topodb 0.0.18

use crate::counters::AccessStats;
use crate::error::TopoError;
use crate::feed::ChangeEvent;
use crate::ids::{EdgeId, NodeId, Scope, ScopeSet};
use crate::index::IndexSpec;
use crate::op::Op;
use crate::state::NodeRecord;
use crate::storage::{AppliedBatch, Storage};
use crossbeam_channel::{bounded, Receiver, Sender};
use std::path::Path;
use std::sync::atomic::{AtomicBool, Ordering};
use std::sync::{Arc, Mutex};
use std::time::{SystemTime, UNIX_EPOCH};

/// Tuning knobs for [`Db::open_with_options`]. Additive: every field
/// defaults to `None`, under which redb's own default is used, so a fresh
/// `DbOptions::default()` behaves identically to `Db::open`/`Db::open_with`.
#[derive(Debug, Clone, Copy, Default)]
pub struct DbOptions {
    /// Threaded straight to `redb::Builder::set_cache_size`. `None` leaves
    /// redb's own default (1 GiB, split 90/10 read/write) in place.
    pub cache_size_bytes: Option<usize>,
    /// HNSW graph-maintenance tuning (F8): build threshold, `m`/`m0`,
    /// `ef_construction`, rebuild ratio. `None` resolves to
    /// `HnswParams::default()` at open (`Storage::open_with_options`),
    /// validated before any IO — an invalid override rejects the whole
    /// open rather than silently falling back. `pub(crate)` type kept
    /// crate-private; re-exported as `topodb::HnswParams` from `lib.rs` so
    /// tests (and any future host tuning this at open time) can construct
    /// one without reaching into the crate.
    pub hnsw_params: Option<crate::hnsw::HnswParams>,
}

/// One queued `Job::Apply`'s ops paired with the reply channel its submitter
/// blocks on. The applier's group-drain path (see the `Job::Apply` arm and
/// `apply_group`) collects these directly, having already peeled `at` off
/// (group batches all share one wall-clock `now` — see the arm's doc
/// comment) and dropped the `Job` wrapper.
type ApplyJob = (Vec<Op>, Sender<Result<AppliedBatch, TopoError>>);

/// A unit of work for the single applier thread. Both variants carry a reply
/// channel so the submitting thread blocks until the applier has finished —
/// and, crucially, so the *applier* remains the sole writer of storage.
enum Job {
    Apply {
        ops: Vec<Op>,
        at: Option<i64>,
        reply: Sender<Result<AppliedBatch, TopoError>>,
    },
    Rebuild {
        reply: Sender<Result<(), TopoError>>,
    },
    /// Fire-and-forget batch of access-counter bumps folded into COUNTERS by
    /// the applier. No reply channel: bumps are auxiliary telemetry, so the
    /// applier logs nothing, broadcasts nothing to the change feed, and never
    /// acknowledges. Enqueued only by the bumper thread (see `open_with`).
    BumpCounters { bumps: Vec<(NodeId, u64, i64)> },
    /// Compacts the op log through `keep_from` on the applier thread (the sole
    /// redb writer). Broadcasts nothing — compaction touches no NODES/EDGES
    /// state and emits no change events — and replies the storage result so the
    /// caller blocks until the trim has committed.
    Compact {
        keep_from: u64,
        reply: Sender<Result<(), TopoError>>,
    },
    /// Writes an arbitrary key/value pair into the `META` table on the
    /// applier thread (the sole redb writer). Broadcasts nothing — `META`
    /// writes are out-of-band from the scoped node/edge graph and its
    /// change feed — and replies the storage result so the caller blocks
    /// until the write has committed.
    Meta {
        key: String,
        value: Vec<u8>,
        reply: Sender<Result<(), TopoError>>,
    },
}

/// A handle to an open database. Cloning shares the same underlying storage
/// and applier thread — `Db` is `Send + Sync + Clone`. All writes funnel
/// through a single applier thread (via `submit`/`submit_at`), so batches
/// serialize deterministically even under concurrent callers.
#[derive(Clone)]
pub struct Db {
    inner: Arc<Inner>,
}

// Manual (not derived) so this doesn't force `Debug` on every field of
// `Inner` (several of which — `Storage` among them — don't derive it and
// aren't otherwise worth adding it to). `Db` itself carries no useful
// state to print; this exists so `Result<Db, TopoError>` — e.g. in a test's
// `panic!("{other:?}")` fallback arm — is formattable.
impl std::fmt::Debug for Db {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        f.debug_struct("Db").finish_non_exhaustive()
    }
}

struct Inner {
    // Read directly by `rebuild_state_from_ops`/`debug_dump_*`/every scoped
    // read (`node`, `nodes_by_label`, `traverse`, ...), and kept alive here
    // so the underlying `redb::Database`'s file handle stays open for the
    // lifetime of the `Db`. The read model is disk-resident: there is no
    // separate in-memory snapshot to keep in step with it (see FORMAT.md /
    // the W5 plan task for the snapshot layer this replaced).
    storage: Arc<Storage>,
    // `Sender` half of the job channel. Wrapped in `Option` so `Drop` can
    // `take()` it and actually drop it *before* joining the applier thread —
    // otherwise the applier's `rx.recv()` loop would never see the channel
    // close and `join()` would hang forever.
    tx: Mutex<Option<Sender<Job>>>,
    applier: Mutex<Option<std::thread::JoinHandle<()>>>,
    // `Sender` half of the bump channel feeding the bumper thread. Reads
    // `try_send` `(NodeId, ts)` pairs here; the bumper accumulates and flushes
    // them as batched `Job::BumpCounters`. Wrapped in `Option` so `Drop` can
    // take+drop it *first* (before joining the bumper) — closing this channel
    // is what makes the bumper's `recv_timeout` loop see `Disconnected`, do its
    // final flush, and exit. See `Drop for Inner` for the full ordering.
    bump_tx: Mutex<Option<Sender<(NodeId, i64)>>>,
    bumper: Mutex<Option<std::thread::JoinHandle<()>>>,
    // Change-feed subscriber registry: the bounded `Sender` half of every
    // live `subscribe` channel. The applier clones this `Arc` at spawn and is
    // the *only* broadcaster; `subscribe` pushes a new sender under the mutex.
    // Both hold the lock only briefly (a push, or one non-blocking drain per
    // batch), and nothing else locks it — so it introduces no lock-ordering
    // hazard against the `tx`/`applier` mutexes. Held behind its own `Arc`
    // (not captured via `Inner`) for the same reason as `storage`: the
    // applier must never hold a strong ref back to `Inner`, or `Drop` would
    // deadlock.
    subs: Arc<Mutex<Vec<Sender<ChangeEvent>>>>,
    // Debug-only instrumentation (F8 Task 4): flips to `true` when
    // `vector_store::search_scan`'s per-scope routing took the built-graph
    // branch (`hnsw::search`) and `false` when it took the brute-force scan
    // branch, overwritten by every scope a query touches (so it reflects
    // only the LAST scope's routing decision, not the whole query). Lives on
    // `Inner` (not `Storage`, which this task's brief scopes out of) — never
    // read from replay state or any persisted table. Surfaced via
    // `Db::debug_last_search_used_graph`/`Db::debug_atomic` so tests can pin
    // that the graph path is genuinely live rather than vacuously agreeing
    // with brute force by still silently scanning.
    debug_last_search_used_graph: AtomicBool,
}

/// Which adjacency table to scan: OUT_ADJ (edges FROM a node) or IN_ADJ
/// (edges TO a node). Used by the private `scan_adjacency` helper.
enum EdgeDirection {
    Out,
    In,
}

/// Temporal gate for adjacency scanning: either `open_only` with a time axis
/// (Valid gates on entry.valid_to before fetch, Recorded gates on
/// rec.superseded_at after fetch), or an Allen predicate over the edge's
/// valid interval (gates only on entry fields, pre-fetch). Used by the
/// private `scan_adjacency` helper to parameterize the temporal logic shared
/// by `edges_from`, `edges_to`, `edges_from_interval`, and `edges_to_interval`.
enum EdgeGate<'a> {
    OpenOnly {
        open_only: bool,
        axis: crate::read::TimeAxis,
    },
    Interval(&'a crate::read::ValidInterval),
}

impl<'a> EdgeGate<'a> {
    /// Returns true if the entry should be skipped (filtered out) by this
    /// gate's pre-fetch logic.
    fn should_skip_entry(&self, entry: &crate::adj::AdjEntryDisk) -> bool {
        match self {
            EdgeGate::OpenOnly { open_only, axis } => {
                if !open_only {
                    return false;
                }
                match axis {
                    // Valid axis: gate on the adjacency entry before the fetch
                    // (hot path, byte-unchanged).
                    crate::read::TimeAxis::Valid => entry.valid_to.is_some(),
                    // Recorded axis: the entry doesn't carry `superseded_at`,
                    // so defer the gate to the fetched record.
                    crate::read::TimeAxis::Recorded => false,
                }
            }
            // Interval gate: check the Allen predicate on entry fields.
            EdgeGate::Interval(interval) => !interval.matches(entry.valid_from, entry.valid_to),
        }
    }

    /// Returns true if the record should be skipped (filtered out) by this
    /// gate's post-fetch logic.
    fn should_skip_record(&self, record: &crate::state::EdgeRecord) -> bool {
        match self {
            EdgeGate::OpenOnly { open_only, axis } => {
                if !open_only {
                    return false;
                }
                match axis {
                    // Valid axis: already filtered at entry level.
                    crate::read::TimeAxis::Valid => false,
                    // Recorded axis: gate on the fetched record's
                    // `superseded_at` (nearly free here since every candidate
                    // already needs a full-record fetch for the result).
                    crate::read::TimeAxis::Recorded => record.superseded_at.is_some(),
                }
            }
            // Interval gate: no post-fetch logic.
            EdgeGate::Interval(_) => false,
        }
    }
}

impl Db {
    /// Opens (creating if necessary) the database at `path` and starts its
    /// single applier thread. `submit`/`submit_at` route through this thread;
    /// it is the only place wall-clock time is read (`submit` uses
    /// `SystemTime::now`; `submit_at` is the deterministic test/backdate
    /// seam). Delegates to `open_with` with a default (empty) `IndexSpec`.
    pub fn open(path: impl AsRef<Path>) -> Result<Self, TopoError> {
        Self::open_with(path, IndexSpec::default())
    }

    /// Opens `path` using the `IndexSpec` persisted in its META (written by
    /// `Storage::ensure_index_spec` on every prior open), so callers need not
    /// supply one. A fresh file, or one predating spec persistence (no
    /// `index_spec` key), opens with `IndexSpec::default()`.
    ///
    /// Idempotent: the persisted spec is passed straight back through
    /// `open_with`, so `ensure_index_spec` sees an unchanged text list and no
    /// FTS reindex is triggered — the equality index is declared exactly as
    /// the file was created. A transient extra (read-only) open of the file
    /// is used to peek the spec before the real `open_with`.
    pub fn open_stored(path: impl AsRef<Path>) -> Result<Self, TopoError> {
        let path = path.as_ref();
        let spec = Storage::read_persisted_index_spec(path)?.unwrap_or_default();
        Self::open_with(path, spec)
    }

    /// Like `open`, but with a declared `IndexSpec` governing which
    /// `(label, prop)` pairs get equality/text-indexed. `spec` is validated
    /// (rejecting duplicate declarations) before anything else happens — an
    /// invalid spec never touches storage. Delegates to `open_with_options`
    /// with `DbOptions::default()`.
    pub fn open_with(path: impl AsRef<Path>, spec: IndexSpec) -> Result<Self, TopoError> {
        Self::open_with_options(path, spec, DbOptions::default())
    }

    /// Like `open_with`, but also takes [`DbOptions`] governing storage
    /// tuning knobs (currently just `cache_size_bytes`, threaded to redb's
    /// `Builder::set_cache_size`).
    pub fn open_with_options(
        path: impl AsRef<Path>,
        spec: IndexSpec,
        options: DbOptions,
    ) -> Result<Self, TopoError> {
        spec.validate()?;
        let spec = Arc::new(spec);
        let storage = Arc::new(Storage::open_with_options(path, spec, options)?);
        let (tx, rx) = bounded::<Job>(256);

        // The thread captures its own clones of `storage`/`subs` — never a
        // clone of `Inner` itself (see the comment on `Inner::storage` for
        // why: a strong ref back to `Inner` would create a cycle where
        // `Inner`'s `Drop` never fires).
        let storage_for_applier = storage.clone();
        let subs: Arc<Mutex<Vec<Sender<ChangeEvent>>>> = Arc::new(Mutex::new(Vec::new()));
        let subs_for_applier = subs.clone();
        let applier = std::thread::spawn(move || {
            // `pending` carries a job that was already POPPED off `rx` by a
            // `try_recv` drain (see the `Job::Apply` arm below) but turned
            // out not to belong to the group being drained. crossbeam's
            // `Receiver` has no peek — once popped, a job can't be put back
            // — so it's stashed here and processed on the NEXT loop
            // iteration, ahead of a fresh `rx.recv()`. This is the only
            // reordering the drain introduces: a job that was sitting in the
            // channel behind the triggering `Job::Apply` can end up
            // processed after a whole group of LATER-arriving `Job::Apply`s
            // that happened to be queued ahead of it at drain time — it was
            // concurrent with the group either way, so this is not
            // observable as anything other than ordinary scheduling
            // nondeterminism between concurrent submitters.
            let mut pending: Option<Job> = None;
            loop {
                let job = match pending.take() {
                    Some(job) => job,
                    None => match rx.recv() {
                        Ok(job) => job,
                        Err(_) => break,
                    },
                };
                match job {
                    Job::Apply { ops, at, reply } => {
                        // A `submit_at`-style deterministic timestamp is
                        // never merged into a group: `apply_batches` takes
                        // ONE `now_ms` for the whole group, so honoring a
                        // caller's explicit clock and a group's shared clock
                        // at once isn't possible. This keeps every test
                        // built on `submit_at` determinism exact — those
                        // jobs always take the single-batch path below,
                        // unchanged from pre-Task-6 behavior.
                        let Some(now) = at else {
                            let now = SystemTime::now()
                                .duration_since(UNIX_EPOCH)
                                .expect("system clock before UNIX epoch")
                                .as_millis() as i64;
                            // Drain up to 16 total `Job::Apply` jobs / 4096
                            // total ops (F9c): the common case under
                            // concurrent submitters is many small batches
                            // queued back to back, and sharing one
                            // `apply_batches` commit means they pay one
                            // fsync instead of one each. `try_recv` POPS —
                            // see the `pending` comment above for what
                            // happens to a non-matching popped job.
                            let mut jobs: Vec<ApplyJob> = vec![(ops, reply)];
                            let mut total_ops = jobs[0].0.len();
                            while jobs.len() < 16 && total_ops < 4096 {
                                match rx.try_recv() {
                                    Ok(Job::Apply {
                                        ops,
                                        at: None,
                                        reply,
                                    }) => {
                                        total_ops += ops.len();
                                        jobs.push((ops, reply));
                                    }
                                    Ok(other) => {
                                        pending = Some(other);
                                        break;
                                    }
                                    Err(_) => break,
                                }
                            }
                            if jobs.len() == 1 {
                                let (ops, reply) = jobs.pop().expect("len checked above");
                                apply_one_job(
                                    &storage_for_applier,
                                    &subs_for_applier,
                                    ops,
                                    now,
                                    reply,
                                );
                            } else {
                                apply_group(&storage_for_applier, &subs_for_applier, jobs, now);
                            }
                            continue;
                        };
                        apply_one_job(&storage_for_applier, &subs_for_applier, ops, now, reply);
                    }
                    Job::Rebuild { reply } => {
                        // Rebuild runs on the applier thread — the sole redb
                        // writer — so it serializes with in-flight batch
                        // application: `rebuild_state_from_ops` and
                        // `apply_batch` can never interleave. No separate
                        // vector-index rebuild step anymore (Task 7 deleted
                        // it) — `rebuild_state_from_ops` already rebuilds
                        // `vectors`/`embedding_ref` from the replayed ops via
                        // `apply_op`.
                        let result = storage_for_applier.rebuild_state_from_ops();
                        let _ = reply.send(result);
                    }
                    Job::BumpCounters { bumps } => {
                        // Auxiliary telemetry: fold into COUNTERS and move on.
                        // Deliberately NO op-log append and NO change-feed
                        // broadcast (the feed's broadcast lives only in the
                        // `Job::Apply` success arm above and stays there) — and
                        // no reply, since bumps are fire-and-forget. A failed
                        // write is swallowed: losing best-effort counters must
                        // never take down the applier.
                        let _ = storage_for_applier.merge_counter_bumps(&bumps);
                    }
                    Job::Compact { keep_from, reply } => {
                        // Runs on the applier (sole redb writer), so it
                        // serializes with batch application: no append can
                        // interleave between the delete and the `oldest_seq`
                        // stamp. Compaction touches only the OPS/META tables —
                        // never NODES/EDGES — so there is nothing to
                        // broadcast.
                        let _ = reply.send(storage_for_applier.compact_ops_through(keep_from));
                    }
                    Job::Meta { key, value, reply } => {
                        // Runs on the applier (sole redb writer). Touches
                        // only META — never NODES/EDGES — so there is
                        // nothing to broadcast.
                        let _ = reply.send(storage_for_applier.write_meta(&key, &value));
                    }
                }
            }
        });

        // Bumper thread: owns batching of access-counter bumps so reads never
        // pay a per-hit write. It holds a *clone* of the applier `Sender` and
        // forwards accumulated bumps as `Job::BumpCounters`. Because of that
        // clone, `Drop for Inner` MUST join this thread *before* dropping the
        // applier `tx` — otherwise the applier channel never closes and the
        // applier join hangs (see `Drop for Inner`).
        let (bump_tx, bump_rx) = bounded::<(NodeId, i64)>(4096);
        let applier_tx_for_bumper = tx.clone();
        let bumper = std::thread::spawn(move || {
            let mut pending: std::collections::HashMap<NodeId, (u64, i64)> = Default::default();
            let flush = |pending: &mut std::collections::HashMap<NodeId, (u64, i64)>| {
                if pending.is_empty() {
                    return;
                }
                let bumps: Vec<(NodeId, u64, i64)> =
                    pending.drain().map(|(id, (n, ts))| (id, n, ts)).collect();
                // Applier gone (shutdown race) → drop silently; aux data.
                let _ = applier_tx_for_bumper.send(Job::BumpCounters { bumps });
            };
            loop {
                match bump_rx.recv_timeout(std::time::Duration::from_millis(100)) {
                    Ok((id, ts)) => {
                        let e = pending.entry(id).or_insert((0, 0));
                        e.0 += 1;
                        e.1 = e.1.max(ts);
                        if pending.len() >= 256 {
                            flush(&mut pending);
                        }
                    }
                    Err(crossbeam_channel::RecvTimeoutError::Timeout) => flush(&mut pending),
                    Err(crossbeam_channel::RecvTimeoutError::Disconnected) => {
                        flush(&mut pending);
                        break;
                    }
                }
            }
        });

        Ok(Self {
            inner: Arc::new(Inner {
                storage,
                tx: Mutex::new(Some(tx)),
                applier: Mutex::new(Some(applier)),
                subs,
                bump_tx: Mutex::new(Some(bump_tx)),
                bumper: Mutex::new(Some(bumper)),
                debug_last_search_used_graph: AtomicBool::new(false),
            }),
        })
    }

    /// The underlying storage. Used by `search_text` (in `fts.rs`) to open a
    /// read transaction over the POSTINGS/FTS_DOCS/META tables from an
    /// `impl Db` block in a sibling module that can't touch `self.inner`.
    #[must_use]
    pub(crate) fn storage(&self) -> &Storage {
        &self.inner.storage
    }

    /// The debug-instrumentation atomic (F8 Task 4). Used by `search_vector`
    /// (in `vector.rs`, a sibling module that can't touch `self.inner`) to
    /// pass a stable reference down into `vector_store::search_scan`'s
    /// per-scope routing, which sets it on every branch it takes.
    #[must_use]
    pub(crate) fn debug_atomic(&self) -> &AtomicBool {
        &self.inner.debug_last_search_used_graph
    }

    /// The on-disk format version of the opened file (delegates to
    /// `Storage::format_version`). Added so `topodb-cli`'s `info` can report
    /// it without reaching into crate internals.
    pub fn format_version(&self) -> u32 {
        // `Storage::format_version` only fails on a missing/malformed META
        // row, which `open_with` guarantees exists (it writes it on first
        // create and validates it on every open) — unreachable for a `Db`
        // that has successfully opened.
        self.inner
            .storage
            .format_version()
            .expect("format_version: META row guaranteed by a successful open")
    }

    /// The `IndexSpec` this db is operating under — the one `open_stored`
    /// resolved (or the one passed to `open_with`). Added so `info` can
    /// report it. A clone of `Storage`'s own copy (the source of truth —
    /// there is no longer a separate snapshot-carried copy to read instead).
    #[must_use]
    pub fn index_spec(&self) -> IndexSpec {
        (*self.inner.storage.spec).clone()
    }

    /// Per-table logical byte counts; benchmark/inspection seam.
    #[doc(hidden)]
    pub fn storage_report(&self) -> Result<Vec<crate::storage::TableReport>, TopoError> {
        self.inner.storage.storage_report()
    }

    /// Records an access bump for each id in `ids`, timestamped with a single
    /// wall-clock read taken once per call. Fire-and-forget: each `(id, now)`
    /// is `try_send`'d to the bumper thread, and on a full or closed channel it
    /// is *silently dropped*. Counters are auxiliary telemetry — a read must
    /// never block, retry, or fail because the counter pipeline is saturated or
    /// shutting down. Called from the scoped read paths (`node`,
    /// `nodes_by_label`, `traverse`) with exactly the nodes they returned.
    pub(crate) fn bump(&self, ids: impl IntoIterator<Item = NodeId>) {
        let now = SystemTime::now()
            .duration_since(UNIX_EPOCH)
            .expect("system clock before UNIX epoch")
            .as_millis() as i64;
        // Clone the sender out from under the mutex so we never hold the lock
        // across `try_send`. `None` once `Drop` has taken it — nothing to bump.
        // A poisoned mutex (applier panicked; poisoned-lock policy) also yields
        // `None`: bumps are auxiliary telemetry, so we silently drop them rather
        // than propagate the panic into a read path.
        let tx = self
            .inner
            .bump_tx
            .lock()
            .ok()
            .and_then(|g| g.as_ref().cloned());
        if let Some(tx) = tx {
            for id in ids {
                // Full (bumper backed up) or Disconnected (shutdown) → drop.
                let _ = tx.try_send((id, now));
            }
        }
    }

    /// Auxiliary access statistics for `id`, scoped exactly like [`Db::node`]:
    /// `None` if the node is absent OR out of `scopes` (the two are
    /// indistinguishable by design); `Some(AccessStats::default())` if the node
    /// exists in scope but has never been counted. **Reading stats never
    /// bumps** — this is a pure read of the COUNTERS table gated on node
    /// existence, so callers can inspect recency without perturbing it.
    pub fn access_stats(
        &self,
        scopes: &ScopeSet,
        id: NodeId,
    ) -> Result<Option<AccessStats>, TopoError> {
        // Gate on scoped existence *without* going through `node()` — reading
        // stats must never bump, and `node()` bumps. We replicate its scope
        // filter directly against storage: `None` if absent OR out of scope
        // (indistinguishable, mirroring `node()`).
        let in_scope = self
            .inner
            .storage
            .load_node(id)?
            .is_some_and(|n| scopes.contains(n.scope));
        if !in_scope {
            return Ok(None);
        }
        Ok(Some(
            self.inner.storage.read_counter(id)?.unwrap_or_default(),
        ))
    }

    /// Access count for scoring: NO existence/scope gate (recall's
    /// candidates are already scope-filtered by the legs) and NO bump —
    /// a raw COUNTERS read. 0 on absent or on read error: a counter read
    /// must never fail a recall.
    pub(crate) fn access_count_unbumped(&self, id: NodeId) -> u64 {
        self.inner
            .storage
            .read_counter(id)
            .ok()
            .flatten()
            .map(|s| s.access_count)
            .unwrap_or(0)
    }

    /// Submits a batch of ops for application, blocking until the applier
    /// thread has processed it. Safe to call from any thread; batches from
    /// concurrent callers serialize through the single applier. Uses the
    /// wall clock (`SystemTime::now`) to resolve any unset timestamps.
    pub fn submit(&self, ops: Vec<Op>) -> Result<AppliedBatch, TopoError> {
        self.submit_inner(ops, None)
    }

    /// Like `submit`, but resolves unset timestamps to `now_ms` instead of
    /// the wall clock. Intended for tests and backdating.
    pub fn submit_at(&self, ops: Vec<Op>, now_ms: i64) -> Result<AppliedBatch, TopoError> {
        self.submit_inner(ops, Some(now_ms))
    }

    /// Subscribes to the change feed, returning the `Receiver` half of a fresh
    /// bounded channel (`capacity` slots) registered with the applier. Every
    /// op the applier commits after this call is pushed as a [`ChangeEvent`]
    /// carrying a monotonic op-log `seq`.
    ///
    /// **Unscoped, by spec design.** The change feed is a *host-level*
    /// primitive that powers external consolidation/decay — it must observe
    /// every committed write regardless of scope. Unlike the scoped read APIs
    /// (`node`, `nodes_by_label`, `traverse`), it is not gated by a
    /// `ScopeSet`.
    ///
    /// **Delivery contract (best-effort, never blocks the applier):** if this
    /// subscriber's buffer is full when the applier broadcasts, the event is
    /// **DROPPED** for this subscriber — the applier never blocks on a slow
    /// consumer. The subscriber detects the resulting gap in `seq` and
    /// recovers the missing ops with [`Db::ops_since`]. A receiver that has
    /// been dropped is pruned from the registry on the next broadcast.
    /// Rejected batches, counter flushes, and rebuilds broadcast nothing;
    /// reads never produce events.
    ///
    /// A `capacity` of 0 is clamped to 1 — crossbeam's zero-capacity channels
    /// are rendezvous channels, which would silently drop nearly every event.
    ///
    /// **Anchoring a gap-free live tail:** capture the log position *before*
    /// subscribing, then backfill the window between them once:
    /// `let seq = db.current_seq()?; let rx = db.subscribe(cap);` then replay
    /// `ops_since(seq + 1)` once and dedup by `seq` against the channel. Any op
    /// committed between the two calls appears in both the replay and the live
    /// channel; deduping by `seq` collapses the overlap, and nothing in the gap
    /// is missed. This recipe is seamless across compaction too:
    /// [`current_seq`](Db::current_seq) survives an empty-but-compacted log
    /// (it falls back to the retained floor), so `ops_since(current_seq() +
    /// 1)` never spuriously returns [`TopoError::Compacted`] right after an
    /// emptying compaction — no special-casing needed at the call site.
    #[must_use]
    pub fn subscribe(&self, capacity: usize) -> Receiver<ChangeEvent> {
        let capacity = capacity.max(1);
        let (tx, rx) = bounded::<ChangeEvent>(capacity);
        // Poisoned subs registry ⇒ the applier panicked and the engine is dead
        // (poisoned-lock policy, see vector.rs). Hand back an already-disconnected
        // Receiver rather than propagating the panic: it reports `Disconnected`
        // immediately, the same terminal signal a subscriber sees after shutdown.
        match self.inner.subs.lock() {
            Ok(mut subs) => subs.push(tx),
            Err(_) => {
                let (tx, rx) = bounded(1);
                drop(tx);
                return rx;
            }
        }
        rx
    }

    /// Replays the durable op log from `since_seq` (**INCLUSIVE**), returning
    /// one [`ChangeEvent`] per op in ascending `seq` order. This is the pull
    /// side of the change feed: subscribers that dropped events (buffer full)
    /// call it to recover the gap after noticing a jump in `seq`.
    ///
    /// **Unscoped, by spec design** — same rationale as [`Db::subscribe`]: the
    /// change feed is a host-level primitive that must see every write. This
    /// is a read: it produces no events of its own.
    ///
    /// Reading below the oldest retained seq returns
    /// [`TopoError::Compacted { oldest }`](TopoError::Compacted): the requested
    /// range dips beneath the compaction floor, so a partial replay would
    /// silently drop history. The caller re-anchors from materialized state
    /// (the NODES/EDGES tables, which stay the source of truth after
    /// compaction) rather than trusting a truncated tail. An uncompacted log
    /// has a floor of 1, so any `since_seq` succeeds.
    pub fn ops_since(&self, since_seq: u64) -> Result<Vec<ChangeEvent>, TopoError> {
        let ops = self.inner.storage.read_ops(since_seq)?;
        Ok(ops
            .into_iter()
            .map(|(seq, op)| ChangeEvent {
                seq,
                op: Arc::new(op),
            })
            .collect())
    }

    /// The highest op-log seq committed so far (0 when the log has never been
    /// written). A plain storage read — no applier round-trip — so it is
    /// cheap and safe to call from any thread. Its purpose is to anchor a
    /// gap-free live tail: take it *before* [`subscribe`](Db::subscribe),
    /// then backfill with `ops_since(seq + 1)` (see `subscribe`'s anchoring
    /// recipe).
    ///
    /// Survives compaction: on an empty-but-compacted log the last OPS key is
    /// gone, but this falls back to the retained floor (`oldest_seq - 1`) so
    /// the high-water mark is never lost. The anchoring recipe's
    /// `ops_since(current_seq() + 1)` therefore never spuriously returns
    /// [`TopoError::Compacted`] right after an emptying compaction — it only
    /// returns `Compacted` for a seq genuinely below the retained floor.
    #[must_use = "the seq anchors ops_since"]
    pub fn current_seq(&self) -> Result<u64, TopoError> {
        self.inner.storage.current_seq()
    }

    /// Compacts the durable op log, dropping every entry with seq `< keep_from`
    /// and advancing the retained floor to `keep_from`. After this,
    /// [`ops_since`](Db::ops_since) below `keep_from` returns
    /// [`TopoError::Compacted`] and [`rebuild_state_from_ops`](Db::rebuild_state_from_ops)
    /// refuses (a compacted log is no longer a full history — NODES/EDGES stay
    /// the materialized source of truth).
    ///
    /// **Host-level primitive** (unscoped, like the change feed it serves).
    /// Edge behaviour mirrors `Storage::compact_ops_through`:
    /// `keep_from <= oldest` is a no-op, `keep_from > current_seq + 1` is
    /// rejected, and `keep_from == current_seq + 1` legally empties the log.
    /// Runs on the applier thread and blocks until it commits; `Closed` after
    /// shutdown, same contract as [`submit`](Db::submit).
    pub fn compact_ops(&self, keep_from: u64) -> Result<(), TopoError> {
        let (reply_tx, reply_rx) = bounded(1);
        let tx = self.sender().ok_or(TopoError::Closed)?;
        tx.send(Job::Compact {
            keep_from,
            reply: reply_tx,
        })
        .map_err(|_| TopoError::Closed)?;
        reply_rx.recv().map_err(|_| TopoError::Closed)?
    }

    /// Reads a key from the `META` table — persistent key/value storage
    /// outside the scoped node/edge graph. `None` if the key was never set.
    /// Reads don't need the applier (redb read txns don't race the single
    /// writer), so this goes straight to `Storage`.
    ///
    /// Reserved keys (`format_version`, `hnsw_params`, `index_spec`) are
    /// readable through this too, but callers should namespace their own
    /// keys (e.g. `onboarding:<name>`) to avoid colliding with them.
    pub fn get_meta(&self, key: &str) -> Result<Option<Vec<u8>>, TopoError> {
        self.inner.storage.read_meta(key)
    }

    /// Writes a key into the `META` table. Runs on the applier thread (the
    /// sole redb writer), blocking until the write has committed —
    /// `Closed` after shutdown, same contract as [`submit`](Db::submit).
    ///
    /// See [`get_meta`](Db::get_meta) for the reserved-key caveat.
    pub fn set_meta(&self, key: &str, value: &[u8]) -> Result<(), TopoError> {
        let (reply_tx, reply_rx) = bounded(1);
        let tx = self.sender().ok_or(TopoError::Closed)?;
        tx.send(Job::Meta {
            key: key.to_string(),
            value: value.to_vec(),
            reply: reply_tx,
        })
        .map_err(|_| TopoError::Closed)?;
        reply_rx.recv().map_err(|_| TopoError::Closed)?
    }

    /// Clones the job `Sender` out of the mutex and releases the guard before
    /// the caller does anything blocking with it. `None` once `Drop` has taken
    /// the sender. Holding the guard across a (potentially blocking) `send` on
    /// the bounded channel would needlessly serialize all submitters against
    /// each other on the mutex rather than on the channel.
    fn sender(&self) -> Option<Sender<Job>> {
        // A poisoned mutex (applier panicked; poisoned-lock policy) maps to
        // `None`, which `submit_inner`/`rebuild_state_from_ops` already turn into
        // `TopoError::Closed` — the same result as a shut-down engine.
        self.inner.tx.lock().ok().and_then(|g| g.as_ref().cloned())
    }

    fn submit_inner(&self, ops: Vec<Op>, at: Option<i64>) -> Result<AppliedBatch, TopoError> {
        let (reply_tx, reply_rx) = bounded(1);
        let tx = self.sender().ok_or(TopoError::Closed)?;
        tx.send(Job::Apply {
            ops,
            at,
            reply: reply_tx,
        })
        .map_err(|_| TopoError::Closed)?;
        reply_rx.recv().map_err(|_| TopoError::Closed)?
    }

    /// Test/inspection helper: every edge `(from, to)` currently in storage,
    /// open or closed. `#[doc(hidden)]` — callers should prefer the query
    /// layer once it exists. Full `EdgeRecord`s (props included), resolved
    /// via a bounded OUT_ADJ scan from `from`'s slot (one read transaction) —
    /// never a full-table scan.
    #[doc(hidden)]
    pub fn all_edges_between(&self, from: NodeId, to: NodeId) -> Vec<crate::state::EdgeRecord> {
        self.edges_between(from, to).unwrap_or_default()
    }

    /// Test/inspection helper: the ids of currently-open edges `(from, to)`
    /// (i.e. `valid_to.is_none()`). `#[doc(hidden)]` — see
    /// `all_edges_between`.
    #[doc(hidden)]
    pub fn open_edges_between(&self, from: NodeId, to: NodeId) -> Vec<EdgeId> {
        self.edges_between(from, to)
            .unwrap_or_default()
            .into_iter()
            .filter(|e| e.valid_to.is_none())
            .map(|e| e.id)
            .collect()
    }

    /// Scoped edge listing from `from`: every edge whose source is `from`,
    /// optionally restricted to a target node, an edge type (matched against
    /// the stored type string exactly — normalize before calling if the type
    /// vocabulary is normalized), and to currently-open edges only
    /// (`valid_to.is_none()`). Only edges whose own scope is in `scopes` are
    /// returned (there is no unscoped read); the endpoints' scopes are NOT
    /// re-gated — an in-scope edge already names both endpoint ids, and this
    /// returns edge records, not node records.
    ///
    /// This is the supersession primitive: when a fact changes, list the open
    /// edges of the old fact's type here, close them, and create the new edge
    /// — the discovery step `traverse` is too coarse for. A missing `from`
    /// slot (never existed, or removed) yields an empty result, not an error.
    /// Does not bump access counters — no node record is returned.
    ///
    /// `axis` picks what `open_only` means: `Valid` (unchanged) gates on the
    /// adjacency entry's `valid_to` — the hot path, no extra fetch. `Recorded`
    /// gates on the fetched record's `superseded_at` instead — nearly free
    /// here since every candidate already needs a full-record fetch for the
    /// result.
    pub fn edges_from(
        &self,
        scopes: &ScopeSet,
        from: NodeId,
        to: Option<NodeId>,
        ty: Option<&str>,
        open_only: bool,
        axis: crate::read::TimeAxis,
    ) -> Result<Vec<crate::state::EdgeRecord>, TopoError> {
        self.scan_adjacency(
            scopes,
            from,
            to,
            ty,
            EdgeDirection::Out,
            EdgeGate::OpenOnly { open_only, axis },
        )
    }

    /// Scoped edge listing to `to`: every edge whose target is `to`,
    /// optionally restricted to a source node, an edge type (matched against
    /// the stored type string exactly — normalize before calling if the type
    /// vocabulary is normalized), and to currently-open edges only
    /// (`valid_to.is_none()`). Only edges whose own scope is in `scopes` are
    /// returned (there is no unscoped read); the endpoints' scopes are NOT
    /// re-gated — an in-scope edge already names both endpoint ids, and this
    /// returns edge records, not node records.
    ///
    /// Incoming-edge counterpart to `edges_from`: lists edges pointing TO a
    /// node, reading the reverse adjacency via IN_ADJ (the same mechanism
    /// `traverse` uses for `Direction::In`). A missing `to` slot (never
    /// existed, or removed) yields an empty result, not an error.
    /// Does not bump access counters — no node record is returned.
    ///
    /// `axis` picks what `open_only` means: see `edges_from`.
    pub fn edges_to(
        &self,
        scopes: &ScopeSet,
        to: NodeId,
        from: Option<NodeId>,
        ty: Option<&str>,
        open_only: bool,
        axis: crate::read::TimeAxis,
    ) -> Result<Vec<crate::state::EdgeRecord>, TopoError> {
        self.scan_adjacency(
            scopes,
            to,
            from,
            ty,
            EdgeDirection::In,
            EdgeGate::OpenOnly { open_only, axis },
        )
    }

    /// [`Db::edges_from`], gated by an Allen predicate over the edge's valid
    /// interval `[valid_from, valid_to)` (pragmatic subset — see
    /// [`crate::ValidInterval`]) instead of the `open_only`/`axis` pair. The
    /// predicate REPLACES the open-only gate, so this surface has no
    /// `open_only` parameter — the two cannot be combined here (hosts reject
    /// an explicit `open_only` alongside a predicate); and it gates the valid
    /// axis only, so there is no `axis` parameter either (recorded-axis
    /// intervals are out of scope). Gating happens on the adjacency entries'
    /// interval fields, same place as `edges_from`'s open-only gate — no
    /// extra record fetches. `Rejected` on an inverted or non-positive
    /// interval; everything else (scoping, type filter, missing-slot
    /// behavior, no counter bumps, oldest-first order) matches `edges_from`.
    pub fn edges_from_interval(
        &self,
        scopes: &ScopeSet,
        from: NodeId,
        to: Option<NodeId>,
        ty: Option<&str>,
        valid_interval: crate::read::ValidInterval,
    ) -> Result<Vec<crate::state::EdgeRecord>, TopoError> {
        valid_interval.validate()?;
        self.scan_adjacency(
            scopes,
            from,
            to,
            ty,
            EdgeDirection::Out,
            EdgeGate::Interval(&valid_interval),
        )
    }

    /// Incoming-edge counterpart to [`Db::edges_from_interval`]: every edge
    /// pointing TO `to` whose valid interval satisfies the Allen predicate,
    /// read via IN_ADJ. Same contract as `edges_from_interval` throughout
    /// (predicate replaces the open-only gate, valid axis only, no counter
    /// bumps, oldest-first order).
    pub fn edges_to_interval(
        &self,
        scopes: &ScopeSet,
        to: NodeId,
        from: Option<NodeId>,
        ty: Option<&str>,
        valid_interval: crate::read::ValidInterval,
    ) -> Result<Vec<crate::state::EdgeRecord>, TopoError> {
        valid_interval.validate()?;
        self.scan_adjacency(
            scopes,
            to,
            from,
            ty,
            EdgeDirection::In,
            EdgeGate::Interval(&valid_interval),
        )
    }

    /// Private helper for adjacency scanning shared by `edges_from`,
    /// `edges_to`, `edges_from_interval`, and `edges_to_interval`. Scans the
    /// adjacency table in the given `direction` from `primary_node`,
    /// optionally filtered to `filter_node`, by `ty`, and gated by `gate`.
    /// Returns scope-filtered edge records in deterministic (oldest-first)
    /// order by edge id.
    ///
    /// An edge-type name the dict has never interned has never been written:
    /// match nothing (an empty filter list is still a filter), mirroring
    /// `traverse`'s treatment of unknown type names.
    fn scan_adjacency(
        &self,
        scopes: &ScopeSet,
        primary_node: NodeId,
        filter_node: Option<NodeId>,
        ty: Option<&str>,
        direction: EdgeDirection,
        gate: EdgeGate,
    ) -> Result<Vec<crate::state::EdgeRecord>, TopoError> {
        let storage = self.storage();
        let dicts = storage.dicts.read().expect("dict lock poisoned");
        let scope_registry = storage
            .scope_registry
            .read()
            .expect("scope registry lock poisoned");
        let type_filter: Option<Vec<u32>> = ty.map(|name| {
            dicts
                .id_of(crate::dict::DictKind::EdgeType, name)
                .into_iter()
                .collect()
        });
        let tx = storage.db.begin_read().map_err(crate::error::storage_err)?;
        let node_slots = tx
            .open_table(crate::slots::NODE_SLOTS)
            .map_err(crate::error::storage_err)?;
        let Some(primary_slot) = crate::slots::node_slot(&node_slots, primary_node)? else {
            return Ok(Vec::new());
        };
        let filter_slot = match filter_node {
            None => None,
            Some(node) => match crate::slots::node_slot(&node_slots, node)? {
                // A filter node that has no slot has no edges either.
                None => return Ok(Vec::new()),
                some => some,
            },
        };
        let adj = match direction {
            EdgeDirection::Out => tx
                .open_table(crate::adj::OUT_ADJ)
                .map_err(crate::error::storage_err)?,
            EdgeDirection::In => tx
                .open_table(crate::adj::IN_ADJ)
                .map_err(crate::error::storage_err)?,
        };
        let edges_table = tx
            .open_table(crate::storage::EDGES)
            .map_err(crate::error::storage_err)?;
        let node_ids = tx
            .open_table(crate::slots::NODE_IDS)
            .map_err(crate::error::storage_err)?;
        let mut out = Vec::new();
        for (_ty, entry) in crate::adj::read_adj(&adj, primary_slot, type_filter.as_deref())? {
            if filter_slot.is_some_and(|slot| entry.target != slot) {
                continue;
            }
            if gate.should_skip_entry(&entry) {
                continue;
            }
            let entry_scope = scope_registry.resolve(entry.scope)?;
            if !scopes.contains(entry_scope) {
                continue;
            }
            if let Some(rec) = crate::storage::read_edge_by_slot(
                &edges_table,
                &dicts,
                &scope_registry,
                &node_ids,
                entry.edge,
            )? {
                if gate.should_skip_record(&rec) {
                    continue;
                }
                out.push(rec);
            }
        }
        // Deterministic order: by edge id (ULIDs sort by mint time, so this
        // is oldest-first).
        out.sort_by_key(|e| e.id);
        Ok(out)
    }

    /// Shared implementation for `all_edges_between`/`open_edges_between`: a
    /// bounded OUT_ADJ scan from `from`'s slot, filtered to entries whose
    /// target resolves to `to`, then fetched as full `EdgeRecord`s — all in
    /// one read transaction. A missing `from`/`to` slot (node never existed,
    /// or was removed) yields an empty result, not an error — mirrors
    /// `Db::node`'s "absence is absence" treatment of a storage miss.
    fn edges_between(
        &self,
        from: NodeId,
        to: NodeId,
    ) -> Result<Vec<crate::state::EdgeRecord>, TopoError> {
        let storage = self.storage();
        let dicts = storage.dicts.read().expect("dict lock poisoned");
        let scope_registry = storage
            .scope_registry
            .read()
            .expect("scope registry lock poisoned");
        let tx = storage.db.begin_read().map_err(crate::error::storage_err)?;
        let node_slots = tx
            .open_table(crate::slots::NODE_SLOTS)
            .map_err(crate::error::storage_err)?;
        let Some(from_slot) = crate::slots::node_slot(&node_slots, from)? else {
            return Ok(Vec::new());
        };
        let Some(to_slot) = crate::slots::node_slot(&node_slots, to)? else {
            return Ok(Vec::new());
        };
        let out_adj = tx
            .open_table(crate::adj::OUT_ADJ)
            .map_err(crate::error::storage_err)?;
        let edges_table = tx
            .open_table(crate::storage::EDGES)
            .map_err(crate::error::storage_err)?;
        let node_ids = tx
            .open_table(crate::slots::NODE_IDS)
            .map_err(crate::error::storage_err)?;
        let mut out = Vec::new();
        for (_ty, entry) in crate::adj::read_adj(&out_adj, from_slot, None)? {
            if entry.target != to_slot {
                continue;
            }
            if let Some(rec) = crate::storage::read_edge_by_slot(
                &edges_table,
                &dicts,
                &scope_registry,
                &node_ids,
                entry.edge,
            )? {
                out.push(rec);
            }
        }
        Ok(out)
    }

    /// Rebuilds NODES/EDGES (and the adjacency/index tables derived from
    /// them) from the OPS log — see `Storage::rebuild_state_from_ops`. The
    /// read model is disk-resident, so readers observe the rebuilt state as
    /// soon as this returns — there is no separate in-memory snapshot to
    /// keep in step with it.
    ///
    /// The rebuild is performed *on the applier thread* (via a `Job::Rebuild`
    /// routed through the same channel as `submit`), not on the caller
    /// thread. The applier is the sole redb writer; doing the rebuild there
    /// serializes it with batch application, so `rebuild_state_from_ops` and
    /// an in-flight `apply_batch` can never interleave. Blocks until the
    /// applier replies; `Closed` after shutdown, same contract as `submit`.
    #[doc(hidden)]
    pub fn rebuild_state_from_ops(&self) -> Result<(), TopoError> {
        let (reply_tx, reply_rx) = bounded(1);
        let tx = self.sender().ok_or(TopoError::Closed)?;
        tx.send(Job::Rebuild { reply: reply_tx })
            .map_err(|_| TopoError::Closed)?;
        reply_rx.recv().map_err(|_| TopoError::Closed)?
    }

    /// Test/inspection helper: every node currently in storage, sorted by
    /// id for deterministic comparison. `#[doc(hidden)]` — see
    /// `all_edges_between`.
    #[doc(hidden)]
    pub fn debug_dump_nodes(&self) -> Vec<crate::state::NodeRecord> {
        let mut out = self
            .inner
            .storage
            .all_nodes()
            .expect("debug dump: storage read failed");
        out.sort_by_key(|n| n.id);
        out
    }

    /// Test/inspection helper: every edge currently in storage, sorted by
    /// id for deterministic comparison. `#[doc(hidden)]` — see
    /// `all_edges_between`.
    #[doc(hidden)]
    pub fn debug_dump_edges(&self) -> Vec<crate::state::EdgeRecord> {
        let mut out = self
            .inner
            .storage
            .all_edges()
            .expect("debug dump: storage read failed");
        out.sort_by_key(|e| e.id);
        out
    }

    /// Test/inspection helper: the raw contents of both adjacency tables
    /// (OUT_ADJ and IN_ADJ), every chunk decoded, **open AND closed entries
    /// included**, sorted deterministically. `#[doc(hidden)]` — see
    /// `all_edges_between`. Exists so tests can assert byte-level adjacency
    /// parity (e.g. that `rebuild_state_from_ops` reproduces identical chunk
    /// content, including closed edges' `valid_to`, which no `as_of`-filtered
    /// public read can observe). A full-table iteration — fine here: a debug
    /// dump is inherently a full scan, not a production read path. Rows are
    /// plain tuples ([`AdjacencyDumpRow`]) rather than the crate-internal
    /// `AdjEntryDisk` type.
    #[doc(hidden)]
    pub fn debug_dump_adjacency(&self) -> Result<Vec<AdjacencyDumpRow>, TopoError> {
        use redb::ReadableTable;
        let storage = self.storage();
        let tx = storage.db.begin_read().map_err(crate::error::storage_err)?;
        let mut out = Vec::new();
        for (is_out, table_def) in [(true, crate::adj::OUT_ADJ), (false, crate::adj::IN_ADJ)] {
            let table = tx
                .open_table(table_def)
                .map_err(crate::error::storage_err)?;
            for entry in table.iter().map_err(crate::error::storage_err)? {
                let (k, v) = entry.map_err(crate::error::storage_err)?;
                let key: [u8; 16] = k
                    .value()
                    .try_into()
                    .map_err(|_| TopoError::Encoding("bad adjacency key".into()))?;
                let slot = u64::from_be_bytes(key[..8].try_into().expect("8-byte slice"));
                let edge_type = u32::from_be_bytes(key[8..12].try_into().expect("4-byte slice"));
                let raw = crate::codec::unframe_value(v.value())?;
                for e in crate::adj::decode_block(raw.as_ref())? {
                    out.push((
                        slot,
                        edge_type,
                        is_out,
                        e.target,
                        e.edge,
                        e.scope,
                        e.valid_from,
                        e.valid_to,
                    ));
                }
            }
        }
        out.sort_unstable();
        Ok(out)
    }

    /// Test/inspection helper: every `POSTINGS` chunk row currently in
    /// storage — the raw key bytes (`scope_id.to_be_bytes() ++ term-UTF-8 ++
    /// chunk.to_be_bytes()`, see `fts::chunked_posting_key`) paired with its
    /// decoded `(slot, tf)` entries, sorted by key bytes. `#[doc(hidden)]` —
    /// see `all_edges_between`. The key is left as raw bytes rather than
    /// split into `(scope_id, term, chunk)`: the term's length is variable
    /// and not self-describing from the key alone, so decomposing it here
    /// would need the same disambiguation `fts.rs`'s chunk-key scan already
    /// handles internally — pointless for a byte-parity debug dump, which
    /// only needs the key to compare equal or not. Exists so tests can
    /// assert byte-level postings parity across `rebuild_state_from_ops`,
    /// the same role `debug_dump_adjacency` plays for OUT_ADJ/IN_ADJ.
    #[doc(hidden)]
    pub fn debug_dump_postings(&self) -> Result<Vec<PostingsDumpRow>, TopoError> {
        use redb::ReadableTable;
        let storage = self.storage();
        let tx = storage.db.begin_read().map_err(crate::error::storage_err)?;
        let table = tx
            .open_table(crate::storage::POSTINGS)
            .map_err(crate::error::storage_err)?;
        let mut out = Vec::new();
        for entry in table.iter().map_err(crate::error::storage_err)? {
            let (k, v) = entry.map_err(crate::error::storage_err)?;
            let key = k.value().to_vec();
            let raw = crate::codec::unframe_value(v.value())?;
            let entries = crate::fts::decode_posting_block(raw.as_ref())?;
            out.push((key, entries));
        }
        out.sort_unstable();
        Ok(out)
    }

    /// Test/inspection helper: every `VECTORS` row currently in storage,
    /// decoded as `(model, scope, slot, vector)`, sorted by that same
    /// `(model, scope, slot)` tuple — the same order `vector_store::vector_key`
    /// sorts its keys in. `#[doc(hidden)]` — see `all_edges_between`. Sorted
    /// explicitly (rather than relying on redb's key-order iteration) so the
    /// dump's ordering guarantee doesn't depend on that implementation
    /// detail; `Vec<f32>` isn't `Ord`, so the sort compares only the decoded
    /// key fields, not the vector payload.
    #[doc(hidden)]
    pub fn debug_dump_vectors(&self) -> Result<Vec<VectorsDumpRow>, TopoError> {
        use redb::ReadableTable;
        let storage = self.storage();
        let tx = storage.db.begin_read().map_err(crate::error::storage_err)?;
        let table = tx
            .open_table(crate::vector_store::VECTORS)
            .map_err(crate::error::storage_err)?;
        let mut out = Vec::new();
        for entry in table.iter().map_err(crate::error::storage_err)? {
            let (k, v) = entry.map_err(crate::error::storage_err)?;
            let key: [u8; 16] = k
                .value()
                .try_into()
                .map_err(|_| TopoError::Encoding("bad vectors key".into()))?;
            let model = u32::from_be_bytes(key[0..4].try_into().expect("4-byte slice"));
            let scope = u32::from_be_bytes(key[4..8].try_into().expect("4-byte slice"));
            let slot = u64::from_be_bytes(key[8..16].try_into().expect("8-byte slice"));
            let raw = crate::codec::unframe_value(v.value())?;
            let (scale, codes): (f32, Vec<i8>) = postcard::from_bytes(raw.as_ref())
                .map_err(|e| TopoError::Encoding(e.to_string()))?;
            let vector = crate::quant::dequantize(scale, &codes);
            out.push((model, scope, slot, vector));
        }
        out.sort_by_key(|a| (a.0, a.1, a.2));
        Ok(out)
    }

    /// Test/inspection helper: every `EMBEDDING_REF` row currently in
    /// storage, decoded as `(slot, model, scope)` — a node's CURRENT
    /// embedding pointer (see `vector_store::put_vector`) — sorted by slot.
    /// `#[doc(hidden)]` — see `all_edges_between`.
    #[doc(hidden)]
    pub fn debug_dump_embedding_ref(&self) -> Result<Vec<EmbeddingRefDumpRow>, TopoError> {
        use redb::ReadableTable;
        let storage = self.storage();
        let tx = storage.db.begin_read().map_err(crate::error::storage_err)?;
        let table = tx
            .open_table(crate::vector_store::EMBEDDING_REF)
            .map_err(crate::error::storage_err)?;
        let mut out = Vec::new();
        for entry in table.iter().map_err(crate::error::storage_err)? {
            let (k, v) = entry.map_err(crate::error::storage_err)?;
            let key: [u8; 8] = k
                .value()
                .try_into()
                .map_err(|_| TopoError::Encoding("bad embedding_ref key".into()))?;
            let slot = u64::from_be_bytes(key);
            let (model, scope) = crate::vector_store::decode_ref(v.value())?;
            out.push((slot, model, scope));
        }
        out.sort_unstable();
        Ok(out)
    }

    /// Test/inspection helper: every `VECTOR_DIMS` row currently in storage,
    /// decoded as `(model_id, dim)` — the per-model pinned embedding
    /// dimension enforced by `storage::check_or_pin_dim` — sorted by model
    /// id. `#[doc(hidden)]` — see `all_edges_between`.
    #[doc(hidden)]
    pub fn debug_dump_vector_dims(&self) -> Result<Vec<VectorDimsDumpRow>, TopoError> {
        use redb::ReadableTable;
        let storage = self.storage();
        let tx = storage.db.begin_read().map_err(crate::error::storage_err)?;
        let table = tx
            .open_table(crate::storage::VECTOR_DIMS)
            .map_err(crate::error::storage_err)?;
        let mut out = Vec::new();
        for entry in table.iter().map_err(crate::error::storage_err)? {
            let (k, v) = entry.map_err(crate::error::storage_err)?;
            let key: [u8; 4] = k
                .value()
                .try_into()
                .map_err(|_| TopoError::Encoding("bad vector_dims key".into()))?;
            let model_id = u32::from_be_bytes(key);
            let val: [u8; 4] = v
                .value()
                .try_into()
                .map_err(|_| TopoError::Encoding("bad vector_dims value".into()))?;
            let dim = u32::from_le_bytes(val);
            out.push((model_id, dim));
        }
        out.sort_unstable();
        Ok(out)
    }

    /// Test/inspection helper: every `LABEL_INDEX` row currently in storage,
    /// decoded as `(label_id, scope_id, node_id, slot)` from the raw
    /// `label_id BE ++ scope_id BE ++ node_id BE` key and `u64` slot value
    /// (`storage::label_index_key`) — sorted by that same tuple, which is
    /// also the on-disk key order (mint-time order within a `(label,
    /// scope)` prefix — see `LABEL_INDEX`'s doc comment). `#[doc(hidden)]` —
    /// see `all_edges_between`. Exists so tests can assert byte-level
    /// LABEL_INDEX parity across `rebuild_state_from_ops`, the same role
    /// `debug_dump_postings`/`debug_dump_vectors`/etc. play for the other
    /// derived/recall tables.
    #[doc(hidden)]
    pub fn debug_dump_label_index(&self) -> Result<Vec<LabelIndexDumpRow>, TopoError> {
        use redb::ReadableTable;
        let storage = self.storage();
        let tx = storage.db.begin_read().map_err(crate::error::storage_err)?;
        let table = tx
            .open_table(crate::storage::LABEL_INDEX)
            .map_err(crate::error::storage_err)?;
        let mut out = Vec::new();
        for entry in table.iter().map_err(crate::error::storage_err)? {
            let (k, v) = entry.map_err(crate::error::storage_err)?;
            let key: [u8; 24] = k
                .value()
                .try_into()
                .map_err(|_| TopoError::Encoding("bad label_index key".into()))?;
            let label_id = u32::from_be_bytes(key[0..4].try_into().expect("4-byte slice"));
            let scope_id = u32::from_be_bytes(key[4..8].try_into().expect("4-byte slice"));
            let node_id = u128::from_be_bytes(key[8..24].try_into().expect("16-byte slice"));
            out.push((label_id, scope_id, node_id, v.value()));
        }
        out.sort_unstable();
        Ok(out)
    }

    /// Debug-only instrumentation (F8 Task 4): `true` if the LAST scope
    /// touched by the most recent `search_vector`/`search_vector_unbumped`
    /// call's `vector_store::search_scan` routing took the built-graph
    /// branch (`hnsw::search`), `false` if it took the brute-force scan
    /// branch. Backed by an `AtomicBool` on `Inner` (see its field doc), set unconditionally by
    /// BOTH routing branches on every non-candidates scope iteration — never
    /// part of replay state, never read by any production code path. Exists
    /// so tests can pin that a built cluster's search genuinely dispatches
    /// through the graph rather than vacuously agreeing with brute force by
    /// still silently scanning. `#[doc(hidden)]` — see `all_edges_between`.
    #[doc(hidden)]
    pub fn debug_last_search_used_graph(&self) -> bool {
        self.debug_atomic().load(Ordering::SeqCst)
    }

    /// Test/inspection helper (F8): every `HNSW_META` cluster-header row,
    /// decoded as `(model, scope, format, built, entry_slot, entry_level,
    /// graph_len, stale)` — the `ClusterMeta` fields in declaration order,
    /// with the `(model, scope)` key prefix split out — sorted by `(model,
    /// scope)`. `#[doc(hidden)]` — see `all_edges_between`.
    #[doc(hidden)]
    pub fn debug_dump_hnsw_meta(&self) -> Result<Vec<HnswMetaDumpRow>, TopoError> {
        use redb::ReadableTable;
        let storage = self.storage();
        let tx = storage.db.begin_read().map_err(crate::error::storage_err)?;
        let table = tx
            .open_table(crate::hnsw::HNSW_META)
            .map_err(crate::error::storage_err)?;
        let mut out = Vec::new();
        for entry in table.iter().map_err(crate::error::storage_err)? {
            let (k, v) = entry.map_err(crate::error::storage_err)?;
            let key: [u8; 8] = k
                .value()
                .try_into()
                .map_err(|_| TopoError::Encoding("bad hnsw_meta key".into()))?;
            let model = u32::from_be_bytes(key[0..4].try_into().expect("4-byte slice"));
            let scope = u32::from_be_bytes(key[4..8].try_into().expect("4-byte slice"));
            let meta: crate::hnsw::ClusterMeta =
                postcard::from_bytes(v.value()).map_err(|e| TopoError::Encoding(e.to_string()))?;
            out.push((
                model,
                scope,
                meta.format,
                meta.built,
                meta.entry_slot,
                meta.entry_level,
                meta.graph_len,
                meta.stale,
            ));
        }
        out.sort_by_key(|r| (r.0, r.1));
        Ok(out)
    }

    /// Test/inspection helper (F8): every `HNSW_LINKS` row, decoded as
    /// `(model, scope, slot, level, tomb, neighbors)` from the raw `model BE4
    /// ++ scope BE4 ++ slot BE8 ++ level` key and framed-postcard `LinkRow`
    /// value — sorted by that same `(model, scope, slot, level)` tuple, which
    /// is also on-disk key order. `#[doc(hidden)]` — see `all_edges_between`.
    #[doc(hidden)]
    pub fn debug_dump_hnsw_links(&self) -> Result<Vec<HnswLinksDumpRow>, TopoError> {
        use redb::ReadableTable;
        let storage = self.storage();
        let tx = storage.db.begin_read().map_err(crate::error::storage_err)?;
        let table = tx
            .open_table(crate::hnsw::HNSW_LINKS)
            .map_err(crate::error::storage_err)?;
        let mut out = Vec::new();
        for entry in table.iter().map_err(crate::error::storage_err)? {
            let (k, v) = entry.map_err(crate::error::storage_err)?;
            let key: [u8; 17] = k
                .value()
                .try_into()
                .map_err(|_| TopoError::Encoding("bad hnsw_links key".into()))?;
            let model = u32::from_be_bytes(key[0..4].try_into().expect("4-byte slice"));
            let scope = u32::from_be_bytes(key[4..8].try_into().expect("4-byte slice"));
            let slot = u64::from_be_bytes(key[8..16].try_into().expect("8-byte slice"));
            let level = key[16];
            let raw = crate::codec::unframe_value(v.value())?;
            let row: crate::hnsw::LinkRow =
                postcard::from_bytes(&raw).map_err(|e| TopoError::Encoding(e.to_string()))?;
            out.push((model, scope, slot, level, row.tomb, row.neighbors));
        }
        out.sort_by_key(|r| (r.0, r.1, r.2, r.3));
        Ok(out)
    }
}

/// One decoded adjacency entry from [`Db::debug_dump_adjacency`]:
/// `(slot, edge_type, is_out, target_slot, edge_slot, scope_id, valid_from,
/// valid_to)`. Plain tuples so the crate-internal `AdjEntryDisk` type is not
/// leaked through this `#[doc(hidden)]` debug seam.
#[doc(hidden)]
pub type AdjacencyDumpRow = (u64, u32, bool, u64, u64, u32, i64, Option<i64>);

/// One decoded `POSTINGS` chunk row from [`Db::debug_dump_postings`]: raw key
/// bytes paired with its decoded `(slot, tf)` entries.
#[doc(hidden)]
pub type PostingsDumpRow = (Vec<u8>, Vec<(u64, u32)>);

/// One decoded `VECTORS` row from [`Db::debug_dump_vectors`]: `(model, scope,
/// slot, vector)`.
#[doc(hidden)]
pub type VectorsDumpRow = (u32, u32, u64, Vec<f32>);

/// One decoded `EMBEDDING_REF` row from [`Db::debug_dump_embedding_ref`]:
/// `(slot, model, scope)`.
#[doc(hidden)]
pub type EmbeddingRefDumpRow = (u64, u32, u32);

/// One decoded `VECTOR_DIMS` row from [`Db::debug_dump_vector_dims`]:
/// `(model_id, dim)`.
#[doc(hidden)]
pub type VectorDimsDumpRow = (u32, u32);

/// One decoded `LABEL_INDEX` row from [`Db::debug_dump_label_index`]:
/// `(label_id, scope_id, node_id, slot)`.
#[doc(hidden)]
pub type LabelIndexDumpRow = (u32, u32, u128, u64);

/// One decoded `HNSW_META` row from [`Db::debug_dump_hnsw_meta`]: `(model,
/// scope, format, built, entry_slot, entry_level, graph_len, stale)`.
#[doc(hidden)]
pub type HnswMetaDumpRow = (u32, u32, u8, bool, u64, u8, u64, u64);

/// One decoded `HNSW_LINKS` row from [`Db::debug_dump_hnsw_links`]: `(model,
/// scope, slot, level, tomb, neighbors)`.
#[doc(hidden)]
pub type HnswLinksDumpRow = (u32, u32, u64, u8, bool, Vec<u64>);

/// The node ids that `validate::prevalidate_edge_scopes` and
/// `validate::prevalidate_create_node_ids` need pre-batch storage state for:
/// `CreateEdge`'s endpoints (scope), and every `CreateNode`'s own id
/// (existence — is this id already taken?). A same-batch `CreateNode` for a
/// `CreateEdge` endpoint id is resolved locally by `prevalidate_edge_scopes`
/// (via its own scan of `ops`) and needs no storage lookup for THAT purpose,
/// but `CreateNode` ids still need a lookup here so
/// `prevalidate_create_node_ids` can see whether the id already exists in
/// storage (or in an earlier batch's overlay, in the group path) before this
/// batch runs.
///
/// Through Task 6 this also covered `SetEmbedding`'s target and
/// `RemoveNode`'s target, for `VectorIndex::prevalidate_dims`/`maintain`
/// (deleted in Task 7 — dim validation now lives entirely inside
/// `apply_batch`'s own transaction, and there is no RAM slab left to
/// maintain), so those two arms are gone from the match below.
fn ids_needing_pre_state(ops: &[Op]) -> std::collections::HashSet<NodeId> {
    let mut ids = std::collections::HashSet::new();
    for op in ops {
        match op {
            Op::CreateEdge { from, to, .. } => {
                ids.insert(*from);
                ids.insert(*to);
            }
            Op::CreateNode { id, .. } => {
                ids.insert(*id);
            }
            _ => {}
        }
    }
    ids
}

/// Rebuilds a `TopoError` of the SAME kind as `e`, with `e`'s message folded
/// into context about the group pre-validation read that produced it.
/// `TopoError` isn't `Clone` (its `Storage` variant boxes a `redb::Error`,
/// which isn't `Clone` either), so this exists to hand every job in a failed
/// group its own error object without flattening every kind into `Rejected`
/// — see the call site in `apply_group`. `Storage`'s inner `redb::Error`
/// can't be reconstructed bit-for-bit without cloning it, so it's rebuilt as
/// a synthetic `redb::Error::Corrupted` carrying the original message; the
/// point isn't to preserve the exact redb sub-error, only the `TopoError`
/// top-level kind (`Storage` vs. `Rejected` vs. ...) that downstream
/// matches (e.g. `topodb-mcp`'s `server.rs`) actually branch on.
fn group_pre_read_error(e: &TopoError) -> TopoError {
    let msg = e.to_string();
    let ctx = |m: String| format!("group pre-validation read failed: {m}");
    match e {
        TopoError::Storage(_) => TopoError::Storage(Box::new(redb::Error::Corrupted(ctx(msg)))),
        TopoError::Busy => TopoError::Busy,
        TopoError::Encoding(m) => TopoError::Encoding(ctx(m.clone())),
        TopoError::Rejected(m) => TopoError::Rejected(ctx(m.clone())),
        TopoError::Compacted { oldest } => TopoError::Compacted { oldest: *oldest },
        TopoError::Closed => TopoError::Closed,
        TopoError::UnsupportedFormat { found, supported } => TopoError::UnsupportedFormat {
            found: *found,
            supported: *supported,
        },
    }
}

/// The single-batch apply path: pre-validate against real storage state,
/// apply, broadcast, reply. This is the ENTIRE pre-Task-6 `Job::Apply`
/// handling, unchanged — used both for a lone (non-grouped) submission and,
/// by `apply_group` below, to replay a group's batches one at a time when
/// the shared optimistic commit fails.
fn apply_one_job(
    storage: &Storage,
    subs: &Arc<Mutex<Vec<Sender<ChangeEvent>>>>,
    ops: Vec<Op>,
    now: i64,
    reply: Sender<Result<AppliedBatch, TopoError>>,
) {
    // Read pre-batch node state (scope) for every CreateEdge endpoint this
    // batch might reference, in ONE storage read, BEFORE `apply_batch` runs
    // — the input `prevalidate_edge_scopes` (below) needs.
    let pre = match storage.load_nodes(&ids_needing_pre_state(&ops)) {
        Ok(m) => m,
        Err(e) => {
            let _ = reply.send(Err(e));
            return;
        }
    };
    // Edge-scope pre-validation has the same contract: reject before
    // `apply_batch` so storage is untouched. It must not live in `apply_op`,
    // which is shared with op-log replay.
    if let Err(e) = crate::validate::prevalidate_edge_scopes(&pre, &ops) {
        let _ = reply.send(Err(e));
        return;
    }
    // Same contract as `prevalidate_edge_scopes` immediately above: reject
    // before `apply_batch` so storage is untouched, and never fold this into
    // `apply_op` (shared with replay, which must stay tolerant of a historic
    // log's duplicate-create ops).
    if let Err(e) = crate::validate::prevalidate_create_node_ids(&pre, &ops) {
        let _ = reply.send(Err(e));
        return;
    }
    match storage.apply_batch(ops, now) {
        Ok(batch) => {
            broadcast_batch(subs, &batch);
            // If the caller already dropped its reply receiver, there's
            // nothing to do with the result — move on.
            let _ = reply.send(Ok(batch));
        }
        Err(e) => {
            let _ = reply.send(Err(e));
        }
    }
}

/// Broadcasts a committed batch's ops to live subscribers. Called *after*
/// the batch has committed (so a subscriber that reacts by reading sees its
/// own event's effect). Best-effort, non-blocking: a full subscriber buffer
/// drops the event (the subscriber detects the `seq` gap and recovers via
/// `ops_since`); a disconnected receiver is pruned. The applier NEVER blocks
/// on a slow subscriber.
fn broadcast_batch(subs: &Arc<Mutex<Vec<Sender<ChangeEvent>>>>, batch: &AppliedBatch) {
    // Wrap each op in an `Arc` ONCE per op for the whole batch (not once per
    // op per subscriber) — every subscriber below then only pays for a
    // cheap `Arc::clone`.
    let ev_ops: Vec<Arc<Op>> = batch
        .resolved
        .iter()
        .map(|op| Arc::new(op.clone()))
        .collect();
    let mut subs = subs.lock().unwrap();
    subs.retain(|s| {
        for (i, ev_op) in ev_ops.iter().enumerate() {
            let ev = ChangeEvent {
                seq: batch.first_seq + i as u64,
                op: ev_op.clone(),
            };
            match s.try_send(ev) {
                Ok(()) => {}
                Err(crossbeam_channel::TrySendError::Full(_)) => {}
                Err(crossbeam_channel::TrySendError::Disconnected(_)) => return false,
            }
        }
        true
    });
}

/// Applies a drained group of 2..=16 `Job::Apply` jobs (all sharing wall
/// clock `now` — see the `Job::Apply` arm's `submit_at` note) via one
/// optimistic `Storage::apply_batches` commit (F9c: one shared fsync instead
/// of one per batch). Falls back to replaying every included job
/// individually through `apply_one_job` — the exact pre-Task-6 per-batch
/// path — whenever pre-validation or the shared commit rejects any batch in
/// the group, so per-batch atomicity is preserved: the group-commit optimism
/// costs nothing beyond one wasted attempt when it doesn't pan out.
fn apply_group(
    storage: &Storage,
    subs: &Arc<Mutex<Vec<Sender<ChangeEvent>>>>,
    jobs: Vec<ApplyJob>,
    now: i64,
) {
    // ONE storage read covering every id any batch in the group might
    // reference as a `CreateEdge` endpoint — same rationale as
    // `apply_one_job`'s pre-read, just widened to the whole group up front
    // so it's paid once instead of per batch.
    let mut all_ids = std::collections::HashSet::new();
    for (ops, _) in &jobs {
        all_ids.extend(ids_needing_pre_state(ops));
    }
    let base_pre = match storage.load_nodes(&all_ids) {
        Ok(m) => m,
        Err(e) => {
            // The read itself failed — not a per-batch validation rejection.
            // `TopoError` isn't `Clone`, so every job gets its own error
            // reconstructed from the same underlying failure rather than the
            // original value. Earlier this collapsed every kind to
            // `Rejected`, which reaches submitters (e.g. topodb-mcp's
            // `server.rs`) as an `invalid_params`-shaped client error even
            // when the real cause was a storage/IO failure — the same
            // single-batch path (`apply_one_job`, just above) forwards `e`
            // untouched, so the group path must not downgrade its kind.
            // `group_pre_read_error` keeps the top-level variant (and any
            // structured fields callers match on, like `Compacted::oldest`)
            // and only the message gains this context.
            for (_, reply) in jobs {
                let _ = reply.send(Err(group_pre_read_error(&e)));
            }
            return;
        }
    };

    // Accumulated group-scope overlay (F9c pre-validation): the
    // scope-affecting effect of every ALREADY-INCLUDED prior batch in this
    // group (`CreateNode`/`RemoveNode`), keyed by node id, layered on top of
    // `base_pre` (the real storage read taken before the group started) so
    // batch N's pre-validation sees batch N-1's same-group `CreateNode`s and
    // `RemoveNode`s without a fresh storage read. `Some(scope)` = as of this
    // point in the group, the id resolves to a node with this scope (either
    // just created by an earlier batch, or untouched and still whatever
    // `base_pre`/absence said). `None` = an earlier batch in the group
    // removed this id, so it must be treated as absent even though
    // `base_pre` still has it.
    let mut overlay: std::collections::HashMap<NodeId, Option<Scope>> =
        std::collections::HashMap::new();
    let mut included: Vec<ApplyJob> = Vec::with_capacity(jobs.len());

    for (ops, reply) in jobs {
        let mut effective_pre = base_pre.clone();
        for (&id, ov) in &overlay {
            match ov {
                Some(scope) => {
                    // Only `.scope` is ever read by `prevalidate_edge_scopes`
                    // — the rest of the record is a synthesized, unused
                    // placeholder.
                    effective_pre.insert(
                        id,
                        NodeRecord {
                            id,
                            scope: *scope,
                            label: Default::default(),
                            props: Default::default(),
                            embedding: None,
                        },
                    );
                }
                None => {
                    effective_pre.remove(&id);
                }
            }
        }

        if let Err(e) = crate::validate::prevalidate_edge_scopes(&effective_pre, &ops) {
            let _ = reply.send(Err(e));
            continue;
        }
        // Same-group defense-in-depth for Finding 1/3: `effective_pre`
        // already carries prior same-group `CreateNode`s (via `overlay`, set
        // to `Some(scope)` below) and omits prior same-group `RemoveNode`s
        // (`None`), so this sees a duplicate-id create against BOTH real
        // storage state and everything this group has done so far.
        if let Err(e) = crate::validate::prevalidate_create_node_ids(&effective_pre, &ops) {
            let _ = reply.send(Err(e));
            continue;
        }

        // Record this batch's effect in the overlay BEFORE moving `ops`
        // into `included` — subsequent batches in the group must see it.
        for op in &ops {
            match op {
                Op::CreateNode { id, scope, .. } => {
                    overlay.insert(*id, Some(*scope));
                }
                Op::RemoveNode { id } => {
                    overlay.insert(*id, None);
                }
                _ => {}
            }
        }

        included.push((ops, reply));
    }

    if included.is_empty() {
        return;
    }
    if included.len() == 1 {
        let (ops, reply) = included.pop().expect("len checked above");
        apply_one_job(storage, subs, ops, now, reply);
        return;
    }

    let groups: Vec<Vec<Op>> = included.iter().map(|(ops, _)| ops.clone()).collect();
    let results = storage.apply_batches(groups, now);

    if results.iter().all(Result::is_ok) {
        // Happy path: one shared commit landed every batch. Broadcast +
        // reply per batch, IN SUBMISSION ORDER — `included` was built (and
        // `groups` derived from it) in the order jobs were drained off the
        // channel, and `apply_batches` returns results in that same order.
        for ((_, reply), result) in included.into_iter().zip(results) {
            let batch = result.expect("checked all Ok above");
            broadcast_batch(subs, &batch);
            let _ = reply.send(Ok(batch));
        }
    } else {
        // Any batch failed inside the shared txn: `apply_batches` aborted
        // the WHOLE group — nothing in it committed (see its doc comment).
        // Discard every entry in `results` (none reflect committed state,
        // by construction) and replay each included batch individually
        // through the exact pre-Task-6 per-batch path: real storage reads,
        // real `apply_batch` calls, in submission order. Per-batch
        // atomicity is exactly what it always was; the group-commit
        // optimism just cost one wasted shared-txn attempt.
        for (ops, reply) in included {
            apply_one_job(storage, subs, ops, now, reply);
        }
    }
}

impl Drop for Inner {
    fn drop(&mut self) {
        // Shutdown order is load-bearing because the bumper thread holds a
        // *clone* of the applier `tx`. It must be, in exactly this sequence:
        //
        //   1. take+drop `bump_tx` — closes the bump channel so the bumper's
        //      `recv_timeout` loop sees `Disconnected`, does its FINAL flush
        //      (enqueuing one last `Job::BumpCounters` into the applier
        //      channel), and returns.
        //   2. join the bumper — waits for that final flush to be enqueued and
        //      for the bumper's clone of the applier `tx` to be dropped.
        //   3. take+drop `tx` — only now, with the bumper's clone gone, does
        //      the applier channel actually close.
        //   4. join the applier — its `rx.recv()` loop finally sees the closed
        //      channel (after draining the final flush) and exits.
        //
        // Reorder these and you either deadlock (drop `tx` while the bumper's
        // clone keeps the applier channel open → applier join hangs) or lose
        // the final flush (join applier before the bumper has enqueued it).
        // Shutdown must proceed even if a mutex was poisoned by an applier panic
        // (poisoned-lock policy, see vector.rs) — otherwise the host leaks the
        // applier/bumper threads on drop. Recover the guard via `into_inner`.
        self.bump_tx
            .lock()
            .unwrap_or_else(std::sync::PoisonError::into_inner)
            .take();
        if let Some(h) = self
            .bumper
            .lock()
            .unwrap_or_else(std::sync::PoisonError::into_inner)
            .take()
        {
            let _ = h.join();
        }
        self.tx
            .lock()
            .unwrap_or_else(std::sync::PoisonError::into_inner)
            .take();
        if let Some(h) = self
            .applier
            .lock()
            .unwrap_or_else(std::sync::PoisonError::into_inner)
            .take()
        {
            let _ = h.join();
        }
    }
}

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

    #[test]
    fn dropped_receiver_is_pruned_on_next_broadcast() {
        let dir = tempfile::tempdir().unwrap();
        let db = Db::open(dir.path().join("t.redb")).unwrap();
        let rx = db.subscribe(4);
        drop(rx);
        db.submit(vec![crate::Op::CreateNode {
            id: crate::NodeId::new(),
            scope: crate::Scope::Id(crate::ScopeId::new()),
            label: "M".into(),
            props: Default::default(),
        }])
        .unwrap();
        assert_eq!(
            db.inner.subs.lock().unwrap().len(),
            0,
            "disconnected sender must be pruned"
        );
    }

    #[test]
    fn meta_roundtrip_survives_reopen() {
        let dir = tempfile::tempdir().unwrap();
        let path = dir.path().join("m.redb");
        {
            let db = Db::open_with(&path, IndexSpec::default()).unwrap();
            assert_eq!(db.get_meta("onboarding:test").unwrap(), None);
            db.set_meta("onboarding:test", b"hello").unwrap();
            assert_eq!(
                db.get_meta("onboarding:test").unwrap().as_deref(),
                Some(&b"hello"[..])
            );
        }
        // reopen — value persisted
        let db = Db::open_stored(&path).unwrap();
        assert_eq!(
            db.get_meta("onboarding:test").unwrap().as_deref(),
            Some(&b"hello"[..])
        );
    }
}