::num::NonZeroUsize, ApiError> {
    std::num::NonZeroUsize::new(limit.unwrap_or(DEFAULT_KEY_PAGE_LIMIT.get()))
        .filter(|limit| limit.get() <= tollgate_store::MAX_KEY_PAGE_LIMIT)
        .ok_or_else(|| ApiError {
            status: StatusCode::UNPROCESSABLE_ENTITY,
            code: "invalid-limit",
            title: format!(
                "credential page limit must be between 1 and {}",
                tollgate_store::MAX_KEY_PAGE_LIMIT
            ),
            generation: None,
            balance_exhaustion: None,
            balance_shortfall: None,
        })
}

fn one() -> std::num::NonZeroUsize {
    std::num::NonZeroUsize::new(1).expect("one is nonzero")
}

async fn set_capacity_class<S: Backend>(
    operator: OperatorIdentity,
    State(state): State<ServerState<S>>,
    ApiPath(account): ApiPath<AccountId>,
    ApiJson(request): ApiJson<SetCapacityClassRequest>,
) -> Result<Json<SetStatusResponse>, ApiError> {
    // 200 with the blast radius, for the reason `set_status` returns one: the
    // ledger half is one row and the snapshot half is however many credentials
    // the account has (GL-99).
    let change = operator
        .run(
            "set_capacity_class",
            account,
            state.clock.as_ref(),
            state
                .store
                .set_capacity_class(account, request.capacity_class),
        )
        .await?;
    Ok(Json(SetStatusResponse {
        republished: change.republished,
        unreadable: change.unreadable,
    }))
}

async fn publish_snapshot<S: Backend>(
    operator: OperatorIdentity,
    State(state): State<ServerState<S>>,
    ApiPath(principal): ApiPath<Principal>,
    ApiJson(request): ApiJson<PublishSnapshotRequest>,
) -> Result<StatusCode, ApiError> {
    let snapshot = PublishableSnapshot::try_new(request.snapshot)?;
    operator
        .run(
            "publish_snapshot",
            principal,
            state.clock.as_ref(),
            state.store.publish_snapshot(principal, snapshot),
        )
        .await?;
    Ok(StatusCode::NO_CONTENT)
}

async fn remove_snapshot<S: Backend>(
    operator: OperatorIdentity,
    State(state): State<ServerState<S>>,
    ApiPath(principal): ApiPath<Principal>,
) -> Result<StatusCode, ApiError> {
    operator
        .run(
            "remove_snapshot",
            principal,
            state.clock.as_ref(),
            state.store.remove_snapshot(principal),
        )
        .await?;
    Ok(StatusCode::NO_CONTENT)
}