mongreldb_core/

database.rs

//! Multi-table `Database` container (spec §5, §6, §10).
//!
//! Owns the shared services — catalog, dual-counter epoch authority, shared
//! raw/decoded page caches, snapshot-retention registry, and the DB-wide KEK —
//! and mounts per-table [`Table`] engines under `tables/<id>/` that borrow them.
//! P1 scope: per-table WALs remain (collapsed into one shared WAL in P2); the
//! win here is one consistent commit clock across tables and one reopen path.

use crate::catalog::{self, Catalog, CatalogEntry, TableState, META_DEK_LEN};
use crate::engine::{SharedCtx, Table};
use crate::epoch::{Epoch, EpochAuthority, Snapshot};
use crate::error::{MongrelError, Result};
use crate::external_table::ExternalTableEntry;
use crate::memtable::Value;
use crate::procedure::{
    ProcedureCallOutput, ProcedureCallResult, ProcedureCallRow, ProcedureCondition, ProcedureEntry,
    ProcedureStep, ProcedureValue, StoredProcedure,
};
use crate::retention::{OwnedSnapshotGuard, SnapshotGuard, SnapshotRegistry};
use crate::rowid::RowId;
use crate::schema::{AlterColumn, ColumnDef, Schema};
use crate::trigger::{
    StoredTrigger, TriggerCondition, TriggerConfig, TriggerEntry, TriggerEvent, TriggerExpr,
    TriggerRaiseAction, TriggerStep, TriggerTarget, TriggerTiming, TriggerValue,
};
use parking_lot::{Mutex, RwLock};
use std::collections::HashMap;
use std::io::Write;
use std::path::{Path, PathBuf};
use std::sync::atomic::{AtomicBool, AtomicU32};
use std::sync::Arc;

pub const TABLES_DIR: &str = "tables";
pub const VTAB_DIR: &str = "_vtab";
pub const META_DIR: &str = "_meta";
pub const KEYS_FILENAME: &str = "keys";

/// Sentinel `table_id` for `CheckIssue`s that concern the shared WAL rather
/// than any table. `u64::MAX` is never allocated to a real table (the catalog
/// mints ids from 0 upward), so [`Database::doctor`] can safely skip them.
pub const WAL_TABLE_ID: u64 = u64::MAX;
/// Sentinel `table_id` for `CheckIssue`s that concern external-table module
/// state instead of an ordinary table.
pub const EXTERNAL_TABLE_ID: u64 = u64::MAX - 1;

#[derive(Debug, Clone)]
pub enum ExternalTriggerWrite {
    Insert {
        table: String,
        cells: Vec<(u16, Value)>,
    },
    UpdateByPk {
        table: String,
        pk: Value,
        cells: Vec<(u16, Value)>,
    },
    DeleteByPk {
        table: String,
        pk: Value,
    },
}

impl ExternalTriggerWrite {
    fn table(&self) -> &str {
        match self {
            Self::Insert { table, .. }
            | Self::UpdateByPk { table, .. }
            | Self::DeleteByPk { table, .. } => table,
        }
    }
}

#[derive(Debug, Clone, PartialEq)]
pub enum ExternalTriggerBaseWrite {
    Put {
        table: String,
        cells: Vec<(u16, Value)>,
    },
    Delete {
        table: String,
        row_id: RowId,
    },
}

#[derive(Debug, Clone, PartialEq)]
pub struct ExternalTriggerWriteResult {
    pub state: Vec<u8>,
    pub base_writes: Vec<ExternalTriggerBaseWrite>,
}

impl ExternalTriggerWriteResult {
    pub fn new(state: Vec<u8>) -> Self {
        Self {
            state,
            base_writes: Vec::new(),
        }
    }
}

pub trait ExternalTriggerBridge {
    fn apply_trigger_external_write(
        &self,
        entry: &ExternalTableEntry,
        base_state: Vec<u8>,
        op: ExternalTriggerWrite,
    ) -> Result<ExternalTriggerWriteResult>;
}

/// A pending uniform-epoch run written during a large transaction (spec §8.5).
struct SpilledRun {
    table_id: u64,
    run_id: u128,
    pending_path: PathBuf,
    rows: Vec<crate::memtable::Row>,
    row_count: u64,
    min_rid: u64,
    max_rid: u64,
}

#[derive(Debug, Clone)]
struct TriggerRowImage {
    columns: HashMap<u16, Value>,
}

impl TriggerRowImage {
    fn from_row(row: crate::memtable::Row) -> Self {
        Self {
            columns: row.columns,
        }
    }

    fn from_cells(cells: &[(u16, Value)]) -> Self {
        Self {
            columns: cells.iter().cloned().collect(),
        }
    }
}

#[derive(Debug, Clone)]
struct WriteEvent {
    table: String,
    kind: TriggerEvent,
    old: Option<TriggerRowImage>,
    new: Option<TriggerRowImage>,
    changed_columns: Vec<u16>,
    op_indices: Vec<usize>,
    put_idx: Option<usize>,
    trigger_stack: Vec<String>,
}

#[derive(Default)]
struct TriggerExpansion {
    before: Vec<(u64, crate::txn::Staged)>,
    before_stacks: Vec<Vec<String>>,
    before_external: Vec<ExternalTriggerWrite>,
    after: Vec<(u64, crate::txn::Staged)>,
    after_stacks: Vec<Vec<String>>,
    after_external: Vec<ExternalTriggerWrite>,
    ignored_indices: std::collections::BTreeSet<usize>,
}

struct TriggerProgramOutput<'a> {
    added: &'a mut Vec<(u64, crate::txn::Staged)>,
    added_stacks: &'a mut Vec<Vec<String>>,
    added_external: &'a mut Vec<ExternalTriggerWrite>,
    ignored_indices: &'a mut std::collections::BTreeSet<usize>,
}

#[derive(Debug, Clone, Copy, PartialEq, Eq)]
enum TriggerProgramOutcome {
    Continue,
    Ignore,
}

/// An integrity issue found by [`Database::check`] (spec §16).
#[derive(Debug, Clone)]
pub struct CheckIssue {
    pub table_id: u64,
    pub table_name: String,
    pub severity: String,
    pub description: String,
}

/// A handle to a live table inside a [`Database`]. Writes take the inner lock
/// (P1); P3.3 replaces this with lock-free `ArcSwap` reads + a publish lock for
/// writes.
pub type TableHandle = Arc<Mutex<Table>>;

/// A multi-table database: one catalog, one epoch clock, shared caches, a
/// shared WAL, and a live map of name → `Arc<Table>`.
pub struct Database {
    root: PathBuf,
    catalog: RwLock<Catalog>,
    epoch: Arc<EpochAuthority>,
    snapshots: Arc<SnapshotRegistry>,
    page_cache: Arc<crate::cache::Sharded<crate::cache::PageCache>>,
    decoded_cache: Arc<crate::cache::Sharded<crate::cache::DecodedPageCache>>,
    commit_lock: Arc<Mutex<()>>,
    /// One shared WAL multiplexing every table's records (spec §7.2). Owned
    /// behind a `Mutex` so the transaction layer can append + group-sync. Shared
    /// (via `Arc`) with every mounted `Table` so single-table `put`/`commit`
    /// writes also land in this one WAL (B1 — one WAL per database).
    shared_wal: Arc<Mutex<crate::wal::SharedWal>>,
    /// Monotonic per-open transaction-id counter. Scoped by `open_generation`
    /// in P2.7; here it just needs to be unique within an open. Shared with
    /// mounted tables so their auto-commit txn ids never alias cross-table ones.
    next_txn_id: Arc<Mutex<u64>>,
    tables: RwLock<HashMap<u64, TableHandle>>,
    kek: Option<Arc<crate::encryption::Kek>>,
    /// Serializes DDL (create/drop table); data commits serialize through
    /// `commit_lock` shared via `SharedCtx`.
    ddl_lock: Mutex<()>,
    meta_dek: Option<[u8; META_DEK_LEN]>,
    /// P3.4: when staged bytes per table exceed this, write a uniform-epoch
    /// pending run to `_txn/<txn_id>/` instead of streaming Put records (§8.5).
    spill_threshold: std::sync::atomic::AtomicU64,
    /// P3.1: write-key → commit_epoch for first-committer-wins conflict
    /// detection (spec §9.2).
    conflicts: crate::txn::ConflictIndex,
    /// P3.1: min read_epoch of all in-flight txns, drives conflict-index
    /// pruning (spec §9.2, review fix #12).
    active_txns: crate::txn::ActiveTxns,
    /// P3.2: set on fsync error — all subsequent writes fail fast (spec §9.3e).
    /// Shared with mounted tables so a single-table commit also honors poison.
    poisoned: Arc<std::sync::atomic::AtomicBool>,
    /// P3.2: group-commit coordinator. The sequencer appends under the WAL lock
    /// but defers the fsync to one leader here, so concurrent commits share a
    /// single fsync (spec §9.3). Shared with mounted tables.
    group: Arc<crate::txn::GroupCommit>,
    /// P3.6: txn ids currently spilling into `_txn/<id>/`. GC never deletes a
    /// live spill's pending dir (review fix #14, spec §6.4).
    active_spills: Arc<crate::retention::ActiveSpills>,
    /// Test-only barrier invoked after a transaction writes its spill runs but
    /// before the sequencer/publish, so tests can race `gc()` against an
    /// in-flight spill. `None` in production.
    #[doc(hidden)]
    spill_hook: Mutex<Option<Box<dyn Fn() + Send + Sync>>>,
    trigger_recursive: AtomicBool,
    trigger_max_depth: AtomicU32,
}

impl Database {
    /// Create a fresh plaintext database at `root`.
    pub fn create(root: impl AsRef<Path>) -> Result<Self> {
        Self::create_inner(root, None)
    }

    /// Create a fresh encrypted database, deriving the DB-wide KEK from a
    /// passphrase (Argon2id + HKDF). The salt is persisted at `_meta/keys`.
    #[cfg(feature = "encryption")]
    pub fn create_encrypted(root: impl AsRef<Path>, passphrase: &str) -> Result<Self> {
        let root = root.as_ref();
        std::fs::create_dir_all(root)?;
        std::fs::create_dir_all(root.join(META_DIR))?;
        let salt = crate::encryption::random_salt();
        std::fs::write(root.join(META_DIR).join(KEYS_FILENAME), salt)?;
        let kek = Arc::new(crate::encryption::Kek::derive(passphrase, &salt)?);
        Self::create_inner(root, Some(kek))
    }

    /// Create a fresh encrypted database, deriving the DB-wide KEK from a raw
    /// high-entropy key via HKDF. The salt is persisted at `_meta/keys`.
    #[cfg(feature = "encryption")]
    pub fn create_with_key(root: impl AsRef<Path>, key: &[u8]) -> Result<Self> {
        let root = root.as_ref();
        std::fs::create_dir_all(root)?;
        std::fs::create_dir_all(root.join(META_DIR))?;
        let salt = crate::encryption::random_salt();
        std::fs::write(root.join(META_DIR).join(KEYS_FILENAME), salt)?;
        let kek = Arc::new(crate::encryption::Kek::from_raw_key(key, &salt)?);
        Self::create_inner(root, Some(kek))
    }

    fn create_inner(
        root: impl AsRef<Path>,
        kek: Option<Arc<crate::encryption::Kek>>,
    ) -> Result<Self> {
        let root = root.as_ref().to_path_buf();
        std::fs::create_dir_all(&root)?;
        std::fs::create_dir_all(root.join(TABLES_DIR))?;
        let meta_dek = crate::encryption::meta_dek_for(kek.as_deref());
        let cat = Catalog::empty();
        catalog::write_atomic(&root, &cat, meta_dek.as_ref())?;
        Self::finish_open(root, cat, kek, meta_dek, false)
    }

    /// Open an existing plaintext database.
    pub fn open(root: impl AsRef<Path>) -> Result<Self> {
        Self::open_inner(root, None, None)
    }

    /// Open an existing encrypted database with a passphrase.
    #[cfg(feature = "encryption")]
    pub fn open_encrypted(root: impl AsRef<Path>, passphrase: &str) -> Result<Self> {
        let root = root.as_ref();
        let salt_bytes = std::fs::read(root.join(META_DIR).join(KEYS_FILENAME))
            .map_err(|e| MongrelError::NotFound(format!("encryption salt file: {e}")))?;
        let mut salt = [0u8; crate::encryption::SALT_LEN];
        salt.copy_from_slice(&salt_bytes);
        let kek = Arc::new(crate::encryption::Kek::derive(passphrase, &salt)?);
        Self::open_inner(root, Some(kek), None)
    }

    /// Open an existing encrypted database using a raw high-entropy key.
    #[cfg(feature = "encryption")]
    pub fn open_with_key(root: impl AsRef<Path>, key: &[u8]) -> Result<Self> {
        let root = root.as_ref();
        let salt_path = root.join(META_DIR).join(KEYS_FILENAME);
        let salt_bytes = std::fs::read(&salt_path).map_err(|e| {
            MongrelError::NotFound(format!(
                "encryption salt file {:?}: {e} (database not encrypted, or corrupted)",
                salt_path
            ))
        })?;
        if salt_bytes.len() != crate::encryption::SALT_LEN {
            return Err(MongrelError::InvalidArgument(format!(
                "salt file is {} bytes, expected {}",
                salt_bytes.len(),
                crate::encryption::SALT_LEN
            )));
        }
        let mut salt = [0u8; crate::encryption::SALT_LEN];
        salt.copy_from_slice(&salt_bytes);
        let kek = Arc::new(crate::encryption::Kek::from_raw_key(key, &salt)?);
        Self::open_inner(root, Some(kek), None)
    }

    fn open_inner(
        root: impl AsRef<Path>,
        kek: Option<Arc<crate::encryption::Kek>>,
        _meta_dek_override: Option<[u8; META_DEK_LEN]>,
    ) -> Result<Self> {
        let root = root.as_ref().to_path_buf();
        let meta_dek = crate::encryption::meta_dek_for(kek.as_deref());
        let cat = catalog::read(&root, meta_dek.as_ref())?
            .ok_or_else(|| MongrelError::NotFound(format!("no catalog found at {:?}", root)))?;
        Self::finish_open(root, cat, kek, meta_dek, true)
    }

    fn finish_open(
        root: PathBuf,
        cat: Catalog,
        kek: Option<Arc<crate::encryption::Kek>>,
        meta_dek: Option<[u8; META_DEK_LEN]>,
        existing: bool,
    ) -> Result<Self> {
        let epoch = Arc::new(EpochAuthority::new(cat.db_epoch));
        let snapshots = Arc::new(SnapshotRegistry::new());
        let page_cache = Arc::new(crate::cache::Sharded::new(
            crate::cache::CACHE_SHARDS,
            || {
                crate::cache::PageCache::new(
                    crate::engine::PAGE_CACHE_CAPACITY / crate::cache::CACHE_SHARDS as u64,
                )
            },
        ));
        let decoded_cache = Arc::new(crate::cache::Sharded::new(
            crate::cache::CACHE_SHARDS,
            || {
                crate::cache::DecodedPageCache::new(
                    crate::engine::DECODED_CACHE_CAPACITY / crate::cache::CACHE_SHARDS as u64,
                )
            },
        ));
        let commit_lock = Arc::new(Mutex::new(()));
        let wal_dek = crate::encryption::wal_dek_for(kek.as_deref());
        let shared_wal = Arc::new(Mutex::new(if existing {
            crate::wal::SharedWal::open(&root, Epoch(cat.db_epoch), wal_dek.clone())?
        } else {
            crate::wal::SharedWal::create_with_dek(&root, Epoch(cat.db_epoch), wal_dek.clone())?
        }));
        // Shared write-path state handed to every mounted table so single-table
        // `put`/`commit` writes route through the one shared WAL, the one group-
        // commit coordinator, and the one poison flag (B1).
        let poisoned = Arc::new(std::sync::atomic::AtomicBool::new(false));
        let group = Arc::new(crate::txn::GroupCommit::new(
            shared_wal.lock().durable_seq(),
        ));
        // Final base value is set after the open-generation bump below; tables
        // only draw ids once the user issues a write (post-open), so the
        // placeholder is never observed.
        let txn_ids = Arc::new(Mutex::new(1u64));

        // Recover DDL from the shared WAL BEFORE opening tables (spec §15,
        // review fix #16). A crash between WAL fsync and the catalog
        // checkpoint leaves committed DDL durable in the WAL but absent from
        // the on-disk catalog; replay it here so the table-mounting loop and
        // data-record recovery see a correct catalog.
        let mut cat = cat;
        if existing {
            recover_ddl_from_wal(&root, &mut cat, meta_dek.as_ref(), wal_dek.as_ref())?;
        }

        // Open every live table against the shared context. Mounted tables have
        // no private WAL (B1) — `open_in` just loads the manifest/runs and
        // advances the shared epoch authority to its manifest epoch, so the
        // final shared watermark is the max across all tables. All of a mounted
        // table's committed records are replayed below from the shared WAL.
        let mut tables: HashMap<u64, TableHandle> = HashMap::new();
        for entry in &cat.tables {
            if !matches!(entry.state, TableState::Live) {
                continue;
            }
            let tdir = root.join(TABLES_DIR).join(entry.table_id.to_string());
            let ctx = SharedCtx {
                epoch: Arc::clone(&epoch),
                page_cache: Arc::clone(&page_cache),
                decoded_cache: Arc::clone(&decoded_cache),
                snapshots: Arc::clone(&snapshots),
                kek: kek.clone(),
                commit_lock: Arc::clone(&commit_lock),
                shared: Some(crate::engine::SharedWalCtx {
                    wal: Arc::clone(&shared_wal),
                    group: Arc::clone(&group),
                    poisoned: Arc::clone(&poisoned),
                    txn_ids: Arc::clone(&txn_ids),
                }),
            };
            let t = Table::open_in(&tdir, ctx)?;
            tables.insert(entry.table_id, Arc::new(Mutex::new(t)));
        }

        // Recover transaction writes from the shared WAL (spec §15). This is the
        // single durability source for mounted tables: it applies every committed
        // record — both single-table `Table::commit` writes and cross-table
        // transactions — gated by each table's `flushed_epoch` (records already
        // durable in a run are not re-applied).
        if existing {
            recover_shared_wal(&root, &tables, &epoch, wal_dek.as_ref())?;
            // P3.4: sweep stale `_txn/<txn_id>/` dirs left by aborted/crashed
            // large transactions (spec §8.5, review fix #14).
            sweep_pending_txn_dirs(&root, &cat);
        }

        // Bump `open_generation` on every open and scope transaction ids by it
        // (`txn_id = (generation << 32) | counter`), so ids never alias across
        // reopens (review fix #11). Persist the bumped generation to the catalog.
        if existing {
            cat.open_generation = cat.open_generation.wrapping_add(1);
            catalog::write_atomic(&root, &cat, meta_dek.as_ref())?;
        }
        let next_txn_id = (cat.open_generation << 32) | 1;
        // Seed the shared txn-id allocator now that the generation is final.
        *txn_ids.lock() = next_txn_id;

        Ok(Self {
            root,
            catalog: RwLock::new(cat),
            epoch,
            snapshots,
            page_cache,
            decoded_cache,
            commit_lock,
            shared_wal,
            next_txn_id: txn_ids,
            tables: RwLock::new(tables),
            kek,
            ddl_lock: Mutex::new(()),
            meta_dek,
            conflicts: crate::txn::ConflictIndex::new(),
            active_txns: crate::txn::ActiveTxns::new(),
            poisoned,
            group,
            spill_threshold: std::sync::atomic::AtomicU64::new(64 * 1024 * 1024),
            active_spills: Arc::new(crate::retention::ActiveSpills::new()),
            spill_hook: Mutex::new(None),
            trigger_recursive: AtomicBool::new(TriggerConfig::default().recursive_triggers),
            trigger_max_depth: AtomicU32::new(TriggerConfig::default().max_depth),
        })
    }

    /// The current reader-visible epoch.
    pub fn visible_epoch(&self) -> Epoch {
        self.epoch.visible()
    }

    /// Clone the in-memory catalog (for diagnostics / tests).
    pub fn catalog_snapshot(&self) -> Catalog {
        self.catalog.read().clone()
    }

    /// The filesystem root this database was opened/created at.
    pub fn root(&self) -> &Path {
        &self.root
    }

    /// Resolve a table name → id (live tables only). pub(crate) so the
    /// transaction layer can stage by name.
    pub fn table_id(&self, name: &str) -> Result<u64> {
        let cat = self.catalog.read();
        cat.live(name)
            .map(|e| e.table_id)
            .ok_or_else(|| MongrelError::NotFound(format!("table {name:?} not found")))
    }

    pub fn procedures(&self) -> Vec<StoredProcedure> {
        self.catalog
            .read()
            .procedures
            .iter()
            .map(|p| p.procedure.clone())
            .collect()
    }

    pub fn procedure(&self, name: &str) -> Option<StoredProcedure> {
        self.catalog
            .read()
            .procedures
            .iter()
            .find(|p| p.procedure.name == name)
            .map(|p| p.procedure.clone())
    }

    pub fn create_procedure(&self, mut procedure: StoredProcedure) -> Result<StoredProcedure> {
        let _g = self.ddl_lock.lock();
        procedure.validate()?;
        self.validate_procedure_references(&procedure)?;
        {
            let cat = self.catalog.read();
            if cat
                .procedures
                .iter()
                .any(|p| p.procedure.name == procedure.name)
            {
                return Err(MongrelError::InvalidArgument(format!(
                    "procedure {:?} already exists",
                    procedure.name
                )));
            }
        }
        let commit_lock = Arc::clone(&self.commit_lock);
        let _c = commit_lock.lock();
        let epoch = self.epoch.bump_assigned();
        procedure.created_epoch = epoch.0;
        procedure.updated_epoch = epoch.0;
        {
            let mut cat = self.catalog.write();
            cat.procedures.push(ProcedureEntry::from(procedure.clone()));
            cat.db_epoch = epoch.0;
        }
        catalog::write_atomic(&self.root, &self.catalog.read(), self.meta_dek.as_ref())?;
        self.advance_visible(epoch);
        Ok(procedure)
    }

    pub fn create_or_replace_procedure(
        &self,
        procedure: StoredProcedure,
    ) -> Result<StoredProcedure> {
        let _g = self.ddl_lock.lock();
        procedure.validate()?;
        self.validate_procedure_references(&procedure)?;
        let commit_lock = Arc::clone(&self.commit_lock);
        let _c = commit_lock.lock();
        let epoch = self.epoch.bump_assigned();
        let replaced = {
            let mut cat = self.catalog.write();
            let next = match cat
                .procedures
                .iter()
                .position(|p| p.procedure.name == procedure.name)
            {
                Some(idx) => {
                    let next = cat.procedures[idx]
                        .procedure
                        .replaced(procedure.clone(), epoch.0)?;
                    cat.procedures[idx] = ProcedureEntry::from(next.clone());
                    next
                }
                None => {
                    let mut next = procedure;
                    next.created_epoch = epoch.0;
                    next.updated_epoch = epoch.0;
                    cat.procedures.push(ProcedureEntry::from(next.clone()));
                    next
                }
            };
            cat.db_epoch = epoch.0;
            next
        };
        catalog::write_atomic(&self.root, &self.catalog.read(), self.meta_dek.as_ref())?;
        self.advance_visible(epoch);
        Ok(replaced)
    }

    pub fn drop_procedure(&self, name: &str) -> Result<()> {
        let _g = self.ddl_lock.lock();
        let commit_lock = Arc::clone(&self.commit_lock);
        let _c = commit_lock.lock();
        let epoch = self.epoch.bump_assigned();
        {
            let mut cat = self.catalog.write();
            let before = cat.procedures.len();
            cat.procedures.retain(|p| p.procedure.name != name);
            if cat.procedures.len() == before {
                return Err(MongrelError::NotFound(format!(
                    "procedure {name:?} not found"
                )));
            }
            cat.db_epoch = epoch.0;
        }
        catalog::write_atomic(&self.root, &self.catalog.read(), self.meta_dek.as_ref())?;
        self.advance_visible(epoch);
        Ok(())
    }

    pub fn triggers(&self) -> Vec<StoredTrigger> {
        self.catalog
            .read()
            .triggers
            .iter()
            .map(|t| t.trigger.clone())
            .collect()
    }

    pub fn trigger(&self, name: &str) -> Option<StoredTrigger> {
        self.catalog
            .read()
            .triggers
            .iter()
            .find(|t| t.trigger.name == name)
            .map(|t| t.trigger.clone())
    }

    pub fn create_trigger(&self, mut trigger: StoredTrigger) -> Result<StoredTrigger> {
        let _g = self.ddl_lock.lock();
        trigger.validate()?;
        self.validate_trigger_references(&trigger)?;
        {
            let cat = self.catalog.read();
            if cat.triggers.iter().any(|t| t.trigger.name == trigger.name) {
                return Err(MongrelError::InvalidArgument(format!(
                    "trigger {:?} already exists",
                    trigger.name
                )));
            }
        }
        let commit_lock = Arc::clone(&self.commit_lock);
        let _c = commit_lock.lock();
        let epoch = self.epoch.bump_assigned();
        trigger.created_epoch = epoch.0;
        trigger.updated_epoch = epoch.0;
        {
            let mut cat = self.catalog.write();
            cat.triggers.push(TriggerEntry::from(trigger.clone()));
            cat.db_epoch = epoch.0;
        }
        catalog::write_atomic(&self.root, &self.catalog.read(), self.meta_dek.as_ref())?;
        self.advance_visible(epoch);
        Ok(trigger)
    }

    pub fn create_or_replace_trigger(&self, trigger: StoredTrigger) -> Result<StoredTrigger> {
        let _g = self.ddl_lock.lock();
        trigger.validate()?;
        self.validate_trigger_references(&trigger)?;
        let commit_lock = Arc::clone(&self.commit_lock);
        let _c = commit_lock.lock();
        let epoch = self.epoch.bump_assigned();
        let replaced = {
            let mut cat = self.catalog.write();
            let next = match cat
                .triggers
                .iter()
                .position(|t| t.trigger.name == trigger.name)
            {
                Some(idx) => {
                    let next = cat.triggers[idx]
                        .trigger
                        .replaced(trigger.clone(), epoch.0)?;
                    cat.triggers[idx] = TriggerEntry::from(next.clone());
                    next
                }
                None => {
                    let mut next = trigger;
                    next.created_epoch = epoch.0;
                    next.updated_epoch = epoch.0;
                    cat.triggers.push(TriggerEntry::from(next.clone()));
                    next
                }
            };
            cat.db_epoch = epoch.0;
            next
        };
        catalog::write_atomic(&self.root, &self.catalog.read(), self.meta_dek.as_ref())?;
        self.advance_visible(epoch);
        Ok(replaced)
    }

    pub fn drop_trigger(&self, name: &str) -> Result<()> {
        let _g = self.ddl_lock.lock();
        let commit_lock = Arc::clone(&self.commit_lock);
        let _c = commit_lock.lock();
        let epoch = self.epoch.bump_assigned();
        {
            let mut cat = self.catalog.write();
            let before = cat.triggers.len();
            cat.triggers.retain(|t| t.trigger.name != name);
            if cat.triggers.len() == before {
                return Err(MongrelError::NotFound(format!(
                    "trigger {name:?} not found"
                )));
            }
            cat.db_epoch = epoch.0;
        }
        catalog::write_atomic(&self.root, &self.catalog.read(), self.meta_dek.as_ref())?;
        self.advance_visible(epoch);
        Ok(())
    }

    pub fn external_tables(&self) -> Vec<ExternalTableEntry> {
        self.catalog.read().external_tables.clone()
    }

    pub fn external_table(&self, name: &str) -> Option<ExternalTableEntry> {
        self.catalog
            .read()
            .external_tables
            .iter()
            .find(|t| t.name == name)
            .cloned()
    }

    pub fn create_external_table(
        &self,
        mut entry: ExternalTableEntry,
    ) -> Result<ExternalTableEntry> {
        let _g = self.ddl_lock.lock();
        entry.validate()?;
        {
            let cat = self.catalog.read();
            if cat.live(&entry.name).is_some()
                || cat.external_tables.iter().any(|t| t.name == entry.name)
            {
                return Err(MongrelError::InvalidArgument(format!(
                    "table {:?} already exists",
                    entry.name
                )));
            }
        }
        let commit_lock = Arc::clone(&self.commit_lock);
        let _c = commit_lock.lock();
        let epoch = self.epoch.bump_assigned();
        entry.created_epoch = epoch.0;
        {
            let mut cat = self.catalog.write();
            cat.external_tables.push(entry.clone());
            cat.db_epoch = epoch.0;
        }
        catalog::write_atomic(&self.root, &self.catalog.read(), self.meta_dek.as_ref())?;
        self.advance_visible(epoch);
        Ok(entry)
    }

    pub fn drop_external_table(&self, name: &str) -> Result<()> {
        let _g = self.ddl_lock.lock();
        let commit_lock = Arc::clone(&self.commit_lock);
        let _c = commit_lock.lock();
        let epoch = self.epoch.bump_assigned();
        {
            let mut cat = self.catalog.write();
            let before = cat.external_tables.len();
            cat.external_tables.retain(|t| t.name != name);
            if cat.external_tables.len() == before {
                return Err(MongrelError::NotFound(format!(
                    "external table {name:?} not found"
                )));
            }
            cat.db_epoch = epoch.0;
        }
        let state_dir = self.root.join(VTAB_DIR).join(name);
        if state_dir.exists() {
            std::fs::remove_dir_all(state_dir)?;
        }
        catalog::write_atomic(&self.root, &self.catalog.read(), self.meta_dek.as_ref())?;
        self.advance_visible(epoch);
        Ok(())
    }

    pub fn commit_external_table_state(&self, name: &str, state: &[u8]) -> Result<Epoch> {
        let txn_id = self.alloc_txn_id();
        self.commit_transaction_with_external_states(
            txn_id,
            self.epoch.visible(),
            Vec::new(),
            vec![(name.to_string(), state.to_vec())],
            None,
        )
    }

    pub fn trigger_config(&self) -> TriggerConfig {
        use std::sync::atomic::Ordering;
        TriggerConfig {
            recursive_triggers: self.trigger_recursive.load(Ordering::Relaxed),
            max_depth: self.trigger_max_depth.load(Ordering::Relaxed),
        }
    }

    pub fn set_trigger_config(&self, config: TriggerConfig) -> Result<()> {
        use std::sync::atomic::Ordering;
        if config.max_depth == 0 {
            return Err(MongrelError::InvalidArgument(
                "trigger max_depth must be greater than 0".into(),
            ));
        }
        self.trigger_recursive
            .store(config.recursive_triggers, Ordering::Relaxed);
        self.trigger_max_depth
            .store(config.max_depth, Ordering::Relaxed);
        Ok(())
    }

    pub fn set_recursive_triggers(&self, recursive: bool) {
        use std::sync::atomic::Ordering;
        self.trigger_recursive.store(recursive, Ordering::Relaxed);
    }

    pub fn call_procedure(
        &self,
        name: &str,
        args: HashMap<String, crate::Value>,
    ) -> Result<ProcedureCallResult> {
        let procedure = self
            .procedure(name)
            .ok_or_else(|| MongrelError::NotFound(format!("procedure {name:?} not found")))?;
        let args = bind_procedure_args(&procedure, args)?;
        let has_writes = procedure.body.steps.iter().any(ProcedureStep::is_write);
        let mut outputs: HashMap<String, ProcedureCallOutput> = HashMap::new();
        if has_writes {
            let mut tx = self.begin();
            let run = (|| {
                for step in &procedure.body.steps {
                    let output =
                        self.execute_procedure_step(step, &args, &outputs, Some(&mut tx))?;
                    outputs.insert(step.id().to_string(), output);
                }
                eval_return_output(&procedure.body.return_value, &args, &outputs)
            })();
            match run {
                Ok(output) => {
                    let epoch = tx.commit()?.0;
                    Ok(ProcedureCallResult {
                        epoch: Some(epoch),
                        output,
                    })
                }
                Err(e) => {
                    tx.rollback();
                    Err(e)
                }
            }
        } else {
            for step in &procedure.body.steps {
                let output = self.execute_procedure_step(step, &args, &outputs, None)?;
                outputs.insert(step.id().to_string(), output);
            }
            Ok(ProcedureCallResult {
                epoch: None,
                output: eval_return_output(&procedure.body.return_value, &args, &outputs)?,
            })
        }
    }

    fn execute_procedure_step(
        &self,
        step: &ProcedureStep,
        args: &HashMap<String, crate::Value>,
        outputs: &HashMap<String, ProcedureCallOutput>,
        tx: Option<&mut crate::txn::Transaction<'_>>,
    ) -> Result<ProcedureCallOutput> {
        match step {
            ProcedureStep::NativeQuery {
                table,
                conditions,
                projection,
                limit,
                ..
            } => {
                let mut q = crate::Query::new();
                for condition in conditions {
                    q = q.and(eval_condition(condition, args, outputs)?);
                }
                let handle = self.table(table)?;
                let mut rows = handle.lock().query(&q)?;
                if let Some(limit) = limit {
                    rows.truncate(*limit);
                }
                let projection = projection.as_ref();
                Ok(ProcedureCallOutput::Rows(
                    rows.into_iter()
                        .map(|row| ProcedureCallRow {
                            row_id: Some(row.row_id),
                            columns: match projection {
                                Some(ids) => row
                                    .columns
                                    .into_iter()
                                    .filter(|(id, _)| ids.contains(id))
                                    .collect(),
                                None => row.columns,
                            },
                        })
                        .collect(),
                ))
            }
            ProcedureStep::Put {
                table,
                cells,
                returning,
                ..
            } => {
                let tx = tx.ok_or_else(|| {
                    MongrelError::InvalidArgument(
                        "write procedure step requires a transaction".into(),
                    )
                })?;
                let cells = eval_cells(cells, args, outputs)?;
                if *returning {
                    let out = tx.put_returning(table, cells)?;
                    Ok(ProcedureCallOutput::Row(ProcedureCallRow {
                        row_id: None,
                        columns: out.row.columns.into_iter().collect(),
                    }))
                } else {
                    tx.put(table, cells)?;
                    Ok(ProcedureCallOutput::Null)
                }
            }
            ProcedureStep::Upsert {
                table,
                cells,
                update_cells,
                returning,
                ..
            } => {
                let tx = tx.ok_or_else(|| {
                    MongrelError::InvalidArgument(
                        "write procedure step requires a transaction".into(),
                    )
                })?;
                let cells = eval_cells(cells, args, outputs)?;
                let action = match update_cells {
                    Some(update_cells) => {
                        crate::UpsertAction::DoUpdate(eval_cells(update_cells, args, outputs)?)
                    }
                    None => crate::UpsertAction::DoNothing,
                };
                let out = tx.upsert(table, cells, action)?;
                if *returning {
                    Ok(ProcedureCallOutput::Row(ProcedureCallRow {
                        row_id: None,
                        columns: out.row.columns.into_iter().collect(),
                    }))
                } else {
                    Ok(ProcedureCallOutput::Null)
                }
            }
            ProcedureStep::DeleteByPk { table, pk, .. } => {
                let tx = tx.ok_or_else(|| {
                    MongrelError::InvalidArgument(
                        "write procedure step requires a transaction".into(),
                    )
                })?;
                let pk = eval_value(pk, args, outputs)?;
                let handle = self.table(table)?;
                let row_id = handle.lock().lookup_pk(&pk.encode_key()).ok_or_else(|| {
                    MongrelError::NotFound("procedure delete_by_pk target not found".into())
                })?;
                tx.delete(table, row_id)?;
                Ok(ProcedureCallOutput::Scalar(crate::Value::Bool(true)))
            }
            ProcedureStep::DeleteRows { .. } => Err(MongrelError::InvalidArgument(
                "DeleteRows procedure step is not supported by the core executor yet".into(),
            )),
            ProcedureStep::SqlQuery { .. } => Err(MongrelError::InvalidArgument(
                "SqlQuery procedure step must be executed by mongreldb-query".into(),
            )),
        }
    }

    fn validate_procedure_references(&self, procedure: &StoredProcedure) -> Result<()> {
        let cat = self.catalog.read();
        for step in &procedure.body.steps {
            let Some(table_name) = step.table() else {
                continue;
            };
            let schema = &cat
                .live(table_name)
                .ok_or_else(|| {
                    MongrelError::InvalidArgument(format!(
                        "procedure {:?} references unknown table {table_name:?}",
                        procedure.name
                    ))
                })?
                .schema;
            match step {
                ProcedureStep::NativeQuery {
                    conditions,
                    projection,
                    ..
                } => {
                    for condition in conditions {
                        validate_condition_columns(condition, schema)?;
                    }
                    if let Some(projection) = projection {
                        for id in projection {
                            validate_column_id(*id, schema)?;
                        }
                    }
                }
                ProcedureStep::Put { cells, .. } => {
                    for cell in cells {
                        validate_column_id(cell.column_id, schema)?;
                    }
                }
                ProcedureStep::Upsert {
                    cells,
                    update_cells,
                    ..
                } => {
                    for cell in cells {
                        validate_column_id(cell.column_id, schema)?;
                    }
                    if let Some(update_cells) = update_cells {
                        for cell in update_cells {
                            validate_column_id(cell.column_id, schema)?;
                        }
                    }
                }
                ProcedureStep::DeleteByPk { .. } => {
                    if schema.primary_key().is_none() {
                        return Err(MongrelError::InvalidArgument(format!(
                            "procedure {:?} references DeleteByPk on table {table_name:?} without a primary key",
                            procedure.name
                        )));
                    }
                }
                ProcedureStep::DeleteRows { .. } | ProcedureStep::SqlQuery { .. } => {}
            }
        }
        Ok(())
    }

    fn validate_trigger_references(&self, trigger: &StoredTrigger) -> Result<()> {
        let cat = self.catalog.read();
        let target_schema = match &trigger.target {
            TriggerTarget::Table(target_name) => cat
                .live(target_name)
                .ok_or_else(|| {
                    MongrelError::InvalidArgument(format!(
                        "trigger {:?} references unknown target table {target_name:?}",
                        trigger.name
                    ))
                })?
                .schema
                .clone(),
            TriggerTarget::View(_) => Schema {
                columns: trigger.target_columns.clone(),
                ..Schema::default()
            },
        };
        for col in &trigger.update_of {
            if target_schema.column(col).is_none() {
                return Err(MongrelError::InvalidArgument(format!(
                    "trigger {:?} UPDATE OF references unknown column {col:?}",
                    trigger.name
                )));
            }
        }
        if let Some(expr) = &trigger.when {
            validate_trigger_expr(expr, &target_schema, trigger.event)?;
        }
        for step in &trigger.program.steps {
            if matches!(step, TriggerStep::SetNew { .. }) && trigger.timing != TriggerTiming::Before
            {
                return Err(MongrelError::InvalidArgument(
                    "SetNew trigger steps are only valid in BEFORE triggers".into(),
                ));
            }
            validate_trigger_step(step, &cat, &target_schema, trigger.event)?;
        }
        Ok(())
    }

    /// Begin a new transaction reading at the current visible epoch.
    pub fn begin(&self) -> crate::txn::Transaction<'_> {
        let txn_id = self.alloc_txn_id();
        let read = Snapshot::at(self.epoch.visible());
        crate::txn::Transaction::new(self, txn_id, read)
    }

    /// Begin a transaction whose trigger programs may route external-table DML
    /// through an application/query-layer module bridge.
    pub fn begin_with_external_trigger_bridge<'a>(
        &'a self,
        bridge: &'a dyn ExternalTriggerBridge,
    ) -> crate::txn::Transaction<'a> {
        let txn_id = self.alloc_txn_id();
        let read = Snapshot::at(self.epoch.visible());
        crate::txn::Transaction::new(self, txn_id, read).with_external_trigger_bridge(bridge)
    }

    /// Run `f` in a transaction; commit on `Ok`, rollback on `Err`.
    pub fn transaction<T>(
        &self,
        f: impl FnOnce(&mut crate::txn::Transaction) -> Result<T>,
    ) -> Result<T> {
        let mut tx = self.begin();
        match f(&mut tx) {
            Ok(out) => {
                tx.commit()?;
                Ok(out)
            }
            Err(e) => {
                tx.rollback();
                Err(e)
            }
        }
    }

    /// Run `f` in a transaction with an external-trigger bridge; commit on
    /// `Ok`, rollback on `Err`.
    pub fn transaction_with_external_trigger_bridge<'a, T>(
        &'a self,
        bridge: &'a dyn ExternalTriggerBridge,
        f: impl FnOnce(&mut crate::txn::Transaction) -> Result<T>,
    ) -> Result<T> {
        let mut tx = self.begin_with_external_trigger_bridge(bridge);
        match f(&mut tx) {
            Ok(out) => {
                tx.commit()?;
                Ok(out)
            }
            Err(e) => {
                tx.rollback();
                Err(e)
            }
        }
    }

    /// Register a txn in `ActiveTxns` (spec §9.2, review fix #12). Called from
    /// `Transaction::new` so registration happens **before** any read.
    pub(crate) fn register_active(&self, epoch: Epoch) -> crate::txn::ActiveTxnGuard<'_> {
        self.active_txns.register(epoch)
    }

    fn fill_auto_increment_for_staging(
        &self,
        staging: &mut [(u64, crate::txn::Staged)],
    ) -> Result<()> {
        let tables = self.tables.read();
        for (table_id, staged) in staging {
            if let crate::txn::Staged::Put(cells) = staged {
                if let Some(handle) = tables.get(table_id) {
                    let mut t = handle.lock();
                    t.fill_auto_inc(cells)?;
                }
            }
        }
        Ok(())
    }

    fn expand_table_triggers(
        &self,
        staging: &mut Vec<(u64, crate::txn::Staged)>,
        read_epoch: Epoch,
        external_trigger_bridge: Option<&dyn ExternalTriggerBridge>,
        external_states: &mut Vec<(String, Vec<u8>)>,
    ) -> Result<()> {
        let mut external_writes = Vec::new();
        let config = self.trigger_config();
        if config.recursive_triggers {
            let chunk = std::mem::take(staging);
            let stacks = vec![Vec::new(); chunk.len()];
            *staging = self.expand_trigger_chunk(
                chunk,
                stacks,
                read_epoch,
                0,
                config.max_depth,
                &mut external_writes,
            )?;
            self.apply_external_trigger_writes(
                external_writes,
                external_trigger_bridge,
                external_states,
                staging,
            )?;
            return Ok(());
        }

        let mut expansion = self.expand_table_triggers_once(staging, read_epoch, None)?;
        if !expansion.before.is_empty() {
            let mut final_staging = expansion.before;
            final_staging.extend(filter_ignored_staging(
                std::mem::take(staging),
                &expansion.ignored_indices,
            ));
            *staging = final_staging;
        } else if !expansion.ignored_indices.is_empty() {
            *staging = filter_ignored_staging(std::mem::take(staging), &expansion.ignored_indices);
        }
        staging.append(&mut expansion.after);
        external_writes.append(&mut expansion.before_external);
        external_writes.append(&mut expansion.after_external);
        self.apply_external_trigger_writes(
            external_writes,
            external_trigger_bridge,
            external_states,
            staging,
        )?;
        Ok(())
    }

    fn expand_trigger_chunk(
        &self,
        mut chunk: Vec<(u64, crate::txn::Staged)>,
        stacks: Vec<Vec<String>>,
        read_epoch: Epoch,
        depth: u32,
        max_depth: u32,
        external_writes: &mut Vec<ExternalTriggerWrite>,
    ) -> Result<Vec<(u64, crate::txn::Staged)>> {
        if chunk.is_empty() {
            return Ok(Vec::new());
        }
        self.fill_auto_increment_for_staging(&mut chunk)?;
        let expansion = self.expand_table_triggers_once(&mut chunk, read_epoch, Some(&stacks))?;
        if depth >= max_depth && (!expansion.before.is_empty() || !expansion.after.is_empty()) {
            let stack = expansion
                .before_stacks
                .first()
                .or_else(|| expansion.after_stacks.first())
                .cloned()
                .unwrap_or_default();
            return Err(MongrelError::Conflict(format!(
                "trigger recursion exceeded max depth {max_depth}; trigger stack: {}",
                Self::format_trigger_stack(&stack)
            )));
        }

        let mut out = Vec::new();
        external_writes.extend(expansion.before_external);
        out.extend(self.expand_trigger_chunk(
            expansion.before,
            expansion.before_stacks,
            read_epoch,
            depth + 1,
            max_depth,
            external_writes,
        )?);
        out.extend(filter_ignored_staging(chunk, &expansion.ignored_indices));
        external_writes.extend(expansion.after_external);
        out.extend(self.expand_trigger_chunk(
            expansion.after,
            expansion.after_stacks,
            read_epoch,
            depth + 1,
            max_depth,
            external_writes,
        )?);
        Ok(out)
    }

    fn apply_external_trigger_writes(
        &self,
        writes: Vec<ExternalTriggerWrite>,
        bridge: Option<&dyn ExternalTriggerBridge>,
        external_states: &mut Vec<(String, Vec<u8>)>,
        staging: &mut Vec<(u64, crate::txn::Staged)>,
    ) -> Result<()> {
        if writes.is_empty() {
            return Ok(());
        }
        let bridge = bridge.ok_or_else(|| {
            MongrelError::InvalidArgument(
                "trigger program wrote an external table, but this transaction has no external trigger bridge".into(),
            )
        })?;
        for write in writes {
            let table = write.table().to_string();
            let entry = self.external_table(&table).ok_or_else(|| {
                MongrelError::NotFound(format!("external table {table:?} not found"))
            })?;
            let base_state = current_external_state_bytes(&self.root, external_states, &table)?;
            let result = bridge.apply_trigger_external_write(&entry, base_state, write)?;
            external_states.push((table, result.state));
            for base_write in result.base_writes {
                match base_write {
                    ExternalTriggerBaseWrite::Put { table, cells } => {
                        let table_id = self.table_id(&table)?;
                        staging.push((table_id, crate::txn::Staged::Put(cells)));
                    }
                    ExternalTriggerBaseWrite::Delete { table, row_id } => {
                        let table_id = self.table_id(&table)?;
                        staging.push((table_id, crate::txn::Staged::Delete(row_id)));
                    }
                }
            }
        }
        dedup_external_states_in_place(external_states);
        Ok(())
    }

    fn expand_table_triggers_once(
        &self,
        staging: &mut Vec<(u64, crate::txn::Staged)>,
        read_epoch: Epoch,
        trigger_stacks: Option<&[Vec<String>]>,
    ) -> Result<TriggerExpansion> {
        let triggers: Vec<StoredTrigger> = self
            .catalog
            .read()
            .triggers
            .iter()
            .filter(|entry| {
                entry.trigger.enabled
                    && matches!(
                        entry.trigger.timing,
                        TriggerTiming::Before | TriggerTiming::After
                    )
                    && matches!(entry.trigger.target, TriggerTarget::Table(_))
            })
            .map(|entry| entry.trigger.clone())
            .collect();
        if triggers.is_empty() || staging.is_empty() {
            return Ok(TriggerExpansion::default());
        }

        let before_triggers = triggers
            .iter()
            .filter(|trigger| trigger.timing == TriggerTiming::Before)
            .cloned()
            .collect::<Vec<_>>();
        let after_triggers = triggers
            .iter()
            .filter(|trigger| trigger.timing == TriggerTiming::After)
            .cloned()
            .collect::<Vec<_>>();

        let mut before_added = Vec::new();
        let mut before_stacks = Vec::new();
        let mut before_external = Vec::new();
        let mut ignored_indices = std::collections::BTreeSet::new();
        if !before_triggers.is_empty() {
            let before_events =
                self.trigger_events_for_staging(staging, read_epoch, trigger_stacks)?;
            let mut out = TriggerProgramOutput {
                added: &mut before_added,
                added_stacks: &mut before_stacks,
                added_external: &mut before_external,
                ignored_indices: &mut ignored_indices,
            };
            self.execute_triggers_for_events(
                &before_triggers,
                &before_events,
                Some(staging),
                &mut out,
            )?;
        }

        let after_events = if after_triggers.is_empty() {
            Vec::new()
        } else {
            self.trigger_events_for_staging(staging, read_epoch, trigger_stacks)?
                .into_iter()
                .filter(|event| {
                    !event
                        .op_indices
                        .iter()
                        .any(|idx| ignored_indices.contains(idx))
                })
                .collect()
        };

        let mut after_added = Vec::new();
        let mut after_stacks = Vec::new();
        let mut after_external = Vec::new();
        let mut out = TriggerProgramOutput {
            added: &mut after_added,
            added_stacks: &mut after_stacks,
            added_external: &mut after_external,
            ignored_indices: &mut ignored_indices,
        };
        self.execute_triggers_for_events(&after_triggers, &after_events, None, &mut out)?;
        Ok(TriggerExpansion {
            before: before_added,
            before_stacks,
            before_external,
            after: after_added,
            after_stacks,
            after_external,
            ignored_indices,
        })
    }

    fn execute_triggers_for_events(
        &self,
        triggers: &[StoredTrigger],
        events: &[WriteEvent],
        mut staging: Option<&mut Vec<(u64, crate::txn::Staged)>>,
        out: &mut TriggerProgramOutput<'_>,
    ) -> Result<()> {
        for event in events {
            for trigger in triggers {
                if event
                    .op_indices
                    .iter()
                    .any(|idx| out.ignored_indices.contains(idx))
                {
                    break;
                }
                let matches = {
                    let cat = self.catalog.read();
                    trigger_matches_event(trigger, event, &cat)?
                };
                if !matches {
                    continue;
                }
                if let Some(when) = &trigger.when {
                    if !eval_trigger_expr(when, event)? {
                        continue;
                    }
                }
                let trigger_stack = Self::trigger_stack_with(&event.trigger_stack, &trigger.name);
                if event.trigger_stack.iter().any(|name| name == &trigger.name) {
                    return Err(MongrelError::Conflict(format!(
                        "trigger recursion cycle detected; trigger stack: {}",
                        Self::format_trigger_stack(&trigger_stack)
                    )));
                }
                let outcome = match staging.as_mut() {
                    Some(staging) => self.execute_trigger_program(
                        trigger,
                        event,
                        Some(&mut **staging),
                        out,
                        &trigger_stack,
                    )?,
                    None => {
                        self.execute_trigger_program(trigger, event, None, out, &trigger_stack)?
                    }
                };
                if outcome == TriggerProgramOutcome::Ignore {
                    out.ignored_indices.extend(event.op_indices.iter().copied());
                    break;
                }
            }
        }
        Ok(())
    }

    fn trigger_events_for_staging(
        &self,
        staging: &[(u64, crate::txn::Staged)],
        read_epoch: Epoch,
        trigger_stacks: Option<&[Vec<String>]>,
    ) -> Result<Vec<WriteEvent>> {
        use crate::txn::Staged;
        use std::collections::{HashMap, VecDeque};

        let snapshot = Snapshot::at(read_epoch);
        let cat = self.catalog.read();
        let mut table_names = HashMap::new();
        let mut table_schemas = HashMap::new();
        for entry in cat
            .tables
            .iter()
            .filter(|entry| matches!(entry.state, TableState::Live))
        {
            table_names.insert(entry.table_id, entry.name.clone());
            table_schemas.insert(entry.table_id, entry.schema.clone());
        }
        drop(cat);

        let mut old_rows: HashMap<usize, TriggerRowImage> = HashMap::new();
        let mut delete_by_key: HashMap<(u64, Vec<u8>), VecDeque<usize>> = HashMap::new();
        let mut put_by_key: HashMap<(u64, Vec<u8>), VecDeque<usize>> = HashMap::new();

        for (idx, (table_id, staged)) in staging.iter().enumerate() {
            let Some(schema) = table_schemas.get(table_id) else {
                continue;
            };
            let Some(pk) = schema.primary_key() else {
                continue;
            };
            match staged {
                Staged::Delete(row_id) => {
                    let handle = self.table_by_id(*table_id)?;
                    let Some(row) = handle.lock().get(*row_id, snapshot) else {
                        continue;
                    };
                    let Some(pk_value) = row.columns.get(&pk.id) else {
                        continue;
                    };
                    old_rows.insert(idx, TriggerRowImage::from_row(row.clone()));
                    delete_by_key
                        .entry((*table_id, pk_value.encode_key()))
                        .or_default()
                        .push_back(idx);
                }
                Staged::Put(cells) => {
                    if let Some((_, value)) = cells.iter().find(|(id, _)| *id == pk.id) {
                        put_by_key
                            .entry((*table_id, value.encode_key()))
                            .or_default()
                            .push_back(idx);
                    }
                }
                Staged::Truncate => {}
            }
        }

        let mut paired_delete = std::collections::HashSet::new();
        let mut paired_put = std::collections::HashSet::new();
        let mut events = Vec::new();

        for (key, deletes) in delete_by_key.iter_mut() {
            let Some(puts) = put_by_key.get_mut(key) else {
                continue;
            };
            while let (Some(delete_idx), Some(put_idx)) = (deletes.pop_front(), puts.pop_front()) {
                paired_delete.insert(delete_idx);
                paired_put.insert(put_idx);
                let (table_id, _) = &staging[put_idx];
                let Some(table_name) = table_names.get(table_id).cloned() else {
                    continue;
                };
                let old = old_rows.get(&delete_idx).cloned();
                let new = match &staging[put_idx].1 {
                    Staged::Put(cells) => Some(TriggerRowImage::from_cells(cells)),
                    _ => None,
                };
                let changed_columns = changed_columns(old.as_ref(), new.as_ref());
                events.push(WriteEvent {
                    table: table_name,
                    kind: TriggerEvent::Update,
                    old,
                    new,
                    changed_columns,
                    op_indices: vec![delete_idx, put_idx],
                    put_idx: Some(put_idx),
                    trigger_stack: Self::trigger_stack_for_indices(
                        trigger_stacks,
                        &[delete_idx, put_idx],
                    ),
                });
            }
        }

        for (idx, (table_id, staged)) in staging.iter().enumerate() {
            let Some(table_name) = table_names.get(table_id).cloned() else {
                continue;
            };
            match staged {
                Staged::Put(cells) if !paired_put.contains(&idx) => {
                    let new = Some(TriggerRowImage::from_cells(cells));
                    let changed_columns = cells.iter().map(|(id, _)| *id).collect();
                    events.push(WriteEvent {
                        table: table_name,
                        kind: TriggerEvent::Insert,
                        old: None,
                        new,
                        changed_columns,
                        op_indices: vec![idx],
                        put_idx: Some(idx),
                        trigger_stack: Self::trigger_stack_for_indices(trigger_stacks, &[idx]),
                    });
                }
                Staged::Delete(row_id) if !paired_delete.contains(&idx) => {
                    let old = match old_rows.get(&idx).cloned() {
                        Some(old) => Some(old),
                        None => {
                            let handle = self.table_by_id(*table_id)?;
                            let row = handle.lock().get(*row_id, snapshot);
                            row.map(TriggerRowImage::from_row)
                        }
                    };
                    let Some(old) = old else {
                        continue;
                    };
                    let changed_columns = old.columns.keys().copied().collect();
                    events.push(WriteEvent {
                        table: table_name,
                        kind: TriggerEvent::Delete,
                        old: Some(old),
                        new: None,
                        changed_columns,
                        op_indices: vec![idx],
                        put_idx: None,
                        trigger_stack: Self::trigger_stack_for_indices(trigger_stacks, &[idx]),
                    });
                }
                Staged::Truncate => {}
                _ => {}
            }
        }

        Ok(events)
    }

    fn execute_trigger_program(
        &self,
        trigger: &StoredTrigger,
        event: &WriteEvent,
        mut staging: Option<&mut Vec<(u64, crate::txn::Staged)>>,
        out: &mut TriggerProgramOutput<'_>,
        trigger_stack: &[String],
    ) -> Result<TriggerProgramOutcome> {
        let mut event = event.clone();
        for step in &trigger.program.steps {
            match step {
                TriggerStep::SetNew { cells } => {
                    if trigger.timing != TriggerTiming::Before {
                        return Err(MongrelError::InvalidArgument(
                            "SetNew trigger step is only valid in BEFORE triggers".into(),
                        ));
                    }
                    let put_idx = event.put_idx.ok_or_else(|| {
                        MongrelError::InvalidArgument(
                            "SetNew trigger step requires INSERT or UPDATE NEW row".into(),
                        )
                    })?;
                    let staging = staging.as_deref_mut().ok_or_else(|| {
                        MongrelError::InvalidArgument(
                            "SetNew trigger step requires mutable trigger staging".into(),
                        )
                    })?;
                    let Some((_, crate::txn::Staged::Put(row_cells))) = staging.get_mut(put_idx)
                    else {
                        return Err(MongrelError::InvalidArgument(
                            "SetNew trigger step target row is not mutable".into(),
                        ));
                    };
                    for (column_id, value) in eval_trigger_cells(cells, &event)? {
                        row_cells.retain(|(id, _)| *id != column_id);
                        row_cells.push((column_id, value.clone()));
                        if let Some(new) = &mut event.new {
                            new.columns.insert(column_id, value);
                        }
                    }
                    row_cells.sort_by_key(|(id, _)| *id);
                }
                TriggerStep::Insert { table, cells } => {
                    let cells = eval_trigger_cells(cells, &event)?;
                    if let Ok(table_id) = self.table_id(table) {
                        out.added.push((table_id, crate::txn::Staged::Put(cells)));
                        out.added_stacks.push(trigger_stack.to_vec());
                    } else if self.external_table(table).is_some() {
                        out.added_external.push(ExternalTriggerWrite::Insert {
                            table: table.clone(),
                            cells,
                        });
                    } else {
                        return Err(MongrelError::NotFound(format!(
                            "trigger {:?} insert target {table:?} not found",
                            trigger.name
                        )));
                    }
                }
                TriggerStep::UpdateByPk { table, pk, cells } => {
                    let pk = eval_trigger_value(pk, &event)?;
                    let cells = eval_trigger_cells(cells, &event)?;
                    if self.external_table(table).is_some() {
                        out.added_external.push(ExternalTriggerWrite::UpdateByPk {
                            table: table.clone(),
                            pk,
                            cells,
                        });
                    } else {
                        let row_id = self
                            .table(table)?
                            .lock()
                            .lookup_pk(&pk.encode_key())
                            .ok_or_else(|| {
                                MongrelError::NotFound(format!(
                                    "trigger {:?} update target not found",
                                    trigger.name
                                ))
                            })?;
                        let handle = self.table(table)?;
                        let snapshot = Snapshot::at(self.epoch.visible());
                        let old = handle.lock().get(row_id, snapshot).ok_or_else(|| {
                            MongrelError::NotFound(format!(
                                "trigger {:?} update target not visible",
                                trigger.name
                            ))
                        })?;
                        let mut merged = old.columns;
                        for (column_id, value) in cells {
                            merged.insert(column_id, value);
                        }
                        out.added
                            .push((self.table_id(table)?, crate::txn::Staged::Delete(row_id)));
                        out.added_stacks.push(trigger_stack.to_vec());
                        out.added.push((
                            self.table_id(table)?,
                            crate::txn::Staged::Put(merged.into_iter().collect()),
                        ));
                        out.added_stacks.push(trigger_stack.to_vec());
                    }
                }
                TriggerStep::DeleteByPk { table, pk } => {
                    let pk = eval_trigger_value(pk, &event)?;
                    if self.external_table(table).is_some() {
                        out.added_external.push(ExternalTriggerWrite::DeleteByPk {
                            table: table.clone(),
                            pk,
                        });
                    } else {
                        let row_id = self
                            .table(table)?
                            .lock()
                            .lookup_pk(&pk.encode_key())
                            .ok_or_else(|| {
                                MongrelError::NotFound(format!(
                                    "trigger {:?} delete target not found",
                                    trigger.name
                                ))
                            })?;
                        out.added
                            .push((self.table_id(table)?, crate::txn::Staged::Delete(row_id)));
                        out.added_stacks.push(trigger_stack.to_vec());
                    }
                }
                TriggerStep::Select { .. } => {}
                TriggerStep::Raise { action, message } => match action {
                    TriggerRaiseAction::Ignore => return Ok(TriggerProgramOutcome::Ignore),
                    TriggerRaiseAction::Abort
                    | TriggerRaiseAction::Fail
                    | TriggerRaiseAction::Rollback => {
                        let message = eval_trigger_value(message, &event)?;
                        return Err(MongrelError::Conflict(format!(
                            "trigger {:?} raised: {}; trigger stack: {}",
                            trigger.name,
                            trigger_message(message),
                            Self::format_trigger_stack(trigger_stack)
                        )));
                    }
                },
            }
        }
        Ok(TriggerProgramOutcome::Continue)
    }

    fn trigger_stack_for_indices(stacks: Option<&[Vec<String>]>, indices: &[usize]) -> Vec<String> {
        let Some(stacks) = stacks else {
            return Vec::new();
        };
        let mut out = Vec::new();
        for idx in indices {
            let Some(stack) = stacks.get(*idx) else {
                continue;
            };
            for name in stack {
                if !out.iter().any(|existing| existing == name) {
                    out.push(name.clone());
                }
            }
        }
        out
    }

    fn trigger_stack_with(stack: &[String], trigger_name: &str) -> Vec<String> {
        let mut out = stack.to_vec();
        out.push(trigger_name.to_string());
        out
    }

    fn format_trigger_stack(stack: &[String]) -> String {
        if stack.is_empty() {
            "<root>".into()
        } else {
            stack.join(" -> ")
        }
    }

    /// Authoritatively validate every declared constraint on the staged write
    /// set under the transaction's read snapshot, AND expand ON DELETE CASCADE /
    /// SET NULL actions into explicit child ops. Called from
    /// [`Self::commit_transaction`] outside the WAL mutex. Returns the first
    /// violation as an `Err`, aborting the commit atomically. This is the
    /// server-side authority point: concurrent remote writers that each pass
    /// their own client-side checks still cannot both commit a violating batch.
    ///
    /// Scope: CHECK (full, three-valued), UNIQUE beyond the PK (existence scan +
    /// intra-transaction dedup; concurrent-txn races are additionally caught by
    /// `WriteKey::Unique`), and FK insert-side parent existence + ON DELETE
    /// {RESTRICT, CASCADE, SET NULL}. CASCADE appends child deletes (transitive
    /// fixpoint); SET NULL appends child updates (FK columns nulled). Truncate is
    /// RESTRICT-only (cascade-truncate is unsupported).
    fn validate_constraints(
        &self,
        staging: &mut Vec<(u64, crate::txn::Staged)>,
        read_epoch: Epoch,
    ) -> Result<()> {
        use crate::constraint::{encode_composite_key, validate_checks, FkAction};
        use crate::memtable::Row;
        use crate::txn::Staged;
        use std::collections::HashSet;

        let snapshot = Snapshot::at(read_epoch);
        let cat = self.catalog.read();

        // Collect live (id, name, constraints-bearing?) for staged tables.
        let live: Vec<(u64, &str, &crate::schema::Schema)> = cat
            .tables
            .iter()
            .filter(|e| matches!(e.state, TableState::Live))
            .map(|e| (e.table_id, e.name.as_str(), &e.schema))
            .collect();

        // Fast path: bail if no live table declares any constraints at all.
        let any_constraints = live.iter().any(|(_, _, s)| !s.constraints.is_empty());
        if !any_constraints {
            return Ok(());
        }

        // Lazily-loaded visible rows per table, shared across checks.
        let mut rows_cache: HashMap<u64, Vec<Row>> = HashMap::new();
        let mut load_rows = |table_id: u64| -> Result<Vec<Row>> {
            if let Some(r) = rows_cache.get(&table_id) {
                return Ok(r.clone());
            }
            let handle = self.table_by_id(table_id)?;
            let rows = handle.lock().visible_rows(snapshot)?;
            rows_cache.insert(table_id, rows.clone());
            Ok(rows)
        };

        // ── Phase A: expand ON DELETE CASCADE / SET NULL into explicit child
        // ops (transitive fixpoint). RESTRICT is not expanded here — it is
        // enforced as a violation in Phase B. `cascaded` records every delete
        // we have already expanded so a self-referential CASCADE FK cannot loop.
        let mut cascaded: HashSet<(u64, u64)> = HashSet::new();
        loop {
            let mut new_ops: Vec<(u64, Staged)> = Vec::new();
            let deletes: Vec<(u64, crate::rowid::RowId)> = staging
                .iter()
                .filter_map(|(t, op)| match op {
                    Staged::Delete(rid) => Some((*t, *rid)),
                    _ => None,
                })
                .collect();
            for (table_id, rid) in deletes {
                if !cascaded.insert((table_id, rid.0)) {
                    continue;
                }
                let Some(tname) = live
                    .iter()
                    .find(|(t, _, _)| *t == table_id)
                    .map(|(_, n, _)| *n)
                else {
                    continue;
                };
                let parent_handle = self.table_by_id(table_id)?;
                let Some(parent_row) = parent_handle.lock().get(rid, snapshot) else {
                    continue;
                };
                for (child_id, _child_name, child_schema) in &live {
                    for fk in &child_schema.constraints.foreign_keys {
                        if fk.ref_table != tname {
                            continue;
                        }
                        let Some(parent_key) =
                            encode_composite_key(&fk.ref_columns, &parent_row.columns)
                        else {
                            continue;
                        };
                        match fk.on_delete {
                            FkAction::Restrict => continue,
                            FkAction::Cascade => {
                                let child_rows = load_rows(*child_id)?;
                                for cr in &child_rows {
                                    if !cascaded.contains(&(*child_id, cr.row_id.0))
                                        && encode_composite_key(&fk.columns, &cr.columns).as_deref()
                                            == Some(parent_key.as_slice())
                                    {
                                        new_ops.push((*child_id, Staged::Delete(cr.row_id)));
                                    }
                                }
                            }
                            FkAction::SetNull => {
                                let child_rows = load_rows(*child_id)?;
                                for cr in &child_rows {
                                    if !cascaded.contains(&(*child_id, cr.row_id.0))
                                        && encode_composite_key(&fk.columns, &cr.columns).as_deref()
                                            == Some(parent_key.as_slice())
                                    {
                                        // Re-emit the child row with the FK
                                        // columns set to NULL (delete + put).
                                        let mut cells: Vec<(u16, crate::memtable::Value)> = cr
                                            .columns
                                            .iter()
                                            .map(|(k, v)| (*k, v.clone()))
                                            .collect();
                                        for cid in &fk.columns {
                                            cells.retain(|(k, _)| k != cid);
                                            cells.push((*cid, crate::memtable::Value::Null));
                                        }
                                        new_ops.push((*child_id, Staged::Delete(cr.row_id)));
                                        new_ops.push((*child_id, Staged::Put(cells)));
                                    }
                                }
                            }
                        }
                    }
                }
            }
            if new_ops.is_empty() {
                break;
            }
            staging.extend(new_ops);
        }

        // Rows staged for deletion in THIS transaction (now including cascaded
        // deletes). Used to exclude the old version of an updated row from
        // unique-existence scans.
        let staged_deletes: HashSet<(u64, u64)> = staging
            .iter()
            .filter_map(|(t, op)| match op {
                Staged::Delete(rid) => Some((*t, rid.0)),
                _ => None,
            })
            .collect();

        // Intra-transaction unique-key dedup: (table_id, uc_id, key).
        let mut seen_unique: HashSet<(u64, u16, Vec<u8>)> = HashSet::new();

        // ── Phase B: validate the fully-expanded staging set.
        for (table_id, op) in staging.iter() {
            let Some((_, tname, schema)) = live.iter().find(|(t, _, _)| t == table_id).copied()
            else {
                continue;
            };
            let cells_map: HashMap<u16, crate::memtable::Value>;
            match op {
                Staged::Put(cells) => {
                    cells_map = cells.iter().cloned().collect();

                    // CHECK constraints.
                    if !schema.constraints.checks.is_empty() {
                        validate_checks(&schema.constraints.checks, &cells_map)?;
                    }

                    // UNIQUE (non-PK) constraints.
                    for uc in &schema.constraints.uniques {
                        let Some(key) = encode_composite_key(&uc.columns, &cells_map) else {
                            continue; // NULL in a constrained column → skip (SQL).
                        };
                        let marker = (*table_id, uc.id, key.clone());
                        if !seen_unique.insert(marker) {
                            return Err(MongrelError::Conflict(format!(
                                "UNIQUE constraint '{}' on table '{tname}' violated within batch",
                                uc.name
                            )));
                        }
                        let rows = load_rows(*table_id)?;
                        for r in &rows {
                            // Skip rows this same transaction is deleting (the
                            // old version of an updated/cascade-deleted row).
                            if staged_deletes.contains(&(*table_id, r.row_id.0)) {
                                continue;
                            }
                            if let Some(theirs) = encode_composite_key(&uc.columns, &r.columns) {
                                if theirs == key {
                                    return Err(MongrelError::Conflict(format!(
                                        "UNIQUE constraint '{}' on table '{tname}' violated",
                                        uc.name
                                    )));
                                }
                            }
                        }
                    }

                    // FK insert-side: parent must exist.
                    for fk in &schema.constraints.foreign_keys {
                        let Some(child_key) = encode_composite_key(&fk.columns, &cells_map) else {
                            continue; // NULL FK component → not checked (SQL).
                        };
                        let Some(parent_id) = cat
                            .tables
                            .iter()
                            .find(|t| t.name == fk.ref_table)
                            .map(|t| t.table_id)
                        else {
                            return Err(MongrelError::InvalidArgument(format!(
                                "FOREIGN KEY '{}' references unknown table '{}'",
                                fk.name, fk.ref_table
                            )));
                        };
                        let parent_rows = load_rows(parent_id)?;
                        let mut found = false;
                        for r in &parent_rows {
                            if staged_deletes.contains(&(parent_id, r.row_id.0)) {
                                continue;
                            }
                            if let Some(pkey) = encode_composite_key(&fk.ref_columns, &r.columns) {
                                if pkey == child_key {
                                    found = true;
                                    break;
                                }
                            }
                        }
                        if !found {
                            return Err(MongrelError::Conflict(format!(
                                "FOREIGN KEY '{}' on table '{tname}' has no matching parent in '{}'",
                                fk.name, fk.ref_table
                            )));
                        }
                    }
                }
                Staged::Delete(rid) => {
                    // FK ON DELETE RESTRICT: a child row (whose FK action is
                    // RESTRICT) referencing this parent blocks the delete.
                    // CASCADE/SET NULL children were expanded in Phase A.
                    let parent_handle = self.table_by_id(*table_id)?;
                    let Some(parent_row) = parent_handle.lock().get(*rid, snapshot) else {
                        continue;
                    };
                    for (child_id, child_name, child_schema) in &live {
                        for fk in &child_schema.constraints.foreign_keys {
                            if fk.ref_table != tname || fk.on_delete != FkAction::Restrict {
                                continue;
                            }
                            let Some(parent_key) =
                                encode_composite_key(&fk.ref_columns, &parent_row.columns)
                            else {
                                continue;
                            };
                            let child_rows = load_rows(*child_id)?;
                            for r in &child_rows {
                                // A child already being deleted by this txn
                                // (cascade/inline) is not a restrict violation.
                                if staged_deletes.contains(&(*child_id, r.row_id.0)) {
                                    continue;
                                }
                                if let Some(ck) = encode_composite_key(&fk.columns, &r.columns) {
                                    if ck == parent_key {
                                        return Err(MongrelError::Conflict(format!(
                                            "FOREIGN KEY '{}' on table '{child_name}' restricts delete (parent referenced)",
                                            fk.name
                                        )));
                                    }
                                }
                            }
                        }
                    }
                }
                Staged::Truncate => {
                    // Truncate is RESTRICT-only: reject if any child references
                    // this table (any FK action), since cascade-truncate is
                    // unsupported.
                    for (child_id, child_name, child_schema) in &live {
                        for fk in &child_schema.constraints.foreign_keys {
                            if fk.ref_table != tname {
                                continue;
                            }
                            let child_rows = load_rows(*child_id)?;
                            if child_rows
                                .iter()
                                .any(|r| encode_composite_key(&fk.columns, &r.columns).is_some())
                            {
                                return Err(MongrelError::Conflict(format!(
                                    "FOREIGN KEY '{}' on table '{child_name}' restricts truncate of '{tname}'",
                                    fk.name
                                )));
                            }
                        }
                    }
                }
            }
        }
        Ok(())
    }

    /// Seal a transaction (spec §9.3):
    /// 1. Prepare — derive write keys, allocate row ids (brief table locks).
    /// 2. Sequencer — validate-first under the WAL mutex; abort on conflict
    ///    with no epoch consumed; assign epoch, append data records + TxnCommit,
    ///    group-sync, record conflict keys.
    /// 3. Publish — apply to tables, advance visible in-order.
    pub(crate) fn commit_transaction_with_external_states(
        &self,
        txn_id: u64,
        read_epoch: Epoch,
        mut staging: Vec<(u64, crate::txn::Staged)>,
        external_states: Vec<(String, Vec<u8>)>,
        external_trigger_bridge: Option<&dyn ExternalTriggerBridge>,
    ) -> Result<Epoch> {
        use crate::memtable::Row;
        use crate::txn::{Staged, StagedOp, WriteKey};
        use crate::wal::Op;
        use std::collections::hash_map::DefaultHasher;
        use std::hash::{Hash, Hasher};
        use std::sync::atomic::Ordering;

        if self.poisoned.load(Ordering::Relaxed) {
            return Err(MongrelError::Other(
                "database poisoned by fsync error".into(),
            ));
        }
        let mut external_states = dedup_external_states(external_states);
        if !external_states.is_empty() {
            let cat = self.catalog.read();
            for (name, _) in &external_states {
                if !cat.external_tables.iter().any(|entry| entry.name == *name) {
                    return Err(MongrelError::NotFound(format!(
                        "external table {name:?} not found"
                    )));
                }
            }
        }

        // ── 1. Prepare: fill generated values, expand triggers, validate, then
        // derive write keys from the final atomic write set.
        self.fill_auto_increment_for_staging(&mut staging)?;
        self.expand_table_triggers(
            &mut staging,
            read_epoch,
            external_trigger_bridge,
            &mut external_states,
        )?;
        self.fill_auto_increment_for_staging(&mut staging)?;
        external_states = dedup_external_states(external_states);

        // Validate declarative constraints (unique / FK / check) under the read
        // snapshot, outside the WAL mutex. Trigger-produced writes are included
        // here, so the batch either satisfies every declared constraint or is
        // rejected atomically.
        self.validate_constraints(&mut staging, read_epoch)?;

        let write_keys = {
            let cat = self.catalog.read();
            let mut keys: Vec<WriteKey> = Vec::new();
            for (table_id, staged) in &staging {
                match staged {
                    Staged::Put(cells) => {
                        if let Some(entry) = cat.tables.iter().find(|t| t.table_id == *table_id) {
                            for col in &entry.schema.columns {
                                if col.flags.contains(crate::schema::ColumnFlags::PRIMARY_KEY) {
                                    if let Some((_, val)) =
                                        cells.iter().find(|(id, _)| *id == col.id)
                                    {
                                        let mut h = DefaultHasher::new();
                                        val.encode_key().hash(&mut h);
                                        keys.push(WriteKey::Unique {
                                            table_id: *table_id,
                                            index_id: 0,
                                            key_hash: h.finish(),
                                        });
                                    }
                                }
                            }
                            // Declared non-PK unique constraints register a
                            // `WriteKey::Unique` (namespace-separated from the
                            // PK's index_id==0 by setting the high bit) so two
                            // concurrent transactions inserting the same key
                            // cannot both commit. Rows with any NULL constrained
                            // column are skipped (SQL semantics).
                            for uc in &entry.schema.constraints.uniques {
                                if let Some(key_bytes) = crate::constraint::encode_composite_key(
                                    &uc.columns,
                                    &cells.iter().cloned().collect(),
                                ) {
                                    let mut h = DefaultHasher::new();
                                    key_bytes.hash(&mut h);
                                    keys.push(WriteKey::Unique {
                                        table_id: *table_id,
                                        index_id: uc.id | 0x8000,
                                        key_hash: h.finish(),
                                    });
                                }
                            }
                        }
                    }
                    Staged::Delete(rid) => keys.push(WriteKey::Row {
                        table_id: *table_id,
                        row_id: rid.0,
                    }),
                    Staged::Truncate => keys.push(WriteKey::Table {
                        table_id: *table_id,
                    }),
                }
            }
            for (name, _) in &external_states {
                let mut h = DefaultHasher::new();
                name.hash(&mut h);
                keys.push(WriteKey::Unique {
                    table_id: EXTERNAL_TABLE_ID,
                    index_id: 0,
                    key_hash: h.finish(),
                });
            }
            keys
        };

        // Opportunistic pruning.
        let min_active = self.active_txns.min_read_epoch();
        if min_active < u64::MAX {
            self.conflicts.prune_below(Epoch(min_active));
        }

        // ── 1a. Pre-validate the full write-set OUTSIDE the sequencer (spec
        // §8.5, review fix #17). Snapshot the conflict-index version so the
        // sequencer only re-checks if new commits arrived in the interim.
        if self.conflicts.conflicts(&write_keys, read_epoch) {
            return Err(MongrelError::Conflict(
                "write-write conflict (pre-validate, first-committer-wins)".into(),
            ));
        }
        let pre_validate_version = self.conflicts.version();

        // ── 1b. Spill: if a table's staged puts exceed the threshold, write a
        // uniform-epoch pending run (spec §8.5). Rows in the run are NOT
        // streamed as Put records; they are linked at publish time.
        let mut spilled: Vec<SpilledRun> = Vec::new();
        let mut spilled_tables: std::collections::HashSet<u64> = std::collections::HashSet::new();
        // Protect this txn's `_txn/<id>/` dir from a concurrent `gc()` for as long
        // as the spill runs are live (registered on first spill, dropped at the
        // end of this function on commit/abort/error).
        let mut spill_guard: Option<crate::retention::SpillGuard> = None;
        {
            let mut table_bytes: HashMap<u64, usize> = HashMap::new();
            for (table_id, staged) in &staging {
                if let Staged::Put(cells) = staged {
                    *table_bytes.entry(*table_id).or_default() += cells.len() * 16;
                }
            }
            let tables = self.tables.read();
            for (&table_id, &bytes) in &table_bytes {
                if bytes as u64
                    <= self
                        .spill_threshold
                        .load(std::sync::atomic::Ordering::Relaxed)
                {
                    continue;
                }
                let Some(handle) = tables.get(&table_id) else {
                    continue;
                };
                spill_guard.get_or_insert_with(|| self.active_spills.register(txn_id));
                let mut t = handle.lock();
                let tdir = t.table_dir().to_path_buf();
                let txn_dir = tdir.join("_txn").join(txn_id.to_string());
                std::fs::create_dir_all(&txn_dir)?;
                let run_id = t.alloc_run_id() as u128;
                let pending_path = txn_dir.join(format!("r-{run_id}.sr"));

                let mut rows: Vec<Row> = Vec::new();
                for (tid, staged) in &staging {
                    if *tid != table_id {
                        continue;
                    }
                    if let Staged::Put(cells) = staged {
                        t.validate_cells_not_null(cells)?;
                        let row_id = t.alloc_row_id();
                        let mut row = Row::new(row_id, Epoch(0));
                        for (c, v) in cells {
                            row.columns.insert(*c, v.clone());
                        }
                        rows.push(row);
                    }
                }
                let schema = t.schema_ref().clone();
                let kek = t.kek_ref().cloned();
                let specs = t.indexable_column_specs();
                drop(t);

                let mut writer = crate::sorted_run::RunWriter::new(&schema, run_id, Epoch(0), 0)
                    .uniform_epoch(true);
                if let Some(ref kek) = kek {
                    writer = writer.with_encryption(kek.as_ref(), specs);
                }
                let header = writer.write(&pending_path, &rows)?;
                let row_count = header.row_count;
                let min_rid = rows.first().map(|r| r.row_id.0).unwrap_or(0);
                let max_rid = rows.last().map(|r| r.row_id.0).unwrap_or(0);

                spilled.push(SpilledRun {
                    table_id,
                    run_id,
                    pending_path,
                    rows,
                    row_count,
                    min_rid,
                    max_rid,
                });
                spilled_tables.insert(table_id);
            }
        }

        // Test seam: let a test race `gc()` against this in-flight spill.
        if spill_guard.is_some() {
            if let Some(hook) = self.spill_hook.lock().as_ref() {
                hook();
            }
        }

        // ── 1c. Pre-build non-spilled put rows OUTSIDE the WAL critical section.
        // Allocating row ids + building the rows here (lock order: table handle →
        // nothing) means the sequencer never locks a table handle while holding
        // the shared-WAL mutex. That matters because `Table::commit`/`flush` lock
        // the table handle THEN the shared WAL; if the sequencer did the reverse
        // (WAL then handle) the two paths would deadlock (review fix: B1).
        // Aligned 1:1 with `staging`; `None` for deletes and spilled puts.
        // Row ids are allocated here, before the sequencer's delta conflict
        // re-check, so a losing txn leaks the ids it reserved — harmless, the
        // u64 row-id space is monotonic and gaps are expected (spills do the same).
        let mut prebuilt: Vec<Option<Row>> = Vec::with_capacity(staging.len());
        {
            let tables = self.tables.read();
            for (table_id, staged) in &staging {
                match staged {
                    Staged::Put(cells) if !spilled_tables.contains(table_id) => {
                        let handle = tables.get(table_id).ok_or_else(|| {
                            MongrelError::NotFound(format!("table {table_id} not mounted"))
                        })?;
                        let mut t = handle.lock();
                        t.validate_cells_not_null(cells)?;
                        let row_id = t.alloc_row_id();
                        drop(t);
                        let mut row = Row::new(row_id, Epoch(0));
                        for (c, v) in cells {
                            row.columns.insert(*c, v.clone());
                        }
                        prebuilt.push(Some(row));
                    }
                    Staged::Put(_) | Staged::Delete(_) | Staged::Truncate => prebuilt.push(None),
                }
            }
        }

        let mut prepared_external = Vec::with_capacity(external_states.len());
        for (name, state) in &external_states {
            let pending = prepare_external_state_file(&self.root, name, state, txn_id)?;
            prepared_external.push((name.clone(), state.clone(), pending));
        }

        // ── 2. Sequencer: validate-first → assign → append → sync → record ──
        let added_runs: Vec<crate::wal::AddedRun> = spilled
            .iter()
            .map(|s| crate::wal::AddedRun {
                table_id: s.table_id,
                run_id: s.run_id,
                row_count: s.row_count,
                level: 0,
                min_row_id: s.min_rid,
                max_row_id: s.max_rid,
                content_hash: [0u8; 32],
            })
            .collect();
        let (new_epoch, applies, commit_seq) = {
            let mut wal = self.shared_wal.lock();

            // Re-check only if the conflict index advanced since pre-validation
            // (bounded delta — spec §8.5, review fix #17). If the version is
            // unchanged, the pre-check result is still valid and the sequencer
            // does O(1) work regardless of write-set size.
            if self.conflicts.version() != pre_validate_version
                && self.conflicts.conflicts(&write_keys, read_epoch)
            {
                // Abort: this txn assigned no epoch yet, so drop the quarantined
                // spill runs we wrote during prepare instead of leaking them in
                // `_txn/` until the next GC/reopen sweep.
                drop(wal);
                for s in &spilled {
                    if let Some(parent) = s.pending_path.parent() {
                        let _ = std::fs::remove_dir_all(parent);
                    }
                }
                for (_, _, pending) in &prepared_external {
                    let _ = std::fs::remove_file(pending);
                }
                return Err(MongrelError::Conflict(
                    "write-write conflict (sequencer delta re-check)".into(),
                ));
            }

            let new_epoch = self.epoch.bump_assigned();
            let mut applies: Vec<(u64, Vec<StagedOp>)> = Vec::new();

            for (idx, (table_id, staged)) in staging.iter().enumerate() {
                // Skip puts for tables that were spilled — their data is in a
                // pending run, not in streamed Put records.
                if spilled_tables.contains(table_id) && matches!(staged, Staged::Put(_)) {
                    continue;
                }
                let mut ops = Vec::new();
                match staged {
                    Staged::Put(_) => {
                        // Stamp the pre-built row at the real assigned epoch.
                        let mut row = prebuilt[idx].take().expect("prebuilt put row");
                        row.committed_epoch = new_epoch;
                        let payload = bincode::serialize(&vec![row.clone()])
                            .map_err(|e| MongrelError::Other(format!("row serialize: {e}")))?;
                        wal.append(
                            txn_id,
                            *table_id,
                            Op::Put {
                                table_id: *table_id,
                                rows: payload,
                            },
                        )?;
                        ops.push(StagedOp::Put(row));
                    }
                    Staged::Delete(rid) => {
                        wal.append(
                            txn_id,
                            *table_id,
                            Op::Delete {
                                table_id: *table_id,
                                row_ids: vec![*rid],
                            },
                        )?;
                        ops.push(StagedOp::Delete(*rid));
                    }
                    Staged::Truncate => {
                        wal.append(
                            txn_id,
                            *table_id,
                            Op::TruncateTable {
                                table_id: *table_id,
                            },
                        )?;
                        ops.push(StagedOp::Truncate);
                    }
                }
                applies.push((*table_id, ops));
            }

            for (name, state, _) in &prepared_external {
                wal.append(
                    txn_id,
                    EXTERNAL_TABLE_ID,
                    Op::ExternalTableState {
                        name: name.clone(),
                        state: state.clone(),
                    },
                )?;
            }

            let commit_seq = wal.append_commit(txn_id, new_epoch, &added_runs)?;

            // Record the conflict + assign the epoch under the WAL lock so commit
            // order == WAL append order, but DO NOT fsync here (P3.2): the fsync
            // moves out of this critical section to the group-commit coordinator
            // so concurrent committers share a single leader fsync.
            self.conflicts.record(&write_keys, new_epoch);
            (new_epoch, applies, commit_seq)
        };

        // ── 2b. Durability: one leader fsync serves this whole batch (P3.2). ──
        self.group
            .await_durable(&self.shared_wal, commit_seq)
            .inspect_err(|_| {
                self.poisoned.store(true, Ordering::Relaxed);
            })?;

        // ── 3. Publish: link spilled runs + apply non-spilled ops ──
        {
            let tables = self.tables.read();
            // Link spilled runs first.
            for s in &spilled {
                if let Some(handle) = tables.get(&s.table_id) {
                    let mut t = handle.lock();
                    let dest = t.run_path(s.run_id as u64);
                    std::fs::rename(&s.pending_path, &dest)?;
                    // Clean up the now-empty `_txn/<txn_id>/` dir.
                    if let Some(parent) = s.pending_path.parent() {
                        let _ = std::fs::remove_dir_all(parent);
                    }
                    t.link_run(crate::manifest::RunRef {
                        run_id: s.run_id,
                        level: 0,
                        epoch_created: new_epoch.0,
                        row_count: s.row_count,
                    });
                    // Update indexes + live_count from the spilled rows WITHOUT
                    // materializing them in the memtable (P3.4): they are served
                    // from the linked uniform-epoch run, read at `new_epoch` via
                    // the RunRef. This keeps peak memory bounded for large txns.
                    t.apply_run_metadata(&s.rows)?;
                    t.invalidate_pending_cache();
                    t.persist_manifest(new_epoch)?;
                }
            }
            // Apply non-spilled ops.
            for (table_id, ops) in applies {
                if let Some(handle) = tables.get(&table_id) {
                    let mut t = handle.lock();
                    for op in ops {
                        match op {
                            StagedOp::Put(row) => t.apply_put_rows(vec![row])?,
                            StagedOp::Delete(rid) => t.apply_delete(rid, new_epoch),
                            StagedOp::Truncate => t.apply_truncate(new_epoch)?,
                        }
                    }
                    t.invalidate_pending_cache();
                    t.persist_manifest(new_epoch)?;
                }
            }
        }
        for (name, _, pending) in &prepared_external {
            publish_external_state_file(&self.root, name, pending)?;
        }

        self.advance_visible(new_epoch);
        Ok(new_epoch)
    }

    /// Advance `visible` in-order: epoch E becomes visible only once E and all
    /// prior unpublished epochs have finished publishing (spec §9.3e). The
    /// in-order gate lives on the shared [`EpochAuthority`] so this path, the
    /// single-table `Table::commit` path, and DDL all share one watermark and
    /// can never publish out of assigned order under concurrency.
    fn advance_visible(&self, published: Epoch) {
        self.epoch.publish_in_order(published);
    }

    /// Register a read snapshot at the current visible epoch and return it with
    /// a guard that retains it for GC until dropped.
    pub fn snapshot(&self) -> (Snapshot, SnapshotGuard<'_>) {
        let e = self.epoch.visible();
        let g = self.snapshots.register(e);
        (Snapshot::at(e), g)
    }

    /// Owned (clonable-handle) variant of [`Self::snapshot`] for cross-thread
    /// retention.
    pub fn snapshot_owned(&self) -> (Snapshot, OwnedSnapshotGuard) {
        let e = self.epoch.visible();
        let g = self.snapshots.register_owned(e);
        (Snapshot::at(e), g)
    }

    /// Names of all live tables.
    pub fn table_names(&self) -> Vec<String> {
        self.catalog
            .read()
            .tables
            .iter()
            .filter(|t| matches!(t.state, TableState::Live))
            .map(|t| t.name.clone())
            .collect()
    }

    /// Best-effort flush-on-close (§4.4): force-flush every mounted table
    /// that has pending writes to a `.sr` sorted run, so WAL segments can be
    /// reaped on the next open. Call this as the last action before a
    /// short-lived process (CLI, one-shot script) exits. The daemon does not
    /// need this — its background auto-compactor handles run management.
    pub fn close(&self) -> Result<()> {
        for name in self.table_names() {
            if let Ok(handle) = self.table(&name) {
                if let Err(e) = handle.lock().close() {
                    eprintln!("[close] flush failed for {name}: {e}");
                }
            }
        }
        Ok(())
    }

    /// Compact every mounted table: merge all sorted runs into one clean run
    /// so query cost stays flat (single-run fast path) instead of growing
    /// with run count. Tables with < 2 runs are skipped (no-op). Each table
    /// is locked individually for its own compaction; snapshot retention is
    /// honored by `Table::compact`. Returns `(tables_compacted, tables_skipped)`.
    pub fn compact(&self) -> Result<(usize, usize)> {
        let mut compacted = 0;
        let mut skipped = 0;
        for name in self.table_names() {
            let Ok(handle) = self.table(&name) else {
                continue;
            };
            {
                let mut t = handle.lock();
                let before = t.run_count();
                if before < 2 {
                    skipped += 1;
                    continue;
                }
                match t.compact() {
                    Ok(()) => {
                        let after = t.run_count();
                        compacted += 1;
                        eprintln!("[compact] {name}: {before} -> {after} runs");
                    }
                    Err(e) => {
                        eprintln!("[compact] {name}: compaction failed: {e}");
                        skipped += 1;
                    }
                }
            }
        }
        Ok((compacted, skipped))
    }

    /// Compact a single table by name. Returns `Ok(true)` if it was
    /// compacted, `Ok(false)` if skipped (< 2 runs).
    pub fn compact_table(&self, name: &str) -> Result<bool> {
        let handle = self.table(name)?;
        let mut t = handle.lock();
        let before = t.run_count();
        if before < 2 {
            return Ok(false);
        }
        t.compact()?;
        Ok(t.run_count() < before)
    }

    /// Look up a live table by name.
    pub fn table(&self, name: &str) -> Result<TableHandle> {
        let cat = self.catalog.read();
        let entry = cat
            .live(name)
            .ok_or_else(|| MongrelError::NotFound(format!("table {name:?} not found")))?;
        let id = entry.table_id;
        drop(cat);
        self.tables
            .read()
            .get(&id)
            .cloned()
            .ok_or_else(|| MongrelError::NotFound(format!("table {name:?} not mounted")))
    }

    /// Resolve a live table id → mounted handle (used by the constraint
    /// validation pass and other id-qualified internal paths).
    fn table_by_id(&self, id: u64) -> Result<TableHandle> {
        self.tables
            .read()
            .get(&id)
            .cloned()
            .ok_or_else(|| MongrelError::NotFound(format!("table id {id} not mounted")))
    }

    /// Create a new table. The DDL is first logged to the shared WAL
    /// (`Op::Ddl(CreateTable)` + `TxnCommit`) and group-synced so it is durable
    /// BEFORE the in-memory catalog and table map are mutated; the catalog
    /// checkpoint is rewritten afterwards (spec §15, review fix #16). A reopen
    /// that sees a stale catalog still recovers the table by replaying the Ddl.
    pub fn create_table(&self, name: &str, schema: Schema) -> Result<u64> {
        use crate::wal::DdlOp;
        use std::sync::atomic::Ordering;

        if self.poisoned.load(Ordering::Relaxed) {
            return Err(MongrelError::Other(
                "database poisoned by fsync error".into(),
            ));
        }

        let _g = self.ddl_lock.lock();
        {
            let cat = self.catalog.read();
            if cat.live(name).is_some() {
                return Err(MongrelError::InvalidArgument(format!(
                    "table {name:?} already exists"
                )));
            }
        }

        // Allocate id + epoch + txn id under the commit lock so the DDL commit
        // is serialized with data commits (in-order publish).
        let commit_lock = Arc::clone(&self.commit_lock);
        let _c = commit_lock.lock();
        let table_id = {
            let mut cat = self.catalog.write();
            let id = cat.next_table_id;
            cat.next_table_id += 1;
            id
        };
        let epoch = self.epoch.bump_assigned();
        let txn_id = self.alloc_txn_id();

        // Stamp the schema_id with the unique table_id so every table in the
        // database has a distinct schema_id (caller-provided values are
        // ignored to prevent collisions).
        let mut schema = schema;
        schema.schema_id = table_id;
        // Defense in depth: reject an invalid schema BEFORE any durable
        // side-effect. `Table::create_in` re-validates, but by then the DDL has
        // already been appended to the shared WAL; a failing create_in would
        // leave a dangling entry that `recover_ddl_from_wal` replays without
        // re-validating, corrupting the catalog on reopen. Validating here
        // keeps the WAL free of schemas that can never be opened.
        schema.validate_auto_increment()?;

        // 1. Log the DDL + commit marker to the shared WAL, then make it durable
        //    via the group-commit coordinator (no fsync under the WAL lock — P3.2).
        let schema_json = DdlOp::encode_schema(&schema)?;
        let commit_seq = {
            let mut wal = self.shared_wal.lock();
            wal.append(
                txn_id,
                table_id,
                crate::wal::Op::Ddl(DdlOp::CreateTable {
                    table_id,
                    name: name.to_string(),
                    schema_json,
                }),
            )?;
            wal.append_commit(txn_id, epoch, &[])?
        };
        self.group
            .await_durable(&self.shared_wal, commit_seq)
            .inspect_err(|_| {
                self.poisoned.store(true, Ordering::Relaxed);
            })?;

        // 2. Create the on-disk table dir + manifest.
        let tdir = self.root.join(TABLES_DIR).join(table_id.to_string());
        std::fs::create_dir_all(&tdir)?;
        let ctx = SharedCtx {
            epoch: Arc::clone(&self.epoch),
            page_cache: Arc::clone(&self.page_cache),
            decoded_cache: Arc::clone(&self.decoded_cache),
            snapshots: Arc::clone(&self.snapshots),
            kek: self.kek.clone(),
            commit_lock: Arc::clone(&self.commit_lock),
            shared: Some(crate::engine::SharedWalCtx {
                wal: Arc::clone(&self.shared_wal),
                group: Arc::clone(&self.group),
                poisoned: Arc::clone(&self.poisoned),
                txn_ids: Arc::clone(&self.next_txn_id),
            }),
        };
        let table = Table::create_in(&tdir, schema.clone(), table_id, ctx)?;

        // 3. Mutate the in-memory catalog + mount the table, then rewrite the
        //    catalog checkpoint (lazy: outside the WAL critical section).
        {
            let mut cat = self.catalog.write();
            cat.tables.push(CatalogEntry {
                table_id,
                name: name.to_string(),
                schema,
                state: TableState::Live,
                created_epoch: epoch.0,
            });
        }
        catalog::write_atomic(&self.root, &self.catalog.read(), self.meta_dek.as_ref())?;
        self.tables
            .write()
            .insert(table_id, Arc::new(Mutex::new(table)));

        self.advance_visible(epoch);
        Ok(table_id)
    }

    /// Logically drop a table, logging the DDL through the shared WAL first.
    pub fn drop_table(&self, name: &str) -> Result<()> {
        use crate::wal::DdlOp;
        use std::sync::atomic::Ordering;

        if self.poisoned.load(Ordering::Relaxed) {
            return Err(MongrelError::Other(
                "database poisoned by fsync error".into(),
            ));
        }

        let _g = self.ddl_lock.lock();
        let table_id = {
            let cat = self.catalog.read();
            cat.live(name)
                .ok_or_else(|| MongrelError::NotFound(format!("table {name:?} not found")))?
                .table_id
        };

        let commit_lock = Arc::clone(&self.commit_lock);
        let _c = commit_lock.lock();
        let epoch = self.epoch.bump_assigned();
        let txn_id = self.alloc_txn_id();
        let commit_seq = {
            let mut wal = self.shared_wal.lock();
            wal.append(
                txn_id,
                table_id,
                crate::wal::Op::Ddl(DdlOp::DropTable { table_id }),
            )?;
            wal.append_commit(txn_id, epoch, &[])?
        };
        self.group
            .await_durable(&self.shared_wal, commit_seq)
            .inspect_err(|_| {
                self.poisoned.store(true, Ordering::Relaxed);
            })?;

        {
            let mut cat = self.catalog.write();
            let entry = cat
                .tables
                .iter_mut()
                .find(|t| t.table_id == table_id)
                .ok_or_else(|| MongrelError::NotFound(format!("table {name:?} not found")))?;
            entry.state = TableState::Dropped { at_epoch: epoch.0 };
            cat.triggers.retain(|trigger| {
                !matches!(
                    &trigger.trigger.target,
                    TriggerTarget::Table(target) if target == name
                )
            });
        }
        catalog::write_atomic(&self.root, &self.catalog.read(), self.meta_dek.as_ref())?;
        self.tables.write().remove(&table_id);

        self.advance_visible(epoch);
        Ok(())
    }

    /// Rename a live table. `name` must exist and `new_name` must not collide
    /// with any live table; both checks run under `ddl_lock` so they are atomic
    /// with the rename and with concurrent `create_table` existence checks (no
    /// TOCTOU window). A no-op rename (`name == new_name`) succeeds without
    /// side-effects. The rename is logged to the shared WAL as
    /// `DdlOp::RenameTable` and recovered on reopen; the `table_id`, schema,
    /// and on-disk layout are unchanged (the table is keyed by `table_id`, so
    /// the in-memory object does not move — only the catalog name changes).
    pub fn rename_table(&self, name: &str, new_name: &str) -> Result<()> {
        use crate::wal::DdlOp;
        use std::sync::atomic::Ordering;

        if self.poisoned.load(Ordering::Relaxed) {
            return Err(MongrelError::Other(
                "database poisoned by fsync error".into(),
            ));
        }

        // A no-op rename short-circuits before any locking, so it can never
        // trip the "target already exists" check (the source *is* that name).
        if name == new_name {
            return Ok(());
        }
        if new_name.is_empty() {
            return Err(MongrelError::InvalidArgument(
                "rename_table: new name must not be empty".into(),
            ));
        }

        let _g = self.ddl_lock.lock();
        let table_id = {
            let cat = self.catalog.read();
            let src = cat
                .live(name)
                .ok_or_else(|| MongrelError::NotFound(format!("table {name:?} not found")))?;
            // Target must be free. Checked under ddl_lock, which every other
            // DDL (create/rename/drop) also holds, so a concurrent operation
            // cannot claim `new_name` between this check and the catalog write.
            if cat.live(new_name).is_some() {
                return Err(MongrelError::InvalidArgument(format!(
                    "rename_table: a table named {new_name:?} already exists"
                )));
            }
            src.table_id
        };

        let commit_lock = Arc::clone(&self.commit_lock);
        let _c = commit_lock.lock();
        let epoch = self.epoch.bump_assigned();
        let txn_id = self.alloc_txn_id();
        let commit_seq = {
            let mut wal = self.shared_wal.lock();
            wal.append(
                txn_id,
                table_id,
                crate::wal::Op::Ddl(DdlOp::RenameTable {
                    table_id,
                    new_name: new_name.to_string(),
                }),
            )?;
            wal.append_commit(txn_id, epoch, &[])?
        };
        self.group
            .await_durable(&self.shared_wal, commit_seq)
            .inspect_err(|_| {
                self.poisoned.store(true, Ordering::Relaxed);
            })?;

        {
            let mut cat = self.catalog.write();
            let entry = cat
                .tables
                .iter_mut()
                .find(|t| t.table_id == table_id)
                .ok_or_else(|| MongrelError::NotFound(format!("table {name:?} not found")))?;
            entry.name = new_name.to_string();
            for trigger in &mut cat.triggers {
                if matches!(
                    &trigger.trigger.target,
                    TriggerTarget::Table(target) if target == name
                ) {
                    trigger.trigger = trigger.trigger.retarget_table(new_name, epoch.0)?;
                }
            }
        }
        catalog::write_atomic(&self.root, &self.catalog.read(), self.meta_dek.as_ref())?;
        // The in-memory table object is keyed by table_id, not name, so it does
        // not move and live TableHandles remain valid.
        self.advance_visible(epoch);
        Ok(())
    }

    pub fn alter_column(
        &self,
        table_name: &str,
        column_name: &str,
        change: AlterColumn,
    ) -> Result<ColumnDef> {
        use crate::wal::DdlOp;
        use std::sync::atomic::Ordering;

        if self.poisoned.load(Ordering::Relaxed) {
            return Err(MongrelError::Other(
                "database poisoned by fsync error".into(),
            ));
        }

        let _g = self.ddl_lock.lock();
        let table_id = {
            let cat = self.catalog.read();
            cat.live(table_name)
                .ok_or_else(|| MongrelError::NotFound(format!("table {table_name:?} not found")))?
                .table_id
        };
        let handle =
            self.tables.read().get(&table_id).cloned().ok_or_else(|| {
                MongrelError::NotFound(format!("table {table_name:?} not mounted"))
            })?;
        let mut table = handle.lock();
        let column = table.prepare_alter_column(column_name, &change)?;
        let renamed_column = (column.name != column_name).then(|| column.name.clone());
        if table
            .schema()
            .columns
            .iter()
            .find(|c| c.id == column.id)
            .is_some_and(|c| c == &column)
        {
            return Ok(column);
        }

        let commit_lock = Arc::clone(&self.commit_lock);
        let _c = commit_lock.lock();
        let epoch = self.epoch.bump_assigned();
        let txn_id = self.alloc_txn_id();
        let column_json = DdlOp::encode_column(&column)?;
        let commit_seq = {
            let mut wal = self.shared_wal.lock();
            wal.append(
                txn_id,
                table_id,
                crate::wal::Op::Ddl(DdlOp::AlterTable {
                    table_id,
                    column_json,
                }),
            )?;
            wal.append_commit(txn_id, epoch, &[])?
        };
        self.group
            .await_durable(&self.shared_wal, commit_seq)
            .inspect_err(|_| {
                self.poisoned.store(true, Ordering::Relaxed);
            })?;

        table.apply_altered_column(column.clone())?;
        let schema = table.schema().clone();
        drop(table);

        {
            let mut cat = self.catalog.write();
            let entry = cat
                .tables
                .iter_mut()
                .find(|t| t.table_id == table_id)
                .ok_or_else(|| MongrelError::NotFound(format!("table {table_name:?} not found")))?;
            entry.schema = schema;
            if let Some(new_column_name) = renamed_column {
                for trigger in &mut cat.triggers {
                    if matches!(
                        &trigger.trigger.target,
                        TriggerTarget::Table(target) if target == table_name
                    ) {
                        trigger.trigger = trigger.trigger.renamed_update_column(
                            column_name,
                            new_column_name.clone(),
                            epoch.0,
                        )?;
                    }
                }
            }
        }
        catalog::write_atomic(&self.root, &self.catalog.read(), self.meta_dek.as_ref())?;

        self.advance_visible(epoch);
        Ok(column)
    }

    /// Retention-gated garbage collection (spec §6.4, §7.4, §16). Deletes:
    /// - Dropped-table subdirs whose `at_epoch < min_active_snapshot`.
    /// - Stale `_txn/` dirs (aborted/crashed large-txn pending runs).
    ///
    /// Returns the number of items reclaimed.
    pub fn gc(&self) -> Result<usize> {
        let min_active = self.snapshots.min_active(self.epoch.visible()).0;
        let mut reclaimed = 0;

        // Reclaim dropped-table dirs where no pinned snapshot still needs them.
        let cat = self.catalog.read();
        for entry in &cat.tables {
            if let TableState::Dropped { at_epoch } = entry.state {
                if at_epoch <= min_active {
                    let tdir = self.root.join(TABLES_DIR).join(entry.table_id.to_string());
                    if tdir.exists() {
                        std::fs::remove_dir_all(&tdir)?;
                        reclaimed += 1;
                    }
                }
            }
        }
        drop(cat);

        // Sweep stale _txn/<id>/ dirs on remaining live tables — but NEVER an
        // in-flight spill's dir (deleting it would lose the pending run and fail
        // the commit, review fix #14). Each `_txn/` subdir is named by its txn id;
        // skip any id still registered in `active_spills`.
        let cat = self.catalog.read();
        for entry in &cat.tables {
            if !matches!(entry.state, TableState::Live) {
                continue;
            }
            let txn_dir = self
                .root
                .join(TABLES_DIR)
                .join(entry.table_id.to_string())
                .join("_txn");
            if !txn_dir.exists() {
                continue;
            }
            for sub in std::fs::read_dir(&txn_dir)? {
                let sub = sub?;
                let name = sub.file_name();
                let Some(name) = name.to_str() else { continue };
                // A non-numeric entry can't belong to a live txn — sweep it.
                let is_active = name
                    .parse::<u64>()
                    .map(|id| self.active_spills.is_active(id))
                    .unwrap_or(false);
                if is_active {
                    continue;
                }
                std::fs::remove_dir_all(sub.path())?;
                reclaimed += 1;
            }
        }
        drop(cat);

        let external_names = {
            let cat = self.catalog.read();
            cat.external_tables
                .iter()
                .map(|entry| entry.name.clone())
                .collect::<std::collections::HashSet<_>>()
        };
        let vtab_dir = self.root.join(VTAB_DIR);
        if vtab_dir.exists() {
            for entry in std::fs::read_dir(&vtab_dir)? {
                let entry = entry?;
                let name = entry.file_name();
                let Some(name) = name.to_str() else { continue };
                if external_names.contains(name) {
                    continue;
                }
                let path = entry.path();
                if path.is_dir() {
                    std::fs::remove_dir_all(path)?;
                } else {
                    std::fs::remove_file(path)?;
                }
                reclaimed += 1;
            }
        }

        // Reap compaction-superseded runs whose retire epoch no pinned snapshot
        // can still need (spec §6.4). Each table deletes its own retired files
        // gated on `min_active` and persists its manifest.
        let tables = self.tables.read();
        for handle in tables.values() {
            reclaimed += handle.lock().reap_retiring(Epoch(min_active))?;
        }

        // WAL-segment GC (spec §6.4/§16). `SharedWal::open` mints a fresh active
        // segment on every reopen without truncating the prior ones, so rotated
        // segments accumulate. Once every live table's committed data is durable
        // in runs (no in-memory rows) and no in-flight spill is open, all rotated
        // (non-active) segments are redundant for recovery and safe to delete —
        // an in-flight txn only ever appends to the active segment, which is
        // never deleted.
        let all_durable = self.active_spills.is_idle()
            && tables.values().all(|h| {
                let g = h.lock();
                g.memtable_len() == 0 && g.mutable_run_len() == 0
            });
        drop(tables);
        if all_durable {
            reclaimed += self.shared_wal.lock().gc_segments(u64::MAX)?;
        }

        Ok(reclaimed)
    }
    fn alloc_txn_id(&self) -> u64 {
        let mut g = self.next_txn_id.lock();
        let id = *g;
        *g = g.wrapping_add(1);
        id
    }

    /// Set the per-table spill threshold (bytes). When a transaction's staged
    /// bytes for a single table exceed this, the rows are written as a
    /// uniform-epoch pending run instead of streamed Put records (spec §8.5).
    pub fn set_spill_threshold(&self, bytes: u64) {
        self.spill_threshold
            .store(bytes, std::sync::atomic::Ordering::Relaxed);
    }

    /// Test-only: install a hook invoked after a transaction writes its spill
    /// runs but before the sequencer, so a test can race `gc()` against an
    /// in-flight spill. Not part of the stable API.
    #[doc(hidden)]
    pub fn __set_spill_hook(&self, f: impl Fn() + Send + Sync + 'static) {
        *self.spill_hook.lock() = Some(Box::new(f));
    }

    /// Number of WAL fsyncs issued so far (test/diagnostic). With group commit
    /// this stays well below the number of committed transactions when commits
    /// are concurrent (one leader fsync covers a whole batch — spec §9.3).
    #[doc(hidden)]
    pub fn __wal_group_sync_count(&self) -> u64 {
        self.shared_wal.lock().group_sync_count()
    }

    /// Force the poisoned state (test-only) to verify the §9.3e fail-fast
    /// contract that an fsync error would trigger in production.
    #[doc(hidden)]
    pub fn __poison(&self) {
        self.poisoned
            .store(true, std::sync::atomic::Ordering::Relaxed);
    }

    /// Verify multi-table integrity (spec §16). For every live table this:
    /// authenticates the manifest; opens each `RunRef`'s file through
    /// [`RunReader`](crate::sorted_run::RunReader), which verifies the run footer
    /// checksum and — for encrypted DBs — the keyed run-metadata MAC; checks each
    /// run's physical row count against its `RunRef`; flags `RunRef`s whose file
    /// is missing (dangling) and `.sr` files on disk that no `RunRef` references
    /// (orphan); and verifies `flushed_epoch <= current_epoch`. Returns the list
    /// of issues found (empty = healthy). Orphans are `warning`-severity; all
    /// other findings are `error`-severity (so [`Self::doctor`] quarantines them).
    ///
    /// Cost: O(total run bytes) — the footer checksum is verified over each run's
    /// full body, so this is an integrity tool, not a hot path.
    pub fn check(&self) -> Vec<CheckIssue> {
        let mut issues = Vec::new();
        let cat = self.catalog.read();
        let manifest_meta_dek = crate::encryption::meta_dek_for(self.kek.as_deref());
        for entry in &cat.tables {
            if !matches!(entry.state, TableState::Live) {
                continue;
            }
            let tdir = self.root.join(TABLES_DIR).join(entry.table_id.to_string());
            let mut err = |sev: &str, desc: String| {
                issues.push(CheckIssue {
                    table_id: entry.table_id,
                    table_name: entry.name.clone(),
                    severity: sev.into(),
                    description: desc,
                });
            };
            let m = match crate::manifest::read(&tdir, manifest_meta_dek.as_ref()) {
                Ok(m) => m,
                Err(e) => {
                    err("error", format!("manifest read failed: {e}"));
                    continue;
                }
            };
            if m.flushed_epoch > m.current_epoch {
                err(
                    "error",
                    format!(
                        "flushed_epoch {} exceeds current_epoch {} (impossible)",
                        m.flushed_epoch, m.current_epoch
                    ),
                );
            }

            let runs_dir = tdir.join(crate::engine::RUNS_DIR);
            let mut referenced: std::collections::HashSet<u128> = std::collections::HashSet::new();
            for rr in &m.runs {
                referenced.insert(rr.run_id);
                let run_path = runs_dir.join(format!("r-{}.sr", rr.run_id));
                if !run_path.exists() {
                    err("error", format!("missing run file: r-{}.sr", rr.run_id));
                    continue;
                }
                match crate::sorted_run::RunReader::open(
                    &run_path,
                    entry.schema.clone(),
                    self.kek.clone(),
                ) {
                    Ok(reader) => {
                        if reader.row_count() as u64 != rr.row_count {
                            err(
                                "error",
                                format!(
                                    "run r-{} row count mismatch: manifest {} vs run {}",
                                    rr.run_id,
                                    rr.row_count,
                                    reader.row_count()
                                ),
                            );
                        }
                    }
                    Err(e) => {
                        err(
                            "error",
                            format!("run r-{} integrity check failed: {e}", rr.run_id),
                        );
                    }
                }
            }

            // Compaction-superseded runs awaiting retention-gated deletion are
            // tracked in `retiring`; their files are expected on disk, so they
            // are not orphans.
            for r in &m.retiring {
                referenced.insert(r.run_id);
            }

            // Orphan `.sr` files present on disk but absent from the manifest.
            if let Ok(rd) = std::fs::read_dir(&runs_dir) {
                for ent in rd.flatten() {
                    let p = ent.path();
                    if p.extension().and_then(|s| s.to_str()) != Some("sr") {
                        continue;
                    }
                    let run_id = p
                        .file_stem()
                        .and_then(|s| s.to_str())
                        .and_then(|s| s.strip_prefix("r-"))
                        .and_then(|s| s.parse::<u128>().ok());
                    if let Some(id) = run_id {
                        if !referenced.contains(&id) {
                            err(
                                "warning",
                                format!("orphan run file r-{id}.sr not referenced by the manifest"),
                            );
                        }
                    }
                }
            }
        }

        let external_names = cat
            .external_tables
            .iter()
            .map(|entry| entry.name.clone())
            .collect::<std::collections::HashSet<_>>();
        let vtab_dir = self.root.join(VTAB_DIR);
        if let Ok(entries) = std::fs::read_dir(&vtab_dir) {
            for entry in entries.flatten() {
                let name = entry.file_name();
                let Some(name) = name.to_str() else { continue };
                if !external_names.contains(name) {
                    issues.push(CheckIssue {
                        table_id: EXTERNAL_TABLE_ID,
                        table_name: name.to_string(),
                        severity: "warning".into(),
                        description: format!(
                            "orphan external table state entry {:?} not referenced by the catalog",
                            entry.path()
                        ),
                    });
                }
            }
        }

        // WAL retention / integrity invariant (spec §16): every on-disk WAL
        // segment must open (header magic + version, and the frame cipher must
        // be derivable for an encrypted WAL). A segment that won't open is
        // corrupt or truncated and would break crash recovery. `table_id` is
        // the reserved `WAL_TABLE_ID` sentinel (u64::MAX) so [`Self::doctor`]
        // never confuses a WAL issue with a real table.
        for (seg, msg) in self.shared_wal.lock().verify_segments() {
            issues.push(CheckIssue {
                table_id: WAL_TABLE_ID,
                table_name: "<wal>".into(),
                severity: "error".into(),
                description: format!("WAL segment seg-{seg:06}.wal failed integrity check: {msg}"),
            });
        }
        issues
    }

    /// Quarantine unreadable tables (spec §16). Moves corrupt table dirs to
    /// `_quarantine/<table_id>/`, marks them dropped in the catalog, and
    /// unmounts them from the live table map so the DB still opens.
    pub fn doctor(&self) -> Result<Vec<u64>> {
        // Hold the DDL lock for the whole operation to prevent concurrent
        // create_table/drop_table from racing the catalog/dir mutation.
        let _ddl = self.ddl_lock.lock();
        let issues = self.check();
        // A corrupt WAL segment is reported as an error but is NOT a table
        // problem — quarantining an innocent table cannot fix it (and the first
        // real table is id 0, so the WAL sentinel WAL_TABLE_ID = u64::MAX keeps
        // them disjoint). The admin must address WAL corruption manually.
        let bad_tables: std::collections::HashSet<u64> = issues
            .iter()
            .filter(|i| {
                i.severity == "error"
                    && i.table_id != WAL_TABLE_ID
                    && i.table_id != EXTERNAL_TABLE_ID
            })
            .map(|i| i.table_id)
            .collect();
        if bad_tables.is_empty() {
            return Ok(Vec::new());
        }

        let qdir = self.root.join("_quarantine");
        std::fs::create_dir_all(&qdir)?;
        let mut quarantined = Vec::new();
        for &table_id in &bad_tables {
            let tdir = self.root.join(TABLES_DIR).join(table_id.to_string());
            if tdir.exists() {
                let dest = qdir.join(table_id.to_string());
                std::fs::rename(&tdir, &dest)?;
            }
            {
                let mut cat = self.catalog.write();
                if let Some(entry) = cat.tables.iter_mut().find(|t| t.table_id == table_id) {
                    entry.state = TableState::Dropped {
                        at_epoch: self.epoch.visible().0,
                    };
                }
            }
            // Unmount the live handle so no further access reaches the moved dir.
            self.tables.write().remove(&table_id);
            quarantined.push(table_id);
        }
        catalog::write_atomic(&self.root, &self.catalog.read(), self.meta_dek.as_ref())?;
        Ok(quarantined)
    }

    /// The DB-wide KEK (if encrypted).
    #[allow(dead_code)]
    pub(crate) fn kek(&self) -> Option<&Arc<crate::encryption::Kek>> {
        self.kek.as_ref()
    }

    /// Shared epoch authority (used by the transaction layer in P2).
    #[allow(dead_code)]
    pub(crate) fn epoch_authority(&self) -> &Arc<EpochAuthority> {
        &self.epoch
    }

    /// Shared snapshot registry (used by GC in P3.6).
    #[allow(dead_code)]
    pub(crate) fn snapshots(&self) -> &Arc<SnapshotRegistry> {
        &self.snapshots
    }
}

fn external_state_dir(root: &Path, name: &str) -> PathBuf {
    root.join(VTAB_DIR).join(name)
}

fn filter_ignored_staging(
    staging: Vec<(u64, crate::txn::Staged)>,
    ignored_indices: &std::collections::BTreeSet<usize>,
) -> Vec<(u64, crate::txn::Staged)> {
    if ignored_indices.is_empty() {
        return staging;
    }
    staging
        .into_iter()
        .enumerate()
        .filter_map(|(idx, staged)| (!ignored_indices.contains(&idx)).then_some(staged))
        .collect()
}

fn external_state_file(root: &Path, name: &str) -> PathBuf {
    external_state_dir(root, name).join("state.json")
}

fn read_external_state_file(root: &Path, name: &str) -> Result<Vec<u8>> {
    let path = external_state_file(root, name);
    match std::fs::read(path) {
        Ok(bytes) => Ok(bytes),
        Err(e) if e.kind() == std::io::ErrorKind::NotFound => Ok(Vec::new()),
        Err(e) => Err(e.into()),
    }
}

fn current_external_state_bytes(
    root: &Path,
    external_states: &[(String, Vec<u8>)],
    name: &str,
) -> Result<Vec<u8>> {
    for (table, state) in external_states.iter().rev() {
        if table == name {
            return Ok(state.clone());
        }
    }
    read_external_state_file(root, name)
}

fn dedup_external_states(external_states: Vec<(String, Vec<u8>)>) -> Vec<(String, Vec<u8>)> {
    let mut out = external_states;
    dedup_external_states_in_place(&mut out);
    out
}

fn dedup_external_states_in_place(external_states: &mut Vec<(String, Vec<u8>)>) {
    let mut seen = std::collections::HashSet::new();
    let mut out = Vec::with_capacity(external_states.len());
    for (name, state) in std::mem::take(external_states).into_iter().rev() {
        if seen.insert(name.clone()) {
            out.push((name, state));
        }
    }
    out.reverse();
    *external_states = out;
}

fn prepare_external_state_file(
    root: &Path,
    name: &str,
    state: &[u8],
    txn_id: u64,
) -> Result<PathBuf> {
    let dir = external_state_dir(root, name);
    std::fs::create_dir_all(&dir)?;
    let pending = dir.join(format!("state.json.{txn_id}.tmp"));
    {
        let mut file = std::fs::File::create(&pending)?;
        file.write_all(state)?;
        file.sync_all()?;
    }
    Ok(pending)
}

fn publish_external_state_file(root: &Path, name: &str, pending: &Path) -> Result<()> {
    let path = external_state_file(root, name);
    std::fs::rename(pending, &path)?;
    if let Ok(dir) = std::fs::File::open(external_state_dir(root, name)) {
        let _ = dir.sync_all();
    }
    Ok(())
}

fn write_external_state_file(root: &Path, name: &str, state: &[u8]) -> Result<()> {
    let pending = prepare_external_state_file(root, name, state, 0)?;
    publish_external_state_file(root, name, &pending)
}

/// Two-pass, `flushed_epoch`-gated recovery of the shared WAL (spec §15).
///
/// Pass 1 scans every `TxnCommit` marker and records `txn_id → commit_epoch`
/// (the per-txn outcome; aborted / in-flight / torn-tail txns are absent). Pass
/// 2 applies each committed data record (Put/Delete) to its table at the commit
/// epoch, skipping records whose `commit_epoch <= table.flushed_epoch` (already
/// durable in a sorted run). Finally the shared epoch authority is raised to the
/// max committed epoch so the next commit continues monotonically.
fn recover_shared_wal(
    root: &Path,
    tables: &HashMap<u64, TableHandle>,
    epoch: &EpochAuthority,
    wal_dek: Option<&zeroize::Zeroizing<[u8; 32]>>,
) -> Result<()> {
    use crate::memtable::Row;
    use crate::rowid::RowId;
    use crate::wal::{Op, SharedWal};

    let records = SharedWal::replay_with_dek(root, wal_dek)?;

    // Pass 1: committed-txn outcomes + collect spilled-run info.
    let mut committed: HashMap<u64, u64> = HashMap::new();
    let mut spilled_to_link: Vec<(
        u64, /*txn_id*/
        u64, /*epoch*/
        Vec<crate::wal::AddedRun>,
    )> = Vec::new();
    for r in &records {
        if let Op::TxnCommit {
            epoch: ce,
            ref added_runs,
        } = r.op
        {
            committed.insert(r.txn_id, ce);
            if !added_runs.is_empty() {
                spilled_to_link.push((r.txn_id, ce, added_runs.clone()));
            }
        }
    }

    // Pass 2: stage data per table, gated by flushed_epoch.
    type TableStage = (Vec<Row>, Vec<(RowId, Epoch)>, Option<Epoch>, Epoch);
    let mut stage: HashMap<u64, TableStage> = HashMap::new();
    let mut max_epoch = epoch.visible().0;
    for r in records {
        let Some(&ce) = committed.get(&r.txn_id) else {
            continue; // aborted / in-flight — discard
        };
        let commit_epoch = Epoch(ce);
        max_epoch = max_epoch.max(ce);
        match r.op {
            Op::Put { table_id, rows } => {
                // Skip if this table already flushed past the commit epoch.
                let skip = tables
                    .get(&table_id)
                    .map(|h| h.lock().flushed_epoch() >= ce)
                    .unwrap_or(true);
                if skip {
                    continue;
                }
                let rows: Vec<Row> = match bincode::deserialize(&rows) {
                    Ok(v) => v,
                    Err(_) => continue,
                };
                // Re-stamp each row at the txn commit epoch (rows are pre-stamped
                // at pending_epoch which equals the commit epoch, but be robust).
                let rows: Vec<Row> = rows
                    .into_iter()
                    .map(|mut row| {
                        row.committed_epoch = commit_epoch;
                        row
                    })
                    .collect();
                let entry = stage
                    .entry(table_id)
                    .or_insert_with(|| (Vec::new(), Vec::new(), None, commit_epoch));
                entry.0.extend(rows);
                entry.3 = commit_epoch;
            }
            Op::Delete { table_id, row_ids } => {
                let skip = tables
                    .get(&table_id)
                    .map(|h| h.lock().flushed_epoch() >= ce)
                    .unwrap_or(true);
                if skip {
                    continue;
                }
                let dels = row_ids.into_iter().map(|rid| (rid, commit_epoch));
                let entry = stage
                    .entry(table_id)
                    .or_insert_with(|| (Vec::new(), Vec::new(), None, commit_epoch));
                entry.1.extend(dels);
                entry.3 = commit_epoch;
            }
            Op::TruncateTable { table_id } => {
                let skip = tables
                    .get(&table_id)
                    .map(|h| h.lock().flushed_epoch() >= ce)
                    .unwrap_or(true);
                if skip {
                    continue;
                }
                stage.insert(
                    table_id,
                    (Vec::new(), Vec::new(), Some(commit_epoch), commit_epoch),
                );
            }
            Op::ExternalTableState { name, state } => {
                write_external_state_file(root, &name, &state)?;
            }
            Op::Flush { .. } | Op::TxnCommit { .. } | Op::TxnAbort | Op::Ddl(_) => {}
        }
    }
    for (table_id, (rows, deletes, truncate_epoch, table_epoch)) in stage {
        let Some(handle) = tables.get(&table_id) else {
            continue;
        };
        let mut t = handle.lock();
        if let Some(epoch) = truncate_epoch {
            t.apply_truncate(epoch)?;
        }
        t.recover_apply(rows, deletes)?;
        if truncate_epoch.is_some() {
            let rows = t.visible_rows(Snapshot::at(Epoch(u64::MAX)))?;
            t.live_count = rows.len() as u64;
            t.persist_manifest(table_epoch)?;
        }
    }

    // Pass 3: link spilled runs from committed txns (spec §8.5). A crash
    // between TxnCommit sync and the publish phase leaves the run in
    // `_txn/<txn_id>/`. Move it to `_runs/` and add the RunRef.
    for (txn_id, ce, added_runs) in &spilled_to_link {
        for ar in added_runs {
            let Some(handle) = tables.get(&ar.table_id) else {
                continue;
            };
            let mut t = handle.lock();
            let dest = t.run_path(ar.run_id as u64);
            if !dest.exists() {
                let pending = root
                    .join(TABLES_DIR)
                    .join(ar.table_id.to_string())
                    .join("_txn")
                    .join(txn_id.to_string())
                    .join(format!("r-{}.sr", ar.run_id));
                if pending.exists() {
                    if let Some(parent) = pending.parent() {
                        std::fs::rename(&pending, &dest)?;
                        let _ = std::fs::remove_dir_all(parent);
                    }
                }
            }
            // Only link a run whose file is actually present, and never re-link
            // one the publish phase already persisted into the manifest (which is
            // the common clean-reopen case, since the `TxnCommit` lives in the WAL
            // until segment GC). `recover_spilled_run` is idempotent + reconciles
            // `live_count`/indexes only when the run is genuinely new.
            if t.run_path(ar.run_id as u64).exists() {
                t.recover_spilled_run(crate::manifest::RunRef {
                    run_id: ar.run_id,
                    level: ar.level,
                    epoch_created: *ce,
                    row_count: ar.row_count,
                });
            }
        }
    }

    epoch.advance_recovered(Epoch(max_epoch));
    Ok(())
}

fn validate_condition_columns(condition: &ProcedureCondition, schema: &Schema) -> Result<()> {
    match condition {
        ProcedureCondition::Pk { .. } => {
            if schema.primary_key().is_none() {
                return Err(MongrelError::InvalidArgument(
                    "procedure condition Pk references a table without a primary key".into(),
                ));
            }
        }
        ProcedureCondition::BitmapEq { column_id, .. }
        | ProcedureCondition::BitmapIn { column_id, .. }
        | ProcedureCondition::Range { column_id, .. }
        | ProcedureCondition::RangeF64 { column_id, .. }
        | ProcedureCondition::IsNull { column_id }
        | ProcedureCondition::IsNotNull { column_id }
        | ProcedureCondition::FmContains { column_id, .. } => {
            validate_column_id(*column_id, schema)?;
        }
    }
    Ok(())
}

fn bind_procedure_args(
    procedure: &StoredProcedure,
    mut args: HashMap<String, crate::Value>,
) -> Result<HashMap<String, crate::Value>> {
    let mut out = HashMap::new();
    for param in &procedure.params {
        let value = match args.remove(&param.name) {
            Some(value) => value,
            None => param.default.clone().ok_or_else(|| {
                MongrelError::InvalidArgument(format!(
                    "missing required procedure parameter {:?}",
                    param.name
                ))
            })?,
        };
        if !param.nullable && matches!(value, crate::Value::Null) {
            return Err(MongrelError::InvalidArgument(format!(
                "procedure parameter {:?} must not be NULL",
                param.name
            )));
        }
        if !matches!(value, crate::Value::Null) && !value_matches_type(&value, param.ty) {
            return Err(MongrelError::InvalidArgument(format!(
                "procedure parameter {:?} has wrong type",
                param.name
            )));
        }
        out.insert(param.name.clone(), value);
    }
    if let Some(extra) = args.keys().next() {
        return Err(MongrelError::InvalidArgument(format!(
            "unknown procedure parameter {extra:?}"
        )));
    }
    Ok(out)
}

fn value_matches_type(value: &crate::Value, ty: crate::TypeId) -> bool {
    matches!(
        (value, ty),
        (crate::Value::Bool(_), crate::TypeId::Bool)
            | (crate::Value::Int64(_), crate::TypeId::Int8)
            | (crate::Value::Int64(_), crate::TypeId::Int16)
            | (crate::Value::Int64(_), crate::TypeId::Int32)
            | (crate::Value::Int64(_), crate::TypeId::Int64)
            | (crate::Value::Int64(_), crate::TypeId::UInt8)
            | (crate::Value::Int64(_), crate::TypeId::UInt16)
            | (crate::Value::Int64(_), crate::TypeId::UInt32)
            | (crate::Value::Int64(_), crate::TypeId::UInt64)
            | (crate::Value::Int64(_), crate::TypeId::TimestampNanos)
            | (crate::Value::Int64(_), crate::TypeId::Date32)
            | (crate::Value::Float64(_), crate::TypeId::Float32)
            | (crate::Value::Float64(_), crate::TypeId::Float64)
            | (crate::Value::Bytes(_), crate::TypeId::Bytes)
            | (crate::Value::Embedding(_), crate::TypeId::Embedding { .. })
    )
}

fn eval_cells(
    cells: &[crate::procedure::ProcedureCell],
    args: &HashMap<String, crate::Value>,
    outputs: &HashMap<String, ProcedureCallOutput>,
) -> Result<Vec<(u16, crate::Value)>> {
    cells
        .iter()
        .map(|cell| Ok((cell.column_id, eval_value(&cell.value, args, outputs)?)))
        .collect()
}

fn eval_condition(
    condition: &ProcedureCondition,
    args: &HashMap<String, crate::Value>,
    outputs: &HashMap<String, ProcedureCallOutput>,
) -> Result<crate::Condition> {
    Ok(match condition {
        ProcedureCondition::Pk { value } => {
            crate::Condition::Pk(eval_value(value, args, outputs)?.encode_key())
        }
        ProcedureCondition::BitmapEq { column_id, value } => crate::Condition::BitmapEq {
            column_id: *column_id,
            value: eval_value(value, args, outputs)?.encode_key(),
        },
        ProcedureCondition::BitmapIn { column_id, values } => crate::Condition::BitmapIn {
            column_id: *column_id,
            values: values
                .iter()
                .map(|value| Ok(eval_value(value, args, outputs)?.encode_key()))
                .collect::<Result<Vec<_>>>()?,
        },
        ProcedureCondition::Range { column_id, lo, hi } => crate::Condition::Range {
            column_id: *column_id,
            lo: expect_i64(eval_value(lo, args, outputs)?)?,
            hi: expect_i64(eval_value(hi, args, outputs)?)?,
        },
        ProcedureCondition::RangeF64 {
            column_id,
            lo,
            lo_inclusive,
            hi,
            hi_inclusive,
        } => crate::Condition::RangeF64 {
            column_id: *column_id,
            lo: expect_f64(eval_value(lo, args, outputs)?)?,
            lo_inclusive: *lo_inclusive,
            hi: expect_f64(eval_value(hi, args, outputs)?)?,
            hi_inclusive: *hi_inclusive,
        },
        ProcedureCondition::IsNull { column_id } => crate::Condition::IsNull {
            column_id: *column_id,
        },
        ProcedureCondition::IsNotNull { column_id } => crate::Condition::IsNotNull {
            column_id: *column_id,
        },
        ProcedureCondition::FmContains { column_id, pattern } => crate::Condition::FmContains {
            column_id: *column_id,
            pattern: expect_bytes(eval_value(pattern, args, outputs)?)?,
        },
    })
}

fn eval_value(
    value: &ProcedureValue,
    args: &HashMap<String, crate::Value>,
    outputs: &HashMap<String, ProcedureCallOutput>,
) -> Result<crate::Value> {
    match value {
        ProcedureValue::Literal(value) => Ok(value.clone()),
        ProcedureValue::Param(name) => args.get(name).cloned().ok_or_else(|| {
            MongrelError::InvalidArgument(format!("unknown procedure parameter {name:?}"))
        }),
        ProcedureValue::StepScalar(id) => match outputs.get(id) {
            Some(ProcedureCallOutput::Scalar(value)) => Ok(value.clone()),
            _ => Err(MongrelError::InvalidArgument(format!(
                "procedure step {id:?} did not return a scalar"
            ))),
        },
        ProcedureValue::StepRows(_) | ProcedureValue::StepRow(_) => {
            Err(MongrelError::InvalidArgument(
                "row-valued procedure reference cannot be used as a scalar".into(),
            ))
        }
        ProcedureValue::Object(_) | ProcedureValue::Array(_) => Err(MongrelError::InvalidArgument(
            "structured procedure value cannot be used as a scalar cell".into(),
        )),
    }
}

fn eval_return_output(
    value: &ProcedureValue,
    args: &HashMap<String, crate::Value>,
    outputs: &HashMap<String, ProcedureCallOutput>,
) -> Result<ProcedureCallOutput> {
    match value {
        ProcedureValue::Literal(value) => Ok(ProcedureCallOutput::Scalar(value.clone())),
        ProcedureValue::Param(name) => Ok(ProcedureCallOutput::Scalar(
            args.get(name).cloned().ok_or_else(|| {
                MongrelError::InvalidArgument(format!("unknown procedure parameter {name:?}"))
            })?,
        )),
        ProcedureValue::StepRows(id)
        | ProcedureValue::StepRow(id)
        | ProcedureValue::StepScalar(id) => outputs.get(id).cloned().ok_or_else(|| {
            MongrelError::InvalidArgument(format!("unknown procedure step output {id:?}"))
        }),
        ProcedureValue::Object(fields) => {
            let mut out = Vec::with_capacity(fields.len());
            for (name, value) in fields {
                out.push((name.clone(), eval_return_output(value, args, outputs)?));
            }
            Ok(ProcedureCallOutput::Object(out))
        }
        ProcedureValue::Array(values) => {
            let mut out = Vec::with_capacity(values.len());
            for value in values {
                out.push(eval_return_output(value, args, outputs)?);
            }
            Ok(ProcedureCallOutput::Array(out))
        }
    }
}

fn expect_i64(value: crate::Value) -> Result<i64> {
    match value {
        crate::Value::Int64(value) => Ok(value),
        _ => Err(MongrelError::InvalidArgument(
            "procedure value must be Int64".into(),
        )),
    }
}

fn expect_f64(value: crate::Value) -> Result<f64> {
    match value {
        crate::Value::Float64(value) => Ok(value),
        _ => Err(MongrelError::InvalidArgument(
            "procedure value must be Float64".into(),
        )),
    }
}

fn expect_bytes(value: crate::Value) -> Result<Vec<u8>> {
    match value {
        crate::Value::Bytes(value) => Ok(value),
        _ => Err(MongrelError::InvalidArgument(
            "procedure value must be Bytes".into(),
        )),
    }
}

fn validate_column_id(column_id: u16, schema: &Schema) -> Result<()> {
    if schema.columns.iter().any(|c| c.id == column_id) {
        Ok(())
    } else {
        Err(MongrelError::InvalidArgument(format!(
            "unknown column id {column_id}"
        )))
    }
}

fn trigger_matches_event(
    trigger: &StoredTrigger,
    event: &WriteEvent,
    cat: &Catalog,
) -> Result<bool> {
    if trigger.event != event.kind {
        return Ok(false);
    }
    let TriggerTarget::Table(target) = &trigger.target else {
        return Ok(false);
    };
    if target != &event.table {
        return Ok(false);
    }
    if trigger.event == TriggerEvent::Update && !trigger.update_of.is_empty() {
        let schema = &cat
            .live(target)
            .ok_or_else(|| {
                MongrelError::InvalidArgument(format!(
                    "trigger {:?} references unknown table {target:?}",
                    trigger.name
                ))
            })?
            .schema;
        let mut watched = Vec::with_capacity(trigger.update_of.len());
        for name in &trigger.update_of {
            let col = schema.column(name).ok_or_else(|| {
                MongrelError::InvalidArgument(format!(
                    "trigger {:?} references unknown UPDATE OF column {name:?}",
                    trigger.name
                ))
            })?;
            watched.push(col.id);
        }
        if !event
            .changed_columns
            .iter()
            .any(|column_id| watched.contains(column_id))
        {
            return Ok(false);
        }
    }
    Ok(true)
}

fn changed_columns(old: Option<&TriggerRowImage>, new: Option<&TriggerRowImage>) -> Vec<u16> {
    let mut ids = std::collections::BTreeSet::new();
    if let Some(old) = old {
        ids.extend(old.columns.keys().copied());
    }
    if let Some(new) = new {
        ids.extend(new.columns.keys().copied());
    }
    ids.into_iter()
        .filter(|id| {
            old.and_then(|row| row.columns.get(id)) != new.and_then(|row| row.columns.get(id))
        })
        .collect()
}

fn eval_trigger_cells(
    cells: &[crate::trigger::TriggerCell],
    event: &WriteEvent,
) -> Result<Vec<(u16, Value)>> {
    cells
        .iter()
        .map(|cell| Ok((cell.column_id, eval_trigger_value(&cell.value, event)?)))
        .collect()
}

fn eval_trigger_expr(expr: &TriggerExpr, event: &WriteEvent) -> Result<bool> {
    match expr {
        TriggerExpr::Value(value) => match eval_trigger_value(value, event)? {
            Value::Bool(value) => Ok(value),
            Value::Null => Ok(false),
            other => Err(MongrelError::InvalidArgument(format!(
                "trigger WHEN value must be boolean, got {other:?}"
            ))),
        },
        TriggerExpr::Eq { left, right } => Ok(values_equal(
            &eval_trigger_value(left, event)?,
            &eval_trigger_value(right, event)?,
        )),
        TriggerExpr::NotEq { left, right } => Ok(!values_equal(
            &eval_trigger_value(left, event)?,
            &eval_trigger_value(right, event)?,
        )),
        TriggerExpr::IsNull(value) => Ok(matches!(eval_trigger_value(value, event)?, Value::Null)),
        TriggerExpr::IsNotNull(value) => {
            Ok(!matches!(eval_trigger_value(value, event)?, Value::Null))
        }
    }
}

fn eval_trigger_value(value: &TriggerValue, event: &WriteEvent) -> Result<Value> {
    match value {
        TriggerValue::Literal(value) => Ok(value.clone()),
        TriggerValue::NewColumn(column_id) => event
            .new
            .as_ref()
            .and_then(|row| row.columns.get(column_id))
            .cloned()
            .ok_or_else(|| MongrelError::InvalidArgument("NEW column is not available".into())),
        TriggerValue::OldColumn(column_id) => event
            .old
            .as_ref()
            .and_then(|row| row.columns.get(column_id))
            .cloned()
            .ok_or_else(|| MongrelError::InvalidArgument("OLD column is not available".into())),
    }
}

fn values_equal(left: &Value, right: &Value) -> bool {
    match (left, right) {
        (Value::Null, Value::Null) => true,
        (Value::Bool(a), Value::Bool(b)) => a == b,
        (Value::Int64(a), Value::Int64(b)) => a == b,
        (Value::Float64(a), Value::Float64(b)) => a.to_bits() == b.to_bits(),
        (Value::Bytes(a), Value::Bytes(b)) => a == b,
        (Value::Embedding(a), Value::Embedding(b)) => {
            a.len() == b.len()
                && a.iter()
                    .zip(b.iter())
                    .all(|(a, b)| a.to_bits() == b.to_bits())
        }
        _ => false,
    }
}

fn trigger_message(value: Value) -> String {
    match value {
        Value::Null => "NULL".into(),
        Value::Bool(value) => value.to_string(),
        Value::Int64(value) => value.to_string(),
        Value::Float64(value) => value.to_string(),
        Value::Bytes(value) => String::from_utf8_lossy(&value).into_owned(),
        Value::Embedding(value) => format!("{value:?}"),
    }
}

fn validate_trigger_step(
    step: &TriggerStep,
    cat: &Catalog,
    target_schema: &Schema,
    event: TriggerEvent,
) -> Result<()> {
    match step {
        TriggerStep::SetNew { cells } => {
            if event == TriggerEvent::Delete {
                return Err(MongrelError::InvalidArgument(
                    "SetNew trigger step is not valid for DELETE triggers".into(),
                ));
            }
            for cell in cells {
                validate_column_id(cell.column_id, target_schema)?;
                validate_trigger_value(&cell.value, target_schema, event)?;
            }
        }
        TriggerStep::Insert { table, cells } => {
            let schema = trigger_write_schema(cat, table, "insert")?;
            for cell in cells {
                validate_column_id(cell.column_id, schema)?;
                validate_trigger_value(&cell.value, target_schema, event)?;
            }
        }
        TriggerStep::UpdateByPk { table, pk, cells } => {
            let schema = trigger_write_schema(cat, table, "update")?;
            if schema.primary_key().is_none() {
                return Err(MongrelError::InvalidArgument(format!(
                    "trigger update_by_pk references table {table:?} without a primary key"
                )));
            }
            validate_trigger_value(pk, target_schema, event)?;
            for cell in cells {
                validate_column_id(cell.column_id, schema)?;
                validate_trigger_value(&cell.value, target_schema, event)?;
            }
        }
        TriggerStep::DeleteByPk { table, pk } => {
            let schema = trigger_write_schema(cat, table, "delete")?;
            if schema.primary_key().is_none() {
                return Err(MongrelError::InvalidArgument(format!(
                    "trigger delete_by_pk references table {table:?} without a primary key"
                )));
            }
            validate_trigger_value(pk, target_schema, event)?;
        }
        TriggerStep::Select {
            table, conditions, ..
        } => {
            let schema = trigger_read_schema(cat, table)?;
            for condition in conditions {
                validate_trigger_condition(condition, schema, target_schema, event)?;
            }
        }
        TriggerStep::Raise { message, .. } => {
            validate_trigger_value(message, target_schema, event)?
        }
    }
    Ok(())
}

fn trigger_write_schema<'a>(cat: &'a Catalog, table: &str, op: &str) -> Result<&'a Schema> {
    if let Some(entry) = cat.live(table) {
        return Ok(&entry.schema);
    }
    if let Some(entry) = cat.external_tables.iter().find(|entry| entry.name == table) {
        let allowed = match op {
            "insert" => entry.capabilities.writable || entry.capabilities.insert_only,
            "update" | "delete" => entry.capabilities.writable,
            _ => false,
        };
        if !allowed {
            return Err(MongrelError::InvalidArgument(format!(
                "trigger {op} references external table {table:?}, but module {:?} is not writable for that operation",
                entry.module
            )));
        }
        if !entry.capabilities.transaction_safe {
            return Err(MongrelError::InvalidArgument(format!(
                "trigger {op} references external table {table:?}, but module {:?} is not transaction-safe",
                entry.module
            )));
        }
        return Ok(&entry.declared_schema);
    }
    Err(MongrelError::InvalidArgument(format!(
        "trigger references unknown table {table:?}"
    )))
}

fn trigger_read_schema<'a>(cat: &'a Catalog, table: &str) -> Result<&'a Schema> {
    if let Some(entry) = cat.live(table) {
        return Ok(&entry.schema);
    }
    if let Some(entry) = cat.external_tables.iter().find(|entry| entry.name == table) {
        if entry.capabilities.trigger_safe {
            return Ok(&entry.declared_schema);
        }
        return Err(MongrelError::InvalidArgument(format!(
            "trigger reads external table {table:?}, but module {:?} is not trigger-safe",
            entry.module
        )));
    }
    Err(MongrelError::InvalidArgument(format!(
        "trigger references unknown table {table:?}"
    )))
}

fn validate_trigger_condition(
    condition: &TriggerCondition,
    schema: &Schema,
    target_schema: &Schema,
    event: TriggerEvent,
) -> Result<()> {
    match condition {
        TriggerCondition::Pk { value } => {
            if schema.primary_key().is_none() {
                return Err(MongrelError::InvalidArgument(
                    "trigger condition Pk references a table without a primary key".into(),
                ));
            }
            validate_trigger_value(value, target_schema, event)
        }
        TriggerCondition::Eq { column_id, value } => {
            validate_column_id(*column_id, schema)?;
            validate_trigger_value(value, target_schema, event)
        }
        TriggerCondition::IsNull { column_id } | TriggerCondition::IsNotNull { column_id } => {
            validate_column_id(*column_id, schema)
        }
    }
}

fn validate_trigger_expr(expr: &TriggerExpr, schema: &Schema, event: TriggerEvent) -> Result<()> {
    match expr {
        TriggerExpr::Value(value) | TriggerExpr::IsNull(value) | TriggerExpr::IsNotNull(value) => {
            validate_trigger_value(value, schema, event)
        }
        TriggerExpr::Eq { left, right } | TriggerExpr::NotEq { left, right } => {
            validate_trigger_value(left, schema, event)?;
            validate_trigger_value(right, schema, event)
        }
    }
}

fn validate_trigger_value(
    value: &TriggerValue,
    schema: &Schema,
    event: TriggerEvent,
) -> Result<()> {
    match value {
        TriggerValue::Literal(_) => Ok(()),
        TriggerValue::NewColumn(id) => {
            if event == TriggerEvent::Delete {
                return Err(MongrelError::InvalidArgument(
                    "DELETE triggers cannot reference NEW".into(),
                ));
            }
            validate_column_id(*id, schema)
        }
        TriggerValue::OldColumn(id) => {
            if event == TriggerEvent::Insert {
                return Err(MongrelError::InvalidArgument(
                    "INSERT triggers cannot reference OLD".into(),
                ));
            }
            validate_column_id(*id, schema)
        }
    }
}

/// Replay committed `Op::Ddl` records from the shared WAL into the catalog
/// (spec §15, review fix #16). A crash between WAL group-sync and the catalog
/// checkpoint leaves DDL durable in the WAL but absent from the on-disk
/// catalog. This pass closes that window by reconstructing missing entries
/// (and marking committed drops) before tables are mounted.
fn recover_ddl_from_wal(
    root: &Path,
    cat: &mut Catalog,
    meta_dek: Option<&[u8; META_DEK_LEN]>,
    wal_dek: Option<&zeroize::Zeroizing<[u8; 32]>>,
) -> Result<()> {
    use crate::wal::{DdlOp, Op, SharedWal};

    let records = match SharedWal::replay_with_dek(root, wal_dek) {
        Ok(r) => r,
        Err(_) => return Ok(()),
    };

    let mut committed: HashMap<u64, u64> = HashMap::new();
    for r in &records {
        if let Op::TxnCommit { epoch: ce, .. } = r.op {
            committed.insert(r.txn_id, ce);
        }
    }

    let mut changed = false;
    for r in records {
        let Some(&ce) = committed.get(&r.txn_id) else {
            continue;
        };
        match r.op {
            Op::Ddl(DdlOp::CreateTable {
                table_id,
                ref name,
                ref schema_json,
            }) => {
                if cat.tables.iter().any(|t| t.table_id == table_id) {
                    continue;
                }
                let schema = DdlOp::decode_schema(schema_json)?;
                let tdir = root.join(TABLES_DIR).join(table_id.to_string());
                if !tdir.exists() {
                    std::fs::create_dir_all(tdir.join(crate::engine::WAL_DIR))?;
                    std::fs::create_dir_all(tdir.join(crate::engine::RUNS_DIR))?;
                    crate::engine::write_schema(&tdir, &schema)?;
                    // The DB-wide meta DEK is also the per-table manifest meta
                    // DEK (both derive from the KEK via `derive_meta_key`), so a
                    // reconstructed manifest must be sealed with it — otherwise
                    // the follow-up `Table::open_in` cannot authenticate it on an
                    // encrypted DB and the table becomes permanently unopenable.
                    let mut m = crate::manifest::Manifest::new(table_id, schema.schema_id);
                    crate::manifest::write_atomic(&tdir, &mut m, meta_dek)?;
                }
                cat.tables.push(CatalogEntry {
                    table_id,
                    name: name.clone(),
                    schema,
                    state: TableState::Live,
                    created_epoch: ce,
                });
                cat.next_table_id = cat.next_table_id.max(table_id + 1);
                changed = true;
            }
            Op::Ddl(DdlOp::DropTable { table_id }) => {
                if let Some(entry) = cat.tables.iter_mut().find(|t| t.table_id == table_id) {
                    if matches!(entry.state, TableState::Live) {
                        entry.state = TableState::Dropped { at_epoch: ce };
                        changed = true;
                    }
                }
            }
            Op::Ddl(DdlOp::RenameTable {
                table_id,
                ref new_name,
            }) => {
                if let Some(entry) = cat.tables.iter_mut().find(|t| t.table_id == table_id) {
                    if entry.name != *new_name {
                        entry.name = new_name.clone();
                        changed = true;
                    }
                }
                // If the entry is absent, its CreateTable was already
                // checkpointed carrying the post-rename name, so there is
                // nothing to apply — a no-op, not an error.
            }
            Op::Ddl(DdlOp::AlterTable {
                table_id,
                ref column_json,
            }) => {
                let column = DdlOp::decode_column(column_json)?;
                if let Some(entry) = cat.tables.iter_mut().find(|t| t.table_id == table_id) {
                    if apply_recovered_column_def(&mut entry.schema, column) {
                        let tdir = root.join(TABLES_DIR).join(table_id.to_string());
                        if tdir.exists() {
                            crate::engine::write_schema(&tdir, &entry.schema)?;
                        }
                        changed = true;
                    }
                }
            }
            _ => {}
        }
    }

    if changed {
        catalog::write_atomic(root, cat, meta_dek)?;
    }
    Ok(())
}

fn apply_recovered_column_def(schema: &mut Schema, column: ColumnDef) -> bool {
    match schema.columns.iter_mut().find(|c| c.id == column.id) {
        Some(existing) if *existing == column => false,
        Some(existing) => {
            *existing = column;
            schema.schema_id = schema.schema_id.saturating_add(1);
            true
        }
        None => {
            schema.columns.push(column);
            schema.schema_id = schema.schema_id.saturating_add(1);
            true
        }
    }
}

/// Sweep stale `_txn/<txn_id>/` dirs from every table (spec §8.5, review fix
/// #14). These dirs hold pending uniform-epoch runs from large transactions
/// that were aborted or crashed before commit. On open, all such dirs are safe
/// to remove — committed txns moved their runs to `_runs/` at publish time.
fn sweep_pending_txn_dirs(root: &Path, cat: &Catalog) {
    for entry in &cat.tables {
        let txn_dir = root
            .join(TABLES_DIR)
            .join(entry.table_id.to_string())
            .join("_txn");
        if txn_dir.exists() {
            let _ = std::fs::remove_dir_all(&txn_dir);
        }
    }
}