quiver_server/

lib.rs

// SPDX-License-Identifier: AGPL-3.0-only
//! The Quiver daemon: gRPC and REST over the embeddable [`Database`], with
//! API-key auth and secure-by-default configuration.
//!
//! Both transports are thin shells over the same shared engine operations; the
//! engine is synchronous and CPU/`fsync`-bound, so every call is offloaded with
//! `spawn_blocking`. Access is guarded by a reader–writer lock (ADR-0057): reads
//! take the shared lock and run **concurrently**, writes take the exclusive lock,
//! and the single-writer model is unchanged (ADR-0002/0006). A read that finds a
//! collection's index stale (a prior write deferred its rebuild) serves the prior
//! snapshot and schedules an **off-lock** rebuild (ADR-0062): the index is rebuilt
//! under the shared lock and swapped in under a brief write lock, so a rebuild never
//! stalls concurrent readers.
//!
//! Auth is by scoped API key (Bearer / gRPC `authorization` metadata) with
//! default-deny RBAC: each key carries a role (read ⊆ write ⊆ admin) and a
//! collection scope, enforced on every operation at the shared op layer
//! (ADR-0011/0013, the `auth` module). Encryption-at-rest is on by default
//! (ADR-0010): unless `insecure` is set, an `encryption_key` is required and the
//! engine is opened through `quiver-crypto`'s AEAD codec; payloads may also be
//! client-side-encrypted (ADR-0012). TLS-in-transit uses `rustls` over the
//! audited `ring` provider — REST via `axum-server`, gRPC via tonic's `tls-ring`
//! — and a non-loopback bind requires it; setting a client CA additionally
//! requires mutual TLS. Mutating and administrative operations, and every
//! access-control denial, are recorded to an append-only audit log (ADR-0011,
//! the `audit` module) when `audit_log` is set. Per-request cost limits bound
//! the work any single authenticated request can demand (ADR-0040, the
//! [`Limits`] type), and an opt-in per-key token-bucket rate limiter bounds the
//! request *rate* (ADR-0049, the [`RateLimiter`] type); per-tenant engine
//! partitioning is a later phase. Design: `docs/api/rest-grpc.md`.

mod audit;
mod auth;
mod cluster;
mod coordinator;
mod error;
mod grpc;
mod metrics;
mod otlp;
mod rate_limit;
mod replication;
mod rest;

use std::collections::{HashMap, HashSet};
use std::future::Future;
use std::net::{IpAddr, Ipv4Addr, SocketAddr};
use std::path::PathBuf;
use std::pin::Pin;
use std::sync::{Arc, Mutex, RwLock};

use axum_server::tls_rustls::RustlsConfig;
use figment::Figment;
use figment::providers::{Env, Format, Serialized, Toml};
use serde::{Deserialize, Serialize};
use serde_json::Value;
use tokio::net::TcpListener;
use tokio::sync::broadcast;
use tokio_stream::wrappers::TcpListenerStream;
use tonic::transport::{Certificate, Identity, ServerTlsConfig};

use quiver_crypto::AeadCodec;
use quiver_embed::{
    Database, Descriptor, DistanceMetric, Dtype, FilterableField, IndexSpec, SearchParams,
    SnapshotInfo, SparseVector, TEXT_KEY, VectorEncryption, WalEntry, WalOp,
};
use quiver_query::Filter;

pub use auth::{Action, ApiKey, CollectionScope};
pub use error::Error;
pub use otlp::OtlpConfig;
// The embedding/rerank seam lives in its own lean crate (ADR-0058) so the MCP
// server can share it; re-exported here so the server's public API is unchanged.
pub use quiver_providers::{
    EmbedRegistry, EmbeddingConfig, EmbeddingProvider, ProviderError, ProviderKind, RerankConfig,
    RerankProvider,
};
pub use rate_limit::{RateDecision, RateLimitConfig, RateLimitSnapshot, RateLimiter};

use audit::{AuditLog, Outcome};
use auth::Principal;

/// Per-request cost limits (ADR-0040). Each cap bounds the work a single
/// authenticated request can demand, so one oversized request cannot exhaust the
/// node under the single-writer model (ADR-0006). Over-limit requests are
/// **rejected** with HTTP 400 / gRPC `InvalidArgument` rather than silently
/// clamped — a truncated `k` or `ef_search` would return surprising, lower-quality
/// results with no signal. Defaults are generous; raise a cap with a `[limits]`
/// table in `quiver.toml` or the matching `QUIVER_MAX_*` environment variable.
#[derive(Debug, Clone, Copy, Serialize, Deserialize)]
#[serde(default)]
pub struct Limits {
    /// Maximum `k` (top-k) for `search` / `search_multi_vector`.
    pub max_k: usize,
    /// Maximum search beam width (`ef_search`).
    pub max_ef_search: usize,
    /// Maximum `fetch` page size.
    pub max_fetch_limit: usize,
    /// Maximum vector dimension: the declared collection dimension at creation
    /// and the length of any query vector (per token, for multi-vector).
    pub max_vector_dim: usize,
    /// Maximum serialized-JSON payload size per point, in bytes.
    pub max_payload_bytes: usize,
    /// Maximum number of points / documents in a single upsert request.
    pub max_batch_size: usize,
    /// Maximum HTTP request body size, in bytes (enforced by the REST layer via
    /// axum's `DefaultBodyLimit`; gRPC is bounded by tonic's decode limit).
    pub max_request_body_bytes: usize,
    /// Maximum number of non-zero terms in a hybrid-search sparse query (ADR-0043),
    /// bounding the posting-list scan.
    pub max_sparse_terms: usize,
    /// Maximum number of points in a single bulk upsert (`POST …/points:bulk`,
    /// ADR-0045). Larger than `max_batch_size` because the bulk path defers index
    /// maintenance to one rebuild; the request is still bounded by
    /// `max_request_body_bytes`, so raise that too for very large bulk loads.
    pub max_bulk_batch_size: usize,
}

impl Default for Limits {
    fn default() -> Self {
        Self {
            max_k: 10_000,
            max_ef_search: 4_096,
            max_fetch_limit: 10_000,
            max_vector_dim: 8_192,
            max_payload_bytes: 65_536,
            max_batch_size: 1_000,
            max_request_body_bytes: 32 * 1024 * 1024,
            max_sparse_terms: 4_096,
            max_bulk_batch_size: 50_000,
        }
    }
}

impl Limits {
    // Apply `QUIVER_MAX_*` overrides after figment extraction (ADR-0013 env
    // layer). The flat env keys do not nest under figment's `limits` table, so
    // they are read explicitly here; a malformed value is a hard config error.
    fn apply_env_overrides(&mut self) -> Result<(), Error> {
        let slots: [(&str, &mut usize); 9] = [
            ("QUIVER_MAX_K", &mut self.max_k),
            ("QUIVER_MAX_EF_SEARCH", &mut self.max_ef_search),
            ("QUIVER_MAX_FETCH_LIMIT", &mut self.max_fetch_limit),
            ("QUIVER_MAX_VECTOR_DIM", &mut self.max_vector_dim),
            ("QUIVER_MAX_PAYLOAD_BYTES", &mut self.max_payload_bytes),
            ("QUIVER_MAX_BATCH_SIZE", &mut self.max_batch_size),
            (
                "QUIVER_MAX_REQUEST_BODY_BYTES",
                &mut self.max_request_body_bytes,
            ),
            ("QUIVER_MAX_SPARSE_TERMS", &mut self.max_sparse_terms),
            ("QUIVER_MAX_BULK_BATCH_SIZE", &mut self.max_bulk_batch_size),
        ];
        for (key, slot) in slots {
            if let Ok(raw) = std::env::var(key) {
                *slot = raw.parse().map_err(|_| {
                    Error::Config(format!("{key} must be a positive integer, got {raw:?}"))
                })?;
            }
        }
        Ok(())
    }

    // Reject a zero cap (a `0` limit would refuse every request).
    fn validate(&self) -> Result<(), Error> {
        let named = [
            ("max_k", self.max_k),
            ("max_ef_search", self.max_ef_search),
            ("max_fetch_limit", self.max_fetch_limit),
            ("max_vector_dim", self.max_vector_dim),
            ("max_payload_bytes", self.max_payload_bytes),
            ("max_batch_size", self.max_batch_size),
            ("max_request_body_bytes", self.max_request_body_bytes),
            ("max_sparse_terms", self.max_sparse_terms),
            ("max_bulk_batch_size", self.max_bulk_batch_size),
        ];
        if let Some((name, _)) = named.into_iter().find(|&(_, v)| v == 0) {
            return Err(Error::Config(format!(
                "limits.{name} must be greater than zero"
            )));
        }
        Ok(())
    }

    fn check_search(&self, k: usize, ef_search: usize) -> Result<(), Error> {
        if k > self.max_k {
            return Err(Error::BadRequest(format!(
                "k ({k}) exceeds the maximum of {} (raise QUIVER_MAX_K)",
                self.max_k
            )));
        }
        if ef_search > self.max_ef_search {
            return Err(Error::BadRequest(format!(
                "ef_search ({ef_search}) exceeds the maximum of {} (raise QUIVER_MAX_EF_SEARCH)",
                self.max_ef_search
            )));
        }
        Ok(())
    }

    fn check_sparse_terms(&self, n: usize) -> Result<(), Error> {
        if n > self.max_sparse_terms {
            return Err(Error::BadRequest(format!(
                "sparse query has {n} terms, exceeding the maximum of {} (raise QUIVER_MAX_SPARSE_TERMS)",
                self.max_sparse_terms
            )));
        }
        Ok(())
    }

    fn check_fetch(&self, limit: usize) -> Result<(), Error> {
        if limit > self.max_fetch_limit {
            return Err(Error::BadRequest(format!(
                "limit ({limit}) exceeds the maximum of {} (raise QUIVER_MAX_FETCH_LIMIT)",
                self.max_fetch_limit
            )));
        }
        Ok(())
    }

    fn check_dim(&self, dim: usize) -> Result<(), Error> {
        if dim > self.max_vector_dim {
            return Err(Error::BadRequest(format!(
                "dimension ({dim}) exceeds the maximum of {} (raise QUIVER_MAX_VECTOR_DIM)",
                self.max_vector_dim
            )));
        }
        Ok(())
    }

    fn check_vector_len(&self, len: usize) -> Result<(), Error> {
        if len > self.max_vector_dim {
            return Err(Error::BadRequest(format!(
                "vector length ({len}) exceeds the maximum of {} (raise QUIVER_MAX_VECTOR_DIM)",
                self.max_vector_dim
            )));
        }
        Ok(())
    }

    fn check_batch(&self, n: usize) -> Result<(), Error> {
        if n > self.max_batch_size {
            return Err(Error::BadRequest(format!(
                "batch of {n} exceeds the maximum of {} (raise QUIVER_MAX_BATCH_SIZE)",
                self.max_batch_size
            )));
        }
        Ok(())
    }

    fn check_bulk_batch(&self, n: usize) -> Result<(), Error> {
        if n > self.max_bulk_batch_size {
            return Err(Error::BadRequest(format!(
                "bulk batch of {n} exceeds the maximum of {} (raise QUIVER_MAX_BULK_BATCH_SIZE)",
                self.max_bulk_batch_size
            )));
        }
        Ok(())
    }

    fn check_payload(&self, payload: &Value) -> Result<(), Error> {
        let size = serde_json::to_vec(payload)
            .map(|v| v.len())
            .map_err(|e| Error::Internal(format!("payload serialization: {e}")))?;
        if size > self.max_payload_bytes {
            return Err(Error::BadRequest(format!(
                "payload of {size} bytes exceeds the maximum of {} (raise QUIVER_MAX_PAYLOAD_BYTES)",
                self.max_payload_bytes
            )));
        }
        Ok(())
    }
}

/// Server configuration, layered defaults → `quiver.toml` → `QUIVER_*` env and
/// validated at startup (ADR-0013).
#[derive(Debug, Clone, Serialize, Deserialize)]
#[serde(default)]
pub struct Config {
    /// Data directory for the storage engine.
    pub data_dir: PathBuf,
    /// REST (HTTP/1.1) bind address.
    pub rest_addr: SocketAddr,
    /// gRPC (HTTP/2) bind address.
    pub grpc_addr: SocketAddr,
    /// Accepted API keys with their RBAC scopes (ADR-0011). A bare secret —
    /// from a comma-separated `QUIVER_API_KEYS` (the env form) or a plain TOML
    /// array entry — is an all-collections admin key; a structured
    /// `{secret, role, collections}` entry pins a narrower scope. Empty is
    /// allowed only with `insecure = true`.
    #[serde(default, deserialize_with = "auth::de_api_keys")]
    pub api_keys: Vec<ApiKey>,
    /// Hex-encoded 256-bit **master key** for encryption-at-rest (64 hex
    /// characters). It wraps per-collection data-encryption keys (ADR-0010).
    /// Required unless `insecure = true` or [`master_key_file`] is set; source it
    /// from the environment or a secret store, never the committed config. `None`
    /// ⇒ data is stored unencrypted (only valid in `insecure` mode).
    ///
    /// [`master_key_file`]: Config::master_key_file
    pub encryption_key: Option<String>,
    /// Path to a file holding the hex master key, as an alternative to
    /// [`encryption_key`] (set exactly one). Lets the key arrive as a mounted
    /// secret (Docker/Kubernetes) or a KMS-decrypted file rather than an
    /// environment variable. It should be mode `0600`; a group/world-accessible
    /// file is warned about at startup.
    ///
    /// [`encryption_key`]: Config::encryption_key
    pub master_key_file: Option<PathBuf>,
    /// Path to the PEM-encoded TLS certificate chain. Must be set together with
    /// `tls_key`. Required for a non-loopback bind unless `insecure = true`.
    pub tls_cert: Option<PathBuf>,
    /// Path to the PEM-encoded TLS private key. Must be set together with
    /// `tls_cert`.
    pub tls_key: Option<PathBuf>,
    /// Path to a PEM-encoded CA certificate that signs accepted client
    /// certificates. When set, both transports require **mutual TLS**: a client
    /// must present a certificate chaining to this CA to connect (ADR-0011).
    /// Requires `tls_cert`/`tls_key`; bearer API keys still carry the RBAC scope.
    pub tls_client_ca: Option<PathBuf>,
    /// Path to an append-only audit log file (ADR-0011). When set, every
    /// mutating and administrative operation and every access-control denial is
    /// appended as one JSON object per line (JSON Lines); records are always
    /// also emitted as `tracing` events. Unset ⇒ tracing only. Secrets are never
    /// written — see `docs/security/audit.md`.
    pub audit_log: Option<PathBuf>,
    /// Run as a **read-replica follower** (ADR-0030): connect to a leader's gRPC
    /// endpoint at this URL (e.g. `http://leader:6334`) and continuously apply its
    /// committed operations, serving reads. A follower **refuses writes**. Unset ⇒
    /// this node is a normal read-write leader. (Plaintext `http://` for now; TLS
    /// to the leader is a follow-up — run replication over a trusted network.)
    pub leader_url: Option<String>,
    /// API key the follower presents to the leader's admin-scoped `Replicate`
    /// stream (used with `leader_url`). Source it like any secret.
    pub leader_api_key: Option<String>,
    /// Opt out of the secure defaults (no auth, no encryption-at-rest, allow a
    /// non-loopback bind without TLS). For local development only; never the
    /// default.
    pub insecure: bool,
    /// Per-request cost limits (ADR-0040). Set with a `[limits]` table in
    /// `quiver.toml` or the `QUIVER_MAX_*` environment variables.
    pub limits: Limits,
    /// Opt-in server-side embedding providers, keyed by collection name (ADR-0047).
    /// Configured with `[embedding.<collection>]` tables in `quiver.toml`; default
    /// empty, so the engine stays model-agnostic and library mode is unaffected.
    /// API keys are referenced by env-var *name* and resolved at startup, never
    /// stored. Enables `search_text` / `upsert_text` for the named collections.
    #[serde(default)]
    pub embedding: HashMap<String, EmbeddingConfig>,
    /// Opt-in server-side rerank providers, keyed by collection name (ADR-0047).
    /// Configured with `[rerank.<collection>]` tables; enables the one-call
    /// retrieve→rerank stage of `search_text`.
    #[serde(default)]
    pub rerank: HashMap<String, RerankConfig>,
    /// Opt-in per-key rate limiting (ADR-0049). Set a `[rate_limit]` table in
    /// `quiver.toml` or the `QUIVER_RATE_LIMIT_*` env vars;
    /// `requests_per_second = 0` (the default) disables it.
    #[serde(default)]
    pub rate_limit: RateLimitConfig,
    /// Opt-in OpenTelemetry traces export (ADR-0059). Set an `[otlp]` table or the
    /// `QUIVER_OTLP_*` env vars; an empty `endpoint` (the default) disables it.
    /// Export additionally requires the server's `otlp` build feature.
    #[serde(default)]
    pub otlp: OtlpConfig,
    /// Lock-free MVCC reads (ADR-0064), **experimental and default-off**. Serves
    /// reads of single-vector, in-memory collections from an `arc-swap` snapshot so
    /// pure-vector reads never block on a concurrent writer. Also settable via
    /// `QUIVER_MVCC_READS`. See `docs/adr/0064-mvcc-reads-implementation.md`.
    #[serde(default)]
    pub mvcc_reads: bool,
    /// Opt-in **cluster router** mode (ADR-0065): a non-empty list of shard base
    /// URLs (e.g. `["http://s1:6333","http://s2:6333"]`, or the bracketed
    /// `QUIVER_CLUSTER_SHARDS` env array `[http://s1:6333,http://s2:6333]`) makes
    /// this server a stateless router that shards writes and scatter-gathers
    /// searches across the shards. Empty (the default) = an ordinary single-node
    /// server.
    #[serde(default)]
    pub cluster_shards: Vec<String>,
    /// Optional per-shard **read replicas** (ADR-0065 increment 2). Each entry is
    /// `"<shard_index>=<replica_url>"` (repeat the index for several replicas of one
    /// shard), e.g. `["0=http://s1b:6333", "1=http://s2b:6333"]` or the bracketed
    /// `QUIVER_CLUSTER_REPLICAS` env list. A replica is an ordinary follower
    /// (ADR-0030) of its shard's primary; the router fans searches across
    /// `{primary} ∪ replicas` to spread read load. A shard with no entry is
    /// primary-only. Empty (the default) = no replicas.
    #[serde(default)]
    pub cluster_replicas: Vec<String>,
    /// Optional API key the router presents to its shards (a cluster runs over a
    /// trusted network). `None` = shards are unauthenticated.
    #[serde(default)]
    pub cluster_shard_key: Option<String>,
    /// Run this process as the **cluster coordinator** (ADR-0066) instead of a
    /// normal server: it owns the authoritative versioned shard map and serves the
    /// `/cluster/*` membership API, off the data path. It seeds its map from
    /// `cluster_shards`/`cluster_replicas` (or recovers it from `coordinator_state`).
    /// Default `false`.
    #[serde(default)]
    pub coordinator: bool,
    /// Base URL of the cluster coordinator (ADR-0066). When set on a **router**, the
    /// router refreshes its shard map from `GET {url}/cluster/map` on an interval, so
    /// a membership change propagates without a restart. `None` = a static map from
    /// `cluster_shards` (increments 1–2 behaviour).
    #[serde(default)]
    pub coordinator_url: Option<String>,
    /// Where the **coordinator** persists its versioned map + id counter, so a
    /// restart recovers the exact membership (and never reuses a shard id). `None` =
    /// in-memory only (lost on restart). Only meaningful with `coordinator = true`.
    #[serde(default)]
    pub coordinator_state: Option<PathBuf>,
}

impl Default for Config {
    fn default() -> Self {
        Self {
            data_dir: PathBuf::from("./quiver-data"),
            rest_addr: SocketAddr::new(IpAddr::V4(Ipv4Addr::LOCALHOST), 6333),
            grpc_addr: SocketAddr::new(IpAddr::V4(Ipv4Addr::LOCALHOST), 6334),
            api_keys: Vec::new(),
            encryption_key: None,
            master_key_file: None,
            tls_cert: None,
            tls_key: None,
            tls_client_ca: None,
            audit_log: None,
            leader_url: None,
            leader_api_key: None,
            insecure: false,
            limits: Limits::default(),
            embedding: HashMap::new(),
            rerank: HashMap::new(),
            rate_limit: RateLimitConfig::default(),
            otlp: OtlpConfig::default(),
            mvcc_reads: false,
            cluster_shards: Vec::new(),
            cluster_replicas: Vec::new(),
            cluster_shard_key: None,
            coordinator: false,
            coordinator_url: None,
            coordinator_state: None,
        }
    }
}

impl Config {
    /// Load configuration from defaults, an optional `quiver.toml`, and
    /// `QUIVER_*` environment variables.
    pub fn load() -> Result<Self, Error> {
        let mut config: Config = Figment::from(Serialized::defaults(Config::default()))
            .merge(Toml::file("quiver.toml"))
            .merge(Env::prefixed("QUIVER_"))
            .extract()
            .map_err(|e| Error::Config(e.to_string()))?;
        // The flat `QUIVER_MAX_*` env keys do not nest under the `limits` table
        // figment builds, so apply them explicitly (ADR-0040).
        config.limits.apply_env_overrides()?;
        // Same for the flat `QUIVER_RATE_LIMIT_*` keys (ADR-0049).
        config
            .rate_limit
            .apply_env_overrides()
            .map_err(Error::Config)?;
        // Same for the flat `QUIVER_OTLP_*` keys (ADR-0059).
        config.otlp.apply_env_overrides().map_err(Error::Config)?;
        Ok(config)
    }

    /// Reject insecure configurations unless explicitly opted out (ADR-0013):
    /// no anonymous access, encryption-at-rest on by default with a valid key,
    /// and no non-loopback bind without TLS.
    pub fn validate(&self) -> Result<(), Error> {
        if self.api_keys.is_empty() && !self.insecure {
            return Err(Error::Config(
                "no api_keys configured: set QUIVER_API_KEYS (comma-separated) or \
                 set insecure=true for local development"
                    .to_owned(),
            ));
        }
        // Resolve the master key from the env var or a key file (exactly one).
        let master_key = self.master_key_hex()?;
        if master_key.is_none() && !self.insecure {
            return Err(Error::Config(
                "no encryption key configured: encryption-at-rest is on by default — \
                 set QUIVER_ENCRYPTION_KEY to a 64-hex-character (256-bit) key (or \
                 QUIVER_MASTER_KEY_FILE to a file holding it), or set insecure=true to \
                 store data unencrypted (development only)"
                    .to_owned(),
            ));
        }
        // Fail fast on a malformed key rather than at first write.
        if let Some(key) = &master_key {
            AeadCodec::from_hex(key)
                .map_err(|e| Error::Config(format!("invalid master key: {e}")))?;
        }
        // TLS certificate and key are set together or not at all.
        if self.tls_cert.is_some() != self.tls_key.is_some() {
            return Err(Error::Config(
                "tls_cert and tls_key must be set together".to_owned(),
            ));
        }
        // mTLS layers on top of server TLS: a client CA needs a server cert/key.
        if self.tls_client_ca.is_some() && !(self.tls_cert.is_some() && self.tls_key.is_some()) {
            return Err(Error::Config(
                "tls_client_ca (mutual TLS) requires tls_cert and tls_key".to_owned(),
            ));
        }
        let tls_enabled = self.tls_cert.is_some() && self.tls_key.is_some();
        let non_loopback = !self.rest_addr.ip().is_loopback() || !self.grpc_addr.ip().is_loopback();
        if non_loopback && !tls_enabled && !self.insecure {
            return Err(Error::Config(
                "non-loopback bind requires TLS: set tls_cert and tls_key (PEM files), \
                 or insecure=true for local development"
                    .to_owned(),
            ));
        }
        // Reject a nonsensical cost limit (a `0` cap would refuse every request).
        self.limits.validate()?;
        Ok(())
    }

    /// The effective hex master key: from [`master_key_file`] when set (read and
    /// trimmed), otherwise [`encryption_key`]. `None` means no key is configured
    /// (only valid with `insecure`).
    ///
    /// # Errors
    /// [`Error::Config`] if both sources are set, or the key file cannot be read.
    ///
    /// [`master_key_file`]: Config::master_key_file
    /// [`encryption_key`]: Config::encryption_key
    pub(crate) fn master_key_hex(&self) -> Result<Option<String>, Error> {
        let env_key = self
            .encryption_key
            .as_deref()
            .map(str::trim)
            .filter(|k| !k.is_empty());
        match (&self.master_key_file, env_key) {
            (Some(_), Some(_)) => Err(Error::Config(
                "set either encryption_key (QUIVER_ENCRYPTION_KEY) or master_key_file \
                 (QUIVER_MASTER_KEY_FILE), not both"
                    .to_owned(),
            )),
            (Some(path), None) => {
                warn_if_world_readable(path);
                let hex = std::fs::read_to_string(path).map_err(|e| {
                    Error::Config(format!("reading master_key_file {}: {e}", path.display()))
                })?;
                Ok(Some(hex.trim().to_owned()))
            }
            (None, Some(key)) => Ok(Some(key.to_owned())),
            (None, None) => Ok(None),
        }
    }
}

// Warn (don't fail — Docker/Kubernetes secrets often mount group/world-readable)
// when a master-key file is more permissive than `0600`.
#[cfg(unix)]
fn warn_if_world_readable(path: &std::path::Path) {
    use std::os::unix::fs::PermissionsExt;
    if let Ok(meta) = std::fs::metadata(path)
        && meta.permissions().mode() & 0o077 != 0
    {
        tracing::warn!(
            path = %path.display(),
            mode = format!("{:o}", meta.permissions().mode() & 0o777),
            "master key file is group/world-accessible; restrict it to 0600"
        );
    }
}

#[cfg(not(unix))]
fn warn_if_world_readable(_path: &std::path::Path) {}

/// Shared server state: the engine behind a single-writer lock, the accepted
/// API keys with their RBAC scopes, and the audit log.
#[derive(Clone)]
pub(crate) struct AppState {
    // The engine behind a reader–writer lock (ADR-0057): reads take the shared
    // lock and run concurrently; writes take the exclusive lock. A read that finds
    // a collection's index stale upgrades to the write lock once to rebuild it,
    // then serves concurrently again.
    db: Arc<RwLock<Database>>,
    keys: Arc<Vec<ApiKey>>,
    audit: Arc<AuditLog>,
    // Fan-out of every committed op to replication followers (ADR-0030). The
    // commit observer publishes here; each `Replicate` stream subscribes.
    replication_tx: broadcast::Sender<WalEntry>,
    // True on a replication follower: external writes are refused; the engine's
    // state is owned by the stream it applies from the leader (ADR-0030).
    read_only: bool,
    // Per-request cost limits, enforced at this op layer so both transports are
    // covered by one implementation (ADR-0040).
    limits: Limits,
    // Opt-in, provider-agnostic server-side embedding/rerank providers, keyed by
    // collection (ADR-0047). Empty on the common path; `search_text`/`upsert_text`
    // require a configured embedder for the target collection.
    embed: Arc<EmbedRegistry>,
    // Opt-in per-key token-bucket rate limiter (ADR-0049). A no-op when disabled.
    rate_limiter: Arc<RateLimiter>,
    // Prometheus metrics registry (ADR-0014/0054), scraped at `GET /metrics`.
    metrics: Arc<metrics::Metrics>,
    // Collections with an off-lock index rebuild in flight (ADR-0062). A search that
    // observes a deferred rebuild schedules one here, single-flighted so concurrent
    // readers never kick off duplicate builds for the same collection.
    rebuilding: Arc<Mutex<HashSet<String>>>,
    // Lock-free MVCC read cache (ADR-0064 increment 3): cached `arc-swap` snapshot
    // cells for MVCC-served collections. A pure-vector search loads the cell and
    // searches it **without taking the database lock**, so it never blocks on a
    // concurrent writer. Populated under the read lock the first time a collection
    // is seen fresh; the map itself changes only on create/drop, never on a data
    // write. Empty unless `QUIVER_MVCC_READS` is on.
    snapshot_cells: Arc<RwLock<HashMap<String, quiver_embed::SnapshotCell>>>,
    // Snapshot of the engine's `QUIVER_MVCC_READS` flag, read once at startup. When
    // on, the writer drives off-lock consolidation rebuilds (ADR-0064 increment 3)
    // because pure-vector fast-path reads bypass the reader-driven scheduler. Off by
    // default, so the non-MVCC path is byte-identical.
    mvcc: bool,
    // Opt-in cluster router (ADR-0065): when `Some`, every collection/point op
    // fans out to the shards instead of touching the local engine. `None` = an
    // ordinary single-node server (the engine-backed path, untouched).
    cluster: Option<Arc<cluster::Cluster>>,
}

/// A collection's metadata.
pub(crate) struct CollectionInfo {
    pub name: String,
    pub dim: u32,
    pub metric: DistanceMetric,
    pub count: u64,
    pub index: IndexSpec,
    pub filterable: Vec<FilterableField>,
    pub multivector: bool,
    pub vector_encryption: VectorEncryption,
}

/// A point to upsert.
pub(crate) struct PointIn {
    pub id: String,
    pub vector: Vec<f32>,
    pub payload: Value,
}

/// A text point to embed server-side and upsert (ADR-0047).
pub(crate) struct TextPointIn {
    pub id: String,
    pub text: String,
    pub payload: Value,
}

/// Default candidate pool size a rerank stage over-fetches before reordering to
/// the requested `k` (ADR-0047).
const RERANK_CANDIDATES: usize = 50;

/// The document text a rerank stage scores: the original text stored under
/// [`TEXT_KEY`] by `upsert_text`, else the whole payload as a string so the
/// reranker still has something to compare.
fn doc_text(payload: Option<&Value>) -> String {
    match payload {
        Some(Value::Object(map)) => map
            .get(TEXT_KEY)
            .and_then(Value::as_str)
            .map_or_else(|| Value::Object(map.clone()).to_string(), str::to_owned),
        Some(v) => v.to_string(),
        None => String::new(),
    }
}

/// A fetched point.
pub(crate) struct PointOut {
    pub id: String,
    pub vector: Option<Vec<f32>>,
    pub payload: Value,
}

/// A search hit.
pub(crate) struct MatchOut {
    pub id: String,
    pub score: f32,
    pub payload: Option<Value>,
    pub vector: Option<Vec<f32>>,
}

/// A multi-vector document to upsert.
pub(crate) struct DocumentIn {
    pub id: String,
    pub vectors: Vec<Vec<f32>>,
    pub payload: Value,
}

/// A multi-vector document hit (MaxSim).
pub(crate) struct DocumentMatchOut {
    pub id: String,
    pub score: f32,
    pub payload: Option<Value>,
    pub vectors: Option<Vec<Vec<f32>>>,
}

impl AppState {
    /// Authenticate a presented bearer token to its [`Principal`], or `None`
    /// (a 401). An empty key set means `insecure` mode (validated at startup),
    /// which admits any caller as an all-collections admin.
    pub(crate) fn authenticate(&self, presented: Option<&str>) -> Option<Principal> {
        auth::authenticate(&self.keys, presented)
    }

    /// Consume one rate-limit token for `actor` (ADR-0049). A no-op `Allowed` when
    /// rate limiting is disabled. Both transports call this at their auth choke
    /// point so the limiter is enforced by one implementation.
    pub(crate) fn rate_limit(&self, actor: &str) -> RateDecision {
        self.rate_limiter.check(actor)
    }

    /// Whether the per-key rate limiter is active (lets a transport skip the work
    /// entirely on the common, disabled path).
    pub(crate) fn rate_limit_enabled(&self) -> bool {
        self.rate_limiter.enabled()
    }

    // Run a **mutating** engine op behind the exclusive write lock, off the async
    // runtime (the engine is synchronous and CPU/IO-bound). The single writer is
    // unchanged from the prior single-mutex model (ADR-0006/0057).
    async fn write_blocking<T, F>(&self, f: F) -> Result<T, Error>
    where
        T: Send + 'static,
        F: FnOnce(&mut Database) -> quiver_embed::Result<T> + Send + 'static,
    {
        let db = Arc::clone(&self.db);
        tokio::task::spawn_blocking(move || -> Result<T, Error> {
            let mut guard = db
                .write()
                .map_err(|_| Error::Internal("database lock poisoned".to_owned()))?;
            f(&mut guard).map_err(Error::Engine)
        })
        .await
        .map_err(|e| Error::Internal(format!("blocking task failed: {e}")))?
    }

    // Run a **read-only** engine op behind the shared read lock — many of these run
    // concurrently (ADR-0057). The closure gets `&Database`, so it can only call the
    // `&self` reads (`*_snapshot`, `fetch`, accessors).
    async fn read_blocking<T, F>(&self, f: F) -> Result<T, Error>
    where
        T: Send + 'static,
        F: FnOnce(&Database) -> quiver_embed::Result<T> + Send + 'static,
    {
        let db = Arc::clone(&self.db);
        tokio::task::spawn_blocking(move || -> Result<T, Error> {
            let guard = db
                .read()
                .map_err(|_| Error::Internal("database lock poisoned".to_owned()))?;
            f(&guard).map_err(Error::Engine)
        })
        .await
        .map_err(|e| Error::Internal(format!("blocking task failed: {e}")))?
    }

    // Run a snapshot read for `collection` (ADR-0057/0062). Takes only the shared
    // read lock, so concurrent searches run in parallel. When a prior write deferred
    // this collection's rebuild, the snapshot read serves the **prior** snapshot
    // (snapshot-isolated, slightly stale) rather than blocking, and we schedule an
    // **off-lock** rebuild so the next read is fresh — the reader never waits on a
    // rebuild under the exclusive lock.
    async fn search_blocking<T, F>(&self, collection: String, f: F) -> Result<T, Error>
    where
        T: Send + 'static,
        F: FnOnce(&Database) -> quiver_embed::Result<T> + Send + 'static,
    {
        let db = Arc::clone(&self.db);
        let cells = Arc::clone(&self.snapshot_cells);
        let coll = collection.clone();
        let (result, stale) = tokio::task::spawn_blocking(move || -> Result<(T, bool), Error> {
            let guard = db
                .read()
                .map_err(|_| Error::Internal("database lock poisoned".to_owned()))?;
            let result = f(&guard).map_err(Error::Engine)?;
            // Report staleness so the caller can schedule a background rebuild; a
            // missing collection (raced a drop) is simply "nothing to rebuild".
            let stale = guard.needs_rebuild(&coll).unwrap_or(false);
            // Warm the lock-free read cache (ADR-0064 increment 3): once an
            // MVCC-served collection is fresh (its base snapshot is published), cache
            // its cell so subsequent pure-vector reads skip the lock. The cell
            // self-updates as the writer republishes, so it is never re-fetched.
            if !stale
                && let Ok(Some(cell)) = guard.mvcc_cell(&coll)
                && let Ok(mut map) = cells.write()
            {
                map.entry(coll.clone()).or_insert(cell);
            }
            Ok((result, stale))
        })
        .await
        .map_err(|e| Error::Internal(format!("blocking task failed: {e}")))??;
        if stale {
            self.schedule_rebuild(collection);
        }
        Ok(result)
    }

    // The cached lock-free snapshot cell for `collection`, if one has been warmed
    // (ADR-0064 increment 3). The map read contends only with create/drop, never
    // with a data write — that is what makes the pure-vector read lock-free.
    fn cached_cell(&self, collection: &str) -> Option<quiver_embed::SnapshotCell> {
        self.snapshot_cells.read().ok()?.get(collection).cloned()
    }

    // After an MVCC-mode data write, schedule an off-lock consolidation rebuild if
    // the write left the overlay stale (ADR-0064 increment 3): pure-vector fast-path
    // reads bypass the reader-driven scheduler, so the writer drives consolidation.
    // A brief read-lock staleness check; the rebuild itself runs off-lock and
    // single-flighted. No-op (and not called) when MVCC is off.
    async fn schedule_if_stale(&self, collection: &str) {
        let db = Arc::clone(&self.db);
        let coll = collection.to_owned();
        let stale = tokio::task::spawn_blocking(move || {
            db.read()
                .ok()
                .and_then(|g| g.needs_rebuild(&coll).ok())
                .unwrap_or(false)
        })
        .await
        .unwrap_or(false);
        if stale {
            self.schedule_rebuild(collection.to_owned());
        }
    }

    // Schedule a single off-lock rebuild for `collection` (ADR-0062), deduplicated
    // so concurrent readers that all observe the same deferred rebuild start exactly
    // one. A no-op if one is already in flight.
    fn schedule_rebuild(&self, collection: String) {
        {
            let mut inflight = match self.rebuilding.lock() {
                Ok(g) => g,
                Err(_) => return,
            };
            if !inflight.insert(collection.clone()) {
                return; // already rebuilding this collection
            }
        }
        let state = self.clone();
        tokio::spawn(async move {
            state.run_rebuild(&collection).await;
            if let Ok(mut inflight) = state.rebuilding.lock() {
                inflight.remove(&collection);
            }
        });
    }

    // Drive a collection's off-lock rebuild to completion (ADR-0062): capture the
    // inputs under the shared read lock, build with no lock held, install under a
    // brief write lock, and repeat while a write landed during the build (the commit
    // reports the collection still stale). Each phase is a `spawn_blocking` so the
    // CPU-bound build never stalls the async runtime; errors end the attempt (the
    // collection stays stale and the next read reschedules).
    async fn run_rebuild(&self, collection: &str) {
        loop {
            let db = Arc::clone(&self.db);
            let coll = collection.to_owned();
            let inputs = tokio::task::spawn_blocking(move || {
                let guard = db.read().ok()?;
                guard.snapshot_rebuild_inputs(&coll).ok().flatten()
            })
            .await
            .ok()
            .flatten();
            let Some(inputs) = inputs else { return };

            let Ok(Ok(rebuilt)) = tokio::task::spawn_blocking(move || inputs.build()).await else {
                return;
            };

            let db = Arc::clone(&self.db);
            let still_stale = tokio::task::spawn_blocking(move || {
                let mut guard = db.write().ok()?;
                guard.commit_rebuild(rebuilt).ok()
            })
            .await
            .ok()
            .flatten();
            match still_stale {
                Some(true) => continue, // a write landed during the build — rebuild again
                _ => return,
            }
        }
    }

    // Authorize `action` on `resource`, recording a denial in the audit log
    // before propagating it. The shared choke point for both transports.
    fn authorize(
        &self,
        principal: &Principal,
        action: Action,
        op: &str,
        resource: &str,
    ) -> Result<(), Error> {
        principal
            .require(action, Some(resource))
            .inspect_err(|_| self.audit.deny(principal.actor(), op, resource))
    }

    // Authorize a collection-agnostic operation (listing): only the role is
    // checked; a denial is recorded against the `*` resource.
    fn authorize_global(
        &self,
        principal: &Principal,
        action: Action,
        op: &str,
    ) -> Result<(), Error> {
        principal
            .require(action, None)
            .inspect_err(|_| self.audit.deny(principal.actor(), op, "*"))
    }

    /// Open a replication stream (ADR-0030): authorize (admin), then — in a single
    /// engine critical section so no commit can interleave — subscribe to the live
    /// commit tail and snapshot current state. The caller streams the snapshot,
    /// then the tail from the receiver; because the subscription is taken under the
    /// same lock as the snapshot, every post-snapshot op arrives on the receiver
    /// and none is missed or duplicated.
    pub(crate) async fn open_replication(
        &self,
        principal: &Principal,
    ) -> Result<(Vec<WalOp>, broadcast::Receiver<WalEntry>), Error> {
        self.authorize_global(principal, Action::Admin, "replicate")?;
        let tx = self.replication_tx.clone();
        self.read_blocking(move |db| {
            let rx = tx.subscribe();
            let snapshot = db.replication_snapshot()?;
            Ok((snapshot, rx))
        })
        .await
    }

    /// Apply a replicated op received from the leader (ADR-0030). Internal to the
    /// follower stream — deliberately NOT gated by `read_only`, which only refuses
    /// *external* client writes.
    pub(crate) async fn apply_replicated(&self, op: WalOp) -> Result<(), Error> {
        self.write_blocking(move |db| db.apply_replicated(op)).await
    }

    // Refuse a mutating operation on a read-only replication follower (ADR-0030);
    // its state is owned by the leader's stream, not by external clients.
    fn ensure_writable(&self, op: &str) -> Result<(), Error> {
        if self.read_only {
            return Err(Error::Forbidden(format!(
                "{op}: this node is a read-only replication follower"
            )));
        }
        Ok(())
    }

    #[allow(clippy::too_many_arguments)]
    pub(crate) async fn create_collection(
        &self,
        principal: &Principal,
        name: String,
        dim: u32,
        metric: DistanceMetric,
        index: IndexSpec,
        filterable: Vec<FilterableField>,
        multivector: bool,
        vector_encryption: VectorEncryption,
    ) -> Result<CollectionInfo, Error> {
        self.ensure_writable("create_collection")?;
        self.authorize(principal, Action::Admin, "create_collection", &name)?;
        self.limits.check_dim(dim as usize)?;
        if let Some(c) = &self.cluster {
            return c
                .create_collection(
                    name,
                    dim,
                    metric,
                    index,
                    filterable,
                    multivector,
                    vector_encryption,
                )
                .await;
        }
        let descriptor = Descriptor::new(dim, Dtype::F32, metric)
            .with_index(index)
            .with_filterable(filterable.clone())
            .with_multivector(multivector)
            .with_vector_encryption(vector_encryption);
        let owned = name.clone();
        let result = self
            .write_blocking(move |db| db.create_collection(&owned, descriptor))
            .await;
        self.audit.record(
            principal.actor(),
            "create_collection",
            &name,
            Outcome::of(&result),
        );
        result?;
        Ok(CollectionInfo {
            name,
            dim,
            metric,
            count: 0,
            index,
            filterable,
            multivector,
            vector_encryption,
        })
    }

    pub(crate) async fn get_collection(
        &self,
        principal: &Principal,
        name: String,
    ) -> Result<CollectionInfo, Error> {
        self.authorize(principal, Action::Read, "get_collection", &name)?;
        self.read_blocking(move |db| {
            let descriptor = db
                .descriptor(&name)
                .cloned()
                .ok_or_else(|| quiver_embed::Error::CollectionNotFound(name.clone()))?;
            // A multi-vector collection reports its document count, not its
            // (much larger) token-row count.
            let count = if descriptor.multivector {
                db.document_count(&name)? as u64
            } else {
                db.len(&name)? as u64
            };
            Ok(CollectionInfo {
                name,
                dim: descriptor.dim,
                metric: descriptor.metric,
                count,
                index: descriptor.index,
                filterable: descriptor.filterable,
                multivector: descriptor.multivector,
                vector_encryption: descriptor.vector_encryption,
            })
        })
        .await
    }

    pub(crate) async fn list_collections(
        &self,
        principal: &Principal,
    ) -> Result<Vec<CollectionInfo>, Error> {
        self.authorize_global(principal, Action::Read, "list_collections")?;
        let mut infos = self
            .read_blocking(|db| {
                let mut out = Vec::new();
                for name in db.collection_names() {
                    if let Some(descriptor) = db.descriptor(&name).cloned() {
                        let count = if descriptor.multivector {
                            db.document_count(&name)? as u64
                        } else {
                            db.len(&name)? as u64
                        };
                        out.push(CollectionInfo {
                            name,
                            dim: descriptor.dim,
                            metric: descriptor.metric,
                            count,
                            index: descriptor.index,
                            filterable: descriptor.filterable,
                            multivector: descriptor.multivector,
                            vector_encryption: descriptor.vector_encryption,
                        });
                    }
                }
                Ok(out)
            })
            .await?;
        // Never reveal collections outside the caller's scope.
        infos.retain(|info| principal.can_see(&info.name));
        Ok(infos)
    }

    pub(crate) async fn delete_collection(
        &self,
        principal: &Principal,
        name: String,
    ) -> Result<bool, Error> {
        self.ensure_writable("delete_collection")?;
        self.authorize(principal, Action::Admin, "delete_collection", &name)?;
        if let Some(c) = &self.cluster {
            return c.drop_collection(&name).await;
        }
        let resource = name.clone();
        let result = self
            .write_blocking(move |db| db.drop_collection(&name))
            .await;
        self.audit.record(
            principal.actor(),
            "delete_collection",
            &resource,
            Outcome::of(&result),
        );
        // Evict the dropped collection's cached lock-free cell (ADR-0064 increment 3)
        // so a later same-named collection never serves a stale cell.
        if matches!(result, Ok(true))
            && let Ok(mut map) = self.snapshot_cells.write()
        {
            map.remove(&resource);
        }
        result
    }

    #[tracing::instrument(skip_all, fields(collection = %collection, points = points.len()))]
    pub(crate) async fn upsert(
        &self,
        principal: &Principal,
        collection: String,
        points: Vec<PointIn>,
    ) -> Result<u64, Error> {
        self.ensure_writable("upsert")?;
        self.authorize(principal, Action::Write, "upsert", &collection)?;
        self.limits.check_batch(points.len())?;
        for p in &points {
            self.limits.check_vector_len(p.vector.len())?;
            self.limits.check_payload(&p.payload)?;
        }
        if let Some(c) = &self.cluster {
            return c.upsert(&collection, points).await;
        }
        let resource = collection.clone();
        let result = self
            .write_blocking(move |db| {
                let records: Vec<(&str, &[f32], &serde_json::Value)> = points
                    .iter()
                    .map(|p| (p.id.as_str(), p.vector.as_slice(), &p.payload))
                    .collect();
                db.upsert_batch(&collection, &records)
            })
            .await;
        self.audit
            .record(principal.actor(), "upsert", &resource, Outcome::of(&result));
        if self.mvcc && result.is_ok() {
            self.schedule_if_stale(&resource).await;
        }
        result
    }

    // Bulk upsert for a load-then-query workload (ADR-0045): one WAL fsync plus a
    // single deferred index-build pass, with the larger `max_bulk_batch_size` cap.
    pub(crate) async fn upsert_bulk(
        &self,
        principal: &Principal,
        collection: String,
        points: Vec<PointIn>,
    ) -> Result<u64, Error> {
        self.ensure_writable("upsert")?;
        self.authorize(principal, Action::Write, "upsert", &collection)?;
        self.limits.check_bulk_batch(points.len())?;
        for p in &points {
            self.limits.check_vector_len(p.vector.len())?;
            self.limits.check_payload(&p.payload)?;
        }
        if let Some(c) = &self.cluster {
            return c.upsert_bulk(&collection, points).await;
        }
        let resource = collection.clone();
        let result = self
            .write_blocking(move |db| {
                let records: Vec<(&str, &[f32], &serde_json::Value)> = points
                    .iter()
                    .map(|p| (p.id.as_str(), p.vector.as_slice(), &p.payload))
                    .collect();
                db.upsert_bulk(&collection, &records)
            })
            .await;
        self.audit.record(
            principal.actor(),
            "upsert_bulk",
            &resource,
            Outcome::of(&result),
        );
        if self.mvcc && result.is_ok() {
            self.schedule_if_stale(&resource).await;
        }
        result
    }

    /// Take a consistent online snapshot of the whole database into a
    /// server-local `destination` directory (ADR-0050). A global admin
    /// operation; runs the checkpoint + copy on the blocking pool.
    #[tracing::instrument(skip_all)]
    pub(crate) async fn snapshot(
        &self,
        principal: &Principal,
        destination: String,
    ) -> Result<SnapshotInfo, Error> {
        self.ensure_writable("snapshot")?;
        self.authorize_global(principal, Action::Admin, "snapshot")?;
        let dest = std::path::PathBuf::from(&destination);
        let result = self.write_blocking(move |db| db.snapshot(&dest)).await;
        self.audit.record(
            principal.actor(),
            "snapshot",
            &destination,
            Outcome::of(&result),
        );
        result
    }

    pub(crate) async fn delete_points(
        &self,
        principal: &Principal,
        collection: String,
        ids: Vec<String>,
    ) -> Result<u64, Error> {
        self.ensure_writable("delete_points")?;
        self.authorize(principal, Action::Write, "delete_points", &collection)?;
        if let Some(c) = &self.cluster {
            return c.delete_points(&collection, ids).await;
        }
        let resource = collection.clone();
        let result = self
            .write_blocking(move |db| {
                let mut count = 0u64;
                for id in &ids {
                    if db.delete(&collection, id)? {
                        count += 1;
                    }
                }
                Ok(count)
            })
            .await;
        self.audit.record(
            principal.actor(),
            "delete_points",
            &resource,
            Outcome::of(&result),
        );
        if self.mvcc && result.is_ok() {
            self.schedule_if_stale(&resource).await;
        }
        result
    }

    pub(crate) async fn get_points(
        &self,
        principal: &Principal,
        collection: String,
        ids: Vec<String>,
        with_vector: bool,
    ) -> Result<Vec<PointOut>, Error> {
        self.authorize(principal, Action::Read, "get_points", &collection)?;
        if let Some(c) = &self.cluster {
            return c.get_points(&collection, ids, with_vector).await;
        }
        self.read_blocking(move |db| {
            let mut out = Vec::new();
            for id in &ids {
                if let Some(m) = db.get(&collection, id)? {
                    out.push(PointOut {
                        id: m.id,
                        vector: if with_vector { m.vector } else { None },
                        payload: m.payload.unwrap_or(Value::Null),
                    });
                }
            }
            Ok(out)
        })
        .await
    }

    #[allow(clippy::too_many_arguments)]
    #[tracing::instrument(skip_all, fields(collection = %collection, k, filtered = filter.is_some()))]
    pub(crate) async fn search(
        &self,
        principal: &Principal,
        collection: String,
        vector: Vec<f32>,
        k: usize,
        filter: Option<Filter>,
        ef_search: usize,
        with_payload: bool,
        with_vector: bool,
    ) -> Result<Vec<MatchOut>, Error> {
        self.authorize(principal, Action::Read, "search", &collection)?;
        self.limits.check_search(k, ef_search)?;
        self.limits.check_vector_len(vector.len())?;

        // Cluster router (ADR-0065): scatter-gather across the shards, merge top-k.
        if let Some(c) = &self.cluster {
            return c
                .search(
                    &collection,
                    vector,
                    k,
                    filter,
                    ef_search,
                    with_payload,
                    with_vector,
                )
                .await;
        }

        // Lock-free fast path (ADR-0064 increment 3): a **pure-vector** read of an
        // MVCC-served collection loads the cached snapshot cell and searches it with
        // **no database lock**, so it never blocks on a concurrent writer. Payload,
        // filter, and vector results need the store (behind the lock) and fall
        // through to `search_blocking`.
        if filter.is_none()
            && !with_payload
            && !with_vector
            && let Some(cell) = self.cached_cell(&collection)
        {
            return tokio::task::spawn_blocking(move || {
                let matches = cell.load().search(&vector, k, ef_search)?;
                Ok::<_, quiver_embed::Error>(
                    matches
                        .into_iter()
                        .map(|m| MatchOut {
                            id: m.id,
                            score: m.score,
                            payload: None,
                            vector: None,
                        })
                        .collect(),
                )
            })
            .await
            .map_err(|e| Error::Internal(format!("blocking task failed: {e}")))?
            .map_err(Error::Engine);
        }

        let params = SearchParams {
            k,
            filter,
            ef_search,
            with_payload,
            with_vector,
        };
        let coll = collection.clone();
        self.search_blocking(coll, move |db| {
            let matches = db.search_snapshot(&collection, &vector, &params)?;
            Ok(matches
                .into_iter()
                .map(|m| MatchOut {
                    id: m.id,
                    score: m.score,
                    payload: m.payload,
                    vector: m.vector,
                })
                .collect())
        })
        .await
    }

    #[allow(clippy::too_many_arguments)]
    pub(crate) async fn hybrid_search(
        &self,
        principal: &Principal,
        collection: String,
        dense: Option<Vec<f32>>,
        sparse: Option<(Vec<u32>, Vec<f32>)>,
        text: Option<String>,
        k: usize,
        filter: Option<Filter>,
        ef_search: usize,
        rrf_k0: f32,
        with_payload: bool,
        with_vector: bool,
    ) -> Result<Vec<MatchOut>, Error> {
        self.authorize(principal, Action::Read, "hybrid_search", &collection)?;
        self.limits.check_search(k, ef_search)?;
        if let Some(v) = &dense {
            self.limits.check_vector_len(v.len())?;
        }
        if let Some((indices, values)) = &sparse {
            self.limits.check_sparse_terms(indices.len())?;
            if indices.len() != values.len() {
                return Err(Error::BadRequest(format!(
                    "sparse query indices ({}) and values ({}) length mismatch",
                    indices.len(),
                    values.len()
                )));
            }
        }
        let params = SearchParams {
            k,
            filter,
            ef_search,
            with_payload,
            with_vector,
        };
        let sv = sparse.map(|(indices, values)| SparseVector { indices, values });
        let coll = collection.clone();
        self.search_blocking(coll, move |db| {
            let matches = db.hybrid_search_snapshot(
                &collection,
                dense.as_deref(),
                sv.as_ref(),
                text.as_deref(),
                &params,
                rrf_k0,
            )?;
            Ok(matches
                .into_iter()
                .map(|m| MatchOut {
                    id: m.id,
                    score: m.score,
                    payload: m.payload,
                    vector: m.vector,
                })
                .collect())
        })
        .await
    }

    /// Embed `text` with the collection's configured provider and run a dense (or
    /// dense ⊕ BM25, if the collection has text) search, optionally reranking the
    /// candidates in one call (ADR-0047). The text is also passed to the BM25 side,
    /// so a `upsert_text` corpus yields hybrid lexical+semantic retrieval for free.
    #[allow(clippy::too_many_arguments)]
    pub(crate) async fn search_text(
        &self,
        principal: &Principal,
        collection: String,
        text: String,
        k: usize,
        filter: Option<Filter>,
        ef_search: usize,
        rrf_k0: f32,
        with_payload: bool,
        with_vector: bool,
        rerank: bool,
    ) -> Result<Vec<MatchOut>, Error> {
        self.authorize(principal, Action::Read, "search_text", &collection)?;
        self.limits.check_search(k, ef_search)?;
        let embedder = self.embed.embedder(&collection).ok_or_else(|| {
            Error::BadRequest(format!(
                "collection {collection:?} has no embedding provider configured \
                 (set an [embedding.{collection}] table in quiver.toml — ADR-0047)"
            ))
        })?;
        // Embed off the async runtime: the provider call is blocking network I/O.
        let query = text.clone();
        let vector = tokio::task::spawn_blocking(move || embedder.embed(&[query]))
            .await
            .map_err(|e| Error::Internal(format!("embedding task failed: {e}")))?
            .map_err(|e| Error::Upstream(e.to_string()))?
            .into_iter()
            .next()
            .ok_or_else(|| Error::Upstream("embedding provider returned no vector".to_owned()))?;
        self.limits.check_vector_len(vector.len())?;

        let reranker = if rerank {
            self.embed.reranker(&collection)
        } else {
            None
        };
        // Over-fetch when reranking so the reranker can reorder a wide candidate set
        // down to the requested `k`; fetch payloads when reranking (we need the doc
        // text) even if the caller did not ask for them.
        let need_payload = with_payload || reranker.is_some();
        let fetch_k = if reranker.is_some() {
            k.max(RERANK_CANDIDATES)
        } else {
            k
        };

        let mut hits = self
            .hybrid_search(
                principal,
                collection,
                Some(vector),
                None,
                Some(text.clone()),
                fetch_k,
                filter,
                ef_search,
                rrf_k0,
                need_payload,
                with_vector,
            )
            .await?;

        if let Some(rr) = reranker {
            let docs: Vec<String> = hits.iter().map(|h| doc_text(h.payload.as_ref())).collect();
            let query = text;
            let scores = tokio::task::spawn_blocking(move || rr.rerank(&query, &docs))
                .await
                .map_err(|e| Error::Internal(format!("rerank task failed: {e}")))?
                .map_err(|e| Error::Upstream(e.to_string()))?;
            // Re-score each hit and sort by the rerank score, descending.
            let mut scored: Vec<(f32, MatchOut)> = scores
                .into_iter()
                .zip(hits)
                .map(|(s, mut h)| {
                    h.score = s;
                    (s, h)
                })
                .collect();
            scored.sort_by(|a, b| b.0.partial_cmp(&a.0).unwrap_or(std::cmp::Ordering::Equal));
            hits = scored.into_iter().map(|(_, h)| h).collect();
        }

        hits.truncate(k);
        // Drop payloads we only fetched for reranking if the caller didn't want them.
        if !with_payload {
            for h in &mut hits {
                h.payload = None;
            }
        }
        Ok(hits)
    }

    /// Embed each point's `text` with the collection's provider and upsert it as a
    /// dense point, co-populating the `__quiver_text__` payload key (ADR-0046) so the
    /// same text is indexed for BM25 — one call feeds both the dense and lexical
    /// sides (ADR-0047).
    pub(crate) async fn upsert_text(
        &self,
        principal: &Principal,
        collection: String,
        points: Vec<TextPointIn>,
    ) -> Result<u64, Error> {
        self.ensure_writable("upsert_text")?;
        self.authorize(principal, Action::Write, "upsert_text", &collection)?;
        self.limits.check_batch(points.len())?;
        for p in &points {
            if !matches!(p.payload, Value::Object(_) | Value::Null) {
                return Err(Error::BadRequest(
                    "upsert_text payload must be a JSON object or null".to_owned(),
                ));
            }
        }
        let embedder = self.embed.embedder(&collection).ok_or_else(|| {
            Error::BadRequest(format!(
                "collection {collection:?} has no embedding provider configured \
                 (set an [embedding.{collection}] table in quiver.toml — ADR-0047)"
            ))
        })?;
        let texts: Vec<String> = points.iter().map(|p| p.text.clone()).collect();
        let vectors = tokio::task::spawn_blocking(move || embedder.embed(&texts))
            .await
            .map_err(|e| Error::Internal(format!("embedding task failed: {e}")))?
            .map_err(|e| Error::Upstream(e.to_string()))?;
        if vectors.len() != points.len() {
            return Err(Error::Upstream(format!(
                "embedding provider returned {} vectors for {} inputs",
                vectors.len(),
                points.len()
            )));
        }
        let dense: Vec<PointIn> = points
            .into_iter()
            .zip(vectors)
            .map(|(p, vector)| {
                let mut payload = match p.payload {
                    Value::Object(map) => map,
                    _ => serde_json::Map::new(),
                };
                // Don't clobber a caller-supplied text key.
                payload
                    .entry(TEXT_KEY.to_owned())
                    .or_insert_with(|| Value::String(p.text.clone()));
                PointIn {
                    id: p.id,
                    vector,
                    payload: Value::Object(payload),
                }
            })
            .collect();
        self.upsert(principal, collection, dense).await
    }

    #[allow(clippy::too_many_arguments)]
    pub(crate) async fn fetch(
        &self,
        principal: &Principal,
        collection: String,
        filter: Option<Filter>,
        offset: usize,
        limit: usize,
        with_payload: bool,
        with_vector: bool,
    ) -> Result<Vec<MatchOut>, Error> {
        self.authorize(principal, Action::Read, "fetch", &collection)?;
        self.limits.check_fetch(limit)?;
        self.read_blocking(move |db| {
            let matches = db.fetch(
                &collection,
                filter.as_ref(),
                offset,
                limit,
                with_payload,
                with_vector,
            )?;
            Ok(matches
                .into_iter()
                .map(|m| MatchOut {
                    id: m.id,
                    score: m.score,
                    payload: m.payload,
                    vector: m.vector,
                })
                .collect())
        })
        .await
    }

    pub(crate) async fn upsert_documents(
        &self,
        principal: &Principal,
        collection: String,
        documents: Vec<DocumentIn>,
    ) -> Result<u64, Error> {
        self.ensure_writable("upsert_documents")?;
        self.authorize(principal, Action::Write, "upsert_documents", &collection)?;
        self.limits.check_batch(documents.len())?;
        for doc in &documents {
            self.limits.check_payload(&doc.payload)?;
            for token in &doc.vectors {
                self.limits.check_vector_len(token.len())?;
            }
        }
        let resource = collection.clone();
        let result = self
            .write_blocking(move |db| {
                let mut count = 0u64;
                for doc in &documents {
                    db.upsert_document(&collection, &doc.id, &doc.vectors, &doc.payload)?;
                    count += 1;
                }
                Ok(count)
            })
            .await;
        self.audit.record(
            principal.actor(),
            "upsert_documents",
            &resource,
            Outcome::of(&result),
        );
        result
    }

    pub(crate) async fn delete_documents(
        &self,
        principal: &Principal,
        collection: String,
        ids: Vec<String>,
    ) -> Result<u64, Error> {
        self.ensure_writable("delete_documents")?;
        self.authorize(principal, Action::Write, "delete_documents", &collection)?;
        let resource = collection.clone();
        let result = self
            .write_blocking(move |db| {
                let mut count = 0u64;
                for id in &ids {
                    if db.delete_document(&collection, id)? {
                        count += 1;
                    }
                }
                Ok(count)
            })
            .await;
        self.audit.record(
            principal.actor(),
            "delete_documents",
            &resource,
            Outcome::of(&result),
        );
        result
    }

    #[allow(clippy::too_many_arguments)]
    pub(crate) async fn search_multi_vector(
        &self,
        principal: &Principal,
        collection: String,
        query: Vec<Vec<f32>>,
        k: usize,
        filter: Option<Filter>,
        ef_search: usize,
        with_payload: bool,
        with_vector: bool,
    ) -> Result<Vec<DocumentMatchOut>, Error> {
        self.authorize(principal, Action::Read, "search_multi_vector", &collection)?;
        self.limits.check_search(k, ef_search)?;
        for token in &query {
            self.limits.check_vector_len(token.len())?;
        }
        let params = SearchParams {
            k,
            filter,
            ef_search,
            with_payload,
            with_vector,
        };
        let coll = collection.clone();
        self.search_blocking(coll, move |db| {
            let matches = db.search_multi_vector_snapshot(&collection, &query, &params)?;
            Ok(matches
                .into_iter()
                .map(|m| DocumentMatchOut {
                    id: m.id,
                    score: m.score,
                    payload: m.payload,
                    vectors: m.vectors,
                })
                .collect())
        })
        .await
    }
}

/// How many recently-committed ops the leader buffers for replication followers
/// (ADR-0030). A follower that falls further behind than this re-bootstraps.
const REPLICATION_BUFFER: usize = 1024;

/// Run the server from `config` until a shutdown signal (Ctrl-C).
pub async fn run(config: Config) -> Result<(), Error> {
    config.validate()?;
    let rest_listener = TcpListener::bind(config.rest_addr)
        .await
        .map_err(Error::Io)?;
    let grpc_listener = TcpListener::bind(config.grpc_addr)
        .await
        .map_err(Error::Io)?;
    tracing::info!(rest = %config.rest_addr, grpc = %config.grpc_addr, "quiver listening");
    tokio::select! {
        result = serve(config, rest_listener, grpc_listener) => result,
        () = shutdown_signal() => {
            tracing::info!("shutdown signal received");
            Ok(())
        }
    }
}

/// Serve REST and gRPC on the given (already-bound) listeners until a transport
/// error. Exposed so tests can bind ephemeral ports.
pub async fn serve(
    config: Config,
    rest_listener: TcpListener,
    grpc_listener: TcpListener,
) -> Result<(), Error> {
    // Coordinator mode (ADR-0066) is a wholly separate, data-plane-free service: it
    // owns the versioned shard map and serves the membership API on the REST
    // listener. It opens no database and runs no gRPC.
    if config.coordinator {
        drop(grpc_listener);
        return coordinator::serve_coordinator(config, rest_listener).await;
    }
    let mut db = open_database(&config)?;
    let audit = Arc::new(AuditLog::open(config.audit_log.as_deref())?);
    // Publish every committed op to replication followers (ADR-0030). The observer
    // runs inside the engine's write critical section; `broadcast::Sender::send` is
    // non-blocking, so it never stalls a write.
    let (replication_tx, _) = broadcast::channel(REPLICATION_BUFFER);
    {
        let tx = replication_tx.clone();
        db.set_commit_observer(Arc::new(move |entry: &WalEntry| {
            let _ = tx.send(entry.clone());
        }));
    }
    // Build the opt-in embedding/rerank providers, resolving each `api_key_env`
    // from the environment now (ADR-0047) so a missing key fails fast at startup
    // rather than on the first request.
    let embed = EmbedRegistry::from_config(&config.embedding, &config.rerank)
        .map_err(|e| Error::Config(e.to_string()))?;

    // Enable lock-free MVCC reads if requested (ADR-0064): `config.mvcc_reads` or
    // `QUIVER_MVCC_READS` (the latter also defaults the engine flag at open).
    if config.mvcc_reads {
        db.set_mvcc_reads(true);
    }
    let mvcc = db.mvcc_reads();

    // Opt-in cluster router (ADR-0065): build the shard fan-out when configured.
    let cluster = if config.cluster_shards.is_empty() {
        None
    } else {
        let c = cluster::Cluster::new(
            config.cluster_shards.clone(),
            config.cluster_replicas.clone(),
            config.cluster_shard_key.clone(),
        )?;
        tracing::info!(shards = c.shard_count(), "quiver cluster router enabled");
        let c = Arc::new(c);
        // With a coordinator (ADR-0066), refresh the shard map from it so a
        // membership change propagates without a restart. The seed above is the
        // bootstrap map; the coordinator's map (≥ version 0) takes over on refresh.
        if let Some(coord) = config.coordinator_url.clone() {
            let router = c.clone();
            tokio::spawn(async move {
                loop {
                    if let Err(e) = router.refresh_from(&coord).await {
                        tracing::debug!(error = %e, "shard-map refresh failed; will retry");
                    }
                    tokio::time::sleep(cluster::MAP_REFRESH_INTERVAL).await;
                }
            });
        }
        Some(c)
    };

    let state = AppState {
        db: Arc::new(RwLock::new(db)),
        keys: Arc::new(config.api_keys.clone()),
        audit,
        replication_tx,
        read_only: config.leader_url.is_some(),
        limits: config.limits,
        embed: Arc::new(embed),
        rate_limiter: Arc::new(RateLimiter::new(config.rate_limit)),
        metrics: Arc::new(metrics::Metrics::default()),
        rebuilding: Arc::new(Mutex::new(HashSet::new())),
        snapshot_cells: Arc::new(RwLock::new(HashMap::new())),
        mvcc,
        cluster,
    };

    // A follower continuously applies the leader's committed-op stream (ADR-0030).
    if let Some(leader_url) = config.leader_url.clone() {
        replication::spawn_follower(state.clone(), leader_url, config.leader_api_key.clone());
    }

    let app = rest::router(state.clone());
    let grpc = grpc::service(state);

    let tls = load_tls(&config)?;

    // REST: terminate TLS with axum-server when configured, else serve plaintext.
    let rest_fut: Pin<Box<dyn Future<Output = Result<(), Error>> + Send>> = match &tls {
        Some(material) => {
            let rustls_config = RustlsConfig::from_config(Arc::clone(&material.rest_config));
            let std_listener = rest_listener.into_std().map_err(Error::Io)?;
            let server =
                axum_server::from_tcp_rustls(std_listener, rustls_config).map_err(Error::Io)?;
            Box::pin(async move {
                server
                    .serve(app.into_make_service())
                    .await
                    .map_err(Error::Io)
            })
        }
        None => Box::pin(async move { axum::serve(rest_listener, app).await.map_err(Error::Io) }),
    };

    // gRPC: tonic terminates TLS itself (ring provider) when an identity is set.
    let mut grpc_builder = tonic::transport::Server::builder();
    if let Some(material) = &tls {
        let identity = Identity::from_pem(&material.cert_pem, &material.key_pem);
        let mut tls_config = ServerTlsConfig::new().identity(identity);
        // Require client certificates chaining to the configured CA (mTLS).
        if let Some(ca_pem) = &material.client_ca_pem {
            tls_config = tls_config.client_ca_root(Certificate::from_pem(ca_pem));
        }
        grpc_builder = grpc_builder
            .tls_config(tls_config)
            .map_err(|e| Error::Internal(format!("grpc tls config: {e}")))?;
    }
    let grpc_fut = async move {
        grpc_builder
            .add_service(grpc)
            .serve_with_incoming(TcpListenerStream::new(grpc_listener))
            .await
            .map_err(|e| Error::Internal(format!("grpc server: {e}")))
    };

    tokio::try_join!(rest_fut, grpc_fut)?;
    Ok(())
}

async fn shutdown_signal() {
    let _ = tokio::signal::ctrl_c().await;
}

// Open the engine, enabling encryption-at-rest when a key is configured. The
// configured key is the **master key** of an envelope key-ring (ADR-0010): it
// wraps a per-collection data-encryption key, so dropping a collection
// crypto-shreds it. With no key (only valid in `insecure` mode, enforced by
// `Config::validate`) the engine is opened in plaintext.
fn open_database(config: &Config) -> Result<Database, Error> {
    let master_key = config.master_key_hex()?;
    let keyring =
        quiver_crypto::open_keyring(&config.data_dir, master_key.as_deref(), config.insecure)
            .map_err(|e| Error::Config(e.to_string()))?;
    let db = match keyring {
        Some(keyring) => Database::open_with_keyring(&config.data_dir, keyring)?,
        None => Database::open(&config.data_dir)?,
    };
    Ok(db)
}

// TLS material shared by both transports: the raw PEM (for tonic's `Identity`
// and `client_ca_root`) and a parsed rustls server config (for axum-server's
// REST acceptor). `client_ca_pem` is set only when mutual TLS is configured.
struct TlsMaterial {
    cert_pem: Vec<u8>,
    key_pem: Vec<u8>,
    client_ca_pem: Option<Vec<u8>>,
    rest_config: Arc<rustls::ServerConfig>,
}

// Read the configured certificate, key, and optional client CA, returning `None`
// when TLS is not configured. `Config::validate` already enforces that the cert
// and key are set together, that a client CA requires them, and that a
// non-loopback bind requires TLS.
fn load_tls(config: &Config) -> Result<Option<TlsMaterial>, Error> {
    match (&config.tls_cert, &config.tls_key) {
        (Some(cert_path), Some(key_path)) => {
            let cert_pem = std::fs::read(cert_path).map_err(Error::Io)?;
            let key_pem = std::fs::read(key_path).map_err(Error::Io)?;
            let client_ca_pem = config
                .tls_client_ca
                .as_ref()
                .map(std::fs::read)
                .transpose()
                .map_err(Error::Io)?;
            let rest_config = Arc::new(rustls_server_config(
                &cert_pem,
                &key_pem,
                client_ca_pem.as_deref(),
            )?);
            Ok(Some(TlsMaterial {
                cert_pem,
                key_pem,
                client_ca_pem,
                rest_config,
            }))
        }
        (None, None) => Ok(None),
        _ => Err(Error::Config(
            "tls_cert and tls_key must be set together".to_owned(),
        )),
    }
}

// Build a rustls server config from PEM bytes over the audited `ring` provider
// (no OpenSSL, no aws-lc-rs C toolchain). TLS 1.3 and 1.2 are offered. When a
// client CA is supplied, client certificates chaining to it are required (mTLS).
fn rustls_server_config(
    cert_pem: &[u8],
    key_pem: &[u8],
    client_ca_pem: Option<&[u8]>,
) -> Result<rustls::ServerConfig, Error> {
    use rustls_pki_types::pem::PemObject;
    use rustls_pki_types::{CertificateDer, PrivateKeyDer};

    let certs = CertificateDer::pem_slice_iter(cert_pem)
        .collect::<std::result::Result<Vec<_>, _>>()
        .map_err(|e| Error::Config(format!("parsing tls_cert: {e}")))?;
    if certs.is_empty() {
        return Err(Error::Config(
            "tls_cert contains no certificates".to_owned(),
        ));
    }
    let key = PrivateKeyDer::from_pem_slice(key_pem)
        .map_err(|e| Error::Config(format!("parsing tls_key: {e}")))?;
    let provider = Arc::new(rustls::crypto::ring::default_provider());
    let builder = rustls::ServerConfig::builder_with_provider(Arc::clone(&provider))
        .with_safe_default_protocol_versions()
        .map_err(|e| Error::Internal(format!("tls protocol versions: {e}")))?;
    let builder = match client_ca_pem {
        Some(ca_pem) => {
            let mut roots = rustls::RootCertStore::empty();
            for cert in CertificateDer::pem_slice_iter(ca_pem) {
                let cert =
                    cert.map_err(|e| Error::Config(format!("parsing tls_client_ca: {e}")))?;
                roots
                    .add(cert)
                    .map_err(|e| Error::Config(format!("adding tls_client_ca: {e}")))?;
            }
            let verifier = rustls::server::WebPkiClientVerifier::builder_with_provider(
                Arc::new(roots),
                provider,
            )
            .build()
            .map_err(|e| Error::Config(format!("client certificate verifier: {e}")))?;
            builder.with_client_cert_verifier(verifier)
        }
        None => builder.with_no_client_auth(),
    };
    builder
        .with_single_cert(certs, key)
        .map_err(|e| Error::Config(format!("tls certificate/key: {e}")))
}

/// Initialize structured logging from `RUST_LOG` (defaulting to `info`). Safe to
/// call once at startup; a second call is ignored.
pub fn init_tracing() {
    init_observability(&Config::default());
}

/// Install the global tracing subscriber: an `RUST_LOG`-driven `fmt` layer plus,
/// when the `otlp` feature is built **and** an OTLP endpoint is configured
/// (ADR-0059), an OpenTelemetry traces export layer. Safe to call once at
/// startup; a second call is a no-op. A failure to build the OTLP exporter logs a
/// warning and falls back to `fmt`-only rather than taking the server down.
#[cfg_attr(not(feature = "otlp"), allow(unused_variables))]
pub fn init_observability(config: &Config) {
    use tracing_subscriber::EnvFilter;
    use tracing_subscriber::prelude::*;
    let filter = EnvFilter::try_from_default_env().unwrap_or_else(|_| EnvFilter::new("info"));
    let registry = tracing_subscriber::registry()
        .with(filter)
        .with(tracing_subscriber::fmt::layer());

    #[cfg(feature = "otlp")]
    if config.otlp.is_enabled() {
        match otlp::build_provider(&config.otlp) {
            Ok(provider) => {
                use opentelemetry::trace::TracerProvider as _;
                let tracer = provider.tracer("quiver");
                otlp::store_provider(provider);
                let _ = registry
                    .with(tracing_opentelemetry::layer().with_tracer(tracer))
                    .try_init();
                return;
            }
            Err(e) => eprintln!("OTLP traces export disabled: {e}"),
        }
    }

    let _ = registry.try_init();
}

/// Flush and shut down the OpenTelemetry exporter, if one was installed. A no-op
/// without the `otlp` feature or when no endpoint was configured. Call once on
/// server shutdown so batched spans are not lost.
pub fn shutdown_observability() {
    #[cfg(feature = "otlp")]
    otlp::shutdown();
}

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

    // A valid 64-hex-character (256-bit) test key.
    const TEST_KEY: &str = "00112233445566778899aabbccddeeff00112233445566778899aabbccddeeff";

    #[test]
    fn config_rejects_missing_keys_unless_insecure() {
        let mut config = Config::default();
        assert!(config.validate().is_err());
        config.insecure = true;
        assert!(config.validate().is_ok());
        config.insecure = false;
        config.api_keys = vec!["secret".into()];
        config.encryption_key = Some(TEST_KEY.to_owned());
        assert!(config.validate().is_ok());
    }

    #[test]
    fn config_requires_encryption_key_unless_insecure() {
        let mut config = Config {
            api_keys: vec!["secret".into()],
            ..Config::default()
        };
        // API key set but no encryption key, not insecure ⇒ rejected.
        assert!(config.validate().is_err());
        config.encryption_key = Some(TEST_KEY.to_owned());
        assert!(config.validate().is_ok());
        // A malformed key is rejected up front, not at first write.
        config.encryption_key = Some("not-a-valid-hex-key".to_owned());
        assert!(config.validate().is_err());
        // Insecure mode may run without encryption-at-rest.
        config.insecure = true;
        config.encryption_key = None;
        assert!(config.validate().is_ok());
    }

    #[test]
    fn master_key_file_is_an_alternative_to_the_env_key() {
        let dir = tempfile::tempdir().unwrap();
        let path = dir.path().join("master.key");
        // A trailing newline (as editors and `echo` add) is trimmed off.
        std::fs::write(&path, format!("{TEST_KEY}\n")).unwrap();

        let mut config = Config {
            api_keys: vec!["secret".into()],
            master_key_file: Some(path.clone()),
            ..Config::default()
        };
        // The file alone satisfies encryption-at-rest and resolves to the key.
        assert!(config.validate().is_ok());
        assert_eq!(config.master_key_hex().unwrap().as_deref(), Some(TEST_KEY));

        // Setting both the env key and a file is rejected as ambiguous.
        config.encryption_key = Some(TEST_KEY.to_owned());
        assert!(config.validate().is_err());

        // A file holding malformed hex is rejected up front.
        config.encryption_key = None;
        std::fs::write(&path, "not-a-valid-key").unwrap();
        assert!(config.validate().is_err());
    }

    #[test]
    fn config_rejects_public_bind_without_optout() {
        let mut config = Config {
            api_keys: vec!["secret".into()],
            encryption_key: Some(TEST_KEY.to_owned()),
            ..Config::default()
        };
        config.rest_addr = SocketAddr::new(IpAddr::V4(Ipv4Addr::UNSPECIFIED), 6333);
        // Auth and encryption are satisfied, so the only failure is the bind rule.
        assert!(config.validate().is_err());
        config.insecure = true;
        assert!(config.validate().is_ok());
    }

    #[test]
    fn config_public_bind_allowed_with_tls() {
        let config = Config {
            api_keys: vec!["secret".into()],
            encryption_key: Some(TEST_KEY.to_owned()),
            tls_cert: Some(PathBuf::from("cert.pem")),
            tls_key: Some(PathBuf::from("key.pem")),
            rest_addr: SocketAddr::new(IpAddr::V4(Ipv4Addr::UNSPECIFIED), 6333),
            ..Config::default()
        };
        // TLS configured ⇒ a non-loopback bind is allowed without insecure.
        assert!(config.validate().is_ok());
    }

    #[test]
    fn config_tls_cert_and_key_must_pair() {
        let mut config = Config {
            api_keys: vec!["secret".into()],
            encryption_key: Some(TEST_KEY.to_owned()),
            tls_cert: Some(PathBuf::from("cert.pem")),
            ..Config::default()
        };
        // Cert without key ⇒ rejected.
        assert!(config.validate().is_err());
        config.tls_key = Some(PathBuf::from("key.pem"));
        assert!(config.validate().is_ok());
    }
}