tsoracle-driver-openraft 2.1.0

//
//  ░▀█▀░█▀▀░█▀█░█▀▄░█▀█░█▀▀░█░░░█▀▀
//  ░░█░░▀▀█░█░█░█▀▄░█▀█░█░░░█░░░█▀▀
//  ░░▀░░▀▀▀░▀▀▀░▀░▀░▀░▀░▀▀▀░▀▀▀░▀▀▀
//
//  tsoracle — Distributed Timestamp Oracle
//  https://www.tsoracle.rs
//
//  Copyright (c) 2026 Prisma Risk
//
//  Licensed under the Apache License, Version 2.0 (the "License");
//  you may not use this file except in compliance with the License.
//  You may obtain a copy of the License at
//
//      https://www.apache.org/licenses/LICENSE-2.0
//
//  Unless required by applicable law or agreed to in writing, software
//  distributed under the License is distributed on an "AS IS" BASIS,
//  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
//  See the License for the specific language governing permissions and
//  limitations under the License.
//

// #[PerformanceCriticalPath]
//! `RaftStateMachine` for the high-water counter with pluggable snapshot
//! persistence.
//!
//! State is one `u64` plus the openraft-required apply-progress metadata
//! (`last_applied`, `last_membership`). Snapshots are encoded with the toolkit's
//! version-prefixed codec ([`tsoracle_openraft_toolkit::encode`] /
//! [`tsoracle_openraft_toolkit::decode`] at the active write version) — a leading
//! version byte followed by the postcard body — written through to a
//! [`SnapshotStore`] so they survive a
//! process restart. The default store is in-memory
//! ([`HighWaterStateMachine::new`]); production deployments construct via
//! [`HighWaterStateMachine::with_store`] and supply a durable backend such as
//! [`crate::snapshot_store::RocksdbSnapshotStore`].
//!
//! All state lives inside an `Arc<Mutex<Core>>` so the state machine is
//! cheaply cloneable — required because openraft's
//! `RaftStateMachine::SnapshotBuilder = Self` design hands out clones to drive
//! `build_snapshot` concurrently with `apply`. The `Arc<dyn SnapshotStore>` is
//! shared the same way, so every clone sees the same persisted snapshot.

use std::io;
use std::io::Cursor;
use std::sync::Arc;

use futures::StreamExt;
use openraft::EntryPayload;
use openraft::RaftSnapshotBuilder;
use openraft::StoredMembership;
use openraft::storage::EntryResponder;
use openraft::storage::RaftStateMachine;
use openraft::storage::Snapshot;
use openraft::type_config::alias::{LogIdOf, SnapshotMetaOf, SnapshotOf, StoredMembershipOf};
use parking_lot::Mutex;
use serde::{Deserialize, Serialize};

use tsoracle_codec::{
    VersionedCodec, decode_framed, decode_postcard_exact, encode_framed, encode_postcard,
};
use tsoracle_openraft_toolkit::{
    ActiveWriteVersion, BASELINE_WRITE_VERSION, BATCH_WRITE_VERSION, DENSE_WRITE_VERSION,
    MAX_READABLE_VERSION, MIN_READABLE_VERSION, codec_io_error,
};

use crate::log_entry::{HighWaterCommand, SetFormatVersionPayload};
use crate::snapshot_store::{InMemorySnapshotStore, SnapshotStore};
use crate::type_config::{ApplyOutcome, HighWaterApplied, TypeConfig};

type LogId = LogIdOf<TypeConfig>;
type SnapMeta = SnapshotMetaOf<TypeConfig>;
type SnapOf = SnapshotOf<TypeConfig>;
type SnapData = Cursor<Vec<u8>>;
type StoredMem = StoredMembershipOf<TypeConfig>;

/// Snapshot payload. The persisted/streamed bytes are version-framed as
/// `[active write version | postcard(Self)]`, so decode them through the
/// toolkit `decode_framed` over the readable range
/// `[MIN_READABLE_VERSION, MAX_READABLE_VERSION]` rather than raw
/// `postcard::from_bytes`.
///
/// Exposed at the crate root so callers building tooling around the snapshot
/// format (e.g. inspectors, migration tools) can decode it without re-deriving
/// the layout.
#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
pub struct HighWaterStateMachineSnapshot {
    pub current_value: u64,
    pub last_applied: Option<LogId>,
    pub last_membership: StoredMem,
    /// Per-key dense counters (v5+). Empty in a v4 snapshot (pre-dense); the
    /// restore path keeps the genesis cap seeded from the constructor.
    pub dense: std::collections::BTreeMap<String, u64>,
    /// Genesis cardinality cap (v5+). Zero in a v4 snapshot ("absent
    /// sentinel"); the restore path keeps the cap seeded from the constructor.
    pub dense_cap: u64,
}

/// The v4 on-disk snapshot layout, frozen. Used only to decode
/// `BASELINE_WRITE_VERSION` bytes written before the dense fields existed;
/// converted into the current payload (empty dense map, sentinel cap 0 =
/// "absent") on decode. Do not edit its field set — it must remain
/// byte-identical to the pre-dense layout.
#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
struct HighWaterStateMachineSnapshotV4 {
    current_value: u64,
    last_applied: Option<LogId>,
    last_membership: StoredMem,
}

/// On-disk envelope written to the [`SnapshotStore`]: pairs the openraft
/// snapshot meta with the version-framed payload, where `data` holds
/// `[active write version | postcard(HighWaterStateMachineSnapshot)]`. Kept
/// private — embedders that need to inspect persisted snapshots decode the
/// inner `data` blob through the toolkit `decode_framed` over the readable
/// range.
#[derive(Debug, Clone, Serialize, Deserialize)]
struct PersistedSnapshot {
    meta: SnapMeta,
    data: Vec<u8>,
}

impl VersionedCodec for HighWaterStateMachineSnapshot {
    fn decode_version(version: u8, body: &[u8]) -> Result<Self, tsoracle_codec::CodecError> {
        match version {
            // v4 (BASELINE_WRITE_VERSION): decode the frozen old layout, lift
            // into the current payload. Pre-dense snapshots carry no dense
            // bytes; dense is empty and cap is the sentinel 0 ("absent") — the
            // restore path keeps the genesis cap seeded from the constructor.
            v if v == BASELINE_WRITE_VERSION => {
                let old: HighWaterStateMachineSnapshotV4 = decode_postcard_exact(body)?;
                Ok(HighWaterStateMachineSnapshot {
                    current_value: old.current_value,
                    last_applied: old.last_applied,
                    last_membership: old.last_membership,
                    dense: std::collections::BTreeMap::new(),
                    dense_cap: 0,
                })
            }
            // v5 (DENSE_WRITE_VERSION): the current layout decodes directly.
            // Ungated — production reads dense snapshots from v5-activated
            // clusters without any feature flag.
            v if v == DENSE_WRITE_VERSION => decode_postcard_exact(body),
            // v6 (BATCH_WRITE_VERSION): snapshot layout is byte-identical to v5;
            // batch adds a log command, not snapshot state, so the same decoder
            // applies.
            v if v == BATCH_WRITE_VERSION => decode_postcard_exact(body),
            other => Err(tsoracle_codec::CodecError::VersionUnsupported {
                min: MIN_READABLE_VERSION,
                max: MAX_READABLE_VERSION,
                actual: other,
            }),
        }
    }

    fn encode_version(&self, version: u8) -> Result<Vec<u8>, tsoracle_codec::CodecError> {
        match version {
            // v4: project to the frozen old layout (dense is always empty before
            // activation, so dropping it is lossless). This is what
            // build_snapshot emits while the active write version is still 4.
            v if v == BASELINE_WRITE_VERSION => {
                // Fail loud in ALL build profiles (not just debug): a non-empty
                // dense map at this point means write-version 5 data is being
                // encoded into a v4 snapshot, which would silently drop it.
                // The invariant (dense only populated after activation flips the
                // write version to 5) is maintained by the rollout gate, so
                // reaching here with a non-empty map indicates a protocol
                // violation that must be surfaced rather than silently lost.
                if !self.dense.is_empty() {
                    return Err(tsoracle_codec::CodecError::NotRepresentable { version });
                }
                encode_postcard(&HighWaterStateMachineSnapshotV4 {
                    current_value: self.current_value,
                    last_applied: self.last_applied,
                    last_membership: self.last_membership.clone(),
                })
            }
            // v5: the current (dense) layout encodes directly. Ungated.
            v if v == DENSE_WRITE_VERSION => encode_postcard(self),
            // v6 (BATCH_WRITE_VERSION): snapshot layout is byte-identical to v5;
            // batch adds a log command, not snapshot state, so the same encoder
            // applies.
            v if v == BATCH_WRITE_VERSION => encode_postcard(self),
            other => Err(tsoracle_codec::CodecError::VersionUnsupported {
                min: MIN_READABLE_VERSION,
                max: MAX_READABLE_VERSION,
                actual: other,
            }),
        }
    }
}

impl VersionedCodec for PersistedSnapshot {
    fn decode_version(version: u8, body: &[u8]) -> Result<Self, tsoracle_codec::CodecError> {
        match version {
            // v4: envelope layout unchanged — opaque `data` bytes hold the
            // version-framed payload.
            v if v == BASELINE_WRITE_VERSION => decode_postcard_exact(body),
            // v5: envelope layout still unchanged (only the payload data blob
            // inside gains dense bytes). Ungated — production reads v5
            // envelopes once activated.
            v if v == DENSE_WRITE_VERSION => decode_postcard_exact(body),
            // v6 (BATCH_WRITE_VERSION): envelope layout is byte-identical to v5;
            // batch adds a log command, not snapshot state, so the same decoder
            // applies.
            v if v == BATCH_WRITE_VERSION => decode_postcard_exact(body),
            other => Err(tsoracle_codec::CodecError::VersionUnsupported {
                min: MIN_READABLE_VERSION,
                max: MAX_READABLE_VERSION,
                actual: other,
            }),
        }
    }

    fn encode_version(&self, version: u8) -> Result<Vec<u8>, tsoracle_codec::CodecError> {
        match version {
            v if v == BASELINE_WRITE_VERSION => encode_postcard(self),
            // v5: envelope layout unchanged; real ungated arm for production.
            v if v == DENSE_WRITE_VERSION => encode_postcard(self),
            // v6 (BATCH_WRITE_VERSION): envelope layout is byte-identical to v5;
            // batch adds a log command, not snapshot state, so the same encoder
            // applies.
            v if v == BATCH_WRITE_VERSION => encode_postcard(self),
            other => Err(tsoracle_codec::CodecError::VersionUnsupported {
                min: MIN_READABLE_VERSION,
                max: MAX_READABLE_VERSION,
                actual: other,
            }),
        }
    }
}

/// Mutable core guarded by a single mutex. `parking_lot::Mutex` because the
/// apply path is hot and the state is fully re-derivable from the raft log
/// (poison semantics would just mask the originating panic).
struct Core {
    current_value: u64,
    /// Per-key dense counters. Deterministic iteration (BTreeMap) so snapshots
    /// are byte-identical across replicas.
    dense: std::collections::BTreeMap<String, u64>,
    /// Immutable, cluster-wide genesis cardinality cap. Seeded from the SM
    /// constructor (identical config on every node) or restored from a
    /// snapshot; never mutated at runtime. Stored in the replicated snapshot so
    /// replay/restore reproduces the same accept/reject decisions (spec §6.2).
    dense_cap: u64,
    last_applied: Option<LogId>,
    last_membership: StoredMem,
    /// Snapshot index counter, used to make snapshot ids unique even when two
    /// snapshots are produced at the same `last_applied` log id.
    snapshot_idx: u64,
    /// The most-recently built or installed snapshot, retained in memory so
    /// `get_current_snapshot` does not need to rebuild on every call.
    current_snapshot: Option<StoredSnapshot>,
}

#[derive(Clone)]
struct StoredSnapshot {
    meta: SnapMeta,
    data: Vec<u8>,
}

/// Whether a snapshot at `incoming` may replace one currently published at
/// `published` without regressing the applied log id.
///
/// Equal ids may replace (a fresh rebuild at the same `last_applied` carries
/// the same state, only a new `snapshot_id`). `None` — no prior snapshot, or a
/// pre-genesis snapshot — is the minimum, so a `None` incoming never displaces
/// a `Some` published. This is the monotone guard that closes the
/// `build_snapshot`/`install_snapshot` publish TOCTOU.
fn supersedes_published(incoming: Option<LogId>, published: Option<LogId>) -> bool {
    incoming >= published
}

/// `RaftStateMachine` for the high-water counter, with pluggable snapshot
/// persistence.
///
/// Clone-cheap: the `Arc<Mutex<Core>>`, the `Arc<dyn SnapshotStore>`, and the
/// `Arc<Mutex<()>>` persist lock are all shared by clone. Required by
/// openraft's `get_snapshot_builder` contract which uses
/// `SnapshotBuilder = Self`.
pub struct HighWaterStateMachine {
    core: Arc<Mutex<Core>>,
    store: Arc<dyn SnapshotStore>,
    /// Serializes the persist-then-publish sequence of `build_snapshot` and
    /// `install_snapshot` across all clones. openraft drives `build_snapshot`
    /// on a clone concurrently with `install_snapshot` on the main handle;
    /// without this lock a slow build could write the store and publish
    /// `current_snapshot` *after* a newer install, rolling both the durable
    /// and in-memory snapshot back to a stale `last_log_id`. Distinct from
    /// `core` precisely so the hot `apply` path never serializes against
    /// snapshot I/O — `core` is never held across `store.save`.
    persist: Arc<Mutex<()>>,
    /// Shared active write version: the version this SM stamps onto snapshots
    /// it builds and installs. Shares the one cell with the log store (and,
    /// in a later phase, the wire sender) so all writers emit the same
    /// version. Mutated only by a successful activation apply (later phase);
    /// defaults to BASELINE.
    active_write_version: ActiveWriteVersion,
}

impl Clone for HighWaterStateMachine {
    fn clone(&self) -> Self {
        Self {
            core: Arc::clone(&self.core),
            store: Arc::clone(&self.store),
            persist: Arc::clone(&self.persist),
            active_write_version: self.active_write_version.clone(),
        }
    }
}

impl Default for HighWaterStateMachine {
    fn default() -> Self {
        Self::new()
    }
}

impl HighWaterStateMachine {
    /// Build a state machine backed by an in-memory snapshot store.
    ///
    /// Equivalent to
    /// `HighWaterStateMachine::with_store(Arc::new(InMemorySnapshotStore::new()))`.
    /// Snapshots survive within a process but not across restarts; for that
    /// use [`HighWaterStateMachine::with_store`] with a durable backend such
    /// as [`crate::snapshot_store::RocksdbSnapshotStore`].
    #[expect(
        clippy::expect_used,
        reason = "`with_store` only fails when `store.load()` returns Err; \
                  `InMemorySnapshotStore::load` is `Ok(None)` for a fresh \
                  store, so this branch is unreachable."
    )]
    pub fn new() -> Self {
        let store: Arc<dyn SnapshotStore> = Arc::new(InMemorySnapshotStore::new());
        Self::with_store(store).expect("InMemorySnapshotStore::load is infallible")
    }

    /// Build a state machine with a custom genesis cardinality cap, backed by
    /// an in-memory snapshot store. The cap is immutable after construction and
    /// must be identical on every cluster member. Intended for tests that need
    /// a small cap to exercise `DenseCardinalityExceeded`; production callers
    /// use [`new`](Self::new) which defaults to
    /// [`DEFAULT_DENSE_CARDINALITY_CAP`](crate::DEFAULT_DENSE_CARDINALITY_CAP).
    #[cfg(test)]
    #[expect(
        clippy::expect_used,
        reason = "`new_with_dense_cap` only fails when `store.load()` returns Err; \
                  `InMemorySnapshotStore::load` is `Ok(None)` for a fresh \
                  store, so this branch is unreachable."
    )]
    fn new_with_dense_cap(dense_cap: u64) -> Self {
        let store: Arc<dyn SnapshotStore> = Arc::new(InMemorySnapshotStore::new());
        let sm = Self::with_store(store).expect("InMemorySnapshotStore::load is infallible");
        sm.core.lock().dense_cap = dense_cap;
        sm
    }

    /// Build a state machine backed by `store` and rehydrated from whatever
    /// snapshot the store currently holds.
    ///
    /// If `store.load()` returns a snapshot, the state machine's
    /// `current_value`, `last_applied`, `last_membership`, dense state, and
    /// `get_current_snapshot()` all reflect that snapshot on return. This is
    /// the contract that lets openraft re-enable the default snapshot policy:
    /// after a restart the log store's `last_purged_log_id` may sit above
    /// index 0, and the state machine must already cover those purged entries
    /// or openraft will panic during recovery.
    ///
    /// The active write version is recovered to at least the persisted
    /// snapshot's leading version byte, so a store holding a v5/dense snapshot
    /// does not come back at the baseline (which would trip the fail-loud
    /// `NotRepresentable` guard on the next `build_snapshot`). For full
    /// max-of-evidence recovery — folding in the highest version byte among
    /// durable *log* records as well — seed the cell via
    /// [`recover_active_write_version`](tsoracle_openraft_toolkit::recover_active_write_version)
    /// and pass it to [`with_store_and_active_version`](Self::with_store_and_active_version);
    /// `with_store` only sees the snapshot store.
    pub fn with_store(store: Arc<dyn SnapshotStore>) -> io::Result<Self> {
        Self::with_store_and_active_version(store, ActiveWriteVersion::default())
    }

    /// Build a state machine backed by `store`, sharing `active_write_version`
    /// with the log store (and, in a later phase, the wire sender). Bootstrap
    /// constructs and seeds the cell once, then threads the same clone here
    /// and into the log store; non-bootstrap callers use
    /// [`with_store`](Self::with_store), which supplies a fresh BASELINE cell.
    pub fn with_store_and_active_version(
        store: Arc<dyn SnapshotStore>,
        active_write_version: ActiveWriteVersion,
    ) -> io::Result<Self> {
        let mut core = Core {
            current_value: 0,
            dense: std::collections::BTreeMap::new(),
            dense_cap: crate::DEFAULT_DENSE_CARDINALITY_CAP,
            last_applied: None,
            last_membership: StoredMembership::default(),
            snapshot_idx: 0,
            current_snapshot: None,
        };
        if let Some(bytes) = store.load()? {
            let persisted: PersistedSnapshot =
                decode_framed(MIN_READABLE_VERSION, MAX_READABLE_VERSION, &bytes)
                    .map_err(|e| codec_io_error("persisted snapshot envelope decode", e))?;
            let payload: HighWaterStateMachineSnapshot =
                decode_framed(MIN_READABLE_VERSION, MAX_READABLE_VERSION, &persisted.data)
                    .map_err(|e| codec_io_error("persisted snapshot payload decode", e))?;
            core.current_value = payload.current_value;
            core.last_applied = payload.last_applied;
            core.last_membership = payload.last_membership;
            core.dense = payload.dense;
            // A v4 snapshot carries dense_cap == 0 (absent sentinel); keep the
            // constructor-seeded genesis cap. A v5 snapshot carries the real cap.
            if payload.dense_cap != 0 {
                core.dense_cap = payload.dense_cap;
            }
            core.current_snapshot = Some(StoredSnapshot {
                meta: persisted.meta,
                data: persisted.data,
            });
            // Recover the active write version from the persisted envelope's
            // leading version byte, mirroring `install_snapshot`'s
            // max-of-evidence rule (and `recover_active_write_version`'s).
            // Without this, the bare `with_store` path — which hands in a fresh
            // BASELINE cell — would rehydrate a populated dense map (from a v5
            // snapshot) while leaving the cell at v4, then trip the fail-loud
            // `NotRepresentable` guard on its next `build_snapshot`. The byte is
            // a safe floor: a v5 snapshot proves a `SetFormatVersion` to at
            // least that version already committed cluster-wide. Only ever
            // advances, so a caller (e.g. standalone bootstrap) that already
            // seeded the cell from richer log+snapshot evidence is unaffected.
            // `decode_framed` accepted the byte above, so it is within
            // `[MIN, MAX_READABLE_VERSION]` and never exceeds what this binary
            // can write.
            if let Some(&snapshot_version) = bytes.first() {
                if snapshot_version > active_write_version.get() {
                    active_write_version.set(snapshot_version);
                }
            }
        }
        let state_machine = Self {
            core: Arc::new(Mutex::new(core)),
            store,
            persist: Arc::new(Mutex::new(())),
            active_write_version,
        };
        // Format-migration observability: publish the compile-time readable
        // bounds and the recovered active write version so a freshly-started
        // node reports its version state before any apply runs.
        crate::observability::record_readable_bounds(MIN_READABLE_VERSION, MAX_READABLE_VERSION);
        crate::observability::record_active_write_version(state_machine.active_write_version());
        Ok(state_machine)
    }

    /// The version this SM currently stamps onto snapshots.
    pub fn active_write_version(&self) -> u8 {
        self.active_write_version.get()
    }

    /// Read the current high-water value without going through raft.
    ///
    /// Returns the value most-recently written by `apply` or
    /// `install_snapshot`. This is a state-machine-local read; callers that
    /// need linearizability must coordinate a read barrier through `Raft`
    /// before calling.
    pub async fn current_value(&self) -> u64 {
        self.core.lock().current_value
    }

    /// Read a dense counter by key, returning 0 if absent. State-machine-local
    /// read; callers needing linearizability must issue a read barrier first.
    pub fn dense_value(&self, key: &str) -> u64 {
        self.core.lock().dense.get(key).copied().unwrap_or(0)
    }

    fn snapshot_id_for(last_applied: Option<&LogId>, idx: u64) -> String {
        let log_index = last_applied.map(|l| l.index).unwrap_or(0);
        format!("{log_index}-{idx}")
    }

    /// Durably persist `envelope`, then publish `(meta, data)` as the current
    /// in-memory snapshot, running `on_adopt` against the core under the same
    /// publish lock — but only if `meta.last_log_id` does not regress the
    /// snapshot already published.
    ///
    /// The whole sequence is serialized across clones by `self.persist` so a
    /// slow `build_snapshot` cannot interleave its store write or its publish
    /// behind a newer `install_snapshot` (or vice-versa) and roll the durable
    /// and in-memory snapshot back to a stale `last_log_id`. The published
    /// `last_log_id` is re-read *inside* the lock, immediately before the
    /// store write, so the decision can't be invalidated by a concurrent
    /// commit. `self.core` is taken only in brief bursts and never held across
    /// `store.save`, so `apply` never serializes against snapshot I/O.
    ///
    /// Returns `Ok(true)` if the snapshot was adopted, `Ok(false)` if it was
    /// dropped as stale because a newer snapshot is already durable and
    /// published.
    fn commit_snapshot(
        &self,
        meta: SnapMeta,
        data: Vec<u8>,
        envelope: &[u8],
        on_adopt: impl FnOnce(&mut Core),
    ) -> io::Result<bool> {
        let _persist = self.persist.lock();

        let published = self
            .core
            .lock()
            .current_snapshot
            .as_ref()
            .and_then(|s| s.meta.last_log_id);
        if !supersedes_published(meta.last_log_id, published) {
            // `debug!`, not `warn!`: this file is on the per-entry apply hot
            // path (`#[PerformanceCriticalPath]`), where info-or-higher logging
            // is banned. A discarded stale publish is also a benign, expected
            // race resolution — the monotone gate doing its job — not an
            // operational fault, so debug is the right level on the merits too.
            tracing::debug!(
                incoming.last_log_id = ?meta.last_log_id,
                published.last_log_id = ?published,
                snapshot_id = %meta.snapshot_id,
                "discarding stale snapshot publish: a newer snapshot is \
                 already durable and published",
            );
            return Ok(false);
        }

        self.store.save(envelope)?;

        let mut core = self.core.lock();
        on_adopt(&mut core);
        core.current_snapshot = Some(StoredSnapshot { meta, data });
        Ok(true)
    }
}

impl RaftSnapshotBuilder<TypeConfig> for HighWaterStateMachine {
    async fn build_snapshot(&mut self) -> Result<SnapOf, io::Error> {
        // Build the payload + meta under the lock, then release it before
        // calling the store: `SnapshotStore::save` may do disk I/O (rocksdb
        // sync write), and holding a `parking_lot::Mutex` across that would
        // serialize `apply` against snapshot persistence unnecessarily.
        let (snapshot_payload, meta) = {
            let mut core = self.core.lock();
            core.snapshot_idx += 1;
            let payload = HighWaterStateMachineSnapshot {
                current_value: core.current_value,
                last_applied: core.last_applied,
                last_membership: core.last_membership.clone(),
                dense: core.dense.clone(),
                dense_cap: core.dense_cap,
            };
            let snapshot_id = Self::snapshot_id_for(core.last_applied.as_ref(), core.snapshot_idx);
            let meta = SnapMeta {
                last_log_id: core.last_applied,
                last_membership: core.last_membership.clone(),
                snapshot_id,
            };
            let bytes = encode_framed(self.active_write_version.get(), &payload)
                .map_err(|e| codec_io_error("snapshot payload serialize", e))?;
            (bytes, meta)
        };

        let persisted = PersistedSnapshot {
            meta: meta.clone(),
            data: snapshot_payload.clone(),
        };
        let envelope = encode_framed(self.active_write_version.get(), &persisted)
            .map_err(|e| codec_io_error("snapshot envelope serialize", e))?;
        // Persist + publish through the monotone, serialized commit path. The
        // store write happens BEFORE `current_snapshot` is updated, so a
        // failed write leaves the prior (already-persisted) snapshot in place
        // and a follower streaming via `get_current_snapshot` never observes a
        // snapshot we could not durably write. A build whose `last_applied`
        // was captured before a newer install bumped the published snapshot is
        // dropped here rather than rolling the durable + in-memory snapshot
        // back to its stale `last_log_id`. We still return the snapshot we
        // built — openraft asked for one, and the bytes faithfully capture the
        // state we observed.
        self.commit_snapshot(meta.clone(), snapshot_payload.clone(), &envelope, |_| {})?;

        Ok(Snapshot {
            meta,
            snapshot: Cursor::new(snapshot_payload),
        })
    }
}

impl RaftStateMachine<TypeConfig> for HighWaterStateMachine {
    type SnapshotBuilder = Self;

    async fn applied_state(&mut self) -> Result<(Option<LogId>, StoredMem), io::Error> {
        let core = self.core.lock();
        Ok((core.last_applied, core.last_membership.clone()))
    }

    async fn apply<Strm>(&mut self, mut entries: Strm) -> Result<(), io::Error>
    where
        Strm: futures::Stream<Item = Result<EntryResponder<TypeConfig>, io::Error>>
            + Unpin
            + openraft::OptionalSend,
    {
        while let Some(item) = entries.next().await {
            let (entry, responder_opt) = item?;
            let log_id = entry.log_id;

            let applied = match &entry.payload {
                EntryPayload::Blank => {
                    let mut core = self.core.lock();
                    core.last_applied = Some(log_id);
                    HighWaterApplied {
                        value: core.current_value,
                        outcome: ApplyOutcome::Advanced,
                    }
                }
                EntryPayload::Normal(HighWaterCommand::Advance(advance)) => {
                    let mut core = self.core.lock();
                    core.current_value = advance.merge(core.current_value);
                    core.last_applied = Some(log_id);
                    HighWaterApplied {
                        value: core.current_value,
                        outcome: ApplyOutcome::Advanced,
                    }
                }
                EntryPayload::Normal(HighWaterCommand::SetFormatVersion(
                    SetFormatVersionPayload {
                        target,
                        gated_members,
                    },
                )) => {
                    let target = *target;
                    // Defense-in-depth ahead of the subset check: refuse to
                    // flip the cell to a target outside the LOCAL binary's
                    // readable range. The gate at proposal time
                    // (`target_in_local_readable_range` in `capabilities`)
                    // is the primary safety mechanism, but the apply arm
                    // re-checks here so an older (buggy) binary's committed
                    // out-of-range entry cannot poison a fixed binary's
                    // active write version on replay. Reaching this branch
                    // means a protocol invariant was already broken; the
                    // apply pipeline still must not panic (that would tear
                    // down the leader) — emit a NO-OP outcome and a
                    // distinct counter so the operator sees it.
                    if !(MIN_READABLE_VERSION..=MAX_READABLE_VERSION).contains(&target) {
                        // `debug!` not `error!`: hot path, same rule as the
                        // subset arms below. The operator-visible surface
                        // is the `noop_target_out_of_range` counter — a
                        // non-zero value means an older binary committed
                        // an out-of-range bump (or a peer violated
                        // protocol) and the apply arm's defense-in-depth
                        // contained it. The gate's `warn!` at
                        // `run_activation_gate` is the loud, operator-
                        // facing surface for the in-process path; this
                        // branch only fires on replay of an already-
                        // committed entry, where loud logging on every
                        // restart would just spam.
                        tracing::debug!(
                            target = target,
                            min = MIN_READABLE_VERSION,
                            max = MAX_READABLE_VERSION,
                            "SetFormatVersion apply rejected: target outside local readable range"
                        );
                        crate::observability::record_noop_target_out_of_range();
                        let mut core = self.core.lock();
                        core.last_applied = Some(log_id);
                        HighWaterApplied {
                            value: core.current_value,
                            outcome: ApplyOutcome::FormatActivationTargetOutOfRange { target },
                        }
                    } else {
                        // Evaluate the subset against the membership committed as
                        // of this entry's log position. Apply folds the log in
                        // index order, so `core.last_membership` is exactly that
                        // membership (a `Membership` entry at a lower index has
                        // already updated it; a later one has not). Cover BOTH
                        // voters and learners: openraft replicates/snapshots
                        // learners, so an un-gated learner must force a no-op.
                        let mut core = self.core.lock();
                        let committed_members: std::collections::BTreeSet<u64> = core
                            .last_membership
                            .membership()
                            .nodes()
                            .map(|(node_id, _node)| *node_id)
                            .collect();
                        let outcome = if committed_members.is_subset(gated_members) {
                            // Successful (non-no-op) apply: set the
                            // process-shared active-write-version cell. NO meta
                            // write — durability is the raft log (deterministic
                            // replay re-runs this exact check) plus the snapshot
                            // frame byte.
                            self.active_write_version.set(target);
                            // `debug!` not `info!`: this is on the per-entry apply
                            // hot path (this file carries the
                            // `#[PerformanceCriticalPath]` marker), so
                            // info-or-higher logging is banned. The
                            // operator-visible surface is the `applied` counter
                            // and the `active_write_version` gauge.
                            tracing::debug!(
                                target = target,
                                gated_members = ?gated_members,
                                committed_members = ?committed_members,
                                "SetFormatVersion applied: active write version flipped"
                            );
                            crate::observability::record_applied();
                            crate::observability::record_active_write_version(target);
                            ApplyOutcome::FormatActivated { target }
                        } else {
                            // Membership grew an un-gated member between the gate
                            // and this entry's position. Leave the cell
                            // untouched; the operator re-gates and re-issues. A
                            // no-op writes nothing at `target`, so it cannot
                            // resurrect on restart.
                            //
                            // `debug!` not `warn!`: hot path, same rule as the
                            // success arm above. The operator-visible surface is
                            // the `noop_membership_subset` counter — a non-zero
                            // value during an activation means a membership
                            // change raced the bump.
                            tracing::debug!(
                                target = target,
                                gated_members = ?gated_members,
                                committed_members = ?committed_members,
                                "SetFormatVersion no-op: committed membership is not a subset of the gated set"
                            );
                            crate::observability::record_noop_membership_subset();
                            ApplyOutcome::FormatActivationNoop { target }
                        };
                        core.last_applied = Some(log_id);
                        HighWaterApplied {
                            value: core.current_value,
                            outcome,
                        }
                    }
                }
                EntryPayload::Normal(HighWaterCommand::AdvanceDense { key, count }) => {
                    let mut core = self.core.lock();
                    let key_str = key.as_str();
                    let present = core.dense.contains_key(key_str);
                    let outcome = if !present && core.dense.len() as u64 >= core.dense_cap {
                        ApplyOutcome::DenseCardinalityExceeded {
                            cap: core.dense_cap,
                        }
                    } else {
                        let start = core.dense.get(key_str).copied().unwrap_or(0);
                        match start.checked_add(u64::from(*count)) {
                            Some(next) => {
                                core.dense.insert(key_str.to_string(), next);
                                ApplyOutcome::DenseAdvanced { start }
                            }
                            None => ApplyOutcome::DenseOverflow,
                        }
                    };
                    core.last_applied = Some(log_id);
                    HighWaterApplied {
                        value: core.current_value,
                        outcome,
                    }
                }
                EntryPayload::Normal(HighWaterCommand::AdvanceDenseBatch { entries }) => {
                    let mut core = self.core.lock();
                    // Phase 1: cardinality. Count distinct keys not already
                    // present; a duplicate entry on the same key counts once
                    // (set semantics), so the new-key count uses a BTreeSet.
                    let new_keys: std::collections::BTreeSet<&str> = entries
                        .iter()
                        .map(|e| e.key.as_str())
                        .filter(|k| !core.dense.contains_key(*k))
                        .collect();
                    let outcome =
                        if core.dense.len() as u64 + new_keys.len() as u64 > core.dense_cap {
                            ApplyOutcome::DenseBatchCardinalityExceeded {
                                cap: core.dense_cap,
                            }
                        } else {
                            // Phase 2: sequential fold into a scratch map seeded
                            // lazily from current counters. Each entry's start is the
                            // running value in the scratch map (or the stored value if
                            // not yet in the scratch map); the running value
                            // accumulates across repeats, so a duplicate key yields
                            // adjacent non-overlapping blocks and the overflow check
                            // is against the ACCUMULATED value — not the shared
                            // pre-batch counter. Any overflow rejects the whole batch.
                            let mut scratch: std::collections::BTreeMap<&str, u64> =
                                std::collections::BTreeMap::new();
                            let mut starts: Vec<u64> = Vec::with_capacity(entries.len());
                            let mut overflow = false;
                            for entry in entries {
                                let key_str = entry.key.as_str();
                                let running = scratch
                                    .get(key_str)
                                    .copied()
                                    .or_else(|| core.dense.get(key_str).copied())
                                    .unwrap_or(0);
                                starts.push(running);
                                match running.checked_add(u64::from(entry.count)) {
                                    Some(next) => {
                                        scratch.insert(key_str, next);
                                    }
                                    None => {
                                        overflow = true;
                                        break;
                                    }
                                }
                            }
                            if overflow {
                                ApplyOutcome::DenseBatchOverflow
                            } else {
                                // Phase 3: commit. Only now mutate the durable map.
                                for (key_str, next) in scratch {
                                    core.dense.insert(key_str.to_string(), next);
                                }
                                ApplyOutcome::DenseBatchAdvanced { starts }
                            }
                        };
                    core.last_applied = Some(log_id);
                    HighWaterApplied {
                        value: core.current_value,
                        outcome,
                    }
                }
                EntryPayload::Membership(membership) => {
                    let mut core = self.core.lock();
                    core.last_membership = StoredMembership::new(Some(log_id), membership.clone());
                    core.last_applied = Some(log_id);
                    HighWaterApplied {
                        value: core.current_value,
                        outcome: ApplyOutcome::Advanced,
                    }
                }
            };

            if let Some(responder) = responder_opt {
                responder.send(applied);
            }
        }
        Ok(())
    }

    async fn get_snapshot_builder(&mut self) -> Self::SnapshotBuilder {
        self.clone()
    }

    async fn begin_receiving_snapshot(&mut self) -> Result<SnapData, io::Error> {
        Ok(Cursor::new(Vec::new()))
    }

    async fn install_snapshot(
        &mut self,
        meta: &SnapMeta,
        snapshot: SnapData,
    ) -> Result<(), io::Error> {
        let bytes = snapshot.into_inner();
        let payload: HighWaterStateMachineSnapshot =
            decode_framed(MIN_READABLE_VERSION, MAX_READABLE_VERSION, &bytes)
                .map_err(|e| codec_io_error("snapshot payload decode", e))?;

        // `meta.last_log_id` is read back by `get_current_snapshot` while
        // `payload.last_applied` is read back by `applied_state`. `build_snapshot`
        // derives both from the same `last_applied`, so they always agree for an
        // honest peer. Reject any disagreement before persisting or mutating: a
        // mismatch would otherwise desync those two reader methods permanently
        // and silently. Checked before `store.save` so a rejected install leaves
        // both the store and in-memory state untouched for openraft to retry.
        if meta.last_log_id != payload.last_applied {
            return Err(io::Error::new(
                io::ErrorKind::InvalidData,
                format!(
                    "snapshot meta/payload disagree on last_log_id: \
                     meta.last_log_id={:?}, payload.last_applied={:?}",
                    meta.last_log_id, payload.last_applied
                ),
            ));
        }

        // Re-validate every dense key before adopting the snapshot. The dense
        // map is persisted as raw `String` keys, so — unlike the replicated
        // `AdvanceDense` log command, whose `SeqKey` re-validates on decode — a
        // crafted or corrupted snapshot can carry an empty or oversized key that
        // `decode_framed` above accepts. Enforce `SeqKey`'s invariant here, fail
        // loud, and do it before the `active_write_version` bump and the persist
        // below so a rejected install leaves the store and in-memory state
        // untouched for openraft to retry — exactly like the meta/payload check
        // above.
        for key in payload.dense.keys() {
            if let Err(e) = tsoracle_core::SeqKey::try_new(key.as_str()) {
                return Err(io::Error::new(
                    io::ErrorKind::InvalidData,
                    format!("snapshot dense key fails the SeqKey invariant: {e}"),
                ));
            }
        }

        // The snapshot's leading version byte is evidence that the sending
        // leader's active write version was at least this high — which means a
        // `SetFormatVersion` activation to that version already committed
        // cluster-wide (the all-members gate required every member, including
        // this one, to be able to read it). Advance our cell to match, mirroring
        // `recover_active_write_version`'s max-of-evidence rule. Without this, a
        // follower that receives a v5 snapshot whose `SetFormatVersion` entry was
        // purged into the snapshot would keep `active_write_version` at the
        // baseline while holding a populated dense map, then trip the fail-loud
        // v4-encode `NotRepresentable` guard on its next `build_snapshot`. The
        // byte is already within `[MIN, MAX_READABLE_VERSION]` (decode_framed
        // accepted it above), so this never exceeds what this binary can write.
        if let Some(&snapshot_version) = bytes.first() {
            if snapshot_version > self.active_write_version.get() {
                self.active_write_version.set(snapshot_version);
                crate::observability::record_active_write_version(snapshot_version);
            }
        }

        let persisted = PersistedSnapshot {
            meta: meta.clone(),
            data: bytes.clone(),
        };
        let envelope = encode_framed(self.active_write_version.get(), &persisted)
            .map_err(|e| codec_io_error("snapshot envelope serialize", e))?;
        // Persist, apply the snapshot's state, and publish atomically through
        // the monotone, serialized commit path — the same one `build_snapshot`
        // uses, so neither can clobber the other. The store write happens
        // before the in-memory state is advanced, so a failed write leaves the
        // SM at its prior state for openraft to retry. A stale install (a
        // lower `last_log_id` than the already-published snapshot) is dropped
        // as an accepted no-op: we already cover at least this snapshot, so
        // adopting it would regress both the durable snapshot and the applied
        // state.
        self.commit_snapshot(meta.clone(), bytes, &envelope, |core| {
            core.current_value = payload.current_value;
            core.last_applied = payload.last_applied;
            core.last_membership = payload.last_membership;
            core.dense = payload.dense;
            // A v4 snapshot carries dense_cap == 0 (absent sentinel); keep the
            // constructor-seeded genesis cap. A v5 snapshot carries the real cap.
            if payload.dense_cap != 0 {
                core.dense_cap = payload.dense_cap;
            }
        })?;
        Ok(())
    }

    async fn get_current_snapshot(&mut self) -> Result<Option<SnapOf>, io::Error> {
        let core = self.core.lock();
        Ok(core.current_snapshot.as_ref().map(|s| Snapshot {
            meta: s.meta.clone(),
            snapshot: Cursor::new(s.data.clone()),
        }))
    }
}

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

    use std::collections::BTreeSet;

    use futures::stream;
    use openraft::EntryPayload;
    use openraft::entry::RaftEntry;
    use openraft::storage::EntryResponder;
    use openraft::type_config::alias::EntryOf;

    use crate::log_entry::HighWaterCommand;
    use crate::type_config::TypeConfig;
    use tsoracle_consensus::AdvancePayload;

    // --- Test helpers ---

    /// Install a process-wide `DEBUG`-level subscriber so the `tracing::debug!`
    /// on the stale-publish reject path actually evaluates its fields (without
    /// an interested subscriber the macro short-circuits and the argument
    /// expressions are never executed). Idempotent across tests via `try_init`.
    fn enable_tracing() {
        let _ = tracing_subscriber::fmt()
            .with_max_level(tracing::Level::DEBUG)
            .with_test_writer()
            .try_init();
    }

    /// Build a `LogId` for the toolkit's default leader-id layout
    /// (`LeaderId<u64, u64>`) with term=1, node_id=1, and the given index.
    fn log_id(index: u64) -> LogIdOf<TypeConfig> {
        openraft::testing::log_id::<TypeConfig>(1, 1, index)
    }

    fn entry(
        index: u64,
        payload: EntryPayload<HighWaterCommand, u64, crate::type_config::OpenraftPeer>,
    ) -> EntryResponder<TypeConfig> {
        let e: EntryOf<TypeConfig> = match payload {
            EntryPayload::Blank => EntryOf::<TypeConfig>::new_blank(log_id(index)),
            EntryPayload::Normal(d) => EntryOf::<TypeConfig>::new_normal(log_id(index), d),
            EntryPayload::Membership(m) => EntryOf::<TypeConfig>::new_membership(log_id(index), m),
        };
        (e, None)
    }

    async fn apply_one(
        sm: &mut HighWaterStateMachine,
        index: u64,
        payload: EntryPayload<HighWaterCommand, u64, crate::type_config::OpenraftPeer>,
    ) {
        sm.apply(stream::iter([Ok(entry(index, payload))]))
            .await
            .expect("apply");
    }

    // --- Tests ---

    #[tokio::test]
    async fn apply_blank_updates_only_log_id() {
        let mut sm = HighWaterStateMachine::new();
        apply_one(&mut sm, 1, EntryPayload::Blank).await;
        assert_eq!(sm.current_value().await, 0);
        let (last, _) = sm.applied_state().await.unwrap();
        assert_eq!(last.map(|l| l.index), Some(1));
    }

    #[tokio::test]
    async fn apply_normal_advances_value() {
        let mut sm = HighWaterStateMachine::new();
        apply_one(
            &mut sm,
            1,
            EntryPayload::Normal(HighWaterCommand::Advance(AdvancePayload { at_least: 100 })),
        )
        .await;
        assert_eq!(sm.current_value().await, 100);
    }

    #[tokio::test]
    async fn apply_normal_holds_monotonic_under_stale_target() {
        let mut sm = HighWaterStateMachine::new();
        apply_one(
            &mut sm,
            1,
            EntryPayload::Normal(HighWaterCommand::Advance(AdvancePayload { at_least: 100 })),
        )
        .await;
        apply_one(
            &mut sm,
            2,
            EntryPayload::Normal(HighWaterCommand::Advance(AdvancePayload { at_least: 50 })),
        )
        .await;
        assert_eq!(sm.current_value().await, 100);
    }

    #[tokio::test]
    async fn apply_normal_equal_target_holds_value() {
        let mut sm = HighWaterStateMachine::new();
        apply_one(
            &mut sm,
            1,
            EntryPayload::Normal(HighWaterCommand::Advance(AdvancePayload { at_least: 100 })),
        )
        .await;
        apply_one(
            &mut sm,
            2,
            EntryPayload::Normal(HighWaterCommand::Advance(AdvancePayload { at_least: 100 })),
        )
        .await;
        assert_eq!(sm.current_value().await, 100);
    }

    #[tokio::test]
    async fn apply_membership_updates_membership_only() {
        let mut sm = HighWaterStateMachine::new();
        apply_one(
            &mut sm,
            1,
            EntryPayload::Normal(HighWaterCommand::Advance(AdvancePayload { at_least: 42 })),
        )
        .await;
        let mem = openraft::Membership::new_with_defaults(vec![BTreeSet::from([1u64])], [1u64]);
        apply_one(&mut sm, 2, EntryPayload::Membership(mem)).await;
        assert_eq!(sm.current_value().await, 42);
        let (last, _) = sm.applied_state().await.unwrap();
        assert_eq!(last.map(|l| l.index), Some(2));
    }

    // ---- SetFormatVersion apply tests ----
    //
    // The apply path keys the flip off a successful (non-no-op) apply: the
    // subset check `committed_members ⊆ gated_members` runs against
    // `core.last_membership` evaluated at the bump's own log position (apply
    // folds the log in index order). On a hit the shared cell is set; on a
    // miss the cell is untouched and a no-op outcome is returned. Direct
    // cell observation is sufficient since `SetFormatVersionPayload`'s effect
    // is exactly "did the cell move?"; we don't need to capture the
    // ApplyOutcome through the responder channel for these checks.

    /// A membership whose voter config is `voters`, as a `Membership` payload.
    /// The concrete `Membership<NID, N>` type is inferred from the call site
    /// (e.g. `EntryPayload::Membership(...)` infers it for `TypeConfig`).
    fn voters_membership(
        voters: &[u64],
    ) -> openraft::Membership<u64, crate::type_config::OpenraftPeer> {
        openraft::Membership::new_with_defaults(
            vec![voters.iter().copied().collect::<BTreeSet<u64>>()],
            voters.to_vec(),
        )
    }

    #[tokio::test]
    async fn set_format_version_subset_match_sets_cell() {
        // The flip is observable because DENSE_WRITE_VERSION (5) is a real
        // target distinct from BASELINE (4) and within the readable range
        // [MIN, MAX] = [4, 5]. The behavior pinned by this test — "subset OK +
        // in-range target ⇒ cell flips" — uses the real dense activation target.
        let mut sm = HighWaterStateMachine::new();
        let target = tsoracle_openraft_toolkit::DENSE_WRITE_VERSION;
        // Establish membership {1, 2} at index 1.
        apply_one(
            &mut sm,
            1,
            EntryPayload::Membership(voters_membership(&[1, 2])),
        )
        .await;
        // Bump gated on a superset of the committed membership → subset holds.
        apply_one(
            &mut sm,
            2,
            EntryPayload::Normal(HighWaterCommand::SetFormatVersion(
                SetFormatVersionPayload {
                    target,
                    gated_members: BTreeSet::from([1u64, 2u64, 3u64]),
                },
            )),
        )
        .await;
        assert_eq!(
            sm.active_write_version(),
            target,
            "cell set on successful subset apply"
        );
    }

    #[tokio::test]
    async fn set_format_version_subset_match_out_of_range_target_is_apply_noop() {
        // Mirror of the (feature-gated) "subset match" test but for the
        // default build: a subset-OK bump whose target is OUT of the
        // local readable range must be no-op'd by the apply-arm
        // defense-in-depth, regardless of subset success. Pins the
        // ordering invariant — the range check runs BEFORE the subset
        // check, so an OOR target never sneaks through on a gate
        // committed by an older buggy binary.
        let mut sm = HighWaterStateMachine::new();
        let before = sm.active_write_version();
        apply_one(
            &mut sm,
            1,
            EntryPayload::Membership(voters_membership(&[1, 2])),
        )
        .await;
        apply_one(
            &mut sm,
            2,
            EntryPayload::Normal(HighWaterCommand::SetFormatVersion(
                SetFormatVersionPayload {
                    target: 7, // out of range under default features.
                    gated_members: BTreeSet::from([1u64, 2u64, 3u64]),
                },
            )),
        )
        .await;
        assert_eq!(
            sm.active_write_version(),
            before,
            "subset-OK + OOR target ⇒ apply-arm defense-in-depth no-op"
        );
    }

    // ---- Apply-arm defense-in-depth: out-of-range target is a no-op ----
    //
    // The all-members gate at proposal time is the primary safety
    // mechanism, but the apply arm independently refuses to flip the
    // shared cell to a target outside the LOCAL binary's
    // [MIN_READABLE_VERSION, MAX_READABLE_VERSION]. This guards against
    // two paths the gate cannot:
    //
    //   1. Replay of a log committed by an older (buggy) binary that
    //      let an out-of-range target through. A fixed binary replaying
    //      that log must NOT propagate the corruption into its in-memory
    //      cell, since the codec arms reject encode/decode at that
    //      version and the cluster wedges.
    //   2. A protocol-violating peer somehow committing such an entry
    //      (e.g. a forged proposal). The fail-safe at apply contains
    //      blast radius to "the flip didn't take effect" instead of
    //      "all subsequent log appends fail".
    //
    // The behavior is a NO-OP (cell untouched + `FormatActivationTargetOutOfRange`
    // outcome), never a panic. The apply pipeline is on the critical
    // path; refusing to apply would tear the leader down rather than
    // contain the bad entry.

    #[tokio::test]
    async fn set_format_version_target_above_max_is_apply_noop() {
        let mut sm = HighWaterStateMachine::new();
        let before = sm.active_write_version();
        // Membership matches the gated set (subset OK) so the only thing
        // that can hold the flip back is the range check.
        apply_one(
            &mut sm,
            1,
            EntryPayload::Membership(voters_membership(&[1, 2])),
        )
        .await;
        let above_max = tsoracle_openraft_toolkit::MAX_READABLE_VERSION.saturating_add(1);
        if above_max == tsoracle_openraft_toolkit::MAX_READABLE_VERSION {
            return; // MAX == u8::MAX (impossible in practice).
        }
        apply_one(
            &mut sm,
            2,
            EntryPayload::Normal(HighWaterCommand::SetFormatVersion(
                SetFormatVersionPayload {
                    target: above_max,
                    gated_members: BTreeSet::from([1u64, 2u64]),
                },
            )),
        )
        .await;
        assert_eq!(
            sm.active_write_version(),
            before,
            "apply must refuse to flip to a target above MAX_READABLE_VERSION"
        );
    }

    #[tokio::test]
    async fn set_format_version_target_below_min_is_apply_noop() {
        let mut sm = HighWaterStateMachine::new();
        let before = sm.active_write_version();
        apply_one(
            &mut sm,
            1,
            EntryPayload::Membership(voters_membership(&[1, 2])),
        )
        .await;
        // MIN_READABLE_VERSION is at least 1 in any sensible build (the
        // codec's compile-time `MIN <= MAX` plus the framing layer's
        // version byte both effectively forbid MIN == 0), so MIN - 1 is
        // always representable and out of range. This is the bug class
        // the finding hit: a target below MIN that the (pre-fix) gate
        // accepted because it only checked the upper bound.
        let below_min = tsoracle_openraft_toolkit::MIN_READABLE_VERSION - 1;
        apply_one(
            &mut sm,
            2,
            EntryPayload::Normal(HighWaterCommand::SetFormatVersion(
                SetFormatVersionPayload {
                    target: below_min,
                    gated_members: BTreeSet::from([1u64, 2u64]),
                },
            )),
        )
        .await;
        assert_eq!(
            sm.active_write_version(),
            before,
            "apply must refuse to flip to a target below MIN_READABLE_VERSION"
        );
    }

    #[tokio::test]
    async fn set_format_version_target_zero_is_apply_noop() {
        // Target 0 is the worst case: 0 is also the legacy-unframed wire
        // sentinel, so a 0-stamped record collides with that
        // interpretation. The apply arm must refuse the flip rather than
        // let it through (the codec would later reject every encode).
        let mut sm = HighWaterStateMachine::new();
        let before = sm.active_write_version();
        apply_one(
            &mut sm,
            1,
            EntryPayload::Membership(voters_membership(&[1, 2])),
        )
        .await;
        apply_one(
            &mut sm,
            2,
            EntryPayload::Normal(HighWaterCommand::SetFormatVersion(
                SetFormatVersionPayload {
                    target: 0,
                    gated_members: BTreeSet::from([1u64, 2u64]),
                },
            )),
        )
        .await;
        assert_eq!(
            sm.active_write_version(),
            before,
            "apply must refuse to flip to a 0 target"
        );
    }

    #[tokio::test]
    async fn set_format_version_membership_not_subset_is_noop() {
        let mut sm = HighWaterStateMachine::new();
        let before = sm.active_write_version();
        // Membership {1, 2, 9} at index 1; the gate only covered {1, 2}.
        apply_one(
            &mut sm,
            1,
            EntryPayload::Membership(voters_membership(&[1, 2, 9])),
        )
        .await;
        apply_one(
            &mut sm,
            2,
            EntryPayload::Normal(HighWaterCommand::SetFormatVersion(
                SetFormatVersionPayload {
                    target: 7,
                    gated_members: BTreeSet::from([1u64, 2u64]),
                },
            )),
        )
        .await;
        assert_eq!(
            sm.active_write_version(),
            before,
            "no-op must leave the shared cell untouched"
        );
    }

    #[tokio::test]
    async fn set_format_version_covers_learners_not_only_voters() {
        let mut sm = HighWaterStateMachine::new();
        let before = sm.active_write_version();
        // Voter {1}, learner {2}. `new_with_defaults(voter_groups, all_ids)`
        // treats node ids in `all_ids` but absent from `voter_groups` as
        // learners. A voter-only gate `{1}` must NOT satisfy the subset —
        // the learner is a current member and openraft replicates to it,
        // so an un-gated learner forces a no-op.
        let membership =
            openraft::Membership::new_with_defaults(vec![BTreeSet::from([1u64])], [1u64, 2u64]);
        apply_one(&mut sm, 1, EntryPayload::Membership(membership)).await;
        apply_one(
            &mut sm,
            2,
            EntryPayload::Normal(HighWaterCommand::SetFormatVersion(
                SetFormatVersionPayload {
                    target: 7,
                    gated_members: BTreeSet::from([1u64]),
                },
            )),
        )
        .await;
        assert_eq!(
            sm.active_write_version(),
            before,
            "un-gated learner must force a no-op"
        );
    }

    // ---- Replay / recovery confirmation tests ----
    //
    // These model openraft's deterministic recovery: a fresh state machine
    // (fresh BASELINE-seeded cell) re-applies the committed log in index
    // order, so the SetFormatVersion subset check re-runs per entry. A
    // no-op replays as a no-op (no record was ever written at the higher
    // version); a success re-establishes the cell.

    #[derive(Clone)]
    enum ReplayEntry {
        Membership(Vec<u64>),
        Bump { target: u8, gated: Vec<u64> },
    }

    async fn replay(sm: &mut HighWaterStateMachine, log: &[(u64, ReplayEntry)]) {
        for (index, kind) in log {
            let payload = match kind {
                ReplayEntry::Membership(voters) => {
                    EntryPayload::Membership(voters_membership(voters))
                }
                ReplayEntry::Bump { target, gated } => EntryPayload::Normal(
                    HighWaterCommand::SetFormatVersion(SetFormatVersionPayload {
                        target: *target,
                        gated_members: gated.iter().copied().collect(),
                    }),
                ),
            };
            apply_one(sm, *index, payload).await;
        }
    }

    #[tokio::test]
    async fn successful_activation_survives_replay() {
        let target = tsoracle_openraft_toolkit::DENSE_WRITE_VERSION;
        let log = vec![
            (1u64, ReplayEntry::Membership(vec![1, 2])),
            (
                2u64,
                ReplayEntry::Bump {
                    target,
                    gated: vec![1, 2, 3],
                },
            ),
        ];
        // Live apply sets the cell.
        let mut live = HighWaterStateMachine::new();
        replay(&mut live, &log).await;
        assert_eq!(live.active_write_version(), target);

        // "Restart": a fresh SM with a fresh BASELINE-seeded cell, replaying
        // the same committed log, re-establishes target via the re-run
        // subset check. NO meta key consulted.
        let mut recovered = HighWaterStateMachine::new();
        assert_ne!(
            recovered.active_write_version(),
            target,
            "fresh cell starts at baseline"
        );
        replay(&mut recovered, &log).await;
        assert_eq!(
            recovered.active_write_version(),
            target,
            "replay re-applies the flip"
        );
    }

    #[tokio::test]
    async fn out_of_range_bump_replays_as_noop_across_restart() {
        // Replay-determinism contract for an out-of-range bump committed
        // by an older (buggy) binary: every replay reaches the same
        // state (cell stays at baseline) because the apply arm's range
        // check uses ONLY the local binary's constants + the entry
        // data, never timing or per-replay side state. The fix CANNOT
        // resurrect the bad flip on restart.
        let log = vec![
            (1u64, ReplayEntry::Membership(vec![1, 2])),
            (
                2u64,
                ReplayEntry::Bump {
                    target: 7, // out of range under default features.
                    gated: vec![1, 2, 3],
                },
            ),
        ];
        let mut live = HighWaterStateMachine::new();
        let baseline = live.active_write_version();
        replay(&mut live, &log).await;
        assert_eq!(
            live.active_write_version(),
            baseline,
            "live apply no-ops on OOR target"
        );

        let mut recovered = HighWaterStateMachine::new();
        replay(&mut recovered, &log).await;
        assert_eq!(
            recovered.active_write_version(),
            baseline,
            "replay is deterministic: OOR bump stays a no-op across restart"
        );
    }

    #[tokio::test]
    async fn noop_bump_never_advances_across_restart() {
        let log = vec![
            // Membership has an un-gated member 9; the gate only covered {1, 2}.
            (1u64, ReplayEntry::Membership(vec![1, 2, 9])),
            (
                2u64,
                ReplayEntry::Bump {
                    target: 7,
                    gated: vec![1, 2],
                },
            ),
        ];
        let mut live = HighWaterStateMachine::new();
        let baseline = live.active_write_version();
        replay(&mut live, &log).await;
        assert_eq!(
            live.active_write_version(),
            baseline,
            "no-op leaves the cell"
        );

        // Restart: replay the same log; the no-op replays as a no-op (the
        // subset check fails identically), so the cell stays at baseline.
        // The committed entry's mere presence never advances the version.
        let mut recovered = HighWaterStateMachine::new();
        replay(&mut recovered, &log).await;
        assert_eq!(
            recovered.active_write_version(),
            baseline,
            "a no-op'd bump cannot resurrect on restart"
        );
    }

    #[tokio::test]
    async fn replay_is_deterministic() {
        let in_range = tsoracle_openraft_toolkit::DENSE_WRITE_VERSION;
        let log = vec![
            (1u64, ReplayEntry::Membership(vec![1, 2])),
            (
                2u64,
                ReplayEntry::Bump {
                    target: in_range,
                    gated: vec![1, 2, 3],
                },
            ),
            // A later bump gated on a set that excludes a now-added member.
            (3u64, ReplayEntry::Membership(vec![1, 2, 5])),
            (
                4u64,
                ReplayEntry::Bump {
                    target: in_range,
                    gated: vec![1, 2],
                },
            ),
        ];
        let mut first = HighWaterStateMachine::new();
        replay(&mut first, &log).await;
        let mut second = HighWaterStateMachine::new();
        replay(&mut second, &log).await;
        // Index-2 bump succeeded (membership {1,2} ⊆ {1,2,3}); index-4
        // bump no-ops (membership {1,2,5} ⊄ {1,2}), so the cell stays at
        // `in_range`.
        assert_eq!(first.active_write_version(), in_range);
        assert_eq!(
            first.active_write_version(),
            second.active_write_version(),
            "two replays of the same committed log yield the same cell"
        );
    }

    #[tokio::test]
    async fn replay_is_deterministic_default_features() {
        // Default-features replay-determinism check: any mix of
        // out-of-range bumps and membership entries produces the same
        // cell state across two independent replays, because the apply
        // arm folds the log deterministically (range check + subset
        // check are both pure functions of entry data + local binary
        // constants).
        let log = vec![
            (1u64, ReplayEntry::Membership(vec![1, 2])),
            (
                2u64,
                ReplayEntry::Bump {
                    target: 7,
                    gated: vec![1, 2, 3],
                },
            ),
            (3u64, ReplayEntry::Membership(vec![1, 2, 5])),
            (
                4u64,
                ReplayEntry::Bump {
                    target: 8,
                    gated: vec![1, 2],
                },
            ),
        ];
        let mut first = HighWaterStateMachine::new();
        let baseline = first.active_write_version();
        replay(&mut first, &log).await;
        let mut second = HighWaterStateMachine::new();
        replay(&mut second, &log).await;
        assert_eq!(first.active_write_version(), baseline);
        assert_eq!(
            first.active_write_version(),
            second.active_write_version(),
            "two replays of the same committed log yield the same cell"
        );
    }

    // ---- Snapshot active-write-version stamping (format migration) ----
    //
    // These tests pin the contract that the snapshot builder reads the
    // node's CURRENT active write version (the shared cell) at build time,
    // rather than a hard-coded constant. Today that is BASELINE_WRITE_VERSION
    // since no activation has flipped the cell; the assertions are written
    // against the accessor (`sm.active_write_version()`) so they remain
    // correct after a real activation moves the cell forward and the next
    // build emits at the new version.

    #[tokio::test]
    async fn build_snapshot_stamps_active_write_version() {
        // The persisted envelope's leading version byte must equal what the
        // accessor reports — not a literal. This proves the plumbing is
        // active-version-driven and would catch a regression to a constant.
        let store: Arc<dyn SnapshotStore> = Arc::new(InMemorySnapshotStore::new());
        let mut sm = HighWaterStateMachine::with_store(store.clone()).expect("with_store");
        apply_one(
            &mut sm,
            1,
            EntryPayload::Normal(HighWaterCommand::Advance(AdvancePayload { at_least: 321 })),
        )
        .await;

        let active = sm.active_write_version();
        sm.build_snapshot().await.expect("build_snapshot");

        let envelope_bytes = store.load().expect("load").expect("snapshot present");
        assert_eq!(
            envelope_bytes[0], active,
            "envelope leading version byte must equal the active write version"
        );

        // The inner payload blob is the on-disk record openraft replays from
        // on snapshot install; it must be stamped at the active version too
        // so a follower receiving it can decode against its own version
        // window.
        let persisted: PersistedSnapshot = tsoracle_codec::decode_framed(
            tsoracle_openraft_toolkit::MIN_READABLE_VERSION,
            tsoracle_openraft_toolkit::MAX_READABLE_VERSION,
            &envelope_bytes,
        )
        .expect("decode envelope");
        assert_eq!(
            persisted.data[0], active,
            "inner snapshot payload leading version byte must equal the active write version"
        );
    }

    #[tokio::test]
    async fn old_version_snapshot_is_read_and_reemitted_at_active_version() {
        // Migration-on-next-write: a snapshot persisted at version V is
        // readable across the multi-version codec and the next
        // build_snapshot re-emits at the cluster's ACTIVE write version.
        //
        // SCOPE: this drives the same-version axis — persist at the active
        // write version (BASELINE = 4, since no activation flip is injected
        // here) and assert the reopen reads it back and re-emits at the active
        // version. The genuine cross-version rewrite (install v4, activate v5,
        // assert the rebuild is v5) is the orthogonal axis; the readable range
        // is now [4, 5], so the v4 lower version genuinely exists on disk
        // pre-activation.
        let store: Arc<dyn SnapshotStore> = Arc::new(InMemorySnapshotStore::new());

        // Persist a snapshot via a first SM instance, then drop it.
        {
            let mut writer = HighWaterStateMachine::with_store(store.clone()).expect("writer SM");
            apply_one(
                &mut writer,
                3,
                EntryPayload::Normal(HighWaterCommand::Advance(AdvancePayload { at_least: 777 })),
            )
            .await;
            writer
                .build_snapshot()
                .await
                .expect("build initial snapshot");
        }

        // Reopen: the recovery path decodes across the readable range and
        // restores state.
        let mut sm = HighWaterStateMachine::with_store(store.clone()).expect("reopened SM");
        assert_eq!(
            sm.current_value().await,
            777,
            "recovered value across reopen"
        );

        // The next build re-emits at the active write version.
        let active = sm.active_write_version();
        apply_one(
            &mut sm,
            4,
            EntryPayload::Normal(HighWaterCommand::Advance(AdvancePayload { at_least: 888 })),
        )
        .await;
        sm.build_snapshot().await.expect("rebuild snapshot");

        let rebuilt = store.load().expect("load").expect("snapshot present");
        assert_eq!(
            rebuilt[0], active,
            "rebuilt snapshot must be emitted at the active write version"
        );
        let payload: PersistedSnapshot = tsoracle_codec::decode_framed(
            tsoracle_openraft_toolkit::MIN_READABLE_VERSION,
            tsoracle_openraft_toolkit::MAX_READABLE_VERSION,
            &rebuilt,
        )
        .expect("decode rebuilt envelope");
        assert_eq!(payload.meta.last_log_id.map(|l| l.index), Some(4));
    }

    #[test]
    fn snapshot_codec_accepts_full_readable_range() {
        // The migration-seam invariant in isolation: the snapshot decoder
        // must accept any version in [MIN_READABLE_VERSION,
        // MAX_READABLE_VERSION] and reject anything outside it. This is
        // what makes an OLD-version on-disk snapshot readable after the
        // active version moves forward. Asserted against the codec range
        // directly so it covers the full readable range [MIN = 4, MAX = 6] —
        // iterating v4 (decode-and-lift), v5 (direct decode), and v6
        // (byte-identical to v5).
        // Empty dense map + cap 0 so the v4 encode arm (which projects to the
        // frozen V4 struct) is lossless; assert_eq!(decoded, payload) holds
        // for every version in the readable range because the decode-and-lift
        // path restores dense=empty, dense_cap=0 from a v4 blob.
        let payload = HighWaterStateMachineSnapshot {
            current_value: 5,
            last_applied: Some(log_id(2)),
            last_membership: StoredMem::default(),
            dense: std::collections::BTreeMap::new(),
            dense_cap: 0,
        };
        for version in tsoracle_openraft_toolkit::MIN_READABLE_VERSION
            ..=tsoracle_openraft_toolkit::MAX_READABLE_VERSION
        {
            let framed = tsoracle_codec::encode_framed(version, &payload).expect("encode in range");
            let decoded: HighWaterStateMachineSnapshot = tsoracle_codec::decode_framed(
                tsoracle_openraft_toolkit::MIN_READABLE_VERSION,
                tsoracle_openraft_toolkit::MAX_READABLE_VERSION,
                &framed,
            )
            .expect("decode in range");
            assert_eq!(decoded, payload);
        }

        // One past the readable max must be rejected loudly.
        let above = tsoracle_openraft_toolkit::MAX_READABLE_VERSION.saturating_add(1);
        let mut framed_above = vec![above];
        framed_above.extend_from_slice(&tsoracle_codec::encode_postcard(&payload).expect("body"));
        assert!(
            tsoracle_codec::decode_framed::<HighWaterStateMachineSnapshot>(
                tsoracle_openraft_toolkit::MIN_READABLE_VERSION,
                tsoracle_openraft_toolkit::MAX_READABLE_VERSION,
                &framed_above,
            )
            .is_err(),
            "a version above the readable max must be rejected"
        );
    }

    #[tokio::test]
    async fn active_write_version_survives_reopen_and_drives_writes() {
        // The durable active write version is recovered on reopen and is
        // the version stamped onto snapshot writes. On `main` the recovered
        // value is BASELINE because no activation has flipped it; this
        // test pins that (a) the accessor is stable across a reopen and
        // (b) the snapshot write uses exactly that recovered value.
        let store: Arc<dyn SnapshotStore> = Arc::new(InMemorySnapshotStore::new());
        let recovered_active;
        {
            let mut sm = HighWaterStateMachine::with_store(store.clone()).expect("first SM");
            apply_one(
                &mut sm,
                1,
                EntryPayload::Normal(HighWaterCommand::Advance(AdvancePayload { at_least: 55 })),
            )
            .await;
            sm.build_snapshot().await.expect("build_snapshot");
            recovered_active = sm.active_write_version();
        }

        let mut reopened = HighWaterStateMachine::with_store(store.clone()).expect("reopened SM");
        assert_eq!(
            reopened.active_write_version(),
            recovered_active,
            "active write version must be stable across reopen"
        );

        apply_one(
            &mut reopened,
            2,
            EntryPayload::Normal(HighWaterCommand::Advance(AdvancePayload { at_least: 66 })),
        )
        .await;
        reopened.build_snapshot().await.expect("rebuild");
        let bytes = store.load().expect("load").expect("snapshot present");
        assert_eq!(
            bytes[0],
            reopened.active_write_version(),
            "rebuilt snapshot stamped with the recovered active write version"
        );
    }

    // ---- Snapshot tests ----

    #[tokio::test]
    async fn build_snapshot_round_trips_payload() {
        let mut sm = HighWaterStateMachine::new();
        apply_one(
            &mut sm,
            1,
            EntryPayload::Normal(HighWaterCommand::Advance(AdvancePayload { at_least: 500 })),
        )
        .await;

        let snap = sm.build_snapshot().await.expect("build_snapshot");
        let bytes = snap.snapshot.into_inner();
        // Use decode_framed (VersionedCodec) so the v4 bytes are decoded via
        // the frozen-V4-struct + lift path, not a raw postcard deserialize.
        let payload: HighWaterStateMachineSnapshot = tsoracle_codec::decode_framed(
            tsoracle_openraft_toolkit::MIN_READABLE_VERSION,
            tsoracle_openraft_toolkit::MAX_READABLE_VERSION,
            &bytes,
        )
        .expect("decode snapshot");
        assert_eq!(payload.current_value, 500);
        assert_eq!(payload.last_applied.map(|l| l.index), Some(1));
    }

    #[tokio::test]
    async fn build_snapshot_uses_fresh_id_each_time() {
        let mut sm = HighWaterStateMachine::new();
        apply_one(
            &mut sm,
            1,
            EntryPayload::Normal(HighWaterCommand::Advance(AdvancePayload { at_least: 7 })),
        )
        .await;

        let a = sm.build_snapshot().await.expect("build_snapshot a");
        let b = sm.build_snapshot().await.expect("build_snapshot b");
        assert_ne!(
            a.meta.snapshot_id, b.meta.snapshot_id,
            "two snapshots at same last_applied must have distinct ids"
        );
    }

    #[tokio::test]
    async fn install_snapshot_replaces_state() {
        let mut sm = HighWaterStateMachine::new();
        let payload = HighWaterStateMachineSnapshot {
            current_value: 999,
            last_applied: Some(log_id(5)),
            last_membership: StoredMem::default(),
            dense: std::collections::BTreeMap::new(),
            dense_cap: 0,
        };
        let bytes = tsoracle_codec::encode_framed(
            tsoracle_openraft_toolkit::BASELINE_WRITE_VERSION,
            &payload,
        )
        .expect("serialize payload");

        let meta = SnapMeta {
            last_log_id: payload.last_applied,
            last_membership: payload.last_membership.clone(),
            snapshot_id: "test-install-1".to_string(),
        };
        sm.install_snapshot(&meta, std::io::Cursor::new(bytes))
            .await
            .expect("install_snapshot");

        assert_eq!(sm.current_value().await, 999);
        let (last, _) = sm.applied_state().await.unwrap();
        assert_eq!(last.map(|l| l.index), Some(5));

        let current = sm
            .get_current_snapshot()
            .await
            .expect("get_current_snapshot")
            .expect("snapshot present");
        assert_eq!(current.meta.snapshot_id, "test-install-1");
    }

    /// Build a v5 (dense) snapshot whose `dense` map carries one raw-`String`
    /// key, paired with a meta whose `last_log_id` agrees with the payload so
    /// the install reaches the dense-key validation guard (not the earlier
    /// meta/payload-disagreement check). Returns `(meta, framed bytes)`.
    fn dense_snapshot_with_key(key: &str) -> (SnapMeta, Vec<u8>) {
        let last_applied = Some(log_id(9));
        let payload = HighWaterStateMachineSnapshot {
            current_value: 999,
            last_applied,
            last_membership: StoredMem::default(),
            dense: std::collections::BTreeMap::from([(key.to_string(), 7u64)]),
            dense_cap: 64,
        };
        let bytes =
            tsoracle_codec::encode_framed(tsoracle_openraft_toolkit::DENSE_WRITE_VERSION, &payload)
                .expect("serialize v5 dense payload");
        let meta = SnapMeta {
            last_log_id: last_applied,
            last_membership: StoredMem::default(),
            snapshot_id: "test-invalid-dense-key".to_string(),
        };
        (meta, bytes)
    }

    #[tokio::test]
    async fn install_snapshot_rejects_empty_dense_key() {
        let mut sm = HighWaterStateMachine::new();
        let before = sm.current_value().await;
        let (meta, bytes) = dense_snapshot_with_key("");

        let err = sm
            .install_snapshot(&meta, std::io::Cursor::new(bytes))
            .await
            .expect_err("install of a snapshot with an empty dense key must be rejected");
        assert_eq!(err.kind(), std::io::ErrorKind::InvalidData);

        // Rejected before any mutation: the wholesale state replacement (which
        // would have set current_value to 999) did not happen, and the invalid
        // key is absent from the dense map.
        assert_eq!(sm.current_value().await, before);
        assert_eq!(sm.dense_value(""), 0);
    }

    #[tokio::test]
    async fn install_snapshot_rejects_oversized_dense_key() {
        let mut sm = HighWaterStateMachine::new();
        let before = sm.current_value().await;
        let oversized = "a".repeat(tsoracle_core::MAX_SEQ_KEY_LEN + 1);
        let (meta, bytes) = dense_snapshot_with_key(&oversized);

        let err = sm
            .install_snapshot(&meta, std::io::Cursor::new(bytes))
            .await
            .expect_err("install of a snapshot with an oversized dense key must be rejected");
        assert_eq!(err.kind(), std::io::ErrorKind::InvalidData);

        assert_eq!(sm.current_value().await, before);
        assert_eq!(sm.dense_value(&oversized), 0);
    }

    #[tokio::test]
    async fn install_snapshot_accepts_valid_dense_key() {
        // The guard must not reject an honest v5 snapshot: a valid dense key
        // installs and is observable afterwards.
        let mut sm = HighWaterStateMachine::new();
        let (meta, bytes) = dense_snapshot_with_key("orders");

        sm.install_snapshot(&meta, std::io::Cursor::new(bytes))
            .await
            .expect("install of a snapshot with a valid dense key must succeed");

        assert_eq!(sm.current_value().await, 999);
        assert_eq!(sm.dense_value("orders"), 7);
    }

    #[tokio::test]
    async fn get_current_snapshot_initially_none() {
        let mut sm = HighWaterStateMachine::new();
        let s = sm
            .get_current_snapshot()
            .await
            .expect("get_current_snapshot");
        assert!(s.is_none());
    }

    // ---- SnapshotStore integration ----

    use crate::snapshot_store::{InMemorySnapshotStore, SnapshotStore};

    #[tokio::test]
    async fn build_snapshot_writes_through_store() {
        let store: Arc<dyn SnapshotStore> = Arc::new(InMemorySnapshotStore::new());
        let mut sm = HighWaterStateMachine::with_store(store.clone()).expect("with_store");
        apply_one(
            &mut sm,
            1,
            EntryPayload::Normal(HighWaterCommand::Advance(AdvancePayload { at_least: 42 })),
        )
        .await;
        sm.build_snapshot().await.expect("build_snapshot");
        assert!(
            store.load().expect("load").is_some(),
            "store must contain a persisted snapshot after build_snapshot"
        );
    }

    #[tokio::test]
    async fn with_store_recovers_prior_snapshot_on_construction() {
        let store: Arc<dyn SnapshotStore> = Arc::new(InMemorySnapshotStore::new());
        {
            let mut sm = HighWaterStateMachine::with_store(store.clone()).expect("first SM");
            apply_one(
                &mut sm,
                1,
                EntryPayload::Normal(HighWaterCommand::Advance(AdvancePayload { at_least: 99 })),
            )
            .await;
            sm.build_snapshot().await.expect("build_snapshot");
        }
        let mut sm = HighWaterStateMachine::with_store(store).expect("reopened SM");
        assert_eq!(sm.current_value().await, 99);
        // `applied_state` must report the snapshot's last_log_id after reopen —
        // without this, openraft re-applies from index 0 and panics on missing
        // log entries that the snapshot already covered.
        let (last, _) = sm.applied_state().await.unwrap();
        assert_eq!(last.map(|l| l.index), Some(1));
        let snap = sm
            .get_current_snapshot()
            .await
            .expect("get_current_snapshot")
            .expect("snapshot present after reopen");
        assert_eq!(snap.meta.last_log_id.map(|l| l.index), Some(1));
    }

    #[tokio::test]
    async fn default_constructor_matches_new() {
        // `Default` is the canonical "no-arg" entry point for embedders that
        // build the SM via `..Default::default()`; pinning behavior here keeps
        // the in-memory snapshot store as the unsurprising default.
        let mut sm = HighWaterStateMachine::default();
        assert_eq!(sm.current_value().await, 0);
        let snap = sm
            .get_current_snapshot()
            .await
            .expect("get_current_snapshot");
        assert!(snap.is_none());
    }

    #[tokio::test]
    async fn state_machine_defaults_to_baseline_active_write_version() {
        let store: Arc<dyn SnapshotStore> = Arc::new(InMemorySnapshotStore::new());
        let sm = HighWaterStateMachine::with_store(store).expect("with_store");
        assert_eq!(
            sm.active_write_version(),
            tsoracle_openraft_toolkit::BASELINE_WRITE_VERSION
        );
    }

    #[tokio::test]
    async fn state_machine_reads_the_same_shared_cell() {
        // The SM and the log store hold clones of ONE cell. A set() on the
        // cell (a later phase's activation apply will do this) is observed
        // identically by every clone.
        let cell = tsoracle_openraft_toolkit::ActiveWriteVersion::default();
        let store: Arc<dyn SnapshotStore> = Arc::new(InMemorySnapshotStore::new());
        let sm = HighWaterStateMachine::with_store_and_active_version(store, cell.clone())
            .expect("with_store_and_active_version");
        assert_eq!(
            sm.active_write_version(),
            tsoracle_openraft_toolkit::BASELINE_WRITE_VERSION
        );
        cell.set(7);
        assert_eq!(sm.active_write_version(), 7);
    }

    #[tokio::test]
    async fn begin_receiving_snapshot_returns_empty_cursor() {
        // openraft hands the returned cursor to the snapshot-receiving network
        // path; the contract is "empty, writable buffer." Anything non-empty
        // would corrupt the install on the receiving side.
        let mut sm = HighWaterStateMachine::new();
        let cursor = sm
            .begin_receiving_snapshot()
            .await
            .expect("begin_receiving_snapshot");
        assert!(cursor.into_inner().is_empty());
    }

    #[tokio::test]
    async fn with_store_errors_on_malformed_persisted_envelope() {
        // Hardens the recovery path against on-disk corruption: a snapshot
        // blob that doesn't decode as `PersistedSnapshot` must surface as a
        // structured `io::Error` from `with_store`, not a silent reset to
        // the default state (which would lose the value across restart).
        let store: Arc<dyn SnapshotStore> = Arc::new(InMemorySnapshotStore::new());
        store.save(b"not a postcard envelope").unwrap();
        let Err(err) = HighWaterStateMachine::with_store(store) else {
            panic!("with_store must reject malformed envelope");
        };
        let msg = err.to_string();
        assert!(msg.contains("envelope decode"));
    }

    #[tokio::test]
    async fn with_store_errors_on_malformed_inner_payload() {
        // Envelope decodes but the inner `HighWaterStateMachineSnapshot` does
        // not — also surfaces as `io::Error` rather than silent state reset.
        let envelope = PersistedSnapshot {
            meta: SnapMeta::default(),
            data: b"not a postcard payload".to_vec(),
        };
        let bytes = tsoracle_openraft_toolkit::encode(
            tsoracle_openraft_toolkit::BASELINE_WRITE_VERSION,
            &envelope,
        )
        .unwrap();
        let store: Arc<dyn SnapshotStore> = Arc::new(InMemorySnapshotStore::new());
        store.save(&bytes).unwrap();
        let Err(err) = HighWaterStateMachine::with_store(store) else {
            panic!("with_store must reject malformed inner payload");
        };
        let msg = err.to_string();
        assert!(msg.contains("payload decode"));
    }

    #[tokio::test]
    async fn install_snapshot_writes_through_store() {
        let store: Arc<dyn SnapshotStore> = Arc::new(InMemorySnapshotStore::new());
        let mut sm = HighWaterStateMachine::with_store(store.clone()).expect("with_store");
        let payload = HighWaterStateMachineSnapshot {
            current_value: 700,
            last_applied: Some(log_id(10)),
            last_membership: StoredMem::default(),
            dense: std::collections::BTreeMap::new(),
            dense_cap: 0,
        };
        let bytes = tsoracle_codec::encode_framed(
            tsoracle_openraft_toolkit::BASELINE_WRITE_VERSION,
            &payload,
        )
        .unwrap();
        let meta = SnapMeta {
            last_log_id: payload.last_applied,
            last_membership: payload.last_membership.clone(),
            snapshot_id: "install-1".into(),
        };
        sm.install_snapshot(&meta, std::io::Cursor::new(bytes))
            .await
            .expect("install_snapshot");
        // Reopen with the same store: install must have written through.
        let mut sm2 = HighWaterStateMachine::with_store(store).expect("reopened SM");
        assert_eq!(sm2.current_value().await, 700);
        let (last, _) = sm2.applied_state().await.unwrap();
        assert_eq!(last.map(|l| l.index), Some(10));
    }

    /// Encode a `(meta, framed-payload-bytes)` pair for an install at
    /// `(value, last_applied)`, with `meta.last_log_id` matching the payload so
    /// the meta/payload consistency check passes and the install reaches the
    /// publish path.
    fn install_payload(value: u64, last_applied: LogId) -> (SnapMeta, Vec<u8>) {
        let payload = HighWaterStateMachineSnapshot {
            current_value: value,
            last_applied: Some(last_applied),
            last_membership: StoredMem::default(),
            dense: std::collections::BTreeMap::new(),
            dense_cap: 0,
        };
        let bytes = tsoracle_codec::encode_framed(
            tsoracle_openraft_toolkit::BASELINE_WRITE_VERSION,
            &payload,
        )
        .expect("serialize payload");
        let meta = SnapMeta {
            last_log_id: payload.last_applied,
            last_membership: payload.last_membership.clone(),
            snapshot_id: format!("snap-{}", last_applied.index),
        };
        (meta, bytes)
    }

    #[tokio::test]
    async fn install_snapshot_does_not_regress_published_snapshot() {
        // `build_snapshot` publishes both the durable store (:256) and the
        // in-memory `current_snapshot` (:258) AFTER releasing the core lock,
        // and openraft can drive a stale build on a clone concurrently with a
        // newer install. The publish must therefore be monotone by
        // `last_log_id`: a later-arriving, lower-indexed snapshot must not roll
        // the durable + in-memory snapshot back to a stale state (which a
        // subsequent restart would then recover, after openraft has already
        // purged its log past the newer snapshot). We exercise that invariant
        // through the install path here — install and build share the same
        // monotone publish, so a regressing install stands in for a stale build
        // that finishes its write last.
        enable_tracing();
        let store: Arc<dyn SnapshotStore> = Arc::new(InMemorySnapshotStore::new());
        let mut sm = HighWaterStateMachine::with_store(store.clone()).expect("with_store");

        let (newer_meta, newer_bytes) = install_payload(80, log_id(8));
        sm.install_snapshot(&newer_meta, Cursor::new(newer_bytes))
            .await
            .expect("install newer");

        // A stale snapshot (lower last_log_id) arriving afterwards is a no-op:
        // we are already at least this far, so it must neither error nor
        // regress state.
        let (older_meta, older_bytes) = install_payload(50, log_id(5));
        sm.install_snapshot(&older_meta, Cursor::new(older_bytes))
            .await
            .expect("stale install must be an accepted no-op, not an error");

        assert_eq!(
            sm.current_value().await,
            80,
            "value must not regress to the stale snapshot"
        );
        let (last, _) = sm.applied_state().await.unwrap();
        assert_eq!(
            last.map(|l| l.index),
            Some(8),
            "last_applied must not regress"
        );
        let current = sm
            .get_current_snapshot()
            .await
            .expect("get_current_snapshot")
            .expect("snapshot present");
        assert_eq!(
            current.meta.last_log_id.map(|l| l.index),
            Some(8),
            "in-memory current_snapshot must not regress"
        );

        // The durable store must also still hold the newer snapshot — this is
        // the recovery-critical half: reopening must not resurrect the stale
        // state.
        let mut reopened = HighWaterStateMachine::with_store(store).expect("reopen");
        assert_eq!(
            reopened.current_value().await,
            80,
            "durable store must not have been rolled back to the stale snapshot"
        );
        let (last, _) = reopened.applied_state().await.unwrap();
        assert_eq!(last.map(|l| l.index), Some(8));
    }

    #[test]
    fn supersedes_published_is_monotone_by_last_log_id() {
        // The exact rule the publish gate enforces. Equal or greater
        // last_log_id supersedes; a lower one — or `None` against a `Some` —
        // does not. `None` is the minimum: a fresh/pre-genesis snapshot never
        // displaces an established one, but anything fills an empty slot.
        assert!(
            supersedes_published(Some(log_id(5)), None),
            "any Some fills an empty slot"
        );
        assert!(
            supersedes_published(None, None),
            "first publish into an empty slot is allowed"
        );
        assert!(
            supersedes_published(Some(log_id(8)), Some(log_id(5))),
            "higher index supersedes"
        );
        assert!(
            supersedes_published(Some(log_id(5)), Some(log_id(5))),
            "equal index may republish (same state, fresh snapshot_id)"
        );
        assert!(
            !supersedes_published(Some(log_id(5)), Some(log_id(8))),
            "lower index must not regress"
        );
        assert!(
            !supersedes_published(None, Some(log_id(1))),
            "None must not regress an established snapshot"
        );
    }

    #[tokio::test]
    async fn commit_snapshot_drops_a_stale_publish() {
        // Exercise the build-side of the TOCTOU directly at the seam
        // build_snapshot shares with install. Once a newer snapshot is durable
        // and published, a commit carrying a lower last_log_id — a build whose
        // state was captured before the newer install but whose write lands
        // last — must be dropped: it reports `Ok(false)`, runs no adopt
        // callback, and leaves both the store and current_snapshot byte-for-byte
        // untouched. (The full build path can only reach this state under a
        // real scheduling race, so we drive the shared commit seam directly.)
        enable_tracing();
        let store: Arc<dyn SnapshotStore> = Arc::new(InMemorySnapshotStore::new());
        let mut sm = HighWaterStateMachine::with_store(store.clone()).expect("with_store");

        let (newer_meta, newer_bytes) = install_payload(80, log_id(8));
        sm.install_snapshot(&newer_meta, Cursor::new(newer_bytes))
            .await
            .expect("install newer");
        let durable_after_newer = store
            .load()
            .expect("load")
            .expect("durable snapshot present");

        let (stale_meta, stale_data) = install_payload(50, log_id(5));
        let envelope = tsoracle_openraft_toolkit::encode(
            tsoracle_openraft_toolkit::BASELINE_WRITE_VERSION,
            &PersistedSnapshot {
                meta: stale_meta.clone(),
                data: stale_data.clone(),
            },
        )
        .expect("encode envelope");

        let adopt_ran = std::cell::Cell::new(false);
        let adopted = sm
            .commit_snapshot(stale_meta, stale_data, &envelope, |_| adopt_ran.set(true))
            .expect("commit_snapshot");

        assert!(!adopted, "stale commit must report it was not adopted");
        assert!(
            !adopt_ran.get(),
            "adopt callback must not run for a dropped commit"
        );
        assert_eq!(
            store.load().expect("load").as_deref(),
            Some(durable_after_newer.as_slice()),
            "durable snapshot must be unchanged — a stale commit must not write"
        );
        let current = sm
            .get_current_snapshot()
            .await
            .expect("get_current_snapshot")
            .expect("snapshot present");
        assert_eq!(
            current.meta.last_log_id.map(|l| l.index),
            Some(8),
            "in-memory current_snapshot must be unchanged"
        );
    }

    #[tokio::test]
    async fn install_snapshot_rejects_meta_payload_log_id_mismatch() {
        // `meta.last_log_id` (read back by `get_current_snapshot`) and
        // `payload.last_applied` (read back by `applied_state`) must agree —
        // honest peers always build them from the same `last_applied`, so a
        // disagreement signals corruption or a peer bug. Installing it anyway
        // would silently desync the two reader methods, so the install must be
        // rejected, leave state untouched, and persist nothing.
        let store: Arc<dyn SnapshotStore> = Arc::new(InMemorySnapshotStore::new());
        let mut sm = HighWaterStateMachine::with_store(store.clone()).expect("with_store");

        let payload = HighWaterStateMachineSnapshot {
            current_value: 123,
            last_applied: Some(log_id(5)),
            last_membership: StoredMem::default(),
            dense: std::collections::BTreeMap::new(),
            dense_cap: 0,
        };
        let bytes = tsoracle_codec::encode_framed(
            tsoracle_openraft_toolkit::BASELINE_WRITE_VERSION,
            &payload,
        )
        .expect("serialize payload");

        let meta = SnapMeta {
            last_log_id: Some(log_id(6)),
            last_membership: payload.last_membership.clone(),
            snapshot_id: "mismatch-1".into(),
        };

        let Err(err) = sm
            .install_snapshot(&meta, std::io::Cursor::new(bytes))
            .await
        else {
            panic!("install_snapshot must reject meta/payload last_log_id mismatch");
        };
        let msg = err.to_string();
        assert!(
            msg.contains("last_log_id"),
            "error should name the mismatched field: {msg}"
        );

        assert_eq!(sm.current_value().await, 0);
        let (last, _) = sm.applied_state().await.unwrap();
        assert_eq!(last, None);
        assert!(
            sm.get_current_snapshot().await.unwrap().is_none(),
            "rejected install must not publish a current snapshot"
        );
        assert!(
            store.load().expect("load").is_none(),
            "rejected install must not persist an envelope"
        );
    }

    // ---- Property tests ----

    use proptest::prelude::*;

    fn rt() -> tokio::runtime::Runtime {
        tokio::runtime::Builder::new_current_thread()
            .enable_all()
            .build()
            .unwrap()
    }

    proptest! {
        /// Monotonicity invariant: applying an arbitrary sequence of `Advance`
        /// values leaves the state machine at `max(0, max(at_leasts))`, and
        /// `current_value` is non-decreasing at every intermediate step.
        #[test]
        fn p1_monotonicity_under_arbitrary_bumps(
            targets in prop::collection::vec(any::<u64>(), 0..=64)
        ) {
            rt().block_on(async {
                let mut sm = HighWaterStateMachine::new();
                let mut prev = 0u64;
                let mut idx = 0u64;
                for t in &targets {
                    idx += 1;
                    apply_one(
                        &mut sm,
                        idx,
                        EntryPayload::Normal(HighWaterCommand::Advance(AdvancePayload { at_least: *t })),
                    )
                    .await;
                    let now = sm.current_value().await;
                    prop_assert!(now >= prev, "value went backwards: prev={prev} now={now}");
                    prev = now;
                }
                let expected = targets.iter().copied().max().unwrap_or(0);
                prop_assert_eq!(sm.current_value().await, expected);
                Ok(())
            })?;
        }

        /// Snapshot payload round-trip: build_snapshot -> install_snapshot
        /// preserves (current_value, last_applied, last_membership) across
        /// arbitrary apply sequences.
        #[test]
        fn p2_snapshot_payload_round_trip(
            bumps in prop::collection::vec(any::<u64>(), 0..=32)
        ) {
            rt().block_on(async {
                let mut sm = HighWaterStateMachine::new();
                let mut idx = 0u64;
                for t in &bumps {
                    idx += 1;
                    apply_one(
                        &mut sm,
                        idx,
                        EntryPayload::Normal(HighWaterCommand::Advance(AdvancePayload { at_least: *t })),
                    )
                    .await;
                }

                let snap = sm.build_snapshot().await.expect("build_snapshot");
                let meta = snap.meta.clone();
                let bytes = snap.snapshot.into_inner();

                let mut sm2 = HighWaterStateMachine::new();
                sm2.install_snapshot(&meta, std::io::Cursor::new(bytes))
                    .await
                    .expect("install_snapshot");

                prop_assert_eq!(sm2.current_value().await, sm.current_value().await);
                let (a_last, a_mem) = sm.applied_state().await.unwrap();
                let (b_last, b_mem) = sm2.applied_state().await.unwrap();
                prop_assert_eq!(a_last, b_last);
                prop_assert_eq!(a_mem, b_mem);
                Ok(())
            })?;
        }
    }

    #[test]
    fn snapshot_payload_versioned_codec_matches_legacy_frame() {
        use tsoracle_codec::{decode_framed, encode_framed};
        // The v4 encode arm projects to the frozen `HighWaterStateMachineSnapshotV4`
        // struct (3 fields, no dense/cap). This test proves:
        //
        //  1. The v4 framed bytes are `[4, 7, 0, 0, 0, 0]` — the on-disk v4
        //     layout did not change from the pre-dense era.
        //  2. Those bytes equal what the frozen V4 struct produces when encoded
        //     directly — a real pre-dense reader can still decode the output.
        //  3. A v4-decode-and-lift roundtrip restores the original value with
        //     empty dense map + cap 0 ("absent" sentinel).
        let payload = HighWaterStateMachineSnapshot {
            current_value: 7,
            last_applied: None,
            last_membership: StoredMembership::default(),
            dense: std::collections::BTreeMap::new(),
            dense_cap: 0,
        };
        let via_seam = encode_framed(tsoracle_openraft_toolkit::BASELINE_WRITE_VERSION, &payload)
            .expect("encode_framed");

        // Prove the v4 body equals encoding the frozen V4 struct directly.
        let v4_direct = tsoracle_openraft_toolkit::encode(
            tsoracle_openraft_toolkit::BASELINE_WRITE_VERSION,
            &HighWaterStateMachineSnapshotV4 {
                current_value: 7,
                last_applied: None,
                last_membership: StoredMembership::default(),
            },
        )
        .expect("v4 direct encode");
        assert_eq!(
            via_seam, v4_direct,
            "v4 encode must project to the frozen V4 struct"
        );
        assert_eq!(via_seam, vec![4, 7, 0, 0, 0, 0]);

        // decode-and-lift: v4 bytes decode into the current layout with empty
        // dense map and cap 0 (the "absent" sentinel), which matches `payload`.
        let back: HighWaterStateMachineSnapshot = decode_framed(
            tsoracle_openraft_toolkit::BASELINE_WRITE_VERSION,
            tsoracle_openraft_toolkit::BASELINE_WRITE_VERSION,
            &via_seam,
        )
        .expect("decode_framed");
        assert_eq!(back, payload);
    }

    #[test]
    fn snapshot_payload_versioned_codec_rejects_foreign_version() {
        use tsoracle_codec::{CodecError, decode_framed};
        let framed = vec![0xFFu8, 7, 0, 0, 0, 0];
        let err = decode_framed::<HighWaterStateMachineSnapshot>(
            tsoracle_openraft_toolkit::BASELINE_WRITE_VERSION,
            tsoracle_openraft_toolkit::BASELINE_WRITE_VERSION,
            &framed,
        )
        .expect_err("must reject");
        assert!(matches!(
            err,
            CodecError::VersionUnsupported { actual: 0xFF, .. }
        ));
    }

    #[test]
    fn snapshot_payload_pins_v4_layout() {
        use tsoracle_codec::encode_framed;
        // Hand-built v4 frame: [BASELINE_WRITE_VERSION | postcard(V4_payload)],
        // produced through the VersionedCodec seam that build_snapshot uses
        // when active_write_version == BASELINE_WRITE_VERSION. The v4 encode
        // arm projects to the frozen `HighWaterStateMachineSnapshotV4` struct
        // (3 fields, no dense/cap), so the body bytes are unchanged from the
        // pre-dense era. Leading byte advanced 3 -> 4 when OpenraftPeer gained
        // the admin_endpoint field (a breaking on-disk change for membership);
        // the postcard body is unchanged. Body [7, 0, 0, 0, 0] =
        // current_value 7, last_applied None, then default StoredMembership
        // (None log id + empty configs + empty nodes). Reordering or inserting
        // a field in the V4 struct changes these bytes and trips this test,
        // forcing a deliberate review.
        let payload = HighWaterStateMachineSnapshot {
            current_value: 7,
            last_applied: None,
            last_membership: StoredMembership::default(),
            dense: std::collections::BTreeMap::new(),
            dense_cap: 0,
        };
        let framed = encode_framed(tsoracle_openraft_toolkit::BASELINE_WRITE_VERSION, &payload)
            .expect("encode");
        // v4 encode projects to the frozen V4 struct: body is still
        // [7, 0, 0, 0, 0] — the dense/cap bytes are NOT in the output.
        assert_eq!(framed, vec![4, 7, 0, 0, 0, 0]);
    }

    #[test]
    fn log_entry_pins_v4_layout() {
        use crate::log_codec::OpenraftLogCodec;
        use tsoracle_openraft_toolkit::LogStoreCodec;
        // The bytes RocksdbLogStore<TypeConfig> persists per entry: the
        // active write version byte (prepended by the store) +
        // OpenraftLogCodec entry body — the v4 frame around a Normal entry
        // carrying Advance(AdvancePayload { at_least: 5 })
        // at log id (term 1, node 1, index 1). Body [1,1,1,1,0,5] = leader
        // (term 1, node 1), index 1, EntryPayload::Normal tag (1), Advance
        // variant (0), at_least 5 — byte-identical to the pre-seam layout.
        let lid = openraft::testing::log_id::<TypeConfig>(1, 1, 1);
        let entry: EntryOf<TypeConfig> = EntryOf::<TypeConfig>::new_normal(
            lid,
            HighWaterCommand::Advance(AdvancePayload { at_least: 5 }),
        );
        let body = <OpenraftLogCodec as LogStoreCodec<TypeConfig>>::encode_entry(
            tsoracle_openraft_toolkit::BASELINE_WRITE_VERSION,
            &entry,
        )
        .expect("encode entry body");
        let mut framed = vec![tsoracle_openraft_toolkit::BASELINE_WRITE_VERSION];
        framed.extend_from_slice(&body);
        assert_eq!(framed, vec![4, 1, 1, 1, 1, 0, 5]);
    }

    #[test]
    fn meta_vote_pins_v4_layout() {
        use crate::log_codec::OpenraftLogCodec;
        use tsoracle_openraft_toolkit::LogStoreCodec;
        // The bytes RocksdbLogStore<TypeConfig> persists in the meta column for a
        // Vote: active write version byte + OpenraftLogCodec vote body. Body [7, 3, 1] =
        // leader (term 7, node 3), committed flag true. A layout change to Vote
        // trips this test. This is the recovery-critical field the framing
        // protects: a foreign version loud-rejects instead of misdecoding.
        let vote: openraft::type_config::alias::VoteOf<TypeConfig> =
            openraft::Vote::new_committed(7, 3);
        let body = <OpenraftLogCodec as LogStoreCodec<TypeConfig>>::encode_vote(
            tsoracle_openraft_toolkit::BASELINE_WRITE_VERSION,
            &vote,
        )
        .expect("encode vote body");
        let mut framed = vec![tsoracle_openraft_toolkit::BASELINE_WRITE_VERSION];
        framed.extend_from_slice(&body);
        assert_eq!(framed, vec![4, 7, 3, 1]);
    }

    #[tokio::test]
    async fn install_snapshot_rejects_foreign_schema_version() {
        // A streamed snapshot framed with a foreign version must be rejected
        // as InvalidData before any store write or state mutation.
        let store: Arc<dyn SnapshotStore> = Arc::new(InMemorySnapshotStore::new());
        let mut sm = HighWaterStateMachine::with_store(store.clone()).expect("with_store");
        let meta = SnapMeta {
            last_log_id: None,
            last_membership: StoredMembership::default(),
            snapshot_id: "test".to_string(),
        };
        let foreign = vec![0xFF, 7, 0, 0, 0, 0];
        let err = sm
            .install_snapshot(&meta, Cursor::new(foreign))
            .await
            .expect_err("must reject foreign version");
        assert_eq!(err.kind(), io::ErrorKind::InvalidData);

        assert_eq!(sm.current_value().await, 0);
        let (last, _) = sm.applied_state().await.unwrap();
        assert_eq!(last, None);
        assert!(
            sm.get_current_snapshot().await.unwrap().is_none(),
            "rejected install must not publish a current snapshot"
        );
        assert!(
            store.load().expect("load").is_none(),
            "rejected install must not persist an envelope"
        );
    }

    // ---- Snapshot codec cross-version tests (Task 4) ----
    //
    // These tests pin the postcard correctness guarantees added by Task 4:
    //  - v5 round-trip: encode a dense snapshot, decode back equal.
    //  - v4 forward-compat: encode an empty-dense snapshot at v4; the bytes
    //    equal what the frozen V4 struct produces; decode lifts to empty dense.
    //  - v4 old-bytes: hand-encode a V4 struct, decode as current → empty dense.

    #[test]
    fn snapshot_codec_v5_roundtrip_with_nonempty_dense_map() {
        use tsoracle_codec::{decode_framed, encode_framed};
        use tsoracle_openraft_toolkit::DENSE_WRITE_VERSION;

        let mut dense = std::collections::BTreeMap::new();
        dense.insert("orders".to_string(), 42u64);
        dense.insert("invoices".to_string(), 1000u64);

        let payload = HighWaterStateMachineSnapshot {
            current_value: 777,
            last_applied: Some(log_id(9)),
            last_membership: StoredMem::default(),
            dense: dense.clone(),
            dense_cap: 500,
        };

        let framed = encode_framed(DENSE_WRITE_VERSION, &payload).expect("v5 encode");
        assert_eq!(framed[0], DENSE_WRITE_VERSION, "leading byte must be v5");

        let decoded: HighWaterStateMachineSnapshot = decode_framed(
            tsoracle_openraft_toolkit::MIN_READABLE_VERSION,
            tsoracle_openraft_toolkit::MAX_READABLE_VERSION,
            &framed,
        )
        .expect("v5 decode");
        assert_eq!(
            decoded, payload,
            "v5 roundtrip must produce the original payload"
        );
        assert_eq!(decoded.dense, dense, "dense map must survive roundtrip");
        assert_eq!(decoded.dense_cap, 500, "dense_cap must survive roundtrip");
    }

    #[test]
    fn snapshot_payload_round_trips_at_v6() {
        use tsoracle_codec::{decode_framed, encode_framed};
        use tsoracle_openraft_toolkit::BATCH_WRITE_VERSION;

        let mut dense = std::collections::BTreeMap::new();
        dense.insert("orders".to_string(), 9u64);

        let payload = HighWaterStateMachineSnapshot {
            current_value: 777,
            last_applied: Some(log_id(9)),
            last_membership: StoredMem::default(),
            dense: dense.clone(),
            dense_cap: 32,
        };

        let framed = encode_framed(BATCH_WRITE_VERSION, &payload).expect("v6 encode");
        assert_eq!(framed[0], BATCH_WRITE_VERSION, "leading byte must be v6");

        let decoded: HighWaterStateMachineSnapshot = decode_framed(
            tsoracle_openraft_toolkit::MIN_READABLE_VERSION,
            tsoracle_openraft_toolkit::MAX_READABLE_VERSION,
            &framed,
        )
        .expect("v6 decode");
        assert_eq!(
            decoded, payload,
            "v6 roundtrip must produce the original payload"
        );
        assert_eq!(decoded.dense, dense, "dense map must survive v6 roundtrip");
        assert_eq!(decoded.dense_cap, 32, "dense_cap must survive v6 roundtrip");
    }

    #[test]
    fn snapshot_codec_v4_forward_compat_empty_dense() {
        use tsoracle_codec::{decode_framed, encode_framed};

        // A pre-dense snapshot: dense is empty, cap is 0.
        let payload = HighWaterStateMachineSnapshot {
            current_value: 42,
            last_applied: Some(log_id(3)),
            last_membership: StoredMem::default(),
            dense: std::collections::BTreeMap::new(),
            dense_cap: 0,
        };

        let framed = encode_framed(tsoracle_openraft_toolkit::BASELINE_WRITE_VERSION, &payload)
            .expect("v4 encode");

        // The v4 bytes must equal encoding the frozen V4 struct directly,
        // so a pre-dense reader can still decode them.
        let v4_struct_bytes = tsoracle_openraft_toolkit::encode(
            tsoracle_openraft_toolkit::BASELINE_WRITE_VERSION,
            &HighWaterStateMachineSnapshotV4 {
                current_value: 42,
                last_applied: Some(log_id(3)),
                last_membership: StoredMem::default(),
            },
        )
        .expect("v4 struct direct encode");
        assert_eq!(
            framed, v4_struct_bytes,
            "v4 encode of empty-dense payload must be byte-identical to frozen V4 struct"
        );

        // Decode: the v4 bytes lift into the current layout with empty dense, cap 0.
        let decoded: HighWaterStateMachineSnapshot = decode_framed(
            tsoracle_openraft_toolkit::MIN_READABLE_VERSION,
            tsoracle_openraft_toolkit::MAX_READABLE_VERSION,
            &framed,
        )
        .expect("v4 decode");
        assert_eq!(
            decoded, payload,
            "v4 forward-compat decode must produce empty-dense payload"
        );
        assert!(decoded.dense.is_empty(), "decoded v4 dense must be empty");
        assert_eq!(
            decoded.dense_cap, 0,
            "decoded v4 cap must be 0 (absent sentinel)"
        );
    }

    #[test]
    fn snapshot_codec_v4_old_bytes_decode_lifts_to_empty_dense() {
        use tsoracle_codec::{decode_framed, encode_postcard};

        // Hand-encode a V4 struct (the 3-field pre-dense layout) as postcard,
        // then frame it as a v4 record. This simulates bytes written by an old
        // (pre-dense) node.
        let v4 = HighWaterStateMachineSnapshotV4 {
            current_value: 99,
            last_applied: Some(log_id(7)),
            last_membership: StoredMem::default(),
        };
        let v4_body = encode_postcard(&v4).expect("postcard encode V4");
        let mut framed = vec![tsoracle_openraft_toolkit::BASELINE_WRITE_VERSION];
        framed.extend_from_slice(&v4_body);

        // Decode through the VersionedCodec path.
        let decoded: HighWaterStateMachineSnapshot = decode_framed(
            tsoracle_openraft_toolkit::MIN_READABLE_VERSION,
            tsoracle_openraft_toolkit::MAX_READABLE_VERSION,
            &framed,
        )
        .expect("decode old v4 bytes");

        assert_eq!(decoded.current_value, 99, "current_value lifts correctly");
        assert_eq!(
            decoded.last_applied.map(|l| l.index),
            Some(7),
            "last_applied lifts correctly"
        );
        assert!(
            decoded.dense.is_empty(),
            "old v4 bytes must lift to empty dense map"
        );
        assert_eq!(
            decoded.dense_cap, 0,
            "old v4 bytes must lift to cap 0 (absent)"
        );
    }

    // ---- AdvanceDense apply tests ----
    //
    // Drive the `AdvanceDense` apply arm directly via `apply_one` and verify
    // the observable side effects through `dense_value()` (the pre/post
    // counter), the `last_applied` log position, and the fact that the
    // high-water is NOT touched by a dense advance.
    //
    // The four cases the plan mandates:
    //  1. First advance of "orders" by 5 → start=0, dense["orders"]=5.
    //  2. Second advance of "orders" by 3 → start=5, dense["orders"]=8.
    //  3. A new key when the cap is full → DenseCardinalityExceeded: counter
    //     stays 0 and the map size does not grow.
    //  4. A counter at u64::MAX with count=1 → DenseOverflow: counter
    //     stays at u64::MAX.
    //
    // We verify start and outcome by observing `dense_value` BEFORE the apply
    // (which gives `start`) and AFTER (which gives `start + count` on success,
    // unchanged on rejection). The high-water is pinned at 0 throughout (no
    // `Advance` entries), which confirms the apply arm does NOT touch it.

    fn dense_key(s: &str) -> tsoracle_core::SeqKey {
        tsoracle_core::SeqKey::try_new(s).expect("valid key")
    }

    #[tokio::test]
    async fn dense_advance_first_key_starts_at_zero() {
        let mut sm = HighWaterStateMachine::new();
        let key = dense_key("orders");

        // Pre-condition: key is absent → start would be 0.
        assert_eq!(sm.dense_value("orders"), 0);

        apply_one(
            &mut sm,
            1,
            EntryPayload::Normal(HighWaterCommand::AdvanceDense {
                key: key.clone(),
                count: 5,
            }),
        )
        .await;

        // Post-condition: counter advanced by 5; start was 0.
        assert_eq!(
            sm.dense_value("orders"),
            5,
            "first advance by 5: dense[orders] = 5 (start=0)"
        );
        // High-water must be untouched.
        assert_eq!(
            sm.current_value().await,
            0,
            "AdvanceDense must not touch the high-water"
        );
        let (last, _) = sm.applied_state().await.unwrap();
        assert_eq!(last.map(|l| l.index), Some(1));
    }

    #[tokio::test]
    async fn dense_advance_second_apply_starts_at_previous_value() {
        let mut sm = HighWaterStateMachine::new();
        let key = dense_key("orders");

        // First advance: start=0, next=5.
        apply_one(
            &mut sm,
            1,
            EntryPayload::Normal(HighWaterCommand::AdvanceDense {
                key: key.clone(),
                count: 5,
            }),
        )
        .await;
        assert_eq!(sm.dense_value("orders"), 5);

        // Capture start before second apply.
        let start_before = sm.dense_value("orders"); // == 5

        apply_one(
            &mut sm,
            2,
            EntryPayload::Normal(HighWaterCommand::AdvanceDense {
                key: key.clone(),
                count: 3,
            }),
        )
        .await;

        // The second block starts at 5 (the pre-advance value).
        assert_eq!(
            start_before, 5,
            "second advance: start must be the prior counter value (5)"
        );
        assert_eq!(
            sm.dense_value("orders"),
            8,
            "second advance by 3: dense[orders] = 8"
        );
    }

    #[tokio::test]
    async fn dense_advance_cardinality_cap_rejects_new_key() {
        // Cap set to 1: one key is allowed; a second new key must be rejected.
        let mut sm = HighWaterStateMachine::new_with_dense_cap(1);
        let k1 = dense_key("orders");
        let k2 = dense_key("users");

        // First key is accepted (map was empty, cap=1 → len 0 < 1 → OK).
        apply_one(
            &mut sm,
            1,
            EntryPayload::Normal(HighWaterCommand::AdvanceDense {
                key: k1.clone(),
                count: 1,
            }),
        )
        .await;
        assert_eq!(sm.dense_value("orders"), 1, "first key accepted");

        // Second NEW key: map len==1 >= cap=1 → DenseCardinalityExceeded.
        // The counter for the new key must stay at 0 (rejection is a no-op on
        // the dense map).
        apply_one(
            &mut sm,
            2,
            EntryPayload::Normal(HighWaterCommand::AdvanceDense {
                key: k2.clone(),
                count: 1,
            }),
        )
        .await;
        assert_eq!(
            sm.dense_value("users"),
            0,
            "new key rejected by cardinality cap: dense[users] must be 0"
        );
        // Existing key still advances (already present → cap check skipped).
        apply_one(
            &mut sm,
            3,
            EntryPayload::Normal(HighWaterCommand::AdvanceDense {
                key: k1.clone(),
                count: 4,
            }),
        )
        .await;
        assert_eq!(
            sm.dense_value("orders"),
            5,
            "existing key still advances despite cap being full"
        );
    }

    #[tokio::test]
    async fn dense_advance_overflow_leaves_counter_unchanged() {
        // Seed the counter at u64::MAX by using a large initial count and then
        // bridging up. Since count is u32 (max 2^32-1) we need multiple applies
        // to reach u64::MAX. It is simpler to directly manipulate the core in
        // the test module (same file, so private access is allowed).
        let sm = HighWaterStateMachine::new();
        {
            let mut core = sm.core.lock();
            core.dense.insert("x".to_string(), u64::MAX);
        }
        // Apply AdvanceDense with count=1: checked_add(u64::MAX, 1) overflows.
        let mut sm_mut = sm;
        apply_one(
            &mut sm_mut,
            1,
            EntryPayload::Normal(HighWaterCommand::AdvanceDense {
                key: dense_key("x"),
                count: 1,
            }),
        )
        .await;
        // Overflow: counter must remain at u64::MAX (rejection is a no-op).
        assert_eq!(
            sm_mut.dense_value("x"),
            u64::MAX,
            "overflow must not modify the counter"
        );
    }

    // ---- AdvanceDenseBatch apply tests ----
    //
    // These tests pin the all-or-nothing semantics of the batch apply arm:
    //  1. A successful batch advances ALL counters atomically; starts are
    //     the pre-advance values in request order (including duplicate-key
    //     blocks which are adjacent and non-overlapping within a single batch).
    //  2. A cardinality violation rejects the ENTIRE batch — no counter moves,
    //     not even keys already in the map.
    //  3. An overflow caused by ACCUMULATED advances across duplicate entries
    //     rejects the ENTIRE batch — a per-entry-against-prestate check would
    //     pass both entries individually, so the sequential fold is the only
    //     correct approach.
    //
    // Outcome is verified by observing `dense_value` before and after each
    // apply (the same mechanism the single-key AdvanceDense tests use), since
    // `ApplyOutcome` is only delivered via the openraft responder channel,
    // which the test harness does not wire.

    use crate::log_entry::DenseAdvance;

    #[tokio::test]
    async fn dense_batch_applies_atomically_and_gaplessly() {
        // A two-key batch: both keys start absent (counter = 0) and the batch
        // advances "orders" by 5 and "users" by 2. After the batch both counters
        // must reflect exactly those advances — gaplessly starting from 0.
        let mut sm = HighWaterStateMachine::new_with_dense_cap(8);

        // Pre-condition: both keys absent.
        assert_eq!(sm.dense_value("orders"), 0);
        assert_eq!(sm.dense_value("users"), 0);

        apply_one(
            &mut sm,
            1,
            EntryPayload::Normal(HighWaterCommand::AdvanceDenseBatch {
                entries: vec![
                    DenseAdvance {
                        key: dense_key("orders"),
                        count: 5,
                    },
                    DenseAdvance {
                        key: dense_key("users"),
                        count: 2,
                    },
                ],
            }),
        )
        .await;

        // Post-condition: both counters advanced from 0.
        assert_eq!(
            sm.dense_value("orders"),
            5,
            "first batch: orders must advance from 0 to 5"
        );
        assert_eq!(
            sm.dense_value("users"),
            2,
            "first batch: users must advance from 0 to 2"
        );
        // High-water must be untouched.
        assert_eq!(
            sm.current_value().await,
            0,
            "AdvanceDenseBatch must not touch the high-water"
        );

        // Second batch: advance "orders" by 3 — start must be 5 (current value).
        apply_one(
            &mut sm,
            2,
            EntryPayload::Normal(HighWaterCommand::AdvanceDenseBatch {
                entries: vec![DenseAdvance {
                    key: dense_key("orders"),
                    count: 3,
                }],
            }),
        )
        .await;

        assert_eq!(
            sm.dense_value("orders"),
            8,
            "second batch: orders must advance from 5 to 8"
        );
        assert_eq!(
            sm.dense_value("users"),
            2,
            "users must be unchanged by a batch that does not mention it"
        );
    }

    #[tokio::test]
    async fn dense_batch_cardinality_is_atomic() {
        // Cap = 1: only one distinct key is permitted. Seed "a" with a first
        // batch (1 key, accepted), then submit a batch of ["a", "b"] — "b" is
        // new and would push the count to 2 > cap. The ENTIRE batch must be
        // rejected atomically: "a" must NOT advance and "b" must remain absent.
        let mut sm = HighWaterStateMachine::new_with_dense_cap(1);

        // Seed "a" at 1.
        apply_one(
            &mut sm,
            1,
            EntryPayload::Normal(HighWaterCommand::AdvanceDenseBatch {
                entries: vec![DenseAdvance {
                    key: dense_key("a"),
                    count: 1,
                }],
            }),
        )
        .await;
        assert_eq!(sm.dense_value("a"), 1, "seed apply must succeed");

        // The cardinality-exceeded batch: "a" already exists (no new key there)
        // but "b" is new and would push the distinct-key count to 2 > cap=1.
        apply_one(
            &mut sm,
            2,
            EntryPayload::Normal(HighWaterCommand::AdvanceDenseBatch {
                entries: vec![
                    DenseAdvance {
                        key: dense_key("a"),
                        count: 1,
                    },
                    DenseAdvance {
                        key: dense_key("b"),
                        count: 1,
                    },
                ],
            }),
        )
        .await;

        // Atomic rejection: neither "a" nor "b" must move.
        assert_eq!(
            sm.dense_value("a"),
            1,
            "cardinality rejection must not advance existing key a"
        );
        assert_eq!(
            sm.dense_value("b"),
            0,
            "cardinality rejection must leave new key b absent"
        );
    }

    #[tokio::test]
    async fn dense_batch_overflow_is_atomic_and_accumulates_duplicates() {
        // Seed "k" near the ceiling so that two entries on "k" within a single
        // batch — each advancing by 4 — ACCUMULATE to an overflow: the running
        // value after the first entry is (u64::MAX - 5 + 4) = u64::MAX - 1,
        // and after the second entry checked_add(u64::MAX - 1, 4) overflows.
        // A naive per-entry check against the pre-batch stored value (u64::MAX - 5)
        // would pass both (each count=4 fits), so the sequential fold over a
        // scratch map is the only correct implementation.
        //
        // The batch must be rejected entirely: "k" must remain at u64::MAX - 5.
        let mut sm = HighWaterStateMachine::new_with_dense_cap(8);
        {
            let mut core = sm.core.lock();
            core.dense.insert("k".to_string(), u64::MAX - 5);
        }

        apply_one(
            &mut sm,
            1,
            EntryPayload::Normal(HighWaterCommand::AdvanceDenseBatch {
                entries: vec![
                    DenseAdvance {
                        key: dense_key("k"),
                        count: 4,
                    },
                    DenseAdvance {
                        key: dense_key("k"),
                        count: 4,
                    },
                ],
            }),
        )
        .await;

        // Entire batch must be rejected: k stays at u64::MAX - 5.
        assert_eq!(
            sm.dense_value("k"),
            u64::MAX - 5,
            "overflow from accumulated duplicate advances must not modify any counter"
        );
    }

    // ---- Regression test: install_snapshot advances active_write_version ----
    //
    // Bug: a follower that received a v5 snapshot whose `SetFormatVersion` entry
    // had already been purged into the snapshot would keep `active_write_version`
    // at `BASELINE_WRITE_VERSION` (4) while now holding a populated dense map.
    // The next `build_snapshot` call would then hit the fail-loud
    // `NotRepresentable` guard in `encode_version`'s v4 arm and return an error.
    //
    // Fix: after decoding the incoming snapshot, `install_snapshot` now advances
    // the `active_write_version` cell to the snapshot's leading version byte if
    // it exceeds the current cell value.

    #[tokio::test]
    async fn install_snapshot_advances_active_write_version_to_snapshot_byte() {
        // ---- Step 1: build a v5-framed snapshot on a source SM ----
        //
        // Create a source SM whose active_write_version is DENSE_WRITE_VERSION
        // (5) and whose dense map is non-empty — the combination that would have
        // triggered the NotRepresentable guard on the follower.
        let v5_cell = ActiveWriteVersion::default();
        v5_cell.set(DENSE_WRITE_VERSION);

        let store_src: Arc<dyn SnapshotStore> = Arc::new(InMemorySnapshotStore::new());
        let mut src =
            HighWaterStateMachine::with_store_and_active_version(store_src, v5_cell.clone())
                .expect("source SM");

        // Populate the dense map and advance the high-water so the snapshot is
        // non-trivial.  `with_store_and_active_version` starts at the genesis
        // cardinality cap; apply an AdvanceDense to ensure dense is non-empty.
        apply_one(
            &mut src,
            1,
            EntryPayload::Normal(HighWaterCommand::AdvanceDense {
                key: dense_key("orders"),
                count: 42,
            }),
        )
        .await;

        assert_eq!(src.active_write_version(), DENSE_WRITE_VERSION);
        assert_eq!(src.dense_value("orders"), 42, "source dense populated");

        // Build the v5 snapshot: the bytes handed back by openraft are what
        // gets streamed to the follower via `begin_receiving_snapshot` +
        // `install_snapshot`.
        let snap = src
            .build_snapshot()
            .await
            .expect("build_snapshot on source");
        let snap_meta = snap.meta.clone();
        // `snap.snapshot` is a `Cursor<Vec<u8>>`; the inner Vec is the framed
        // payload bytes that `install_snapshot` receives.
        let snap_bytes = snap.snapshot.into_inner();

        // Confirm the snapshot is v5-framed (the first byte is the version).
        assert_eq!(
            snap_bytes[0], DENSE_WRITE_VERSION,
            "source snapshot must be v5-framed"
        );

        // ---- Step 2: build a fresh target SM at BASELINE_WRITE_VERSION ----
        //
        // This models a newly-joining follower that never applied the
        // SetFormatVersion entry (it was already purged into the snapshot it is
        // about to receive).
        let mut target = HighWaterStateMachine::new();
        assert_eq!(
            target.active_write_version(),
            BASELINE_WRITE_VERSION,
            "target must start at baseline (the pre-bug state)"
        );
        assert_eq!(target.dense_value("orders"), 0, "target dense starts empty");

        // ---- Step 3: install the v5 snapshot onto the target ----
        target
            .install_snapshot(&snap_meta, std::io::Cursor::new(snap_bytes))
            .await
            .expect("install_snapshot must succeed");

        // ---- Step 4: regression assertions ----

        // PRIMARY: the cell must advance to DENSE_WRITE_VERSION after install.
        // Before the fix this stayed at BASELINE_WRITE_VERSION (4) — the exact
        // bug being pinned by this test.
        assert_eq!(
            target.active_write_version(),
            DENSE_WRITE_VERSION,
            "install_snapshot must advance active_write_version to the snapshot's \
             leading version byte (regression: was left at BASELINE before fix)"
        );

        // Dense state must be restored from the snapshot.
        assert_eq!(
            target.dense_value("orders"),
            42,
            "dense map must be restored from the installed snapshot"
        );

        // CRITICAL: build_snapshot on the target must now SUCCEED.  Before the
        // fix, this would error with NotRepresentable because the v4 encode arm
        // would be selected (cell=4) while dense is non-empty.
        target
            .build_snapshot()
            .await
            .expect("build_snapshot on target after install must NOT return NotRepresentable");
    }

    // ---- Regression test: with_store recovers active_write_version ----
    //
    // Sibling of `install_snapshot_advances_active_write_version_to_snapshot_byte`.
    // Bug: the public `with_store` / `with_store_and_active_version` load path
    // rehydrated `core.dense` (and `dense_cap`) from a persisted v5 snapshot but
    // left the freshly-defaulted `active_write_version` cell at
    // `BASELINE_WRITE_VERSION` (4). The next `build_snapshot` would then select
    // the v4 encode arm with a non-empty dense map and hit the fail-loud
    // `NotRepresentable` guard; `advance_dense` would also stay gated.
    //
    // Production dodges this because standalone bootstrap recovers the cell from
    // log + snapshot evidence *before* constructing the SM, but the public
    // `with_store` API documents durable-snapshot rehydration, so the bare path
    // must recover the version byte too (a safe floor: the snapshot byte proves
    // an activation to at least that version committed cluster-wide).
    //
    // Fix: after `store.load()`, advance the cell to the persisted envelope's
    // leading version byte when it exceeds the cell's current value.

    #[tokio::test]
    async fn with_store_recovers_active_write_version_from_v5_snapshot() {
        // ---- Step 1: persist a v5/dense snapshot through a source SM ----
        let v5_cell = ActiveWriteVersion::default();
        v5_cell.set(DENSE_WRITE_VERSION);

        let store: Arc<dyn SnapshotStore> = Arc::new(InMemorySnapshotStore::new());
        {
            let mut src = HighWaterStateMachine::with_store_and_active_version(
                store.clone(),
                v5_cell.clone(),
            )
            .expect("source SM");
            apply_one(
                &mut src,
                1,
                EntryPayload::Normal(HighWaterCommand::AdvanceDense {
                    key: dense_key("orders"),
                    count: 42,
                }),
            )
            .await;
            src.build_snapshot().await.expect("build v5 snapshot");
        }

        // Confirm the persisted envelope is v5-framed.
        let persisted_bytes = store.load().expect("load").expect("snapshot present");
        assert_eq!(
            persisted_bytes[0], DENSE_WRITE_VERSION,
            "persisted envelope must be v5-framed"
        );

        // ---- Step 2: reopen via the bare public API (fresh BASELINE cell) ----
        // This models an embedder that uses `with_store` with a durable backend
        // holding v5 data — the pre-bug state left the cell at baseline.
        let mut reopened =
            HighWaterStateMachine::with_store(store.clone()).expect("reopen via with_store");

        // ---- Step 3: regression assertions ----

        // PRIMARY: the cell must recover to DENSE_WRITE_VERSION from the
        // snapshot's leading byte. Before the fix this stayed at BASELINE (4).
        assert_eq!(
            reopened.active_write_version(),
            DENSE_WRITE_VERSION,
            "with_store must recover active_write_version from the persisted \
             snapshot's leading version byte (regression: was left at BASELINE)"
        );

        // Dense state must be rehydrated from the snapshot.
        assert_eq!(
            reopened.dense_value("orders"),
            42,
            "dense map must be rehydrated from the persisted snapshot"
        );

        // CRITICAL: build_snapshot must now SUCCEED. Before the fix, the v4
        // encode arm (cell=4) with a non-empty dense map hit NotRepresentable.
        reopened
            .build_snapshot()
            .await
            .expect("build_snapshot after reopen must NOT return NotRepresentable");
    }
}