vcon_server/

store.rs

//! One tenant's vCons: a sealed object file per kept version and a catalog
//! row for it. The catalog is the truth: a version exists once its row is
//! committed, and an object file with no row is garbage.

use std::path::{Path, PathBuf};
use std::sync::{Arc, Mutex, MutexGuard};

use rusqlite::{Connection, OptionalExtension, Transaction, TransactionBehavior, params};
use uuid::Uuid;
use vcon::Inspection;

use crate::audit::{self, Action, Actor, Entry};
use crate::catalog;
use crate::clock::{Clock, nanos};
use crate::envelope::{self, Context, Key, WrapLimit};
use crate::fs::Fs;
use crate::ids::{KeyId, TenantId, VersionId};
use crate::retention::Retention;

mod examine;
mod legal;
mod list;
mod purge;
mod rekey;

pub use examine::Examination;
pub use legal::{Legal, VersionInfo};
pub use list::{Cursor, ListQuery, Listed, Listing};
pub use purge::{Kept, PurgeReport, due};

/// The tables of a tenant's catalog, besides its audit log. A key column
/// says NOT NULL because SQLite lets a PRIMARY KEY other than an INTEGER
/// one hold NULL.
const TABLES: &str = "
CREATE TABLE vcon (
    uuid BLOB PRIMARY KEY NOT NULL,
    current BLOB,
    deleted_at INTEGER,
    hold INTEGER NOT NULL DEFAULT 0,
    stored_at INTEGER NOT NULL DEFAULT 0
);
CREATE INDEX vcon_listed ON vcon (stored_at, uuid);
CREATE TABLE version (
    version BLOB PRIMARY KEY NOT NULL,
    uuid BLOB NOT NULL,
    form TEXT NOT NULL,
    gzip INTEGER NOT NULL,
    size INTEGER NOT NULL,
    stored_at INTEGER NOT NULL,
    hidden_at INTEGER,
    tenant_key_id BLOB NOT NULL,
    wrapped_key BLOB NOT NULL
);
CREATE INDEX version_uuid ON version (uuid);
CREATE TABLE tombstone (
    uuid BLOB PRIMARY KEY NOT NULL,
    purged_at INTEGER NOT NULL
);
CREATE TABLE key_use (
    tenant_key_id BLOB PRIMARY KEY NOT NULL,
    wraps INTEGER NOT NULL
);
";

/// Who is calling, for the audit record.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub struct Caller {
    pub key: KeyId,
}

/// What `create` did.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum Created {
    /// Stored as a new vCon.
    New(VersionId),
    /// The same bytes were already current; nothing was written.
    Same(VersionId),
}

/// A version as `get` returns it.
#[derive(Debug, Clone, PartialEq, Eq)]
pub struct Stored {
    pub version: VersionId,
    pub form: vcon::Form,
    /// Whether the body is gzipped, as it arrived.
    pub gzip: bool,
    /// The body exactly as it arrived.
    pub body: Vec<u8>,
}

/// What a tenant's Store keeps, as removal checks it.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub struct Counts {
    /// Versions kept, hidden ones included.
    pub versions: u64,
    /// Uuids under a legal hold.
    pub held: u64,
}

/// Why a Store call was refused or failed.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum StoreError {
    Conflict,
    Gone,
    NotFound,
    NotDeleted,
    PreconditionFailed,
    Immutable,
    /// The tenant key is at its wrap limit and must be rotated.
    LimitReached,
    /// A row's object file is missing or does not open.
    Damaged,
    /// The audit record could not be written, so nothing was done.
    AuditFailed,
    Io,
}

/// One tenant's vCons.
pub struct Store {
    tenant: TenantId,
    tenant_key: Key,
    objects: PathBuf,
    conn: Mutex<Connection>,
    fs: Arc<dyn Fs>,
    clock: Arc<dyn Clock>,
    #[cfg(test)]
    hooks: tests::Hooks,
}

impl Store {
    /// Opens the tenant's catalog under its catalog key, making it if new,
    /// then runs start-up recovery.
    pub fn open(
        dir: &Path,
        tenant: TenantId,
        tenant_key: Key,
        fs: Arc<dyn Fs>,
        clock: Arc<dyn Clock>,
    ) -> Result<Store, StoreError> {
        let store = Store::open_beside(dir, tenant, tenant_key, fs, clock)?;
        store.recover()?;
        Ok(store)
    }

    /// Opens the store as a command running beside a server does: without
    /// start-up recovery, which would take the server's writes in progress
    /// for a crash's leftovers.
    pub fn open_beside(
        dir: &Path,
        tenant: TenantId,
        tenant_key: Key,
        fs: Arc<dyn Fs>,
        clock: Arc<dyn Clock>,
    ) -> Result<Store, StoreError> {
        let catalog_key = envelope::catalog_key(&tenant_key).map_err(|_| StoreError::Io)?;
        let schema = format!("{TABLES}{}", audit::SCHEMA);
        let conn = catalog::open(&dir.join("catalog.sqlite"), &catalog_key, &schema)
            .map_err(|_| StoreError::Io)?;
        let objects = dir.join("objects");
        fs.create_dir_all(&objects).map_err(|_| StoreError::Io)?;
        // A new directory survives a crash only once the directory that
        // holds it is synced; without this a crash could take the objects
        // directory, synced files and all, while the catalog keeps its rows.
        fs.sync_dir(dir).map_err(|_| StoreError::Io)?;
        Ok(Store {
            tenant,
            tenant_key,
            objects,
            conn: Mutex::new(conn),
            fs,
            clock,
            #[cfg(test)]
            hooks: tests::Hooks::default(),
        })
    }

    /// Start-up recovery: the WAL is checkpointed with truncation, and
    /// temporary files and object files with no row are removed. A name
    /// that is not a version id, with or without `.tmp`, is not ours and
    /// is left alone.
    fn recover(&self) -> Result<(), StoreError> {
        let conn = self.conn();
        let busy: i64 = conn
            .query_row("PRAGMA wal_checkpoint(TRUNCATE)", [], |r| r.get(0))
            .map_err(io)?;
        if busy != 0 {
            return Err(StoreError::Io);
        }
        let mut removed = false;
        for name in self.fs.list(&self.objects).map_err(io)? {
            let Some(name) = name.to_str() else { continue };
            let (id, temporary) = match name.strip_suffix(".tmp") {
                Some(id) => (id, true),
                None => (name, false),
            };
            let Some(version) = VersionId::parse(id) else {
                continue;
            };
            let has_row = conn
                .query_row(
                    "SELECT 1 FROM version WHERE version = ?1",
                    [version.0],
                    |_| Ok(()),
                )
                .optional()
                .map_err(io)?
                .is_some();
            if temporary || !has_row {
                self.fs.remove(&self.objects.join(name)).map_err(io)?;
                removed = true;
            }
        }
        if removed {
            self.fs.sync_dir(&self.objects).map_err(io)?;
        }
        Ok(())
    }

    /// Stores a new vCon. The uuid's state is checked before anything is
    /// sealed, the body is written in the order below, and the
    /// commit checks the state again, so of two racing writers one wins
    /// and the other's file is removed.
    pub fn create(
        &self,
        who: &Caller,
        found: &Inspection,
        body: &[u8],
    ) -> Result<Created, StoreError> {
        let uuid = found.uuid;
        let entry = |status, version| Entry {
            actor: Actor::Key(who.key),
            action: Action::Create,
            target: Some(uuid.to_string()),
            version,
            status,
        };
        let key_id = self.tenant_key.id();
        {
            let mut conn = self.conn();
            let tx = immediate(&mut conn)?;
            let refused = match state(&tx, uuid)? {
                State::Gone => Some((410, None, StoreError::Gone)),
                State::Live(current) => match self.read_version(&tx, uuid, current) {
                    Ok(stored) if stored == body => {
                        self.record(&tx, &entry(200, Some(current)))?;
                        commit(tx)?;
                        return Ok(Created::Same(current));
                    }
                    Ok(_) => Some((409, None, StoreError::Conflict)),
                    Err(StoreError::Damaged) => Some((500, Some(current), StoreError::Damaged)),
                    Err(e) => return Err(e),
                },
                State::Absent if envelope::wrap_limit(wraps(&tx, key_id)?) == WrapLimit::Refuse => {
                    Some((503, None, StoreError::LimitReached))
                }
                // The wrap is counted before it is made, so every wrap is
                // counted once whether this write wins or loses.
                State::Absent => {
                    add_wrap(&tx, key_id)?;
                    None
                }
            };
            if let Some((status, version, error)) = refused {
                self.record(&tx, &entry(status, version))?;
                commit(tx)?;
                return Err(error);
            }
            commit(tx)?;
        }

        let (version, wrapped, file) = self.write_version(uuid, body)?;
        self.pause();

        let committed = (|| {
            let mut conn = self.conn();
            let tx = immediate(&mut conn)?;
            match state(&tx, uuid)? {
                State::Absent => {}
                lost => {
                    let (status, error) = match lost {
                        State::Gone => (410, StoreError::Gone),
                        _ => (409, StoreError::Conflict),
                    };
                    self.record(&tx, &entry(status, None))?;
                    commit(tx)?;
                    return Err(error);
                }
            }
            // The version's time, and the uuid's first: the list keeps
            // its place by it through replace, delete and restore.
            let stored_at = nanos(self.clock.now());
            tx.execute(
                "INSERT INTO version
                     (version, uuid, form, gzip, size, stored_at, hidden_at, tenant_key_id, wrapped_key)
                 VALUES (?1, ?2, ?3, ?4, ?5, ?6, NULL, ?7, ?8)",
                params![
                    version.0,
                    uuid.as_bytes(),
                    form_name(found.form),
                    found.gzip,
                    i64::try_from(body.len()).map_err(io)?,
                    stored_at,
                    key_id.bytes(),
                    wrapped,
                ],
            )
            .map_err(io)?;
            tx.execute(
                "INSERT INTO vcon (uuid, current, stored_at) VALUES (?1, ?2, ?3)",
                params![uuid.as_bytes(), version.0, stored_at],
            )
            .map_err(io)?;
            self.record(&tx, &entry(201, Some(version)))?;
            commit(tx)?;
            Ok(Created::New(version))
        })();
        if committed.is_err() {
            // Best effort: start-up recovery removes any file left over.
            let _ = self.fs.remove(&file);
        }
        committed
    }

    /// Reads the current version of `uuid`. The body is decrypted whole,
    /// then one transaction checks that the version is still current and
    /// commits the audit record; only then is the body returned. If the
    /// version changed meanwhile, the read starts again.
    pub fn get(&self, who: &Caller, uuid: Uuid) -> Result<Stored, StoreError> {
        let entry = |status, version| Entry {
            actor: Actor::Key(who.key),
            action: Action::Read,
            target: Some(uuid.to_string()),
            version,
            status,
        };
        let refuse = |tx: Transaction, state: State| {
            let (status, error) = match state {
                State::Gone => (410, StoreError::Gone),
                _ => (404, StoreError::NotFound),
            };
            self.record(&tx, &entry(status, None))?;
            commit(tx)?;
            Err(error)
        };
        loop {
            let (version, row) = {
                let mut conn = self.conn();
                let tx = immediate(&mut conn)?;
                match state(&tx, uuid)? {
                    State::Live(version) => (version, version_row(&tx, version)),
                    other => return refuse(tx, other),
                }
            };
            let opened = row.and_then(|row| {
                let body = self.open_version(uuid, version, &row.wrapped)?;
                Ok(Stored {
                    version,
                    form: row.form,
                    gzip: row.gzip,
                    body,
                })
            });
            self.pause();
            let mut conn = self.conn();
            let tx = immediate(&mut conn)?;
            match state(&tx, uuid)? {
                State::Live(now) if now == version => {
                    let status = if opened.is_ok() { 200 } else { 500 };
                    self.record(&tx, &entry(status, Some(version)))?;
                    commit(tx)?;
                    return opened;
                }
                State::Live(_) => continue,
                other => return refuse(tx, other),
            }
        }
    }

    /// Replaces the current version of `found.uuid`, if it is `if_match`
    /// and neither signed nor encrypted. The caller has checked that the
    /// body's uuid is the one asked for. The old version is hidden, and
    /// kept for legal reads until it is purged. As in `create`, the state
    /// is checked before sealing and again at the commit, so of two racing
    /// replaces one wins and the other's file is removed.
    pub fn replace(
        &self,
        who: &Caller,
        if_match: VersionId,
        found: &Inspection,
        body: &[u8],
    ) -> Result<VersionId, StoreError> {
        let uuid = found.uuid;
        let entry = |status, version| Entry {
            actor: Actor::Key(who.key),
            action: Action::Replace,
            target: Some(uuid.to_string()),
            version,
            status,
        };
        let key_id = self.tenant_key.id();
        // Why the state does not allow this replace, if it does not.
        let refusal = |tx: &Transaction| -> Result<Option<(u16, StoreError)>, StoreError> {
            Ok(match state(tx, uuid)? {
                State::Absent => Some((404, StoreError::NotFound)),
                State::Gone => Some((410, StoreError::Gone)),
                State::Live(current) if version_row(tx, current)?.form != vcon::Form::Unsigned => {
                    Some((409, StoreError::Immutable))
                }
                State::Live(current) if current != if_match => {
                    Some((412, StoreError::PreconditionFailed))
                }
                State::Live(_) => None,
            })
        };
        {
            let mut conn = self.conn();
            let tx = immediate(&mut conn)?;
            let refused = match refusal(&tx)? {
                None if envelope::wrap_limit(wraps(&tx, key_id)?) == WrapLimit::Refuse => {
                    Some((503, StoreError::LimitReached))
                }
                other => other,
            };
            if let Some((status, error)) = refused {
                self.record(&tx, &entry(status, None))?;
                commit(tx)?;
                return Err(error);
            }
            add_wrap(&tx, key_id)?;
            commit(tx)?;
        }

        let (version, wrapped, file) = self.write_version(uuid, body)?;
        self.pause();

        let committed = (|| {
            let mut conn = self.conn();
            let tx = immediate(&mut conn)?;
            if let Some((status, error)) = refusal(&tx)? {
                self.record(&tx, &entry(status, None))?;
                commit(tx)?;
                return Err(error);
            }
            let now = nanos(self.clock.now());
            tx.execute(
                "INSERT INTO version
                     (version, uuid, form, gzip, size, stored_at, hidden_at, tenant_key_id, wrapped_key)
                 VALUES (?1, ?2, ?3, ?4, ?5, ?6, NULL, ?7, ?8)",
                params![
                    version.0,
                    uuid.as_bytes(),
                    form_name(found.form),
                    found.gzip,
                    i64::try_from(body.len()).map_err(io)?,
                    now,
                    key_id.bytes(),
                    wrapped,
                ],
            )
            .map_err(io)?;
            tx.execute(
                "UPDATE version SET hidden_at = ?1 WHERE version = ?2",
                params![now, if_match.0],
            )
            .map_err(io)?;
            tx.execute(
                "UPDATE vcon SET current = ?1 WHERE uuid = ?2",
                params![version.0, uuid.as_bytes()],
            )
            .map_err(io)?;
            self.record(&tx, &entry(200, Some(version)))?;
            commit(tx)?;
            Ok(version)
        })();
        if committed.is_err() {
            // Best effort: start-up recovery removes any file left over.
            let _ = self.fs.remove(&file);
        }
        committed
    }

    /// Deletes `uuid`: one transaction marks it deleted, empties its
    /// current version and hides every kept version, with the audit
    /// record. Its versions stay for legal reads until they are purged;
    /// any already as old as `retention` is purged at once, so a zero
    /// period erases the vCon inside the request.
    /// Deleting a deleted or purged uuid succeeds again. A legal hold does
    /// not stop a delete, only a purge.
    pub fn delete(&self, who: &Caller, uuid: Uuid, retention: Retention) -> Result<(), StoreError> {
        let entry = |status| Entry {
            actor: Actor::Key(who.key),
            action: Action::Delete,
            target: Some(uuid.to_string()),
            version: None,
            status,
        };
        let mut conn = self.conn();
        let tx = immediate(&mut conn)?;
        let outcome = match state(&tx, uuid)? {
            State::Absent => Err(StoreError::NotFound),
            State::Gone => Ok(()),
            State::Live(_) => {
                let now = nanos(self.clock.now());
                tx.execute(
                    "UPDATE vcon SET current = NULL, deleted_at = ?1 WHERE uuid = ?2",
                    params![now, uuid.as_bytes()],
                )
                .map_err(io)?;
                tx.execute(
                    "UPDATE version SET hidden_at = ?1 WHERE uuid = ?2 AND hidden_at IS NULL",
                    params![now, uuid.as_bytes()],
                )
                .map_err(io)?;
                Ok(())
            }
        };
        let status = if outcome.is_ok() { 204 } else { 404 };
        self.record(&tx, &entry(status))?;
        commit(tx)?;
        outcome?;
        drop(conn);
        // A version already as old as `retention` is erased now, not at
        // the next sweep.
        self.purge_hidden(Actor::Key(who.key), retention, Some(uuid))?;
        Ok(())
    }

    /// Every version whose object file is missing. Such a version answers
    /// 500 when read; it is reported, never dropped.
    pub fn check(&self) -> Result<Vec<VersionId>, StoreError> {
        let conn = self.conn();
        let mut rows = conn
            .prepare("SELECT version FROM version ORDER BY version")
            .map_err(io)?;
        let versions = rows
            .query_map([], |r| r.get::<_, [u8; 16]>(0))
            .map_err(io)?
            .collect::<Result<Vec<_>, _>>()
            .map_err(io)?;
        let present: std::collections::BTreeSet<_> = self
            .fs
            .list(&self.objects)
            .map_err(io)?
            .into_iter()
            .collect();
        Ok(versions
            .into_iter()
            .map(VersionId)
            .filter(|v| !present.contains(std::ffi::OsStr::new(&v.to_string())))
            .collect())
    }

    /// Seals `body` as a new version of `uuid` under a new data key, wraps
    /// that key under the tenant key, and writes the object file. Returns
    /// the version id, the wrapped key and the file's path.
    fn write_version(
        &self,
        uuid: Uuid,
        body: &[u8],
    ) -> Result<(VersionId, Vec<u8>, PathBuf), StoreError> {
        let version = VersionId::generate();
        let ctx = Context::Version {
            tenant: self.tenant,
            uuid,
            version,
        };
        let data_key = Key::generate().map_err(|_| StoreError::Io)?;
        let blob = envelope::seal(&data_key, &ctx, body);
        let wrapped = envelope::wrap(&self.tenant_key, &ctx, &data_key);
        let file = self.write_object(version, &blob)?;
        Ok((version, wrapped, file))
    }

    /// In tests, runs the hook set for the gap between work done outside
    /// the catalog lock and the transaction that confirms it.
    fn pause(&self) {
        #[cfg(test)]
        {
            let hook = self.hooks.between.lock().expect("hook").take();
            if let Some(hook) = hook {
                hook();
            }
        }
    }

    /// Writes a sealed body in order: a temporary file, synced,
    /// renamed to its version id, and the directory synced. Returns the
    /// file's path.
    fn write_object(&self, version: VersionId, blob: &[u8]) -> Result<PathBuf, StoreError> {
        let tmp = self.objects.join(format!("{version}.tmp"));
        let file = self.objects.join(version.to_string());
        let written = self
            .fs
            .write_new(&tmp, blob)
            .and_then(|()| self.fs.sync_file(&tmp))
            .and_then(|()| self.fs.rename(&tmp, &file))
            .and_then(|()| self.fs.sync_dir(&self.objects));
        if written.is_err() {
            // Best effort: start-up recovery removes any file left over.
            let _ = self.fs.remove(&tmp);
            let _ = self.fs.remove(&file);
            return Err(StoreError::Io);
        }
        Ok(file)
    }

    /// The plaintext of a stored version. A missing row or file, or one
    /// that does not open, is damage.
    fn read_version(
        &self,
        tx: &Transaction,
        uuid: Uuid,
        version: VersionId,
    ) -> Result<Vec<u8>, StoreError> {
        let row = version_row(tx, version)?;
        self.open_version(uuid, version, &row.wrapped)
    }

    /// Opens a version's object file with its wrapped data key.
    fn open_version(
        &self,
        uuid: Uuid,
        version: VersionId,
        wrapped: &[u8],
    ) -> Result<Vec<u8>, StoreError> {
        let ctx = Context::Version {
            tenant: self.tenant,
            uuid,
            version,
        };
        let data_key =
            envelope::unwrap(&self.tenant_key, &ctx, wrapped).map_err(|_| StoreError::Damaged)?;
        let blob = self
            .fs
            .read(&self.objects.join(version.to_string()))
            .map_err(|_| StoreError::Damaged)?;
        envelope::open(&data_key, &ctx, &blob).map_err(|_| StoreError::Damaged)
    }

    /// How many versions the Store keeps and how many uuids are held.
    pub fn counts(&self) -> Result<Counts, StoreError> {
        let conn = self.conn();
        let count = |sql| -> Result<u64, StoreError> {
            let n: i64 = conn.query_row(sql, [], |r| r.get(0)).map_err(io)?;
            u64::try_from(n).map_err(io)
        };
        Ok(Counts {
            versions: count("SELECT count(*) FROM version")?,
            held: count("SELECT count(*) FROM vcon WHERE hold != 0")?,
        })
    }

    /// Writes the tenant's log after `since` as JSON lines, oldest first.
    /// Not recorded: reading the log changes nothing.
    pub fn export(
        &self,
        since: Option<i64>,
        out: &mut dyn std::io::Write,
    ) -> Result<(), audit::AuditError> {
        audit::export(&self.conn(), since, out)
    }

    /// A page of the tenant's log as JSON lines, the lines `export` writes
    /// for the same records. The read is recorded as audit_read in the same
    /// log and transaction, after the page is read, so a page never holds
    /// its own record.
    pub fn audit_page(
        &self,
        who: &Caller,
        query: &audit::AuditQuery,
    ) -> Result<Vec<u8>, StoreError> {
        let mut conn = self.conn();
        let tx = immediate(&mut conn)?;
        let mut out = Vec::new();
        let read = audit::page(&tx, query, &mut out);
        self.record(
            &tx,
            &Entry {
                actor: Actor::Key(who.key),
                action: Action::AuditRead,
                target: None,
                version: None,
                status: if read.is_ok() { 200 } else { 500 },
            },
        )?;
        commit(tx)?;
        read.map(|()| out).map_err(io)
    }

    /// Records, in a transaction of its own, an action on the tenant taken
    /// elsewhere, such as a change to its retention period.
    pub fn log(&self, entry: &Entry) -> Result<(), StoreError> {
        let mut conn = self.conn();
        let tx = immediate(&mut conn)?;
        self.record(&tx, entry)?;
        commit(tx)
    }

    /// Makes every audit insert on this Store's connection fail, for tests.
    #[cfg(test)]
    pub(crate) fn break_audit(&self) {
        self.conn()
            .execute_batch(
                "CREATE TEMP TRIGGER no_audit BEFORE INSERT ON audit
                 BEGIN SELECT RAISE(ABORT, 'no audit'); END;",
            )
            .expect("trigger");
    }

    /// The catalog connection, for tests that damage or inspect it.
    #[cfg(test)]
    pub(crate) fn conn_for_test(&self) -> MutexGuard<'_, Connection> {
        self.conn()
    }

    /// The tenant's log as `export` writes it, for tests.
    #[cfg(test)]
    pub(crate) fn records(&self) -> Vec<serde_json::Value> {
        crate::audit::testing::exported(&self.conn(), None)
    }

    /// Counts, in a transaction of its own, an event the tenant's log
    /// keeps one record per minute for, such as a failed check of one of
    /// its keys.
    pub fn count(&self, entry: &Entry) -> Result<(), StoreError> {
        let mut conn = self.conn();
        let tx = immediate(&mut conn)?;
        audit::count(&tx, entry, self.clock.now()).map_err(|_| StoreError::AuditFailed)?;
        commit(tx)
    }

    /// Writes an audit record in `tx`; if it cannot, nothing is done.
    fn record(&self, tx: &Transaction, entry: &Entry) -> Result<(), StoreError> {
        audit::record(tx, entry, self.clock.now()).map_err(|_| StoreError::AuditFailed)
    }

    fn conn(&self) -> MutexGuard<'_, Connection> {
        self.conn
            .lock()
            .expect("no thread panicked holding the catalog")
    }
}

/// What the catalog says about a uuid.
enum State {
    /// Never stored.
    Absent,
    /// Stored, with this current version.
    Live(VersionId),
    /// Deleted or purged.
    Gone,
}

fn state(tx: &Transaction, uuid: Uuid) -> Result<State, StoreError> {
    let purged = tx
        .query_row(
            "SELECT 1 FROM tombstone WHERE uuid = ?1",
            [uuid.as_bytes()],
            |_| Ok(()),
        )
        .optional()
        .map_err(io)?;
    if purged.is_some() {
        return Ok(State::Gone);
    }
    let row: Option<(Option<[u8; 16]>, Option<i64>)> = tx
        .query_row(
            "SELECT current, deleted_at FROM vcon WHERE uuid = ?1",
            [uuid.as_bytes()],
            |r| Ok((r.get(0)?, r.get(1)?)),
        )
        .optional()
        .map_err(io)?;
    match row {
        None => Ok(State::Absent),
        Some((_, Some(_))) => Ok(State::Gone),
        Some((Some(current), None)) => Ok(State::Live(VersionId(current))),
        // A live row always names its current version.
        Some((None, None)) => Err(StoreError::Damaged),
    }
}

/// What a version row holds that a read needs.
struct VersionRow {
    form: vcon::Form,
    gzip: bool,
    wrapped: Vec<u8>,
}

/// A version's row. One that is missing or unreadable is damage.
fn version_row(tx: &Transaction, version: VersionId) -> Result<VersionRow, StoreError> {
    let row: Option<(String, bool, Vec<u8>)> = tx
        .query_row(
            "SELECT form, gzip, wrapped_key FROM version WHERE version = ?1",
            [version.0],
            |r| Ok((r.get(0)?, r.get(1)?, r.get(2)?)),
        )
        .optional()
        .map_err(io)?;
    let (form, gzip, wrapped) = row.ok_or(StoreError::Damaged)?;
    let form = match form.as_str() {
        "unsigned" => vcon::Form::Unsigned,
        "signed" => vcon::Form::Signed,
        "encrypted" => vcon::Form::Encrypted,
        _ => return Err(StoreError::Damaged),
    };
    Ok(VersionRow {
        form,
        gzip,
        wrapped,
    })
}

/// Data keys wrapped so far under the tenant key `key`.
fn wraps(tx: &Transaction, key: KeyId) -> Result<u64, StoreError> {
    let wraps: Option<i64> = tx
        .query_row(
            "SELECT wraps FROM key_use WHERE tenant_key_id = ?1",
            [key.bytes()],
            |r| r.get(0),
        )
        .optional()
        .map_err(io)?;
    u64::try_from(wraps.unwrap_or(0)).map_err(io)
}

fn add_wrap(tx: &Transaction, key: KeyId) -> Result<(), StoreError> {
    tx.execute(
        "INSERT INTO key_use (tenant_key_id, wraps) VALUES (?1, 1)
         ON CONFLICT (tenant_key_id) DO UPDATE SET wraps = wraps + 1",
        [key.bytes()],
    )
    .map_err(io)?;
    Ok(())
}

fn immediate(conn: &mut Connection) -> Result<Transaction<'_>, StoreError> {
    conn.transaction_with_behavior(TransactionBehavior::Immediate)
        .map_err(io)
}

fn commit(tx: Transaction) -> Result<(), StoreError> {
    tx.commit().map_err(io)
}

fn io<E>(_: E) -> StoreError {
    StoreError::Io
}

fn form_name(form: vcon::Form) -> &'static str {
    match form {
        vcon::Form::Unsigned => "unsigned",
        vcon::Form::Signed => "signed",
        vcon::Form::Encrypted => "encrypted",
    }
}

/// Stored vCons, for tests.
#[cfg(test)]
pub(crate) mod testing {
    use std::collections::BTreeSet;

    use vcon::{Form, Inspection, Limits, inspect};

    /// An unsigned vCon with this uuid, and what `inspect` makes of it.
    pub(crate) fn vcon(uuid: &str, subject: &str) -> (Inspection, Vec<u8>) {
        let body = format!(
            r#"{{"uuid": "{uuid}", "created_at": "2026-10-23T12:00:00Z", "subject": "{subject}", "parties": []}}"#
        )
        .into_bytes();
        let found = inspect(&body, None, &Limits { inflated: 1 << 20 }, &BTreeSet::new())
            .expect("a valid vCon");
        assert_eq!(found.form, Form::Unsigned);
        (found, body)
    }
}

#[cfg(test)]
mod tests {
    use std::collections::BTreeSet;

    use serde_json::Value;
    use time::macros::datetime;
    use vcon::{Limits, inspect};

    use super::testing::vcon;
    use super::*;
    use crate::audit::testing::exported;
    use crate::catalog::testing::columns_taking_null;
    use crate::clock::ManualClock;
    use crate::fs::{CrashFs, Op};

    const UUID: &str = "01a07da8-c2bb-83e5-b9a2-279e0d16bc46";

    /// Test-only points where a test can step in.
    #[derive(Default)]
    pub(super) struct Hooks {
        /// Runs once, between work done outside the catalog lock (a
        /// write's files, a read's decryption) and the transaction that
        /// confirms it.
        pub(super) between: Mutex<Option<Box<dyn FnOnce() + Send>>>,
    }

    pub(super) struct Fixture {
        pub(super) dir: tempfile::TempDir,
        pub(super) fs: Arc<CrashFs>,
        pub(super) clock: Arc<ManualClock>,
        /// The tenant key, wrapped under `master`, so a test can open the
        /// store again with the same key.
        master: Key,
        wrapped: Vec<u8>,
        pub(super) store: Store,
        pub(super) caller: Caller,
    }

    pub(super) const TENANT: TenantId = TenantId([0x11; 16]);

    impl Fixture {
        pub(super) fn open(&self) -> Result<Store, StoreError> {
            let ctx = Context::Tenant { tenant: TENANT };
            let tenant_key = envelope::unwrap(&self.master, &ctx, &self.wrapped).expect("unwraps");
            Store::open(
                self.dir.path(),
                TENANT,
                tenant_key,
                self.fs.clone(),
                self.clock.clone(),
            )
        }

        /// The tenant key the Store was made under.
        pub(super) fn tenant_key(&self) -> Key {
            let ctx = Context::Tenant { tenant: TENANT };
            envelope::unwrap(&self.master, &ctx, &self.wrapped).expect("unwraps")
        }

        pub(super) fn tenant(&self) -> TenantId {
            TENANT
        }

        /// Opens the store again as a command beside a server would.
        fn open_beside(&self) -> Result<Store, StoreError> {
            let ctx = Context::Tenant { tenant: TENANT };
            let tenant_key = envelope::unwrap(&self.master, &ctx, &self.wrapped).expect("unwraps");
            Store::open_beside(
                self.dir.path(),
                TENANT,
                tenant_key,
                self.fs.clone(),
                self.clock.clone(),
            )
        }

        /// Closes the store and opens it again, as a restart would.
        pub(super) fn reopen(&mut self) {
            let placeholder = self.open().expect("opens");
            drop(std::mem::replace(&mut self.store, placeholder));
            self.store = self.open().expect("reopens");
        }
    }

    pub(super) fn fixture() -> Fixture {
        let dir = tempfile::tempdir().expect("temp dir");
        let fs = Arc::new(CrashFs::new(dir.path()));
        let clock = Arc::new(ManualClock::at(datetime!(2026-10-23 12:00 UTC)));
        let master = Key::generate().expect("master key");
        let tenant_key = Key::generate().expect("tenant key");
        let wrapped = envelope::wrap(&master, &Context::Tenant { tenant: TENANT }, &tenant_key);
        let store =
            Store::open(dir.path(), TENANT, tenant_key, fs.clone(), clock.clone()).expect("opens");
        let caller = Caller {
            key: Key::generate().expect("caller key").id(),
        };
        Fixture {
            dir,
            fs,
            clock,
            master,
            wrapped,
            store,
            caller,
        }
    }

    fn audit_log(store: &Store) -> Vec<Value> {
        exported(&store.conn(), None)
    }

    fn count(store: &Store, sql: &str) -> i64 {
        store.conn().query_row(sql, [], |r| r.get(0)).expect(sql)
    }

    fn wraps(store: &Store) -> i64 {
        store
            .conn()
            .query_row("SELECT coalesce(sum(wraps), 0) FROM key_use", [], |r| {
                r.get(0)
            })
            .expect("wraps")
    }

    /// The body the store holds for `version`, opened as `get` will.
    fn stored_body(f: &Fixture, uuid: Uuid, version: VersionId) -> Vec<u8> {
        let wrapped: Vec<u8> = f
            .store
            .conn()
            .query_row(
                "SELECT wrapped_key FROM version WHERE version = ?1",
                [version.0],
                |r| r.get(0),
            )
            .expect("row");
        let ctx = Context::Version {
            tenant: f.store.tenant,
            uuid,
            version,
        };
        let data_key = envelope::unwrap(&f.store.tenant_key, &ctx, &wrapped).expect("unwraps");
        let blob =
            f.fs.read(&f.store.objects.join(version.to_string()))
                .expect("file");
        envelope::open(&data_key, &ctx, &blob).expect("opens")
    }

    fn audit_page(f: &Fixture, after: i64, limit: u32) -> Result<Vec<u8>, StoreError> {
        f.store
            .audit_page(&f.caller, &audit::AuditQuery { after, limit })
    }

    fn export_text(store: &Store, since: Option<i64>) -> Vec<u8> {
        let mut out = Vec::new();
        store.export(since, &mut out).expect("export");
        out
    }

    #[test]
    fn an_audit_page_is_export_s_lines_and_records_its_read_after_them() {
        let f = fixture();
        for n in 0..3 {
            let (found, body) = vcon(&format!("01a07da8-c2bb-83e5-b9a2-27900000000{n}"), "x");
            f.store.create(&f.caller, &found, &body).expect("stored");
        }
        let before = export_text(&f.store, None);
        assert_eq!(audit_page(&f, 0, 1000), Ok(before.clone()));
        let read = audit_log(&f.store);
        assert_eq!(read.len(), 4, "three creates and the read");
        assert_eq!(read[3]["action"], "audit_read");
        assert_eq!(read[3]["actor"], f.caller.key.to_string());
        assert_eq!(read[3]["status"], 200);
        assert_eq!(read[3]["seq"], 4);
        assert_eq!(read[3]["target"], Value::Null);
        // A page of two from seq 1: the second and third creates.
        let lines: Vec<&[u8]> = before.split_inclusive(|b| *b == b'\n').collect();
        assert_eq!(audit_page(&f, 1, 2), Ok(lines[1..3].concat()));
        // Past the end: nothing but the earlier reads' records.
        let tail = audit_page(&f, 4, 1000).expect("a page");
        assert_eq!(tail, export_text(&f.store, Some(4))[..tail.len()]);
        assert_eq!(String::from_utf8(tail).expect("UTF-8").lines().count(), 1);
    }

    /// A record that cannot be read back (a status past any HTTP status)
    /// fails the page, and the read is still recorded, as 500.
    #[test]
    fn an_audit_page_that_cannot_be_read_is_a_failure_and_recorded() {
        let f = fixture();
        f.store
            .conn_for_test()
            .execute(
                "INSERT INTO audit (at, actor, action, status, count)
                 VALUES ('2026-10-23T12:00:00Z', 'cli', 'read', 70000, 1)",
                [],
            )
            .expect("damaged");
        assert_eq!(audit_page(&f, 0, 10), Err(StoreError::Io));
        let mut recorded = Vec::new();
        f.store.export(Some(1), &mut recorded).expect("export");
        let recorded: Value = serde_json::from_slice(&recorded).expect("one record");
        assert_eq!(recorded["action"], "audit_read");
        assert_eq!(recorded["status"], 500);
    }

    #[test]
    fn an_audit_page_whose_record_fails_is_refused() {
        let f = fixture();
        f.store.break_audit();
        assert_eq!(audit_page(&f, 0, 10), Err(StoreError::AuditFailed));
    }

    #[test]
    fn a_new_vcon_is_written_in_order_and_recorded() {
        let f = fixture();
        let (found, body) = vcon(UUID, "first");
        let ops = f.fs.log().len();
        let Ok(Created::New(version)) = f.store.create(&f.caller, &found, &body) else {
            panic!("a new vCon")
        };
        let (tmp, done) = (
            f.store.objects.join(format!("{version}.tmp")),
            f.store.objects.join(version.to_string()),
        );
        assert_eq!(
            f.fs.log()[ops..],
            [
                Op::Write(tmp.clone()),
                Op::SyncFile(tmp.clone()),
                Op::Rename(tmp, done),
                Op::SyncDir(f.store.objects.clone()),
            ]
        );
        assert_eq!(stored_body(&f, found.uuid, version), body);
        assert_eq!(
            count(
                &f.store,
                "SELECT count(*) FROM version WHERE hidden_at IS NULL"
            ),
            1
        );
        let current: Vec<u8> = f
            .store
            .conn()
            .query_row(
                "SELECT current FROM vcon WHERE uuid = ?1",
                [found.uuid.as_bytes()],
                |r| r.get(0),
            )
            .expect("vcon row");
        assert_eq!(current, version.0);
        assert_eq!(wraps(&f.store), 1, "one data key wrapped");
        let log = audit_log(&f.store);
        assert_eq!(log.len(), 1);
        assert_eq!(log[0]["action"], "create");
        assert_eq!(log[0]["actor"], f.caller.key.to_string());
        assert_eq!(log[0]["target"], UUID);
        assert_eq!(log[0]["version"], version.to_string());
        assert_eq!(log[0]["status"], 201);
    }

    #[test]
    fn the_same_body_again_is_same_and_writes_nothing() {
        let f = fixture();
        let (found, body) = vcon(UUID, "first");
        let Ok(Created::New(version)) = f.store.create(&f.caller, &found, &body) else {
            panic!("a new vCon")
        };
        let ops = f.fs.log().len();
        assert_eq!(
            f.store.create(&f.caller, &found, &body),
            Ok(Created::Same(version))
        );
        assert_eq!(f.fs.log().len(), ops, "no file written");
        assert_eq!(count(&f.store, "SELECT count(*) FROM version"), 1);
        assert_eq!(wraps(&f.store), 1, "nothing wrapped");
        let log = audit_log(&f.store);
        assert_eq!(
            (log[1]["status"].clone(), log[1]["version"].clone()),
            (200.into(), version.to_string().into())
        );
    }

    #[test]
    fn another_body_for_a_stored_uuid_is_a_conflict_and_writes_nothing() {
        let f = fixture();
        let (found, body) = vcon(UUID, "first");
        f.store.create(&f.caller, &found, &body).expect("stored");
        let ops = f.fs.log().len();
        // The same length as the first body, so only the bytes differ.
        let (found, other) = vcon(UUID, "other");
        assert_eq!(
            f.store.create(&f.caller, &found, &other),
            Err(StoreError::Conflict)
        );
        assert_eq!(f.fs.log().len(), ops, "no file written");
        assert_eq!(wraps(&f.store), 1, "nothing wrapped");
        assert_eq!(audit_log(&f.store)[1]["status"], 409);
    }

    #[test]
    fn a_deleted_or_purged_uuid_is_gone_and_writes_nothing() {
        for sql in [
            "INSERT INTO vcon (uuid, current, deleted_at) VALUES (?1, NULL, 1)",
            "INSERT INTO tombstone (uuid, purged_at) VALUES (?1, 1)",
        ] {
            let f = fixture();
            let (found, body) = vcon(UUID, "first");
            f.store
                .conn()
                .execute(sql, [found.uuid.as_bytes()])
                .expect(sql);
            let ops = f.fs.log().len();
            assert_eq!(
                f.store.create(&f.caller, &found, &body),
                Err(StoreError::Gone),
                "{sql}"
            );
            assert_eq!(f.fs.log().len(), ops, "no file written");
            assert_eq!(wraps(&f.store), 0, "nothing wrapped");
            assert_eq!(audit_log(&f.store)[0]["status"], 410);
        }
    }

    #[test]
    fn counts_include_hidden_versions_and_held_uuids() {
        let f = fixture();
        let none = Counts {
            versions: 0,
            held: 0,
        };
        assert_eq!(f.store.counts(), Ok(none));
        let (found, old) = stored(&f, "first");
        let (_, newer) = vcon(UUID, "second");
        f.store
            .replace(&f.caller, old, &found, &newer)
            .expect("replaced");
        f.store
            .delete(&f.caller, found.uuid, Retention::Years(7))
            .expect("deleted");
        let (other, body) = vcon("01a07da8-c2bb-83e5-b9a2-279e0d16bc47", "other");
        f.store.create(&f.caller, &other, &body).expect("stored");
        assert_eq!(
            f.store.counts(),
            Ok(Counts {
                versions: 3,
                held: 0
            }),
            "two hidden versions and a live one"
        );
        f.store
            .conn()
            .execute(
                "UPDATE vcon SET hold = 1 WHERE uuid = ?1",
                [found.uuid.as_bytes()],
            )
            .expect("a hold");
        assert_eq!(f.store.counts().map(|c| c.held), Ok(1));
    }

    #[test]
    fn a_negative_wrap_count_is_damage_and_stores_nothing() {
        let f = fixture();
        f.store
            .conn()
            .execute(
                "INSERT INTO key_use (tenant_key_id, wraps) VALUES (?1, -1)",
                [f.store.tenant_key.id().bytes()],
            )
            .expect("a damaged count");
        let (found, body) = vcon(UUID, "first");
        let ops = f.fs.log().len();
        assert_eq!(
            f.store.create(&f.caller, &found, &body),
            Err(StoreError::Io)
        );
        assert_eq!(f.fs.log().len(), ops, "no file written");
    }

    #[test]
    fn a_tenant_key_at_its_wrap_limit_refuses_new_vcons() {
        let f = fixture();
        f.store
            .conn()
            .execute(
                "INSERT INTO key_use (tenant_key_id, wraps) VALUES (?1, ?2)",
                params![f.store.tenant_key.id().bytes(), 1_i64 << 32],
            )
            .expect("at the limit");
        let (found, body) = vcon(UUID, "first");
        let ops = f.fs.log().len();
        assert_eq!(
            f.store.create(&f.caller, &found, &body),
            Err(StoreError::LimitReached)
        );
        assert_eq!(f.fs.log().len(), ops, "no file written");
        assert_eq!(audit_log(&f.store)[0]["status"], 503);
    }

    #[test]
    fn a_missing_current_file_is_damage_not_a_conflict() {
        let f = fixture();
        let (found, body) = vcon(UUID, "first");
        let Ok(Created::New(version)) = f.store.create(&f.caller, &found, &body) else {
            panic!("a new vCon")
        };
        f.fs.remove(&f.store.objects.join(version.to_string()))
            .expect("remove");
        assert_eq!(
            f.store.create(&f.caller, &found, &body),
            Err(StoreError::Damaged)
        );
        assert_eq!(audit_log(&f.store)[1]["status"], 500);
    }

    #[test]
    fn a_racing_create_that_loses_gets_a_conflict_and_leaves_no_file() {
        let f = Arc::new(fixture());
        let (found, body) = vcon(UUID, "first");
        let winner = f.clone();
        *f.store.hooks.between.lock().expect("hook") = Some(Box::new(move || {
            let (found, other) = vcon(UUID, "second");
            assert!(matches!(
                winner.store.create(&winner.caller, &found, &other),
                Ok(Created::New(_))
            ));
        }));
        assert_eq!(
            f.store.create(&f.caller, &found, &body),
            Err(StoreError::Conflict)
        );
        assert_eq!(
            count(&f.store, "SELECT count(*) FROM version"),
            1,
            "only the winner"
        );
        assert_eq!(
            f.fs.list(&f.store.objects).expect("list").len(),
            1,
            "the loser's file is gone"
        );
        assert_eq!(wraps(&f.store), 2, "both wraps are counted");
        let statuses: Vec<_> = audit_log(&f.store)
            .iter()
            .map(|r| r["status"].clone())
            .collect();
        assert_eq!(statuses, [201, 409]);
    }

    #[test]
    fn a_create_whose_audit_record_fails_leaves_no_row_and_no_file() {
        let f = fixture();
        // Every audit insert now fails, as a full disk or a broken log would.
        f.store
            .conn()
            .execute_batch(
                "CREATE TEMP TRIGGER no_audit BEFORE INSERT ON audit
                 BEGIN SELECT RAISE(ABORT, 'audit fails'); END;",
            )
            .expect("trigger");
        let (found, body) = vcon(UUID, "first");
        assert_eq!(
            f.store.create(&f.caller, &found, &body),
            Err(StoreError::AuditFailed)
        );
        assert_eq!(count(&f.store, "SELECT count(*) FROM version"), 0);
        assert_eq!(count(&f.store, "SELECT count(*) FROM vcon"), 0);
        assert_eq!(
            f.fs.list(&f.store.objects).expect("list"),
            Vec::<std::ffi::OsString>::new()
        );
    }

    fn get_entry_status(f: &Fixture) -> Value {
        audit_log(&f.store).last().expect("a record")["status"].clone()
    }

    #[test]
    fn a_stored_vcon_reads_back_byte_for_byte_and_is_recorded() {
        let f = fixture();
        let (found, body) = vcon(UUID, "first");
        let Ok(Created::New(version)) = f.store.create(&f.caller, &found, &body) else {
            panic!("a new vCon")
        };
        let stored = f.store.get(&f.caller, found.uuid).expect("reads");
        assert_eq!(
            stored,
            Stored {
                version,
                form: vcon::Form::Unsigned,
                gzip: false,
                body
            }
        );
        let log = audit_log(&f.store);
        let read = log.last().expect("a record");
        assert_eq!(read["action"], "read");
        assert_eq!(read["status"], 200);
        assert_eq!(read["version"], version.to_string());
        assert_eq!(read["target"], UUID);
    }

    #[test]
    fn a_uuid_never_stored_is_not_found() {
        let f = fixture();
        let (found, _) = vcon(UUID, "first");
        assert_eq!(
            f.store.get(&f.caller, found.uuid),
            Err(StoreError::NotFound)
        );
        assert_eq!(get_entry_status(&f), 404);
    }

    #[test]
    fn a_deleted_or_purged_uuid_is_gone() {
        for sql in [
            "UPDATE vcon SET current = NULL, deleted_at = 1 WHERE uuid = ?1",
            "DELETE FROM vcon WHERE uuid = ?1; INSERT INTO tombstone (uuid, purged_at) VALUES (?1, 1)",
        ] {
            let f = fixture();
            let (found, body) = vcon(UUID, "first");
            f.store.create(&f.caller, &found, &body).expect("stored");
            for statement in sql.split("; ") {
                f.store
                    .conn()
                    .execute(statement, [found.uuid.as_bytes()])
                    .expect(statement);
            }
            assert_eq!(
                f.store.get(&f.caller, found.uuid),
                Err(StoreError::Gone),
                "{sql}"
            );
            assert_eq!(get_entry_status(&f), 410);
        }
    }

    #[test]
    fn a_missing_or_damaged_file_is_a_500_never_an_empty_body() {
        let f = fixture();
        let (found, body) = vcon(UUID, "first");
        let Ok(Created::New(version)) = f.store.create(&f.caller, &found, &body) else {
            panic!("a new vCon")
        };
        let file = f.store.objects.join(version.to_string());
        let blob = f.fs.read(&file).expect("file");
        f.fs.remove(&file).expect("remove");
        assert_eq!(
            f.store.get(&f.caller, found.uuid),
            Err(StoreError::Damaged),
            "missing"
        );
        assert_eq!(get_entry_status(&f), 500);
        let mut bad = blob.clone();
        let last = bad.len() - 1;
        bad[last] ^= 1;
        f.fs.write_new(&file, &bad).expect("write");
        assert_eq!(
            f.store.get(&f.caller, found.uuid),
            Err(StoreError::Damaged),
            "tampered"
        );
    }

    #[test]
    fn a_read_racing_a_delete_never_returns_the_body() {
        let f = Arc::new(fixture());
        let (found, body) = vcon(UUID, "first");
        f.store.create(&f.caller, &found, &body).expect("stored");
        let deleter = f.clone();
        *f.store.hooks.between.lock().expect("hook") = Some(Box::new(move || {
            deleter
                .store
                .conn()
                .execute(
                    "UPDATE vcon SET current = NULL, deleted_at = 1 WHERE uuid = ?1",
                    [found.uuid.as_bytes()],
                )
                .expect("delete");
        }));
        assert_eq!(f.store.get(&f.caller, found.uuid), Err(StoreError::Gone));
        assert_eq!(get_entry_status(&f), 410);
    }

    #[test]
    fn a_read_racing_a_new_version_starts_again_and_returns_the_new_one() {
        let f = Arc::new(fixture());
        let (found, body) = vcon(UUID, "first");
        f.store.create(&f.caller, &found, &body).expect("stored");
        let (_, newer) = vcon(UUID, "newer");
        let (version, wrapped, _) = f.store.write_version(found.uuid, &newer).expect("written");
        let writer = f.clone();
        *f.store.hooks.between.lock().expect("hook") = Some(Box::new(move || {
            let conn = writer.store.conn();
            conn.execute(
                "INSERT INTO version
                     (version, uuid, form, gzip, size, stored_at, tenant_key_id, wrapped_key)
                 VALUES (?1, ?2, 'unsigned', 0, 0, 0, ?3, ?4)",
                params![
                    version.0,
                    found.uuid.as_bytes(),
                    writer.store.tenant_key.id().bytes(),
                    wrapped
                ],
            )
            .expect("new version");
            conn.execute(
                "UPDATE vcon SET current = ?1 WHERE uuid = ?2",
                params![version.0, found.uuid.as_bytes()],
            )
            .expect("made current");
        }));
        let stored = f.store.get(&f.caller, found.uuid).expect("reads");
        assert_eq!((stored.version, stored.body), (version, newer));
        let reads = audit_log(&f.store)
            .iter()
            .filter(|r| r["action"] == "read")
            .count();
        assert_eq!(reads, 1, "one record for the one request");
    }

    #[test]
    fn every_form_and_the_gzip_flag_read_back_as_stored() {
        use base64::Engine;
        use std::io::Write;
        let f = fixture();
        let uuids = [
            "01a07da8-c2bb-83e5-b9a2-279e0d16bc01",
            "01a07da8-c2bb-83e5-b9a2-279e0d16bc02",
            "01a07da8-c2bb-83e5-b9a2-279e0d16bc03",
        ];
        let (_, unsigned) = vcon(uuids[0], "plain");
        let mut gz = flate2::write::GzEncoder::new(Vec::new(), flate2::Compression::default());
        gz.write_all(&unsigned).expect("gzip");
        let gzipped = gz.finish().expect("gzip");
        let (_, inner) = vcon(uuids[1], "signed");
        let signed = format!(
            r#"{{"payload": "{}", "signatures": [{{"header": {{}}}}]}}"#,
            base64::engine::general_purpose::URL_SAFE_NO_PAD.encode(inner)
        )
        .into_bytes();
        let encrypted = format!(
            r#"{{"unprotected": {{"uuid": "{}"}}, "ciphertext": "x", "recipients": []}}"#,
            uuids[2]
        )
        .into_bytes();
        let cases = [
            (gzipped, vcon::Form::Unsigned, true),
            (signed, vcon::Form::Signed, false),
            (encrypted, vcon::Form::Encrypted, false),
        ];
        for (body, form, gzip) in cases {
            let found = inspect(&body, None, &Limits { inflated: 1 << 20 }, &BTreeSet::new())
                .expect("a valid vCon");
            assert_eq!((found.form, found.gzip), (form, gzip));
            f.store.create(&f.caller, &found, &body).expect("stored");
            let stored = f.store.get(&f.caller, found.uuid).expect("reads");
            assert_eq!((stored.form, stored.gzip, stored.body), (form, gzip, body));
        }
    }

    /// Stores `subject` under UUID and returns its version.
    fn stored(f: &Fixture, subject: &str) -> (Inspection, VersionId) {
        let (found, body) = vcon(UUID, subject);
        let Ok(Created::New(version)) = f.store.create(&f.caller, &found, &body) else {
            panic!("a new vCon")
        };
        (found, version)
    }

    #[test]
    fn a_replace_makes_a_new_current_version_and_hides_the_old_one() {
        let f = fixture();
        let (found, old) = stored(&f, "first");
        let (_, newer) = vcon(UUID, "second");
        let version = f
            .store
            .replace(&f.caller, old, &found, &newer)
            .expect("replaced");
        assert_ne!(version, old);
        let stored = f.store.get(&f.caller, found.uuid).expect("reads");
        assert_eq!((stored.version, stored.body), (version, newer));
        let hidden: Option<i64> = f
            .store
            .conn()
            .query_row(
                "SELECT hidden_at FROM version WHERE version = ?1",
                [old.0],
                |r| r.get(0),
            )
            .expect("the old row is kept");
        assert!(hidden.is_some(), "the old version is hidden");
        assert!(
            f.fs.read(&f.store.objects.join(old.to_string())).is_ok(),
            "its file is kept"
        );
        let record = audit_log(&f.store)
            .into_iter()
            .find(|r| r["action"] == "replace")
            .expect("a replace record");
        assert_eq!(
            (record["status"].clone(), record["version"].clone()),
            (200.into(), version.to_string().into())
        );
    }

    #[test]
    fn a_replace_of_a_missing_or_gone_uuid_is_refused() {
        let f = fixture();
        let (found, body) = vcon(UUID, "first");
        let any = VersionId([7; 16]);
        assert_eq!(
            f.store.replace(&f.caller, any, &found, &body),
            Err(StoreError::NotFound)
        );
        assert_eq!(get_entry_status(&f), 404);
        let (found, old) = stored(&f, "first");
        f.store
            .conn()
            .execute(
                "UPDATE vcon SET current = NULL, deleted_at = 1 WHERE uuid = ?1",
                [found.uuid.as_bytes()],
            )
            .expect("delete");
        assert_eq!(
            f.store.replace(&f.caller, old, &found, &body),
            Err(StoreError::Gone)
        );
        assert_eq!(get_entry_status(&f), 410);
    }

    #[test]
    fn a_stale_if_match_is_refused_and_writes_nothing() {
        let f = fixture();
        let (found, _) = stored(&f, "first");
        let ops = f.fs.log().len();
        let (_, newer) = vcon(UUID, "second");
        assert_eq!(
            f.store
                .replace(&f.caller, VersionId([7; 16]), &found, &newer),
            Err(StoreError::PreconditionFailed)
        );
        assert_eq!(f.fs.log().len(), ops, "no file written");
        assert_eq!(wraps(&f.store), 1, "nothing wrapped");
        assert_eq!(get_entry_status(&f), 412);
    }

    #[test]
    fn a_signed_or_encrypted_current_version_is_immutable() {
        use base64::Engine;
        let f = fixture();
        let (found, old) = stored(&f, "first");
        let (_, inner) = vcon(UUID, "signed");
        let signed = format!(
            r#"{{"payload": "{}", "signatures": [{{"header": {{}}}}]}}"#,
            base64::engine::general_purpose::URL_SAFE_NO_PAD.encode(inner)
        )
        .into_bytes();
        let signed_found = inspect(
            &signed,
            None,
            &Limits { inflated: 1 << 20 },
            &BTreeSet::new(),
        )
        .expect("a signed vCon");
        let version = f
            .store
            .replace(&f.caller, old, &signed_found, &signed)
            .expect("an unsigned vCon may be replaced by a signed one");
        let (_, newer) = vcon(UUID, "second");
        let ops = f.fs.log().len();
        assert_eq!(
            f.store.replace(&f.caller, version, &found, &newer),
            Err(StoreError::Immutable)
        );
        assert_eq!(f.fs.log().len(), ops, "no file written");
        assert_eq!(get_entry_status(&f), 409);

        let g = fixture();
        let encrypted = format!(
            r#"{{"unprotected": {{"uuid": "{UUID}"}}, "ciphertext": "x", "recipients": []}}"#
        )
        .into_bytes();
        let enc_found = inspect(
            &encrypted,
            None,
            &Limits { inflated: 1 << 20 },
            &BTreeSet::new(),
        )
        .expect("an encrypted vCon");
        let Ok(Created::New(enc_version)) = g.store.create(&g.caller, &enc_found, &encrypted)
        else {
            panic!("stored")
        };
        assert_eq!(
            g.store.replace(&g.caller, enc_version, &found, &newer),
            Err(StoreError::Immutable)
        );
    }

    #[test]
    fn a_replace_at_the_wrap_limit_is_refused() {
        let f = fixture();
        let (found, old) = stored(&f, "first");
        f.store
            .conn()
            .execute("UPDATE key_use SET wraps = ?1", [1_i64 << 32])
            .expect("at the limit");
        let (_, newer) = vcon(UUID, "second");
        assert_eq!(
            f.store.replace(&f.caller, old, &found, &newer),
            Err(StoreError::LimitReached)
        );
        assert_eq!(get_entry_status(&f), 503);
    }

    #[test]
    fn of_two_replaces_racing_on_one_version_the_second_to_commit_loses() {
        let f = Arc::new(fixture());
        let (found, old) = stored(&f, "first");
        let winner = f.clone();
        let racer = found;
        *f.store.hooks.between.lock().expect("hook") = Some(Box::new(move || {
            let (_, body) = vcon(UUID, "winner");
            winner
                .store
                .replace(&winner.caller, old, &racer, &body)
                .expect("wins");
        }));
        let (_, body) = vcon(UUID, "loser");
        assert_eq!(
            f.store.replace(&f.caller, old, &found, &body),
            Err(StoreError::PreconditionFailed)
        );
        assert_eq!(
            f.fs.list(&f.store.objects).expect("list").len(),
            2,
            "first and winner only"
        );
        assert_eq!(wraps(&f.store), 3, "every wrap counted");
        assert_eq!(get_entry_status(&f), 412);
    }

    #[test]
    fn two_threads_replacing_with_the_same_if_match_one_wins() {
        let f = Arc::new(fixture());
        let (_, old) = stored(&f, "first");
        let threads: Vec<_> = ["one", "two"]
            .into_iter()
            .map(|subject| {
                let f = f.clone();
                std::thread::spawn(move || {
                    let (found, body) = vcon(UUID, subject);
                    f.store.replace(&f.caller, old, &found, &body)
                })
            })
            .collect();
        let results: Vec<_> = threads
            .into_iter()
            .map(|t| t.join().expect("thread"))
            .collect();
        assert_eq!(
            results.iter().filter(|r| r.is_ok()).count(),
            1,
            "{results:?}"
        );
        assert!(
            results.contains(&Err(StoreError::PreconditionFailed)),
            "{results:?}"
        );
    }

    #[test]
    fn a_replace_whose_audit_record_fails_changes_nothing() {
        let f = fixture();
        let (found, old) = stored(&f, "first");
        f.store
            .conn()
            .execute_batch(
                "CREATE TEMP TRIGGER no_audit BEFORE INSERT ON audit
                 BEGIN SELECT RAISE(ABORT, 'audit fails'); END;",
            )
            .expect("trigger");
        let (_, newer) = vcon(UUID, "second");
        assert_eq!(
            f.store.replace(&f.caller, old, &found, &newer),
            Err(StoreError::AuditFailed)
        );
        f.store
            .conn()
            .execute_batch("DROP TRIGGER no_audit")
            .expect("drop trigger");
        assert_eq!(
            f.store.get(&f.caller, found.uuid).expect("reads").version,
            old
        );
        assert_eq!(
            f.fs.list(&f.store.objects).expect("list").len(),
            1,
            "only the first file"
        );
    }

    fn hidden_versions(f: &Fixture) -> i64 {
        count(
            &f.store,
            "SELECT count(*) FROM version WHERE hidden_at IS NOT NULL",
        )
    }

    #[test]
    fn a_deleted_uuid_answers_gone_and_keeps_its_versions_hidden() {
        let f = fixture();
        let (found, old) = stored(&f, "first");
        let (_, newer) = vcon(UUID, "second");
        f.clock.set(datetime!(2026-10-23 13:00 UTC));
        let current = f
            .store
            .replace(&f.caller, old, &found, &newer)
            .expect("replaced");
        f.clock.set(datetime!(2026-10-23 14:00 UTC));
        assert_eq!(
            f.store.delete(&f.caller, found.uuid, Retention::Years(7)),
            Ok(())
        );
        let record = audit_log(&f.store).pop().expect("a record");
        assert_eq!(
            (record["action"].clone(), record["status"].clone()),
            ("delete".into(), 204.into())
        );
        let hidden_at = |version: VersionId| -> i64 {
            f.store
                .conn()
                .query_row(
                    "SELECT hidden_at FROM version WHERE version = ?1",
                    [version.0],
                    |r| r.get(0),
                )
                .expect("hidden")
        };
        assert_eq!(
            hidden_at(old),
            nanos(datetime!(2026-10-23 13:00 UTC)),
            "hidden by the replace"
        );
        assert_eq!(
            hidden_at(current),
            nanos(datetime!(2026-10-23 14:00 UTC)),
            "hidden by the delete"
        );
        assert_eq!(f.store.get(&f.caller, found.uuid), Err(StoreError::Gone));
        let (_, body) = vcon(UUID, "third");
        assert_eq!(
            f.store.create(&f.caller, &found, &body),
            Err(StoreError::Gone)
        );
        assert_eq!(hidden_versions(&f), 2, "both versions hidden");
        assert_eq!(
            f.fs.list(&f.store.objects).expect("list").len(),
            2,
            "files kept for legal reads"
        );
        let (current, deleted_at): (Option<Vec<u8>>, Option<i64>) = f
            .store
            .conn()
            .query_row(
                "SELECT current, deleted_at FROM vcon WHERE uuid = ?1",
                [found.uuid.as_bytes()],
                |r| Ok((r.get(0)?, r.get(1)?)),
            )
            .expect("row");
        assert_eq!(current, None);
        assert_eq!(deleted_at, Some(nanos(datetime!(2026-10-23 14:00 UTC))));
    }

    #[test]
    fn deleting_a_deleted_or_purged_uuid_succeeds_again() {
        let f = fixture();
        let (found, _) = stored(&f, "first");
        f.store
            .delete(&f.caller, found.uuid, Retention::Years(7))
            .expect("deleted");
        assert_eq!(
            f.store.delete(&f.caller, found.uuid, Retention::Years(7)),
            Ok(())
        );
        assert_eq!(get_entry_status(&f), 204);
        let g = fixture();
        g.store
            .conn()
            .execute(
                "INSERT INTO tombstone (uuid, purged_at) VALUES (?1, 1)",
                [found.uuid.as_bytes()],
            )
            .expect("purged");
        assert_eq!(
            g.store.delete(&g.caller, found.uuid, Retention::Years(7)),
            Ok(())
        );
        assert_eq!(get_entry_status(&g), 204);
    }

    #[test]
    fn deleting_a_uuid_never_stored_is_not_found() {
        let f = fixture();
        let (found, _) = vcon(UUID, "first");
        assert_eq!(
            f.store.delete(&f.caller, found.uuid, Retention::Years(7)),
            Err(StoreError::NotFound)
        );
        assert_eq!(get_entry_status(&f), 404);
    }

    #[test]
    fn a_hold_does_not_stop_a_delete() {
        let f = fixture();
        let (found, _) = stored(&f, "first");
        f.store
            .conn()
            .execute(
                "UPDATE vcon SET hold = 1 WHERE uuid = ?1",
                [found.uuid.as_bytes()],
            )
            .expect("hold");
        assert_eq!(
            f.store.delete(&f.caller, found.uuid, Retention::Years(7)),
            Ok(())
        );
        assert_eq!(f.store.get(&f.caller, found.uuid), Err(StoreError::Gone));
    }

    #[test]
    fn a_delete_whose_audit_record_fails_changes_nothing() {
        let f = fixture();
        let (found, version) = stored(&f, "first");
        f.store
            .conn()
            .execute_batch(
                "CREATE TEMP TRIGGER no_audit BEFORE INSERT ON audit
                 BEGIN SELECT RAISE(ABORT, 'audit fails'); END;",
            )
            .expect("trigger");
        assert_eq!(
            f.store.delete(&f.caller, found.uuid, Retention::Years(7)),
            Err(StoreError::AuditFailed)
        );
        f.store
            .conn()
            .execute_batch("DROP TRIGGER no_audit")
            .expect("drop trigger");
        assert_eq!(
            f.store
                .get(&f.caller, found.uuid)
                .expect("still live")
                .version,
            version
        );
        assert_eq!(hidden_versions(&f), 0);
    }

    #[test]
    fn only_the_optional_catalog_columns_may_be_empty() {
        let f = fixture();
        let conn = f.store.conn();
        let tables: [(&str, &[(&str, &str)]); 5] = [
            (
                "vcon",
                &[
                    ("uuid", "x'01'"),
                    ("current", "x'02'"),
                    ("deleted_at", "1"),
                    ("hold", "0"),
                ],
            ),
            (
                "version",
                &[
                    ("version", "x'02'"),
                    ("uuid", "x'01'"),
                    ("form", "'unsigned'"),
                    ("gzip", "0"),
                    ("size", "1"),
                    ("stored_at", "1"),
                    ("hidden_at", "1"),
                    ("tenant_key_id", "x'03'"),
                    ("wrapped_key", "x'04'"),
                ],
            ),
            ("tombstone", &[("uuid", "x'01'"), ("purged_at", "1")]),
            ("key_use", &[("tenant_key_id", "x'03'"), ("wraps", "1")]),
            (
                "audit",
                &[
                    ("at", "'2026-10-23T12:00Z'"),
                    ("actor", "'cli'"),
                    ("action", "'create'"),
                    ("target", "'x'"),
                    ("version", "'y'"),
                    ("status", "201"),
                    ("count", "1"),
                ],
            ),
        ];
        let may_be_empty: Vec<String> = tables
            .iter()
            .flat_map(|(table, row)| columns_taking_null(&conn, table, row))
            .collect();
        assert_eq!(
            may_be_empty,
            [
                "vcon.current",
                "vcon.deleted_at",
                "version.hidden_at",
                "audit.target",
                "audit.version"
            ]
        );
    }

    #[test]
    fn a_new_stores_objects_directory_survives_a_crash() {
        let f = fixture();
        f.fs.crash();
        assert_eq!(
            f.fs.list(f.dir.path()).expect("list"),
            [std::ffi::OsString::from("objects")]
        );
    }

    #[test]
    fn opening_removes_temporary_and_orphaned_files_and_nothing_else() {
        let mut f = fixture();
        let (found, version) = stored(&f, "first");
        let orphan = VersionId([0x0e; 16]);
        let objects = f.store.objects.clone();
        for name in [
            format!("{orphan}"),
            format!("{orphan}.tmp"),
            format!("{version}.tmp"),
            "notes.txt".to_owned(),
        ] {
            f.fs.write_new(&objects.join(name), b"x").expect("write");
        }
        f.reopen();
        let names: Vec<String> =
            f.fs.list(&objects)
                .expect("list")
                .into_iter()
                .map(|n| n.to_string_lossy().into_owned())
                .collect();
        assert_eq!(names, [version.to_string(), "notes.txt".to_owned()]);
        assert_eq!(
            f.store
                .get(&f.caller, found.uuid)
                .expect("still reads")
                .version,
            version
        );
        let log = f.fs.log();
        let last_removal = log
            .iter()
            .rposition(|op| matches!(op, Op::Remove(_)))
            .expect("files were removed");
        assert!(
            log[last_removal..].contains(&Op::SyncDir(objects)),
            "the removals are synced"
        );
    }

    /// A command beside a running server must not take the server's
    /// writes in progress for a crash's leftovers.
    #[test]
    fn opening_beside_a_server_removes_nothing() {
        let f = fixture();
        let (found, version) = stored(&f, "first");
        let orphan = VersionId([0x0e; 16]);
        let objects = f.store.objects.clone();
        let mut written = vec![version.to_string()];
        for name in [
            format!("{orphan}"),
            format!("{orphan}.tmp"),
            format!("{version}.tmp"),
        ] {
            f.fs.write_new(&objects.join(&name), b"x").expect("write");
            written.push(name);
        }
        written.sort();
        let ops = f.fs.log().len();
        let beside = f.open_beside().expect("opens beside");
        let mut names: Vec<String> =
            f.fs.list(&objects)
                .expect("list")
                .into_iter()
                .map(|n| n.to_string_lossy().into_owned())
                .collect();
        names.sort();
        assert_eq!(names, written);
        assert!(
            !f.fs.log()[ops..]
                .iter()
                .any(|op| matches!(op, Op::Remove(_))),
            "nothing removed"
        );
        assert_eq!(
            beside.get(&f.caller, found.uuid).expect("reads").version,
            version
        );
    }

    #[test]
    fn opening_checkpoints_the_wal_with_truncation() {
        let mut f = fixture();
        stored(&f, "first");
        let wal = f.dir.path().join("catalog.sqlite-wal");
        assert!(
            std::fs::metadata(&wal).expect("a WAL").len() > 0,
            "the WAL holds the write"
        );
        f.reopen();
        assert_eq!(std::fs::metadata(&wal).expect("a WAL").len(), 0);
    }

    #[test]
    fn check_reports_a_version_whose_file_is_missing() {
        let f = fixture();
        let (_, version) = stored(&f, "first");
        assert_eq!(f.store.check(), Ok(vec![]));
        f.fs.remove(&f.store.objects.join(version.to_string()))
            .expect("remove");
        assert_eq!(f.store.check(), Ok(vec![version]));
    }

    #[test]
    fn opening_refuses_while_a_reader_stops_the_wal_from_being_truncated() {
        let f = fixture();
        stored(&f, "first");
        let ctx = Context::Tenant { tenant: TENANT };
        let tenant_key = envelope::unwrap(&f.master, &ctx, &f.wrapped).expect("unwraps");
        let catalog_key = envelope::catalog_key(&tenant_key).expect("catalog key");
        let mut reader = catalog::open(&f.dir.path().join("catalog.sqlite"), &catalog_key, "")
            .expect("a second connection");
        let read = reader.transaction().expect("read transaction");
        let _: i64 = read
            .query_row("SELECT count(*) FROM vcon", [], |r| r.get(0))
            .expect("a read holds a snapshot");
        assert!(matches!(f.open(), Err(StoreError::Io)));
        drop(read);
        assert!(f.open().is_ok(), "opens once the reader is done");
    }

    /// The filesystem changes a create or replace makes when nothing
    /// fails: write, sync, rename, sync the directory.
    const WRITE_STEPS: usize = 4;

    /// After a restart, no row lacks its file and every file has a row.
    fn assert_whole(f: &Fixture, step: usize) {
        assert_eq!(
            f.store.check(),
            Ok(vec![]),
            "a row without its file at step {step}"
        );
        for name in f.fs.list(&f.store.objects).expect("list") {
            let name = name.to_string_lossy().into_owned();
            let version = VersionId::parse(&name)
                .unwrap_or_else(|| panic!("{name} left over at step {step}"));
            let rows = count(
                &f.store,
                &format!("SELECT count(*) FROM version WHERE hex(version) = upper('{version}')"),
            );
            assert_eq!(rows, 1, "{name} has no row at step {step}");
        }
    }

    /// Makes the next audit insert fail for every connection, as if the
    /// process died before its commit; `unblock` removes it again.
    pub(super) fn block_commits(f: &Fixture) -> Connection {
        let ctx = Context::Tenant { tenant: TENANT };
        let tenant_key = envelope::unwrap(&f.master, &ctx, &f.wrapped).expect("unwraps");
        let catalog_key = envelope::catalog_key(&tenant_key).expect("catalog key");
        catalog::open(&f.dir.path().join("catalog.sqlite"), &catalog_key, "")
            .expect("second connection")
    }

    fn die_before_commit(f: &Fixture) {
        let other = block_commits(f);
        let fs = f.fs.clone();
        *f.store.hooks.between.lock().expect("hook") = Some(Box::new(move || {
            fs.fail_from(0);
            other
                .execute_batch(
                    "CREATE TRIGGER dead BEFORE INSERT ON audit
                     BEGIN SELECT RAISE(ABORT, 'the process has died'); END;",
                )
                .expect("trigger");
        }));
    }

    pub(super) fn restart(f: &mut Fixture) {
        f.fs.crash();
        f.store
            .conn()
            .execute_batch("DROP TRIGGER IF EXISTS dead")
            .expect("drop trigger");
        f.reopen();
    }

    #[test]
    fn a_create_cut_off_at_any_step_leaves_no_vcon_or_the_whole_vcon() {
        let (found, body) = vcon(UUID, "first");
        for step in 0..=WRITE_STEPS {
            let mut f = fixture();
            f.fs.fail_from(f.fs.log().len() + step);
            let result = f.store.create(&f.caller, &found, &body);
            assert_eq!(
                result.is_ok(),
                step == WRITE_STEPS,
                "step {step}: {result:?}"
            );
            restart(&mut f);
            match f.store.get(&f.caller, found.uuid) {
                Ok(stored) => assert_eq!(stored.body, body, "step {step}"),
                Err(StoreError::NotFound) => {
                    assert!(result.is_err(), "acknowledged, then lost, at step {step}")
                }
                other => panic!("half a vCon at step {step}: {other:?}"),
            }
            assert_whole(&f, step);
        }
    }

    #[test]
    fn a_create_cut_off_between_its_file_and_its_commit_leaves_nothing() {
        let mut f = fixture();
        let (found, body) = vcon(UUID, "first");
        die_before_commit(&f);
        assert!(f.store.create(&f.caller, &found, &body).is_err());
        restart(&mut f);
        assert_eq!(
            f.store.get(&f.caller, found.uuid),
            Err(StoreError::NotFound)
        );
        assert_whole(&f, WRITE_STEPS);
    }

    #[test]
    fn a_replace_cut_off_at_any_step_leaves_the_old_version_or_the_new() {
        let (found, old_body) = vcon(UUID, "first");
        let (_, new_body) = vcon(UUID, "second");
        for step in 0..=WRITE_STEPS {
            let mut f = fixture();
            let Ok(Created::New(old)) = f.store.create(&f.caller, &found, &old_body) else {
                panic!("stored")
            };
            f.fs.fail_from(f.fs.log().len() + step);
            let result = f.store.replace(&f.caller, old, &found, &new_body);
            assert_eq!(
                result.is_ok(),
                step == WRITE_STEPS,
                "step {step}: {result:?}"
            );
            restart(&mut f);
            let stored = f
                .store
                .get(&f.caller, found.uuid)
                .unwrap_or_else(|e| panic!("step {step}: {e:?}"));
            match result {
                Ok(new) => assert_eq!(
                    (stored.version, stored.body),
                    (new, new_body.clone()),
                    "step {step}"
                ),
                Err(_) => assert_eq!(
                    (stored.version, stored.body),
                    (old, old_body.clone()),
                    "step {step}"
                ),
            }
            assert_whole(&f, step);
        }
    }

    #[test]
    fn a_replace_cut_off_between_its_file_and_its_commit_keeps_the_old_version() {
        let mut f = fixture();
        let (found, old) = stored(&f, "first");
        let (_, newer) = vcon(UUID, "second");
        die_before_commit(&f);
        assert!(f.store.replace(&f.caller, old, &found, &newer).is_err());
        restart(&mut f);
        assert_eq!(
            f.store.get(&f.caller, found.uuid).expect("reads").version,
            old
        );
        assert_eq!(hidden_versions(&f), 0);
        assert_whole(&f, WRITE_STEPS);
    }

    #[test]
    fn a_delete_cut_off_before_its_commit_leaves_the_vcon_live() {
        let mut f = fixture();
        let (found, version) = stored(&f, "first");
        let other = block_commits(&f);
        other
            .execute_batch(
                "CREATE TRIGGER dead BEFORE INSERT ON audit
                 BEGIN SELECT RAISE(ABORT, 'the process has died'); END;",
            )
            .expect("trigger");
        assert!(
            f.store
                .delete(&f.caller, found.uuid, Retention::Years(7))
                .is_err()
        );
        drop(other);
        restart(&mut f);
        assert_eq!(
            f.store
                .get(&f.caller, found.uuid)
                .expect("still live")
                .version,
            version
        );
        assert_eq!(hidden_versions(&f), 0);
    }
}