ethrex_storage/

store.rs

#[cfg(feature = "rocksdb")]
use crate::backend::rocksdb::RocksDBBackend;
use crate::{
    STORE_METADATA_FILENAME, STORE_SCHEMA_VERSION,
    api::{
        StorageBackend, StorageReadView, StorageWriteBatch,
        tables::{
            ACCOUNT_CODE_METADATA, ACCOUNT_CODES, ACCOUNT_FLATKEYVALUE, ACCOUNT_TRIE_NODES,
            BAD_BLOCKS, BLOCK_ACCESS_LISTS, BLOCK_NUMBERS, BODIES, CANONICAL_BLOCK_HASHES,
            CHAIN_DATA, EXECUTION_WITNESSES, FULLSYNC_HEADERS, HEADERS, INVALID_CHAINS,
            MISC_VALUES, PENDING_BLOCKS, RECEIPTS_V2, SNAP_STATE, STATE_HISTORY,
            STORAGE_FLATKEYVALUE, STORAGE_TRIE_NODES, TRANSACTION_LOCATIONS,
        },
    },
    apply_prefix,
    backend::in_memory::InMemoryBackend,
    block_data_buffer::BlockDataBuffer,
    error::StoreError,
    journal::{FlatDiff, JournalEntry},
    layering::{Overlay, OverlayCf, TrieLayerCache, TrieWrapper},
    rlp::{BlockBodyRLP, BlockHeaderRLP, BlockRLP},
    trie::{BackendTrieDB, BackendTrieDBLocked, classify_trie_key},
    utils::{ChainDataIndex, SnapStateIndex},
};

use ethrex_common::{
    Address, H256, U256,
    types::{
        AccountInfo, AccountState, AccountUpdate, Block, BlockBody, BlockHash, BlockHeader,
        BlockNumber, ChainConfig, Code, CodeMetadata, ForkId, Genesis, GenesisAccount, Index,
        Receipt, Transaction,
        block_access_list::BlockAccessList,
        block_execution_witness::{ExecutionWitness, RpcExecutionWitness},
    },
    utils::keccak,
};
use ethrex_crypto::{NativeCrypto, keccak::keccak_hash};
use ethrex_rlp::{
    decode::{RLPDecode, decode_bytes, decode_rlp_item},
    encode::RLPEncode,
};
use ethrex_trie::{EMPTY_TRIE_HASH, Nibbles, Trie, TrieLogger, TrieNode, TrieWitness};
use ethrex_trie::{Node, NodeRLP};
use lru::LruCache;
use rayon::prelude::*;
use rustc_hash::FxBuildHasher;
use serde::{Deserialize, Serialize};
use std::{
    collections::{BTreeMap, HashMap, HashSet, hash_map::Entry},
    fmt::Debug,
    io::Write,
    num::NonZeroUsize,
    path::{Path, PathBuf},
    sync::{
        Arc, Condvar, Mutex, RwLock,
        atomic::{AtomicUsize, Ordering},
        mpsc::{SyncSender, TryRecvError, sync_channel},
    },
    thread::JoinHandle,
};
use tracing::{debug, error, info, warn};

/// Maximum number of execution witnesses to keep in the database
pub const MAX_WITNESSES: u64 = 128;

// We use one constant for in-memory and another for on-disk backends.
// This is due to tests requiring state older than 128 blocks.
// TODO: unify these
pub const DB_COMMIT_THRESHOLD: usize = 128;
const IN_MEMORY_COMMIT_THRESHOLD: usize = 10000;

/// Depth-only commit threshold for batch execution (full sync / block import). Each batch
/// layer holds ~1024 blocks of trie diffs (~1 GB), so we flush after a few layers to bound
/// memory. The canonical `head - DB_COMMIT_THRESHOLD` safe-commit root never lands on a batch
/// layer boundary, so batch mode commits by depth instead; this is sound because full sync and
/// import only ever extend a single canonical chain (no competing forks to mis-commit).
pub const BATCH_COMMIT_THRESHOLD: usize = 4;

/// Sorted, sharded `multi_get` tuning, shared by the account, storage, and
/// trie-node batch read paths. A single `multi_get` runs at queue depth 1
/// (async_io is off), so cold batches are split into contiguous sorted shards
/// read on separate blocking threads. `KEYS_PER_SHARD` targets shard size,
/// `MAX_SHARDS` caps concurrency; scattered keys make 64 the sweet spot (128
/// over-fragments; coldbench).
const KEYS_PER_SHARD: usize = 256;
const MAX_SHARDS: usize = 64;

/// Shard count for a sorted, sharded `multi_get` over `n` keys.
fn shard_count(n: usize) -> usize {
    n.div_ceil(KEYS_PER_SHARD).clamp(1, MAX_SHARDS)
}

/// Ceiling on the RocksDB shared block cache: 12 GiB.
///
/// This cache holds both data blocks AND the index/bloom-filter blocks for every
/// open SST file (because we enable `cache_index_and_filter_blocks`), so its size
/// is the effective upper bound on RocksDB's resident memory footprint. 12 GiB
/// keeps the filter/index working set resident plus hot EVM state; a sweep on a
/// synced mainnet node (32 GiB cap) found 8-16 GiB all keep up with head-following,
/// with larger giving no gain (the OS page cache backstops the uncompressed state
/// CFs) and ~8 GiB the floor where the filter set starts to thrash.
pub const MAX_ROCKSDB_BLOCK_CACHE_SIZE_BYTES: usize = 12 * 1024 * 1024 * 1024;

/// Floor on the RocksDB shared block cache: 512 MiB. Below this the per-SST
/// index and filter blocks no longer stay resident and every trie read pays an
/// extra disk seek, which is worse than the memory it saves. Applied even when
/// it exceeds [`ROCKSDB_BLOCK_CACHE_MEMORY_PERCENT`] of a very small host, and
/// even when it exceeds the detected limit outright: a node given less than this
/// much memory cannot follow the chain anyway, so the floor is the more useful
/// failure mode than a cache too small to hold the filters.
pub const MIN_ROCKSDB_BLOCK_CACHE_SIZE_BYTES: usize = 512 * 1024 * 1024;

/// Share of the host's (or cgroup's) memory the block cache may claim by default.
///
/// The rest has to cover the in-memory trie-layer backlog, block execution, the
/// mempool, peer buffers and allocator slack, so the cache cannot have most of the
/// machine. At 40% a 32 GiB host lands on the 12 GiB ceiling and a 16 GiB host gets
/// ~6.4 GiB — under the ~8 GiB thrash floor, but that is the memory-constrained
/// tradeoff the ceiling's docs describe, and it leaves the node room not to be
/// OOM-killed.
pub const ROCKSDB_BLOCK_CACHE_MEMORY_PERCENT: usize = 40;

/// Default RocksDB shared block cache size, derived from the memory this process
/// is actually allowed to use.
///
/// A fixed default cannot serve both a 64 GiB validator and a 16 GiB CI runner: the
/// ceiling alone is 71% of a 16 GiB host, which leaves no headroom for the rest of
/// the node. Sizes to [`ROCKSDB_BLOCK_CACHE_MEMORY_PERCENT`] of the detected limit —
/// the smaller of physical memory and the cgroup limit, so a container is sized
/// against its own allowance rather than the machine it lands on — clamped to
/// [`MIN_ROCKSDB_BLOCK_CACHE_SIZE_BYTES`] ..=
/// [`MAX_ROCKSDB_BLOCK_CACHE_SIZE_BYTES`]. Falls back to the ceiling when the limit
/// cannot be detected, preserving the previous behavior.
pub fn default_rocksdb_block_cache_size() -> usize {
    let physical = physical_memory_bytes();
    let cgroup = cgroup_memory_limit_bytes();
    let limit = [physical, cgroup].into_iter().flatten().min();
    let size = rocksdb_block_cache_size_for(limit);
    // The cache size used to be a knowable constant; now it depends on the
    // environment, on a code path that exists because a mis-sized cache was
    // only ever discovered via an OOM kill. Leave the whole derivation in the
    // log: which detector won, what it read, and what that resolved to.
    let source = match (physical, cgroup) {
        (None, None) => "undetected, falling back to the ceiling",
        (Some(_), None) => "physical memory",
        (None, Some(_)) => "cgroup limit",
        (Some(p), Some(c)) => {
            if c <= p {
                "cgroup limit"
            } else {
                "physical memory"
            }
        }
    };
    info!(
        detected_limit_bytes = ?limit,
        source,
        cache_bytes = size,
        "sized RocksDB block cache"
    );
    size
}

/// Pure part of [`default_rocksdb_block_cache_size`]: the clamp, with the detected
/// memory limit supplied by the caller. `None` means detection failed.
pub fn rocksdb_block_cache_size_for(memory_limit: Option<usize>) -> usize {
    let Some(limit) = memory_limit else {
        return MAX_ROCKSDB_BLOCK_CACHE_SIZE_BYTES;
    };
    (limit.saturating_mul(ROCKSDB_BLOCK_CACHE_MEMORY_PERCENT) / 100).clamp(
        MIN_ROCKSDB_BLOCK_CACHE_SIZE_BYTES,
        MAX_ROCKSDB_BLOCK_CACHE_SIZE_BYTES,
    )
}

/// `MemTotal` from `/proc/meminfo`, in bytes.
fn physical_memory_bytes() -> Option<usize> {
    let meminfo = std::fs::read_to_string("/proc/meminfo").ok()?;
    let kib: usize = meminfo
        .lines()
        .find_map(|line| line.strip_prefix("MemTotal:"))?
        .split_whitespace()
        .next()?
        .parse()
        .ok()?;
    kib.checked_mul(1024)
}

/// The cgroup memory limit that applies to this process, in bytes: `memory.max`
/// under cgroup v2, falling back to `memory.limit_in_bytes` under v1. Both report
/// "no limit" out of band — v2 with the literal `max`, v1 with a sentinel near
/// `u64::MAX` — which yields `None` here (v1's sentinel simply exceeds any real
/// `MemTotal`, so the `min` above discards it).
///
/// The limit may sit on any cgroup between the one the process belongs to and the
/// mount root — a systemd slice, a Kubernetes pod's parent, an outer container — and
/// the mount root itself is usually unlimited, so reading the root alone would miss
/// it. The process's own cgroup comes from `/proc/self/cgroup` and every level up to
/// the root is read, with the smallest limit winning.
fn cgroup_memory_limit_bytes() -> Option<usize> {
    let cgroups = std::fs::read_to_string("/proc/self/cgroup").ok()?;
    let v2 = cgroups
        .lines()
        .find_map(|line| line.strip_prefix("0::"))
        .and_then(|relative| {
            cgroup_limit_up_to_root(Path::new("/sys/fs/cgroup"), relative, "memory.max")
        });
    v2.or_else(|| {
        let relative = cgroups.lines().find_map(|line| {
            // `hierarchy-ID:controller-list:cgroup-path`
            let mut fields = line.splitn(3, ':');
            let controllers = fields.nth(1)?;
            let path = fields.next()?;
            controllers
                .split(',')
                .any(|controller| controller == "memory")
                .then_some(path)
        })?;
        cgroup_limit_up_to_root(
            Path::new("/sys/fs/cgroup/memory"),
            relative,
            "memory.limit_in_bytes",
        )
    })
}

/// Smallest value of `file` across `<mount>/<relative>` and each of its ancestors up
/// to `mount`, skipping levels whose file is absent or reports no limit.
fn cgroup_limit_up_to_root(mount: &Path, relative: &str, file: &str) -> Option<usize> {
    let mut dir = mount.join(relative.trim_start_matches('/'));
    let mut limit: Option<usize> = None;
    loop {
        if let Some(value) = std::fs::read_to_string(dir.join(file))
            .ok()
            .and_then(|contents| contents.trim().parse::<usize>().ok())
        {
            limit = Some(limit.map_or(value, |current| current.min(value)));
        }
        if dir == mount || !dir.pop() {
            return limit;
        }
    }
}

/// Tunable configuration for [`Store::new_with_config`] and related constructors.
///
/// Use [`StoreConfig::default()`] for production-tuned defaults; callers that
/// don't need to override anything should keep calling [`Store::new`] directly.
#[derive(Debug, Clone, Copy)]
pub struct StoreConfig {
    /// Size in bytes of the RocksDB shared block cache. With
    /// `cache_index_and_filter_blocks` enabled (the ethrex default), this is
    /// the effective ceiling on RocksDB's resident memory. Ignored for
    /// in-memory backends.
    pub rocksdb_block_cache_size: usize,
    /// Bound on the persist worker's channel: number of staged (acked) live
    /// messages whose flush may still be in flight. Once full, the next send
    /// blocks — that is the backpressure that throttles `newPayload`.
    /// Clamped to `max(1)` at construction (0 would make a rendezvous channel).
    pub persist_channel_capacity: usize,
}

impl StoreConfig {
    /// Every tuned default except the block cache, which the caller supplies.
    /// An explicit cache size must come through here rather than overwriting a
    /// [`StoreConfig::default()`], which would run (and log) the memory
    /// detection only to discard its result.
    pub fn with_rocksdb_block_cache_size(rocksdb_block_cache_size: usize) -> Self {
        Self {
            rocksdb_block_cache_size,
            persist_channel_capacity: DEFAULT_PERSIST_CHANNEL_CAPACITY,
        }
    }
}

impl Default for StoreConfig {
    fn default() -> Self {
        Self::with_rocksdb_block_cache_size(default_rocksdb_block_cache_size())
    }
}

/// Control messages for the FlatKeyValue generator
#[derive(Debug, PartialEq)]
enum FKVGeneratorControlMessage {
    Stop,
    Continue,
}

/// Byte budget for the in-memory bytecode cache.
///
/// Bytecode is a small fraction of the working set next to trie nodes and flat
/// key-values (which the RocksDB block cache serves), but a cold code read costs a
/// blob-file fetch, so caching contracts pays for itself: 256 MiB holds ~10k max-size
/// (24 KiB) contracts, or ~100k of typical size.
const CODE_CACHE_MAX_SIZE: u64 = 256 * 1024 * 1024;

/// Entry bound for [`Store::code_metadata_cache`], derived from a 16 MiB ceiling at the
/// ~64 B an `LruCache` entry costs (32 B key + 8 B value + list/table overhead). Sized
/// against the code cache it shadows: at 256 MiB that holds ~10k max-size or ~100k
/// typical contracts, so this keeps a length for every code that could be resident and
/// then some, while bounding what an `EXTCODESIZE` sweep over unique contracts can pin.
const CODE_METADATA_CACHE_MAX_ENTRIES: usize = (16 * 1024 * 1024) / 64;

/// Key used to persist the `flushed_upto` block number in `MISC_VALUES`.
const FLUSHED_UPTO_KEY: &[u8] = b"bodies_flushed_upto";

/// Single key under which the bounded list of bad blocks is stored in `BAD_BLOCKS`.
const BAD_BLOCKS_KEY: &[u8] = b"bad_blocks";

/// Maximum number of bad blocks retained for `debug_getBadBlocks`.
const MAX_BAD_BLOCKS: usize = 16;

#[derive(Debug)]
struct CodeCache {
    inner_cache: LruCache<H256, Code, FxBuildHasher>,
    cache_size: u64,
    max_size: u64,
}

impl Default for CodeCache {
    fn default() -> Self {
        Self {
            inner_cache: LruCache::unbounded_with_hasher(FxBuildHasher),
            cache_size: 0,
            max_size: CODE_CACHE_MAX_SIZE,
        }
    }
}

impl CodeCache {
    fn get(&mut self, code_hash: &H256) -> Result<Option<Code>, StoreError> {
        Ok(self.inner_cache.get(code_hash).cloned())
    }

    fn insert(&mut self, code: &Code) -> Result<(), StoreError> {
        // A hash already cached must not be added to `cache_size` again, or the counter
        // drifts up permanently and evicts entries that fit. `get` also refreshes
        // recency, which is what a repeated read should do.
        if self.inner_cache.get(&code.hash).is_some() {
            return Ok(());
        }

        self.cache_size += code.size() as u64;
        self.inner_cache.put(code.hash, code.clone());

        while self.cache_size > self.max_size {
            if let Some((_, code)) = self.inner_cache.pop_lru() {
                self.cache_size -= code.size() as u64;
            } else {
                break;
            }
        }

        let cache_len = self.inner_cache.len();
        let current_size = self.cache_size;
        debug!(
            "[ACCOUNT CODE CACHE] cache elements (): {cache_len}, total size: {current_size} bytes"
        );
        Ok(())
    }
}

/// Main storage interface for the ethrex client.
///
/// `Store` is `Clone` and thread-safe; all clones share the same backend and
/// caches via `Arc`. Reads consult an in-memory block-data buffer before disk
/// so not-yet-flushed blocks are always visible.
#[derive(Debug, Clone)]
pub struct Store {
    /// Path to the database directory.
    db_path: PathBuf,
    /// Storage backend (InMemory or RocksDB).
    backend: Arc<dyn StorageBackend>,
    /// Chain configuration (fork schedule, chain ID, etc.).
    chain_config: ChainConfig,
    /// Cache for trie nodes from recent blocks.
    trie_cache: Arc<RwLock<Arc<TrieLayerCache>>>,
    /// Channel for controlling the FlatKeyValue generator background task.
    flatkeyvalue_control_tx: std::sync::mpsc::SyncSender<FKVGeneratorControlMessage>,
    /// In-memory overlay of block data not yet flushed to disk.
    block_data_buffer: Arc<RwLock<Arc<BlockDataBuffer>>>,
    /// Channel to the single persist worker (`apply_updates` → `PersistMessage::Block`,
    /// `wait_for_persistence_idle` → `PersistMessage::Ping`). The worker is the
    /// sole mutator of `block_data_buffer` in production.
    persist_tx: std::sync::mpsc::SyncSender<PersistMessage>,
    /// Roots whose trie diff-layer is being built but not yet installed in
    /// `trie_cache`. Trie opens block on these so a just-added block's state is
    /// never read as stale before its layer lands.
    pending_trie_roots: Arc<PendingTrieRoots>,
    /// Cached latest canonical block header. May be slightly stale, which is
    /// acceptable for RPC "latest" queries and sync operations.
    latest_block_header: LatestBlockHeaderCache,
    /// Last computed FlatKeyValue for incremental updates.
    last_computed_flatkeyvalue: Arc<RwLock<Vec<u8>>>,

    /// Cache for account bytecodes, keyed by the bytecode hash.
    /// Note that we don't remove entries on account code changes, since
    /// those changes already affect the code hash stored in the account, and only
    /// may result in this cache having useless data.
    account_code_cache: Arc<Mutex<CodeCache>>,

    /// Cache for code metadata (code length), keyed by the bytecode hash.
    ///
    /// Bounded: `EXTCODESIZE` reads this on the execution path, so an unbounded map
    /// would grow by one entry per distinct contract ever asked for a length and never
    /// give the memory back. See [`CODE_METADATA_CACHE_MAX_ENTRIES`].
    code_metadata_cache: Arc<Mutex<LruCache<H256, CodeMetadata, FxBuildHasher>>>,

    /// Serializes concurrent `forkchoice_update` callers so that the cache
    /// update and the DB write transaction remain mutually ordered.
    fcu_lock: Arc<tokio::sync::Mutex<()>>,

    /// Canonical safe-commit state root, computed after each forkchoice update.
    ///
    /// Shared with the [`TrieLayerCache`] so that the Store can update the cell without
    /// replacing the cache Arc. `H256::zero()` means "no safe commit point yet".
    /// Cloning `Store` shares this cell across all clones, which is required and correct.
    safe_commit_root: Arc<RwLock<H256>>,

    /// While set, `forkchoice_update_inner` skips STATE_HISTORY pruning (the
    /// finalized-number update still lands). Set by `Blockchain::enter_reorg` for
    /// the duration of a deep-reorg apply pass: `Overlay::from_journal` reads
    /// journal entries one by one with no snapshot isolation, and syncer-driven
    /// forkchoice updates are not gated by the reorg mutex, so a concurrent
    /// finality advance could otherwise prune entries out from under overlay
    /// construction (spurious `MissingEntry`) or between a case-1 attempt and its
    /// retry. Pruning catches up on the first finality advance after the pass
    /// ends (`delete_range` is cumulative).
    journal_pruning_paused: Arc<std::sync::atomic::AtomicBool>,

    background_threads: Arc<ThreadList>,
}

#[derive(Debug, Default)]
struct ThreadList {
    list: Vec<JoinHandle<()>>,
}

impl Drop for ThreadList {
    fn drop(&mut self) {
        for handle in self.list.drain(..) {
            let _ = handle.join();
        }
    }
}

/// Storage trie nodes grouped by account address hash.
///
/// Each entry contains the hashed account address and the trie nodes
/// for that account's storage trie.
pub type StorageTrieNodes = Vec<(H256, Vec<(Nibbles, Vec<u8>)>)>;
type StorageTries = HashMap<Address, (TrieWitness, Trie)>;

/// Storage backend type selection.
///
/// Used when creating a new [`Store`] to specify which backend to use.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum EngineType {
    /// In-memory storage, non-persistent. Suitable for testing.
    InMemory,
    /// RocksDB storage, persistent. Suitable for production.
    #[cfg(feature = "rocksdb")]
    RocksDB,
}

/// Batch of updates to apply to the store atomically.
///
/// Used during block execution to collect all state changes before
/// committing them to the database in a single transaction.
pub struct UpdateBatch {
    /// New nodes to add to the state trie.
    pub account_updates: Vec<TrieNode>,
    /// Storage trie updates per account (keyed by hashed address).
    pub storage_updates: Vec<(H256, Vec<TrieNode>)>,
    /// Blocks to store.
    pub blocks: Vec<Block>,
    /// Receipts to store, grouped by block hash.
    pub receipts: Vec<(H256, Vec<Receipt>)>,
    /// Contract code updates (code hash -> bytecode).
    pub code_updates: Vec<(H256, Code)>,
    /// Commit gate for this batch's trie layers (independent of `wait_for_flush`).
    ///
    /// - `None`: live path (`newPayload`). The persist worker commits by the canonical
    ///   `head - DB_COMMIT_THRESHOLD` safe-commit root.
    /// - `Some(depth)`: single-canonical-chain execution (batch import, full sync, startup
    ///   state regeneration). The persist worker commits every layer deeper than `depth`
    ///   (see [`Trie::get_commitable_by_depth`]), which bounds resident trie layers to ~`depth`.
    pub commit_depth: Option<usize>,
    /// When the persist worker acks (independent of `commit_depth`).
    ///
    /// - `false`: ack after staging, so the caller's next-block execution overlaps this
    ///   block's disk flush. Used by the live path and the per-block re-execution paths
    ///   (regen / full-sync fallback / import tail) — their memory is already bounded by
    ///   `commit_depth`'s depth gate and the persist channel capacity, so waiting per block
    ///   would only serialize CPU and I/O for no benefit.
    /// - `true`: ack after flush, bounding in-flight work to ~1 message. Used by the bespoke
    ///   batch path, where a single message carries ~1024 blocks (~1 GB of trie diff) and two
    ///   in flight would be a real memory cost.
    pub wait_for_flush: bool,
}

/// A block's historical chain data to backfill below the snap-sync pivot: the
/// (already-canonical) header, its body, and its receipts.
///
/// `index_transactions` controls whether the transaction-lookup index is written
/// for this block; the backfill task drives it from the `--history.transactions`
/// horizon so old blocks can be excluded from the index while still storing their
/// bodies and receipts.
pub struct BackfilledBlock {
    pub header: BlockHeader,
    pub body: BlockBody,
    pub receipts: Vec<Receipt>,
    pub index_transactions: bool,
}

/// Storage trie updates grouped by account address hash.
pub type StorageUpdates = Vec<(H256, Vec<(Nibbles, Vec<u8>)>)>;

/// Collection of account state changes from block execution.
///
/// Contains all the data needed to update the state trie after
/// executing a block: account updates, storage updates, and code deployments.
pub struct AccountUpdatesList {
    /// Root hash of the state trie after applying these updates.
    pub state_trie_hash: H256,
    /// State trie node updates (path -> RLP-encoded node).
    pub state_updates: Vec<(Nibbles, Vec<u8>)>,
    /// Storage trie updates per account.
    pub storage_updates: StorageUpdates,
    /// New contract bytecode deployments.
    pub code_updates: Vec<(H256, Code)>,
}

/// Encodes a tx-location entry as the operand passed to `merge_cf`.
///
/// The operand uses the **same encoding as the stored value** — a
/// `Vec<(BlockNumber, BlockHash, Index)>` with a single element. This is
/// required for an *associative* merge operator: RocksDB folds operands
/// together with PartialMerge (during compaction, without a base value), and
/// the result becomes an operand for a later merge. If the operand format
/// differed from the merge output (e.g. operand = bare tuple, output = Vec),
/// the re-fed result would fail to decode and entries would be silently
/// dropped. Keeping both as `Vec` makes the merge truly associative.
pub(crate) fn encode_tx_location_operand(
    block_number: BlockNumber,
    block_hash: BlockHash,
    index: Index,
) -> Vec<u8> {
    vec![(block_number, block_hash, index)].encode_to_vec()
}

/// Merge function for the `TRANSACTION_LOCATIONS` column family.
///
/// The CF stores `Vec<(BlockNumber, BlockHash, Index)>` keyed by tx hash.
/// Both stored values and operands use this same `Vec` encoding — this
/// associativity requirement is mandatory: RocksDB folds operands together
/// during compaction without a base value (PartialMerge), then feeds that
/// result back into a later merge. A differing format would silently drop
/// entries. See `encode_tx_location_operand`.
///
/// Within the fold, a later entry with the same `block_hash` replaces an
/// earlier one (reorg dedupe). On decode failure the merge returns `None`
/// so RocksDB surfaces a corruption error rather than silently dropping
/// locations.
///
/// Merge instead of read-modify-write avoids the ~5–20 ms/block per-tx point
/// lookup on the write path; consolidation is deferred to compaction or the
/// next read.
pub fn tx_locations_merge(
    existing: Option<&[u8]>,
    operands: impl IntoIterator<Item = impl AsRef<[u8]>>,
) -> Option<Vec<u8>> {
    // Fold one RLP-encoded `Vec` chunk into `list`, deduping by block_hash
    // (later entry wins). Returns false on decode failure so the caller can
    // abort the whole merge.
    fn fold_chunk(
        list: &mut Vec<(BlockNumber, BlockHash, Index)>,
        bytes: &[u8],
        what: &str,
    ) -> bool {
        match <Vec<(BlockNumber, BlockHash, Index)>>::decode(bytes) {
            Ok(entries) => {
                for (bn, bh, idx) in entries {
                    list.retain(|(_, existing_bh, _)| *existing_bh != bh);
                    list.push((bn, bh, idx));
                }
                true
            }
            Err(e) => {
                error!(
                    "tx_locations_merge: failed to decode {what} ({} bytes): {e}; \
                     aborting merge to avoid silent data loss",
                    bytes.len()
                );
                false
            }
        }
    }

    let mut list: Vec<(BlockNumber, BlockHash, Index)> = Vec::new();

    // Order matters: RocksDB delivers operands oldest-first.
    if let Some(bytes) = existing
        && !fold_chunk(&mut list, bytes, "existing value")
    {
        return None;
    }
    for op in operands {
        if !fold_chunk(&mut list, op.as_ref(), "operand") {
            return None;
        }
    }
    Some(list.encode_to_vec())
}

impl Store {
    /// Block until the persist worker has fully processed all previously-sent
    /// `Block` messages (staged, trie-layer built, flushed, evicted).
    ///
    /// Uses an ack-based `Ping` rather than a bare send because the channel is
    /// buffered — a bare send proves nothing about prior message completion. The
    /// worker is FIFO, so it handles the `Ping` only after every earlier `Block`
    /// is done.
    ///
    /// Concurrent-producer caveat: if another thread sends a `Block` after the
    /// `Ping` is enqueued, that block may not be flushed by the time this returns.
    pub async fn wait_for_persistence_idle(&self) -> Result<(), StoreError> {
        let tx = self.persist_tx.clone();
        tokio::task::spawn_blocking(move || {
            let (ack_tx, ack_rx) = sync_channel::<Result<(), StoreError>>(1);
            tx.send(PersistMessage::Ping(ack_tx))
                .map_err(|e| StoreError::Custom(format!("wait_for_persistence_idle send: {e}")))?;
            ack_rx
                .recv()
                .map_err(|e| StoreError::Custom(format!("wait_for_persistence_idle ack: {e}")))?
        })
        .await
        .map_err(|e| StoreError::Custom(format!("wait_for_persistence_idle join: {e}")))?
    }

    /// Flushes all in-memory state to disk for a clean shutdown.
    ///
    /// Sends a `Shutdown` handshake to the persist worker, which (being FIFO)
    /// first drains every queued `Block`, then force-flushes the block-data
    /// buffer to disk. Once the worker acks, this syncs the backend (memtables +
    /// WAL) so the next process start needs no WAL recovery.
    ///
    /// The in-memory trie diff-layers are intentionally *not* force-committed.
    /// The on-disk trie is a single-version, path-based store, so folding the
    /// non-finalized tail into it would leave a post-restart reorg unable to
    /// reconstruct the overwritten ancestor state (the node would wedge). The
    /// recent (< `DB_COMMIT_THRESHOLD`) layers are dropped and re-executed on the
    /// next start from the deep, reorg-safe on-disk base — exactly as after any
    /// restart today.
    ///
    /// After this returns the persist worker has exited; the store must not be
    /// used for further writes. Idempotent only in the sense that a second call
    /// errors on the closed channel — call it exactly once, on shutdown.
    pub async fn shutdown(&self) -> Result<(), StoreError> {
        let tx = self.persist_tx.clone();
        let backend = self.backend.clone();
        tokio::task::spawn_blocking(move || {
            let (ack_tx, ack_rx) = sync_channel::<Result<(), StoreError>>(1);
            tx.send(PersistMessage::Shutdown { ack: ack_tx })
                .map_err(|e| StoreError::Custom(format!("shutdown send: {e}")))?;
            ack_rx
                .recv()
                .map_err(|e| StoreError::Custom(format!("shutdown ack: {e}")))??;
            // Worker has flushed block data to the WAL/memtables; make it durable
            // and recovery-free.
            backend.flush()
        })
        .await
        .map_err(|e| StoreError::Custom(format!("shutdown join: {e}")))?
    }

    /// Add a block in a single transaction.
    /// This will store -> BlockHeader, BlockBody, BlockTransactions, BlockNumber.
    pub async fn add_block(&self, block: Block) -> Result<(), StoreError> {
        self.add_blocks(vec![block]).await
    }

    /// Add a batch of blocks in a single transaction.
    /// This will store -> BlockHeader, BlockBody, BlockTransactions, BlockNumber.
    pub async fn add_blocks(&self, blocks: Vec<Block>) -> Result<(), StoreError> {
        let db = self.backend.clone();
        tokio::task::spawn_blocking(move || {
            let mut tx = db.begin_write()?;

            for block in blocks {
                write_block_data(
                    tx.as_mut(),
                    block.header.number,
                    block.hash(),
                    &block.header,
                    &block.body,
                )?;
            }

            tx.commit()
        })
        .await
        .map_err(|e| StoreError::Custom(format!("Task panicked: {}", e)))?
    }

    /// Add block header
    pub async fn add_block_header(
        &self,
        block_hash: BlockHash,
        block_header: BlockHeader,
    ) -> Result<(), StoreError> {
        let hash_key = block_hash.encode_to_vec();
        let header_value = BlockHeaderRLP::from(block_header).into_vec();
        self.write_async(HEADERS, hash_key, header_value).await
    }

    /// Add a batch of block headers
    pub async fn add_block_headers(
        &self,
        block_headers: Vec<BlockHeader>,
    ) -> Result<(), StoreError> {
        let mut txn = self.backend.begin_write()?;

        for header in block_headers {
            let block_hash = header.hash();
            let block_number = header.number;
            let hash_key = block_hash.encode_to_vec();
            let header_value = BlockHeaderRLP::from(header).into_vec();

            txn.put(HEADERS, &hash_key, &header_value)?;

            let number_key = block_number.to_le_bytes().to_vec();
            txn.put(BLOCK_NUMBERS, &hash_key, &number_key)?;
        }
        txn.commit()?;
        Ok(())
    }

    /// Obtain canonical block header
    pub fn get_block_header(
        &self,
        block_number: BlockNumber,
    ) -> Result<Option<BlockHeader>, StoreError> {
        let latest = self.latest_block_header.get();
        if block_number == latest.number {
            return Ok(Some((*latest).clone()));
        }
        // Resolve the canonical hash, then read through the buffer-aware by-hash
        // path so a canonical-but-still-buffered block is visible (mirrors
        // `get_block_body`). `load_block_header` is disk-only and would return
        // `None` for a block whose header has not been flushed yet.
        let Some(block_hash) = self.get_canonical_block_hash_sync(block_number)? else {
            return Ok(None);
        };
        self.get_block_header_by_hash(block_hash)
    }

    /// Add block body
    pub async fn add_block_body(
        &self,
        block_hash: BlockHash,
        block_body: BlockBody,
    ) -> Result<(), StoreError> {
        let hash_key = block_hash.encode_to_vec();
        let body_value = BlockBodyRLP::from(block_body).into_vec();
        self.write_async(BODIES, hash_key, body_value).await
    }

    /// Backfill historical chain data for a contiguous range of already-canonical
    /// blocks (bodies, receipts, and optionally the transaction-lookup index),
    /// then lower `earliest_block_number` to `new_earliest` — all in one write
    /// transaction.
    ///
    /// This is the storage half of `--history.chain` backfill. It does NOT touch
    /// state, the trie layer cache, or fork choice: the headers and canonical
    /// mapping already exist from snap-sync header backfill, so only the
    /// body / receipt / tx-index tables are filled. Writing the data and advancing
    /// the frontier atomically preserves the invariant that every block in
    /// `[earliest_block_number, head]` has a stored body — a crash mid-batch
    /// leaves both the data and the frontier unchanged, so resume never skips a
    /// hole.
    ///
    /// `new_earliest` must be the lowest block number covered by `blocks` (the
    /// bottom of the contiguous filled range); the caller fills downward without
    /// gaps.
    pub async fn add_backfilled_blocks(
        &self,
        blocks: Vec<BackfilledBlock>,
        new_earliest: BlockNumber,
    ) -> Result<(), StoreError> {
        let backend = self.backend.clone();
        tokio::task::spawn_blocking(move || {
            let mut txn = backend.begin_write()?;
            for block in &blocks {
                let hash = block.header.hash();
                let hash_key = hash.encode_to_vec();
                txn.put(
                    BODIES,
                    &hash_key,
                    &BlockBodyRLP::from(block.body.clone()).into_vec(),
                )?;
                for (index, receipt) in block.receipts.iter().enumerate() {
                    txn.put(
                        RECEIPTS_V2,
                        &receipt_key(&hash, index as u64),
                        &receipt.encode_storage(),
                    )?;
                }
                if block.index_transactions {
                    for (index, transaction) in block.body.transactions.iter().enumerate() {
                        txn.merge(
                            TRANSACTION_LOCATIONS,
                            transaction.hash(&NativeCrypto).as_bytes(),
                            &encode_tx_location_operand(block.header.number, hash, index as u64),
                        )?;
                    }
                }
            }
            txn.put(
                CHAIN_DATA,
                &chain_data_key(ChainDataIndex::EarliestBlockNumber),
                &new_earliest.to_le_bytes(),
            )?;
            txn.commit()
        })
        .await
        .map_err(|e| StoreError::Custom(format!("Backfill write task panicked: {e}")))?
    }

    /// RocksDB observability snapshot (per-column-family sizes/keys/files, block
    /// cache, compaction). `None` for the in-memory backend. Read by the metrics
    /// collector; cheap (RocksDB property reads).
    pub fn rocksdb_stats(&self) -> Option<crate::api::RocksDbStats> {
        self.backend.db_stats()
    }

    /// Obtain canonical block body
    pub async fn get_block_body(
        &self,
        block_number: BlockNumber,
    ) -> Result<Option<BlockBody>, StoreError> {
        let Some(block_hash) = self.get_canonical_block_hash_sync(block_number)? else {
            return Ok(None);
        };

        self.get_block_body_by_hash(block_hash).await
    }

    /// Remove canonical block
    pub async fn remove_block(&self, block_number: BlockNumber) -> Result<(), StoreError> {
        let Some(hash) = self.get_canonical_block_hash_sync(block_number)? else {
            return Ok(());
        };

        let backend = self.backend.clone();
        tokio::task::spawn_blocking(move || {
            let hash_key = hash.encode_to_vec();

            let mut txn = backend.begin_write()?;
            txn.delete(
                CANONICAL_BLOCK_HASHES,
                block_number.to_le_bytes().as_slice(),
            )?;
            txn.delete(BODIES, &hash_key)?;
            txn.delete(HEADERS, &hash_key)?;
            txn.delete(BLOCK_NUMBERS, &hash_key)?;
            txn.commit()
        })
        .await
        .map_err(|e| StoreError::Custom(format!("Task panicked: {}", e)))?
    }

    /// Obtain canonical block bodies in from..=to
    pub async fn get_block_bodies(
        &self,
        from: BlockNumber,
        to: BlockNumber,
    ) -> Result<Vec<Option<BlockBody>>, StoreError> {
        // TODO: Implement read bulk
        let buffer = self.buffer()?;
        let backend = self.backend.clone();
        tokio::task::spawn_blocking(move || {
            let numbers: Vec<BlockNumber> = (from..=to).collect();
            let mut block_bodies = Vec::new();

            let txn = backend.begin_read()?;
            for number in numbers {
                let Some(hash) = txn
                    .get(CANONICAL_BLOCK_HASHES, number.to_le_bytes().as_slice())?
                    .map(|bytes| H256::decode(bytes.as_slice()))
                    .transpose()?
                else {
                    block_bodies.push(None);
                    continue;
                };
                // Consult the in-memory buffer first so a not-yet-flushed body
                // is not reported as missing (mirrors get_block_bodies_by_hash).
                if let Some(body) = buffer.get_body(&hash) {
                    block_bodies.push(Some(body));
                    continue;
                }
                let hash_key = hash.encode_to_vec();
                let block_body_opt = txn
                    .get(BODIES, &hash_key)?
                    .map(|bytes| BlockBodyRLP::from_bytes(bytes).to())
                    .transpose()
                    .map_err(StoreError::from)?;

                block_bodies.push(block_body_opt);
            }

            Ok(block_bodies)
        })
        .await
        .map_err(|e| StoreError::Custom(format!("Task panicked: {}", e)))?
    }

    /// Obtain block bodies from a list of hashes
    pub async fn get_block_bodies_by_hash(
        &self,
        hashes: Vec<BlockHash>,
    ) -> Result<Vec<BlockBody>, StoreError> {
        let buffer = self.buffer()?;
        let backend = self.backend.clone();
        // TODO: Implement read bulk
        tokio::task::spawn_blocking(move || {
            let txn = backend.begin_read()?;
            let mut block_bodies = Vec::new();
            for hash in hashes {
                // Consult the in-memory buffer first, like the single-hash reader,
                // so a not-yet-flushed body is not reported as missing.
                if let Some(body) = buffer.get_body(&hash) {
                    block_bodies.push(body);
                    continue;
                }
                let hash_key = hash.encode_to_vec();

                let Some(block_body) = txn
                    .get(BODIES, &hash_key)?
                    .map(|bytes| BlockBodyRLP::from_bytes(bytes).to())
                    .transpose()
                    .map_err(StoreError::from)?
                else {
                    return Err(StoreError::Custom(format!(
                        "Block body not found for hash: {hash}"
                    )));
                };
                block_bodies.push(block_body);
            }
            Ok(block_bodies)
        })
        .await
        .map_err(|e| StoreError::Custom(format!("Task panicked: {}", e)))?
    }

    /// Obtain any block body using the hash
    pub async fn get_block_body_by_hash(
        &self,
        block_hash: BlockHash,
    ) -> Result<Option<BlockBody>, StoreError> {
        if let Some(b) = self.buffer()?.get_body(&block_hash) {
            return Ok(Some(b));
        }
        self.read_async(BODIES, block_hash.encode_to_vec())
            .await?
            .map(|bytes| BlockBodyRLP::from_bytes(bytes).to())
            .transpose()
            .map_err(StoreError::from)
    }

    pub fn get_block_header_by_hash(
        &self,
        block_hash: BlockHash,
    ) -> Result<Option<BlockHeader>, StoreError> {
        let latest = self.latest_block_header.get();
        if block_hash == latest.hash() {
            return Ok(Some((*latest).clone()));
        }
        if let Some(h) = self.buffer()?.get_header(&block_hash) {
            return Ok(Some(h));
        }
        self.load_block_header_by_hash(block_hash)
    }

    pub fn add_pending_block(&self, block: Block) -> Result<(), StoreError> {
        let block_hash = block.hash();
        let block_value = BlockRLP::from(block).into_vec();
        self.write(PENDING_BLOCKS, block_hash.as_bytes().to_vec(), block_value)
    }

    pub async fn get_pending_block(
        &self,
        block_hash: BlockHash,
    ) -> Result<Option<Block>, StoreError> {
        self.read_async(PENDING_BLOCKS, block_hash.as_bytes().to_vec())
            .await?
            .map(|bytes| BlockRLP::from_bytes(bytes).to())
            .transpose()
            .map_err(StoreError::from)
    }

    /// Add block number for a given hash
    pub async fn add_block_number(
        &self,
        block_hash: BlockHash,
        block_number: BlockNumber,
    ) -> Result<(), StoreError> {
        let number_value = block_number.to_le_bytes().to_vec();
        self.write_async(BLOCK_NUMBERS, block_hash.encode_to_vec(), number_value)
            .await
    }

    /// Obtain block number for a given hash
    pub async fn get_block_number(
        &self,
        block_hash: BlockHash,
    ) -> Result<Option<BlockNumber>, StoreError> {
        if let Some(n) = self.buffer()?.get_number(&block_hash) {
            return Ok(Some(n));
        }
        self.read_async(BLOCK_NUMBERS, block_hash.encode_to_vec())
            .await?
            .map(decode_block_number)
            .transpose()
    }

    /// Store transaction location (block number and index of the transaction within the block)
    pub async fn add_transaction_location(
        &self,
        transaction_hash: H256,
        block_number: BlockNumber,
        block_hash: BlockHash,
        index: Index,
    ) -> Result<(), StoreError> {
        self.add_transaction_locations(vec![(transaction_hash, block_number, block_hash, index)])
            .await
    }

    /// Store transaction locations in batch (one db transaction for all)
    pub async fn add_transaction_locations(
        &self,
        locations: Vec<(H256, BlockNumber, BlockHash, Index)>,
    ) -> Result<(), StoreError> {
        let db = self.backend.clone();
        tokio::task::spawn_blocking(move || {
            let mut tx = db.begin_write()?;
            for (tx_hash, block_number, block_hash, index) in locations {
                tx.merge(
                    TRANSACTION_LOCATIONS,
                    tx_hash.as_bytes(),
                    &encode_tx_location_operand(block_number, block_hash, index),
                )?;
            }
            tx.commit()
        })
        .await
        .map_err(|e| StoreError::Custom(format!("Task panicked: {}", e)))?
    }

    /// Obtain transaction location (block hash and index)
    pub async fn get_transaction_location(
        &self,
        transaction_hash: H256,
    ) -> Result<Option<(BlockNumber, BlockHash, Index)>, StoreError> {
        let buffered = self.buffer()?.get_tx_locations(&transaction_hash);
        let db = self.backend.clone();
        tokio::task::spawn_blocking(move || {
            let tx = db.begin_read()?;
            let mut locations = buffered;
            if let Some(bytes) = tx.get(TRANSACTION_LOCATIONS, transaction_hash.as_bytes())? {
                locations.extend(<Vec<(BlockNumber, BlockHash, Index)>>::decode(&bytes)?);
            }
            for (block_number, block_hash, index) in locations {
                let canonical_hash = tx
                    .get(
                        CANONICAL_BLOCK_HASHES,
                        block_number.to_le_bytes().as_slice(),
                    )?
                    .map(|bytes| H256::decode(bytes.as_slice()))
                    .transpose()?;
                if canonical_hash == Some(block_hash) {
                    return Ok(Some((block_number, block_hash, index)));
                }
            }
            Ok(None)
        })
        .await
        .map_err(|e| StoreError::Custom(format!("Task panicked: {}", e)))?
    }

    /// Add receipt
    pub async fn add_receipt(
        &self,
        block_hash: BlockHash,
        index: Index,
        receipt: Receipt,
    ) -> Result<(), StoreError> {
        let key = receipt_key(&block_hash, index);
        // Storage codec (NOT wire/consensus): preserves frame-receipt
        // `succeeded` + aggregated logs; identical to encode_to_vec for
        // non-frame receipts.
        let value = receipt.encode_storage();
        self.write_async(RECEIPTS_V2, key, value).await
    }

    /// Add receipts
    pub async fn add_receipts(
        &self,
        block_hash: BlockHash,
        receipts: Vec<Receipt>,
    ) -> Result<(), StoreError> {
        let batch_items: Vec<_> = receipts
            .into_iter()
            .enumerate()
            .map(|(index, receipt)| {
                let key = receipt_key(&block_hash, index as u64);
                let value = receipt.encode_storage();
                (key, value)
            })
            .collect();
        self.write_batch_async(RECEIPTS_V2, batch_items).await
    }

    /// Obtain receipt for a canonical block represented by the block number.
    pub async fn get_receipt(
        &self,
        block_number: BlockNumber,
        index: Index,
    ) -> Result<Option<Receipt>, StoreError> {
        // FIXME (#4353)
        let Some(block_hash) = self.get_canonical_block_hash(block_number).await? else {
            return Ok(None);
        };
        self.get_receipt_by_block_hash(block_hash, index).await
    }

    /// Obtain receipt by block hash and index
    async fn get_receipt_by_block_hash(
        &self,
        block_hash: BlockHash,
        index: Index,
    ) -> Result<Option<Receipt>, StoreError> {
        if let Some(r) = self.buffer()?.get_receipt(&block_hash, index) {
            return Ok(Some(r));
        }
        let key = receipt_key(&block_hash, index);
        self.read_async(RECEIPTS_V2, key)
            .await?
            .map(|bytes| Receipt::decode_storage(bytes.as_slice()))
            .transpose()
            .map_err(StoreError::from)
    }

    /// Capacity of the bytecode cache, in [`Code::size`] bytes, which is how much
    /// bytecode this node is configured to keep resident and so bounds what a block warm
    /// reads speculatively.
    ///
    /// A poisoned cache reports zero, which stops warming rather than warming into a
    /// cache no reader can reach.
    pub fn code_cache_budget_bytes(&self) -> u64 {
        self.account_code_cache
            .lock()
            .map(|cache| cache.max_size)
            .unwrap_or(0)
    }

    /// Get account code by its hash.
    ///
    /// Checks the in-memory block-data buffer first, then the LRU cache
    /// (`account_code_cache`), and finally the database.  Code that has been
    /// inserted via `engine_newPayload` but not yet flushed to disk is therefore
    /// visible to callers without an explicit flush.
    pub fn get_account_code(&self, code_hash: H256) -> Result<Option<Code>, StoreError> {
        if let Some(code) = self.buffer()?.get_code(&code_hash) {
            return Ok(Some(code));
        }
        // check cache first
        if let Some(code) = self
            .account_code_cache
            .lock()
            .map_err(|_| StoreError::LockError)?
            .get(&code_hash)?
        {
            return Ok(Some(code));
        }

        let Some(bytes) = self
            .backend
            .begin_read()?
            .get(ACCOUNT_CODES, code_hash.as_bytes())?
        else {
            return Ok(None);
        };
        let (bytecode_slice, jumpdests) = decode_bytes(&bytes)?;
        let (jumpdests, was_legacy) = decode_jumpdests(bytecode_slice, jumpdests)?;
        let code = Code::from_parts_unchecked(code_hash, bytecode_slice, jumpdests);

        // Self-heal a legacy (pre-bitmap) row so the jumpdest recompute above is
        // paid at most once. Fire-and-forget, and only when a Tokio runtime is
        // reachable: this read runs on rayon execution-path workers, where
        // `spawn` would panic. Skipping just leaves the next read to recompute
        // again; there is no correctness dependency (matches the code-metadata
        // backfill in `get_code_metadata`).
        if was_legacy && let Ok(handle) = tokio::runtime::Handle::try_current() {
            let buf = encode_code(&code);
            let hash_key = code_hash.0.to_vec();
            let backend = self.backend.clone();
            handle.spawn(async move {
                if let Err(e) = async {
                    let mut tx = backend.begin_write()?;
                    tx.put(ACCOUNT_CODES, &hash_key, &buf)?;
                    tx.commit()
                }
                .await
                {
                    tracing::warn!("Failed to rewrite legacy jumpdests during backfill: {e}");
                }
            });
        }

        // insert into cache and evict if needed
        self.account_code_cache
            .lock()
            .map_err(|_| StoreError::LockError)?
            .insert(&code)?;

        Ok(Some(code))
    }

    /// Batched [`Self::get_account_code`].
    ///
    /// Resolves the buffer and the LRU first, then reads whatever is left by whichever
    /// of two strategies gets more of those reads in flight for a batch this size: a
    /// parallel fan-out of point gets, or sorted keys split into contiguous shards read
    /// concurrently. See the comment on the read below for how the choice is made. The
    /// LRU is locked once for the whole batch rather than twice per code.
    ///
    /// Results are returned in the order of `code_hashes`. Duplicate hashes are read
    /// once. `None` means the hash is absent from the database.
    pub fn get_account_codes_batch(
        &self,
        code_hashes: &[H256],
    ) -> Result<Vec<Option<Code>>, StoreError> {
        let mut out: Vec<Option<Code>> = vec![None; code_hashes.len()];
        // Positions to fill per distinct hash, so a repeated hash costs one read.
        let mut pending: HashMap<H256, Vec<usize>> = HashMap::new();

        {
            let buffer = self.buffer()?;
            let mut cache = self
                .account_code_cache
                .lock()
                .map_err(|_| StoreError::LockError)?;
            for (i, hash) in code_hashes.iter().enumerate() {
                if let Some(code) = buffer.get_code(hash) {
                    out[i] = Some(code);
                } else if let Some(code) = cache.get(hash)? {
                    out[i] = Some(code);
                } else {
                    pending.entry(*hash).or_default().push(i);
                }
            }
        }

        if pending.is_empty() {
            return Ok(out);
        }

        let mut missing: Vec<H256> = pending.keys().copied().collect();
        missing.sort_unstable();

        // Cold blob reads here are latency-bound, so what matters is how many are in
        // flight. Two ways to get there, and which one wins depends on the batch size:
        //
        // * A parallel fan-out of point gets reaches queue depth ~= core count. Cheap
        //   for any batch: rayon's pool is already warm.
        // * Contiguous shards of the SORTED keys, one blocking thread each, reach queue
        //   depth ~= shard count and share RocksDB blocks within a shard. This is the
        //   only way past core count (async_io is OFF in this build, so a single
        //   `multi_get` runs the whole batch at queue depth 1), but it pays a thread
        //   spawn per shard.
        //
        // So shard only once sharding can actually beat the fan-out, i.e. once the batch
        // is wide enough to form more shards than there are cores. Below that the
        // fan-out is both deeper and cheaper, and a single serial `multi_get` would be
        // far worse than either.
        //
        // The shard cap has to sit above the core count, or on a host with at least that
        // many cores the fan-out would always win and this path would be unreachable.
        // Twice the cores guarantees that, with a floor so a small host keeps the depth
        // it can already reach: these threads block on I/O rather than compute, so more
        // of them than cores is the point.
        const KEYS_PER_SHARD: usize = 256;
        const MIN_SHARD_CAP: usize = 64;
        let parallelism = std::thread::available_parallelism().map_or(8, |p| p.get());
        let max_shards = parallelism.saturating_mul(2).max(MIN_SHARD_CAP);
        let shards = missing.len().div_ceil(KEYS_PER_SHARD).min(max_shards);
        let read_view = self.backend.begin_read()?;
        // Both paths decode in whatever thread did the read, so rebuilding the jumpdest
        // bitmap for a legacy entry stays off the caller's thread and stays parallel.
        let decode = |hash: &H256, value: Option<Vec<u8>>| -> Result<Option<Code>, StoreError> {
            let Some(bytes) = value else { return Ok(None) };
            let (bytecode_slice, jumpdests) = decode_bytes(&bytes)?;
            // This is the bulk prefetch warm path; a legacy row is recomputed here
            // and self-healed later when `get_account_code` reads it on the hot path.
            let (jumpdests, _was_legacy) = decode_jumpdests(bytecode_slice, jumpdests)?;
            Ok(Some(Code::from_parts_unchecked(
                *hash,
                bytecode_slice,
                jumpdests,
            )))
        };
        let decoded: Vec<Result<Option<Code>, StoreError>> = if shards > parallelism {
            let chunk = missing.len().div_ceil(shards);
            let rv = read_view.as_ref();
            let read_shard = |hashes: &[H256]| -> Vec<Result<Option<Code>, StoreError>> {
                let keys: Vec<&[u8]> = hashes.iter().map(|h| h.as_bytes()).collect();
                rv.multi_get(ACCOUNT_CODES, &keys)
                    .into_iter()
                    .zip(hashes.iter())
                    .map(|(value, hash)| decode(hash, value?))
                    .collect()
            };
            std::thread::scope(|scope| {
                let handles: Vec<_> = missing
                    .chunks(chunk)
                    .map(|ck| (ck, scope.spawn(move || read_shard(ck))))
                    .collect();
                handles
                    .into_iter()
                    .flat_map(|(shard, handle)| {
                        // A panicked shard becomes an `Err` per key it covered, keeping
                        // the results aligned with `missing` and leaving the caller's
                        // best-effort handling to decide. Re-panicking here would
                        // escalate a warm into taking down whatever runs it.
                        handle.join().unwrap_or_else(|_| {
                            shard
                                .iter()
                                .map(|_| {
                                    Err(StoreError::Custom(
                                        "account code shard panicked".to_string(),
                                    ))
                                })
                                .collect()
                        })
                    })
                    .collect()
            })
        } else {
            // One key per read, so `get` rather than a single-key `multi_get`: the
            // batched call sets up its own result buffers, which is pure overhead here.
            let rv = read_view.as_ref();
            missing
                .par_iter()
                .map(|hash| decode(hash, rv.get(ACCOUNT_CODES, hash.as_bytes())?))
                .collect()
        };

        let mut fetched: Vec<Code> = Vec::new();
        for (hash, code) in missing.iter().zip(decoded.into_iter()) {
            let Some(code) = code? else { continue };
            for &i in pending.get(hash).into_iter().flatten() {
                out[i] = Some(code.clone());
            }
            fetched.push(code);
        }

        if !fetched.is_empty() {
            let mut cache = self
                .account_code_cache
                .lock()
                .map_err(|_| StoreError::LockError)?;
            for code in &fetched {
                cache.insert(code)?;
            }
        }

        Ok(out)
    }

    /// Check if account code exists by its hash, without constructing the full `Code` struct.
    /// More efficient than `get_account_code` for existence checks since it skips
    /// RLP decoding and `Code` struct construction (no jumpdest-bitmap decoding).
    /// Note: The underlying `get()` still reads the value from RocksDB (including blob files).
    pub fn code_exists(&self, code_hash: H256) -> Result<bool, StoreError> {
        // Code introduced by a not-yet-flushed block lives only in the buffer; check
        // it first so a contract created in the current block is visible (matches
        // get_account_code / get_code_metadata).
        if self.buffer()?.get_code(&code_hash).is_some() {
            return Ok(true);
        }
        // Check cache first
        if self
            .account_code_cache
            .lock()
            .map_err(|_| StoreError::LockError)?
            .get(&code_hash)?
            .is_some()
        {
            return Ok(true);
        }
        // Check DB without reading the full value
        Ok(self
            .backend
            .begin_read()?
            .get(ACCOUNT_CODES, code_hash.as_bytes())?
            .is_some())
    }

    /// Get code metadata (length) by its hash.
    ///
    /// Checks cache first, falls back to database. If metadata is missing,
    /// falls back to loading full code and extracts length (auto-migration).
    pub fn get_code_metadata(&self, code_hash: H256) -> Result<Option<CodeMetadata>, StoreError> {
        use ethrex_common::constants::EMPTY_KECCAK_HASH;

        // Empty code special case
        if code_hash == *EMPTY_KECCAK_HASH {
            return Ok(Some(CodeMetadata { length: 0 }));
        }

        // Check cache first
        if let Some(metadata) = self
            .code_metadata_cache
            .lock()
            .map_err(|_| StoreError::LockError)?
            .get(&code_hash)
            .copied()
        {
            return Ok(Some(metadata));
        }

        // Try reading from metadata table
        let metadata = if let Some(bytes) = self
            .backend
            .begin_read()?
            .get(ACCOUNT_CODE_METADATA, code_hash.as_bytes())?
        {
            let length =
                u64::from_be_bytes(bytes.try_into().map_err(|_| {
                    StoreError::Custom("Invalid metadata length encoding".to_string())
                })?);
            CodeMetadata { length }
        } else {
            // Fallback: load full code and extract length (auto-migration)
            let Some(code) = self.get_account_code(code_hash)? else {
                return Ok(None);
            };
            let metadata = CodeMetadata {
                length: code.len() as u64,
            };

            // Backfill the row for future reads, fire and forget.
            //
            // Only when a Tokio runtime is reachable: this read is on the execution
            // path (`EXTCODESIZE`), which runs on rayon workers, and
            // `tokio::task::spawn` panics outside a runtime. Skipping the backfill
            // costs one code read the next time that hash is asked for; panicking
            // would take the node down.
            if let Ok(handle) = tokio::runtime::Handle::try_current() {
                let metadata_buf = metadata.length.to_be_bytes().to_vec();
                let hash_key = code_hash.0.to_vec();
                let backend = self.backend.clone();
                handle.spawn(async move {
                    if let Err(e) = async {
                        let mut tx = backend.begin_write()?;
                        tx.put(ACCOUNT_CODE_METADATA, &hash_key, &metadata_buf)?;
                        tx.commit()
                    }
                    .await
                    {
                        tracing::warn!("Failed to write code metadata during backfill: {}", e);
                    }
                });
            }

            metadata
        };

        // Update cache
        self.code_metadata_cache
            .lock()
            .map_err(|_| StoreError::LockError)?
            .put(code_hash, metadata);

        Ok(Some(metadata))
    }

    /// Add account code
    pub async fn add_account_code(&self, code: Code) -> Result<(), StoreError> {
        let hash_key = code.hash.0.to_vec();
        let buf = encode_code(&code);
        let metadata_buf = (code.len() as u64).to_be_bytes();

        // Write both code and metadata atomically
        let backend = self.backend.clone();
        tokio::task::spawn_blocking(move || {
            let mut tx = backend.begin_write()?;
            tx.put(ACCOUNT_CODES, &hash_key, &buf)?;
            tx.put(ACCOUNT_CODE_METADATA, &hash_key, &metadata_buf)?;
            tx.commit()
        })
        .await
        .map_err(|e| StoreError::Custom(format!("Task panicked: {}", e)))?
    }

    /// Clears all checkpoint data created during the last snap sync
    pub async fn clear_snap_state(&self) -> Result<(), StoreError> {
        let db = self.backend.clone();
        tokio::task::spawn_blocking(move || db.clear_table(SNAP_STATE))
            .await
            .map_err(|e| StoreError::Custom(format!("Task panicked: {}", e)))?
    }

    pub async fn get_transaction_by_hash(
        &self,
        transaction_hash: H256,
    ) -> Result<Option<Transaction>, StoreError> {
        let (_block_number, block_hash, index) =
            match self.get_transaction_location(transaction_hash).await? {
                Some(location) => location,
                None => return Ok(None),
            };
        self.get_transaction_by_location(block_hash, index).await
    }

    pub async fn get_transaction_by_location(
        &self,
        block_hash: H256,
        index: u64,
    ) -> Result<Option<Transaction>, StoreError> {
        let block_body = match self.get_block_body_by_hash(block_hash).await? {
            Some(body) => body,
            None => return Ok(None),
        };
        let index: usize = index.try_into()?;
        Ok(block_body.transactions.get(index).cloned())
    }

    pub async fn get_block_by_hash(
        &self,
        block_hash: BlockHash,
    ) -> Result<Option<Block>, StoreError> {
        let header = match self.get_block_header_by_hash(block_hash)? {
            Some(header) => header,
            None => return Ok(None),
        };
        let body = match self.get_block_body_by_hash(block_hash).await? {
            Some(body) => body,
            None => return Ok(None),
        };
        Ok(Some(Block::new(header, body)))
    }

    pub async fn get_block_by_number(
        &self,
        block_number: BlockNumber,
    ) -> Result<Option<Block>, StoreError> {
        let Some(block_hash) = self.get_canonical_block_hash(block_number).await? else {
            return Ok(None);
        };
        self.get_block_by_hash(block_hash).await
    }

    // Get the canonical block hash for a given block number.
    pub async fn get_canonical_block_hash(
        &self,
        block_number: BlockNumber,
    ) -> Result<Option<BlockHash>, StoreError> {
        let last = self.latest_block_header.get();
        if last.number == block_number {
            return Ok(Some(last.hash()));
        }
        let backend = self.backend.clone();
        tokio::task::spawn_blocking(move || {
            backend
                .begin_read()?
                .get(
                    CANONICAL_BLOCK_HASHES,
                    block_number.to_le_bytes().as_slice(),
                )?
                .map(|bytes| H256::decode(bytes.as_slice()))
                .transpose()
                .map_err(StoreError::from)
        })
        .await
        .map_err(|e| StoreError::Custom(format!("Task panicked: {}", e)))?
    }

    /// Stores the chain configuration values, should only be called once after reading the genesis file
    /// Ignores previously stored values if present
    pub async fn set_chain_config(&mut self, chain_config: &ChainConfig) -> Result<(), StoreError> {
        self.chain_config = *chain_config;
        let key = chain_data_key(ChainDataIndex::ChainConfig);
        let value = serde_json::to_string(chain_config)
            .map_err(|_| StoreError::Custom("Failed to serialize chain config".to_string()))?
            .into_bytes();
        self.write_async(CHAIN_DATA, key, value).await
    }

    /// Update earliest block number
    pub async fn update_earliest_block_number(
        &self,
        block_number: BlockNumber,
    ) -> Result<(), StoreError> {
        let key = chain_data_key(ChainDataIndex::EarliestBlockNumber);
        let value = block_number.to_le_bytes().to_vec();
        self.write_async(CHAIN_DATA, key, value).await
    }

    /// Reads a block number stored under the given chain-data index.
    async fn read_chain_data_block_number(
        &self,
        index: ChainDataIndex,
    ) -> Result<Option<BlockNumber>, StoreError> {
        self.read_async(CHAIN_DATA, chain_data_key(index))
            .await?
            .map(decode_block_number)
            .transpose()
    }

    /// Obtain earliest block number
    pub async fn get_earliest_block_number(&self) -> Result<BlockNumber, StoreError> {
        self.read_chain_data_block_number(ChainDataIndex::EarliestBlockNumber)
            .await?
            .ok_or(StoreError::MissingEarliestBlockNumber)
    }

    /// Obtain finalized block number
    pub async fn get_finalized_block_number(&self) -> Result<Option<BlockNumber>, StoreError> {
        self.read_chain_data_block_number(ChainDataIndex::FinalizedBlockNumber)
            .await
    }

    /// Obtain safe block number
    pub async fn get_safe_block_number(&self) -> Result<Option<BlockNumber>, StoreError> {
        self.read_chain_data_block_number(ChainDataIndex::SafeBlockNumber)
            .await
    }

    /// Obtain latest block number
    pub fn get_latest_block_number(&self) -> Result<BlockNumber, StoreError> {
        Ok(self.latest_block_header.get().number)
    }

    /// Update pending block number
    pub async fn update_pending_block_number(
        &self,
        block_number: BlockNumber,
    ) -> Result<(), StoreError> {
        let key = chain_data_key(ChainDataIndex::PendingBlockNumber);
        let value = block_number.to_le_bytes().to_vec();
        self.write_async(CHAIN_DATA, key, value).await
    }

    /// Obtain pending block number
    pub async fn get_pending_block_number(&self) -> Result<Option<BlockNumber>, StoreError> {
        self.read_chain_data_block_number(ChainDataIndex::PendingBlockNumber)
            .await
    }

    /// DB mutation step of `forkchoice_update`.
    ///
    /// Callers MUST hold `fcu_lock` (only `forkchoice_update` should invoke this).
    /// The read of `LatestBlockNumber` below happens outside the write
    /// transaction and would be a TOCTOU window without that serialization.
    /// Applies a fork-choice update atomically: writes canonical hashes, latest/safe/
    /// finalized numbers, and prunes the state-history journal up to the new finalized.
    ///
    /// **Concurrency:** This function must be called with exclusive access. In
    /// production it is gated by `fcu_lock` from [`forkchoice_update`]. The
    /// pre-finalized read used to decide journal pruning happens outside the write
    /// transaction; without serialization, two concurrent callers could race on
    /// `prev_finalized` and double-prune. Tests that call this directly must not
    /// run concurrently with another FCU.
    async fn forkchoice_update_inner(
        &self,
        new_canonical_blocks: Vec<(BlockNumber, BlockHash)>,
        head_number: BlockNumber,
        head_hash: BlockHash,
        safe: Option<BlockNumber>,
        finalized: Option<BlockNumber>,
    ) -> Result<(), StoreError> {
        let latest = self.load_latest_block_number().await?.unwrap_or(0);
        let db = self.backend.clone();
        let journal_pruning_paused = self.journal_pruning_paused.clone();
        tokio::task::spawn_blocking(move || {
            let mut txn = db.begin_write()?;

            for (block_number, block_hash) in new_canonical_blocks {
                let head_key = block_number.to_le_bytes();
                let head_value = block_hash.encode_to_vec();
                txn.put(CANONICAL_BLOCK_HASHES, &head_key, &head_value)?;
            }

            // Delete canonical entries above the new head by enumerating each key.
            // `delete_range` is not safe here: keys are `u64::to_le_bytes()`, and
            // RocksDB's lexicographic comparator does not match LE numeric order
            // (e.g. block 256 = [0x00, 0x01, ..] sorts before block 11 = [0x0B, ..]),
            // so a range-delete would silently miss blocks whose LE first byte is
            // smaller than `head+1`'s first byte.
            for number in (head_number + 1)..=(latest) {
                txn.delete(CANONICAL_BLOCK_HASHES, number.to_le_bytes().as_slice())?;
            }

            // Make head canonical
            let head_key = head_number.to_le_bytes();
            let head_value = head_hash.encode_to_vec();
            txn.put(CANONICAL_BLOCK_HASHES, &head_key, &head_value)?;

            // Update chain data
            let latest_key = chain_data_key(ChainDataIndex::LatestBlockNumber);
            txn.put(CHAIN_DATA, &latest_key, &head_number.to_le_bytes())?;

            if let Some(safe) = safe {
                let safe_key = chain_data_key(ChainDataIndex::SafeBlockNumber);
                txn.put(CHAIN_DATA, &safe_key, &safe.to_le_bytes())?;
            }

            if let Some(finalized) = finalized {
                let finalized_key = chain_data_key(ChainDataIndex::FinalizedBlockNumber);

                // Read the previous finalized number from the same backend before we
                // overwrite it. The journal can only be pruned when finality actually
                // advances; a no-op or backwards FCU must not touch STATE_HISTORY.
                // Pre-merge or fresh chains have no entry; treat as 0.
                //
                // A length mismatch is treated as a hard error rather than a silent
                // fallback to 0: if a future schema change stores this field with a
                // different width, the silent default would make `finalized > 0` true
                // for every FCU and prune the entire journal. Bail out instead.
                let prev_finalized = match db.begin_read()?.get(CHAIN_DATA, &finalized_key)? {
                    Some(bytes) => {
                        let arr: [u8; 8] = bytes.as_slice().try_into().map_err(|_| {
                            StoreError::Custom(format!(
                                "FinalizedBlockNumber has unexpected length {} (want 8)",
                                bytes.len()
                            ))
                        })?;
                        BlockNumber::from_le_bytes(arr)
                    }
                    None => 0,
                };

                txn.put(CHAIN_DATA, &finalized_key, &finalized.to_le_bytes())?;

                // Prune every STATE_HISTORY entry at or below the new finalized number
                // in the same atomic txn. `delete_range` is half-open `[start, end)`,
                // so `end = finalized + 1`. STATE_HISTORY uses big-endian keys, so
                // lexicographic byte order matches numeric order.
                //
                // Skipped while a deep-reorg apply pass is in flight
                // (`journal_pruning_paused`): `Overlay::from_journal` reads entries
                // with no snapshot isolation, so pruning mid-construction fails it
                // with a spurious `MissingEntry`. The finalized-number update above
                // still lands; pruning catches up on the next advance after the
                // pass ends because `delete_range` is cumulative from zero.
                if finalized > prev_finalized
                    && !journal_pruning_paused.load(std::sync::atomic::Ordering::Acquire)
                {
                    let start = 0u64.to_be_bytes();
                    let end = finalized.saturating_add(1).to_be_bytes();
                    txn.delete_range(STATE_HISTORY, &start, &end)?;
                }
            }

            txn.commit()
        })
        .await
        .map_err(|e| StoreError::Custom(format!("Task panicked: {}", e)))?
    }

    pub async fn get_receipts_for_block(
        &self,
        block_hash: &BlockHash,
    ) -> Result<Vec<Receipt>, StoreError> {
        self.get_receipts_for_block_from_index(block_hash, 0, None)
            .await
    }

    /// Retrieves receipts for a block starting from the given index,
    /// optionally limited to `max_count` receipts.
    ///
    /// Uses cursor-based prefix iteration over the 32-byte block hash prefix
    /// for efficient batch retrieval. Used by:
    /// - eth/70 partial receipt requests (EIP-7975) via p2p
    /// - `eth_getTransactionReceipt` RPC with a count limit to avoid
    ///   fetching the entire block's receipts
    pub async fn get_receipts_for_block_from_index(
        &self,
        block_hash: &BlockHash,
        start_index: u64,
        max_count: Option<usize>,
    ) -> Result<Vec<Receipt>, StoreError> {
        if let Some(all) = self.buffer()?.get_receipts(block_hash) {
            let start = start_index as usize;
            let slice = all.into_iter().skip(start);
            return Ok(match max_count {
                Some(max) => slice.take(max).collect(),
                None => slice.collect(),
            });
        }
        let backend = self.backend.clone();
        let block_hash = *block_hash;

        tokio::task::spawn_blocking(move || {
            let txn = backend.begin_read()?;
            let prefix = block_hash.as_bytes().to_vec();
            // Seek directly to block_hash || start_index to avoid O(start_index) scan.
            // Keys are big-endian u64, so lexicographic order matches numeric order.
            let mut seek_key = prefix.clone();
            seek_key.extend_from_slice(&start_index.to_be_bytes());
            let iter = txn.prefix_iterator(RECEIPTS_V2, &seek_key)?;
            let mut receipts = Vec::new();
            for result in iter {
                let (k, v) = result?;
                if !k.starts_with(&prefix) {
                    break;
                }
                if k.len() != 40 {
                    continue;
                }
                receipts.push(Receipt::decode_storage(v.as_ref())?);
                if let Some(max) = max_count
                    && receipts.len() >= max
                {
                    break;
                }
            }
            Ok(receipts)
        })
        .await
        .map_err(|e| StoreError::Custom(format!("Task panicked: {e}")))?
    }

    // Snap State methods

    /// Sets the hash of the last header downloaded during a snap sync
    pub async fn set_header_download_checkpoint(
        &self,
        block_hash: BlockHash,
    ) -> Result<(), StoreError> {
        let key = snap_state_key(SnapStateIndex::HeaderDownloadCheckpoint);
        let value = block_hash.encode_to_vec();
        self.write_async(SNAP_STATE, key, value).await
    }

    /// Gets the hash of the last header downloaded during a snap sync
    pub async fn get_header_download_checkpoint(&self) -> Result<Option<BlockHash>, StoreError> {
        let key = snap_state_key(SnapStateIndex::HeaderDownloadCheckpoint);
        self.backend
            .begin_read()?
            .get(SNAP_STATE, &key)?
            .map(|bytes| H256::decode(bytes.as_slice()))
            .transpose()
            .map_err(StoreError::from)
    }

    /// The `forkchoice_update` and `new_payload` methods require the `latest_valid_hash`
    /// when processing an invalid payload. To provide this, we must track invalid chains.
    ///
    /// We only store the last known valid head upon encountering a bad block,
    /// rather than tracking every subsequent invalid block.
    pub async fn set_latest_valid_ancestor(
        &self,
        bad_block: BlockHash,
        latest_valid: BlockHash,
    ) -> Result<(), StoreError> {
        let value = latest_valid.encode_to_vec();
        self.write_async(INVALID_CHAINS, bad_block.as_bytes().to_vec(), value)
            .await
    }

    /// Returns the latest valid ancestor hash for a given invalid block hash.
    /// Used to provide `latest_valid_hash` in the Engine API when processing invalid payloads.
    pub async fn get_latest_valid_ancestor(
        &self,
        block: BlockHash,
    ) -> Result<Option<BlockHash>, StoreError> {
        self.read_async(INVALID_CHAINS, block.as_bytes().to_vec())
            .await?
            .map(|bytes| H256::decode(bytes.as_slice()))
            .transpose()
            .map_err(StoreError::from)
    }

    /// Records a block that failed validation so it can be served by
    /// `debug_getBadBlocks`. The list is bounded to [`MAX_BAD_BLOCKS`] entries,
    /// kept sorted by descending block number, with the oldest dropped once the
    /// bound is exceeded. Duplicate `(number, hash)` entries are ignored.
    pub async fn add_bad_block(&self, block: Block) -> Result<(), StoreError> {
        let mut bad_blocks = self.get_bad_blocks().await?;
        let block_number = block.header.number;
        let block_hash = block.hash();
        if bad_blocks
            .iter()
            .any(|b| b.header.number == block_number && b.hash() == block_hash)
        {
            return Ok(());
        }
        bad_blocks.push(block);
        bad_blocks.sort_by(|a, b| b.header.number.cmp(&a.header.number));
        bad_blocks.truncate(MAX_BAD_BLOCKS);
        self.write_async(
            BAD_BLOCKS,
            BAD_BLOCKS_KEY.to_vec(),
            bad_blocks.encode_to_vec(),
        )
        .await
    }

    /// Returns the recent bad blocks seen by the client, sorted by descending
    /// block number. Used by `debug_getBadBlocks`.
    pub async fn get_bad_blocks(&self) -> Result<Vec<Block>, StoreError> {
        match self.read_async(BAD_BLOCKS, BAD_BLOCKS_KEY.to_vec()).await? {
            Some(bytes) => Vec::<Block>::decode(&bytes).map_err(StoreError::from),
            None => Ok(Vec::new()),
        }
    }

    /// Obtain block number for a given hash
    pub fn get_block_number_sync(
        &self,
        block_hash: BlockHash,
    ) -> Result<Option<BlockNumber>, StoreError> {
        if let Some(n) = self.buffer()?.get_number(&block_hash) {
            return Ok(Some(n));
        }
        let txn = self.backend.begin_read()?;
        txn.get(BLOCK_NUMBERS, &block_hash.encode_to_vec())?
            .map(decode_block_number)
            .transpose()
    }

    /// Get the canonical block hash for a given block number.
    pub fn get_canonical_block_hash_sync(
        &self,
        block_number: BlockNumber,
    ) -> Result<Option<BlockHash>, StoreError> {
        let last = self.latest_block_header.get();
        if last.number == block_number {
            return Ok(Some(last.hash()));
        }
        let txn = self.backend.begin_read()?;
        txn.get(
            CANONICAL_BLOCK_HASHES,
            block_number.to_le_bytes().as_slice(),
        )?
        .map(|bytes| H256::decode(bytes.as_slice()))
        .transpose()
        .map_err(StoreError::from)
    }

    /// CAUTION: This method writes directly to the underlying database, bypassing any caching layer.
    /// For updating the state after block execution, use [`Self::store_block_updates`].
    pub async fn write_storage_trie_nodes_batch(
        &self,
        storage_trie_nodes: StorageUpdates,
    ) -> Result<(), StoreError> {
        let mut txn = self.backend.begin_write()?;
        tokio::task::spawn_blocking(move || {
            for (address_hash, nodes) in storage_trie_nodes {
                for (node_path, node_data) in nodes {
                    let key = apply_prefix(Some(address_hash), node_path);
                    if node_data.is_empty() {
                        txn.delete(STORAGE_TRIE_NODES, key.as_ref())?;
                    } else {
                        txn.put(STORAGE_TRIE_NODES, key.as_ref(), &node_data)?;
                    }
                }
            }
            txn.commit()
        })
        .await
        .map_err(|e| StoreError::Custom(format!("Task panicked: {}", e)))?
    }

    /// CAUTION: This method writes directly to the underlying database, bypassing any caching layer.
    /// For updating the state after block execution, use [`Self::store_block_updates`].
    pub async fn write_account_code_batch(
        &self,
        account_codes: Vec<(H256, Code)>,
    ) -> Result<(), StoreError> {
        let mut code_batch_items = Vec::new();
        let mut metadata_batch_items = Vec::new();

        for (code_hash, code) in account_codes {
            let buf = encode_code(&code);
            let metadata_buf = (code.len() as u64).to_be_bytes().to_vec();
            code_batch_items.push((code_hash.as_bytes().to_vec(), buf));
            metadata_batch_items.push((code_hash.as_bytes().to_vec(), metadata_buf));
        }

        // Write both batches
        self.write_batch_async(ACCOUNT_CODES, code_batch_items)
            .await?;
        self.write_batch_async(ACCOUNT_CODE_METADATA, metadata_batch_items)
            .await
    }

    /// Returns a snapshot of the current block-data buffer.
    fn buffer(&self) -> Result<Arc<BlockDataBuffer>, StoreError> {
        Ok(self
            .block_data_buffer
            .read()
            .map_err(|_| StoreError::LockError)?
            .clone())
    }

    // Helper methods for async operations with spawn_blocking
    // These methods ensure RocksDB I/O doesn't block the tokio runtime

    /// Helper method for async writes
    /// Spawns blocking task to avoid blocking tokio runtime
    pub fn write(
        &self,
        table: &'static str,
        key: Vec<u8>,
        value: Vec<u8>,
    ) -> Result<(), StoreError> {
        let backend = self.backend.clone();
        let mut txn = backend.begin_write()?;
        txn.put(table, &key, &value)?;
        txn.commit()
    }

    /// Helper method for async writes
    /// Spawns blocking task to avoid blocking tokio runtime
    async fn write_async(
        &self,
        table: &'static str,
        key: Vec<u8>,
        value: Vec<u8>,
    ) -> Result<(), StoreError> {
        let backend = self.backend.clone();

        tokio::task::spawn_blocking(move || {
            let mut txn = backend.begin_write()?;
            txn.put(table, &key, &value)?;
            txn.commit()
        })
        .await
        .map_err(|e| StoreError::Custom(format!("Task panicked: {}", e)))?
    }

    /// Helper method for async reads
    /// Spawns blocking task to avoid blocking tokio runtime
    pub async fn read_async(
        &self,
        table: &'static str,
        key: Vec<u8>,
    ) -> Result<Option<Vec<u8>>, StoreError> {
        let backend = self.backend.clone();

        tokio::task::spawn_blocking(move || {
            let txn = backend.begin_read()?;
            txn.get(table, &key)
        })
        .await
        .map_err(|e| StoreError::Custom(format!("Task panicked: {}", e)))?
    }

    /// Helper method for sync reads
    /// Spawns blocking task to avoid blocking tokio runtime
    pub fn read(&self, table: &'static str, key: Vec<u8>) -> Result<Option<Vec<u8>>, StoreError> {
        let backend = self.backend.clone();
        let txn = backend.begin_read()?;
        txn.get(table, &key)
    }

    /// Helper method for batch writes
    /// Spawns blocking task to avoid blocking tokio runtime
    /// This is the most important optimization for healing performance
    pub async fn write_batch_async(
        &self,
        table: &'static str,
        batch_ops: Vec<(Vec<u8>, Vec<u8>)>,
    ) -> Result<(), StoreError> {
        let backend = self.backend.clone();

        tokio::task::spawn_blocking(move || {
            let mut txn = backend.begin_write()?;
            txn.put_batch(table, batch_ops)?;
            txn.commit()
        })
        .await
        .map_err(|e| StoreError::Custom(format!("Task panicked: {}", e)))?
    }

    /// Helper method for batch writes
    pub fn write_batch(
        &self,
        table: &'static str,
        batch_ops: Vec<(Vec<u8>, Vec<u8>)>,
    ) -> Result<(), StoreError> {
        let backend = self.backend.clone();
        let mut txn = backend.begin_write()?;
        txn.put_batch(table, batch_ops)?;
        txn.commit()
    }

    pub async fn add_fullsync_batch(&self, headers: Vec<BlockHeader>) -> Result<(), StoreError> {
        self.write_batch_async(
            FULLSYNC_HEADERS,
            headers
                .into_iter()
                .map(|header| (header.number.to_le_bytes().to_vec(), header.encode_to_vec()))
                .collect(),
        )
        .await
    }

    pub async fn read_fullsync_batch(
        &self,
        start: BlockNumber,
        limit: u64,
    ) -> Result<Vec<Option<BlockHeader>>, StoreError> {
        let mut res = vec![];
        let read_tx = self.backend.begin_read()?;
        // TODO: use read_bulk here
        for key in start..start + limit {
            let header_opt = read_tx
                .get(FULLSYNC_HEADERS, &key.to_le_bytes())?
                .map(|header| BlockHeader::decode(&header))
                .transpose()?;
            res.push(header_opt);
        }
        Ok(res)
    }

    pub async fn clear_fullsync_headers(&self) -> Result<(), StoreError> {
        self.backend.clear_table(FULLSYNC_HEADERS)
    }

    /// Delete a key from a table
    pub fn delete(&self, table: &'static str, key: Vec<u8>) -> Result<(), StoreError> {
        let mut txn = self.backend.begin_write()?;
        txn.delete(table, &key)?;
        txn.commit()
    }

    pub fn store_block_updates(&self, update_batch: UpdateBatch) -> Result<(), StoreError> {
        self.apply_updates(update_batch)
    }

    /// Compute `(parent_state_root, last_state_root, last_block_number,
    /// last_block_hash)` for a batch's trie update: the state root of the first
    /// block's parent, the last block's own state root, and the last block's
    /// number and hash (the identity the journal records for the committed
    /// layer). Used by `apply_updates` for both the live and full-sync paths
    /// (which share the single persist worker).
    fn batch_state_roots(
        &self,
        update_batch: &UpdateBatch,
    ) -> Result<(H256, H256, BlockNumber, H256), StoreError> {
        let parent_state_root = self
            .get_block_header_by_hash(
                update_batch
                    .blocks
                    .first()
                    .ok_or(StoreError::UpdateBatchNoBlocks)?
                    .header
                    .parent_hash,
            )?
            .map(|header| header.state_root)
            .unwrap_or_default();
        let last_block = update_batch
            .blocks
            .last()
            .ok_or(StoreError::UpdateBatchNoBlocks)?;
        let last_state_root = last_block.header.state_root;
        let last_block_number = last_block.header.number;
        let last_block_hash = last_block.hash();
        Ok((
            parent_state_root,
            last_state_root,
            last_block_number,
            last_block_hash,
        ))
    }

    /// Single path for all updates: hand the whole unit (block data + one aggregate trie
    /// diff) to the SINGLE persist worker and wait for its ack. `commit_depth` selects the
    /// commit gate; `wait_for_flush` selects when the worker acks (see [`UpdateBatch`]).
    fn apply_updates(&self, update_batch: UpdateBatch) -> Result<(), StoreError> {
        let (parent_state_root, last_state_root, last_block_number, last_block_hash) =
            self.batch_state_roots(&update_batch)?;

        let UpdateBatch {
            account_updates,
            storage_updates,
            blocks,
            receipts,
            code_updates,
            commit_depth,
            wait_for_flush,
        } = update_batch;

        // Register before handing off to the worker and before this returns, so
        // any reader opening this root blocks in `gated_snapshot` until the
        // layer is installed rather than snapshotting a stale cache.
        self.pending_trie_roots.register(last_state_root)?;

        // Pair blocks with receipts. Single-block fast path avoids a HashMap
        // allocation; full-sync batch joins by hash.
        let blocks_with_receipts: Vec<(Block, Vec<Receipt>)> = if blocks.len() == 1 {
            let block = blocks.into_iter().next().expect("len == 1");
            let hash = block.hash();
            let r = receipts
                .into_iter()
                .find(|(h, _)| *h == hash)
                .map(|(_, r)| r)
                .unwrap_or_default();
            vec![(block, r)]
        } else {
            let mut receipts_by_hash: std::collections::HashMap<BlockHash, Vec<Receipt>> =
                receipts.into_iter().collect();
            blocks
                .into_iter()
                .map(|b| {
                    let r = receipts_by_hash.remove(&b.hash()).unwrap_or_default();
                    (b, r)
                })
                .collect()
        };

        // Send to the persist worker and wait for its ack.
        // wait_for_flush=false: worker acks after staging; the ack carries the PRIOR flush
        //   result so a disk error surfaces on the next call.
        // wait_for_flush=true: worker acks after flush, bounding in-flight work to ~1.
        let (ack_tx, ack_rx) = sync_channel(1);
        self.persist_tx
            .send(PersistMessage::Block(Box::new(BlockPersist {
                blocks: blocks_with_receipts,
                codes: code_updates,
                parent_state_root,
                child_state_root: last_state_root,
                account_updates,
                storage_updates,
                commit_depth,
                wait_for_flush,
                block_number: last_block_number,
                block_hash: last_block_hash,
                ack: ack_tx,
            })))
            .map_err(|e| StoreError::Custom(format!("failed to send block persist: {e}")))?;
        ack_rx
            .recv()
            .map_err(|e| StoreError::Custom(format!("block persist ack failed: {e}")))??;

        Ok(())
    }

    /// Opens (or creates) a store at `path` with the default [`StoreConfig`].
    ///
    /// Production callers that need to override storage tunables (e.g. the RocksDB
    /// block cache size from a CLI option) should use [`Store::new_with_config`].
    pub fn new(path: impl AsRef<Path>, engine_type: EngineType) -> Result<Self, StoreError> {
        Self::new_with_config(path, engine_type, StoreConfig::default())
    }

    /// Opens (or creates) a store at `path`, applying the supplied [`StoreConfig`].
    pub fn new_with_config(
        path: impl AsRef<Path>,
        engine_type: EngineType,
        // `config` only feeds the RocksDB backend; without that feature it is unused.
        #[cfg_attr(not(feature = "rocksdb"), allow(unused_variables))] config: StoreConfig,
    ) -> Result<Self, StoreError> {
        let db_path = path.as_ref().to_path_buf();

        if engine_type != EngineType::InMemory {
            let version = read_store_schema_version(&db_path)?;

            match version {
                None if db_path.exists() && dir_contains_legacy_db(&db_path)? => {
                    // Pre-metadata DB — cannot migrate safely
                    return Err(StoreError::NotFoundDBVersion);
                }
                None => {
                    // No metadata and no recognizable database files. The directory
                    // may still hold unrelated files (e.g. a JWT secret placed in the
                    // datadir by tooling such as EthDocker, see issue #5680), so treat
                    // this as a fresh datadir and write the initial metadata instead
                    // of erroring out.
                    init_metadata_file(&db_path)?;
                }
                // Neither arm below runs a migration, so neither is a
                // `MigrationFailed`: that variant tells the operator the data may be
                // half-converted, which is false here — the database is untouched.
                Some(v) if v < 1 => {
                    // No ethrex ever wrote a v0 marker; the file is corrupt or hand-edited.
                    return Err(StoreError::IncompatibleDBVersion {
                        found: v,
                        expected: STORE_SCHEMA_VERSION,
                    });
                }
                Some(v) if v > STORE_SCHEMA_VERSION => {
                    // Written by a newer ethrex. Downgrading is unsupported, but the
                    // database is intact: a binary that speaks v{v} opens it as is.
                    return Err(StoreError::IncompatibleDBVersion {
                        found: v,
                        expected: STORE_SCHEMA_VERSION,
                    });
                }
                #[cfg(feature = "rocksdb")]
                Some(v) if v < STORE_SCHEMA_VERSION => {
                    // Open backend, run migrations, then drop obsolete CFs.
                    // Cleanup must happen AFTER migrations so legacy CFs (e.g.
                    // `receipts`) are still readable during the migration.
                    let rocksdb = Arc::new(RocksDBBackend::open(
                        &path,
                        config.rocksdb_block_cache_size,
                    )?);
                    crate::migrations::run_pending_migrations(rocksdb.as_ref(), &db_path, v)?;
                    rocksdb.drop_obsolete_cfs(&path);
                    let backend: Arc<dyn crate::api::StorageBackend> = rocksdb;
                    return Self::from_backend(
                        backend,
                        db_path,
                        DB_COMMIT_THRESHOLD,
                        config.persist_channel_capacity,
                    );
                }
                Some(_) => {
                    // version == STORE_SCHEMA_VERSION, proceed normally.
                    // Without the `rocksdb` feature this also covers v < target,
                    // but that path is unreachable since InMemory is the only
                    // engine type and the outer guard excludes it.
                }
            }
        }

        match engine_type {
            #[cfg(feature = "rocksdb")]
            EngineType::RocksDB => {
                let rocksdb = RocksDBBackend::open(&path, config.rocksdb_block_cache_size)?;
                rocksdb.drop_obsolete_cfs(&path);
                let backend: Arc<dyn StorageBackend> = Arc::new(rocksdb);
                Self::from_backend(
                    backend,
                    db_path,
                    DB_COMMIT_THRESHOLD,
                    config.persist_channel_capacity,
                )
            }
            EngineType::InMemory => {
                let backend = Arc::new(InMemoryBackend::open()?);
                Self::from_backend(
                    backend,
                    db_path,
                    IN_MEMORY_COMMIT_THRESHOLD,
                    config.persist_channel_capacity,
                )
            }
        }
    }

    fn from_backend(
        backend: Arc<dyn StorageBackend>,
        db_path: PathBuf,
        commit_threshold: usize,
        persist_channel_capacity: usize,
    ) -> Result<Self, StoreError> {
        debug!("Initializing Store with {commit_threshold} in-memory diff-layers");
        let (fkv_tx, fkv_rx) = std::sync::mpsc::sync_channel(0);
        let persist_cap = persist_channel_capacity.max(1); // clamp: 0 would be a rendezvous channel
        let (persist_tx, persist_rx) = std::sync::mpsc::sync_channel(persist_cap);

        let (last_written, initial_flushed_upto) = {
            let tx = backend.begin_read()?;
            let last_written = tx
                .get(MISC_VALUES, "last_written".as_bytes())?
                .unwrap_or_else(|| vec![0u8; 64]);
            let last_written = if last_written == [0xff] {
                vec![0xff; 64]
            } else {
                last_written
            };
            let initial_flushed_upto = match tx.get(MISC_VALUES, FLUSHED_UPTO_KEY)? {
                Some(bytes) => decode_flushed_upto(&bytes)?,
                None => 0,
            };
            (last_written, initial_flushed_upto)
        };
        let mut initial_buffer = BlockDataBuffer::new();
        initial_buffer.set_flushed_upto(initial_flushed_upto);

        let mut background_threads = Vec::new();
        let safe_commit_root = Arc::new(RwLock::new(H256::zero()));
        let mut store = Self {
            db_path,
            backend,
            chain_config: Default::default(),
            latest_block_header: Default::default(),
            trie_cache: Arc::new(RwLock::new(Arc::new(TrieLayerCache::new_with_safe_commit(
                commit_threshold,
                safe_commit_root.clone(),
            )))),
            flatkeyvalue_control_tx: fkv_tx,
            block_data_buffer: Arc::new(RwLock::new(Arc::new(initial_buffer))),
            persist_tx,
            pending_trie_roots: Arc::new(PendingTrieRoots::default()),
            last_computed_flatkeyvalue: Arc::new(RwLock::new(last_written)),
            account_code_cache: Arc::new(Mutex::new(CodeCache::default())),
            code_metadata_cache: Arc::new(Mutex::new(LruCache::with_hasher(
                NonZeroUsize::new(CODE_METADATA_CACHE_MAX_ENTRIES).unwrap_or(NonZeroUsize::MIN),
                FxBuildHasher,
            ))),
            fcu_lock: Arc::new(tokio::sync::Mutex::new(())),
            safe_commit_root,
            journal_pruning_paused: Arc::new(std::sync::atomic::AtomicBool::new(false)),
            background_threads: Default::default(),
        };
        let backend_clone = store.backend.clone();
        let last_computed_fkv = store.last_computed_flatkeyvalue.clone();
        background_threads.push(std::thread::spawn(move || {
            let rx = fkv_rx;
            // Wait for the first Continue to start generation
            loop {
                match rx.recv() {
                    Ok(FKVGeneratorControlMessage::Continue) => break,
                    Ok(FKVGeneratorControlMessage::Stop) => {}
                    Err(std::sync::mpsc::RecvError) => {
                        debug!("Closing FlatKeyValue generator.");
                        return;
                    }
                }
            }

            let _ = flatkeyvalue_generator(&backend_clone, &last_computed_fkv, &rx)
                .inspect_err(|err| error!("Error while generating FlatKeyValue: {err}"));
        }));
        // The single persist worker: sole swapper of `block_data_buffer`, sole
        // builder of trie diff-layers. One DB transaction per `Block` message.
        let persist_backend = store.backend.clone();
        let persist_buffer = store.block_data_buffer.clone();
        let persist_trie_cache = store.trie_cache.clone();
        let persist_pending_roots = store.pending_trie_roots.clone();
        let persist_fkv_ctl = store.flatkeyvalue_control_tx.clone();
        background_threads.push(std::thread::spawn(move || {
            let rx = persist_rx;
            // Carries the prior flush result: the live path acks after staging,
            // so a disk failure surfaces on the next message's ack.
            let mut last_flush_result: Result<(), StoreError> = Ok(());
            loop {
                match rx.recv() {
                    Ok(PersistMessage::Block(bp)) => {
                        let bp = *bp;
                        // Stage block data (sole swapper of the buffer; codes
                        // are batch-level and attributed to the first block).
                        let staged = mutate_block_buffer(&persist_buffer, move |b| {
                            let mut codes = Some(bp.codes);
                            for (block, receipts) in bp.blocks {
                                b.insert(block, receipts, codes.take().unwrap_or_default());
                            }
                        });
                        if let Err(e) = staged {
                            // Stage failure is terminal for this message.
                            // Clear the pending root so gated readers are not
                            // blocked forever (apply_trie_phase1, which normally
                            // does this, is skipped when we continue here).
                            persist_pending_roots.clear(bp.child_state_root);
                            let _ = bp.ack.send(Err(e));
                            continue;
                        }
                        // ACK-AFTER-STAGING: ack now, carrying the prior flush result.
                        // NOTE: this acks block validity BEFORE apply_trie_phase1
                        // installs the trie layer below. A phase-1 failure (only
                        // reachable via lock poisoning, which is already fatal) is
                        // therefore deferred to the next block's ack via
                        // last_flush_result rather than attributed to this block;
                        // the pending root is still cleared unconditionally, so
                        // gated readers error rather than hang.
                        if !bp.wait_for_flush {
                            let _ = bp
                                .ack
                                .send(std::mem::replace(&mut last_flush_result, Ok(())));
                        }
                        // Build + install the trie layer; clear the read gate.
                        if let Err(err) = apply_trie_phase1(
                            &persist_trie_cache,
                            &persist_pending_roots,
                            bp.parent_state_root,
                            bp.child_state_root,
                            bp.block_number,
                            bp.block_hash,
                            bp.account_updates,
                            bp.storage_updates,
                        ) {
                            error!("persist worker trie phase-1 failed: {err}");
                            if bp.wait_for_flush {
                                let _ = bp.ack.send(Err(err));
                            } else {
                                last_flush_result = Err(err);
                            }
                            continue;
                        }
                        // Flush block data + commit bottom trie layer when due.
                        let flushed = flush_block_data(persist_backend.as_ref(), &persist_buffer)
                            .inspect_err(|err| error!("flush_block_data failed: {err}"))
                            .and_then(|_| {
                                commit_trie_if_due(
                                    persist_backend.as_ref(),
                                    &persist_trie_cache,
                                    &persist_fkv_ctl,
                                    bp.parent_state_root,
                                    bp.commit_depth,
                                    bp.wait_for_flush,
                                )
                            });
                        // ACK-AFTER-FLUSH: ack now (bounds in-flight work to ~1), folding in
                        // any prior deferred error. ACK-AFTER-STAGING: stash for the next ack.
                        if bp.wait_for_flush {
                            let prior = std::mem::replace(&mut last_flush_result, Ok(()));
                            let _ = bp.ack.send(prior.and(flushed));
                        } else {
                            last_flush_result = flushed;
                        }
                    }
                    Ok(PersistMessage::Commit(root)) => match persist_trie_cache.read() {
                        Ok(guard) => {
                            let trie = guard.clone();
                            drop(guard);
                            // Forkchoice-driven flush is the live (non-batch) path, so
                            // journaling is enabled: pass `is_batch = false`.
                            let _ = commit_to_disk(
                                persist_backend.as_ref(),
                                &persist_fkv_ctl,
                                &persist_trie_cache,
                                &trie,
                                root,
                                false,
                            )
                            .inspect_err(|err| error!("commit_to_disk failed: {err}"));
                        }
                        Err(_) => error!("trie cache lock poisoned during commit"),
                    },
                    Ok(PersistMessage::Ping(ack)) => {
                        // Idle handshake: reached only after all earlier Block
                        // messages are fully processed. Carry the pending flush
                        // result so a live-path failure is not silently dropped.
                        let _ = ack.send(std::mem::replace(&mut last_flush_result, Ok(())));
                    }
                    Ok(PersistMessage::Shutdown { ack }) => {
                        // Graceful shutdown: drain (already guaranteed by FIFO) and
                        // force-flush the not-yet-flushed block-data tail. The trie
                        // diff-layers stay in memory and are dropped on exit: the
                        // on-disk trie is a single-version path store, so committing
                        // the non-finalized tail would make a post-restart reorg
                        // unrecoverable. Those layers re-execute on the next start.
                        let result = flush_block_data(persist_backend.as_ref(), &persist_buffer);
                        let prior = std::mem::replace(&mut last_flush_result, Ok(()));
                        let _ = ack.send(prior.and(result));
                        // No more work will follow a shutdown request.
                        return;
                    }
                    Err(_) => return,
                }
            }
        }));
        store.background_threads = Arc::new(ThreadList {
            list: background_threads,
        });
        Ok(store)
    }

    /// Opens (or creates) a store at `store_path` and seeds it from the
    /// given genesis file, using the default [`StoreConfig`].
    pub async fn new_from_genesis(
        store_path: &Path,
        engine_type: EngineType,
        genesis_path: &str,
    ) -> Result<Self, StoreError> {
        Self::new_from_genesis_with_config(
            store_path,
            engine_type,
            genesis_path,
            StoreConfig::default(),
        )
        .await
    }

    /// Opens (or creates) a store at `store_path` from genesis, applying the
    /// supplied [`StoreConfig`].
    pub async fn new_from_genesis_with_config(
        store_path: &Path,
        engine_type: EngineType,
        genesis_path: &str,
        config: StoreConfig,
    ) -> Result<Self, StoreError> {
        let file = std::fs::File::open(genesis_path)
            .map_err(|error| StoreError::Custom(format!("Failed to open genesis file: {error}")))?;
        let reader = std::io::BufReader::new(file);
        let genesis: Genesis = serde_json::from_reader(reader)
            .map_err(|e| StoreError::Custom(format!("Failed to deserialize genesis file: {e}")))?;
        let mut store = Self::new_with_config(store_path, engine_type, config)?;
        store.add_initial_state(genesis).await?;
        Ok(store)
    }

    pub async fn get_account_info(
        &self,
        block_number: BlockNumber,
        address: Address,
    ) -> Result<Option<AccountInfo>, StoreError> {
        match self.get_canonical_block_hash(block_number).await? {
            Some(block_hash) => self.get_account_info_by_hash(block_hash, address),
            None => Ok(None),
        }
    }

    pub fn get_account_info_by_hash(
        &self,
        block_hash: BlockHash,
        address: Address,
    ) -> Result<Option<AccountInfo>, StoreError> {
        let Some(state_trie) = self.state_trie(block_hash)? else {
            return Ok(None);
        };
        let hashed_address = hash_address_fixed(&address);

        let Some(encoded_state) = state_trie.get(hashed_address.as_bytes())? else {
            return Ok(None);
        };

        let account_state = AccountState::decode(&encoded_state)?;
        Ok(Some(AccountInfo {
            code_hash: account_state.code_hash,
            balance: account_state.balance,
            nonce: account_state.nonce,
        }))
    }

    pub fn get_account_state_by_acc_hash(
        &self,
        block_hash: BlockHash,
        account_hash: H256,
    ) -> Result<Option<AccountState>, StoreError> {
        let Some(state_trie) = self.state_trie(block_hash)? else {
            return Ok(None);
        };
        let Some(encoded_state) = state_trie.get(account_hash.as_bytes())? else {
            return Ok(None);
        };
        let account_state = AccountState::decode(&encoded_state)?;
        Ok(Some(account_state))
    }

    pub async fn get_fork_id(&self) -> Result<ForkId, StoreError> {
        let chain_config = self.get_chain_config();
        let genesis_header = self
            .load_block_header(0)?
            .ok_or(StoreError::MissingEarliestBlockNumber)?;
        let block_header = self.latest_block_header.get();

        Ok(ForkId::new(
            chain_config,
            genesis_header,
            block_header.timestamp,
            block_header.number,
        ))
    }

    pub async fn get_code_by_account_address(
        &self,
        block_number: BlockNumber,
        address: Address,
    ) -> Result<Option<Code>, StoreError> {
        let Some(block_hash) = self.get_canonical_block_hash(block_number).await? else {
            return Ok(None);
        };
        let Some(state_trie) = self.state_trie(block_hash)? else {
            return Ok(None);
        };
        let hashed_address = hash_address_fixed(&address);
        let Some(encoded_state) = state_trie.get(hashed_address.as_bytes())? else {
            return Ok(None);
        };
        let account_state = AccountState::decode(&encoded_state)?;
        self.get_account_code(account_state.code_hash)
    }

    pub async fn get_nonce_by_account_address(
        &self,
        block_number: BlockNumber,
        address: Address,
    ) -> Result<Option<u64>, StoreError> {
        let Some(block_hash) = self.get_canonical_block_hash(block_number).await? else {
            return Ok(None);
        };
        let Some(state_trie) = self.state_trie(block_hash)? else {
            return Ok(None);
        };
        let hashed_address = hash_address_fixed(&address);
        let Some(encoded_state) = state_trie.get(hashed_address.as_bytes())? else {
            return Ok(None);
        };
        let account_state = AccountState::decode(&encoded_state)?;
        Ok(Some(account_state.nonce))
    }

    /// Applies account updates based on the block's latest storage state
    /// and returns the new state root after the updates have been applied.
    pub fn apply_account_updates_batch(
        &self,
        block_hash: BlockHash,
        account_updates: &[AccountUpdate],
    ) -> Result<Option<AccountUpdatesList>, StoreError> {
        let Some(mut state_trie) = self.state_trie(block_hash)? else {
            return Ok(None);
        };

        Ok(Some(self.apply_account_updates_from_trie_batch(
            &mut state_trie,
            account_updates,
        )?))
    }

    pub fn apply_account_updates_from_trie_batch<'a>(
        &self,
        state_trie: &mut Trie,
        account_updates: impl IntoIterator<Item = &'a AccountUpdate>,
    ) -> Result<AccountUpdatesList, StoreError> {
        let mut ret_storage_updates = Vec::new();
        let mut code_updates = Vec::new();
        let state_root = state_trie.hash_no_commit(&NativeCrypto);
        for update in account_updates {
            let hashed_address = hash_address_fixed(&update.address);
            if update.removed {
                // Remove account from trie
                state_trie.remove(hashed_address.as_bytes())?;
                continue;
            }
            // Add or update AccountState in the trie
            // Fetch current state or create a new state to be inserted
            let mut account_state = match state_trie.get(hashed_address.as_bytes())? {
                Some(encoded_state) => AccountState::decode(&encoded_state)?,
                None => AccountState::default(),
            };
            if update.removed_storage {
                account_state.storage_root = EMPTY_TRIE_HASH;
            }
            if let Some(info) = &update.info {
                account_state.nonce = info.nonce;
                account_state.balance = info.balance;
                account_state.code_hash = info.code_hash;
                // Store updated code in DB
                if let Some(code) = &update.code {
                    code_updates.push((info.code_hash, code.clone()));
                }
            }
            // Store the added storage in the account's storage trie and compute its new root
            if !update.added_storage.is_empty() {
                let mut storage_trie =
                    self.open_storage_trie(hashed_address, state_root, account_state.storage_root)?;
                for (storage_key, storage_value) in &update.added_storage {
                    let hashed_key = hash_key(storage_key);
                    if storage_value.is_zero() {
                        storage_trie.remove(&hashed_key)?;
                    } else {
                        storage_trie.insert(hashed_key, storage_value.encode_to_vec())?;
                    }
                }
                let (storage_hash, storage_updates) =
                    storage_trie.collect_changes_since_last_hash(&NativeCrypto);
                account_state.storage_root = storage_hash;
                ret_storage_updates.push((hashed_address, storage_updates));
            }
            state_trie.insert(
                hashed_address.as_bytes().to_vec(),
                account_state.encode_to_vec(),
            )?;
        }
        let (state_trie_hash, state_updates) =
            state_trie.collect_changes_since_last_hash(&NativeCrypto);

        Ok(AccountUpdatesList {
            state_trie_hash,
            state_updates,
            storage_updates: ret_storage_updates,
            code_updates,
        })
    }

    /// Performs the same actions as apply_account_updates_from_trie
    ///  but also returns the used storage tries with witness recorded
    pub fn apply_account_updates_from_trie_with_witness(
        &self,
        mut state_trie: Trie,
        account_updates: &[AccountUpdate],
        mut storage_tries: StorageTries,
    ) -> Result<(StorageTries, AccountUpdatesList), StoreError> {
        let mut ret_storage_updates = Vec::new();

        let mut code_updates = Vec::new();

        let state_root = state_trie.hash_no_commit(&NativeCrypto);

        for update in account_updates.iter() {
            let hashed_address = hash_address(&update.address);

            if update.removed {
                // Remove account from trie
                state_trie.remove(&hashed_address)?;

                continue;
            }

            // Add or update AccountState in the trie
            // Fetch current state or create a new state to be inserted
            let mut account_state = match state_trie.get(&hashed_address)? {
                Some(encoded_state) => AccountState::decode(&encoded_state)?,
                None => AccountState::default(),
            };

            if update.removed_storage {
                account_state.storage_root = EMPTY_TRIE_HASH;
            }

            if let Some(info) = &update.info {
                account_state.nonce = info.nonce;

                account_state.balance = info.balance;

                account_state.code_hash = info.code_hash;

                // Store updated code in DB
                if let Some(code) = &update.code {
                    code_updates.push((info.code_hash, code.clone()));
                }
            }

            // Store the added storage in the account's storage trie and compute its new root
            if !update.added_storage.is_empty() {
                let (_witness, storage_trie) = match storage_tries.entry(update.address) {
                    Entry::Occupied(value) => value.into_mut(),
                    Entry::Vacant(vacant) => {
                        let trie = self.open_storage_trie(
                            H256::from_slice(&hashed_address),
                            state_root,
                            account_state.storage_root,
                        )?;
                        vacant.insert(TrieLogger::open_trie(trie))
                    }
                };

                for (storage_key, storage_value) in &update.added_storage {
                    let hashed_key = hash_key(storage_key);

                    if storage_value.is_zero() {
                        storage_trie.remove(&hashed_key)?;
                    } else {
                        storage_trie.insert(hashed_key, storage_value.encode_to_vec())?;
                    }
                }

                let (storage_hash, storage_updates) =
                    storage_trie.collect_changes_since_last_hash(&NativeCrypto);

                account_state.storage_root = storage_hash;

                ret_storage_updates.push((H256::from_slice(&hashed_address), storage_updates));
            }

            state_trie.insert(hashed_address, account_state.encode_to_vec())?;
        }

        let (state_trie_hash, state_updates) =
            state_trie.collect_changes_since_last_hash(&NativeCrypto);

        let account_updates_list = AccountUpdatesList {
            state_trie_hash,
            state_updates,
            storage_updates: ret_storage_updates,
            code_updates,
        };

        Ok((storage_tries, account_updates_list))
    }

    /// Adds all genesis accounts and returns the genesis block's state_root
    pub async fn setup_genesis_state_trie(
        &self,
        genesis_accounts: BTreeMap<Address, GenesisAccount>,
    ) -> Result<H256, StoreError> {
        let mut storage_trie_nodes = vec![];
        let mut genesis_state_trie = self.open_direct_state_trie(EMPTY_TRIE_HASH)?;
        for (address, account) in genesis_accounts {
            let hashed_address = hash_address(&address);
            let h256_hashed_address = H256::from_slice(&hashed_address);

            // Store account code (as this won't be stored in the trie)
            let code = Code::from_bytecode(account.code, &NativeCrypto);
            let code_hash = code.hash;
            self.add_account_code(code).await?;

            // Store the account's storage in a clean storage trie and compute its root
            let mut storage_trie =
                self.open_direct_storage_trie(h256_hashed_address, EMPTY_TRIE_HASH)?;
            for (storage_key, storage_value) in account.storage {
                if !storage_value.is_zero() {
                    let hashed_key = hash_key(&H256(storage_key.to_big_endian()));
                    storage_trie.insert(hashed_key, storage_value.encode_to_vec())?;
                }
            }

            let (storage_root, storage_nodes) =
                storage_trie.collect_changes_since_last_hash(&NativeCrypto);

            storage_trie_nodes.extend(
                storage_nodes
                    .into_iter()
                    .map(|(path, n)| (apply_prefix(Some(h256_hashed_address), path).into_vec(), n)),
            );

            // Add account to trie
            let account_state = AccountState {
                nonce: account.nonce,
                balance: account.balance,
                storage_root,
                code_hash,
            };
            genesis_state_trie.insert(hashed_address, account_state.encode_to_vec())?;
        }

        let (state_root, account_trie_nodes) =
            genesis_state_trie.collect_changes_since_last_hash(&NativeCrypto);
        let account_trie_nodes = account_trie_nodes
            .into_iter()
            .map(|(path, n)| (apply_prefix(None, path).into_vec(), n))
            .collect::<Vec<_>>();

        let mut tx = self.backend.begin_write()?;
        tx.put_batch(ACCOUNT_TRIE_NODES, account_trie_nodes)?;
        tx.put_batch(STORAGE_TRIE_NODES, storage_trie_nodes)?;
        tx.commit()?;

        Ok(state_root)
    }

    // Key format: block_number (8 bytes, big-endian) + block_hash (32 bytes)
    fn make_witness_key(block_number: u64, block_hash: &BlockHash) -> Vec<u8> {
        let mut composite_key = Vec::with_capacity(8 + 32);
        composite_key.extend_from_slice(&block_number.to_be_bytes());
        composite_key.extend_from_slice(block_hash.as_bytes());
        composite_key
    }

    /// Stores a pre-serialized execution witness for a block.
    ///
    /// The witness is converted to RPC format (RpcExecutionWitness) before storage
    /// to avoid expensive `encode_subtrie` traversal on every read. This pre-computes
    /// the serialization at write time instead of read time.
    pub fn store_witness(
        &self,
        block_hash: BlockHash,
        block_number: u64,
        witness: ExecutionWitness,
    ) -> Result<(), StoreError> {
        // Convert to RPC format once at storage time
        let rpc_witness = RpcExecutionWitness::try_from(witness)?;
        let key = Self::make_witness_key(block_number, &block_hash);
        let value = serde_json::to_vec(&rpc_witness)?;
        self.write(EXECUTION_WITNESSES, key, value)?;
        // Clean up old witnesses (keep only last 128)
        self.cleanup_old_witnesses(block_number)
    }

    fn cleanup_old_witnesses(&self, latest_block_number: u64) -> Result<(), StoreError> {
        // If we have less than 128 blocks, no cleanup needed
        if latest_block_number <= MAX_WITNESSES {
            return Ok(());
        }

        let threshold = latest_block_number - MAX_WITNESSES;

        if let Some(oldest_block_number) = self.get_oldest_witness_number()? {
            let prefix = oldest_block_number.to_be_bytes();
            let mut to_delete = Vec::new();

            {
                let read_txn = self.backend.begin_read()?;
                let iter = read_txn.prefix_iterator(EXECUTION_WITNESSES, &prefix)?;

                // We may have multiple witnesses for the same block number (forks)
                for item in iter {
                    let (key, _value) = item?;
                    let mut block_number_bytes = [0u8; 8];
                    block_number_bytes.copy_from_slice(&key[0..8]);
                    let block_number = u64::from_be_bytes(block_number_bytes);
                    if block_number > threshold {
                        break;
                    }
                    to_delete.push(key.to_vec());
                }
            }

            for key in to_delete {
                self.delete(EXECUTION_WITNESSES, key)?;
            }
        };

        self.update_oldest_witness_number(threshold + 1)?;

        Ok(())
    }

    fn update_oldest_witness_number(&self, oldest_block_number: u64) -> Result<(), StoreError> {
        self.write(
            MISC_VALUES,
            b"oldest_witness_block_number".to_vec(),
            oldest_block_number.to_le_bytes().to_vec(),
        )?;
        Ok(())
    }

    fn get_oldest_witness_number(&self) -> Result<Option<u64>, StoreError> {
        let Some(value) = self.read(MISC_VALUES, b"oldest_witness_block_number".to_vec())? else {
            return Ok(None);
        };

        let array: [u8; 8] = value.as_slice().try_into().map_err(|_| {
            StoreError::Custom("Invalid oldest witness block number bytes".to_string())
        })?;
        Ok(Some(u64::from_le_bytes(array)))
    }

    /// Returns the raw JSON bytes of a cached witness for a block.
    ///
    /// This is the most efficient method for the RPC handler since it avoids
    /// deserialization and re-serialization. The bytes can be parsed directly
    /// as a JSON Value for the RPC response.
    pub fn get_witness_json_bytes(
        &self,
        block_number: u64,
        block_hash: BlockHash,
    ) -> Result<Option<Vec<u8>>, StoreError> {
        let key = Self::make_witness_key(block_number, &block_hash);
        self.read(EXECUTION_WITNESSES, key)
    }

    /// Returns the deserialized RpcExecutionWitness for a block.
    ///
    /// Prefer `get_witness_json_bytes` when you need to return the witness
    /// as JSON (e.g., for RPC responses) to avoid re-serialization.
    pub fn get_witness_by_number_and_hash(
        &self,
        block_number: u64,
        block_hash: BlockHash,
    ) -> Result<Option<RpcExecutionWitness>, StoreError> {
        let key = Self::make_witness_key(block_number, &block_hash);
        match self.read(EXECUTION_WITNESSES, key)? {
            Some(value) => {
                let witness: RpcExecutionWitness = serde_json::from_slice(&value)?;
                Ok(Some(witness))
            }
            None => Ok(None),
        }
    }

    /// Stores a block access list for a given block hash.
    pub fn store_block_access_list(
        &self,
        block_hash: BlockHash,
        bal: &BlockAccessList,
    ) -> Result<(), StoreError> {
        let key = block_hash.as_bytes().to_vec();
        let mut value = vec![];
        bal.encode(&mut value);
        self.write(BLOCK_ACCESS_LISTS, key, value)
    }

    /// Returns the block access list for a given block hash, if stored.
    pub fn get_block_access_list(
        &self,
        block_hash: BlockHash,
    ) -> Result<Option<BlockAccessList>, StoreError> {
        let key = block_hash.as_bytes().to_vec();
        match self.read(BLOCK_ACCESS_LISTS, key)? {
            Some(value) => {
                let bal = BlockAccessList::decode(&value)
                    .map_err(|e| StoreError::Custom(format!("Failed to decode BAL: {e}")))?;
                Ok(Some(bal))
            }
            None => Ok(None),
        }
    }

    pub async fn add_initial_state(&mut self, genesis: Genesis) -> Result<(), StoreError> {
        self.add_initial_state_inner(genesis, false).await
    }

    /// Like [`Store::add_initial_state`], but trusts a pre-existing datadir's
    /// state instead of validating it against the provided genesis. If a genesis
    /// header is already stored, it is kept as-is rather than recomputing the
    /// genesis state root from `genesis.alloc` and rejecting on mismatch. The
    /// chain config from the genesis file is still applied either way.
    ///
    /// Intended for booting a datadir produced out-of-band (e.g. by a state
    /// generator that writes the state trie directly and emits a genesis file
    /// with an empty `alloc`), where the operator vouches for the stored state
    /// root. Has no effect on a fresh datadir: the genesis is built normally.
    pub async fn add_initial_state_skip_validation(
        &mut self,
        genesis: Genesis,
    ) -> Result<(), StoreError> {
        self.add_initial_state_inner(genesis, true).await
    }

    async fn add_initial_state_inner(
        &mut self,
        genesis: Genesis,
        skip_genesis_validation: bool,
    ) -> Result<(), StoreError> {
        debug!("Storing initial state from genesis");

        // Obtain genesis block
        let genesis_block = genesis.get_block();
        let genesis_block_number = genesis_block.header.number;

        let genesis_hash = genesis_block.hash();

        let stored_genesis_header = self.load_block_header(genesis_block_number)?;

        // Always set the chain config from the genesis file. The in-memory
        // `chain_config` starts at `Default::default()` on every boot and is
        // not reloaded from the datadir, so skipping this would leave the store
        // with the wrong chainId and an empty fork schedule. Skip-validation
        // only waives the genesis state-root/header check; the `config` section
        // of the genesis file is still authoritative and must be applied.
        self.set_chain_config(&genesis.config).await?;

        // The cache can't be empty. Clamp the head to the durable block: after a
        // crash, `LatestBlockNumber` can be ahead of `flushed_upto` (FCU writes the
        // head synchronously while block bodies are buffered), so loading the raw
        // latest header would brick boot when its body was never flushed.
        if let Some(latest) = self.load_latest_block_number().await? {
            self.anchor_to_durable_head(latest).await?;
        }

        match stored_genesis_header {
            Some(header) if skip_genesis_validation => {
                info!(
                    stored_genesis = %header.hash(),
                    "Skipping genesis state validation; trusting the genesis header and state already stored in the datadir"
                );
                return Ok(());
            }
            Some(header) if header.hash() == genesis_hash => {
                info!("Received genesis file matching a previously stored one, nothing to do");
                return Ok(());
            }
            Some(_) => {
                error!(
                    "The chain configuration stored in the database is incompatible with the provided configuration. If you intended to switch networks, choose another datadir or clear the database (e.g., run `ethrex removedb`) and try again."
                );
                return Err(StoreError::IncompatibleChainConfig);
            }
            None => {
                self.add_block_header(genesis_hash, genesis_block.header.clone())
                    .await?
            }
        }
        // Store genesis accounts
        // TODO: Should we use this root instead of computing it before the block hash check?
        let genesis_state_root = self.setup_genesis_state_trie(genesis.alloc).await?;
        debug_assert_eq!(genesis_state_root, genesis_block.header.state_root);

        // Store genesis block
        info!(hash = %genesis_hash, "Storing genesis block");

        self.add_block(genesis_block).await?;
        self.update_earliest_block_number(genesis_block_number)
            .await?;
        self.forkchoice_update(vec![], genesis_block_number, genesis_hash, None, None)
            .await?;
        Ok(())
    }

    pub async fn load_initial_state(&self) -> Result<(), StoreError> {
        info!("Loading initial state from DB");
        let Some(latest) = self.load_latest_block_number().await? else {
            return Err(StoreError::MissingLatestBlockNumber);
        };
        // Use the same durable-head clamp as the node boot path so export and the
        // running node agree on the head. The persisted head is only rewritten when
        // it actually moved, so a plain export run does not mutate `CHAIN_DATA`.
        self.anchor_to_durable_head(latest).await?;
        Ok(())
    }

    pub fn get_storage_at(
        &self,
        block_number: BlockNumber,
        address: Address,
        storage_key: H256,
    ) -> Result<Option<U256>, StoreError> {
        match self.get_block_header(block_number)? {
            Some(header) => self.get_storage_at_root(header.state_root, address, storage_key),
            None => Ok(None),
        }
    }

    pub fn get_storage_at_root(
        &self,
        state_root: H256,
        address: Address,
        storage_key: H256,
    ) -> Result<Option<U256>, StoreError> {
        let account_hash = hash_address_fixed(&address);

        // Pre-acquire shared resources once for both trie opens
        let read_view = self.backend.begin_read()?;
        let cache = self.gated_snapshot(state_root)?;
        let last_written = self.last_written()?;
        // While a deep-reorg overlay serves this root, flat-KV reads must
        // go through the trie: journal entries written while the FKV generator was
        // running lack pre-images for keys past the generator frontier, so disk
        // flat-KV may hold the generator's stale values. `TrieWrapper` already
        // forces the trie-node read path in this window; mirror the gate here so
        // the EMPTY_TRIE_HASH shortcut doesn't bypass it.
        let use_fkv = Self::flatkeyvalue_computed_with_last_written(account_hash, &last_written)
            && !cache.overlay_serves(state_root);

        let storage_root = if use_fkv {
            // We will use FKVs, we don't need the root
            EMPTY_TRIE_HASH
        } else {
            let state_trie = self.open_state_trie_shared(
                state_root,
                read_view.clone(),
                cache.clone(),
                last_written.clone(),
            )?;
            let Some(encoded_account) = state_trie.get(account_hash.as_bytes())? else {
                return Ok(None);
            };
            let account = AccountState::decode(&encoded_account)?;
            account.storage_root
        };
        let storage_trie = self.open_storage_trie_shared(
            account_hash,
            state_root,
            storage_root,
            read_view,
            cache,
            last_written,
        )?;

        let hashed_key = hash_key_fixed(&storage_key);
        storage_trie
            .get(&hashed_key)?
            .map(|rlp| U256::decode(&rlp).map_err(StoreError::RLPDecode))
            .transpose()
    }

    /// Gets storage value when the account hash and storage root are already known.
    ///
    /// This skips the state-trie account lookup and account RLP decode done by
    /// [`Self::get_storage_at_root`], and directly opens the account storage trie.
    pub fn get_storage_at_root_with_known_storage_root(
        &self,
        state_root: H256,
        account_hash: H256,
        storage_root: H256,
        storage_key: H256,
    ) -> Result<Option<U256>, StoreError> {
        let read_view = self.backend.begin_read()?;
        let cache = self.gated_snapshot(state_root)?;
        let last_written = self.last_written()?;
        // When FKV is active the real storage root is in the flatkeyvalue store,
        // not in the account's RLP-encoded storage_root field. Use EMPTY_TRIE_HASH
        // so open_storage_trie_shared falls through to the FKV path. While a
        // deep-reorg overlay serves this root, keep the real root instead:
        // disk flat-KV may hold stale generator values, so reads must go through
        // the trie (see `TrieWrapper::flatkeyvalue_computed`).
        let storage_root =
            if Self::flatkeyvalue_computed_with_last_written(account_hash, &last_written)
                && !cache.overlay_serves(state_root)
            {
                EMPTY_TRIE_HASH
            } else {
                storage_root
            };
        let storage_trie = self.open_storage_trie_shared(
            account_hash,
            state_root,
            storage_root,
            read_view,
            cache,
            last_written,
        )?;

        let hashed_key = hash_key_fixed(&storage_key);
        storage_trie
            .get(&hashed_key)?
            .map(|rlp| U256::decode(&rlp).map_err(StoreError::RLPDecode))
            .transpose()
    }

    pub fn get_chain_config(&self) -> ChainConfig {
        self.chain_config
    }

    pub async fn get_latest_canonical_block_hash(&self) -> Result<Option<BlockHash>, StoreError> {
        Ok(Some(self.latest_block_header.get().hash()))
    }

    /// Updates the canonical chain.
    /// Inserts new canonical blocks, removes blocks beyond the new head,
    /// and updates the head, safe, and finalized block pointers.
    /// All operations are performed in a single database transaction.
    pub async fn forkchoice_update(
        &self,
        new_canonical_blocks: Vec<(BlockNumber, BlockHash)>,
        head_number: BlockNumber,
        head_hash: BlockHash,
        safe: Option<BlockNumber>,
        finalized: Option<BlockNumber>,
    ) -> Result<(), StoreError> {
        // Serialize concurrent forkchoice updates. Without this, two callers
        // could interleave their `latest_block_header` cache updates with each
        // other's DB writes, leaving the cache inconsistent with the DB or
        // letting a later caller's write reorder relative to the cache update
        // order (see the TOCTOU discussion around canonical/latest drift).
        let _guard = self.fcu_lock.lock().await;

        // Updates first the latest_block_header to avoid nonce inconsistencies #3927.
        // Snapshot the previous header so we can roll the cache back if the DB
        // write fails — otherwise the cache would point at a block the DB does
        // not consider canonical.
        let previous_head = self.latest_block_header.get();
        let new_head = self
            .get_block_header_by_hash(head_hash)?
            .ok_or_else(|| StoreError::MissingLatestBlockNumber)?;
        self.latest_block_header.update(new_head);
        if let Err(err) = self
            .forkchoice_update_inner(
                new_canonical_blocks,
                head_number,
                head_hash,
                safe,
                finalized,
            )
            .await
        {
            self.latest_block_header.update((*previous_head).clone());
            return Err(err);
        }

        // Refresh the canonical safe-commit root now that the canonical tables reflect the new
        // head. `None` (chain shorter than the threshold, e.g. genesis init at head 0) leaves the
        // cell unchanged so genesis-on-disk is never gated away.
        // No `latest_block_header` rollback on error here (unlike the `inner` failure above): the
        // only error is a poisoned safe-commit `RwLock`, which is an unrecoverable process state.
        if let Some(root) = self.compute_safe_commit_root(head_number)? {
            // Advancing the cell alone does not flush: the commit step (Phase 2) only runs
            // while blocks execute, so an execute-all-then-one-forkchoice flow (e.g. block
            // import) would accumulate every layer and never persist. When the safe-commit
            // root advances, poke the worker to flush the now-committable backlog up to it.
            if self.set_safe_commit_root(root)? {
                let tx = self.persist_tx.clone();
                let _ = tokio::task::spawn_blocking(move || tx.send(PersistMessage::Commit(root)))
                    .await;
            }
        }

        Ok(())
    }

    /// Updates the dedicated safe-commit-root cell with the given state root,
    /// returning `true` if the cell changed (so callers can skip a redundant flush).
    ///
    /// This is a plain synchronous function; it touches only the dedicated cell
    /// and is disjoint from the trie-cache Arc (no cache clone or replacement).
    /// Crate-private: only `forkchoice_update` (post-canonicalization) may set it,
    /// preserving the invariant that the cell only ever holds a canonical state root.
    pub(crate) fn set_safe_commit_root(&self, root: H256) -> Result<bool, StoreError> {
        let mut guard = self
            .safe_commit_root
            .write()
            .map_err(|_| StoreError::LockError)?;
        let changed = *guard != root;
        *guard = root;
        Ok(changed)
    }

    /// Computes the canonical safe-commit state root: the state root of the canonical block
    /// `commit_threshold` layers below `head`.
    ///
    /// Returns `Ok(None)` when the chain is shorter than the threshold (underflow), or when the
    /// target block is not yet canonical / its header is absent. Synchronous getters only; no
    /// await and no lock guard held across one. The threshold is read from the trie cache to
    /// avoid duplicating the IN_MEMORY/DB selection that `from_backend` already made.
    /// Crate-private: only `forkchoice_update` consumes it.
    pub(crate) fn compute_safe_commit_root(
        &self,
        head: BlockNumber,
    ) -> Result<Option<H256>, StoreError> {
        let commit_threshold = self
            .trie_cache
            .read()
            .map_err(|_| StoreError::LockError)?
            .commit_threshold;
        let Some(target) = head.checked_sub(commit_threshold as u64) else {
            return Ok(None);
        };
        let Some(hash) = self.get_canonical_block_hash_sync(target)? else {
            return Ok(None);
        };
        let Some(header) = self.get_block_header_by_hash(hash)? else {
            return Ok(None);
        };
        Ok(Some(header.state_root))
    }

    /// Obtain the storage trie for the given block
    pub fn state_trie(&self, block_hash: BlockHash) -> Result<Option<Trie>, StoreError> {
        let Some(header) = self.get_block_header_by_hash(block_hash)? else {
            return Ok(None);
        };
        Ok(Some(self.open_state_trie(header.state_root)?))
    }

    /// Obtain the storage trie for the given account on the given block
    pub fn storage_trie(
        &self,
        block_hash: BlockHash,
        address: Address,
    ) -> Result<Option<Trie>, StoreError> {
        let Some(header) = self.get_block_header_by_hash(block_hash)? else {
            return Ok(None);
        };
        // Fetch Account from state_trie
        let Some(state_trie) = self.state_trie(block_hash)? else {
            return Ok(None);
        };
        let hashed_address = hash_address_fixed(&address);
        let Some(encoded_account) = state_trie.get(hashed_address.as_bytes())? else {
            return Ok(None);
        };
        let account = AccountState::decode(&encoded_account)?;
        // Open storage_trie
        let storage_root = account.storage_root;
        Ok(Some(self.open_storage_trie(
            hashed_address,
            header.state_root,
            storage_root,
        )?))
    }

    pub async fn get_account_state(
        &self,
        block_number: BlockNumber,
        address: Address,
    ) -> Result<Option<AccountState>, StoreError> {
        let Some(block_hash) = self.get_canonical_block_hash(block_number).await? else {
            return Ok(None);
        };
        let Some(state_trie) = self.state_trie(block_hash)? else {
            return Ok(None);
        };
        self.get_account_state_from_trie(&state_trie, address)
    }

    pub fn get_account_state_by_root(
        &self,
        state_root: H256,
        address: Address,
    ) -> Result<Option<AccountState>, StoreError> {
        let state_trie = self.open_state_trie(state_root)?;
        self.get_account_state_from_trie(&state_trie, address)
    }

    pub fn get_account_state_from_trie(
        &self,
        state_trie: &Trie,
        address: Address,
    ) -> Result<Option<AccountState>, StoreError> {
        let hashed_address = hash_address_fixed(&address);
        let Some(encoded_state) = state_trie.get(hashed_address.as_bytes())? else {
            return Ok(None);
        };
        Ok(Some(AccountState::decode(&encoded_state)?))
    }

    /// Batch lookup of account states by address against a given state root.
    ///
    /// Fast path: for addresses whose hashed path falls within the FKV cursor
    /// (and which are not present in the in-memory diff-layer cache), values
    /// are fetched in a single `multi_get` on `ACCOUNT_FLATKEYVALUE`. Other
    /// addresses fall back to per-address trie walks.
    ///
    /// Results are returned in the same order as the input addresses.
    pub fn get_account_states_batch_by_root(
        &self,
        state_root: H256,
        addresses: &[Address],
    ) -> Result<Vec<Option<AccountState>>, StoreError> {
        if addresses.is_empty() {
            return Ok(Vec::new());
        }

        let last_written = self.last_written()?;
        let trie_cache = self
            .trie_cache
            .read()
            .map_err(|_| StoreError::LockError)?
            .clone();

        let mut results: Vec<Option<AccountState>> = vec![None; addresses.len()];
        // Per-address leaf paths (nibbles + leaf flag). Length 65.
        let leaf_paths: Vec<Vec<u8>> = addresses
            .iter()
            .map(|addr| {
                let hashed = hash_address_fixed(addr);
                Nibbles::from_bytes(hashed.as_bytes()).into_vec()
            })
            .collect();

        let mut fkv_indices: Vec<usize> = Vec::new();
        let mut trie_indices: Vec<usize> = Vec::new();

        // Match `BackendTrieDB::flatkeyvalue_computed` semantics: a path is
        // covered by FKV iff `last_written >= path` as raw nibble bytes. This
        // is the same check `Trie::get` uses; the related helper
        // `Store::flatkeyvalue_computed_with_last_written` slices `[0..64]`
        // and is intentionally more conservative — using that here would
        // unnecessarily fall back to the trie when the cursor sits inside an
        // account's storage sweep (the account leaf is already in FKV at that
        // point; see `flatkeyvalue_generator`).
        //
        // While a deep-reorg overlay serves this root, skip the FKV fast
        // path entirely: this function never consults the overlay, and disk
        // flat-KV may hold values the generator computed against the chain being
        // reorged away (journal entries written during generation lack
        // past-frontier flat-KV pre-images). The per-address trie fallback goes
        // through `TrieWrapper`, which routes reads through the overlay.
        let overlay_active = trie_cache.overlay_serves(state_root);
        let fkv_cursor: &[u8] = last_written.as_slice();
        for (i, path) in leaf_paths.iter().enumerate() {
            if let Some(value) = trie_cache.get(state_root, path.as_slice()) {
                if !value.is_empty() {
                    results[i] = Some(AccountState::decode(&value)?);
                }
                continue;
            }
            if !overlay_active && fkv_cursor >= path.as_slice() {
                fkv_indices.push(i);
            } else {
                trie_indices.push(i);
            }
        }

        if !fkv_indices.is_empty() {
            // Sort by the prefixed (hashed-order) key so adjacent accounts land
            // in the same RocksDB data blocks, then read the sorted keys in
            // contiguous shards concurrently. A single `multi_get` runs the whole
            // batch at queue depth 1 (async_io is off), which on a cold account-
            // heavy block serializes thousands of random reads (coldbench: ~13x
            // slower than this sharded path). Mirrors the storage FKV batch.
            fkv_indices.sort_unstable_by(|&a, &b| leaf_paths[a].cmp(&leaf_paths[b]));
            let read_view = self.backend.begin_read()?;
            let keys: Vec<&[u8]> = fkv_indices
                .iter()
                .map(|&i| leaf_paths[i].as_slice())
                .collect();
            let shards = shard_count(keys.len());
            let raw: Vec<Result<Option<Vec<u8>>, StoreError>> = if shards <= 1 {
                read_view.multi_get(ACCOUNT_FLATKEYVALUE, &keys)
            } else {
                let chunk = keys.len().div_ceil(shards);
                let rv = read_view.as_ref();
                std::thread::scope(|scope| {
                    let handles: Vec<_> = keys
                        .chunks(chunk)
                        .map(|ck| scope.spawn(move || rv.multi_get(ACCOUNT_FLATKEYVALUE, ck)))
                        .collect();
                    handles
                        .into_iter()
                        // A panicked shard yields one Err element rather than
                        // re-panicking the whole node; the consumer's `?` below
                        // surfaces it to the caller. Every caller of this batch
                        // treats a failed warm as best-effort, so the keys are
                        // simply left uncached and read normally later.
                        .flat_map(|h| {
                            h.join().unwrap_or_else(|_| {
                                vec![Err(StoreError::Custom(
                                    "account prefetch shard panicked".into(),
                                ))]
                            })
                        })
                        .collect()
                })
            };
            for (slot, res) in fkv_indices.iter().zip(raw.into_iter()) {
                let Some(encoded) = res? else { continue };
                if encoded.is_empty() {
                    continue;
                }
                results[*slot] = Some(AccountState::decode(&encoded)?);
            }
        }

        if !trie_indices.is_empty() {
            // Fall back to the regular trie path for any addresses whose path
            // hasn't been swept by the FKV generator yet. Parallelized to
            // recover the per-address fan-out the pre-batch `par_iter` path
            // had, which matters during initial sync when most addresses
            // miss FKV.
            let state_trie = self.open_state_trie(state_root)?;
            let fetched: Result<Vec<(usize, Option<AccountState>)>, StoreError> = trie_indices
                .par_iter()
                .map(|&i| {
                    self.get_account_state_from_trie(&state_trie, addresses[i])
                        .map(|s| (i, s))
                })
                .collect();
            for (i, s) in fetched? {
                results[i] = s;
            }
        }

        Ok(results)
    }

    /// Batch lookup of storage slot values against a given state root.
    ///
    /// Each input tuple is `(account_hash, storage_root, slot)` where
    /// `account_hash` is the keccak hash of the account address and `slot` is
    /// the raw (un-hashed) storage key, matching the inputs available to
    /// [`Self::get_storage_at_root_with_known_storage_root`].
    ///
    /// Fast path: for slots whose prefixed FKV leaf path falls within the FKV
    /// cursor (and which are not present in the in-memory diff-layer cache),
    /// values are fetched in a single `multi_get` on `STORAGE_FLATKEYVALUE`.
    /// The batch is sorted by the prefixed FKV key (hashed order) before the
    /// `multi_get` so that adjacent slots share RocksDB data blocks, removing
    /// the random-order read amplification of per-slot lookups. Slots not yet
    /// swept by the FKV generator fall back to per-slot trie walks via
    /// [`Self::get_storage_at_root_with_known_storage_root`], which preserves the
    /// exact FKV / trie / diff-cache ordering of the single-slot path.
    ///
    /// Results are returned in the same order as the input slots. A missing slot
    /// (deleted or never written) yields `None`, identical to the single-get path.
    pub fn get_storage_values_batch_by_root(
        &self,
        state_root: H256,
        slots: &[(H256, H256, H256)],
    ) -> Result<Vec<Option<U256>>, StoreError> {
        if slots.is_empty() {
            return Ok(Vec::new());
        }

        let last_written = self.last_written()?;
        let trie_cache = self
            .trie_cache
            .read()
            .map_err(|_| StoreError::LockError)?
            .clone();

        let mut results: Vec<Option<U256>> = vec![None; slots.len()];
        // Per-slot prefixed FKV leaf paths, length 131: 64 account-hash nibbles
        // + 1 (account leaf flag) + 1 (separator) + 64 hashed-slot nibbles + 1
        // (storage leaf flag). This is the exact key
        // `BackendTrieDB::flatkeyvalue_computed` / `Trie::get` use for a storage
        // leaf (see `apply_prefix`).
        let leaf_paths: Vec<Vec<u8>> = slots
            .iter()
            .map(|(account_hash, _storage_root, slot)| {
                let hashed_slot = hash_key_fixed(slot);
                apply_prefix(Some(*account_hash), Nibbles::from_bytes(&hashed_slot)).into_vec()
            })
            .collect();

        let mut fkv_indices: Vec<usize> = Vec::new();
        let mut trie_indices: Vec<usize> = Vec::new();

        // Same watermark check as the account batch and `BackendTrieDB`: a path
        // is covered by FKV iff `last_written >= prefixed_path` compared as raw
        // nibble bytes. `Nibbles` orders by its inner `data` vec, so the raw
        // slice comparison is byte-identical to the `Nibbles >= Nibbles` check
        // in `BackendTrieDB::flatkeyvalue_computed`.
        //
        // While a deep-reorg overlay serves this root, skip the FKV fast path
        // entirely: this function never consults the overlay, and disk flat-KV may
        // hold values the generator computed against the chain being reorged away
        // (journal entries written during generation lack past-frontier flat-KV
        // pre-images). The per-slot trie fallback reads through `TrieWrapper`, which
        // routes through the overlay. Mirrors the gate in
        // `get_storage_at_root` and `get_account_states_batch_by_root`.
        let overlay_active = trie_cache.overlay_serves(state_root);
        let fkv_cursor: &[u8] = last_written.as_slice();
        for (i, path) in leaf_paths.iter().enumerate() {
            // Diff-layer cache holds the authoritative latest value for this
            // key at this state root, overriding both FKV and trie. Mirror the
            // cache-first semantics of `TrieWrapper::get` / the account batch:
            // an empty cached value means the slot was deleted -> None.
            if let Some(value) = trie_cache.get(state_root, path.as_slice()) {
                if !value.is_empty() {
                    results[i] = Some(U256::decode(&value).map_err(StoreError::RLPDecode)?);
                }
                continue;
            }
            if !overlay_active && fkv_cursor >= path.as_slice() {
                fkv_indices.push(i);
            } else {
                trie_indices.push(i);
            }
        }

        if !fkv_indices.is_empty() {
            // Sort the FKV batch by the prefixed (hashed-order) key so adjacent
            // slots land in the same 4KB RocksDB data blocks. This is the whole
            // point of the batch: it removes the random-order amplification of
            // issuing one point lookup per slot in execution order.
            fkv_indices.sort_unstable_by(|&a, &b| leaf_paths[a].cmp(&leaf_paths[b]));
            let read_view = self.backend.begin_read()?;
            let keys: Vec<&[u8]> = fkv_indices
                .iter()
                .map(|&i| leaf_paths[i].as_slice())
                .collect();
            // Sharded parallel multi_get. A single `multi_get` runs the whole
            // batch serially inside RocksDB (async_io is OFF in our build, so
            // no io_uring => queue depth ~= 1), which on a cold NVMe is purely
            // read-latency-bound: it touches few blocks (sorted) but serializes
            // them. Splitting the SORTED keys into contiguous shards read by
            // separate blocking threads keeps the per-shard block sharing while
            // issuing the shards concurrently, so effective queue depth ~= shard
            // count. On a bloated-account cold read this is ~4.7x a single
            // multi_get and ~4.6x the per-slot `par_iter` baseline (coldbench).
            // Shard count scales with batch size and is capped; small batches
            // run a single in-line multi_get (no thread spawn). All shards share
            // one read view (one snapshot) so the reads stay consistent.
            let shards = shard_count(keys.len());
            let raw: Vec<Result<Option<Vec<u8>>, StoreError>> = if shards <= 1 {
                read_view.multi_get(STORAGE_FLATKEYVALUE, &keys)
            } else {
                let chunk = keys.len().div_ceil(shards);
                let rv = read_view.as_ref();
                std::thread::scope(|scope| {
                    let handles: Vec<_> = keys
                        .chunks(chunk)
                        .map(|ck| scope.spawn(move || rv.multi_get(STORAGE_FLATKEYVALUE, ck)))
                        .collect();
                    handles
                        .into_iter()
                        // A panicked shard yields one Err element rather than
                        // re-panicking the whole node; the consumer's `?` below
                        // surfaces it to the caller. Every caller of this batch
                        // treats a failed warm as best-effort, so the keys are
                        // simply left uncached and read normally later.
                        .flat_map(|h| {
                            h.join().unwrap_or_else(|_| {
                                vec![Err(StoreError::Custom(
                                    "storage prefetch shard panicked".into(),
                                ))]
                            })
                        })
                        .collect()
                })
            };
            for (slot_idx, res) in fkv_indices.iter().zip(raw.into_iter()) {
                let Some(encoded) = res? else { continue };
                if encoded.is_empty() {
                    continue;
                }
                results[*slot_idx] = Some(U256::decode(&encoded).map_err(StoreError::RLPDecode)?);
            }
        }

        if !trie_indices.is_empty() {
            // Fall back to the regular single-slot path for any slot whose path
            // hasn't been swept by the FKV generator yet. Reusing
            // `get_storage_at_root_with_known_storage_root` guarantees the
            // trie-walk fallback is byte-identical to the unbatched read,
            // including its own per-slot FKV / EMPTY_TRIE_HASH handling.
            // Parallelized to recover the per-slot fan-out the pre-batch
            // `par_iter` prefetch path had.
            let fetched: Result<Vec<(usize, Option<U256>)>, StoreError> = trie_indices
                .par_iter()
                .map(|&i| {
                    let (account_hash, storage_root, slot) = slots[i];
                    self.get_storage_at_root_with_known_storage_root(
                        state_root,
                        account_hash,
                        storage_root,
                        slot,
                    )
                    .map(|v| (i, v))
                })
                .collect();
            for (i, v) in fetched? {
                results[i] = v;
            }
        }

        Ok(results)
    }

    /// Best-effort warming of the MPT *internal* nodes the merkleizer will walk,
    /// driven entirely by the BAL working set. The merkle phase reads trie nodes
    /// from `STORAGE_TRIE_NODES` / `ACCOUNT_TRIE_NODES` (see `BackendTrieDB::get`)
    /// by nibble path; the existing flat-KV prefetch warms only the leaf-value
    /// CFs that *execution* reads, so every merkle-walk node read is otherwise
    /// cold. This pulls those node blocks into the page/block cache up front so
    /// the subsequent walk runs warm.
    ///
    /// The keys are not discovered by walking (a walk is QD1 and serial): the
    /// internal nodes on the path to a leaf are all at *prefixes* of that leaf's
    /// nibble path, so for every BAL leaf we speculatively probe every prefix up
    /// to `MAX_PREFETCH_DEPTH`. Prefixes that are not real nodes are rejected by
    /// the CF bloom filter in RAM (no disk read; measured ~free), and the shallow
    /// prefixes collapse on dedup. This turns the cold merkle node reads into one
    /// sorted, sharded, high-queue-depth `multi_get` per CF — the same lever that
    /// fixed the flat-KV `sload` path, applied to the trie-node CFs.
    ///
    /// Results are discarded. Missing a node (e.g. one deeper than the probe
    /// band) is harmless: the merkleizer just cold-reads it. Run this
    /// synchronously up front (mirroring the flat-KV prefetch) so it never races
    /// the merkleizer; it touches different CFs than execution, so it does not
    /// contend with the executor either.
    /// Returns the number of probed keys that hit a real stored node (warmed).
    /// On the warm/no-op path this still reports coverage; a near-zero hit count
    /// against a non-empty working set signals a key-encoding mismatch (the
    /// effectiveness test asserts on this).
    pub fn prefetch_trie_nodes(
        &self,
        storage_slots: &[(Address, H256)],
        accounts: &[Address],
    ) -> Result<usize, StoreError> {
        // Probe band: nodes live at depth <= ~ceil(log16(trie_size)). 8 nibbles
        // covers tries up to ~16^8 entries near-fully; any deeper node is simply
        // cold-read by the walk, so under-covering only loses a little warming,
        // never correctness. Tunable.
        const MAX_PREFETCH_DEPTH: usize = 8;

        // STORAGE_TRIE_NODES: prefixes of apply_prefix(hashed_addr, hashed_slot).
        // Each input pushes MAX_PREFETCH_DEPTH+1 prefixes (depths 0..=max_d).
        let mut storage_keys: Vec<Vec<u8>> =
            Vec::with_capacity(storage_slots.len() * (MAX_PREFETCH_DEPTH + 1));
        for (address, slot) in storage_slots {
            let hashed_address = hash_address_fixed(address);
            let hashed_slot = hash_key_fixed(slot);
            let slot_nibbles = Nibbles::from_bytes(&hashed_slot);
            let max_d = MAX_PREFETCH_DEPTH.min(slot_nibbles.len());
            for d in 0..=max_d {
                let prefix = apply_prefix(Some(hashed_address), slot_nibbles.slice(0, d));
                storage_keys.push(prefix.into_vec());
            }
        }

        // ACCOUNT_TRIE_NODES: prefixes of the hashed address (no account prefix).
        let mut account_keys: Vec<Vec<u8>> =
            Vec::with_capacity(accounts.len() * (MAX_PREFETCH_DEPTH + 1));
        for address in accounts {
            let hashed_address = hash_address_fixed(address);
            let addr_nibbles = Nibbles::from_bytes(hashed_address.as_bytes());
            let max_d = MAX_PREFETCH_DEPTH.min(addr_nibbles.len());
            for d in 0..=max_d {
                account_keys.push(addr_nibbles.slice(0, d).into_vec());
            }
        }

        let hits = self.warm_trie_node_keys(STORAGE_TRIE_NODES, storage_keys)?
            + self.warm_trie_node_keys(ACCOUNT_TRIE_NODES, account_keys)?;
        Ok(hits)
    }

    /// Sorted, sharded, high-queue-depth `multi_get` over `keys` in `table`,
    /// discarding results. Sorting makes adjacent keys share RocksDB data blocks;
    /// sharding across blocking threads issues the reads at queue depth ~= shard
    /// count (async_io is off, so a single multi_get runs at QD1). Mirrors the
    /// storage flat-KV batch path. Dedups first so shared shallow prefixes are
    /// read once.
    fn warm_trie_node_keys(
        &self,
        table: &'static str,
        mut keys: Vec<Vec<u8>>,
    ) -> Result<usize, StoreError> {
        if keys.is_empty() {
            return Ok(0);
        }
        keys.sort_unstable();
        keys.dedup();

        let shards = shard_count(keys.len());
        let read_view = self.backend.begin_read()?;
        let count_hits = |res: Vec<Result<Option<Vec<u8>>, StoreError>>| {
            res.into_iter().filter(|r| matches!(r, Ok(Some(_)))).count()
        };
        if shards <= 1 {
            let refs: Vec<&[u8]> = keys.iter().map(|k| k.as_slice()).collect();
            return Ok(count_hits(read_view.multi_get(table, &refs)));
        }
        let chunk = keys.len().div_ceil(shards);
        let rv = read_view.as_ref();
        let hits = std::thread::scope(|scope| {
            let handles: Vec<_> = keys
                .chunks(chunk)
                .map(|ck| {
                    scope.spawn(move || {
                        let refs: Vec<&[u8]> = ck.iter().map(|k| k.as_slice()).collect();
                        // Warm-only: keep cache population, just tally the hits.
                        count_hits(rv.multi_get(table, &refs))
                    })
                })
                .collect();
            handles
                .into_iter()
                // Best-effort warming: a panicked shard contributes 0 hits rather
                // than re-panicking. The caller treats this prefetch as best-effort
                // (missing nodes are just cold-read by the walk), so a shard hiccup
                // must not unwind block processing.
                .map(|h| h.join().unwrap_or(0))
                .sum::<usize>()
        });
        Ok(hits)
    }

    /// Constructs a merkle proof for the given account address against a given state.
    /// If storage_keys are provided, also constructs the storage proofs for those keys.
    ///
    /// Returns `None` if the state trie is missing, otherwise returns the proof.
    pub async fn get_account_proof(
        &self,
        state_root: H256,
        address: Address,
        storage_keys: &[H256],
    ) -> Result<Option<AccountProof>, StoreError> {
        // TODO: check state root
        // let Some(state_trie) = self.open_state_trie(state_trie)? else {
        //     return Ok(None);
        // };
        let state_trie = self.open_state_trie(state_root)?;
        let address_path = hash_address_fixed(&address);
        let proof = state_trie.get_proof(address_path.as_bytes())?;
        let account_opt = state_trie
            .get(address_path.as_bytes())?
            .map(|encoded_state| AccountState::decode(&encoded_state))
            .transpose()?;

        let mut storage_proof = Vec::with_capacity(storage_keys.len());

        if let Some(account) = &account_opt {
            let storage_trie =
                self.open_storage_trie(address_path, state_root, account.storage_root)?;

            for key in storage_keys {
                let hashed_key = hash_key(key);
                let proof = storage_trie.get_proof(&hashed_key)?;
                let value = storage_trie
                    .get(&hashed_key)?
                    .map(|rlp| U256::decode(&rlp).map_err(StoreError::RLPDecode))
                    .transpose()?
                    .unwrap_or_default();

                let slot_proof = StorageSlotProof {
                    proof,
                    key: *key,
                    value,
                };
                storage_proof.push(slot_proof);
            }
        } else {
            storage_proof.extend(storage_keys.iter().map(|key| StorageSlotProof {
                proof: Vec::new(),
                key: *key,
                value: U256::zero(),
            }));
        }
        let account = account_opt.unwrap_or_default();
        let account_proof = AccountProof {
            proof,
            account,
            storage_proof,
        };
        Ok(Some(account_proof))
    }

    // Returns an iterator across all accounts in the state trie given by the state_root
    // Does not check that the state_root is valid
    pub fn iter_accounts_from(
        &self,
        state_root: H256,
        starting_address: H256,
    ) -> Result<impl Iterator<Item = (H256, AccountState)>, StoreError> {
        let mut iter = self.open_locked_state_trie(state_root)?.into_iter();
        iter.advance(starting_address.0.to_vec())?;
        Ok(iter.content().map_while(|(path, value)| {
            Some((H256::from_slice(&path), AccountState::decode(&value).ok()?))
        }))
    }

    // Returns an iterator across all accounts in the state trie given by the state_root
    // Does not check that the state_root is valid
    pub fn iter_accounts(
        &self,
        state_root: H256,
    ) -> Result<impl Iterator<Item = (H256, AccountState)>, StoreError> {
        self.iter_accounts_from(state_root, H256::zero())
    }

    // Returns an iterator across all accounts in the state trie given by the state_root
    // Does not check that the state_root is valid
    pub fn iter_storage_from(
        &self,
        state_root: H256,
        hashed_address: H256,
        starting_slot: H256,
    ) -> Result<Option<impl Iterator<Item = (H256, U256)>>, StoreError> {
        let state_trie = self.open_locked_state_trie(state_root)?;
        let Some(account_rlp) = state_trie.get(hashed_address.as_bytes())? else {
            return Ok(None);
        };
        let storage_root = AccountState::decode(&account_rlp)?.storage_root;
        let mut iter = self
            .open_locked_storage_trie(hashed_address, state_root, storage_root)?
            .into_iter();
        iter.advance(starting_slot.0.to_vec())?;
        Ok(Some(iter.content().map_while(|(path, value)| {
            Some((H256::from_slice(&path), U256::decode(&value).ok()?))
        })))
    }

    // Returns an iterator across all accounts in the state trie given by the state_root
    // Does not check that the state_root is valid
    pub fn iter_storage(
        &self,
        state_root: H256,
        hashed_address: H256,
    ) -> Result<Option<impl Iterator<Item = (H256, U256)>>, StoreError> {
        self.iter_storage_from(state_root, hashed_address, H256::zero())
    }

    pub fn get_account_range_proof(
        &self,
        state_root: H256,
        starting_hash: H256,
        last_hash: Option<H256>,
    ) -> Result<Vec<Vec<u8>>, StoreError> {
        let state_trie = self.open_state_trie(state_root)?;
        let mut proof = state_trie.get_proof(starting_hash.as_bytes())?;
        if let Some(last_hash) = last_hash {
            proof.extend_from_slice(&state_trie.get_proof(last_hash.as_bytes())?);
        }
        Ok(proof)
    }

    pub fn get_storage_range_proof(
        &self,
        state_root: H256,
        hashed_address: H256,
        starting_hash: H256,
        last_hash: Option<H256>,
    ) -> Result<Option<Vec<Vec<u8>>>, StoreError> {
        let state_trie = self.open_state_trie(state_root)?;
        let Some(account_rlp) = state_trie.get(hashed_address.as_bytes())? else {
            return Ok(None);
        };
        let storage_root = AccountState::decode(&account_rlp)?.storage_root;
        let storage_trie = self.open_storage_trie(hashed_address, state_root, storage_root)?;
        let mut proof = storage_trie.get_proof(starting_hash.as_bytes())?;
        if let Some(last_hash) = last_hash {
            proof.extend_from_slice(&storage_trie.get_proof(last_hash.as_bytes())?);
        }
        Ok(Some(proof))
    }

    /// Receives the root of the state trie and a list of paths where the first path will correspond to a path in the state trie
    /// (aka a hashed account address) and the following paths will be paths in the account's storage trie (aka hashed storage keys)
    /// If only one hash (account) is received, then the state trie node containing the account will be returned.
    /// If more than one hash is received, then the storage trie nodes where each storage key is stored will be returned
    /// For more information check out snap capability message [`GetTrieNodes`](https://github.com/ethereum/devp2p/blob/master/caps/snap.md#gettrienodes-0x06)
    /// The paths can be either full paths (hash) or partial paths (compact-encoded nibbles), if a partial path is given for the account this method will not return storage nodes for it
    pub fn get_trie_nodes(
        &self,
        state_root: H256,
        paths: Vec<Vec<u8>>,
        byte_limit: u64,
    ) -> Result<Vec<Vec<u8>>, StoreError> {
        let Some(account_path) = paths.first() else {
            return Ok(vec![]);
        };
        let state_trie = self.open_state_trie(state_root)?;
        // State Trie Nodes Request
        if paths.len() == 1 {
            // Fetch state trie node
            let node = state_trie.get_node(account_path)?;
            return Ok(vec![node]);
        }
        // Storage Trie Nodes Request
        let Some(account_state) = state_trie
            .get(account_path)?
            .map(|ref rlp| AccountState::decode(rlp))
            .transpose()?
        else {
            return Ok(vec![]);
        };
        // We can't access the storage trie without the account's address hash
        let Ok(hashed_address) = account_path.clone().try_into().map(H256) else {
            return Ok(vec![]);
        };
        let storage_trie =
            self.open_storage_trie(hashed_address, state_root, account_state.storage_root)?;
        // Fetch storage trie nodes
        let mut nodes = vec![];
        let mut bytes_used = 0;
        // The number of sub-paths is bounded upstream by the snap server's per-request lookup
        // cap (`MAX_SERVE_LOOKUPS`), which truncates the pathset before this call; here the
        // byte budget only bounds response size.
        for path in paths.iter().skip(1) {
            if bytes_used >= byte_limit {
                break;
            }
            let node = storage_trie.get_node(path)?;
            bytes_used += node.len() as u64;
            nodes.push(node);
        }
        Ok(nodes)
    }

    // Methods exclusive for trie management during snap-syncing

    /// Snapshot the trie layer cache for reading at `state_root`, blocking until
    /// that root's diff layer has been installed if it is still in-flight (see
    /// [`PendingTrieRoots`]). This is the read barrier for deferred layer builds:
    /// taking it at trie-open time guarantees the snapshot contains the layer, so
    /// a just-added block's state is never read as stale. Roots that are not
    /// pending (already installed, historical/committed, genesis) never block.
    fn gated_snapshot(&self, state_root: H256) -> Result<Arc<TrieLayerCache>, StoreError> {
        self.pending_trie_roots.wait_until_ready(state_root)?;
        Ok(self
            .trie_cache
            .read()
            .map_err(|_| StoreError::LockError)?
            .clone())
    }

    /// Obtain a state trie from the given state root
    /// Doesn't check if the state root is valid
    /// Used for internal store operations
    pub fn open_state_trie(&self, state_root: H256) -> Result<Trie, StoreError> {
        let trie_db = TrieWrapper::new(
            state_root,
            self.gated_snapshot(state_root)?,
            Box::new(BackendTrieDB::new_for_accounts(
                self.backend.clone(),
                self.last_written()?,
            )?),
            None,
        );
        Ok(Trie::open(Box::new(trie_db), state_root))
    }

    /// Obtain a state trie from the given state root
    /// Doesn't check if the state root is valid
    /// Used for internal store operations
    pub fn open_direct_state_trie(&self, state_root: H256) -> Result<Trie, StoreError> {
        Ok(Trie::open(
            Box::new(BackendTrieDB::new_for_accounts(
                self.backend.clone(),
                self.last_written()?,
            )?),
            state_root,
        ))
    }

    /// Obtain a state trie locked for reads from the given state root
    /// Doesn't check if the state root is valid
    /// Used for internal store operations
    pub fn open_locked_state_trie(&self, state_root: H256) -> Result<Trie, StoreError> {
        let trie_db = TrieWrapper::new(
            state_root,
            self.gated_snapshot(state_root)?,
            Box::new(state_trie_locked_backend(
                self.backend.as_ref(),
                self.last_written()?,
            )?),
            None,
        );
        Ok(Trie::open(Box::new(trie_db), state_root))
    }

    /// Obtain a storage trie from the given address and storage_root.
    /// Doesn't check if the account is stored
    pub fn open_storage_trie(
        &self,
        account_hash: H256,
        state_root: H256,
        storage_root: H256,
    ) -> Result<Trie, StoreError> {
        let trie_db = TrieWrapper::new(
            state_root,
            self.gated_snapshot(state_root)?,
            Box::new(BackendTrieDB::new_for_storages(
                self.backend.clone(),
                self.last_written()?,
            )?),
            Some(account_hash),
        );
        Ok(Trie::open(Box::new(trie_db), storage_root))
    }

    /// Open a state trie using pre-acquired shared resources.
    /// Avoids redundant RwLock acquisitions when multiple tries are opened
    /// in the same operation (e.g., state trie + storage trie in get_storage_at_root).
    fn open_state_trie_shared(
        &self,
        state_root: H256,
        read_view: Arc<dyn StorageReadView>,
        cache: Arc<TrieLayerCache>,
        last_written: Vec<u8>,
    ) -> Result<Trie, StoreError> {
        let trie_db = TrieWrapper::new(
            state_root,
            cache,
            Box::new(BackendTrieDB::new_for_accounts_with_view(
                self.backend.clone(),
                read_view,
                last_written,
            )?),
            None,
        );
        Ok(Trie::open(Box::new(trie_db), state_root))
    }

    /// Open a storage trie using pre-acquired shared resources.
    fn open_storage_trie_shared(
        &self,
        account_hash: H256,
        state_root: H256,
        storage_root: H256,
        read_view: Arc<dyn StorageReadView>,
        cache: Arc<TrieLayerCache>,
        last_written: Vec<u8>,
    ) -> Result<Trie, StoreError> {
        let trie_db = TrieWrapper::new(
            state_root,
            cache,
            Box::new(BackendTrieDB::new_for_storages_with_view(
                self.backend.clone(),
                read_view,
                last_written,
            )?),
            Some(account_hash),
        );
        Ok(Trie::open(Box::new(trie_db), storage_root))
    }

    /// Obtain a storage trie from the given address and storage_root.
    /// Doesn't check if the account is stored
    pub fn open_direct_storage_trie(
        &self,
        account_hash: H256,
        storage_root: H256,
    ) -> Result<Trie, StoreError> {
        Ok(Trie::open(
            Box::new(BackendTrieDB::new_for_account_storage(
                self.backend.clone(),
                account_hash,
                self.last_written()?,
            )?),
            storage_root,
        ))
    }

    /// Obtain a read-locked storage trie from the given address and storage_root.
    /// Doesn't check if the account is stored
    pub fn open_locked_storage_trie(
        &self,
        account_hash: H256,
        state_root: H256,
        storage_root: H256,
    ) -> Result<Trie, StoreError> {
        let trie_db = TrieWrapper::new(
            state_root,
            self.gated_snapshot(state_root)?,
            Box::new(state_trie_locked_backend(
                self.backend.as_ref(),
                self.last_written()?,
            )?),
            Some(account_hash),
        );
        Ok(Trie::open(Box::new(trie_db), storage_root))
    }

    pub fn has_state_root(&self, state_root: H256) -> Result<bool, StoreError> {
        // Empty state trie is always available
        if state_root == EMPTY_TRIE_HASH {
            return Ok(true);
        }
        let trie = self.open_state_trie(state_root)?;
        // NOTE: here we hash the root because the trie doesn't check the state root is correct
        let Some(root) = trie.db().get(Nibbles::default())? else {
            return Ok(false);
        };
        let root_hash = ethrex_trie::Node::decode(&root)?
            .compute_hash(&NativeCrypto)
            .finalize(&NativeCrypto);
        Ok(state_root == root_hash)
    }

    // ===========================================================================
    // Deep-reorg primitives (storage side).
    // ===========================================================================

    /// Returns `true` if the in-memory layer cache currently has a layer with
    /// the given `state_root`. Used by the deep-reorg orchestrator to decide
    /// whether the head's state can be reached through forward execution (cache
    /// hit) or whether a deep-reorg path with overlay construction is required.
    pub fn is_state_in_layer_cache(&self, state_root: H256) -> Result<bool, StoreError> {
        let trie = self
            .trie_cache
            .read()
            .map_err(|_| StoreError::LockError)?
            .clone();
        Ok(trie.contains(state_root))
    }

    /// Returns the highest block number with a `STATE_HISTORY` entry; the cache
    /// edge `D` (the deepest block whose post-state is on disk). Returns `None`
    /// if the journal is empty (no commits since boot, or fully pruned by finality).
    ///
    /// O(1) via reverse seek on the column family's last key.
    pub fn highest_state_history_block_number(&self) -> Result<Option<BlockNumber>, StoreError> {
        let read = self.backend.begin_read()?;
        let Some(key) = read.last_key(STATE_HISTORY)? else {
            return Ok(None);
        };
        let arr = <[u8; 8]>::try_from(key.as_slice()).map_err(|_| {
            StoreError::Custom(format!(
                "STATE_HISTORY key has unexpected length: {}",
                key.len()
            ))
        })?;
        Ok(Some(BlockNumber::from_be_bytes(arr)))
    }

    /// Returns the hash of the block whose trie-layer commit the `STATE_HISTORY`
    /// entry at `block_number` journals, or `None` when there is no entry.
    ///
    /// The persist worker stages one entry per committed layer, so the entry at
    /// [`Self::highest_state_history_block_number`] identifies the block whose
    /// post-state is the one on disk.
    pub fn get_state_history_block_hash(
        &self,
        block_number: BlockNumber,
    ) -> Result<Option<BlockHash>, StoreError> {
        let read = self.backend.begin_read()?;
        let Some(bytes) = read.get(STATE_HISTORY, &block_number.to_be_bytes())? else {
            return Ok(None);
        };
        JournalEntry::decode_block_hash(&bytes)
            .map(Some)
            .map_err(|e| StoreError::Custom(format!("STATE_HISTORY entry {block_number}: {e}")))
    }

    /// Returns the lowest block number with a `STATE_HISTORY` entry. Returns `None`
    /// if the journal is empty (no commits since boot, or fully pruned by finality).
    ///
    /// O(1) via forward seek on the column family's first key.
    pub fn lowest_state_history_block_number(&self) -> Result<Option<BlockNumber>, StoreError> {
        let read = self.backend.begin_read()?;
        let Some(key) = read.first_key(STATE_HISTORY)? else {
            return Ok(None);
        };
        let arr = <[u8; 8]>::try_from(key.as_slice()).map_err(|_| {
            StoreError::Custom(format!(
                "STATE_HISTORY key has unexpected length: {}",
                key.len()
            ))
        })?;
        Ok(Some(BlockNumber::from_be_bytes(arr)))
    }

    /// The in-memory diff-layer retention (the layer cache's commit threshold):
    /// the deepest reorg the node can serve straight from the layer cache, with no
    /// journal/overlay reconstruction. RocksDB default 128, in-memory 10000. Used by
    /// `compute_reorg_ceiling` as the physical floor when there is no finality signal.
    pub fn reorg_retention(&self) -> Result<u64, StoreError> {
        let cache = self.trie_cache.read().map_err(|_| StoreError::LockError)?;
        Ok(cache.commit_threshold() as u64)
    }

    /// Test-only: inserts a pre-encoded `STATE_HISTORY` entry at the given block
    /// number. Lets integration tests seed the journal without running enough
    /// commits to trip the in-memory cache's flush threshold.
    #[doc(hidden)]
    pub fn put_state_history_entry_for_test(
        &self,
        block_number: BlockNumber,
        encoded: &[u8],
    ) -> Result<(), StoreError> {
        let mut tx = self.backend.begin_write()?;
        tx.put(STATE_HISTORY, &block_number.to_be_bytes(), encoded)?;
        tx.commit()?;
        Ok(())
    }

    /// Atomically prepares the store for a deep-reorg apply pass.
    ///
    /// Builds an [`Overlay`] from journal entries in `[to_block, from_block]`
    /// (inclusive both ends), verifies each entry's `block_hash` against
    /// `expected_hash`, then swaps the in-memory layer cache for a fresh one
    /// with the overlay installed. After this call:
    ///
    /// - The layer cache contains zero forward layers.
    /// - The overlay is in place; subsequent `TrieWrapper::get` calls cascade
    ///   layer cache -> overlay -> disk.
    /// - The on-disk trie/flat-KV state is unchanged (still at the OLD chain's
    ///   edge `D`).
    ///
    /// Side-chain blocks `[pivot+1 .. new_head]` should now be executed in
    /// chain order through the normal `Blockchain::add_block` path; each
    /// block's reads cascade through the overlay, and each block's commit
    /// produces a new forward layer.
    ///
    /// Errors abort the swap: if overlay construction fails (missing entry,
    /// hash mismatch, decode error), the existing layer cache is left intact.
    pub fn install_overlay_for_reorg(
        &self,
        from_block: BlockNumber,
        to_block: BlockNumber,
        expected_hash: impl Fn(BlockNumber) -> Option<H256>,
    ) -> Result<(), StoreError> {
        // Build the overlay first so any failure aborts before mutating the cache.
        let overlay =
            Overlay::from_journal(self.backend.as_ref(), from_block, to_block, expected_hash)
                .map_err(|e| StoreError::Custom(format!("overlay construction failed: {e}")))?;

        let threshold = {
            let current = self.trie_cache.read().map_err(|_| StoreError::LockError)?;
            current.commit_threshold()
        };
        // Share the Store's safe-commit-root cell so the canonical commit gate keeps
        // working after the cache swap (the cell only advances via forkchoice).
        let mut fresh =
            TrieLayerCache::new_with_safe_commit(threshold, self.safe_commit_root.clone());
        fresh.set_overlay(Arc::new(overlay));

        // Wait for the persist worker to be idle before swapping the cache (see
        // [`Self::rendezvous_persist_worker`]).
        self.rendezvous_persist_worker("install_overlay_for_reorg")?;

        let mut guard = self.trie_cache.write().map_err(|_| StoreError::LockError)?;
        *guard = Arc::new(fresh);
        Ok(())
    }

    /// Waits for the persist worker to be idle before a layer-cache swap. That
    /// worker owns the trie-layer install (`apply_trie_phase1`, run from the
    /// `PersistMessage::Block` handler): it reads `trie_cache`, mutates a local
    /// clone, and RCU-writes it back; if it is mid-flight when we swap, its
    /// write-back can clobber the freshly swapped cache (e.g. drop a just-installed
    /// overlay, or install a side-chain layer over a base that no longer has the
    /// overlay underneath). `PersistMessage::Ping` carries an ack channel and the
    /// worker is FIFO, so its ack proves every earlier `Block` (and thus every
    /// earlier trie install) is fully processed and the worker is back at
    /// `rx.recv()`. This is the synchronous core of [`wait_for_persistence_idle`];
    /// we inline it here because callers are not async. The caller's subsequent
    /// `trie_cache.write()` serialising any future RCU makes the swap safe.
    fn rendezvous_persist_worker(&self, caller: &str) -> Result<(), StoreError> {
        let (ack_tx, ack_rx) = sync_channel::<Result<(), StoreError>>(1);
        self.persist_tx
            .send(PersistMessage::Ping(ack_tx))
            .map_err(|e| {
                StoreError::Custom(format!("{caller}: failed to ping persist worker: {e}"))
            })?;
        ack_rx.recv().map_err(|e| {
            StoreError::Custom(format!("{caller}: persist worker ping ack failed: {e}"))
        })??;
        Ok(())
    }

    /// Returns `(entry_count, byte_size)` of the currently installed overlay, or
    /// `(0, 0)` when no overlay is installed. Used by the observability layer in
    /// `fork_choice.rs` immediately after `install_overlay_for_reorg` to emit
    /// `ethrex_reorg_overlay_entries` / `ethrex_reorg_overlay_bytes`.
    pub fn reorg_overlay_size_hint(&self) -> Result<(usize, usize), StoreError> {
        let guard = self.trie_cache.read().map_err(|_| StoreError::LockError)?;
        match guard.overlay() {
            Some(ov) => Ok((ov.len(), ov.byte_size())),
            None => Ok((0, 0)),
        }
    }

    /// Pauses or resumes STATE_HISTORY pruning at finality advance (see the
    /// `journal_pruning_paused` field). Called by `Blockchain::enter_reorg` /
    /// `ReorgGuard::drop` to bracket a deep-reorg apply pass.
    pub fn set_journal_pruning_paused(&self, paused: bool) {
        self.journal_pruning_paused
            .store(paused, std::sync::atomic::Ordering::Release);
    }

    /// Removes any installed overlay from the layer cache. Called by the
    /// reconciliation path after the first new-chain commit folds
    /// the overlay into disk. Idempotent.
    pub fn clear_reorg_overlay(&self) -> Result<(), StoreError> {
        let mut guard = self.trie_cache.write().map_err(|_| StoreError::LockError)?;
        let mut updated = (**guard).clone();
        updated.clear_overlay();
        *guard = Arc::new(updated);
        Ok(())
    }

    /// Aborts an in-progress deep reorg and resets the layer cache to a fresh
    /// empty state with the same commit threshold. Both the overlay and any
    /// partially-built new-chain layers are discarded. On-disk state is
    /// untouched (still at the OLD chain's `D`), so subsequent FCU evaluations
    /// start from a clean foundation.
    pub fn abort_reorg(&self) -> Result<(), StoreError> {
        // Rendezvous with the persist worker before swapping, exactly like
        // `install_overlay_for_reorg`: the live-path ack fires BEFORE
        // `apply_trie_phase1` installs the layer, so the worker can still be
        // mid-flight with a pre-abort RCU snapshot. Without the rendezvous its
        // write-back could install a side-chain layer into the fresh cache with
        // no overlay underneath — reads at that root would then cascade into
        // old-chain disk state.
        self.rendezvous_persist_worker("abort_reorg")?;
        let mut guard = self.trie_cache.write().map_err(|_| StoreError::LockError)?;
        let threshold = guard.commit_threshold();
        *guard = Arc::new(TrieLayerCache::new_with_safe_commit(
            threshold,
            self.safe_commit_root.clone(),
        ));
        Ok(())
    }

    /// Takes a block hash and returns an iterator to its ancestors. Block headers are returned
    /// in reverse order, starting from the given block and going up to the genesis block.
    pub fn ancestors(&self, block_hash: BlockHash) -> AncestorIterator {
        AncestorIterator {
            store: self.clone(),
            next_hash: block_hash,
        }
    }

    /// Checks if a given block belongs to the current canonical chain. Returns false if the block is not known
    pub fn is_canonical_sync(&self, block_hash: BlockHash) -> Result<bool, StoreError> {
        let Some(block_number) = self.get_block_number_sync(block_hash)? else {
            return Ok(false);
        };
        Ok(self
            .get_canonical_block_hash_sync(block_number)?
            .is_some_and(|h| h == block_hash))
    }

    pub fn generate_flatkeyvalue(&self) -> Result<(), StoreError> {
        self.flatkeyvalue_control_tx
            .send(FKVGeneratorControlMessage::Continue)
            .map_err(|_| StoreError::Custom("FlatKeyValue thread disconnected.".to_string()))
    }

    pub fn create_checkpoint(&self, path: impl AsRef<Path>) -> Result<(), StoreError> {
        self.backend.create_checkpoint(path.as_ref())?;
        init_metadata_file(path.as_ref())?;
        Ok(())
    }

    pub fn get_store_directory(&self) -> Result<PathBuf, StoreError> {
        Ok(self.db_path.clone())
    }

    /// Loads the latest block number stored in the database, bypassing the latest block number cache
    async fn load_latest_block_number(&self) -> Result<Option<BlockNumber>, StoreError> {
        self.read_chain_data_block_number(ChainDataIndex::LatestBlockNumber)
            .await
    }

    fn load_canonical_block_hash(
        &self,
        block_number: BlockNumber,
    ) -> Result<Option<BlockHash>, StoreError> {
        let txn = self.backend.begin_read()?;
        txn.get(
            CANONICAL_BLOCK_HASHES,
            block_number.to_le_bytes().as_slice(),
        )?
        .map(|bytes| H256::decode(bytes.as_slice()))
        .transpose()
        .map_err(StoreError::from)
    }

    fn load_block_header(
        &self,
        block_number: BlockNumber,
    ) -> Result<Option<BlockHeader>, StoreError> {
        let Some(block_hash) = self.load_canonical_block_hash(block_number)? else {
            return Ok(None);
        };
        self.load_block_header_by_hash(block_hash)
    }

    /// Load a block header, bypassing the latest header cache
    fn load_block_header_by_hash(
        &self,
        block_hash: BlockHash,
    ) -> Result<Option<BlockHeader>, StoreError> {
        let txn = self.backend.begin_read()?;
        let hash_key = block_hash.encode_to_vec();
        let header_value = txn.get(HEADERS, hash_key.as_slice())?;
        let mut header = header_value
            .map(|bytes| BlockHeaderRLP::from_bytes(bytes).to())
            .transpose()
            .map_err(StoreError::from)?;
        header.as_mut().inspect(|h| {
            // Set the hash so we avoid recomputing it later
            let _ = h.hash.set(block_hash);
        });
        Ok(header)
    }

    pub fn last_written(&self) -> Result<Vec<u8>, StoreError> {
        let last_computed_flatkeyvalue = self
            .last_computed_flatkeyvalue
            .read()
            .map_err(|_| StoreError::LockError)?;
        Ok(last_computed_flatkeyvalue.clone())
    }

    /// Returns `true` once the flat-key-value generator has finished its full pass.
    ///
    /// Completion is recorded by the 1-byte `[0xff]` sentinel the generator writes to
    /// `MISC_VALUES["last_written"]` on the final iteration (see `flatkeyvalue_generator`);
    /// every non-final frontier value is a real nibble path (bytes `0x00..=0x0f`), so the
    /// sentinel is unambiguous. Reads the durable marker rather than the in-memory frontier,
    /// which is expanded to `[0xff; 64]`/`[0xff; 131]` and would need length-aware handling.
    ///
    /// Used to gate journal-backed deep reorgs: entries journaled while generation is still
    /// in progress omit past-frontier flat-KV pre-images.
    pub fn flatkeyvalue_fully_generated(&self) -> Result<bool, StoreError> {
        let tx = self.backend.begin_read()?;
        let marker = tx.get(MISC_VALUES, "last_written".as_bytes())?;
        Ok(Self::flatkeyvalue_generation_complete(marker.as_deref()))
    }

    /// Pure completeness test for the durable `last_written` marker: complete iff it is the
    /// exact 1-byte `[0xff]` sentinel. Any in-progress frontier is a nibble path (bytes
    /// `0x00..=0x0f`) and an unset marker is absent, so neither matches.
    fn flatkeyvalue_generation_complete(marker: Option<&[u8]>) -> bool {
        marker == Some([0xff].as_slice())
    }

    fn flatkeyvalue_computed_with_last_written(account: H256, last_written: &[u8]) -> bool {
        let account_nibbles = Nibbles::from_bytes(account.as_bytes());
        &last_written[0..64] > account_nibbles.as_ref()
    }

    /// Raises the `flushed_upto` marker to `block_number`, on disk and in the
    /// buffer's mirror, so the next start anchors at that height. The marker is a
    /// durability floor and only moves forward: a value at or below the current
    /// one is a no-op.
    ///
    /// Used when startup adopts a head above the recorded one (interrupted
    /// full-sync batch): the blocks were flushed by the persist worker before the
    /// interruption, but the failed starts in between walked the marker down.
    ///
    /// Must run while the persist worker is idle (at startup, before p2p and RPC
    /// exist): the buffer update is a read-clone-swap with no compare-and-swap,
    /// so a concurrent persist message could lose either mutation.
    pub fn advance_flushed_upto(&self, block_number: BlockNumber) -> Result<(), StoreError> {
        if self
            .read_flushed_upto_opt()?
            .is_some_and(|current| current >= block_number)
        {
            return Ok(());
        }
        let mut tx = self.backend.begin_write()?;
        write_flushed_upto(tx.as_mut(), block_number)?;
        tx.commit()?;
        mutate_block_buffer(&self.block_data_buffer, |b| {
            b.set_flushed_upto(block_number)
        })
    }

    /// Returns the highest block number durably flushed to disk, or `0` when
    /// the marker is absent. Use [`Self::read_flushed_upto_opt`] when you need
    /// to distinguish "absent marker" (legacy DB, everything is durable) from
    /// "marker present and equal to 0".
    pub fn read_flushed_upto(&self) -> Result<BlockNumber, StoreError> {
        Ok(self.read_flushed_upto_opt()?.unwrap_or(0))
    }

    /// Returns `None` when the marker has never been written — a legacy or fresh
    /// DB where everything is durable and the head must not be clamped to 0.
    fn read_flushed_upto_opt(&self) -> Result<Option<BlockNumber>, StoreError> {
        let tx = self.backend.begin_read()?;
        match tx.get(MISC_VALUES, FLUSHED_UPTO_KEY)? {
            Some(bytes) => Ok(Some(decode_flushed_upto(&bytes)?)),
            None => Ok(None),
        }
    }

    /// Insert a block into the in-memory buffer without writing to disk.
    /// For testing only — gates production code off.
    #[cfg(any(test, feature = "testing"))]
    pub fn buffer_block_for_test(&self, block: &Block) {
        mutate_block_buffer(&self.block_data_buffer, |b| {
            b.insert(block.clone(), vec![], vec![])
        })
        .expect("block_data_buffer lock poisoned");
    }

    /// Synchronously flush the block data buffer to disk.
    /// For testing only — gates production code off.
    #[cfg(any(test, feature = "testing"))]
    pub fn flush_block_data_for_test(&self) -> Result<(), StoreError> {
        flush_block_data(self.backend.as_ref(), &self.block_data_buffer)
    }

    /// Read a raw trie node straight from the on-disk account/storage trie-node
    /// table by its committed key. For testing only — lets a reopen assert which
    /// trie diff-layers a shutdown flush did (or did not) commit to disk.
    #[cfg(any(test, feature = "testing"))]
    pub fn get_trie_node_for_test(
        &self,
        is_account: bool,
        key: &[u8],
    ) -> Result<Option<Vec<u8>>, StoreError> {
        let table = if is_account {
            ACCOUNT_TRIE_NODES
        } else {
            STORAGE_TRIE_NODES
        };
        self.backend.begin_read()?.get(table, key)
    }

    /// Insert a block plus associated codes into the in-memory buffer without
    /// writing to disk.  For testing only — proves the buffer overlay resolves
    /// code that has not been persisted yet.
    #[cfg(any(test, feature = "testing"))]
    pub fn buffer_block_with_codes_for_test(&self, block: &Block, codes: Vec<(H256, Code)>) {
        mutate_block_buffer(&self.block_data_buffer, |b| {
            b.insert(block.clone(), vec![], codes)
        })
        .expect("block_data_buffer lock poisoned");
    }

    /// Mark a state root as in-flight (build pending) without doing a build.
    /// For testing only — simulates the window where the persist worker has not
    /// yet installed the layer, so reads at this root must block in
    /// `gated_snapshot`.
    #[cfg(any(test, feature = "testing"))]
    pub fn register_pending_root_for_test(&self, root: H256) -> Result<(), StoreError> {
        self.pending_trie_roots.register(root)
    }

    /// Clear an in-flight state root (simulates the worker having installed the
    /// layer), unblocking readers waiting in `gated_snapshot`. For testing only.
    #[cfg(any(test, feature = "testing"))]
    pub fn clear_pending_root_for_test(&self, root: H256) {
        self.pending_trie_roots.clear(root)
    }

    /// Boot-time recovery: clamp `latest_block_header` to the durable head.
    ///
    /// Durable head = `min(flushed_upto, latest)` when the marker is present
    /// (buffered blocks past `flushed_upto` may be lost after a crash; the CL
    /// re-sends them via `newPayload`). When the marker is absent the DB
    /// predates deferred persistence and everything is on disk — use `latest`
    /// as-is, never rewind to 0. On first boot the marker is seeded so a later
    /// crash clamps against it rather than an absent (→ 0) marker.
    ///
    /// The marker tracks the max flushed *block number*, not which hash is
    /// canonical at that height. A tip reorg inside the flush window — `Na` at
    /// height N is flushed (marker = N), then `newPayload(Nb)` buffers a sibling
    /// and FCU durably repoints `canonical[N]` to the still-unflushed `Nb` — can
    /// leave `canonical[head]` resolving to a header that never reached disk if
    /// we crash before `Nb` flushes. So we walk `head` down to the highest height
    /// whose canonical hash actually resolves on disk rather than bricking with
    /// `MissingLatestBlockNumber`. A legacy DB (no marker) is exempt: everything
    /// there is durable, so a missing header is real corruption and must surface.
    async fn anchor_to_durable_head(&self, latest: BlockNumber) -> Result<(), StoreError> {
        let marker = self.read_flushed_upto_opt()?;
        let start = match marker {
            Some(flushed) => flushed.min(latest),
            None => latest,
        };

        let mut head = start;
        let latest_block_header = loop {
            match self.load_block_header(head)? {
                Some(header) => break header,
                // Legacy/fresh DB: everything is supposed to be durable, so a
                // missing header is real corruption — surface it, don't rewind.
                None if marker.is_none() => return Err(StoreError::MissingLatestBlockNumber),
                None if head == 0 => return Err(StoreError::MissingLatestBlockNumber),
                None => {
                    warn!(
                        "durable head {head}: canonical hash has no on-disk header \
                         (reorg inside flush window); rewinding"
                    );
                    head -= 1;
                }
            }
        };
        self.latest_block_header.update(latest_block_header);

        // Re-anchor the persisted head when we moved below `latest`, and (re)write
        // the marker to the resolved head: an absent marker is seeded to the
        // durable baseline, and a walked-down head lowers the marker so a later
        // crash clamps against a hash known to resolve.
        let reanchor = head != latest;
        let rewrite_marker = marker != Some(head);
        if reanchor || rewrite_marker {
            let mut tx = self.backend.begin_write()?;
            if reanchor {
                // Re-anchor the persisted head so `get_latest_block_number` and
                // every downstream consumer agree with the clamped head.
                let latest_key = chain_data_key(ChainDataIndex::LatestBlockNumber);
                tx.put(CHAIN_DATA, &latest_key, &head.to_le_bytes())?;
            }
            if rewrite_marker {
                write_flushed_upto(tx.as_mut(), head)?;
            }
            tx.commit()?;
        }
        Ok(())
    }
}

/// Writes the `flushed_upto` block number into an open write batch.
///
/// The caller is responsible for committing `tx` afterward.
pub fn write_flushed_upto(
    tx: &mut dyn StorageWriteBatch,
    n: BlockNumber,
) -> Result<(), StoreError> {
    tx.put(MISC_VALUES, FLUSHED_UPTO_KEY, &n.to_le_bytes())
}

/// Decode an 8-byte little-endian `flushed_upto` marker value.
///
/// Returns an error for a present-but-malformed value so on-disk corruption is
/// surfaced loudly rather than silently resetting the durable marker. Single
/// source of truth for both `from_backend` and [`Store::read_flushed_upto`].
fn decode_flushed_upto(bytes: &[u8]) -> Result<BlockNumber, StoreError> {
    let arr: [u8; 8] = bytes
        .try_into()
        .map_err(|_| StoreError::Custom("Invalid flushed_upto bytes".to_string()))?;
    Ok(BlockNumber::from_le_bytes(arr))
}

/// RCU-swap the block-data buffer: read-clone-mutate-swap with no compare-and-swap,
/// so two concurrent callers lose one mutation. In production the persist worker
/// is the only caller while the node runs; [`Store::advance_flushed_upto`] also
/// calls it, but only at startup before the worker has any message in flight.
/// Test helpers call it on one thread.
fn mutate_block_buffer(
    buffer: &Arc<RwLock<Arc<BlockDataBuffer>>>,
    f: impl FnOnce(&mut BlockDataBuffer),
) -> Result<(), StoreError> {
    let mut new_buf = (*buffer.read().map_err(|_| StoreError::LockError)?.clone()).clone();
    f(&mut new_buf);
    *buffer.write().map_err(|_| StoreError::LockError)? = Arc::new(new_buf);
    Ok(())
}

/// Default for [`StoreConfig::persist_channel_capacity`].
const DEFAULT_PERSIST_CHANNEL_CAPACITY: usize = 2;

/// One unit of work for the persist worker: stage block(s), build the trie diff-layer,
/// flush to disk. `commit_depth` selects the commit gate (`None` = canonical safe-commit
/// root, `Some(depth)` = commit layers deeper than `depth`). `wait_for_flush` selects the
/// ack point independently: `false` acks after staging (carrying the prior flush result),
/// `true` acks after flush.
struct BlockPersist {
    blocks: Vec<(Block, Vec<Receipt>)>,
    codes: Vec<(H256, Code)>,
    parent_state_root: H256,
    child_state_root: H256,
    account_updates: TrieNodesUpdate,
    storage_updates: Vec<(H256, TrieNodesUpdate)>,
    commit_depth: Option<usize>,
    wait_for_flush: bool,
    /// Number of the block whose layer this update represents (the last block in
    /// the batch, matching `child_state_root`). Threaded into the trie layer so
    /// the committed-layer identity is available for the journal write path.
    ///
    /// On a multi-block batch this names only the LAST block, which is NOT a usable
    /// journal identity for the batch's layer. `wait_for_flush` does not make that
    /// safe: it decides when THIS message acks, whereas the layer it creates is
    /// journaled or not by whichever LATER commit sweeps it — `TrieLayerCache::commit`
    /// takes the target layer and every ancestor, so a per-block commit can reach a
    /// batch layer. Any journal entry must therefore be gated on the committed
    /// layer covering exactly one block, not on this message's ack mode.
    block_number: BlockNumber,
    /// Hash of the block whose layer this update represents (see `block_number`).
    block_hash: H256,
    ack: std::sync::mpsc::SyncSender<Result<(), StoreError>>,
}

/// Messages for the persist worker. `Ping(ack)` is the idle handshake for
/// [`Store::wait_for_persistence_idle`]: the FIFO worker handles it only after
/// all earlier `Block` messages are fully processed.
enum PersistMessage {
    Block(Box<BlockPersist>),
    /// Flush the committable layer backlog up to and including this state root, then
    /// prune the flushed layers. Sent by `forkchoice_update` when the safe-commit root
    /// advances: the commit step otherwise only runs while blocks execute, so an
    /// execute-all-then-one-forkchoice flow (e.g. block import) would accumulate every
    /// layer and never persist anything to disk.
    Commit(H256),
    Ping(std::sync::mpsc::SyncSender<Result<(), StoreError>>),
    /// Graceful-shutdown handshake. Handled only after every earlier `Block`
    /// (FIFO), so it both drains in-flight work and force-flushes the block-data
    /// buffer to disk. The trie diff-layers are deliberately left in memory (see
    /// [`Store::shutdown`]). The worker acks and exits.
    Shutdown {
        ack: std::sync::mpsc::SyncSender<Result<(), StoreError>>,
    },
}

/// Write one block's header, body, number, and tx locations into an open batch.
/// Shared by [`Store::add_blocks`] (sync import) and [`flush_block_data`]
/// (deferred flush) so the on-disk encoding stays in lockstep. Receipts and codes
/// are written by callers that need them (only `flush_block_data` does).
fn write_block_data(
    tx: &mut dyn StorageWriteBatch,
    number: BlockNumber,
    hash: BlockHash,
    header: &BlockHeader,
    body: &BlockBody,
) -> Result<(), StoreError> {
    let hash_key = hash.encode_to_vec();
    tx.put(
        HEADERS,
        &hash_key,
        BlockHeaderRLP::from(header.clone()).bytes(),
    )?;
    tx.put(
        BODIES,
        &hash_key,
        BlockBodyRLP::from_bytes(body.encode_to_vec()).bytes(),
    )?;
    tx.put(BLOCK_NUMBERS, &hash_key, &number.to_le_bytes())?;
    for (index, transaction) in body.transactions.iter().enumerate() {
        tx.merge(
            TRANSACTION_LOCATIONS,
            transaction.hash(&NativeCrypto).as_bytes(),
            &encode_tx_location_operand(number, hash, index as u64),
        )?;
    }
    Ok(())
}

/// Write all unflushed blocks to disk in one tx, advance `flushed_upto`, then
/// evict. Eviction is gap-safe: blocks stay buffered until the commit succeeds.
fn flush_block_data(
    backend: &dyn StorageBackend,
    buffer: &Arc<RwLock<Arc<BlockDataBuffer>>>,
) -> Result<(), StoreError> {
    let snapshot = buffer.read().map_err(|_| StoreError::LockError)?.clone();
    let to_flush = snapshot.flushable();
    if to_flush.is_empty() {
        return Ok(());
    }
    let hashes: Vec<_> = to_flush.iter().map(|b| b.header.hash()).collect();
    let codes = snapshot.codes_for(&hashes);
    let mut max_number = snapshot.flushed_upto();

    let mut tx = backend.begin_write()?;
    for b in &to_flush {
        let hash = b.header.hash();
        write_block_data(tx.as_mut(), b.number, hash, &b.header, &b.body)?;
        for (index, receipt) in b.receipts.iter().enumerate() {
            tx.put(
                RECEIPTS_V2,
                &receipt_key(&hash, index as u64),
                &receipt.encode_to_vec(),
            )?;
        }
        max_number = max_number.max(b.number);
    }
    for (code_hash, code) in codes {
        let buf = encode_code(&code);
        tx.put(ACCOUNT_CODES, code_hash.as_ref(), &buf)?;
        tx.put(
            ACCOUNT_CODE_METADATA,
            code_hash.as_ref(),
            &(code.len() as u64).to_be_bytes(),
        )?;
    }
    write_flushed_upto(tx.as_mut(), max_number)?;
    tx.commit()?;

    // Phase 3: evict only after the commit succeeded (gap safety).
    mutate_block_buffer(buffer, |b| b.evict_flushed(max_number))
}

type TrieNodesUpdate = Vec<(Nibbles, Vec<u8>)>;

/// Tracks state roots whose trie diff-layer is in-flight (building but not yet
/// installed in `trie_cache`). `apply_updates` registers a root *before*
/// returning; the worker clears it *after* swapping the layer in. This ordering
/// is mandatory: a reader opening a trie at a pending root blocks until the
/// layer is installed, preventing stale on-disk reads.
#[derive(Debug, Default)]
struct PendingTrieRoots {
    /// Fast-path: when zero, nothing is in flight and readers skip the lock.
    count: AtomicUsize,
    roots: Mutex<HashSet<H256>>,
    ready: Condvar,
}

impl PendingTrieRoots {
    /// Mark `root` as in-flight. MUST be called before the build is handed to
    /// the worker (so the worker's `clear` always finds it) and before the head
    /// can advance to `root` (so any reader that can reference it sees it pending).
    fn register(&self, root: H256) -> Result<(), StoreError> {
        let mut roots = self.roots.lock().map_err(|_| StoreError::LockError)?;
        if roots.insert(root) {
            self.count.fetch_add(1, Ordering::Release);
        }
        Ok(())
    }

    /// Mark `root` as installed and wake any waiting readers. MUST be called only
    /// after the layer is swapped into `trie_cache`, so a woken reader sees it.
    /// Best-effort: a poisoned lock means a reader's `wait_until_ready` also errors,
    /// so no reader deadlocks.
    fn clear(&self, root: H256) {
        let Ok(mut roots) = self.roots.lock() else {
            return;
        };
        if roots.remove(&root) {
            self.count.fetch_sub(1, Ordering::Release);
            self.ready.notify_all();
        }
    }

    /// Block until `root` is no longer in-flight (its layer is installed). Returns
    /// immediately on the fast path when nothing is pending.
    fn wait_until_ready(&self, root: H256) -> Result<(), StoreError> {
        if self.count.load(Ordering::Acquire) == 0 {
            return Ok(());
        }
        let mut roots = self.roots.lock().map_err(|_| StoreError::LockError)?;
        while roots.contains(&root) {
            roots = self.ready.wait(roots).map_err(|_| StoreError::LockError)?;
        }
        Ok(())
    }
}

/// Build the trie diff-layer, RCU-swap it into `trie_cache`, then clear the
/// pending root. Swap MUST precede the clear so a woken reader sees the layer.
/// On swap failure the root is still cleared so gated readers error, not deadlock.
#[allow(clippy::too_many_arguments)]
fn apply_trie_phase1(
    trie_cache: &Arc<RwLock<Arc<TrieLayerCache>>>,
    pending_roots: &PendingTrieRoots,
    parent_state_root: H256,
    child_state_root: H256,
    block_number: BlockNumber,
    block_hash: H256,
    account_updates: TrieNodesUpdate,
    storage_updates: Vec<(H256, TrieNodesUpdate)>,
) -> Result<(), StoreError> {
    let build: Result<(), StoreError> = (|| {
        let new_layer = storage_updates
            .into_iter()
            .flat_map(|(account_hash, nodes)| {
                nodes
                    .into_iter()
                    .map(move |(path, node)| (apply_prefix(Some(account_hash), path), node))
            })
            .chain(account_updates)
            .collect();
        let trie = trie_cache
            .read()
            .map_err(|_| StoreError::LockError)?
            .clone();
        let mut trie_mut = (*trie).clone();
        trie_mut.put_batch(
            parent_state_root,
            child_state_root,
            block_number,
            block_hash,
            new_layer,
        );
        *trie_cache.write().map_err(|_| StoreError::LockError)? = Arc::new(trie_mut);
        Ok(())
    })();
    // Always clear the pending root, whether or not the swap succeeded: on success
    // readers see the installed layer; on failure (poisoning) the lock is poisoned
    // so gated readers error rather than read stale, and we must not leave them
    // blocked forever.
    pending_roots.clear(child_state_root);
    build
}

/// Flush and prune the committable trie-layer backlog. No-ops when nothing is committable.
///
/// `commit_depth` selects the gate. `Some(depth)`: single-canonical-chain execution (batch
/// import, full sync, startup regeneration) commits by depth, because the canonical `head - 128`
/// safe-commit root never lands on a batch layer boundary; sound because these paths only ever
/// extend a single canonical chain (no competing forks to mis-commit). `None`: live block-by-block
/// execution uses the canonical safe-commit gate (`TrieLayerCache::get_commitable`) so non-canonical
/// `newPayload` state is never persisted.
///
/// `is_batch` is independent of the gate and only selects journaling (see
/// [`commit_to_disk`]). It tracks `wait_for_flush`, not `commit_depth.is_some()`, so every
/// per-block path journals and only the bespoke batch path skips it. That keeps the
/// full-sync tail and the import tail journaling exactly as they did when `commit_depth`
/// and `wait_for_flush` were a single flag.
///
/// Startup state regeneration is the one path where this is new behavior rather than
/// preserved behavior: it runs before any forkchoice update, so the safe-commit cell is
/// still zero and the canonical gate committed nothing there, which is the unbounded-layer
/// problem the depth gate fixes. Now that it does commit, it also journals. Deriving
/// `is_batch` from `commit_depth` instead would suppress that, but it would also suppress
/// the full-sync and import tails, and it would leave the journal discontiguous with the
/// on-disk root: surviving entries below a non-journaled commit describe pre-images
/// relative to a root the disk has already moved past.
fn commit_trie_if_due(
    backend: &dyn StorageBackend,
    trie_cache: &Arc<RwLock<Arc<TrieLayerCache>>>,
    fkv_ctl: &SyncSender<FKVGeneratorControlMessage>,
    parent_state_root: H256,
    commit_depth: Option<usize>,
    is_batch: bool,
) -> Result<(), StoreError> {
    let trie = trie_cache
        .read()
        .map_err(|_| StoreError::LockError)?
        .clone();
    // Phase 2 + 3: flush and prune the committable backlog.
    let commitable = match commit_depth {
        Some(depth) => trie.get_commitable_by_depth(parent_state_root, depth),
        None => trie.get_commitable(parent_state_root),
    };
    let Some(root) = commitable else {
        // Nothing to commit to disk, move on.
        return Ok(());
    };
    commit_to_disk(backend, fkv_ctl, trie_cache, &trie, root, is_batch)
}

/// Flush the layer at `root` and all older ancestors to disk, then prune them from the
/// in-memory cache (Phases 2 and 3 of the persistence pipeline).
///
/// `trie` must be the current cache snapshot and `root` a committable layer (as returned
/// by [`TrieLayerCache::get_commitable`]). A `root` that is not a layer commits nothing.
///
/// Reused by both the per-block path ([`commit_trie_if_due`]) and the forkchoice-driven
/// flush ([`PersistMessage::Commit`]): without the latter, an execute-all-then-one-forkchoice
/// flow (block import) would never persist, because the commit step only runs while blocks execute.
fn commit_to_disk(
    backend: &dyn StorageBackend,
    fkv_ctl: &SyncSender<FKVGeneratorControlMessage>,
    trie_cache: &Arc<RwLock<Arc<TrieLayerCache>>>,
    trie: &Arc<TrieLayerCache>,
    root: H256,
    is_batch: bool,
) -> Result<(), StoreError> {
    // `root` need not have a layer: the forkchoice `PersistMessage::Commit(root)` path
    // forwards the safe-commit root without consulting the cache, and `put_batch` skips
    // blocks whose state root equals their parent's. Bail before the side effects below.
    if !trie.has_layer(root) {
        debug!(
            root = ?root,
            layers = trie.layer_count(),
            is_batch,
            "Skipping trie commit: state root has no in-memory layer. Expected when the block \
             did not change the state root (empty L2 blocks) or the root was already flushed."
        );
        return Ok(());
    }

    // Stop the flat-key-value generator thread, as the underlying trie is about to change.
    // Ignore the error, if the channel is closed it means there is no worker to notify.
    let _ = fkv_ctl.send(FKVGeneratorControlMessage::Stop);

    // RCU to remove the bottom layer: update step needs to happen after disk layer is updated.
    let mut trie_mut = (**trie).clone();

    // Open the read view BEFORE the write batch so each `.get()` sees disk as it was
    // before our writes. The journal records `(key, prev_value)` so a future rollback
    // can apply diffs directly without reading state.
    //
    // NOTE: `StorageReadView` does not currently promise true snapshot isolation
    // (see the trait docs in `api/mod.rs`). What makes the pre-image read safe here
    // is that `commit_to_disk` only ever runs on the single persist worker thread,
    // `write_tx` is a buffered batch that does not become visible until
    // `write_tx.commit()` at the end of the function, and the other writers to
    // these CFs touch disjoint key space: the flat-KV generator only writes keys
    // strictly past the committed `last_written` frontier, and snap-sync healing
    // completes before block execution commits layers. So every `.get()` below
    // sees on-disk state as of the begin_read call.
    let read_view = backend.begin_read()?;
    let last_written = read_view
        .get(MISC_VALUES, "last_written".as_bytes())?
        .unwrap_or_default();

    let mut write_tx = backend.begin_write()?;

    // Before encoding, accounts have only the account address as their path, while storage keys have
    // the account address (32 bytes) + storage path (up to 32 bytes).

    // Snapshot the overlay (if any) BEFORE commit so reconciliation can fold its entries
    // into this write batch. After a deep reorg, the first
    // new-chain commit advances disk from the OLD chain's edge `D` directly to the new
    // chain's tip `T` in a single atomic write; the overlay supplies the bridge for keys
    // layer_T does not touch. Only meaningful when `!is_batch`: the legacy batch-store path
    // (`wait_for_flush == true`) is the one path that does not journal.
    let overlay_for_reconciliation = if !is_batch {
        trie.overlay().cloned()
    } else {
        None
    };

    // While an overlay is installed, commit only the bottom layer per pass. The
    // The reconciliation below is defined for a single layer at the pivot
    // tip `T`: bridge entries are folded into that layer's writes, [T, D] is
    // delete_ranged, and T's journal entry records pre-images against the
    // old-chain disk state. A multi-layer sweep would journal upper layers'
    // pre-images against old-chain disk instead of the new-chain/bridge state
    // (the intra-batch pre-image map is not seeded with bridge values), silently
    // corrupting a future unwind. The backlog above `T` drains in later passes,
    // after this commit clears the overlay (see `trie_mut.clear_overlay` below).
    let root = if overlay_for_reconciliation.is_some() {
        trie.bottom_layer_root(root)
    } else {
        root
    };

    // `commit` removes the committed layer(s) and returns one `CommittedLayer` per block
    // in oldest-first order. In normal block-by-block operation this is a single layer,
    // one commit-cadence behind the just-added block. A forkchoice-driven flush of an
    // accumulated backlog (e.g. block import) can return several layers at once, so we
    // write one journal entry per block below rather than merging diffs across blocks.
    //
    // `has_layer` above established `root` is a layer and `trie_mut` is a `Clone` that
    // preserves the map, so this must be `Some`. Reaching the error means the clone lost the
    // map, which is corruption, not the ordinary no-op handled above.
    let committed_layers = trie_mut.commit(root).ok_or_else(|| {
        StoreError::Custom(format!(
            "trie layer for state root {root:?} disappeared between the has_layer check and \
             commit (layers={}, is_batch={is_batch}): the TrieLayerCache clone lost the layer \
             map, so this block's state was not written",
            trie.layer_count(),
        ))
    })?;

    // Deep-reorg reconciliation is a single new-chain commit advancing disk to the pivot
    // tip `T`; it must map to exactly one committed layer. A multi-layer commit with an
    // overlay installed would mean a backlog accumulated between overlay install and
    // flush, corrupting the [T, D] delete_range below.
    debug_assert!(
        overlay_for_reconciliation.is_none() || committed_layers.len() == 1,
        "overlay-backed reconciliation must commit exactly one layer (T), got {}",
        committed_layers.len()
    );

    // Reconciliation: overlay entries the new chain has NOT rewritten must be
    // bridged onto disk so disk fully reflects the pivot->T transition. Keys any committed
    // layer touches are skipped (the layer wins, its value is the post-T state). Overlay-only
    // entries with `None` become an empty-value write -> deleted on disk, matching the
    // "absent at pivot" semantics.
    let extra_writes: Vec<(Vec<u8>, Vec<u8>)> = match &overlay_for_reconciliation {
        Some(overlay) => {
            let layer_keys: rustc_hash::FxHashSet<&Vec<u8>> = committed_layers
                .iter()
                .flat_map(|l| l.nodes.iter().map(|(k, _)| k))
                .collect();
            overlay
                .iter_all_entries()
                .filter(|(_, key, _)| !layer_keys.contains(key))
                .map(|(_, key, value)| (key.clone(), value.clone().unwrap_or_default()))
                .collect()
        }
        None => Vec::new(),
    };
    // Pivot heights (T, D) for the reconciliation commit; `None` in steady state.
    let reorg_heights = overlay_for_reconciliation
        .as_ref()
        .map(|ov| (ov.to_block(), ov.from_block()));

    // Intra-batch overlay of values already staged in THIS write batch, so each block's
    // reverse diff records the value as of the *previous* committed block's write, not
    // just the pre-batch on-disk value. `None` means an earlier block deleted the key.
    // For the common single-layer commit this stays empty and every pre-image comes
    // straight from `read_view`. Only consulted/maintained when journaling (`!is_batch`).
    //
    // PERF: the first touch of each key does one synchronous `read_view.get(table, &key)`.
    // For large state diffs this is O(N) extra reads on the per-block critical path.
    // A follow-up could batch these via `multi_get_cf` if profiling shows it's significant.
    let mut overlay: HashMap<Vec<u8>, Option<Vec<u8>>> = HashMap::new();

    let mut result = Ok(());
    'layers: for layer in &committed_layers {
        // Reverse-diff accumulators for this block's journal entry, one per CF. Each entry
        // stores the on-disk key as-is (storage CFs carry their nibble-encoded account-hash
        // prefix), so a future rollback applies diffs directly without interpretation. For
        // the legacy batch-store path (`is_batch == true`) no journal entry is written.
        // Every per-block path journals — including full sync through the unified
        // pipeline, whose interrupted-batch recovery at startup relies on the newest
        // entry naming the block whose state is on disk.
        let mut journal_account_trie: FlatDiff = Vec::new();
        let mut journal_storage_trie: FlatDiff = Vec::new();
        let mut journal_account_flat: FlatDiff = Vec::new();
        let mut journal_storage_flat: FlatDiff = Vec::new();

        // The reconciliation commit is single-layer at `T`; fold the overlay bridge entries
        // into that layer's writes so they land in T's journal entry. `extra` is empty in
        // steady state (no overlay) and for any non-T layer.
        // With the bottom-layer-only rule above, a commit made while the overlay
        // is installed always contains exactly this one (bottom) layer, so it is
        // the reconciliation layer by construction. Matching by block number
        // (`layer.block_number == to_block`) would be fragile: root-preserving
        // blocks create no layer (L2 empty blocks), so the bottom layer can sit
        // above `to_block` and the bridge entries would be neither written nor
        // journaled before the overlay is cleared.
        let is_reconciliation_layer = overlay_for_reconciliation.is_some();
        let extra: &[(Vec<u8>, Vec<u8>)] = if is_reconciliation_layer {
            &extra_writes
        } else {
            &[]
        };

        // Pre-images for the reconciliation layer must be taken at the pivot, not at the
        // old chain's edge `D`. `T`'s journal entry has to reverse disk back to the pivot
        // state, but this batch is built while disk still holds `D`, so `read_view` yields
        // `D`'s values for every key the old chain rewrote in `[T, D]`. The overlay *is*
        // the pivot-vs-`D` difference, so it wins over `read_view` for any key it carries.
        // `None` in the overlay means "absent at the pivot", which is exactly the journal's
        // "delete on rollback" entry.
        let pivot_overlay = overlay_for_reconciliation.as_ref();

        for (key, value) in layer.nodes.iter().chain(extra.iter()) {
            let (is_leaf, is_account) = classify_trie_key(key.len());

            // Keys past the flat-KV generator's frontier aren't written to disk yet, so
            // they must not be journaled either (a `Some(None)` entry recorded here
            // would cause a rollback to delete a key that was never put). The `continue`
            // jumps over both the write and the journal push below.
            if is_leaf && key.as_slice() > last_written.as_slice() {
                continue;
            }
            let table = if is_leaf {
                if is_account {
                    &ACCOUNT_FLATKEYVALUE
                } else {
                    &STORAGE_FLATKEYVALUE
                }
            } else if is_account {
                &ACCOUNT_TRIE_NODES
            } else {
                &STORAGE_TRIE_NODES
            };

            // Pre-image: the intra-batch overlay wins over the pivot overlay, which wins
            // over disk, so multi-layer commits record each block's true pre-state and the
            // reconciliation layer records the pivot's. Skipped for batch (full-sync) commits.
            let prev_value = if !is_batch {
                match overlay.get(key) {
                    Some(v) => Some(v.clone()),
                    None => match pivot_overlay
                        .and_then(|ov| ov.lookup(OverlayCf::classify_by_key_length(key.len()), key))
                    {
                        Some(v) => Some(v),
                        None => match read_view.get(table, key) {
                            Ok(v) => Some(v),
                            Err(e) => {
                                result = Err(e);
                                break 'layers;
                            }
                        },
                    },
                }
            } else {
                None
            };

            let new_value = if value.is_empty() {
                result = write_tx.delete(table, key);
                None
            } else {
                result = write_tx.put(table, key, value);
                Some(value.clone())
            };
            if result.is_err() {
                break 'layers;
            }

            // Record the reverse-diff entry after the put/delete is staged, so a write
            // error doesn't accumulate state we won't persist.
            if let Some(prev) = prev_value {
                let bucket = match (is_leaf, is_account) {
                    (false, true) => &mut journal_account_trie,
                    (false, false) => &mut journal_storage_trie,
                    (true, true) => &mut journal_account_flat,
                    (true, false) => &mut journal_storage_flat,
                };
                bucket.push((key.clone(), prev));
                // Advance the overlay so a later block in this same commit sees this
                // block's write as its pre-image.
                overlay.insert(key.clone(), new_value);
            }
        }

        // Reconciliation: BEFORE writing this block's journal entry, wipe the
        // obsolete OLD-chain entries in `[T, D]` so the new T entry (below) isn't clobbered
        // by the range delete. Fires only for the single reconciliation layer at height T.
        // T = `overlay.to_block()`, D = `overlay.from_block()`.
        if !is_batch && let Some((t, d)) = reorg_heights {
            debug_assert_eq!(
                layer.block_number, t,
                "first new-chain commit must be at the pivot's T height (overlay.to_block)"
            );
            // Only reconcile when this committed layer is exactly the pivot height `t`.
            // The `delete_range` wipes the OLD-chain journal entries `[t, d]`; if the layer
            // is not at `t` the range delete would drop new-chain entries that belong in
            // `[t, d]` and leave gaps that break later reorg recovery. The debug_assert
            // catches this loudly in tests; in release we skip (and log) rather than corrupt
            // STATE_HISTORY.
            if layer.block_number == t {
                let start = t.to_be_bytes();
                let end = d.saturating_add(1).to_be_bytes();
                result = write_tx.delete_range(STATE_HISTORY, &start, &end);
                if result.is_err() {
                    break 'layers;
                }
            } else {
                error!(
                    block_number = layer.block_number,
                    t,
                    d,
                    "deep-reorg reconciliation skipped: committed layer not at pivot height; skipping STATE_HISTORY delete_range to avoid history gaps"
                );
            }
        }

        // Stage this block's journal entry into the same write batch as the trie/flat-KV
        // overwrites. `put` is buffered until `commit`, so all CFs and every block's entry
        // land atomically (or none do on commit failure). Each entry is keyed and identified
        // by its own COMMITTED block, not the in-flight block whose insertion triggered this
        // commit (that block commits later, one cadence behind).
        if !is_batch {
            let entry = JournalEntry {
                block_hash: layer.block_hash,
                parent_state_root: layer.parent_state_root,
                account_trie_diff: journal_account_trie,
                storage_trie_diff: journal_storage_trie,
                account_flat_diff: journal_account_flat,
                storage_flat_diff: journal_storage_flat,
            };
            result = write_tx.put(
                STATE_HISTORY,
                &layer.block_number.to_be_bytes(),
                &entry.encode(),
            );
            if result.is_err() {
                break 'layers;
            }
        }
    }

    if result.is_ok() {
        result = write_tx.commit();
    }
    // We want to send this message even if there was an error during the batch write
    let _ = fkv_ctl.send(FKVGeneratorControlMessage::Continue);
    result?;

    // Reconciliation succeeded: drop the overlay from the cache. Subsequent
    // commits revert to the normal one-block path.
    if overlay_for_reconciliation.is_some() {
        trie_mut.clear_overlay();
    }

    // Phase 3: update diff layers with the removal of bottom layer.
    //
    // SAFETY: `install_overlay_for_reorg` rendezvous-pings this worker before
    // swapping `trie_cache`, so no concurrent RCU can race the write-back here.
    // See the Ping comment in `install_overlay_for_reorg`.
    *trie_cache.write().map_err(|_| StoreError::LockError)? = Arc::new(trie_mut);
    Ok(())
}

// NOTE: we don't receive `Store` here to avoid cyclic dependencies
// with the other end of `control_rx`
fn flatkeyvalue_generator(
    backend: &Arc<dyn StorageBackend>,
    last_computed_fkv: &RwLock<Vec<u8>>,
    control_rx: &std::sync::mpsc::Receiver<FKVGeneratorControlMessage>,
) -> Result<(), StoreError> {
    info!("Generation of FlatKeyValue started.");
    let initial_last_written = backend
        .begin_read()?
        .get(MISC_VALUES, "last_written".as_bytes())?
        .unwrap_or_default();

    if initial_last_written.is_empty() {
        // First time generating the FKV. Remove all FKV entries just in case
        backend.clear_table(ACCOUNT_FLATKEYVALUE)?;
        backend.clear_table(STORAGE_FLATKEYVALUE)?;
    } else if initial_last_written == [0xff] {
        // FKV was already generated
        info!("FlatKeyValue already generated. Skipping.");
        return Ok(());
    }

    loop {
        // Acquire a fresh read view per iteration so updates performed while the
        // generator is paused are visible after a Continue signal.
        let read_tx = backend.begin_read()?;
        let root = read_tx
            .get(ACCOUNT_TRIE_NODES, &[])?
            .ok_or(StoreError::MissingLatestBlockNumber)?;
        let root: Node = ethrex_trie::Node::decode(&root)?;
        let state_root = root.compute_hash(&NativeCrypto).finalize(&NativeCrypto);

        let last_written = read_tx
            .get(MISC_VALUES, "last_written".as_bytes())?
            .unwrap_or_default();
        let last_written_account = last_written
            .get(0..64)
            .map(|v| Nibbles::from_hex(v.to_vec()))
            .unwrap_or_default();
        let mut last_written_storage = last_written
            .get(66..130)
            .map(|v| Nibbles::from_hex(v.to_vec()))
            .unwrap_or_default();

        debug!("Starting FlatKeyValue loop pivot={last_written:?} SR={state_root:x}");

        let mut ctr = 0;
        let mut write_txn = backend.begin_write()?;
        let mut iter = Trie::open(
            Box::new(BackendTrieDB::new_for_accounts_with_view(
                backend.clone(),
                read_tx.clone(),
                last_written.clone(),
            )?),
            state_root,
        )
        .into_iter();
        if last_written_account > Nibbles::default() {
            iter.advance(last_written_account.to_bytes())?;
        }
        let res = iter.try_for_each(|(path, node)| -> Result<(), StoreError> {
            let Node::Leaf(node) = node else {
                return Ok(());
            };
            let account_state = AccountState::decode(&node.value)?;
            let account_hash = H256::from_slice(&path.to_bytes());
            write_txn.put(MISC_VALUES, "last_written".as_bytes(), path.as_ref())?;
            write_txn.put(ACCOUNT_FLATKEYVALUE, path.as_ref(), &node.value)?;
            ctr += 1;
            if ctr > 10_000 {
                write_txn.commit()?;
                write_txn = backend.begin_write()?;
                *last_computed_fkv
                    .write()
                    .map_err(|_| StoreError::LockError)? = path.as_ref().to_vec();
                ctr = 0;
            }

            let mut iter_inner = Trie::open(
                Box::new(BackendTrieDB::new_for_account_storage_with_view(
                    backend.clone(),
                    read_tx.clone(),
                    account_hash,
                    path.as_ref().to_vec(),
                )?),
                account_state.storage_root,
            )
            .into_iter();
            if last_written_storage > Nibbles::default() {
                iter_inner.advance(last_written_storage.to_bytes())?;
                last_written_storage = Nibbles::default();
            }
            iter_inner.try_for_each(|(path, node)| -> Result<(), StoreError> {
                let Node::Leaf(node) = node else {
                    return Ok(());
                };
                let key = apply_prefix(Some(account_hash), path);
                write_txn.put(MISC_VALUES, "last_written".as_bytes(), key.as_ref())?;
                write_txn.put(STORAGE_FLATKEYVALUE, key.as_ref(), &node.value)?;
                ctr += 1;
                if ctr > 10_000 {
                    write_txn.commit()?;
                    write_txn = backend.begin_write()?;
                    *last_computed_fkv
                        .write()
                        .map_err(|_| StoreError::LockError)? = key.into_vec();
                    ctr = 0;
                }
                fkv_check_for_stop_msg(control_rx)?;
                Ok(())
            })?;
            fkv_check_for_stop_msg(control_rx)?;
            Ok(())
        });
        match res {
            Err(StoreError::PivotChanged) => {
                match control_rx.recv() {
                    Ok(FKVGeneratorControlMessage::Continue) => {}
                    Ok(FKVGeneratorControlMessage::Stop) => {
                        return Err(StoreError::Custom("Unexpected Stop message".to_string()));
                    }
                    // If the channel was closed, we stop generation prematurely
                    Err(std::sync::mpsc::RecvError) => {
                        info!("Store closed, stopping FlatKeyValue generation.");
                        return Ok(());
                    }
                }
            }
            Err(err) => return Err(err),
            Ok(()) => {
                write_txn.put(MISC_VALUES, "last_written".as_bytes(), &[0xff])?;
                write_txn.commit()?;
                *last_computed_fkv
                    .write()
                    .map_err(|_| StoreError::LockError)? = vec![0xff; 131];
                info!("FlatKeyValue generation finished.");
                return Ok(());
            }
        };
    }
}

fn fkv_check_for_stop_msg(
    control_rx: &std::sync::mpsc::Receiver<FKVGeneratorControlMessage>,
) -> Result<(), StoreError> {
    match control_rx.try_recv() {
        Ok(FKVGeneratorControlMessage::Stop) | Err(TryRecvError::Disconnected) => {
            return Err(StoreError::PivotChanged);
        }
        Ok(FKVGeneratorControlMessage::Continue) => {
            return Err(StoreError::Custom(
                "Unexpected Continue message".to_string(),
            ));
        }
        Err(TryRecvError::Empty) => {}
    }
    Ok(())
}

fn state_trie_locked_backend(
    backend: &dyn StorageBackend,
    last_written: Vec<u8>,
) -> Result<BackendTrieDBLocked, StoreError> {
    // No address prefix for state trie
    BackendTrieDBLocked::new(backend, last_written)
}

pub struct AccountProof {
    pub proof: Vec<NodeRLP>,
    pub account: AccountState,
    pub storage_proof: Vec<StorageSlotProof>,
}

pub struct StorageSlotProof {
    pub proof: Vec<NodeRLP>,
    pub key: H256,
    pub value: U256,
}

pub struct AncestorIterator {
    store: Store,
    next_hash: BlockHash,
}

impl Iterator for AncestorIterator {
    type Item = Result<(BlockHash, BlockHeader), StoreError>;

    fn next(&mut self) -> Option<Self::Item> {
        let next_hash = self.next_hash;
        // Buffer-aware: a not-yet-flushed ancestor (e.g. on a side branch during
        // a reorg) must be visible here, or a BLOCKHASH opcode resolving through
        // this walk would wrongly reject a valid block.
        match self.store.get_block_header_by_hash(next_hash) {
            Ok(Some(header)) => {
                let ret_hash = self.next_hash;
                self.next_hash = header.parent_hash;
                Some(Ok((ret_hash, header)))
            }
            Ok(None) => None,
            Err(e) => Some(Err(e)),
        }
    }
}

pub fn hash_address(address: &Address) -> Vec<u8> {
    keccak_hash(address.to_fixed_bytes()).to_vec()
}

fn hash_address_fixed(address: &Address) -> H256 {
    keccak(address.to_fixed_bytes())
}

pub fn hash_key(key: &H256) -> Vec<u8> {
    keccak_hash(key.to_fixed_bytes()).to_vec()
}

pub fn hash_key_fixed(key: &H256) -> [u8; 32] {
    keccak_hash(key.to_fixed_bytes())
}

fn chain_data_key(index: ChainDataIndex) -> Vec<u8> {
    (index as u8).encode_to_vec()
}

/// Decodes a block number stored as little-endian bytes.
fn decode_block_number(bytes: Vec<u8>) -> Result<BlockNumber, StoreError> {
    let array: [u8; 8] = bytes
        .try_into()
        .map_err(|_| StoreError::Custom("Invalid BlockNumber bytes".to_string()))?;
    Ok(BlockNumber::from_le_bytes(array))
}

fn snap_state_key(index: SnapStateIndex) -> Vec<u8> {
    (index as u8).encode_to_vec()
}

/// Builds a fixed-width RECEIPTS key: block_hash (32B) || index (8B BE).
pub fn receipt_key(block_hash: &BlockHash, index: u64) -> Vec<u8> {
    let mut key = Vec::with_capacity(40);
    key.extend_from_slice(block_hash.as_bytes());
    key.extend_from_slice(&index.to_be_bytes());
    key
}

fn encode_code(code: &Code) -> Vec<u8> {
    let jumpdests = code.jumpdests();
    let mut buf = Vec::with_capacity(6 + code.len() + jumpdests.len());
    code.code().encode(&mut buf);
    jumpdests.encode(&mut buf);
    buf
}

/// Decodes the JUMPDEST bitmap stored after the bytecode in an [`ACCOUNT_CODES`] value.
///
/// Values written before the bitmap representation hold an RLP *list* of `u32` offsets
/// there instead. The RLP item header distinguishes a list from a byte string, so both
/// forms are readable: for the older one the bitmap is rebuilt from the bytecode, which
/// is cheaper than decoding the list.
/// Decodes the stored jump-destination bitmap. The bool is `true` when the value
/// was in the legacy RLP-list-of-`u32` format and had to be recomputed from
/// `code` — the caller uses it to self-heal the row to the bitmap format so the
/// recompute is paid at most once per entry rather than on every cold read.
fn decode_jumpdests(code: &[u8], encoded: &[u8]) -> Result<(Arc<[u8]>, bool), StoreError> {
    let (is_list, payload, _) = decode_rlp_item(encoded)?;
    if is_list {
        return Ok((Code::compute_jumpdests(code), true));
    }
    Ok((Arc::from(payload), false))
}

#[derive(Debug, Default, Clone)]
struct LatestBlockHeaderCache {
    current: Arc<Mutex<Arc<BlockHeader>>>,
}

impl LatestBlockHeaderCache {
    pub fn get(&self) -> Arc<BlockHeader> {
        self.current.lock().expect("poisoned mutex").clone()
    }

    pub fn update(&self, header: BlockHeader) {
        let new = Arc::new(header);
        *self.current.lock().expect("poisoned mutex") = new;
    }
}

#[derive(Debug, Serialize, Deserialize)]
pub struct StoreMetadata {
    pub schema_version: u64,
}

impl StoreMetadata {
    pub fn new(schema_version: u64) -> Self {
        Self { schema_version }
    }
}

/// Reads the schema version from the metadata file, if it exists.
///
/// Returns `Some(version)` when metadata.json is present and valid,
/// or `None` when the file does not exist.
fn read_store_schema_version(path: &Path) -> Result<Option<u64>, StoreError> {
    let metadata_path = path.join(STORE_METADATA_FILENAME);
    if !metadata_path.exists() {
        return Ok(None);
    }
    if !metadata_path.is_file() {
        return Err(StoreError::Custom(
            "store schema path exists but is not a file".to_string(),
        ));
    }
    let file_contents = std::fs::read_to_string(metadata_path)?;
    let metadata: StoreMetadata = serde_json::from_str(&file_contents)?;
    Ok(Some(metadata.schema_version))
}

fn init_metadata_file(parent_path: &Path) -> Result<(), StoreError> {
    std::fs::create_dir_all(parent_path)?;

    let metadata_path = parent_path.join(STORE_METADATA_FILENAME);
    let metadata = StoreMetadata::new(STORE_SCHEMA_VERSION);
    let serialized_metadata = serde_json::to_string_pretty(&metadata)?;
    let mut new_file = std::fs::File::create_new(metadata_path)?;
    new_file.write_all(serialized_metadata.as_bytes())?;
    Ok(())
}

/// Returns `true` if `path` contains a *legacy* database — one written before
/// the metadata file existed, so it has no `metadata.json` to identify it.
/// Detected by RocksDB's own marker files, as opposed to unrelated files that
/// merely share the datadir. Only meaningful once metadata has been confirmed
/// absent; otherwise prefer `has_valid_db`, which keys off the metadata file.
///
/// Previously the caller treated *any* non-empty directory as such a legacy
/// database, which made startup fail when unrelated files lived alongside the DB
/// — e.g. EthDocker writes the JWT secret into the datadir (issue #5680). We
/// instead look for RocksDB's marker files, so a datadir that only contains such
/// unrelated files is correctly treated as fresh.
fn dir_contains_legacy_db(path: &Path) -> Result<bool, StoreError> {
    // `CURRENT` has a fixed name and is written by every RocksDB instance, so
    // check for it directly instead of scanning a datadir that may hold many
    // unrelated files.
    if path.join("CURRENT").is_file() {
        return Ok(true);
    }
    // The manifest has a numeric suffix (`MANIFEST-<n>`), so it can only be
    // found by scanning. Restrict to plain files: a directory that happens to
    // share the name is not a database marker.
    for entry in std::fs::read_dir(path)? {
        let entry = entry?;
        if !entry.file_type()?.is_file() {
            continue;
        }
        if entry.file_name().to_string_lossy().starts_with("MANIFEST-") {
            return Ok(true);
        }
    }
    Ok(false)
}

/// Checks whether a valid (or migratable) database exists at the given path
/// by looking for a metadata.json file with a schema version between 1 and
/// `STORE_SCHEMA_VERSION` (inclusive).
pub fn has_valid_db(path: &Path) -> bool {
    let metadata_path = path.join(STORE_METADATA_FILENAME);
    if !metadata_path.is_file() {
        return false;
    }
    let Ok(contents) = std::fs::read_to_string(&metadata_path) else {
        return false;
    };
    let Ok(metadata) = serde_json::from_str::<StoreMetadata>(&contents) else {
        return false;
    };
    metadata.schema_version >= 1 && metadata.schema_version <= STORE_SCHEMA_VERSION
}

/// Reads the chain ID from an existing database without performing a full
/// store initialization. Returns `None` if the database doesn't exist or
/// the chain config can't be read. Always returns `None` when compiled
/// without the `rocksdb` feature.
///
/// Each failure mode logs a warning so callers (and operators) can diagnose
/// why an existing database was not usable — previously every error was
/// silently swallowed by `.ok()?`.
pub fn read_chain_id_from_db(path: &Path) -> Option<u64> {
    if !has_valid_db(path) {
        return None;
    }
    #[cfg(feature = "rocksdb")]
    {
        // The cache size is irrelevant for this one-shot chain-id read (the LRU
        // is sized as a ceiling, not pre-allocated), so we use the ceiling.
        let backend = match RocksDBBackend::open(path, MAX_ROCKSDB_BLOCK_CACHE_SIZE_BYTES) {
            Ok(backend) => backend,
            Err(e) => {
                warn!("Failed to open RocksDB at {path:?} to read chain ID: {e}");
                return None;
            }
        };
        let read = match backend.begin_read() {
            Ok(read) => read,
            Err(e) => {
                warn!("Failed to begin read transaction at {path:?}: {e}");
                return None;
            }
        };
        let key = chain_data_key(ChainDataIndex::ChainConfig);
        let bytes = match read.get(CHAIN_DATA, &key) {
            Ok(Some(bytes)) => bytes,
            Ok(None) => {
                warn!("Chain config entry not found in database at {path:?}");
                return None;
            }
            Err(e) => {
                warn!("Failed to read chain config from database at {path:?}: {e}");
                return None;
            }
        };
        // Only extract chain_id here: the stored `ChainConfig` JSON may include
        // fields whose serialization changed across releases (e.g. pre-v10 wrote
        // `terminal_total_difficulty` as a plain number, v10 expects hex string).
        // Deserializing the full struct would reject otherwise-migratable v9 data.
        #[derive(serde::Deserialize)]
        #[serde(rename_all = "camelCase")]
        struct ChainIdOnly {
            chain_id: u64,
        }
        match serde_json::from_slice::<ChainIdOnly>(&bytes) {
            Ok(partial) => Some(partial.chain_id),
            Err(e) => {
                warn!("Failed to deserialize chain ID from database at {path:?}: {e}");
                None
            }
        }
    }
    #[cfg(not(feature = "rocksdb"))]
    {
        let _ = path;
        None
    }
}

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

    const JUMPDEST: u8 = 0x5b;
    const PUSH1: u8 = 0x60;

    /// `EXTCODESIZE` resolves a length through [`Store::get_code_metadata`], and
    /// execution runs on rayon workers, which are not inside a Tokio runtime. A hash
    /// with no metadata row SHALL still answer from the bytecode there rather than
    /// panicking on the backfill spawn. This test is deliberately NOT `#[tokio::test]`:
    /// the absence of a runtime is the condition under test.
    #[test]
    fn metadata_falls_back_to_the_bytecode_without_a_tokio_runtime() {
        use crate::backend::in_memory::InMemoryBackend;

        let backend: Arc<dyn StorageBackend> = Arc::new(InMemoryBackend::open().unwrap());
        let dir = tempfile::tempdir().unwrap();
        let store = Store::from_backend(
            backend.clone(),
            dir.path().to_path_buf(),
            1,
            DEFAULT_PERSIST_CHANNEL_CAPACITY,
        )
        .unwrap();

        // Code present, metadata row absent: the shape of any database written before
        // ACCOUNT_CODE_METADATA existed, which no migration backfills.
        let code = jumpdest_dense_code();
        let mut tx = backend.begin_write().unwrap();
        tx.put(ACCOUNT_CODES, code.hash.as_bytes(), &encode_code(&code))
            .unwrap();
        tx.commit().unwrap();

        let metadata = store
            .get_code_metadata(code.hash)
            .expect("metadata read must not fail off-runtime");
        assert_eq!(metadata.map(|m| m.length), Some(code.len() as u64));
    }

    /// A max-size contract that is almost entirely JUMPDESTs, the shape that makes the
    /// jump-destination representation matter: as a list of offsets it is ~4x the size
    /// of the bytecode it describes.
    fn jumpdest_dense_code() -> Code {
        let mut bytecode = vec![JUMPDEST; 24576];
        bytecode[0] = 0x00;
        Code::from_bytecode_unchecked(bytecode.into(), H256::zero())
    }

    /// Encodes an `ACCOUNT_CODES` value the way it was written before the bitmap: the
    /// bytecode followed by an RLP list of `u32` JUMPDEST offsets.
    fn encode_code_legacy(code: &Code) -> Vec<u8> {
        let offsets: Vec<u32> = (0..code.len())
            .filter(|offset| code.is_valid_jumpdest(*offset))
            .map(|offset| offset as u32)
            .collect();
        let mut buf = Vec::new();
        code.code().encode(&mut buf);
        offsets.encode(&mut buf);
        buf
    }

    #[test]
    fn encoded_value_carries_one_bitmap_bit_per_code_byte() {
        let code = jumpdest_dense_code();
        let encoded = encode_code(&code);

        assert_eq!(encoded.len(), 6 + code.len() + code.len().div_ceil(8));
    }

    #[test]
    fn round_trip_preserves_the_bitmap() {
        let code = jumpdest_dense_code();
        let encoded = encode_code(&code);

        let (bytecode, jumpdests) = decode_bytes(&encoded).unwrap();
        assert_eq!(bytecode, code.code());
        let (rebuilt, was_legacy) = decode_jumpdests(bytecode, jumpdests).unwrap();
        assert_eq!(rebuilt.as_ref(), code.jumpdests());
        assert!(!was_legacy, "bitmap format must not be flagged legacy");
    }

    /// Values written before the bitmap SHALL still decode, with the bitmap rebuilt from
    /// the bytecode, so an existing database needs no rewriting.
    ///
    /// The expectation is built from the offsets the legacy value itself names, not from
    /// `Code::compute_jumpdests`, so this pins the rebuilt bitmap against the old format
    /// rather than against the implementation that produces it.
    #[test]
    fn legacy_offset_list_values_decode_to_the_same_bitmap() {
        for (bytecode, offsets) in [
            (vec![0x00, JUMPDEST, 0x00, JUMPDEST], vec![1u32, 3]),
            // The byte after PUSH1 is its immediate, so only offset 2 is a destination.
            (vec![PUSH1, JUMPDEST, JUMPDEST], vec![2u32]),
            (vec![0x00; 32], vec![]),
        ] {
            let code = Code::from_bytecode_unchecked(bytecode.clone().into(), H256::zero());
            let legacy = encode_code_legacy(&code);

            let mut expected = vec![0u8; bytecode.len().div_ceil(8)];
            for offset in &offsets {
                let index = usize::try_from(*offset).expect("offset fits usize");
                expected[index / 8] |= 1 << (index % 8);
            }
            // A jumpless contract stores no bitmap at all rather than an all-zero one.
            if offsets.is_empty() {
                expected.clear();
            }

            let (decoded_bytecode, jumpdests) = decode_bytes(&legacy).unwrap();
            assert_eq!(decoded_bytecode, code.code());
            let (rebuilt, was_legacy) = decode_jumpdests(decoded_bytecode, jumpdests).unwrap();
            assert_eq!(
                rebuilt.as_ref(),
                expected.as_slice(),
                "rebuilt bitmap disagrees with the legacy offsets for {bytecode:?}"
            );
            assert!(
                was_legacy,
                "legacy RLP-list format must be flagged for self-heal"
            );
        }
    }

    /// Re-inserting a cached hash SHALL NOT grow the accounted size, or the counter
    /// drifts up until the cache evicts entries that fit.
    #[test]
    fn repeated_inserts_do_not_inflate_the_accounted_size() {
        let code = jumpdest_dense_code();
        let mut cache = CodeCache::default();

        cache.insert(&code).unwrap();
        let after_first = cache.cache_size;
        for _ in 0..8 {
            cache.insert(&code).unwrap();
        }

        assert_eq!(cache.cache_size, after_first);
        assert_eq!(cache.inner_cache.len(), 1);
    }

    /// The accounted size SHALL include the bytecode itself, so the cache honors its
    /// memory budget instead of holding orders of magnitude more than it accounts for.
    #[test]
    fn accounted_size_covers_the_bytecode_and_bitmap() {
        let code = jumpdest_dense_code();
        let mut cache = CodeCache::default();
        cache.insert(&code).unwrap();

        assert!(cache.cache_size >= (code.len() + code.jumpdests().len()) as u64);
    }

    #[test]
    fn cache_evicts_down_to_its_budget() {
        let mut cache = CodeCache {
            max_size: 128 * 1024,
            ..Default::default()
        };

        for i in 0..16u8 {
            let mut bytecode = vec![JUMPDEST; 24576];
            bytecode[0] = i;
            cache
                .insert(&Code::from_bytecode_unchecked(
                    bytecode.into(),
                    H256::from_low_u64_be(i.into()),
                ))
                .unwrap();
        }

        assert!(cache.cache_size <= cache.max_size);
        assert!(cache.inner_cache.len() < 16);
    }

    /// The default budget SHALL be the one the cache actually enforces, so a change to
    /// the constant cannot silently leave the cache unbounded.
    #[test]
    fn default_cache_uses_the_configured_budget() {
        assert_eq!(CodeCache::default().max_size, CODE_CACHE_MAX_SIZE);
    }
}

#[cfg(test)]
mod state_history_tests {
    use super::*;
    use crate::api::tables::STATE_HISTORY;
    use crate::backend::in_memory::InMemoryBackend;
    use crate::journal::JournalEntry;
    use ethrex_common::types::{BlockBody, BlockHeader};
    use ethrex_trie::Nibbles;
    use std::time::{Duration, Instant};

    fn make_block(number: BlockNumber, parent_hash: H256, state_root: H256) -> Block {
        let header = BlockHeader {
            number,
            parent_hash,
            state_root,
            ..Default::default()
        };
        Block::new(header, BlockBody::default())
    }

    /// Polls `STATE_HISTORY` for an entry at the given block number, up to `timeout`.
    /// The trie worker commits to disk asynchronously after `store_block_updates`
    /// returns, so a small wait window is required.
    fn await_journal_entry(
        backend: &Arc<dyn StorageBackend>,
        block_number: BlockNumber,
        timeout: Duration,
    ) -> Option<Vec<u8>> {
        let key = block_number.to_be_bytes();
        let deadline = Instant::now() + timeout;
        loop {
            let read = backend.begin_read().ok()?;
            if let Ok(Some(v)) = read.get(STATE_HISTORY, &key) {
                return Some(v);
            }
            if Instant::now() >= deadline {
                return None;
            }
            std::thread::sleep(Duration::from_millis(10));
        }
    }

    /// Asserts no STATE_HISTORY entry materializes for `block_number` within the
    /// given window. Polls repeatedly; if an entry appears at any poll, fails
    /// loudly. Absence at every poll over the full window counts as verified.
    /// More robust than a single fixed sleep under CI load.
    fn assert_no_journal_entry(backend: &Arc<dyn StorageBackend>, block_number: BlockNumber) {
        let window = Duration::from_millis(500);
        let key = block_number.to_be_bytes();
        let deadline = Instant::now() + window;
        loop {
            let read = backend.begin_read().expect("read view");
            let v = read.get(STATE_HISTORY, &key).expect("get");
            assert!(
                v.is_none(),
                "expected no STATE_HISTORY entry for block {block_number}, got {v:?}"
            );
            if Instant::now() >= deadline {
                return;
            }
            std::thread::sleep(Duration::from_millis(20));
        }
    }

    /// Live commits gate on the canonical safe-commit root cell (see `get_commitable`),
    /// which only advances via forkchoice. We simulate that by calling
    /// `set_safe_commit_root` to the parent block's state root before storing the next
    /// block: storing block N+1 then commits block N's layer to disk, producing one
    /// journal entry per committed block. We verify entries for block 1 and block 2.
    #[test]
    fn journal_entry_written_per_block_in_regular_mode() {
        let backend: Arc<dyn StorageBackend> = Arc::new(InMemoryBackend::open().unwrap());
        let dir = tempfile::tempdir().unwrap();
        let store = Store::from_backend(
            backend.clone(),
            dir.path().to_path_buf(),
            1,
            DEFAULT_PERSIST_CHANNEL_CAPACITY,
        )
        .unwrap();

        let state_root_1 = H256::repeat_byte(0x11);
        let block1 = make_block(1, H256::zero(), state_root_1);
        let block1_hash = block1.hash();
        store
            .store_block_updates(UpdateBatch {
                account_updates: vec![(Nibbles::from_raw(&[0x00, 0x01], false), vec![0xab, 0xcd])],
                storage_updates: vec![],
                blocks: vec![block1],
                receipts: vec![],
                code_updates: vec![],
                commit_depth: None,
                wait_for_flush: false,
            })
            .unwrap();

        // Advance the safe-commit root to block 1's state root, then store block 2:
        // the canonical gate now finds block 1's layer committable and flushes it.
        store.set_safe_commit_root(state_root_1).unwrap();
        let state_root_2 = H256::repeat_byte(0x22);
        let block2 = make_block(2, block1_hash, state_root_2);
        let block2_hash = block2.hash();
        store
            .store_block_updates(UpdateBatch {
                account_updates: vec![(Nibbles::from_raw(&[0x00, 0x02], false), vec![0xef, 0x11])],
                storage_updates: vec![],
                blocks: vec![block2],
                receipts: vec![],
                code_updates: vec![],
                commit_depth: None,
                wait_for_flush: false,
            })
            .unwrap();

        let bytes = await_journal_entry(&backend, 1, Duration::from_secs(2))
            .expect("STATE_HISTORY entry for block 1 should appear after block 2 commits it");
        let entry = JournalEntry::decode(&bytes).unwrap();
        assert_eq!(entry.block_hash, block1_hash);
        assert_eq!(entry.parent_state_root, H256::zero());
        assert!(!entry.account_trie_diff.is_empty());
        let (path, prev) = &entry.account_trie_diff[0];
        assert_eq!(prev, &None, "first-time write means previous value is None");
        assert!(path.len() < 65);

        // Advance the safe-commit root to block 2's state root, then store block 3:
        // block 2's layer is now committable.
        store.set_safe_commit_root(state_root_2).unwrap();
        let state_root_3 = H256::repeat_byte(0x33);
        let block3 = make_block(3, block2_hash, state_root_3);
        store
            .store_block_updates(UpdateBatch {
                account_updates: vec![(Nibbles::from_raw(&[0x00, 0x03], false), vec![0x77])],
                storage_updates: vec![],
                blocks: vec![block3],
                receipts: vec![],
                code_updates: vec![],
                commit_depth: None,
                wait_for_flush: false,
            })
            .unwrap();

        let bytes = await_journal_entry(&backend, 2, Duration::from_secs(2))
            .expect("STATE_HISTORY entry for block 2 should appear after block 3 commits it");
        let entry = JournalEntry::decode(&bytes).unwrap();
        assert_eq!(entry.block_hash, block2_hash);
        assert_eq!(entry.parent_state_root, state_root_1);
        assert!(!entry.account_trie_diff.is_empty());
    }

    /// A depth-gated commit that is NOT in batch mode SHALL journal, and its entries
    /// SHALL carry the committed layer's own identity and pre-image.
    ///
    /// `commit_depth: Some(d)` together with `wait_for_flush: false` — the startup
    /// regeneration and sync/import tail paths — is a combination that did not exist while
    /// the commit gate and the ack timing were a single flag, so no other test covers it.
    /// Journaling follows `wait_for_flush`, so deriving it from `commit_depth.is_some()`
    /// instead would silently stop journaling on exactly these paths; this test fails in
    /// that case.
    #[test]
    fn depth_gated_per_block_commit_journals_with_correct_pre_image() {
        let backend: Arc<dyn StorageBackend> = Arc::new(InMemoryBackend::open().unwrap());
        let dir = tempfile::tempdir().unwrap();
        let store = Store::from_backend(
            backend.clone(),
            dir.path().to_path_buf(),
            1,
            DEFAULT_PERSIST_CHANNEL_CAPACITY,
        )
        .unwrap();

        // One shared key rewritten every block, so pre-images are checkable.
        let shared = Nibbles::from_raw(&[0x0a, 0x0b], false);
        let mut prev_hash = H256::zero();
        let mut hashes = vec![H256::zero()];
        let mut roots = vec![H256::zero()];
        for n in 1..=6u64 {
            let state_root = H256::repeat_byte(0xc0 | (n as u8));
            let block = make_block(n, prev_hash, state_root);
            prev_hash = block.hash();
            hashes.push(prev_hash);
            roots.push(state_root);
            store
                .store_block_updates(UpdateBatch {
                    account_updates: vec![(shared.clone(), vec![0xb0 | (n as u8)])],
                    storage_updates: vec![],
                    blocks: vec![block],
                    receipts: vec![],
                    code_updates: vec![],
                    commit_depth: Some(3),
                    wait_for_flush: false,
                })
                .unwrap();
        }

        let bytes = await_journal_entry(&backend, 1, Duration::from_secs(2))
            .expect("a depth-gated per-block commit must journal");
        let entry = JournalEntry::decode(&bytes).unwrap();
        assert_eq!(
            entry.block_hash, hashes[1],
            "entry 1 carries block 1's identity"
        );
        assert_eq!(entry.parent_state_root, H256::zero());
        assert_eq!(
            entry.account_trie_diff[0].1, None,
            "block 1 first-writes the key, so its pre-image is absence"
        );

        let bytes =
            await_journal_entry(&backend, 2, Duration::from_secs(2)).expect("entry for block 2");
        let entry = JournalEntry::decode(&bytes).unwrap();
        assert_eq!(
            entry.block_hash, hashes[2],
            "entry 2 carries block 2's identity"
        );
        assert_eq!(
            entry.parent_state_root, roots[1],
            "entry 2's parent_state_root is block 1's state root"
        );
        assert_eq!(
            entry.account_trie_diff[0].1,
            Some(vec![0xb0 | 1u8]),
            "block 2's pre-image is the value block 1 wrote"
        );
    }

    /// The bespoke batch path SHALL skip the journal entirely. To actually exercise the
    /// gating we push enough batches to trigger a commit under
    /// `BATCH_COMMIT_THRESHOLD = 4`, then verify no STATE_HISTORY entry materializes
    /// despite the commit happening.
    ///
    /// That path is `wait_for_flush: true` with `commit_depth: Some(BATCH_COMMIT_THRESHOLD)`
    /// — the combination the single `batch_mode` flag used to stand for. Journaling follows
    /// `wait_for_flush`, so the depth-gated per-block paths are not covered here.
    #[test]
    fn journal_skipped_in_batch_mode() {
        let backend: Arc<dyn StorageBackend> = Arc::new(InMemoryBackend::open().unwrap());
        let dir = tempfile::tempdir().unwrap();
        let store = Store::from_backend(
            backend.clone(),
            dir.path().to_path_buf(),
            1,
            DEFAULT_PERSIST_CHANNEL_CAPACITY,
        )
        .unwrap();

        let mut prev_hash = H256::zero();
        for n in 1..=5u64 {
            let state_root = H256::repeat_byte(0xa0 | (n as u8));
            let block = make_block(n, prev_hash, state_root);
            prev_hash = block.hash();
            store
                .store_block_updates(UpdateBatch {
                    account_updates: vec![(
                        Nibbles::from_raw(&[n as u8], false),
                        vec![0xde, 0xad, n as u8],
                    )],
                    storage_updates: vec![],
                    blocks: vec![block],
                    receipts: vec![],
                    code_updates: vec![],
                    commit_depth: Some(BATCH_COMMIT_THRESHOLD),
                    wait_for_flush: true,
                })
                .unwrap();
        }

        for n in 1..=5u64 {
            assert_no_journal_entry(&backend, n);
        }
    }

    /// Storage trie updates SHALL appear in `storage_trie_diff` (not
    /// `account_trie_diff`), with their on-disk keys as written.
    #[test]
    fn journal_storage_updates_appear_in_storage_diff() {
        let backend: Arc<dyn StorageBackend> = Arc::new(InMemoryBackend::open().unwrap());
        let dir = tempfile::tempdir().unwrap();
        let store = Store::from_backend(
            backend.clone(),
            dir.path().to_path_buf(),
            1,
            DEFAULT_PERSIST_CHANNEL_CAPACITY,
        )
        .unwrap();

        let account_hash_a = H256::repeat_byte(0xa0);
        let account_hash_b = H256::repeat_byte(0xb0);

        let state_root_1 = H256::repeat_byte(0x33);
        let block1 = make_block(1, H256::zero(), state_root_1);
        let block1_hash = block1.hash();
        store
            .store_block_updates(UpdateBatch {
                account_updates: vec![],
                storage_updates: vec![
                    (
                        account_hash_a,
                        vec![(Nibbles::from_raw(&[0x05], false), vec![0x01])],
                    ),
                    (
                        account_hash_b,
                        vec![(Nibbles::from_raw(&[0x06], false), vec![0x02])],
                    ),
                ],
                blocks: vec![block1],
                receipts: vec![],
                code_updates: vec![],
                commit_depth: None,
                wait_for_flush: false,
            })
            .unwrap();

        // Advance the safe-commit root to block 1's state root, then store block 2:
        // block 1's layer is now committable and flushes to disk.
        store.set_safe_commit_root(state_root_1).unwrap();
        let state_root_2 = H256::repeat_byte(0x44);
        let block2 = make_block(2, block1_hash, state_root_2);
        store
            .store_block_updates(UpdateBatch {
                account_updates: vec![(Nibbles::from_raw(&[0xee], false), vec![0xff])],
                storage_updates: vec![],
                blocks: vec![block2],
                receipts: vec![],
                code_updates: vec![],
                commit_depth: None,
                wait_for_flush: false,
            })
            .unwrap();

        let bytes = await_journal_entry(&backend, 1, Duration::from_secs(2))
            .expect("STATE_HISTORY entry for block 1");
        let entry = JournalEntry::decode(&bytes).unwrap();
        assert_eq!(entry.block_hash, block1_hash);
        assert_eq!(
            entry.storage_trie_diff.len(),
            2,
            "two distinct account hashes must produce two storage_trie entries"
        );
        for (_key, prev) in &entry.storage_trie_diff {
            assert_eq!(prev, &None, "first-time storage write has None pre-image");
        }
        assert!(entry.account_trie_diff.is_empty());
    }

    fn seed_journal_entries(backend: &Arc<dyn StorageBackend>, block_numbers: &[BlockNumber]) {
        let mut tx = backend.begin_write().unwrap();
        for n in block_numbers {
            let entry = JournalEntry {
                block_hash: H256::repeat_byte(*n as u8),
                parent_state_root: H256::zero(),
                account_trie_diff: vec![(vec![*n as u8], None)],
                storage_trie_diff: vec![],
                account_flat_diff: vec![],
                storage_flat_diff: vec![],
            };
            tx.put(STATE_HISTORY, &n.to_be_bytes(), &entry.encode())
                .unwrap();
        }
        tx.commit().unwrap();
    }

    fn journal_entry_exists(backend: &Arc<dyn StorageBackend>, block_number: BlockNumber) -> bool {
        backend
            .begin_read()
            .unwrap()
            .get(STATE_HISTORY, &block_number.to_be_bytes())
            .unwrap()
            .is_some()
    }

    /// Finality advance SHALL prune every STATE_HISTORY entry at or below the
    /// new finalized number, in the same atomic txn as the finalized-number update.
    #[tokio::test]
    async fn finality_advance_prunes_journal_below_boundary() {
        let backend: Arc<dyn StorageBackend> = Arc::new(InMemoryBackend::open().unwrap());
        let dir = tempfile::tempdir().unwrap();
        let store = Store::from_backend(
            backend.clone(),
            dir.path().to_path_buf(),
            1,
            DEFAULT_PERSIST_CHANNEL_CAPACITY,
        )
        .unwrap();

        seed_journal_entries(&backend, &(1..=10).collect::<Vec<_>>());
        for n in 1..=10 {
            assert!(journal_entry_exists(&backend, n), "seed entry {n} present");
        }

        store
            .forkchoice_update_inner(vec![], 100, H256::zero(), None, Some(5))
            .await
            .unwrap();

        for n in 1..=5 {
            assert!(
                !journal_entry_exists(&backend, n),
                "entry {n} should have been pruned (<= finalized)"
            );
        }
        for n in 6..=10 {
            assert!(
                journal_entry_exists(&backend, n),
                "entry {n} should remain (> finalized)"
            );
        }
    }

    /// Forkchoice updates that don't advance finalized SHALL NOT prune the journal.
    #[tokio::test]
    async fn finality_no_op_does_not_prune() {
        let backend: Arc<dyn StorageBackend> = Arc::new(InMemoryBackend::open().unwrap());
        let dir = tempfile::tempdir().unwrap();
        let store = Store::from_backend(
            backend.clone(),
            dir.path().to_path_buf(),
            1,
            DEFAULT_PERSIST_CHANNEL_CAPACITY,
        )
        .unwrap();

        store
            .forkchoice_update_inner(vec![], 100, H256::zero(), None, Some(5))
            .await
            .unwrap();

        seed_journal_entries(&backend, &(6..=10).collect::<Vec<_>>());

        // FCU re-asserting finalized = 5: must not prune anything.
        store
            .forkchoice_update_inner(vec![], 100, H256::zero(), None, Some(5))
            .await
            .unwrap();

        for n in 6..=10 {
            assert!(
                journal_entry_exists(&backend, n),
                "entry {n} should still exist after no-op finality update"
            );
        }

        // FCU with finalized = None: also a no-op for pruning.
        store
            .forkchoice_update_inner(vec![], 100, H256::zero(), None, None)
            .await
            .unwrap();

        for n in 6..=10 {
            assert!(
                journal_entry_exists(&backend, n),
                "entry {n} should still exist when finalized is None"
            );
        }
    }

    /// A malformed (wrong-length) `FinalizedBlockNumber` value SHALL surface as a
    /// hard error rather than a silent fallback to 0, which would over-prune the
    /// journal on the next FCU.
    #[tokio::test]
    async fn malformed_finalized_returns_error_not_silent_zero() {
        let backend: Arc<dyn StorageBackend> = Arc::new(InMemoryBackend::open().unwrap());
        let dir = tempfile::tempdir().unwrap();
        let store = Store::from_backend(
            backend.clone(),
            dir.path().to_path_buf(),
            1,
            DEFAULT_PERSIST_CHANNEL_CAPACITY,
        )
        .unwrap();

        // Plant a 4-byte (instead of 8-byte) FinalizedBlockNumber value.
        let mut tx = backend.begin_write().unwrap();
        let finalized_key = chain_data_key(ChainDataIndex::FinalizedBlockNumber);
        tx.put(CHAIN_DATA, &finalized_key, &[0u8, 0, 0, 0]).unwrap();
        tx.commit().unwrap();

        seed_journal_entries(&backend, &(1..=5).collect::<Vec<_>>());

        let err = store
            .forkchoice_update_inner(vec![], 100, H256::zero(), None, Some(3))
            .await
            .expect_err("malformed finalized must surface as an error");
        let msg = format!("{err}");
        assert!(
            msg.contains("FinalizedBlockNumber has unexpected length"),
            "expected length-mismatch error, got: {msg}"
        );

        // Journal must not have been pruned.
        for n in 1..=5 {
            assert!(
                journal_entry_exists(&backend, n),
                "entry {n} must NOT be pruned when finalized read failed"
            );
        }
    }

    // -----------------------------------------------------------------------
    // Deep-reorg primitive tests.
    // -----------------------------------------------------------------------

    /// `highest_state_history_block_number` SHALL return the max key present in
    /// `STATE_HISTORY`, or `None` when the table is empty.
    #[tokio::test]
    async fn highest_state_history_block_number_finds_max() {
        let backend: Arc<dyn StorageBackend> = Arc::new(InMemoryBackend::open().unwrap());
        let dir = tempfile::tempdir().unwrap();
        let store = Store::from_backend(
            backend.clone(),
            dir.path().to_path_buf(),
            1,
            DEFAULT_PERSIST_CHANNEL_CAPACITY,
        )
        .unwrap();

        assert_eq!(store.highest_state_history_block_number().unwrap(), None);

        seed_journal_entries(&backend, &[3, 7, 5, 11, 8]);
        assert_eq!(
            store.highest_state_history_block_number().unwrap(),
            Some(11),
            "max over present entries"
        );
    }

    /// The interrupted-batch recovery at startup identifies the block whose state is
    /// on disk from the newest `STATE_HISTORY` entry, so the entry's block hash must
    /// read back exactly as journaled and an absent entry must read as `None`.
    #[tokio::test]
    async fn journaled_block_hash_reads_back() {
        let backend: Arc<dyn StorageBackend> = Arc::new(InMemoryBackend::open().unwrap());
        let dir = tempfile::tempdir().unwrap();
        let store = Store::from_backend(
            backend.clone(),
            dir.path().to_path_buf(),
            1,
            DEFAULT_PERSIST_CHANNEL_CAPACITY,
        )
        .unwrap();

        seed_journal_entries(&backend, &[3, 11]);
        assert_eq!(
            store.get_state_history_block_hash(11).unwrap(),
            Some(H256::repeat_byte(11))
        );
        assert_eq!(store.get_state_history_block_hash(4).unwrap(), None);
    }

    /// `flushed_upto` is a durability floor: `advance_flushed_upto` raises it and
    /// ignores a value at or below the current marker.
    #[tokio::test]
    async fn advance_flushed_upto_only_moves_forward() {
        let backend: Arc<dyn StorageBackend> = Arc::new(InMemoryBackend::open().unwrap());
        let dir = tempfile::tempdir().unwrap();
        let store = Store::from_backend(
            backend,
            dir.path().to_path_buf(),
            1,
            DEFAULT_PERSIST_CHANNEL_CAPACITY,
        )
        .unwrap();

        store.advance_flushed_upto(10).unwrap();
        assert_eq!(store.read_flushed_upto().unwrap(), 10);
        store.advance_flushed_upto(5).unwrap();
        assert_eq!(store.read_flushed_upto().unwrap(), 10);
        store.advance_flushed_upto(12).unwrap();
        assert_eq!(store.read_flushed_upto().unwrap(), 12);
    }

    /// `lowest_state_history_block_number` SHALL return the min key present in
    /// `STATE_HISTORY`, or `None` when the table is empty. Phase 2's cap fallback
    /// depends on this when no finalized hash is known.
    #[tokio::test]
    async fn lowest_state_history_block_number_finds_min() {
        let backend: Arc<dyn StorageBackend> = Arc::new(InMemoryBackend::open().unwrap());
        let dir = tempfile::tempdir().unwrap();
        let store = Store::from_backend(
            backend.clone(),
            dir.path().to_path_buf(),
            1,
            DEFAULT_PERSIST_CHANNEL_CAPACITY,
        )
        .unwrap();

        assert_eq!(store.lowest_state_history_block_number().unwrap(), None);

        seed_journal_entries(&backend, &[7, 3, 11, 5, 8]);
        assert_eq!(
            store.lowest_state_history_block_number().unwrap(),
            Some(3),
            "min over present entries"
        );
    }

    /// `install_overlay_for_reorg` SHALL atomically swap the layer cache for a
    /// fresh empty one with the overlay installed; layer-cache hits in the OLD
    /// cache MUST NOT survive the swap.
    #[tokio::test]
    async fn install_overlay_replaces_cache_with_fresh_one() {
        let backend: Arc<dyn StorageBackend> = Arc::new(InMemoryBackend::open().unwrap());
        let dir = tempfile::tempdir().unwrap();
        let store = Store::from_backend(
            backend.clone(),
            dir.path().to_path_buf(),
            4,
            DEFAULT_PERSIST_CHANNEL_CAPACITY,
        )
        .unwrap();

        seed_journal_entries(&backend, &[3, 4]);

        // Pre-populate the cache with a sentinel layer to verify the swap discards it.
        {
            let mut guard = store.trie_cache.write().unwrap();
            let mut updated = (**guard).clone();
            updated.put_batch(
                H256::zero(),
                H256::repeat_byte(0xff),
                42,
                H256::repeat_byte(0x42),
                vec![],
            );
            *guard = Arc::new(updated);
        }
        assert!(
            store
                .is_state_in_layer_cache(H256::repeat_byte(0xff))
                .unwrap()
        );

        store
            .install_overlay_for_reorg(4, 3, |_| None)
            .expect("install must succeed");

        assert!(
            !store
                .is_state_in_layer_cache(H256::repeat_byte(0xff))
                .unwrap(),
            "old cache layer must be gone after swap"
        );
        let cache = store.trie_cache.read().unwrap().clone();
        assert!(cache.overlay().is_some(), "overlay must be installed");
        assert_eq!(
            cache.commit_threshold(),
            4,
            "fresh cache must inherit threshold"
        );
    }

    /// `install_overlay_for_reorg` SHALL leave the existing cache intact when
    /// overlay construction fails (missing journal entry).
    #[tokio::test]
    async fn install_overlay_failure_preserves_cache() {
        let backend: Arc<dyn StorageBackend> = Arc::new(InMemoryBackend::open().unwrap());
        let dir = tempfile::tempdir().unwrap();
        let store = Store::from_backend(
            backend.clone(),
            dir.path().to_path_buf(),
            4,
            DEFAULT_PERSIST_CHANNEL_CAPACITY,
        )
        .unwrap();

        // Only seed block 5; constructor will walk down through 4 and 3 and fail.
        seed_journal_entries(&backend, &[5]);
        {
            let mut guard = store.trie_cache.write().unwrap();
            let mut updated = (**guard).clone();
            updated.put_batch(
                H256::zero(),
                H256::repeat_byte(0xaa),
                7,
                H256::repeat_byte(0x77),
                vec![],
            );
            *guard = Arc::new(updated);
        }
        assert!(
            store
                .is_state_in_layer_cache(H256::repeat_byte(0xaa))
                .unwrap()
        );

        let err = store
            .install_overlay_for_reorg(5, 3, |_| None)
            .expect_err("missing entries 3 and 4 must abort");
        let msg = format!("{err}");
        assert!(
            msg.contains("overlay construction failed"),
            "error should explain construction failure: {msg}"
        );

        // Cache must be intact: the sentinel layer must still be there.
        assert!(
            store
                .is_state_in_layer_cache(H256::repeat_byte(0xaa))
                .unwrap(),
            "cache must survive a failed overlay install"
        );
        let cache = store.trie_cache.read().unwrap().clone();
        assert!(cache.overlay().is_none(), "no overlay should be installed");
    }

    /// `abort_reorg` SHALL discard both the overlay and any partial new-chain
    /// layers, restoring the cache to a fresh empty state with the same
    /// commit threshold.
    #[tokio::test]
    async fn abort_reorg_resets_cache_to_fresh() {
        let backend: Arc<dyn StorageBackend> = Arc::new(InMemoryBackend::open().unwrap());
        let dir = tempfile::tempdir().unwrap();
        let store = Store::from_backend(
            backend.clone(),
            dir.path().to_path_buf(),
            4,
            DEFAULT_PERSIST_CHANNEL_CAPACITY,
        )
        .unwrap();

        seed_journal_entries(&backend, &[3, 4]);
        store.install_overlay_for_reorg(4, 3, |_| None).unwrap();
        // Simulate partial new-chain layer.
        {
            let mut guard = store.trie_cache.write().unwrap();
            let mut updated = (**guard).clone();
            updated.put_batch(
                H256::zero(),
                H256::repeat_byte(0xbb),
                4,
                H256::repeat_byte(0xb4),
                vec![],
            );
            *guard = Arc::new(updated);
        }
        assert!(
            store
                .is_state_in_layer_cache(H256::repeat_byte(0xbb))
                .unwrap()
        );

        store.abort_reorg().unwrap();

        assert!(
            !store
                .is_state_in_layer_cache(H256::repeat_byte(0xbb))
                .unwrap(),
            "partial new-chain layer must be discarded"
        );
        let cache = store.trie_cache.read().unwrap().clone();
        assert!(cache.overlay().is_none(), "overlay must be cleared");
        assert_eq!(
            cache.commit_threshold(),
            4,
            "fresh cache must inherit threshold"
        );
    }

    /// `clear_reorg_overlay` SHALL remove an installed overlay and SHALL be
    /// idempotent (safe to call when no overlay is installed).
    #[tokio::test]
    async fn clear_reorg_overlay_removes_overlay_and_is_idempotent() {
        let backend: Arc<dyn StorageBackend> = Arc::new(InMemoryBackend::open().unwrap());
        let dir = tempfile::tempdir().unwrap();
        let store = Store::from_backend(
            backend.clone(),
            dir.path().to_path_buf(),
            4,
            DEFAULT_PERSIST_CHANNEL_CAPACITY,
        )
        .unwrap();

        // Idempotent when no overlay is installed.
        store.clear_reorg_overlay().unwrap();
        assert!(store.trie_cache.read().unwrap().overlay().is_none());

        seed_journal_entries(&backend, &[7]);
        store.install_overlay_for_reorg(7, 7, |_| None).unwrap();
        assert!(store.trie_cache.read().unwrap().overlay().is_some());

        store.clear_reorg_overlay().unwrap();
        assert!(store.trie_cache.read().unwrap().overlay().is_none());

        // Calling again is still a no-op.
        store.clear_reorg_overlay().unwrap();
        assert!(store.trie_cache.read().unwrap().overlay().is_none());
    }

    /// Regression test: journal entries written while the FKV generator
    /// was running lack flat-KV pre-images for keys past the generator frontier,
    /// and the generator's own flat-KV writes are never journaled. While an
    /// overlay serves the read's state root, every flat-KV fast path must be
    /// disabled so reads fall back to the (always journaled) trie nodes;
    /// otherwise they serve the generator's stale values for the chain being
    /// reorged away. Roots the overlay does not serve must keep the fast path.
    #[tokio::test]
    async fn deep_reorg_overlay_disables_flat_kv_fast_paths() {
        let backend: Arc<dyn StorageBackend> = Arc::new(InMemoryBackend::open().unwrap());

        // Simulate a finished FKV generation BEFORE the Store boots, so its
        // in-memory frontier starts at [0xff; 64] and every path counts as
        // "computed" (the Store expands the durable [0xff] sentinel at open).
        {
            let mut tx = backend.begin_write().unwrap();
            tx.put(MISC_VALUES, "last_written".as_bytes(), &[0xff])
                .unwrap();
            tx.commit().unwrap();
        }

        let dir = tempfile::tempdir().unwrap();
        let store = Store::from_backend(
            backend.clone(),
            dir.path().to_path_buf(),
            1,
            DEFAULT_PERSIST_CHANNEL_CAPACITY,
        )
        .unwrap();

        // The pivot state is empty, so the correct value for every read at the
        // pivot root is None. The old chain created the account/slot after the
        // pivot; the generator then wrote their flat-KV leaves against the old
        // chain — writes that are never journaled.
        let address = Address::from_low_u64_be(0xbeef);
        let account_hash = hash_address_fixed(&address);
        let slot = H256::from_low_u64_be(1);
        let slot_hash = hash_key_fixed(&slot);
        let generator_account = AccountState {
            nonce: 7,
            balance: U256::from(999u64),
            storage_root: H256::zero(),
            code_hash: H256::zero(),
        };
        let generator_slot_value = U256::from(42u64);
        {
            let mut tx = backend.begin_write().unwrap();
            tx.put(
                ACCOUNT_FLATKEYVALUE,
                &Nibbles::from_bytes(account_hash.as_bytes()).into_vec(),
                &generator_account.encode_to_vec(),
            )
            .unwrap();
            tx.put(
                STORAGE_FLATKEYVALUE,
                &apply_prefix(Some(account_hash), Nibbles::from_bytes(&slot_hash)).into_vec(),
                &generator_slot_value.encode_to_vec(),
            )
            .unwrap();
            tx.commit().unwrap();
        }

        // Journal entries for blocks 10 and 11 above the pivot (block 9), as
        // written DURING generation: flat diffs are empty because the keys were
        // past the frontier at commit time (the incomplete-journal hole). The pivot's state
        // root is the empty trie.
        let pivot_root = EMPTY_TRIE_HASH;
        {
            let mut tx = backend.begin_write().unwrap();
            for (n, parent_root) in [(10u64, pivot_root), (11, H256::repeat_byte(0x10))] {
                let entry = JournalEntry {
                    block_hash: H256::repeat_byte(n as u8),
                    parent_state_root: parent_root,
                    account_trie_diff: vec![],
                    storage_trie_diff: vec![],
                    account_flat_diff: vec![],
                    storage_flat_diff: vec![],
                };
                tx.put(STATE_HISTORY, &n.to_be_bytes(), &entry.encode())
                    .unwrap();
            }
            tx.commit().unwrap();
        }

        store
            .install_overlay_for_reorg(11, 10, |_| None)
            .expect("overlay install must succeed");

        // Reads at the pivot root must return the pivot state (account and slot
        // absent), NOT the generator's stale flat-KV values.
        let accounts = store
            .get_account_states_batch_by_root(pivot_root, &[address])
            .unwrap();
        assert_eq!(
            accounts,
            vec![None],
            "batch account read must not serve the generator's stale flat-KV value"
        );

        let slot_value = store
            .get_storage_at_root(pivot_root, address, slot)
            .unwrap();
        assert_eq!(
            slot_value, None,
            "storage read must not serve the generator's stale flat-KV value"
        );

        let batched_slots = store
            .get_storage_values_batch_by_root(
                pivot_root,
                &[(hash_address_fixed(&address), EMPTY_TRIE_HASH, slot)],
            )
            .unwrap();
        assert_eq!(
            batched_slots,
            vec![None],
            "batch storage read must not serve the generator's stale flat-KV value"
        );

        let state_trie = store.open_state_trie(pivot_root).unwrap();
        assert_eq!(
            state_trie.get(account_hash.as_bytes()).unwrap(),
            None,
            "trie read must not serve the generator's stale flat-KV value"
        );

        // Outside the overlay window the fast paths are untouched: a root the
        // overlay does not serve still reads flat-KV straight from disk.
        let unserved_root = H256::repeat_byte(0x99);
        let accounts = store
            .get_account_states_batch_by_root(unserved_root, &[address])
            .unwrap();
        assert_eq!(
            accounts,
            vec![Some(generator_account)],
            "unserved roots must keep the flat-KV fast path"
        );
        let slot_value = store
            .get_storage_at_root(unserved_root, address, slot)
            .unwrap();
        assert_eq!(
            slot_value,
            Some(generator_slot_value),
            "unserved roots must keep the flat-KV fast path"
        );

        let batched_slots = store
            .get_storage_values_batch_by_root(
                unserved_root,
                &[(hash_address_fixed(&address), EMPTY_TRIE_HASH, slot)],
            )
            .unwrap();
        assert_eq!(
            batched_slots,
            vec![Some(generator_slot_value)],
            "unserved roots must keep the flat-KV fast path"
        );
    }

    /// Adversarial regression for the multi-layer reconciliation finding: with an
    /// overlay installed, a commit targeting a root SEVERAL layers above the pivot
    /// must commit ONLY the bottom layer (the pivot tip `T`) per pass. Before the
    /// fix this swept every layer at once: the `debug_assert!(len == 1)` panicked in
    /// debug builds, and in release the upper layers' journal entries recorded
    /// pre-images against the OLD-chain disk state instead of the new-chain/bridge
    /// state — silent corruption for a future unwind.
    #[tokio::test]
    async fn overlay_backed_commit_only_commits_bottom_layer_per_pass() {
        let backend: Arc<dyn StorageBackend> = Arc::new(InMemoryBackend::open().unwrap());
        let dir = tempfile::tempdir().unwrap();
        let store = Store::from_backend(
            backend.clone(),
            dir.path().to_path_buf(),
            4,
            DEFAULT_PERSIST_CHANNEL_CAPACITY,
        )
        .unwrap();

        // Pivot at block 9's root; journal entries for blocks 10 and 11 (the old
        // chain above the pivot). The overlay serves the pivot root.
        let pivot_root = H256::repeat_byte(0x09);
        {
            let mut tx = backend.begin_write().unwrap();
            for (n, parent_root) in [(10u64, pivot_root), (11, H256::repeat_byte(0x0a))] {
                let entry = JournalEntry {
                    block_hash: H256::repeat_byte(n as u8),
                    parent_state_root: parent_root,
                    account_trie_diff: vec![(vec![0x00, n as u8], None)],
                    storage_trie_diff: vec![],
                    account_flat_diff: vec![],
                    storage_flat_diff: vec![],
                };
                tx.put(STATE_HISTORY, &n.to_be_bytes(), &entry.encode())
                    .unwrap();
            }
            tx.commit().unwrap();
        }
        store.install_overlay_for_reorg(11, 10, |_| None).unwrap();

        // Three new-chain layers above the pivot: blocks 10, 11, 12.
        let roots: Vec<H256> = (10..=12u8).map(H256::repeat_byte).collect();
        {
            let mut guard = store.trie_cache.write().unwrap();
            let mut updated = (**guard).clone();
            let mut parent = pivot_root;
            for (i, root) in roots.iter().enumerate() {
                let n = 10 + i as u64;
                updated.put_batch(
                    parent,
                    *root,
                    n,
                    H256::repeat_byte(n as u8),
                    vec![(Nibbles::from_raw(&[0x01, n as u8], false), vec![n as u8])],
                );
                parent = *root;
            }
            *guard = Arc::new(updated);
        }

        // Commit targeting the TOP layer (3 deep). With the overlay installed this
        // MUST reduce to the bottom layer only.
        let trie = store.trie_cache.read().unwrap().clone();
        commit_to_disk(
            store.backend.as_ref(),
            &store.flatkeyvalue_control_tx,
            &store.trie_cache,
            &trie,
            roots[2],
            false,
        )
        .expect("bottom-layer commit must succeed (no debug_assert panic)");

        // Only the bottom layer (block 10) is committed and pruned; 11 and 12 stay resident.
        assert!(
            !store.is_state_in_layer_cache(roots[0]).unwrap(),
            "bottom layer must be committed and pruned"
        );
        assert!(
            store.is_state_in_layer_cache(roots[1]).unwrap()
                && store.is_state_in_layer_cache(roots[2]).unwrap(),
            "upper layers must remain resident after the bottom-layer-only commit"
        );
        // The overlay was consumed by the reconciliation commit.
        assert!(
            store.trie_cache.read().unwrap().overlay().is_none(),
            "reconciliation must clear the overlay"
        );
        // Journal: an entry exists for block 10, none yet for 11/12.
        assert!(
            journal_entry_exists(&backend, 10),
            "bottom layer T must be journaled"
        );
        assert!(!journal_entry_exists(&backend, 11));
        assert!(!journal_entry_exists(&backend, 12));

        // The reconciliation fold MUST land in the bottom layer's journal entry:
        // the overlay's bridge keys (untouched by the layer) appear with their
        // old-chain pre-images. This is what block-number-keyed reconciliation
        // matching loses when the bottom layer sits above `to_block`.
        let entry_bytes = backend
            .begin_read()
            .unwrap()
            .get(STATE_HISTORY, &10u64.to_be_bytes())
            .unwrap()
            .expect("journal entry for block 10");
        let entry = JournalEntry::decode(&entry_bytes).unwrap();
        let diff_keys: Vec<&Vec<u8>> = entry.account_trie_diff.iter().map(|(k, _)| k).collect();
        assert!(
            diff_keys.contains(&&vec![0x00, 10]) && diff_keys.contains(&&vec![0x00, 11]),
            "bridge keys from the overlay must be folded into T's journal entry, got {diff_keys:?}"
        );

        // Second pass: with the overlay gone the remaining backlog commits normally.
        let trie = store.trie_cache.read().unwrap().clone();
        commit_to_disk(
            store.backend.as_ref(),
            &store.flatkeyvalue_control_tx,
            &store.trie_cache,
            &trie,
            roots[2],
            false,
        )
        .unwrap();
        assert!(!store.is_state_in_layer_cache(roots[1]).unwrap());
        assert!(!store.is_state_in_layer_cache(roots[2]).unwrap());
        assert!(journal_entry_exists(&backend, 11));
        assert!(journal_entry_exists(&backend, 12));
    }

    /// Adversarial regression for the journal-pruning race: while a deep-reorg
    /// apply pass holds the pause flag, a finality advance must NOT prune
    /// STATE_HISTORY (a concurrent `Overlay::from_journal` reads entries with no
    /// snapshot isolation). Pruning must catch up once the pause is released.
    #[tokio::test]
    async fn journal_pruning_pauses_during_reorg_and_catches_up_after() {
        let backend: Arc<dyn StorageBackend> = Arc::new(InMemoryBackend::open().unwrap());
        let dir = tempfile::tempdir().unwrap();
        let store = Store::from_backend(
            backend.clone(),
            dir.path().to_path_buf(),
            1,
            DEFAULT_PERSIST_CHANNEL_CAPACITY,
        )
        .unwrap();

        seed_journal_entries(&backend, &(1..=5).collect::<Vec<_>>());

        // Paused: finality advance to 3 must not prune anything.
        store.set_journal_pruning_paused(true);
        store
            .forkchoice_update_inner(vec![], 100, H256::zero(), None, Some(3))
            .await
            .unwrap();
        for n in 1..=5 {
            assert!(
                journal_entry_exists(&backend, n),
                "entry {n} must survive pruning while the reorg pause is held"
            );
        }

        // Released: the next advance prunes cumulatively from zero.
        store.set_journal_pruning_paused(false);
        store
            .forkchoice_update_inner(vec![], 100, H256::zero(), None, Some(4))
            .await
            .unwrap();
        for n in 1..=4 {
            assert!(
                !journal_entry_exists(&backend, n),
                "entry {n} must be pruned once the pause is released"
            );
        }
        assert!(journal_entry_exists(&backend, 5));
    }
}

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

    fn h256(b: u8) -> H256 {
        H256::from_low_u64_be(b as u64)
    }

    fn op(bn: BlockNumber, bh: H256, idx: Index) -> Vec<u8> {
        encode_tx_location_operand(bn, bh, idx)
    }

    fn decode(v: &[u8]) -> Vec<(BlockNumber, BlockHash, Index)> {
        <Vec<(BlockNumber, BlockHash, Index)>>::decode(v).unwrap()
    }

    #[test]
    fn single_operand_on_empty_base() {
        let out = tx_locations_merge(None, vec![op(100, h256(0x10), 0)]).unwrap();
        assert_eq!(decode(&out), vec![(100, h256(0x10), 0)]);
    }

    #[test]
    fn operand_appended_to_existing_base() {
        let base = vec![(100u64, h256(0x10), 0u64)].encode_to_vec();
        let out = tx_locations_merge(Some(&base), vec![op(101, h256(0x11), 5)]).unwrap();
        let mut got = decode(&out);
        got.sort();
        let mut want = vec![(100, h256(0x10), 0), (101, h256(0x11), 5)];
        want.sort();
        assert_eq!(got, want);
    }

    #[test]
    fn multiple_operands_combined() {
        let out = tx_locations_merge(
            None,
            vec![
                op(100, h256(0x10), 0),
                op(100, h256(0x11), 1),
                op(101, h256(0x12), 2),
            ],
        )
        .unwrap();
        assert_eq!(decode(&out).len(), 3);
    }

    #[test]
    fn same_block_hash_is_deduped() {
        // Two operands with the same block_hash: the later one replaces the earlier.
        let out =
            tx_locations_merge(None, vec![op(100, h256(0x10), 0), op(100, h256(0x10), 7)]).unwrap();
        assert_eq!(decode(&out), vec![(100, h256(0x10), 7)]);
    }

    #[test]
    fn malformed_operand_aborts_merge() {
        // Fail loud: a malformed operand must abort the merge (return None), not
        // silently drop it and commit a partial result.
        let out = tx_locations_merge(None, vec![vec![0xff, 0xff], op(100, h256(0x10), 0)]);
        assert!(out.is_none(), "merge must abort on a malformed operand");
    }

    #[test]
    fn malformed_base_value_aborts_merge() {
        let out = tx_locations_merge(Some(&[0xff, 0xff]), vec![op(100, h256(0x10), 0)]);
        assert!(out.is_none(), "merge must abort on a corrupt base value");
    }

    /// Regression for the associative-merge format bug: a PartialMerge result
    /// must be re-mergeable as an operand. RocksDB folds operands together
    /// without a base value during compaction, then feeds that result back into
    /// a later merge. If the operand format differed from the output format,
    /// the re-fed result would fail to decode and entries would be dropped
    /// (observed as 1664 silent drops during a compaction pass on mainnet).
    #[test]
    fn partial_merge_result_is_a_valid_operand() {
        // Step 1: PartialMerge — combine operands with NO base value.
        let partial =
            tx_locations_merge(None, vec![op(100, h256(0x10), 0), op(101, h256(0x11), 1)]).unwrap();

        // Step 2: the partial result is now itself an operand in a later merge,
        // on top of an existing base value. This is the path that used to drop
        // entries.
        let base = vec![(99u64, h256(0x09), 9u64)].encode_to_vec();
        let out = tx_locations_merge(Some(&base), vec![partial]).unwrap();

        let mut got = decode(&out);
        got.sort();
        let mut want = vec![
            (99, h256(0x09), 9),
            (100, h256(0x10), 0),
            (101, h256(0x11), 1),
        ];
        want.sort();
        assert_eq!(
            got, want,
            "no entries may be lost when re-merging a partial result"
        );
    }

    /// Operand and stored-value encodings must be byte-identical types, so a
    /// freshly-encoded operand round-trips through the value decoder.
    #[test]
    fn operand_encoding_matches_value_encoding() {
        let operand = op(100, h256(0x10), 3);
        // Decoding the operand as the stored Vec type must succeed.
        assert_eq!(decode(&operand), vec![(100, h256(0x10), 3)]);
    }

    /// Chained PartialMerges (operand-only folds applied repeatedly) stay valid.
    #[test]
    fn chained_partial_merges() {
        let p1 = tx_locations_merge(None, vec![op(1, h256(0x01), 0)]).unwrap();
        let p2 = tx_locations_merge(None, vec![p1, op(2, h256(0x02), 0)]).unwrap();
        let p3 = tx_locations_merge(None, vec![p2, op(3, h256(0x03), 0)]).unwrap();
        let out = tx_locations_merge(None, vec![p3]).unwrap();
        assert_eq!(decode(&out).len(), 3);
    }
}

#[cfg(test)]
mod backfill_write_tests {
    use super::*;
    use ethrex_common::types::{EIP1559Transaction, TxType};

    fn test_header(number: BlockNumber) -> BlockHeader {
        BlockHeader {
            number,
            ..Default::default()
        }
    }

    fn test_tx(nonce: u64) -> Transaction {
        Transaction::EIP1559Transaction(EIP1559Transaction {
            nonce,
            ..Default::default()
        })
    }

    fn test_receipt() -> Receipt {
        Receipt {
            tx_type: TxType::EIP1559,
            succeeded: true,
            cumulative_gas_used: 21_000,
            logs: vec![],
            payer: None,
            frame_receipts: None,
        }
    }

    /// `add_backfilled_blocks` writes the body + receipts + transaction index and
    /// lowers the frontier (`earliest_block_number`) in one shot; the indexed tx
    /// is then locatable for its (canonical) block.
    #[tokio::test]
    async fn writes_data_indexes_txs_and_lowers_frontier() {
        let store = Store::new("", EngineType::InMemory).expect("in-memory store");
        store.update_earliest_block_number(100).await.unwrap();

        let header = test_header(99);
        let hash = header.hash();
        let tx = test_tx(7);
        let body = BlockBody {
            transactions: vec![tx.clone()],
            ommers: vec![],
            withdrawals: Some(vec![]),
        };
        // In production the header is already stored and canonical from snap
        // header backfill; mirror the canonical mapping so the tx-index lookup
        // (which is canonical-filtered) can resolve.
        store.add_block_headers(vec![header.clone()]).await.unwrap();
        store
            .forkchoice_update(vec![(99, hash)], 99, hash, None, None)
            .await
            .unwrap();

        store
            .add_backfilled_blocks(
                vec![BackfilledBlock {
                    header,
                    body,
                    receipts: vec![test_receipt()],
                    index_transactions: true,
                }],
                99,
            )
            .await
            .unwrap();

        assert_eq!(
            store.get_earliest_block_number().await.unwrap(),
            99,
            "frontier must be lowered atomically with the data write"
        );
        assert!(
            store.get_block_body_by_hash(hash).await.unwrap().is_some(),
            "body must be stored"
        );
        assert_eq!(
            store.get_receipts_for_block(&hash).await.unwrap().len(),
            1,
            "receipts must be stored"
        );
        assert_eq!(
            store
                .get_transaction_location(tx.hash(&NativeCrypto))
                .await
                .unwrap(),
            Some((99, hash, 0)),
            "indexed transaction must be locatable"
        );
    }

    /// With `index_transactions = false` (a block outside the
    /// `--history.transactions` horizon) the body and receipts are still written,
    /// but the transaction index is skipped.
    #[tokio::test]
    async fn skips_tx_index_when_disabled() {
        let store = Store::new("", EngineType::InMemory).expect("in-memory store");
        store.update_earliest_block_number(100).await.unwrap();

        let header = test_header(99);
        let hash = header.hash();
        let tx = test_tx(7);
        let body = BlockBody {
            transactions: vec![tx.clone()],
            ommers: vec![],
            withdrawals: Some(vec![]),
        };
        store.add_block_headers(vec![header.clone()]).await.unwrap();
        store
            .forkchoice_update(vec![(99, hash)], 99, hash, None, None)
            .await
            .unwrap();

        store
            .add_backfilled_blocks(
                vec![BackfilledBlock {
                    header,
                    body,
                    receipts: vec![test_receipt()],
                    index_transactions: false,
                }],
                99,
            )
            .await
            .unwrap();

        assert!(
            store.get_block_body_by_hash(hash).await.unwrap().is_some(),
            "body must still be stored"
        );
        assert_eq!(
            store.get_receipts_for_block(&hash).await.unwrap().len(),
            1,
            "receipts must still be stored"
        );
        assert_eq!(
            store
                .get_transaction_location(tx.hash(&NativeCrypto))
                .await
                .unwrap(),
            None,
            "transaction index must be skipped when index_transactions is false"
        );
    }

    /// End-to-end at the storage/RPC boundary: the getters the RPC handlers use
    /// (`get_block_body`, `get_transaction_by_hash`, `get_receipt`) return nothing
    /// for a headers-only (post-snap) block and real data after backfill.
    #[tokio::test]
    async fn historical_getters_transition_from_null_to_present() {
        let store = Store::new("", EngineType::InMemory).expect("in-memory store");
        store.update_earliest_block_number(100).await.unwrap();

        let header = test_header(99);
        let hash = header.hash();
        let tx = test_tx(7);
        let tx_hash = tx.hash(&NativeCrypto);
        let body = BlockBody {
            transactions: vec![tx],
            ommers: vec![],
            withdrawals: Some(vec![]),
        };
        // Post-snap state: header stored and canonical, but no body/receipts.
        store.add_block_headers(vec![header.clone()]).await.unwrap();
        store
            .forkchoice_update(vec![(99, hash)], 99, hash, None, None)
            .await
            .unwrap();

        // Previously-null historical queries.
        assert!(
            store.get_block_body(99).await.unwrap().is_none(),
            "body absent before backfill"
        );
        assert!(
            store
                .get_transaction_by_hash(tx_hash)
                .await
                .unwrap()
                .is_none(),
            "tx unlocatable before backfill"
        );
        assert!(
            store.get_receipt(99, 0).await.unwrap().is_none(),
            "receipt absent before backfill"
        );

        store
            .add_backfilled_blocks(
                vec![BackfilledBlock {
                    header,
                    body,
                    receipts: vec![test_receipt()],
                    index_transactions: true,
                }],
                99,
            )
            .await
            .unwrap();

        // The same queries now return real data.
        assert!(
            store.get_block_body(99).await.unwrap().is_some(),
            "body present after backfill"
        );
        assert!(
            store
                .get_transaction_by_hash(tx_hash)
                .await
                .unwrap()
                .is_some(),
            "tx locatable after backfill"
        );
        assert!(
            store.get_receipt(99, 0).await.unwrap().is_some(),
            "receipt present after backfill"
        );
    }

    /// Backfill fills downward in batches; the frontier lowers after each batch,
    /// persists (doubling as the resume cursor), and leaves no gaps.
    #[tokio::test]
    async fn resume_fills_contiguously_without_gaps() {
        let store = Store::new("", EngineType::InMemory).expect("in-memory store");
        store.update_earliest_block_number(100).await.unwrap();

        let make_batch = |range: std::ops::Range<u64>| -> Vec<BackfilledBlock> {
            range
                .rev()
                .map(|n| BackfilledBlock {
                    header: test_header(n),
                    body: BlockBody {
                        transactions: vec![],
                        ommers: vec![],
                        withdrawals: Some(vec![]),
                    },
                    receipts: vec![],
                    index_transactions: false,
                })
                .collect()
        };

        // Batch 1: fill [90, 99], lowering the frontier to 90.
        store
            .add_backfilled_blocks(make_batch(90..100), 90)
            .await
            .unwrap();
        assert_eq!(store.get_earliest_block_number().await.unwrap(), 90);

        // Simulate a restart: the task resumes from the persisted frontier.
        let resume_from = store.get_earliest_block_number().await.unwrap();
        assert_eq!(resume_from, 90, "resume reads the persisted frontier");

        // Batch 2: fill [80, 89], lowering the frontier to 80.
        store
            .add_backfilled_blocks(make_batch(80..90), 80)
            .await
            .unwrap();
        assert_eq!(store.get_earliest_block_number().await.unwrap(), 80);

        // No gaps: every block in [80, 99] has a stored body.
        for n in 80..100 {
            assert!(
                store
                    .get_block_body_by_hash(test_header(n).hash())
                    .await
                    .unwrap()
                    .is_some(),
                "block {n} must have a body (no gap left by resume)"
            );
        }
    }
}

/// The schema-version guard runs before any backend is opened, so these tests need
/// a persistent `EngineType` but never touch RocksDB itself.
#[cfg(test)]
#[cfg(feature = "rocksdb")]
mod schema_version_guard_tests {
    use super::*;

    fn write_marker(dir: &Path, schema_version: u64) {
        let metadata = StoreMetadata::new(schema_version);
        std::fs::write(
            dir.join(STORE_METADATA_FILENAME),
            serde_json::to_string_pretty(&metadata).unwrap(),
        )
        .unwrap();
    }

    fn read_marker(dir: &Path) -> u64 {
        read_store_schema_version(dir).unwrap().unwrap()
    }

    #[test]
    fn a_database_ahead_of_the_binary_is_refused_and_left_untouched() {
        // The scenario behind this test: a datadir opened once by a newer build
        // (which stamps its own schema version) and then handed back to an older
        // one. The older build must refuse with a version error — not a
        // "migration failed" — and must not rewrite the marker, so the newer
        // build can still open the database.
        let dir = tempfile::tempdir().unwrap();
        let newer = STORE_SCHEMA_VERSION + 1;
        write_marker(dir.path(), newer);

        let result =
            Store::new_with_config(dir.path(), EngineType::RocksDB, StoreConfig::default());

        assert!(
            matches!(
                result,
                Err(StoreError::IncompatibleDBVersion { found, expected })
                    if found == newer && expected == STORE_SCHEMA_VERSION
            ),
            "expected IncompatibleDBVersion, got {:?}",
            result.err()
        );
        assert_eq!(read_marker(dir.path()), newer);
    }

    #[test]
    fn a_zero_schema_version_is_incompatible_not_a_failed_migration() {
        // No ethrex ever writes v0; the marker is corrupt. Nothing was migrated,
        // so the error must not claim a migration failed.
        let dir = tempfile::tempdir().unwrap();
        write_marker(dir.path(), 0);

        let result =
            Store::new_with_config(dir.path(), EngineType::RocksDB, StoreConfig::default());

        assert!(
            matches!(
                result,
                Err(StoreError::IncompatibleDBVersion { found: 0, expected })
                    if expected == STORE_SCHEMA_VERSION
            ),
            "expected IncompatibleDBVersion, got {:?}",
            result.err()
        );
    }
}

#[cfg(test)]
mod datadir_tests {
    use super::*;
    use std::fs;

    #[test]
    fn empty_dir_has_no_existing_db() {
        let dir = tempfile::tempdir().unwrap();
        assert!(!dir_contains_legacy_db(dir.path()).unwrap());
    }

    #[test]
    fn dir_with_only_unrelated_files_has_no_existing_db() {
        // Regression for #5680: a JWT secret (or any unrelated file) in the
        // datadir must not be mistaken for an existing database.
        let dir = tempfile::tempdir().unwrap();
        fs::write(dir.path().join("jwt.hex"), "0xdeadbeef").unwrap();
        fs::write(dir.path().join("LOG"), "noise").unwrap();
        assert!(!dir_contains_legacy_db(dir.path()).unwrap());
    }

    #[test]
    fn dir_with_rocksdb_markers_has_existing_db() {
        // A `CURRENT` file (and, separately, a `MANIFEST-*` file) marks a real DB.
        let dir = tempfile::tempdir().unwrap();
        fs::write(dir.path().join("CURRENT"), "MANIFEST-000001\n").unwrap();
        assert!(dir_contains_legacy_db(dir.path()).unwrap());

        let dir2 = tempfile::tempdir().unwrap();
        fs::write(dir2.path().join("MANIFEST-000007"), "x").unwrap();
        assert!(dir_contains_legacy_db(dir2.path()).unwrap());
    }

    #[test]
    fn dir_with_marker_named_subdirectories_has_no_existing_db() {
        // A *directory* named like a marker file must not be mistaken for a DB;
        // RocksDB only ever visits these as plain files.
        let dir = tempfile::tempdir().unwrap();
        fs::create_dir(dir.path().join("CURRENT")).unwrap();
        fs::create_dir(dir.path().join("MANIFEST-000001")).unwrap();
        assert!(!dir_contains_legacy_db(dir.path()).unwrap());
    }
}

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

    /// The `last_written` marker signals a finished FKV pass ONLY as the exact 1-byte
    /// `[0xff]` sentinel. This gates journal-backed deep reorgs, so the
    /// boundary must be exact: an unset marker, the initial all-zero frontier, and any
    /// mid-generation nibble path must all read as "not complete".
    #[test]
    fn only_the_one_byte_ff_sentinel_counts_as_complete() {
        // Complete: the durable completion sentinel the generator writes.
        assert!(Store::flatkeyvalue_generation_complete(Some(&[0xff])));

        // Not complete:
        assert!(!Store::flatkeyvalue_generation_complete(None)); // marker never written
        assert!(!Store::flatkeyvalue_generation_complete(Some(&[]))); // empty
        assert!(!Store::flatkeyvalue_generation_complete(Some(&[0u8; 64]))); // initial frontier
        assert!(!Store::flatkeyvalue_generation_complete(Some(&[0x0a; 64]))); // mid-gen nibble path
        // The in-memory frontier is expanded to all-0xff (64/131 bytes), but that is never
        // the durable marker; only the 1-byte form means complete, so these read false.
        assert!(!Store::flatkeyvalue_generation_complete(Some(&[0xff; 64])));
        assert!(!Store::flatkeyvalue_generation_complete(Some(&[
            0xff, 0xff
        ])));
    }
}

#[cfg(test)]
mod cgroup_memory_limit_tests {
    use super::*;
    use std::fs;

    /// Builds `<root>/<relative>` and writes `contents` into the `file` of every
    /// directory the map names, keyed by path relative to `root` ("" is the root).
    fn cgroup_tree(root: &Path, relative: &str, file: &str, limits: &[(&str, &str)]) {
        fs::create_dir_all(root.join(relative)).expect("create cgroup tree");
        for (dir, contents) in limits {
            fs::write(root.join(dir).join(file), contents).expect("write limit");
        }
    }

    /// A limit set on an ancestor cgroup (a systemd slice, a pod's parent) applies to
    /// the process, so it must be found even though the process's own cgroup is
    /// unlimited and the mount root reports `max`.
    #[test]
    fn ancestor_limit_applies() {
        let tmp = tempfile::tempdir().expect("tempdir");
        let root = tmp.path();
        cgroup_tree(
            root,
            "system.slice/ethrex.service",
            "memory.max",
            &[("", "max"), ("system.slice", "4294967296")],
        );

        assert_eq!(
            cgroup_limit_up_to_root(root, "/system.slice/ethrex.service", "memory.max"),
            Some(4 * 1024 * 1024 * 1024)
        );
    }

    /// With limits at several levels the tightest one is what the kernel enforces.
    #[test]
    fn smallest_limit_in_the_chain_wins() {
        let tmp = tempfile::tempdir().expect("tempdir");
        let root = tmp.path();
        cgroup_tree(
            root,
            "kubepods/pod123/container",
            "memory.max",
            &[
                ("kubepods", "8589934592"),
                ("kubepods/pod123", "2147483648"),
                ("kubepods/pod123/container", "4294967296"),
            ],
        );

        assert_eq!(
            cgroup_limit_up_to_root(root, "kubepods/pod123/container", "memory.max"),
            Some(2 * 1024 * 1024 * 1024)
        );
    }

    /// An unlimited chain is "unknown", not zero: v2 writes the literal `max` and v1 a
    /// sentinel near `u64::MAX`, and absent files must not count either.
    #[test]
    fn unlimited_chain_yields_no_limit() {
        let tmp = tempfile::tempdir().expect("tempdir");
        let root = tmp.path();
        cgroup_tree(root, "docker/abc", "memory.max", &[("", "max")]);

        assert_eq!(
            cgroup_limit_up_to_root(root, "docker/abc", "memory.max"),
            None
        );
        // A cgroup path that does not exist under this mount (a v1 controller that is
        // not mounted, a stale line in /proc/self/cgroup) reads as unknown too.
        assert_eq!(
            cgroup_limit_up_to_root(root, "not/here", "memory.max"),
            None
        );
    }

    /// The root cgroup is itself a level: a limit set there still applies.
    #[test]
    fn root_limit_applies_to_the_root_cgroup() {
        let tmp = tempfile::tempdir().expect("tempdir");
        let root = tmp.path();
        cgroup_tree(root, "", "memory.limit_in_bytes", &[("", "536870912")]);

        assert_eq!(
            cgroup_limit_up_to_root(root, "/", "memory.limit_in_bytes"),
            Some(512 * 1024 * 1024)
        );
    }
}