zeppelin_embed/lifecycle/

mod.rs

//! Store lifecycle and memory accounting.

mod budget;
mod cancel;
mod cascade;
mod clock;
mod close;
pub mod durability;
mod expansion;
pub(crate) mod graph_cache;
mod group_count;
mod hybrid;
pub mod lock;
pub(crate) mod materialize;
pub(crate) mod namespace_batch;
pub use cascade::CascadeRule;
#[cfg(any(test, feature = "test-seams"))]
pub use namespace_batch::namespace_delete_cascade_with_steps;
pub use namespace_batch::{namespace_declare_cascade, namespace_delete_cascade};
#[cfg(feature = "graph-cypher")]
mod enable_graph;
#[cfg(feature = "graph-cypher")]
pub(crate) mod native_graph;
#[cfg(feature = "graph-cypher")]
mod pinned_search;
mod pool;
pub(crate) mod prepared;
mod prepared_lexical;
mod reindex;
pub(crate) mod rescored_scan;
mod snapshot;
mod snapshot_copy;
mod snapshot_view;
pub use namespace_batch::{
    LiveNamespaceMutation, NamespaceMutation, NamespaceRootId, namespace_batch,
    namespace_batch_live, namespace_relocate,
};
#[cfg(any(test, feature = "test-seams"))]
#[doc(hidden)]
pub use namespace_batch::{
    namespace_batch_live_on_vfs, namespace_batch_live_with_steps, namespace_batch_with_steps,
    namespace_reclaim, namespace_reclaim_on_vfs, namespace_relocate_with_steps,
};
pub(crate) mod stats;

use crate::diag::{timing_elapsed, timing_start};
pub use cancel::{
    CancelToken, Deadline, DeadlineError, QueryCancellation, QueryControl, QueryError,
};
#[cfg(any(test, feature = "test-seams"))]
pub use clock::ManualMonotonicClock;
pub use clock::{MonotonicClock, SystemMonotonicClock};
#[cfg(any(test, feature = "test-seams"))]
pub use hybrid::{
    HybridScoreTestObservations, begin_hybrid_fresh_round_test_observations,
    begin_hybrid_score_test_observations, take_hybrid_score_test_observations,
};
#[cfg(any(test, feature = "test-seams"))]
pub use materialize::MaterializationTestCounters;
pub use materialize::{
    HybridPreparationError, MaterializationError, MaterializedRow, QueryMaterializer,
};

pub use group_count::{
    DocumentGroup, DocumentGroupCounts, DocumentGroupValue, MAX_DOCUMENT_GROUP_LIMIT,
};
pub use rescored_scan::ScanRescoreOptions;
pub use snapshot::{
    InMemorySegment, InMemorySegmentFactors, PreparedSegment, PublishedSnapshot, SnapshotLease,
};
pub use snapshot_copy::SnapshotTargetReason;
pub use stats::Stats;

use std::collections::HashSet;
#[cfg(any(test, feature = "test-seams"))]
use std::io::IoSlice;
use std::path::{Path, PathBuf};
use std::sync::atomic::{AtomicU64, Ordering};
use std::sync::{Arc, Condvar, Mutex, RwLock};
use std::time::Duration;

use durability::{
    CommitTier, DurabilityMode, DurabilityPolicy, DurabilityPolicyError, SyncRequirement,
};
use lock::{StoreLock, StoreLockError};

use self::close::BackgroundThread;

/// Store-owned infrastructure injected only by deterministic tests.
#[cfg(any(test, feature = "test-seams"))]
#[derive(Clone)]
pub struct StoreTestDependencies {
    pub(crate) vfs: Arc<dyn crate::vfs::Vfs>,
    pub(crate) clock: Arc<dyn MonotonicClock>,
    hybrid_leg_fault: Option<HybridLegTestFault>,
    storage_fault_controller: Option<StorageFaultController>,
    ingest_retention_fault_controller: Option<crate::ingest::IngestRetentionFaultController>,
    metadata_test_controller: Option<Arc<crate::planner::MetadataTestController>>,
    pub(crate) vector_fault_controller: Option<crate::scan::vector_fault::VectorFaultController>,
    pub(crate) kernel_fault_controller: Option<crate::kernels::vector_fault::KernelFaultController>,
    vector_seal_scheme: Option<crate::quant::QuantScheme>,
}

#[cfg(any(test, feature = "test-seams"))]
impl StoreTestDependencies {
    /// Binds one filesystem and monotonic clock to every operation performed
    /// by a test store handle.
    #[must_use]
    pub fn new(vfs: Arc<dyn crate::vfs::Vfs>, clock: Arc<dyn MonotonicClock>) -> Self {
        Self {
            vfs,
            clock,
            hybrid_leg_fault: None,
            storage_fault_controller: None,
            ingest_retention_fault_controller: None,
            metadata_test_controller: None,
            vector_fault_controller: None,
            kernel_fault_controller: None,
            vector_seal_scheme: None,
        }
    }

    /// Arms one Store-owned hybrid fault; it is consumed exactly once.
    #[must_use]
    pub const fn with_hybrid_leg_fault(mut self, fault: HybridLegTestFault) -> Self {
        self.hybrid_leg_fault = Some(fault);
        self
    }

    /// Arms one storage-family fault controller at Store-owned production seams.
    #[must_use]
    pub fn with_storage_fault_controller(mut self, controller: StorageFaultController) -> Self {
        self.storage_fault_controller = Some(controller);
        self
    }

    /// Arms one ingest-retention fault at a Store-owned mutation checkpoint.
    #[must_use]
    pub fn with_ingest_retention_fault_controller(
        mut self,
        controller: crate::ingest::IngestRetentionFaultController,
    ) -> Self {
        self.ingest_retention_fault_controller = Some(controller);
        self
    }

    /// Binds one metadata/planner observation and fault controller to the
    /// public Store query path.
    #[must_use]
    pub fn with_metadata_test_controller(
        mut self,
        controller: Arc<crate::planner::MetadataTestController>,
    ) -> Self {
        self.metadata_test_controller = Some(controller);
        self
    }

    /// Arms one vector execution fault at the Store-owned query path.
    #[must_use]
    pub fn with_vector_fault_controller(
        mut self,
        controller: crate::scan::vector_fault::VectorFaultController,
    ) -> Self {
        self.vector_fault_controller = Some(controller);
        self
    }

    /// Forces and observes one concrete kernel backend through Store search.
    #[must_use]
    pub fn with_kernel_fault_controller(
        mut self,
        controller: crate::kernels::vector_fault::KernelFaultController,
    ) -> Self {
        self.kernel_fault_controller = Some(controller);
        self
    }

    /// Selects the immutable vector codec used by the next public Store seal.
    ///
    /// This test-support-only seam leaves public ingest and shipping defaults
    /// unchanged while exercising document-preserving Int8 publication.
    #[must_use]
    pub const fn with_vector_seal_scheme(mut self, scheme: crate::quant::QuantScheme) -> Self {
        self.vector_seal_scheme = Some(scheme);
        self
    }
}

/// Narrow storage-family faults available only through hidden test dependencies.
#[cfg(any(test, feature = "test-seams"))]
#[derive(Clone, Debug, Eq, PartialEq)]
#[doc(hidden)]
pub enum StorageTestFault {
    TornWalHeader,
    TornWalBody,
    TornWalChecksum,
    PostCommitError,
    ManifestPreRename,
    ManifestPostRename,
    CorruptSegmentRegion,
    WrongManifestObject,
    WrongSegmentObject,
    ListOmission { file_name: String },
    DeleteOmission { file_name: String },
}

#[cfg(any(test, feature = "test-seams"))]
impl StorageTestFault {
    const fn operation(&self) -> &'static str {
        match self {
            Self::TornWalHeader | Self::TornWalBody | Self::TornWalChecksum => "wal-prefix",
            Self::PostCommitError => "retry",
            Self::ManifestPreRename | Self::ManifestPostRename => "publication",
            Self::CorruptSegmentRegion | Self::WrongManifestObject | Self::WrongSegmentObject => {
                "format-check"
            }
            Self::ListOmission { .. } | Self::DeleteOmission { .. } => "orphan-cleanup",
        }
    }

    const fn key(&self) -> &'static str {
        match self {
            Self::TornWalHeader => "torn-wal-header",
            Self::TornWalBody => "torn-wal-body",
            Self::TornWalChecksum => "torn-wal-checksum",
            Self::PostCommitError => "post-commit-error",
            Self::ManifestPreRename => "manifest-pre-rename-crash",
            Self::ManifestPostRename => "manifest-post-rename-crash",
            Self::CorruptSegmentRegion => "corrupt-segment-region",
            Self::WrongManifestObject => "wrong-manifest-object",
            Self::WrongSegmentObject => "wrong-segment-object",
            Self::ListOmission { .. } | Self::DeleteOmission { .. } => "list-delete-omission",
        }
    }
}

/// Exact seed-derived target facts supplied before one storage operation runs.
#[cfg(any(test, feature = "test-seams"))]
#[derive(Clone, Debug, Eq, PartialEq)]
#[doc(hidden)]
pub struct StorageFaultPlan {
    op_index: u32,
    artifact: String,
    offset: Option<u64>,
    segment: Option<crate::segment::SegmentId>,
    region_kind: Option<u16>,
    chunk: Option<u32>,
}

#[cfg(any(test, feature = "test-seams"))]
impl StorageFaultPlan {
    /// Creates an exact artifact plan for one operation index.
    #[must_use]
    pub fn new(op_index: u32, artifact: impl Into<String>) -> Self {
        Self {
            op_index,
            artifact: artifact.into(),
            offset: None,
            segment: None,
            region_kind: None,
            chunk: None,
        }
    }

    /// Pins the exact byte offset selected by the primitive fixture.
    #[must_use]
    pub const fn with_offset(mut self, offset: u64) -> Self {
        self.offset = Some(offset);
        self
    }

    /// Pins the exact segment, region kind, and checksum chunk.
    #[must_use]
    pub const fn with_segment_region(
        mut self,
        segment: crate::segment::SegmentId,
        region_kind: u16,
        chunk: u32,
    ) -> Self {
        self.segment = Some(segment);
        self.region_kind = Some(region_kind);
        self.chunk = Some(chunk);
        self
    }

    /// Pins the exact newly published segment for a manifest checkpoint.
    #[must_use]
    pub const fn with_segment(mut self, segment: crate::segment::SegmentId) -> Self {
        self.segment = Some(segment);
        self
    }

    /// Exact operation index in the generated program.
    #[must_use]
    pub const fn op_index(&self) -> u32 {
        self.op_index
    }

    /// Normalized planned artifact.
    #[must_use]
    pub fn artifact(&self) -> &str {
        &self.artifact
    }

    /// Planned mutation/checkpoint byte offset when applicable.
    #[must_use]
    pub const fn offset(&self) -> Option<u64> {
        self.offset
    }

    /// Planned segment identity when applicable.
    #[must_use]
    pub const fn segment(&self) -> Option<crate::segment::SegmentId> {
        self.segment
    }

    /// Planned region kind when applicable.
    #[must_use]
    pub const fn region_kind(&self) -> Option<u16> {
        self.region_kind
    }

    /// Planned checksum chunk when applicable.
    #[must_use]
    pub const fn chunk(&self) -> Option<u32> {
        self.chunk
    }
}

/// Typed production checkpoint that consumed a storage fault.
#[cfg(any(test, feature = "test-seams"))]
#[derive(Clone, Copy, Debug, Eq, PartialEq)]
#[doc(hidden)]
pub enum StorageReceiptSite {
    WalOpenHeaderValidation,
    WalOpenRecordValidation,
    WalOpenRecordChecksum,
    WalCommitAppendAfterInnerSuccess,
    ManifestCommitBeforeRename,
    ManifestCommitAfterRename,
    SegmentReadRegionChecksum,
    ManifestOpenFamilyValidation,
    SegmentOpenFamilyValidation,
    SegmentOpenObjectIdentity,
    OrphanCleanupList,
    OrphanCleanupDelete,
}

#[cfg(any(test, feature = "test-seams"))]
impl StorageReceiptSite {
    const fn key(self) -> &'static str {
        match self {
            Self::WalOpenHeaderValidation => "WalOpen.HeaderValidation",
            Self::WalOpenRecordValidation => "WalOpen.RecordValidation",
            Self::WalOpenRecordChecksum => "WalOpen.RecordChecksum",
            Self::WalCommitAppendAfterInnerSuccess => "WalCommit.AppendAfterInnerSuccess",
            Self::ManifestCommitBeforeRename => "ManifestCommit.BeforeRename",
            Self::ManifestCommitAfterRename => "ManifestCommit.AfterRename",
            Self::SegmentReadRegionChecksum => "SegmentRead.RegionChecksum",
            Self::ManifestOpenFamilyValidation => "ManifestOpen.FamilyValidation",
            Self::SegmentOpenFamilyValidation => "SegmentOpen.FamilyValidation",
            Self::SegmentOpenObjectIdentity => "SegmentOpen.ObjectIdentity",
            Self::OrphanCleanupList => "OrphanCleanup.List",
            Self::OrphanCleanupDelete => "OrphanCleanup.Delete",
        }
    }
}

/// Exact facts observed at the production checkpoint.
#[cfg(any(test, feature = "test-seams"))]
#[derive(Clone, Debug, Eq, PartialEq)]
#[doc(hidden)]
pub enum StorageReceiptObserved {
    WalHeader {
        artifact: String,
        reason: crate::wal::header::WalHeaderError,
    },
    WalRecord {
        artifact: String,
        offset: u64,
        reason: crate::wal::replay::CorruptionReason,
    },
    WalAppend {
        artifact: String,
        encoded_len: u64,
        first_seq: u64,
        last_seq: u64,
        inner_append_completed: bool,
        caller_saw_error: bool,
    },
    ManifestRename {
        temporary: String,
        committed: String,
        rename_performed: bool,
        new_segment_final: bool,
        directory_sync_returned: bool,
    },
    SegmentChecksum {
        artifact: String,
        segment: crate::segment::SegmentId,
        region_kind: u16,
        chunk: u32,
        expected_checksum: u64,
        actual_checksum: u64,
    },
    Format {
        artifact: String,
        check: crate::format::frame::FormatCheck,
        expected_family: Option<u16>,
        actual_family: Option<u16>,
        expected_id: Option<crate::segment::SegmentId>,
        actual_id: Option<crate::segment::SegmentId>,
    },
    Omission {
        artifact: String,
        deletion_observed: bool,
    },
}

/// One fact-only typed storage receipt emitted by product code.
#[cfg(any(test, feature = "test-seams"))]
#[derive(Clone, Debug, Eq, PartialEq)]
#[doc(hidden)]
pub struct StorageFaultReceipt {
    operation: &'static str,
    fault: &'static str,
    site: StorageReceiptSite,
    plan: StorageFaultPlan,
    observed: StorageReceiptObserved,
    cardinality: u32,
}

#[cfg(any(test, feature = "test-seams"))]
impl StorageFaultReceipt {
    #[must_use]
    pub const fn campaign(&self) -> &'static str {
        "storage-durability"
    }

    #[must_use]
    pub const fn operation(&self) -> &'static str {
        self.operation
    }

    #[must_use]
    pub const fn fault(&self) -> &'static str {
        self.fault
    }

    #[must_use]
    pub const fn site(&self) -> &'static str {
        self.site.key()
    }

    #[must_use]
    pub const fn typed_site(&self) -> StorageReceiptSite {
        self.site
    }

    #[must_use]
    pub const fn cardinality(&self) -> u32 {
        self.cardinality
    }

    #[must_use]
    pub const fn plan(&self) -> &StorageFaultPlan {
        &self.plan
    }

    #[must_use]
    pub const fn observed(&self) -> &StorageReceiptObserved {
        &self.observed
    }

    /// Decodes the exact receipt acknowledged by an aborting publication child.
    ///
    /// The child writes this only after `emit` has stored the typed production
    /// receipt. Parsing remains in product test support so an external harness
    /// cannot manufacture a generic receipt from partial fields.
    pub fn from_manifest_abort_acknowledgment(acknowledgment: &str) -> Result<Self, String> {
        let body = acknowledgment
            .strip_suffix('\n')
            .filter(|body| !body.contains('\n'))
            .ok_or_else(|| {
                "manifest abort acknowledgment must end in exactly one newline".to_owned()
            })?;
        let fields = body
            .split('|')
            .map(|field| {
                field.split_once('=').ok_or_else(|| {
                    format!("manifest abort acknowledgment field lacks '=': {field}")
                })
            })
            .collect::<Result<Vec<_>, _>>()?;
        let expected_keys = [
            "campaign",
            "operation",
            "fault",
            "site",
            "op_index",
            "cardinality",
            "artifact",
            "temporary",
            "committed",
            "rename_performed",
            "new_segment_final",
            "directory_sync_returned",
        ];
        if fields.len() != expected_keys.len()
            || fields
                .iter()
                .map(|(key, _)| *key)
                .ne(expected_keys.iter().copied())
        {
            return Err("manifest abort acknowledgment field inventory differs".to_owned());
        }
        let mut values = fields.into_iter().map(|(_, value)| value);
        let campaign = values
            .next()
            .ok_or_else(|| "manifest abort acknowledgment lacks campaign".to_owned())?;
        let operation = values
            .next()
            .ok_or_else(|| "manifest abort acknowledgment lacks operation".to_owned())?;
        let fault = values
            .next()
            .ok_or_else(|| "manifest abort acknowledgment lacks fault".to_owned())?;
        let site = values
            .next()
            .ok_or_else(|| "manifest abort acknowledgment lacks site".to_owned())?;
        let op_index = values
            .next()
            .ok_or_else(|| "manifest abort acknowledgment lacks op index".to_owned())?
            .parse::<u32>()
            .map_err(|error| format!("manifest abort op index is invalid: {error}"))?;
        let cardinality = values
            .next()
            .ok_or_else(|| "manifest abort acknowledgment lacks cardinality".to_owned())?
            .parse::<u32>()
            .map_err(|error| format!("manifest abort cardinality is invalid: {error}"))?;
        let artifact = values
            .next()
            .ok_or_else(|| "manifest abort acknowledgment lacks artifact".to_owned())?;
        let temporary = values
            .next()
            .ok_or_else(|| "manifest abort acknowledgment lacks temporary path".to_owned())?;
        let committed = values
            .next()
            .ok_or_else(|| "manifest abort acknowledgment lacks committed path".to_owned())?;
        let parse_bool = |field: &str, value: &str| match value {
            "true" => Ok(true),
            "false" => Ok(false),
            _ => Err(format!("manifest abort {field} is not a bool: {value}")),
        };
        let rename_performed = parse_bool(
            "rename_performed",
            values
                .next()
                .ok_or_else(|| "manifest abort acknowledgment lacks rename result".to_owned())?,
        )?;
        let new_segment_final = parse_bool(
            "new_segment_final",
            values.next().ok_or_else(|| {
                "manifest abort acknowledgment lacks final-segment result".to_owned()
            })?,
        )?;
        let directory_sync_returned = parse_bool(
            "directory_sync_returned",
            values.next().ok_or_else(|| {
                "manifest abort acknowledgment lacks directory-sync result".to_owned()
            })?,
        )?;
        if campaign != "storage-durability"
            || operation != "publication"
            || cardinality != 1
            || artifact != crate::manifest::io::MANIFEST_FILE
            || temporary != crate::manifest::io::MANIFEST_TEMP_FILE
            || committed != crate::manifest::io::MANIFEST_FILE
            || !new_segment_final
            || directory_sync_returned
        {
            return Err("manifest abort acknowledgment fixed facts differ".to_owned());
        }
        let (fault, site, expected_rename) = match (fault, site) {
            ("manifest-pre-rename-crash", "ManifestCommit.BeforeRename") => (
                "manifest-pre-rename-crash",
                StorageReceiptSite::ManifestCommitBeforeRename,
                false,
            ),
            ("manifest-post-rename-crash", "ManifestCommit.AfterRename") => (
                "manifest-post-rename-crash",
                StorageReceiptSite::ManifestCommitAfterRename,
                true,
            ),
            _ => return Err("manifest abort fault and production site differ".to_owned()),
        };
        if rename_performed != expected_rename {
            return Err("manifest abort rename result differs from production site".to_owned());
        }
        Ok(Self {
            operation: "publication",
            fault,
            site,
            plan: StorageFaultPlan::new(op_index, artifact),
            observed: StorageReceiptObserved::ManifestRename {
                temporary: temporary.to_owned(),
                committed: committed.to_owned(),
                rename_performed,
                new_segment_final,
                directory_sync_returned,
            },
            cardinality,
        })
    }
}

#[cfg(any(test, feature = "test-seams"))]
#[derive(Debug)]
struct StorageFaultState {
    fault: StorageTestFault,
    plan: StorageFaultPlan,
    consumed: bool,
    receipt: Option<StorageFaultReceipt>,
    cleanup_report: Option<StorageCleanupReport>,
    #[cfg(unix)]
    child_abort_ack: Option<std::os::unix::net::UnixStream>,
}

/// Truthful internal result of Store-owned orphan cleanup.
#[cfg(any(test, feature = "test-seams"))]
#[derive(Clone, Debug, Eq, PartialEq)]
#[doc(hidden)]
pub struct StorageCleanupReport {
    reclaimed_bytes: u64,
    deleted_paths: Vec<String>,
    retained_eligible_paths: Vec<String>,
    directory_synced: bool,
}

#[cfg(any(test, feature = "test-seams"))]
impl StorageCleanupReport {
    /// Exact bytes removed from the filesystem.
    #[must_use]
    pub const fn reclaimed_bytes(&self) -> u64 {
        self.reclaimed_bytes
    }

    /// Normalized relative paths confirmed absent after deletion.
    #[must_use]
    pub fn deleted_paths(&self) -> &[String] {
        &self.deleted_paths
    }

    /// Eligible paths whose deletion was omitted by an armed test fault.
    #[must_use]
    pub fn retained_eligible_paths(&self) -> &[String] {
        &self.retained_eligible_paths
    }

    /// Whether the cleanup issued its required directory synchronization.
    #[must_use]
    pub const fn directory_synced(&self) -> bool {
        self.directory_synced
    }
}

/// Shared handle that arms one storage fault and receives its production receipt.
#[cfg(any(test, feature = "test-seams"))]
#[derive(Clone, Debug)]
#[doc(hidden)]
pub struct StorageFaultController {
    state: Arc<Mutex<StorageFaultState>>,
}

#[cfg(any(test, feature = "test-seams"))]
impl StorageFaultController {
    /// Arms one controller. A clone observes the same one-shot receipt.
    #[must_use]
    pub fn new(fault: StorageTestFault, plan: StorageFaultPlan) -> Self {
        Self {
            state: Arc::new(Mutex::new(StorageFaultState {
                fault,
                plan,
                consumed: false,
                receipt: None,
                cleanup_report: None,
                #[cfg(unix)]
                child_abort_ack: None,
            })),
        }
    }

    /// Supplies the inherited descriptor used to acknowledge a child abort.
    #[cfg(unix)]
    #[must_use]
    pub fn with_child_abort_ack(self, acknowledgment: std::os::unix::net::UnixStream) -> Self {
        if let Ok(mut state) = self.state.lock() {
            state.child_abort_ack = Some(acknowledgment);
        }
        self
    }

    /// Takes the single receipt emitted by the intended production operation.
    pub fn take_receipt(&self) -> Option<StorageFaultReceipt> {
        self.state
            .lock()
            .ok()
            .and_then(|mut state| state.receipt.take())
    }

    /// Takes the exact result of the most recent public-open cleanup pass.
    pub fn take_cleanup_report(&self) -> Option<StorageCleanupReport> {
        self.state
            .lock()
            .ok()
            .and_then(|mut state| state.cleanup_report.take())
    }

    fn record_cleanup_report(&self, report: crate::manifest::io::OrphanCleanupReport) {
        let normalize = |paths: Vec<PathBuf>| {
            paths
                .into_iter()
                .filter_map(|path| {
                    path.file_name()
                        .and_then(|name| name.to_str())
                        .map(str::to_owned)
                })
                .collect::<Vec<_>>()
        };
        if let Ok(mut state) = self.state.lock() {
            state.cleanup_report = Some(StorageCleanupReport {
                reclaimed_bytes: report.reclaimed_bytes,
                deleted_paths: normalize(report.deleted_paths),
                retained_eligible_paths: normalize(report.retained_eligible_paths),
                directory_synced: report.directory_synced,
            });
        }
    }

    fn armed_fault(&self) -> Option<StorageTestFault> {
        self.state
            .lock()
            .ok()
            .and_then(|state| (!state.consumed).then(|| state.fault.clone()))
    }

    fn planned_segment(&self) -> Option<crate::segment::SegmentId> {
        self.state
            .lock()
            .ok()
            .and_then(|state| state.plan.segment())
    }

    fn emit(&self, site: StorageReceiptSite, observed: StorageReceiptObserved) -> bool {
        let Ok(mut state) = self.state.lock() else {
            return false;
        };
        if state.consumed {
            return false;
        }
        state.consumed = true;
        state.receipt = Some(StorageFaultReceipt {
            operation: state.fault.operation(),
            fault: state.fault.key(),
            site,
            plan: state.plan.clone(),
            observed,
            cardinality: 1,
        });
        true
    }

    fn abort_after_manifest_receipt(&self) -> ! {
        #[cfg(unix)]
        if let Ok(mut state) = self.state.lock() {
            let wire_receipt = state
                .receipt
                .as_ref()
                .and_then(storage_manifest_abort_wire_receipt);
            if let (Some(wire_receipt), Some(acknowledgment)) =
                (wire_receipt, state.child_abort_ack.as_mut())
            {
                use std::io::Write as _;
                let _ = acknowledgment.write_all(wire_receipt.as_bytes());
                let _ = acknowledgment.flush();
            }
        }
        std::process::abort()
    }
}

#[cfg(any(test, feature = "test-seams"))]
thread_local! {
    static STORAGE_OPEN_CONTROLLER: std::cell::RefCell<
        Option<std::sync::Weak<Mutex<StorageFaultState>>>
    > = const { std::cell::RefCell::new(None) };
}

#[cfg(any(test, feature = "test-seams"))]
fn with_storage_open_controller<R>(
    controller: Option<&StorageFaultController>,
    operation: impl FnOnce() -> R,
) -> R {
    STORAGE_OPEN_CONTROLLER.with(|slot| {
        let previous = slot.replace(controller.map(|controller| Arc::downgrade(&controller.state)));
        let result = operation();
        let _ = slot.replace(previous);
        result
    })
}

#[cfg(any(test, feature = "test-seams"))]
fn storage_open_controller() -> Option<StorageFaultController> {
    STORAGE_OPEN_CONTROLLER.with(|slot| {
        slot.borrow()
            .as_ref()
            .and_then(std::sync::Weak::upgrade)
            .map(|state| StorageFaultController { state })
    })
}

#[cfg(any(test, feature = "test-seams"))]
pub(crate) fn record_storage_wal_recovery_fault(error: &crate::wal::WalRecoveryError) {
    let Some(controller) = storage_open_controller() else {
        return;
    };
    match (controller.armed_fault(), error) {
        (
            Some(StorageTestFault::TornWalHeader),
            crate::wal::WalRecoveryError::InvalidHeader(reason),
        ) => {
            let _ = controller.emit(
                StorageReceiptSite::WalOpenHeaderValidation,
                StorageReceiptObserved::WalHeader {
                    artifact: "wal.ze".to_owned(),
                    reason: *reason,
                },
            );
        }
        (
            Some(StorageTestFault::TornWalBody),
            crate::wal::WalRecoveryError::CorruptAt {
                offset,
                reason:
                    reason @ crate::wal::replay::CorruptionReason::Record {
                        error:
                            crate::wal::record::RecordError::HeaderTruncated { .. }
                            | crate::wal::record::RecordError::BodyTruncated { .. }
                            | crate::wal::record::RecordError::LengthOverflow { .. },
                        ..
                    },
            },
        ) => record_storage_wal_record_fault(
            &controller,
            StorageReceiptSite::WalOpenRecordValidation,
            *offset,
            *reason,
        ),
        (
            Some(StorageTestFault::TornWalChecksum),
            crate::wal::WalRecoveryError::CorruptAt {
                offset,
                reason:
                    reason @ crate::wal::replay::CorruptionReason::Record {
                        error: crate::wal::record::RecordError::ChecksumMismatch { .. },
                        ..
                    },
            },
        ) => record_storage_wal_record_fault(
            &controller,
            StorageReceiptSite::WalOpenRecordChecksum,
            *offset,
            *reason,
        ),
        _ => {}
    }
}

#[cfg(any(test, feature = "test-seams"))]
fn record_storage_wal_record_fault(
    controller: &StorageFaultController,
    site: StorageReceiptSite,
    offset: usize,
    reason: crate::wal::replay::CorruptionReason,
) {
    let Ok(offset) = u64::try_from(offset) else {
        return;
    };
    let _ = controller.emit(
        site,
        StorageReceiptObserved::WalRecord {
            artifact: "wal.ze".to_owned(),
            offset,
            reason,
        },
    );
}

#[cfg(any(test, feature = "test-seams"))]
pub(crate) fn record_storage_manifest_format_fault(
    error: &crate::format::frame::FormatError,
    actual_family: Option<u16>,
) {
    let Some(controller) = storage_open_controller() else {
        return;
    };
    if matches!(
        controller.armed_fault(),
        Some(StorageTestFault::WrongManifestObject)
    ) && error.check() == crate::format::frame::FormatCheck::Family
    {
        let _ = controller.emit(
            StorageReceiptSite::ManifestOpenFamilyValidation,
            StorageReceiptObserved::Format {
                artifact: storage_artifact_name(error.artifact()).to_owned(),
                check: error.check(),
                expected_family: Some(crate::format::FormatFamily::Manifest.id()),
                actual_family,
                expected_id: None,
                actual_id: None,
            },
        );
    }
}

#[cfg(any(test, feature = "test-seams"))]
pub(crate) fn record_storage_segment_format_fault(
    error: &crate::format::frame::FormatError,
    actual_family: Option<u16>,
    expected_id: crate::segment::SegmentId,
) {
    let Some(controller) = storage_open_controller() else {
        return;
    };
    if matches!(
        controller.armed_fault(),
        Some(StorageTestFault::WrongSegmentObject)
    ) && error.check() == crate::format::frame::FormatCheck::Family
    {
        let _ = controller.emit(
            StorageReceiptSite::SegmentOpenFamilyValidation,
            StorageReceiptObserved::Format {
                artifact: storage_artifact_name(error.artifact()).to_owned(),
                check: error.check(),
                expected_family: Some(crate::format::FormatFamily::Segment.id()),
                actual_family,
                expected_id: Some(expected_id),
                actual_id: None,
            },
        );
    }
}

#[cfg(any(test, feature = "test-seams"))]
pub(crate) fn record_storage_segment_identity_fault(
    artifact: &str,
    expected: crate::segment::SegmentId,
    actual: crate::segment::SegmentId,
) {
    let Some(controller) = storage_open_controller() else {
        return;
    };
    if matches!(
        controller.armed_fault(),
        Some(StorageTestFault::WrongSegmentObject)
    ) {
        let _ = controller.emit(
            StorageReceiptSite::SegmentOpenObjectIdentity,
            StorageReceiptObserved::Format {
                artifact: storage_artifact_name(artifact).to_owned(),
                check: crate::format::frame::FormatCheck::ObjectIdentity,
                expected_family: None,
                actual_family: None,
                expected_id: Some(expected),
                actual_id: Some(actual),
            },
        );
    }
}

/// Renders a manifest-abort receipt for the child-abort acknowledgement wire.
///
/// Gated on `unix` because its only caller is the `#[cfg(unix)]` arm of
/// [`StorageFaultController::abort_after_manifest_receipt`], which writes the
/// receipt down an inherited `UnixStream`. Without the gate this is dead code
/// on every other platform. W11 adds the Windows child-abort channel; the gate
/// widens with it.
#[cfg(unix)]
#[cfg(any(test, feature = "test-seams"))]
fn storage_manifest_abort_wire_receipt(receipt: &StorageFaultReceipt) -> Option<String> {
    let StorageReceiptObserved::ManifestRename {
        temporary,
        committed,
        rename_performed,
        new_segment_final,
        directory_sync_returned,
    } = &receipt.observed
    else {
        return None;
    };
    Some(format!(
        "campaign={}|operation={}|fault={}|site={}|op_index={}|cardinality={}|artifact={}|temporary={temporary}|committed={committed}|rename_performed={rename_performed}|new_segment_final={new_segment_final}|directory_sync_returned={directory_sync_returned}\n",
        receipt.campaign(),
        receipt.operation,
        receipt.fault,
        receipt.site.key(),
        receipt.plan.op_index,
        receipt.cardinality,
        receipt.plan.artifact,
    ))
}

#[cfg(any(test, feature = "test-seams"))]
fn storage_artifact_name(artifact: &str) -> &str {
    Path::new(artifact)
        .file_name()
        .and_then(|name| name.to_str())
        .unwrap_or(artifact)
}

#[cfg(any(test, feature = "test-seams"))]
type StorageSegmentControllerRegistry = Mutex<
    Vec<(
        PathBuf,
        crate::segment::SegmentId,
        std::sync::Weak<Mutex<StorageFaultState>>,
    )>,
>;

#[cfg(any(test, feature = "test-seams"))]
static STORAGE_SEGMENT_CONTROLLERS: std::sync::OnceLock<StorageSegmentControllerRegistry> =
    std::sync::OnceLock::new();

#[cfg(any(test, feature = "test-seams"))]
fn register_storage_segment_controller(
    store_directory: &Path,
    segment: crate::segment::SegmentId,
    controller: &StorageFaultController,
) {
    let registry = STORAGE_SEGMENT_CONTROLLERS.get_or_init(|| Mutex::new(Vec::new()));
    if let Ok(mut entries) = registry.lock() {
        entries.retain(|(_, _, state)| state.strong_count() != 0);
        entries.push((
            store_directory.to_path_buf(),
            segment,
            Arc::downgrade(&controller.state),
        ));
    }
}

#[cfg(any(test, feature = "test-seams"))]
pub(crate) fn record_storage_segment_checksum_fault(
    store_directory: &Path,
    segment: crate::segment::SegmentId,
    kind: crate::segment::layout::RegionKind,
    chunk: u32,
    expected: u64,
    actual: u64,
) {
    let Some(registry) = STORAGE_SEGMENT_CONTROLLERS.get() else {
        return;
    };
    let states = match registry.lock() {
        Ok(mut entries) => {
            entries.retain(|(_, _, state)| state.strong_count() != 0);
            entries
                .iter()
                .filter(|(directory, registered, _)| {
                    directory == store_directory && *registered == segment
                })
                .filter_map(|(_, _, state)| state.upgrade())
                .collect::<Vec<_>>()
        }
        Err(_) => return,
    };
    for state in states {
        let controller = StorageFaultController { state };
        if matches!(
            controller.armed_fault(),
            Some(StorageTestFault::CorruptSegmentRegion)
        ) {
            let _ = controller.emit(
                StorageReceiptSite::SegmentReadRegionChecksum,
                StorageReceiptObserved::SegmentChecksum {
                    artifact: segment.file_name(),
                    segment,
                    region_kind: kind.id(),
                    chunk,
                    expected_checksum: expected,
                    actual_checksum: actual,
                },
            );
        }
    }
}

#[cfg(any(test, feature = "test-seams"))]
struct StorageFaultVfs {
    inner: Arc<dyn crate::vfs::Vfs>,
    controller: StorageFaultController,
}

#[cfg(any(test, feature = "test-seams"))]
struct StorageFaultFile {
    inner: Box<dyn crate::vfs::VfsFile>,
    path: PathBuf,
    controller: StorageFaultController,
}

#[cfg(any(test, feature = "test-seams"))]
impl StorageFaultFile {
    fn after_append(&self, encoded_len: u64, first_seq: u64, last_seq: u64) -> std::io::Result<()> {
        if matches!(
            self.controller.armed_fault(),
            Some(StorageTestFault::PostCommitError)
        ) && self.path.file_name().is_some_and(|name| name == "wal.ze")
            && self.controller.emit(
                StorageReceiptSite::WalCommitAppendAfterInnerSuccess,
                StorageReceiptObserved::WalAppend {
                    artifact: "wal.ze".to_owned(),
                    encoded_len,
                    first_seq,
                    last_seq,
                    inner_append_completed: true,
                    caller_saw_error: true,
                },
            )
        {
            return Err(std::io::Error::other(
                "scheduled post-commit error after inner append",
            ));
        }
        Ok(())
    }
}

#[cfg(any(test, feature = "test-seams"))]
impl crate::vfs::VfsFile for StorageFaultFile {
    fn append(&mut self, bytes: &[u8]) -> std::io::Result<()> {
        let record_facts = storage_wal_append_facts(bytes)?;
        self.inner.append(bytes)?;
        record_facts.map_or(Ok(()), |(encoded_len, first_seq, last_seq)| {
            self.after_append(encoded_len, first_seq, last_seq)
        })
    }

    fn append_vectored(&mut self, buffers: &mut [IoSlice<'_>]) -> std::io::Result<()> {
        let mut aggregate = None::<(u64, u64, u64)>;
        for bytes in buffers.iter() {
            if let Some((encoded_len, first_seq, last_seq)) = storage_wal_append_facts(bytes)? {
                aggregate = Some(match aggregate {
                    None => (encoded_len, first_seq, last_seq),
                    Some((total, first, _)) => (
                        total.checked_add(encoded_len).ok_or_else(|| {
                            std::io::Error::other("WAL record append length overflow")
                        })?,
                        first,
                        last_seq,
                    ),
                });
            }
        }
        self.inner.append_vectored(buffers)?;
        aggregate.map_or(Ok(()), |(encoded_len, first_seq, last_seq)| {
            self.after_append(encoded_len, first_seq, last_seq)
        })
    }

    fn sync(&self, kind: crate::vfs::SyncKind) -> std::io::Result<()> {
        self.inner.sync(kind)
    }
}

#[cfg(any(test, feature = "test-seams"))]
impl crate::vfs::Vfs for StorageFaultVfs {
    fn segment_data_read_counter(&self) -> Option<Arc<AtomicU64>> {
        self.inner.segment_data_read_counter()
    }

    fn ensure_directory(&self, path: &Path, create: bool) -> std::io::Result<bool> {
        self.inner.ensure_directory(path, create)
    }

    fn open(&self, path: &Path) -> std::io::Result<u64> {
        self.inner.open(path)
    }

    fn open_for_map(&self, path: &Path) -> std::io::Result<std::fs::File> {
        self.inner.open_for_map(path)
    }

    fn read(&self, path: &Path) -> std::io::Result<Vec<u8>> {
        self.inner.read(path)
    }

    fn read_range(&self, path: &Path, offset: u64, length: usize) -> std::io::Result<Vec<u8>> {
        self.inner.read_range(path, offset, length)
    }

    fn write(&self, path: &Path, bytes: &[u8]) -> std::io::Result<()> {
        self.inner.write(path, bytes)
    }

    fn open_append(&self, path: &Path) -> std::io::Result<Box<dyn crate::vfs::VfsFile>> {
        Ok(Box::new(StorageFaultFile {
            inner: self.inner.open_append(path)?,
            path: path.to_path_buf(),
            controller: self.controller.clone(),
        }))
    }

    fn rename(&self, from: &Path, to: &Path) -> std::io::Result<()> {
        let is_manifest = from
            .file_name()
            .is_some_and(|name| name == crate::manifest::io::MANIFEST_TEMP_FILE)
            && to
                .file_name()
                .is_some_and(|name| name == crate::manifest::io::MANIFEST_FILE);
        if is_manifest
            && matches!(
                self.controller.armed_fault(),
                Some(StorageTestFault::ManifestPreRename)
            )
            && self.controller.emit(
                StorageReceiptSite::ManifestCommitBeforeRename,
                StorageReceiptObserved::ManifestRename {
                    temporary: crate::manifest::io::MANIFEST_TEMP_FILE.to_owned(),
                    committed: crate::manifest::io::MANIFEST_FILE.to_owned(),
                    rename_performed: false,
                    new_segment_final: storage_planned_segment_is_final(
                        &*self.inner,
                        from,
                        self.controller.planned_segment(),
                    ),
                    directory_sync_returned: false,
                },
            )
        {
            self.controller.abort_after_manifest_receipt();
        }
        self.inner.rename(from, to)?;
        if is_manifest
            && matches!(
                self.controller.armed_fault(),
                Some(StorageTestFault::ManifestPostRename)
            )
            && self.controller.emit(
                StorageReceiptSite::ManifestCommitAfterRename,
                StorageReceiptObserved::ManifestRename {
                    temporary: crate::manifest::io::MANIFEST_TEMP_FILE.to_owned(),
                    committed: crate::manifest::io::MANIFEST_FILE.to_owned(),
                    rename_performed: true,
                    new_segment_final: storage_planned_segment_is_final(
                        &*self.inner,
                        to,
                        self.controller.planned_segment(),
                    ),
                    directory_sync_returned: false,
                },
            )
        {
            self.controller.abort_after_manifest_receipt();
        }
        Ok(())
    }

    fn sync(&self, path: &Path, kind: crate::vfs::SyncKind) -> std::io::Result<()> {
        self.inner.sync(path, kind)
    }

    fn list(&self, directory: &Path) -> std::io::Result<Vec<PathBuf>> {
        let mut paths = self.inner.list(directory)?;
        if let Some(StorageTestFault::ListOmission { file_name }) = self.controller.armed_fault()
            && let Some(position) = paths.iter().position(|path| {
                path.file_name()
                    .is_some_and(|name| name == file_name.as_str())
            })
        {
            let _ = paths.remove(position);
            let _ = self.controller.emit(
                StorageReceiptSite::OrphanCleanupList,
                StorageReceiptObserved::Omission {
                    artifact: file_name,
                    deletion_observed: false,
                },
            );
        }
        Ok(paths)
    }

    fn for_each_direct_child(
        &self,
        directory: &Path,
        visitor: &mut dyn FnMut(&Path) -> std::io::Result<()>,
    ) -> std::io::Result<()> {
        // The omission belongs to OrphanCleanup.List, not the earlier native
        // format admission scan. Only list() consumes that one-shot fault.
        self.inner.for_each_direct_child(directory, visitor)
    }

    fn delete(&self, path: &Path) -> std::io::Result<()> {
        if let Some(StorageTestFault::DeleteOmission { file_name }) = self.controller.armed_fault()
            && path
                .file_name()
                .is_some_and(|name| name == file_name.as_str())
            && self.controller.emit(
                StorageReceiptSite::OrphanCleanupDelete,
                StorageReceiptObserved::Omission {
                    artifact: file_name,
                    deletion_observed: false,
                },
            )
        {
            return Ok(());
        }
        self.inner.delete(path)
    }
}

#[cfg(any(test, feature = "test-seams"))]
fn storage_wal_append_facts(bytes: &[u8]) -> std::io::Result<Option<(u64, u64, u64)>> {
    const RECORD_HEADER_LEN: usize = 14;
    const RECORD_CHECKSUM_LEN: usize = 8;
    let has_wal_header = bytes
        .get(..8)
        .is_some_and(|magic| magic == crate::format::frame::FILE_MAGIC)
        && bytes
            .get(8..10)
            .is_some_and(|family| family == crate::format::FormatFamily::Wal.id().to_le_bytes());
    let mut offset = if has_wal_header {
        crate::wal::header::WAL_HEADER_LEN
    } else {
        0
    };
    if offset == bytes.len() {
        return Ok(None);
    }
    let mut encoded_len = 0_u64;
    let mut first_seq = None;
    let mut last_seq = None;
    while offset < bytes.len() {
        let header_end = offset
            .checked_add(RECORD_HEADER_LEN)
            .ok_or_else(|| std::io::Error::other("WAL record header offset overflow"))?;
        let header = bytes
            .get(offset..header_end)
            .ok_or_else(|| std::io::Error::other("WAL append has a torn record header"))?;
        let payload_len = usize::try_from(u32::from_le_bytes(
            header
                .get(..4)
                .ok_or_else(|| std::io::Error::other("WAL payload length is absent"))?
                .try_into()
                .map_err(|_| std::io::Error::other("WAL payload length width changed"))?,
        ))
        .map_err(|_| std::io::Error::other("WAL payload length does not fit usize"))?;
        let seq = u64::from_le_bytes(
            header
                .get(4..12)
                .ok_or_else(|| std::io::Error::other("WAL sequence is absent"))?
                .try_into()
                .map_err(|_| std::io::Error::other("WAL sequence width changed"))?,
        );
        let record_len = RECORD_HEADER_LEN
            .checked_add(payload_len)
            .and_then(|length| length.checked_add(RECORD_CHECKSUM_LEN))
            .ok_or_else(|| std::io::Error::other("WAL record length overflow"))?;
        offset = offset
            .checked_add(record_len)
            .ok_or_else(|| std::io::Error::other("WAL record end overflow"))?;
        if offset > bytes.len() {
            return Err(std::io::Error::other("WAL append has a torn record body"));
        }
        encoded_len = encoded_len
            .checked_add(
                u64::try_from(record_len)
                    .map_err(|_| std::io::Error::other("WAL record length does not fit u64"))?,
            )
            .ok_or_else(|| std::io::Error::other("WAL record append length overflow"))?;
        first_seq.get_or_insert(seq);
        last_seq = Some(seq);
    }
    match (first_seq, last_seq) {
        (Some(first), Some(last)) => Ok(Some((encoded_len, first, last))),
        _ => Ok(None),
    }
}

#[cfg(any(test, feature = "test-seams"))]
fn storage_planned_segment_is_final(
    vfs: &dyn crate::vfs::Vfs,
    artifact: &Path,
    planned: Option<crate::segment::SegmentId>,
) -> bool {
    let Some(file_name) = planned.map(crate::segment::SegmentId::file_name) else {
        return false;
    };
    artifact.parent().is_some_and(|directory| {
        vfs.list(directory).is_ok_and(|paths| {
            paths.iter().any(|path| {
                path.file_name()
                    .and_then(|name| name.to_str())
                    .is_some_and(|name| name == file_name)
            })
        })
    })
}

/// Narrow test-only fault at a Store-owned hybrid leg seam.
#[cfg(any(test, feature = "test-seams"))]
#[derive(Clone, Copy, Debug, Eq, PartialEq)]
#[doc(hidden)]
pub enum HybridLegTestFault {
    /// Panics after the named leg starts and before it can publish output.
    Panic(crate::fusion::FusionLeg),
}

#[allow(clippy::panic)]
fn maybe_trigger_hybrid_leg_panic(armed: bool, detail: &'static str) {
    #[cfg(any(test, feature = "test-seams"))]
    if armed {
        std::panic::panic_any(detail);
    }
    #[cfg(not(any(test, feature = "test-seams")))]
    let _ = (armed, detail);
}

/// Facts emitted by one Store-owned parallel hybrid execution.
#[cfg(any(test, feature = "test-seams"))]
#[derive(Clone, Debug, Eq, PartialEq)]
#[doc(hidden)]
pub struct HybridExecutionReceipt {
    /// Thread that executed vector work.
    pub vector_thread: std::thread::ThreadId,
    /// Name of the scoped lexical worker.
    pub lexical_thread_name: Option<String>,
    /// Vector leg reached a terminal result before return.
    pub vector_completed: bool,
    /// Lexical leg reached a terminal result before return.
    pub lexical_completed: bool,
}

/// Default grace period given to admitted readers before close cancellation.
pub const DEFAULT_READER_DRAIN_TIMEOUT: Duration = Duration::from_millis(250);

/// Explicit lifecycle state for one store handle.
#[derive(Clone, Copy, Debug, Eq, PartialEq)]
pub enum StoreState {
    /// New calls may be admitted.
    Open,
    /// New calls are rejected while admitted work drains.
    Closing,
    /// Teardown completed and only idempotent close/state calls remain valid.
    Closed,
}

/// Store-open configuration.
#[derive(Clone, Debug, Eq, PartialEq)]
pub struct OpenOptions {
    inspection: bool,
    access_mode: AccessMode,
    durability_mode: DurabilityMode,
    commit_tier: CommitTier,
    reader_drain_timeout: Duration,
    max_resident_bytes: u64,
    max_temp_bytes: u64,
    epoch: Option<crate::epoch::StoreEpoch>,
    tokenizer: Option<crate::fts::tokenizer::TokenizerConfig>,
    schema: Option<crate::meta::Schema>,
    #[cfg(feature = "graph-cypher")]
    graph_document: Option<crate::epoch::EmbeddingTower>,
}

/// Filesystem authority requested for one store handle.
#[derive(Clone, Copy, Debug, Eq, PartialEq)]
pub enum AccessMode {
    /// Own the store's one kernel-enforced writer slot.
    ReadWrite,
    /// Load only the last committed snapshot without filesystem mutation.
    ReadOnly,
}

impl OpenOptions {
    /// Returns the default read-write, derived-durability configuration.
    #[must_use]
    pub const fn new() -> Self {
        Self {
            inspection: false,
            access_mode: AccessMode::ReadWrite,
            durability_mode: DurabilityMode::Derived,
            commit_tier: CommitTier::Ordered,
            reader_drain_timeout: DEFAULT_READER_DRAIN_TIMEOUT,
            max_resident_bytes: u64::MAX,
            max_temp_bytes: u64::MAX,
            epoch: None,
            tokenizer: None,
            schema: None,
            #[cfg(feature = "graph-cypher")]
            graph_document: None,
        }
    }

    /// Returns a pure-read configuration that takes no writer lock.
    #[must_use]
    pub const fn read_only() -> Self {
        Self {
            inspection: false,
            access_mode: AccessMode::ReadOnly,
            durability_mode: DurabilityMode::Derived,
            commit_tier: CommitTier::Ordered,
            reader_drain_timeout: DEFAULT_READER_DRAIN_TIMEOUT,
            max_resident_bytes: u64::MAX,
            max_temp_bytes: u64::MAX,
            epoch: None,
            tokenizer: None,
            schema: None,
            #[cfg(feature = "graph-cypher")]
            graph_document: None,
        }
    }

    /// Selects the existing durability-policy mode and per-commit tier.
    #[must_use]
    pub const fn with_durability(mut self, mode: DurabilityMode, tier: CommitTier) -> Self {
        self.durability_mode = mode;
        self.commit_tier = tier;
        self
    }

    /// Replaces the requested filesystem authority. The graph store facade
    /// owns its access mode, so this stays crate-private.
    #[cfg(feature = "graph-cypher")]
    #[must_use]
    pub(crate) const fn with_access_mode(mut self, access_mode: AccessMode) -> Self {
        self.access_mode = access_mode;
        self
    }

    /// Sets the grace period close gives admitted readers before cancellation.
    #[must_use]
    pub const fn with_reader_drain_timeout(mut self, timeout: Duration) -> Self {
        self.reader_drain_timeout = timeout;
        self
    }

    /// Sets the exact ceiling for live engine-owned anonymous bytes.
    #[must_use]
    pub const fn with_max_resident_bytes(mut self, bytes: u64) -> Self {
        self.max_resident_bytes = bytes;
        self
    }

    /// Declares the typed user-column schema when creating a store.
    ///
    /// On reopen, the declaration must be an additive evolution of the
    /// persisted schema (see [`crate::meta::Schema::additive_evolution`]).
    /// A read-write open commits added nullable columns as one generation;
    /// rows written earlier read them as null. A read-only open, or any
    /// other difference, is [`StoreError::SchemaMismatch`].
    #[must_use]
    pub fn with_schema(mut self, schema: crate::meta::Schema) -> Self {
        self.schema = Some(schema);
        self
    }

    /// Sets the exact ceiling for live temporary anonymous bytes.
    #[must_use]
    pub const fn with_max_temp_bytes(mut self, bytes: u64) -> Self {
        self.max_temp_bytes = bytes;
        self
    }

    /// Declares the embedding and tokenizer identity used by this handle.
    ///
    /// The tokenizer configuration defaults to
    /// [`crate::fts::tokenizer::TokenizerConfig::text_default`].
    /// A non-default tokenizer epoch must be paired with [`Self::with_tokenizer`].
    #[must_use]
    pub fn with_epoch(mut self, epoch: crate::epoch::StoreEpoch) -> Self {
        self.epoch = Some(epoch);
        self
    }

    /// Selects the tokenizer configuration used for indexing and queries.
    ///
    /// Open rejects this configuration when its epoch differs from the
    /// declared or persisted store identity.
    #[must_use]
    pub fn with_tokenizer(mut self, tokenizer: crate::fts::tokenizer::TokenizerConfig) -> Self {
        self.tokenizer = Some(tokenizer);
        self
    }
}

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

/// Traversal controls carried only when a caller explicitly selects the graph tier.
#[derive(Clone, Copy, Debug, Eq, PartialEq)]
pub struct GraphSearchOptions {
    profile: crate::graph::search::GraphSearchProfile,
    ef: Option<usize>,
    seed: u64,
    #[cfg(any(test, feature = "test-seams"))]
    cancel_after_hops: Option<usize>,
}

impl GraphSearchOptions {
    /// Creates an adaptive-width graph request for one dataset-shape profile.
    #[must_use]
    pub const fn new(profile: crate::graph::search::GraphSearchProfile) -> Self {
        Self {
            profile,
            ef: None,
            seed: 0,
            #[cfg(any(test, feature = "test-seams"))]
            cancel_after_hops: None,
        }
    }

    /// Overrides the profile's adaptive traversal width.
    #[must_use]
    pub const fn with_ef(mut self, ef: usize) -> Self {
        self.ef = Some(ef);
        self
    }

    /// Selects the deterministic Bit4 query-preparation seed.
    #[must_use]
    pub const fn with_seed(mut self, seed: u64) -> Self {
        self.seed = seed;
        self
    }

    /// Arms deterministic in-traversal cancellation for integration tests.
    #[cfg(any(test, feature = "test-seams"))]
    #[must_use]
    pub const fn with_cancel_after_hops(mut self, hops: usize) -> Self {
        self.cancel_after_hops = Some(hops);
        self
    }

    pub(crate) const fn profile(self) -> crate::graph::search::GraphSearchProfile {
        self.profile
    }

    pub(crate) const fn ef(self) -> Option<usize> {
        self.ef
    }

    pub(crate) const fn seed(self) -> u64 {
        self.seed
    }

    #[cfg(any(test, feature = "test-seams"))]
    pub(crate) const fn cancel_after_hops(self) -> Option<usize> {
        self.cancel_after_hops
    }
}

impl Default for GraphSearchOptions {
    fn default() -> Self {
        Self::new(crate::graph::search::GraphSearchProfile::SiftClass)
    }
}

pub(crate) fn auto_graph_search_options(
    snapshot: &PublishedSnapshot,
) -> Result<GraphSearchOptions, QueryError> {
    snapshot
        .graph_profile()
        .map(|profile| GraphSearchOptions::new(profile.search_profile()))
        .map_err(crate::graph::search::GraphSearchError::Profile)
        .map_err(QueryError::Graph)
}

/// Per-query execution tier at the public store seam.
#[derive(Clone, Copy, Debug, Default, Eq, PartialEq)]
pub enum SearchTier {
    /// Select graph traversal per sealed segment when that graph is published.
    #[default]
    Auto,
    /// Exhaustively score full-precision vectors. Large scans share at most
    /// four pool workers across callers, bounded further by the scan worker
    /// budget; small scans execute serially.
    Exact,
    /// Preserve the existing exhaustive sealed-segment scan.
    Scan,
    /// Traverse every sealed segment's Vamana graph.
    Graph(GraphSearchOptions),
}

/// Store-search controls, separating scan worker budget from tier selection.
#[derive(Clone, Copy, Debug, Default, Eq, PartialEq)]
pub struct SearchOptions {
    scan: crate::scan::ScanOptions,
    tier: Option<SearchTier>,
    scan_rescore: Option<ScanRescoreOptions>,
}

#[derive(Clone, Copy, Debug, Eq, PartialEq)]
enum GraphBoundMode {
    Shared,
    #[cfg(test)]
    Independent,
}

impl SearchOptions {
    /// Creates options with no tier preference and the supplied scan worker budget.
    #[must_use]
    pub const fn new(scan: crate::scan::ScanOptions) -> Self {
        Self {
            scan,
            tier: None,
            scan_rescore: None,
        }
    }

    /// Explicitly selects one store-search tier.
    #[must_use]
    pub const fn with_tier(mut self, tier: SearchTier) -> Self {
        self.tier = Some(tier);
        self.scan_rescore = None;
        self
    }

    /// Selects quantized Scan candidates followed by exact full-precision scores.
    /// Candidate membership remains approximate unless every eligible row is
    /// rescored. A subsequent `with_tier` call clears these controls.
    #[must_use]
    pub const fn with_scan_rescore(mut self, options: ScanRescoreOptions) -> Self {
        self.tier = Some(SearchTier::Scan);
        self.scan_rescore = Some(options);
        self
    }

    /// Explicit selected-row rescoring controls, absent for ordinary tiers.
    #[must_use]
    pub const fn scan_rescore(self) -> Option<ScanRescoreOptions> {
        self.scan_rescore
    }

    /// Returns the scan worker controls used by scan-tier work.
    #[must_use]
    pub const fn scan(self) -> crate::scan::ScanOptions {
        self.scan
    }

    /// Returns the effective execution tier, using automatic selection when
    /// the caller expressed no preference.
    #[must_use]
    pub const fn tier(self) -> SearchTier {
        match self.tier {
            Some(tier) => tier,
            None => SearchTier::Auto,
        }
    }

    pub(crate) const fn explicit_tier(self) -> Option<SearchTier> {
        self.tier
    }
}

impl From<crate::scan::ScanOptions> for SearchOptions {
    fn from(scan: crate::scan::ScanOptions) -> Self {
        Self::new(scan)
    }
}

/// Bit-set selecting the optional fields copied by [`Store::get_documents`].
#[derive(Clone, Copy, Debug, Default, Eq, PartialEq)]
pub struct DocumentFields(u8);

impl DocumentFields {
    /// Selects no optional fields.
    pub const NONE: Self = Self(0);
    /// Selects the full-precision vector.
    pub const VECTOR: Self = Self(1 << 0);
    /// Selects the stored UTF-8 text.
    pub const TEXT: Self = Self(1 << 1);
    /// Selects the opaque metadata bytes.
    pub const METADATA: Self = Self(1 << 2);
    /// Selects the typed schema attributes.
    pub const ATTRIBUTES: Self = Self(1 << 3);
    /// Selects every optional field.
    pub const ALL: Self =
        Self(Self::VECTOR.0 | Self::TEXT.0 | Self::METADATA.0 | Self::ATTRIBUTES.0);

    /// Returns whether every bit in `fields` is selected.
    #[must_use]
    pub const fn contains(self, fields: Self) -> bool {
        self.0 & fields.0 == fields.0
    }
}

impl std::ops::BitOr for DocumentFields {
    type Output = Self;

    fn bitor(self, rhs: Self) -> Self::Output {
        Self(self.0 | rhs.0)
    }
}

/// One live document copied from a single pinned store generation.
#[derive(Clone, Debug, PartialEq)]
pub struct StoredDocument {
    /// Stable application document identifier.
    pub doc_id: crate::ingest::DocId,
    /// Live monotonic revision.
    pub revision: crate::ingest::Revision,
    /// Canonical clustering timestamp.
    pub timestamp: i64,
    /// Selected full-precision vector, or `None` when not requested.
    pub vector: Option<Vec<f32>>,
    /// Selected stored UTF-8 text, or `None` when absent or not requested.
    pub text: Option<String>,
    /// Selected opaque metadata, or `None` when absent or not requested.
    pub metadata: Option<Vec<u8>>,
    /// Selected non-null typed attributes, or `None` when not requested.
    pub attributes: Option<Vec<(crate::meta::ColumnId, crate::meta::PredicateValue)>>,
}

/// Deterministic ordering used by [`Store::scan_documents`].
#[derive(Clone, Copy, Debug, Default, Eq, PartialEq)]
pub enum ScanOrder {
    /// Immutable segments in snapshot order and then the active segment, with
    /// ascending row numbers inside every source.
    #[default]
    Storage,
    /// Ascending timestamp with ascending document id as the tie breaker.
    TimestampAscending,
    /// Descending timestamp with ascending document id as the tie breaker.
    TimestampDescending,
    /// A declared `U64`, `I64` or `F64` user attribute in `direction`, with
    /// ascending document id as the tie breaker. `F64` compares numerically
    /// and `-0.0` equals `+0.0`. A row whose value is missing or NaN has no
    /// orderable value and sorts after every orderable row in both
    /// directions.
    Attribute {
        /// Declared user attribute that supplies the sort key.
        column: crate::meta::ColumnId,
        /// Direction applied to orderable values.
        direction: ScanDirection,
    },
}

/// Direction of an attribute-ordered document scan.
#[derive(Clone, Copy, Debug, Eq, PartialEq)]
pub enum ScanDirection {
    /// Smallest value first.
    Ascending,
    /// Largest value first.
    Descending,
}

/// Source position named by a document-scan continuation.
#[derive(Clone, Copy, Debug, Eq, PartialEq)]
pub enum DocumentScanSource {
    /// An immutable segment and its stable identifier.
    Sealed(crate::segment::SegmentId),
    /// The active in-memory segment.
    Active,
}

/// Opaque-to-callers continuation for one published store generation.
#[derive(Clone, Copy, Debug, Eq, PartialEq)]
pub struct DocumentScanCursor {
    /// Generation pinned by the page that produced this cursor.
    pub generation: u64,
    /// Order of the scan that issued this cursor; a request with any other
    /// order rejects it.
    pub order: ScanOrder,
    /// Source containing the next matching row.
    pub source: DocumentScanSource,
    /// Source-local row at which the next page starts.
    pub next_row: u32,
}

/// One bounded document-scan request.
#[derive(Clone, Debug)]
pub struct DocumentScanRequest<'a> {
    limit: usize,
    fields: DocumentFields,
    order: ScanOrder,
    cursor: Option<DocumentScanCursor>,
    timestamp_range: Option<(i64, i64)>,
    predicate: Option<&'a crate::meta::Predicate>,
    control: QueryControl,
}

impl<'a> DocumentScanRequest<'a> {
    /// Creates a storage-ordered scan without a filter or timestamp bound.
    #[must_use]
    pub fn new(limit: usize, fields: DocumentFields, control: QueryControl) -> Self {
        Self {
            limit,
            fields,
            order: ScanOrder::Storage,
            cursor: None,
            timestamp_range: None,
            predicate: None,
            control,
        }
    }

    /// Selects the deterministic row ordering.
    #[must_use]
    pub const fn with_order(mut self, order: ScanOrder) -> Self {
        self.order = order;
        self
    }

    /// Resumes at a continuation returned by an earlier page.
    #[must_use]
    pub const fn with_cursor(mut self, cursor: DocumentScanCursor) -> Self {
        self.cursor = Some(cursor);
        self
    }

    /// Restricts timestamps to the half-open interval `[start, end)`.
    #[must_use]
    pub const fn with_timestamp_range(mut self, start: i64, end: i64) -> Self {
        self.timestamp_range = Some((start, end));
        self
    }

    /// Applies one exact typed metadata predicate.
    #[must_use]
    pub const fn with_predicate(mut self, predicate: &'a crate::meta::Predicate) -> Self {
        self.predicate = Some(predicate);
        self
    }
}

/// One bounded page returned by [`Store::scan_documents`].
#[derive(Clone, Debug, PartialEq)]
pub struct DocumentScanPage {
    /// Materialized live documents in the requested deterministic order.
    pub documents: Vec<StoredDocument>,
    /// Store generation pinned for the complete call.
    pub generation: u64,
    /// Position of the next matching row, or `None` on the last page.
    pub continuation: Option<DocumentScanCursor>,
}

/// Exact live-row count from one pinned store generation.
#[derive(Clone, Copy, Debug, Eq, PartialEq)]
pub struct DocumentCount {
    /// Number of live rows matching the requested bounds.
    pub count: u64,
    /// Store generation pinned for the complete count.
    pub generation: u64,
}

/// Largest document page accepted by the core and C ABI.
pub const MAX_DOCUMENT_SCAN_LIMIT: usize = 1 << 20;

/// An open, lifecycle, or close operation was rejected.
#[derive(Debug)]
pub enum StoreError {
    /// A graph rule or coordinated document mutation refused retention.
    #[cfg(feature = "graph-cypher")]
    DocumentMutation(Box<crate::ingest::IngestError>),
    /// A declaration would introduce a namespace ownership cycle.
    CascadeCycle {
        /// Closed path of namespace names, in traversal order.
        cycle: Vec<String>,
    },
    /// A filesystem operation failed for the named store path.
    Io {
        /// Store path being opened.
        path: PathBuf,
        /// Underlying filesystem failure.
        source: std::io::Error,
    },
    /// The supplied path exists but is not a directory.
    NotDirectory {
        /// Invalid store path.
        path: PathBuf,
    },
    /// Another writer owns the store's process or kernel lock.
    StoreBusy {
        /// Busy store directory.
        path: PathBuf,
    },
    /// The writer lock could not be opened or acquired.
    Lock(StoreLockError),
    /// The selected durability mode/tier is unsupported.
    Durability(DurabilityPolicyError),
    /// Runtime kernel selection or an explicit override was invalid.
    Kernel(crate::kernels::KernelInitError),
    /// The committed manifest could not be loaded or validated.
    Manifest(crate::manifest::ManifestError),
    /// The selected tokenizer configuration could not be compiled.
    Tokenizer(crate::fts::tokenizer::TokenizerError),
    /// The caller's declared interpretation differs from the persisted one.
    EpochMismatch(crate::epoch::EpochMismatch),
    /// Persisted bytes name an epoch but the caller did not declare one.
    EpochUndeclared,
    /// A pre-epoch manifest cannot be retroactively assigned an identity.
    EpochUnstamped,
    /// A schema declaration on reopen differed from the committed schema.
    SchemaMismatch {
        /// Schema already committed in the manifest.
        persisted: crate::meta::Schema,
        /// Schema supplied by the opening caller.
        declared: crate::meta::Schema,
    },
    /// A scan continuation names a different published generation.
    ScanStale {
        /// Generation encoded by the caller's continuation.
        cursor_generation: u64,
        /// Generation pinned by this scan call.
        current_generation: u64,
    },
    /// A document-scan request or predicate is invalid.
    InvalidScan {
        /// Precise validation failure.
        detail: String,
    },
    /// A grouped count found more distinct values than its group limit.
    GroupLimitExceeded {
        /// Group limit supplied by the caller.
        limit: usize,
    },
    /// A referenced immutable segment could not be mapped or validated.
    Segment(crate::segment::SegmentError),
    /// More than one live row carried the same stable document id.
    DuplicateLiveDocument {
        /// Stable id whose uniqueness invariant was violated.
        doc_id: crate::ingest::DocId,
    },
    /// The caller selected graph traversal for a sealed segment without a graph.
    GraphUnavailable {
        /// Immutable segment that cannot satisfy the explicit graph-tier request.
        segment_id: crate::segment::SegmentId,
    },
    /// The checked durable WAL prefix could not be opened.
    Wal(crate::wal::WalReadError),
    /// A non-empty WAL did not replay through a clean end boundary.
    WalRecovery(crate::wal::WalRecoveryError),
    /// A replayed record's retained encoded bytes could not be decoded.
    WalRecord {
        /// Sequence assigned to the invalid retained record.
        seq: crate::wal::LogSeq,
        /// Checked record-access failure.
        source: crate::wal::VisibleRecordError,
    },
    /// A checksummed WAL mutation payload violated its versioned contract.
    WalMutation {
        /// Sequence assigned to the invalid mutation.
        seq: crate::wal::LogSeq,
        /// Persisted operation identifier.
        op: u16,
        /// Typed mutation-payload failure.
        source: crate::ingest::wal_payload::PayloadError,
    },
    /// A recovered upsert would make one document's revision non-monotonic.
    WalRevisionOrder {
        /// Sequence assigned to the invalid upsert.
        seq: crate::wal::LogSeq,
        /// Document whose recovered history is invalid.
        doc_id: crate::ingest::DocId,
        /// Revision already rebuilt from the trusted prefix.
        current: crate::ingest::Revision,
        /// Revision carried by this record.
        attempted: crate::ingest::Revision,
    },
    /// A valid mutation kind has no active-segment application path yet.
    UnsupportedWalMutation {
        /// Sequence assigned to the unsupported mutation.
        seq: crate::wal::LogSeq,
        /// Persisted operation identifier.
        op: u16,
    },
    /// A recovered vector passed payload validation but failed quantization.
    WalVector {
        /// Sequence assigned to the invalid vector mutation.
        seq: crate::wal::LogSeq,
        /// Typed vector-contract failure.
        source: crate::quant::QuantError,
    },
    /// The store-owned WAL writer could not be created or advanced.
    WalWrite(crate::wal::WalWriteError),
    /// Retained WAL visibility could not be released through a committed seal.
    WalRetire(crate::wal::WalRetireError),
    /// A kernel probe needed for an exact statistics snapshot failed.
    Statistics {
        /// Counter or residency component that could not be read.
        component: &'static str,
        /// Underlying operating-system failure.
        source: std::io::Error,
    },
    /// An accounted allocation would exceed a configured ceiling.
    BudgetExceeded {
        /// Total live bytes that would be needed if the operation proceeded.
        needed: u64,
        /// Configured ceiling in bytes.
        budget: u64,
        /// Engine component requesting the allocation.
        component: &'static str,
    },
    /// The allocator rejected a checked reservation without aborting.
    AllocationFailed {
        /// Exact requested bytes.
        needed: u64,
        /// Engine component requesting the allocation.
        component: &'static str,
    },
    /// An active vector did not match the collection's established dimension.
    DimensionMismatch {
        /// Established active dimension.
        expected: usize,
        /// Supplied vector dimension.
        actual: usize,
    },
    /// The active segment exceeded its dense u32 row address space.
    ActiveRowOverflow,
    /// An explicit seal was requested without any active rows.
    EmptyActiveSegment,
    /// Caller cancellation stopped a seal before its manifest commit point.
    SealCancelled,
    /// A write-only lifecycle operation was requested from a read-only handle.
    ReadOnly,
    /// WAL replay requires sealed tombstone repair that a read-only open cannot publish.
    SealedTombstoneRecoveryRequired,
    /// A prepared segment was accounted to a different store handle.
    ForeignPreparedSegment,
    /// The current snapshot generation cannot be incremented.
    GenerationOverflow,
    /// The active generation changed while a serialized writer prepared a commit.
    ConcurrentActiveMutation {
        /// Generation snapshotted before preparing the mutation.
        expected_generation: u64,
        /// Generation found when the writer re-acquired the active state.
        actual_generation: u64,
    },
    /// Exact reclaimed-byte reporting overflowed its u64 contract.
    PartitionBytesOverflow,
    /// A durable physical-purge intent could not be resumed during open.
    PurgeRecovery {
        /// Typed purge failure rendered without discarding its actionable values.
        detail: String,
    },
    /// A new operation raced with close after admissions stopped.
    Closing,
    /// The handle has completed teardown.
    Closed,
    /// Close cancelled an admitted read after its drain grace period elapsed.
    ReadCancelled,
    /// The lifecycle background thread could not be created.
    BackgroundStart {
        /// Operating-system thread creation failure.
        source: std::io::Error,
    },
    /// The lifecycle background thread exited before its startup handshake.
    BackgroundHandshake,
    /// The lifecycle background thread panicked before close joined it.
    BackgroundThreadPanicked,
    /// A persistent query worker thread could not be created.
    QueryPoolStart {
        /// Operating-system thread creation failure.
        source: std::io::Error,
    },
    /// A persistent query worker exited before its startup handshake.
    QueryPoolHandshake,
    /// A persistent query worker panicked before close joined it.
    QueryPoolThreadPanicked,
    /// The operating system could not report a usable query-worker count.
    QueryPoolCapacity {
        /// CPU-topology failure text.
        source: String,
    },
    /// A lifecycle synchronization primitive was poisoned.
    Synchronization {
        /// Synchronization component that rejected the operation.
        component: &'static str,
    },
    /// The directory holds a native graph store, which this legacy store
    /// cannot open. Nothing in the directory was changed.
    NativeGraphDirectory {
        /// Native graph store directory.
        path: PathBuf,
    },
    /// A snapshot target cannot receive a snapshot. Nothing was written.
    SnapshotTarget {
        /// Target path as the caller supplied it.
        path: PathBuf,
        /// Why the target was rejected.
        reason: SnapshotTargetReason,
    },
    /// The state pinned for a snapshot failed a consistency check. Nothing
    /// was published.
    SnapshotPin {
        /// The violated pin invariant.
        detail: &'static str,
    },
    /// A physical purge is pending; a snapshot would copy bytes it must
    /// remove. Await the purge, then snapshot. Nothing was written.
    SnapshotPurgePending,
    /// A physical removal must wait for open in-place snapshots to close.
    SnapshotViewsOpen,
}

impl std::fmt::Display for StoreError {
    fn fmt(&self, formatter: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        match self {
            #[cfg(feature = "graph-cypher")]
            Self::DocumentMutation(error) => error.fmt(formatter),
            Self::CascadeCycle { cycle } => {
                write!(formatter, "cascade cycle: {}", cycle.join(" -> "))
            }
            Self::Io { path, source } => {
                write!(formatter, "store I/O {}: {source}", path.display())
            }
            Self::NotDirectory { path } => {
                write!(
                    formatter,
                    "store path is not a directory: {}",
                    path.display()
                )
            }
            Self::SnapshotViewsOpen => {
                formatter.write_str("close open snapshots before physical removal")
            }
            Self::StoreBusy { path } => {
                write!(formatter, "store already has a writer: {}", path.display())
            }
            Self::Lock(error) => error.fmt(formatter),
            Self::Durability(error) => error.fmt(formatter),
            Self::Kernel(error) => error.fmt(formatter),
            Self::Manifest(error) => error.fmt(formatter),
            Self::Tokenizer(error) => write!(formatter, "store tokenizer: {error}"),
            Self::EpochMismatch(error) => error.fmt(formatter),
            Self::EpochUndeclared => {
                formatter.write_str("store epoch is persisted but the caller declared none")
            }
            Self::EpochUnstamped => formatter
                .write_str("store manifest predates epoch identity and cannot adopt a declaration"),
            Self::SchemaMismatch {
                persisted,
                declared,
            } => match persisted.additive_evolution(declared) {
                Err(conflict) => write!(
                    formatter,
                    "declared store schema does not match the persisted schema: {conflict}"
                ),
                // An additive declaration is refused only by a read-only
                // open, which cannot commit the evolved schema.
                Ok(Some(_)) => {
                    formatter.write_str(
                        "declared store schema adds attributes that a read-only open \
                         cannot persist:",
                    )?;
                    for added in declared
                        .columns()
                        .iter()
                        .filter(|column| persisted.column(column.id()).is_none())
                    {
                        write!(formatter, " '{}' (id {})", added.name(), added.id().get())?;
                    }
                    Ok(())
                }
                Ok(None) => write!(
                    formatter,
                    "declared store schema {declared:?} does not match persisted schema \
                     {persisted:?}"
                ),
            },
            Self::ScanStale {
                cursor_generation,
                current_generation,
            } => write!(
                formatter,
                "scan cursor generation {cursor_generation} differs from current generation {current_generation}"
            ),
            Self::InvalidScan { detail } => write!(formatter, "invalid document scan: {detail}"),
            Self::GroupLimitExceeded { limit } => write!(
                formatter,
                "grouped count found more than {limit} distinct values; raise the group limit"
            ),
            Self::Segment(error) => error.fmt(formatter),
            Self::DuplicateLiveDocument { doc_id } => {
                write!(
                    formatter,
                    "document {} has multiple live rows",
                    doc_id.get()
                )
            }
            Self::GraphUnavailable { segment_id } => {
                write!(formatter, "sealed segment {segment_id} has no graph region")
            }
            Self::Wal(error) => error.fmt(formatter),
            Self::WalRecovery(error) => error.fmt(formatter),
            Self::WalRecord { seq, source } => {
                write!(
                    formatter,
                    "WAL sequence {} retained record: {source}",
                    seq.get()
                )
            }
            Self::WalMutation { seq, op, source } => write!(
                formatter,
                "WAL sequence {} operation {op} payload: {source}",
                seq.get()
            ),
            Self::WalRevisionOrder {
                seq,
                doc_id,
                current,
                attempted,
            } => write!(
                formatter,
                "WAL sequence {} document {} revision {} does not follow {}",
                seq.get(),
                doc_id.get(),
                attempted.get(),
                current.get()
            ),
            Self::UnsupportedWalMutation { seq, op } => write!(
                formatter,
                "WAL sequence {} operation {op} has no active-segment recovery path",
                seq.get()
            ),
            Self::WalVector { seq, source } => {
                write!(formatter, "WAL sequence {} vector: {source}", seq.get())
            }
            Self::WalWrite(error) => error.fmt(formatter),
            Self::WalRetire(error) => error.fmt(formatter),
            Self::Statistics { component, source } => {
                write!(formatter, "store statistics {component}: {source}")
            }
            Self::BudgetExceeded {
                needed,
                budget,
                component,
            } => write!(
                formatter,
                "store {component} allocation needs {needed} bytes, budget is {budget} bytes"
            ),
            Self::AllocationFailed { needed, component } => write!(
                formatter,
                "store {component} allocator rejected {needed} bytes"
            ),
            Self::DimensionMismatch { expected, actual } => write!(
                formatter,
                "active vector dimension {actual} does not match {expected}"
            ),
            Self::ActiveRowOverflow => {
                formatter.write_str("active segment row or byte geometry overflow")
            }
            Self::EmptyActiveSegment => formatter.write_str("active segment is empty"),
            Self::SealCancelled => formatter.write_str("seal was cancelled before commit"),
            Self::ReadOnly => formatter.write_str("store handle is read-only"),
            Self::SealedTombstoneRecoveryRequired => formatter.write_str(
                "store requires writable sealed-tombstone recovery before reads are safe",
            ),
            Self::ForeignPreparedSegment => {
                formatter.write_str("prepared segment belongs to another store")
            }
            Self::GenerationOverflow => formatter.write_str("store snapshot generation overflow"),
            Self::ConcurrentActiveMutation {
                expected_generation,
                actual_generation,
            } => write!(
                formatter,
                "active generation changed from {expected_generation} to {actual_generation} while preparing a commit"
            ),
            Self::PartitionBytesOverflow => {
                formatter.write_str("partition reclaimed-byte count overflow")
            }
            Self::PurgeRecovery { detail } => {
                write!(formatter, "physical purge recovery failed: {detail}")
            }
            Self::Closing => formatter.write_str("store is closing"),
            Self::Closed => formatter.write_str("store is closed"),
            Self::ReadCancelled => formatter.write_str("store close cancelled the admitted read"),
            Self::BackgroundStart { source } => {
                write!(
                    formatter,
                    "store lifecycle thread could not start: {source}"
                )
            }
            Self::BackgroundHandshake => {
                formatter.write_str("store lifecycle thread startup handshake failed")
            }
            Self::BackgroundThreadPanicked => {
                formatter.write_str("store lifecycle thread panicked")
            }
            Self::QueryPoolStart { source } => {
                write!(formatter, "store query worker could not start: {source}")
            }
            Self::QueryPoolHandshake => {
                formatter.write_str("store query worker startup handshake failed")
            }
            Self::QueryPoolThreadPanicked => formatter.write_str("store query worker panicked"),
            Self::QueryPoolCapacity { source } => {
                write!(formatter, "store query worker capacity failed: {source}")
            }
            Self::Synchronization { component } => {
                write!(
                    formatter,
                    "store lifecycle synchronization poisoned: {component}"
                )
            }
            Self::NativeGraphDirectory { path } => write!(
                formatter,
                "native graph directory cannot be opened as a legacy store: {}",
                path.display()
            ),
            Self::SnapshotTarget { path, reason } => {
                write!(formatter, "snapshot target {} {reason}", path.display())
            }
            Self::SnapshotPin { detail } => {
                write!(formatter, "snapshot pin is inconsistent: {detail}")
            }
            Self::SnapshotPurgePending => formatter
                .write_str("a physical purge is pending; await it before taking a snapshot"),
        }
    }
}

/// Coarse, exhaustive classification of a [`StoreError`], stable for hosts
/// that map engine failures onto their own status codes. Every variant of
/// [`StoreError`] maps to exactly one kind; adding a variant is a compile
/// error here until it is classified.
#[derive(Clone, Copy, Debug, Eq, Hash, PartialEq)]
pub enum StoreErrorKind {
    /// An operating-system or filesystem operation failed.
    Io,
    /// A caller-supplied path or schema was invalid.
    InvalidArgument,
    /// Another process or handle owns the store's writer lock.
    StoreBusy,
    /// The requested mode or operation is unsupported.
    Unsupported,
    /// Persisted data failed validation.
    Corrupt,
    /// A configured memory, disk, or work budget was exceeded.
    BudgetExceeded,
    /// A checked allocation could not be reserved.
    OutOfMemory,
    /// Vector dimensions disagreed.
    DimensionMismatch,
    /// The declared epoch differs from the persisted identity.
    EpochMismatch,
    /// The store requires an epoch declaration and none was supplied.
    EpochUndeclared,
    /// An epoch was declared for a store with no stamped identity.
    EpochUnstamped,
    /// An internal invariant failed.
    Internal,
    /// A batch or active segment was empty.
    EmptyBatch,
    /// Cooperative cancellation stopped the operation.
    Cancelled,
    /// The store is read-only.
    ReadOnly,
    /// The store is closing.
    Closing,
    /// The store is closed.
    Closed,
    /// A background or query-pool thread panicked.
    Panic,
    /// A synchronization primitive was poisoned or unavailable.
    Synchronization,
}

impl StoreError {
    /// Classifies this error; see [`StoreErrorKind`].
    #[must_use]
    pub fn kind(&self) -> StoreErrorKind {
        match self {
            #[cfg(feature = "graph-cypher")]
            Self::DocumentMutation(error) => match error.as_ref() {
                crate::ingest::IngestError::Store(error) => error.kind(),
                _ => StoreErrorKind::InvalidArgument,
            },
            Self::Io { .. }
            // A failed manifest publication is I/O, not damaged persisted bytes.
            | Self::Manifest(crate::manifest::ManifestError::Io { .. })
            | Self::Lock(_)
            | Self::Statistics { .. }
            | Self::BackgroundStart { .. }
            | Self::QueryPoolStart { .. }
            | Self::WalWrite(_)
            | Self::WalRetire(_) => StoreErrorKind::Io,
            Self::CascadeCycle { .. }
            | Self::NotDirectory { .. }
            | Self::Tokenizer(_)
            | Self::SchemaMismatch { .. }
            | Self::ScanStale { .. }
            | Self::InvalidScan { .. }
            | Self::NativeGraphDirectory { .. }
            | Self::SnapshotTarget { .. } => StoreErrorKind::InvalidArgument,
            Self::StoreBusy { .. } | Self::SnapshotViewsOpen => StoreErrorKind::StoreBusy,
            Self::Manifest(crate::manifest::ManifestError::GraphUnsupportedBuild)
            | Self::Durability(_)
            | Self::Kernel(_)
            | Self::GraphUnavailable { .. }
            | Self::UnsupportedWalMutation { .. }
            | Self::SnapshotPurgePending => StoreErrorKind::Unsupported,
            Self::Manifest(_)
            | Self::Segment(_)
            | Self::DuplicateLiveDocument { .. }
            | Self::Wal(_)
            | Self::WalRecovery(_)
            | Self::WalRecord { .. }
            | Self::WalMutation { .. }
            | Self::WalRevisionOrder { .. }
            | Self::WalVector { .. }
            | Self::PurgeRecovery { .. } => StoreErrorKind::Corrupt,
            Self::BudgetExceeded { .. } | Self::GroupLimitExceeded { .. } => {
                StoreErrorKind::BudgetExceeded
            }
            Self::AllocationFailed { .. } => StoreErrorKind::OutOfMemory,
            Self::DimensionMismatch { .. } => StoreErrorKind::DimensionMismatch,
            Self::EpochMismatch(_) => StoreErrorKind::EpochMismatch,
            Self::EpochUndeclared => StoreErrorKind::EpochUndeclared,
            Self::EpochUnstamped => StoreErrorKind::EpochUnstamped,
            Self::ActiveRowOverflow
            | Self::GenerationOverflow
            | Self::ConcurrentActiveMutation { .. }
            | Self::PartitionBytesOverflow
            | Self::ForeignPreparedSegment
            | Self::BackgroundHandshake
            | Self::QueryPoolHandshake
            | Self::QueryPoolCapacity { .. }
            | Self::SnapshotPin { .. } => StoreErrorKind::Internal,
            Self::EmptyActiveSegment => StoreErrorKind::EmptyBatch,
            Self::SealCancelled | Self::ReadCancelled => StoreErrorKind::Cancelled,
            Self::ReadOnly | Self::SealedTombstoneRecoveryRequired => StoreErrorKind::ReadOnly,
            Self::Closing => StoreErrorKind::Closing,
            Self::Closed => StoreErrorKind::Closed,
            Self::BackgroundThreadPanicked | Self::QueryPoolThreadPanicked => StoreErrorKind::Panic,
            Self::Synchronization { .. } => StoreErrorKind::Synchronization,
        }
    }
}

impl std::error::Error for StoreError {
    fn source(&self) -> Option<&(dyn std::error::Error + 'static)> {
        match self {
            #[cfg(feature = "graph-cypher")]
            Self::DocumentMutation(error) => Some(error.as_ref()),
            Self::Io { source, .. } => Some(source),
            Self::Lock(error) => Some(error),
            Self::Durability(error) => Some(error),
            Self::Kernel(error) => Some(error),
            Self::Manifest(error) => Some(error),
            Self::Tokenizer(error) => Some(error),
            Self::EpochMismatch(error) => Some(error),
            Self::Segment(error) => Some(error),
            Self::Wal(error) => Some(error),
            Self::WalRecovery(error) => Some(error),
            Self::WalRecord { source, .. } => Some(source),
            Self::WalMutation { source, .. } => Some(source),
            Self::WalVector { source, .. } => Some(source),
            Self::WalWrite(error) => Some(error),
            Self::WalRetire(error) => Some(error),
            Self::Statistics { source, .. } => Some(source),
            Self::BackgroundStart { source } => Some(source),
            Self::QueryPoolStart { source } => Some(source),
            Self::CascadeCycle { .. }
            | Self::NotDirectory { .. }
            | Self::StoreBusy { .. }
            | Self::EpochUndeclared
            | Self::EpochUnstamped
            | Self::SchemaMismatch { .. }
            | Self::ScanStale { .. }
            | Self::InvalidScan { .. }
            | Self::GraphUnavailable { .. }
            | Self::DuplicateLiveDocument { .. }
            | Self::WalRevisionOrder { .. }
            | Self::UnsupportedWalMutation { .. }
            | Self::BudgetExceeded { .. }
            | Self::GroupLimitExceeded { .. }
            | Self::AllocationFailed { .. }
            | Self::DimensionMismatch { .. }
            | Self::ActiveRowOverflow
            | Self::EmptyActiveSegment
            | Self::SealCancelled
            | Self::ReadOnly
            | Self::SealedTombstoneRecoveryRequired
            | Self::ForeignPreparedSegment
            | Self::GenerationOverflow
            | Self::ConcurrentActiveMutation { .. }
            | Self::PartitionBytesOverflow
            | Self::PurgeRecovery { .. }
            | Self::Closing
            | Self::Closed
            | Self::ReadCancelled
            | Self::BackgroundHandshake
            | Self::BackgroundThreadPanicked
            | Self::QueryPoolHandshake
            | Self::QueryPoolThreadPanicked
            | Self::QueryPoolCapacity { .. }
            | Self::Synchronization { .. }
            | Self::NativeGraphDirectory { .. }
            | Self::SnapshotTarget { .. }
            | Self::SnapshotPin { .. }
            | Self::SnapshotPurgePending
            | Self::SnapshotViewsOpen => None,
        }
    }
}

/// Changes completed by this open. Format versions remain manifest v2 / WAL v1.
#[derive(Clone, Copy, Debug, Default, Eq, PartialEq)]
pub struct OpenMigrations {
    /// A nullable schema extension was committed.
    pub schema_added: bool,
    /// An incomplete final WAL record was cut atomically.
    pub wal_tail_cut: bool,
    /// Generation visible when open completed.
    pub generation: u64,
}

/// One explicitly closeable embedded-store handle.
pub struct Store {
    open_migrations: OpenMigrations,
    pub(crate) private_preparation: Option<namespace_batch::PrivatePreparation>,
    pub(crate) directory: PathBuf,
    pub(crate) vfs: Arc<dyn crate::vfs::Vfs>,
    pub(crate) clock: Arc<dyn MonotonicClock>,
    pub(crate) state: Mutex<StoreState>,
    pub(crate) state_changed: Condvar,
    pub(crate) background: Mutex<Option<BackgroundThread>>,
    pub(crate) query_pool: Mutex<Option<Arc<pool::QueryPool>>>,
    pub(crate) lexical_worker: Mutex<Option<Arc<pool::LexicalWorker>>>,
    #[cfg(feature = "graph-cypher")]
    pub(crate) native_graph: Arc<native_graph::NativeGraphPublication>,
    #[cfg(feature = "graph-cypher")]
    graph_enable_pending: std::sync::atomic::AtomicBool,
    pub(crate) snapshot: RwLock<Option<Arc<PublishedSnapshot>>>,
    // Drop order is deliberate: immutable mappings, active buffers, and the
    // WAL descriptor all release before the kernel writer lock.
    pub(crate) active: Mutex<Option<crate::ingest::ActiveState>>,
    pub(crate) wal_writer: Mutex<Option<crate::ingest::StoreWal>>,
    pub(crate) writer_lock: Mutex<Option<Arc<StoreLock>>>,
    #[cfg(feature = "graph-cypher")]
    reader_store_lock: Mutex<Option<StoreLock>>,
    snapshot_pins: Arc<AtomicU64>,
    snapshot_pin: Mutex<Option<snapshot_view::SnapshotPin>>,
    logical_writer_lock: Mutex<Option<StoreLock>>,
    reclamation_pin: Mutex<Option<Arc<StoreLock>>>,
    pub(crate) maintenance: Mutex<()>,
    pub(crate) health_state: Mutex<crate::diag::HealthState>,
    pub(crate) durability_policy: DurabilityPolicy,
    pub(crate) reader_drain_timeout: Duration,
    pub(crate) accounting: Arc<stats::Accounting>,
    pub(crate) active_queries: AtomicU64,
    #[cfg(any(test, feature = "test-seams"))]
    text_materialization_work: materialize::TestMaterializationWork,
    pub(crate) epoch: Option<crate::epoch::StoreEpoch>,
    pub(crate) epoch_alias: crate::epoch::EpochAliasCell,
    pub(crate) tokenizer: crate::fts::tokenizer::Analyzer,
    pub(crate) schema: crate::meta::Schema,
    // Holds only `Weak` handles to the snapshot and active segment it was
    // built from, so it never delays their release; see
    // `CachedLexicalAssembly`.
    lexical_index_cache: LexicalIndexCache,
    #[cfg(any(test, feature = "test-seams"))]
    pub(crate) ingest_retention_fault_controller:
        Option<crate::ingest::IngestRetentionFaultController>,
    #[cfg(any(test, feature = "test-seams"))]
    hybrid_leg_fault: Mutex<Option<HybridLegTestFault>>,
    #[cfg(any(test, feature = "test-seams"))]
    hybrid_execution_receipt: Mutex<Option<HybridExecutionReceipt>>,
    #[cfg(any(test, feature = "test-seams"))]
    pub(crate) metadata_test_controller: Option<Arc<crate::planner::MetadataTestController>>,
    #[cfg(any(test, feature = "test-seams"))]
    pub(crate) vector_fault_controller: Option<crate::scan::vector_fault::VectorFaultController>,
    #[cfg(any(test, feature = "test-seams"))]
    pub(crate) kernel_fault_controller: Option<crate::kernels::vector_fault::KernelFaultController>,
    #[cfg(any(test, feature = "test-seams"))]
    pub(crate) vector_seal_scheme: Option<crate::quant::QuantScheme>,
    #[cfg(test)]
    pub(crate) teardown_probe: Arc<close::TeardownProbe>,
}

fn acquire_writer_lock(
    path: &Path,
    access_mode: AccessMode,
) -> Result<Option<StoreLock>, StoreError> {
    Ok(match access_mode {
        AccessMode::ReadWrite => Some(StoreLock::acquire(path).map_err(|error| match error {
            StoreLockError::Io { path, source }
                if source.kind() == std::io::ErrorKind::WouldBlock =>
            {
                StoreError::StoreBusy { path }
            }
            StoreLockError::Io { path, source } => {
                StoreError::Lock(StoreLockError::Io { path, source })
            }
        })?),
        AccessMode::ReadOnly => None,
    })
}

fn refuse_native_graph_directory(vfs: &dyn crate::vfs::Vfs, path: &Path) -> Result<(), StoreError> {
    match vfs.ensure_directory(path, false) {
        Ok(true) => {}
        Ok(false) => return Ok(()),
        Err(source) if source.kind() == std::io::ErrorKind::NotFound => return Ok(()),
        Err(source) => {
            return Err(StoreError::Io {
                path: path.to_path_buf(),
                source,
            });
        }
    }
    // A catalog protected before a v3 commit can be an orphan beside a v2
    // document store after a crash. Only selector/WAL files mark the legacy
    // graph format there; graph-only directories remain refused.
    let document_store = is_legacy_store_directory(vfs, path)?;
    let mut native = false;
    vfs.for_each_direct_child(path, &mut |child| {
        let Some(name) = child.file_name().and_then(|name| name.to_str()) else {
            return Ok(());
        };
        if name == "graph-root.ze"
            || (name.starts_with("graph-wal-") && name.ends_with(".ze"))
            || (!document_store && name.starts_with("graph-") && name.ends_with(".zgraph"))
            || (name.starts_with("graph-root-") && name.ends_with(".tmp"))
        {
            native = true;
        }
        Ok(())
    })
    .map_err(|source| StoreError::Io {
        path: path.to_path_buf(),
        source,
    })?;
    if native {
        return Err(StoreError::NativeGraphDirectory {
            path: path.to_path_buf(),
        });
    }
    Ok(())
}

/// True when `path` is a directory holding a legacy store's manifest or WAL.
/// A missing path or a non-directory is not a legacy store; the graph open
/// path reports those itself. It only opens candidate files for reading.
pub(crate) fn is_legacy_store_directory(
    vfs: &dyn crate::vfs::Vfs,
    path: &Path,
) -> Result<bool, StoreError> {
    for name in [
        crate::manifest::io::MANIFEST_FILE,
        crate::manifest::io::MANIFEST_TEMP_FILE,
        "wal.ze",
    ] {
        let candidate = path.join(name);
        match vfs.open(&candidate) {
            Ok(_) => return Ok(true),
            Err(source)
                if matches!(
                    source.kind(),
                    std::io::ErrorKind::NotFound | std::io::ErrorKind::NotADirectory
                ) => {}
            Err(source) => {
                return Err(StoreError::Io {
                    path: candidate,
                    source,
                });
            }
        }
    }
    Ok(false)
}

/// Returns the schema the opened store serves and whether it is an additive
/// evolution of the committed schema that open must commit before publishing.
fn resolve_open_schema(
    manifest_exists: bool,
    snapshot: &PublishedSnapshot,
    declared: Option<&crate::meta::Schema>,
    access_mode: AccessMode,
) -> Result<(crate::meta::Schema, bool), StoreError> {
    if !manifest_exists {
        return Ok((
            declared
                .cloned()
                .unwrap_or_else(crate::meta::Schema::timestamp_only),
            false,
        ));
    }
    let persisted = snapshot.schema();
    let Some(declared) = declared else {
        return Ok((persisted.clone(), false));
    };
    let mismatch = || StoreError::SchemaMismatch {
        persisted: persisted.clone(),
        declared: declared.clone(),
    };
    match persisted.additive_evolution(declared) {
        Ok(None) => Ok((persisted.clone(), false)),
        Ok(Some(evolved)) if access_mode == AccessMode::ReadWrite => Ok((evolved, true)),
        Ok(Some(_)) | Err(_) => Err(mismatch()),
    }
}

fn validate_persisted_epoch_before_open(
    vfs: &dyn crate::vfs::Vfs,
    path: &Path,
    options: &OpenOptions,
) -> Result<(), StoreError> {
    let manifest_path = path.join(crate::manifest::io::MANIFEST_FILE);
    let bytes = match vfs.read(&manifest_path) {
        Ok(bytes) => bytes,
        Err(source) if source.kind() == std::io::ErrorKind::NotFound => return Ok(()),
        Err(source) => {
            return Err(StoreError::Io {
                path: manifest_path,
                source,
            });
        }
    };
    let manifest = crate::manifest::decode_manifest(&manifest_path.display().to_string(), &bytes)
        .map_err(StoreError::Manifest)?;
    let declared = if options.inspection {
        manifest.epoch_alias
    } else {
        options
            .epoch
            .as_ref()
            .map(crate::epoch::StoreEpoch::identity)
    };
    validate_epoch_identity(manifest.epoch_alias, declared, true, options.access_mode)
}

fn validate_epoch_identity(
    persisted: Option<crate::epoch::EpochIdentity>,
    declared: Option<crate::epoch::EpochIdentity>,
    manifest_exists: bool,
    access_mode: AccessMode,
) -> Result<(), StoreError> {
    match (persisted, declared) {
        (Some(expected), Some(declared)) if expected != declared => {
            Err(StoreError::EpochMismatch(crate::epoch::EpochMismatch {
                expected,
                declared,
            }))
        }
        (Some(_), None) => Err(StoreError::EpochUndeclared),
        (None, Some(_)) if manifest_exists || access_mode == AccessMode::ReadOnly => {
            Err(StoreError::EpochUnstamped)
        }
        (Some(_), Some(_)) | (None, Some(_)) | (None, None) => Ok(()),
    }
}

fn validate_tokenizer_epoch(
    expected: Option<crate::epoch::EpochIdentity>,
    tokenizer: &crate::fts::tokenizer::Analyzer,
) -> Result<(), StoreError> {
    if let Some(expected) = expected
        && tokenizer.epoch() != expected.tokenizer
    {
        return Err(StoreError::EpochMismatch(crate::epoch::EpochMismatch {
            expected,
            declared: crate::epoch::EpochIdentity {
                embedding: expected.embedding,
                tokenizer: tokenizer.epoch(),
            },
        }));
    }
    Ok(())
}

fn cleanup_open_orphans(
    store: &Store,
) -> Result<crate::manifest::io::OrphanCleanupReport, StoreError> {
    let (reachable_segments, graph_enabled) = store
        .snapshot
        .read()
        .map_err(|_| StoreError::Synchronization {
            component: "published snapshot",
        })?
        .as_ref()
        .map(|snapshot| {
            let reachable = snapshot
                .all_segments()
                .iter()
                .map(|segment| store.directory.join(segment.meta().id.file_name()))
                .collect::<HashSet<_>>();
            #[cfg(feature = "graph-cypher")]
            let graph_enabled = snapshot.graph_enabled;
            #[cfg(not(feature = "graph-cypher"))]
            let graph_enabled = false;
            (reachable, graph_enabled)
        })
        .ok_or(StoreError::Closed)?;
    let mut publication = Some(
        store
            .wal_writer
            .lock()
            .map_err(|_| StoreError::Synchronization {
                component: "WAL writer",
            })?
            .as_ref()
            .ok_or(StoreError::ReadOnly)?
            .manifest_publication()?,
    );
    let report = crate::manifest::io::cleanup_store_orphans(
        store.vfs.as_ref(),
        &store.directory,
        &reachable_segments,
        graph_enabled,
        store.durability_policy,
        || {
            // Listing and opening a candidate write nothing. Arm only before
            // the first unlink; retain the fence through the directory sync.
            publication = publication
                .take()
                .map(crate::ingest::ManifestPublication::arm);
        },
    )
    .map_err(StoreError::Manifest)?;
    if let Some(publication) = publication {
        publication.complete();
    }
    Ok(report)
}

impl Store {
    /// Opens an existing diagnostics copy read-only using its persisted identity.
    /// Tokenizer compatibility is still validated; this never repairs or writes.
    pub fn open_for_inspection(path: impl AsRef<Path>) -> Result<Self, StoreError> {
        let mut options = OpenOptions::read_only();
        options.inspection = true;
        Self::open(path, options)
    }

    /// Opens a store directory with the requested access and durability policy.
    pub fn open(path: impl AsRef<Path>, options: OpenOptions) -> Result<Self, StoreError> {
        let path = path.as_ref();
        namespace_batch::validate_store_path(path)?;
        if options.access_mode == AccessMode::ReadWrite {
            namespace_batch::reclaim_for_open(path)?;
        }
        let _admission =
            namespace_batch::reader_admission(path, options.access_mode == AccessMode::ReadWrite)?;
        let selected = namespace_batch::resolve(path)?;
        let logical_lock = if selected != path {
            acquire_writer_lock(path, options.access_mode)?
        } else {
            None
        };
        // A coordinator could commit between resolution and writer admission.
        if logical_lock.is_some() && namespace_batch::resolve(path)? != selected {
            return Err(StoreError::StoreBusy {
                path: path.to_path_buf(),
            });
        }
        if logical_lock.is_some() {
            // Adopt a decision that may have survived an interrupted directory
            // sync before acknowledging any later ordinary namespace write.
            if let Some(root) = path.parent() {
                crate::vfs::Vfs::sync(&crate::vfs::StdVfs, root, crate::vfs::SyncKind::Full)
                    .map_err(|source| StoreError::Io {
                        path: root.to_path_buf(),
                        source,
                    })?;
            }
        }
        let store = Self::open_with_infrastructure(
            &selected,
            options,
            Arc::new(crate::vfs::StdVfs),
            Arc::new(SystemMonotonicClock),
            None,
            #[cfg(any(test, feature = "test-seams"))]
            None,
            #[cfg(any(test, feature = "test-seams"))]
            None,
            #[cfg(any(test, feature = "test-seams"))]
            None,
            #[cfg(any(test, feature = "test-seams"))]
            None,
            #[cfg(any(test, feature = "test-seams"))]
            None,
            #[cfg(any(test, feature = "test-seams"))]
            None,
            #[cfg(any(test, feature = "test-seams"))]
            None,
        )?;
        if store
            .writer_lock
            .lock()
            .map_err(|_| StoreError::Synchronization {
                component: "writer lock",
            })?
            .is_some()
            && namespace_batch::resolve(path)? != selected
        {
            return Err(StoreError::StoreBusy {
                path: path.to_path_buf(),
            });
        }
        *store
            .logical_writer_lock
            .lock()
            .map_err(|_| StoreError::Synchronization {
                component: "logical writer lock",
            })? = logical_lock;
        Ok(store)
    }

    pub(crate) fn open_private_preparation(
        path: &Path,
        options: OpenOptions,
        authority: namespace_batch::PrivatePreparation,
    ) -> Result<Self, StoreError> {
        Self::open_with_infrastructure(
            path,
            options,
            Arc::new(crate::vfs::StdVfs),
            Arc::new(SystemMonotonicClock),
            Some(authority),
            #[cfg(any(test, feature = "test-seams"))]
            None,
            #[cfg(any(test, feature = "test-seams"))]
            None,
            #[cfg(any(test, feature = "test-seams"))]
            None,
            #[cfg(any(test, feature = "test-seams"))]
            None,
            #[cfg(any(test, feature = "test-seams"))]
            None,
            #[cfg(any(test, feature = "test-seams"))]
            None,
            #[cfg(any(test, feature = "test-seams"))]
            None,
        )
    }

    #[cfg(feature = "graph-cypher")]
    pub(crate) fn open_graph_with_infrastructure(
        path: &Path,
        mut options: OpenOptions,
        document: Option<crate::epoch::EmbeddingTower>,
        vfs: Arc<dyn crate::vfs::Vfs>,
        clock: Arc<dyn MonotonicClock>,
    ) -> Result<Self, StoreError> {
        options.graph_document = document;
        options.max_resident_bytes = options
            .max_resident_bytes
            .min(crate::property_graph::resources::MAX_GRAPH_RESIDENT_BYTES);
        Self::open_with_infrastructure(
            path,
            options,
            vfs,
            clock,
            None,
            #[cfg(any(test, feature = "test-seams"))]
            None,
            #[cfg(any(test, feature = "test-seams"))]
            None,
            #[cfg(any(test, feature = "test-seams"))]
            None,
            #[cfg(any(test, feature = "test-seams"))]
            None,
            #[cfg(any(test, feature = "test-seams"))]
            None,
            #[cfg(any(test, feature = "test-seams"))]
            None,
            #[cfg(any(test, feature = "test-seams"))]
            None,
        )
    }

    #[allow(
        clippy::too_many_arguments,
        reason = "test-support controllers are explicit optional infrastructure dependencies"
    )]
    pub(crate) fn open_with_infrastructure(
        path: impl AsRef<Path>,
        options: OpenOptions,
        vfs: Arc<dyn crate::vfs::Vfs>,
        clock: Arc<dyn MonotonicClock>,
        private_preparation: Option<namespace_batch::PrivatePreparation>,
        #[cfg(any(test, feature = "test-seams"))] hybrid_leg_fault: Option<HybridLegTestFault>,
        #[cfg(any(test, feature = "test-seams"))] storage_fault_controller: Option<
            StorageFaultController,
        >,
        #[cfg(any(test, feature = "test-seams"))] ingest_retention_fault_controller: Option<
            crate::ingest::IngestRetentionFaultController,
        >,
        #[cfg(any(test, feature = "test-seams"))] metadata_test_controller: Option<
            Arc<crate::planner::MetadataTestController>,
        >,
        #[cfg(any(test, feature = "test-seams"))] vector_fault_controller: Option<
            crate::scan::vector_fault::VectorFaultController,
        >,
        #[cfg(any(test, feature = "test-seams"))] kernel_fault_controller: Option<
            crate::kernels::vector_fault::KernelFaultController,
        >,
        #[cfg(any(test, feature = "test-seams"))] vector_seal_scheme: Option<
            crate::quant::QuantScheme,
        >,
    ) -> Result<Self, StoreError> {
        let path = path.as_ref();
        #[cfg(any(test, feature = "test-seams"))]
        match kernel_fault_controller.as_ref() {
            Some(controller) => {
                controller
                    .initialize_for_store()
                    .map_err(StoreError::Kernel)?;
            }
            None => {
                crate::kernels::initialize().map_err(StoreError::Kernel)?;
            }
        }
        #[cfg(not(any(test, feature = "test-seams")))]
        crate::kernels::initialize().map_err(StoreError::Kernel)?;
        #[cfg(any(test, feature = "test-seams"))]
        let vfs = match storage_fault_controller.as_ref() {
            Some(controller) => Arc::new(StorageFaultVfs {
                inner: vfs,
                controller: controller.clone(),
            }) as Arc<dyn crate::vfs::Vfs>,
            None => vfs,
        };
        let durability_policy = DurabilityPolicy::new(options.durability_mode, options.commit_tier)
            .map_err(StoreError::Durability)?;
        refuse_native_graph_directory(vfs.as_ref(), path)?;
        let accounting = Arc::new(stats::Accounting::new(
            options.max_resident_bytes,
            options.max_temp_bytes,
        ));
        let is_directory = vfs
            .ensure_directory(path, options.access_mode == AccessMode::ReadWrite)
            .map_err(|source| StoreError::Io {
                path: path.to_path_buf(),
                source,
            })?;
        if !is_directory {
            return Err(StoreError::NotDirectory {
                path: path.to_path_buf(),
            });
        }
        validate_persisted_epoch_before_open(vfs.as_ref(), path, &options)?;
        let _reclamation_admission = if private_preparation.is_some() {
            namespace_batch::validate_private_preparation(
                vfs.as_ref(),
                path,
                private_preparation.as_ref(),
            )?;
            None
        } else {
            namespace_batch::reader_admission(path, options.access_mode == AccessMode::ReadWrite)?
        };
        let reclamation_pin =
            namespace_batch::reader_lease(path, options.access_mode == AccessMode::ReadWrite)?
                .map(Arc::new);
        namespace_batch::refuse_retired(vfs.as_ref(), path)?;
        #[cfg(feature = "graph-cypher")]
        let reader_store_lock = if options.access_mode == AccessMode::ReadOnly
            && vfs
                .read_range(&path.join(crate::manifest::io::MANIFEST_FILE), 10, 2)
                .is_ok_and(|version| version.as_slice() == 3_u16.to_le_bytes())
        {
            Some(
                StoreLock::acquire_shared(path).map_err(|error| match error {
                    StoreLockError::Io { path, source }
                        if source.kind() == std::io::ErrorKind::WouldBlock =>
                    {
                        StoreError::StoreBusy { path }
                    }
                    error => StoreError::Lock(error),
                })?,
            )
        } else {
            None
        };
        let writer_lock = acquire_writer_lock(path, options.access_mode)?;
        if options.access_mode == AccessMode::ReadWrite && private_preparation.is_none() {
            namespace_batch::sync_portable_publication_for_open(vfs.as_ref(), path)?;
        }
        // Recheck after reclamation admission and the writer lock: the first
        // read is only a refusal probe and cannot authorize any mutation.
        validate_persisted_epoch_before_open(vfs.as_ref(), path, &options)?;
        let mut wal_tail_cut = false;
        if options.access_mode == AccessMode::ReadWrite {
            // The single writer owns the WAL tail: a crash mid-append is cut
            // off here so recovery sees whole batches only (ZE-216).
            wal_tail_cut =
                crate::ingest::cut_interrupted_append(vfs.as_ref(), path, durability_policy)?;
        }
        if options.access_mode == AccessMode::ReadWrite && private_preparation.is_none() {
            namespace_batch::adopt_for_open(vfs.as_ref(), path)?;
        }
        let manifest_path = path.join(crate::manifest::io::MANIFEST_FILE);
        let manifest_exists = match vfs.open(&manifest_path) {
            Ok(_) => true,
            Err(source) if source.kind() == std::io::ErrorKind::NotFound => false,
            Err(source) => {
                return Err(StoreError::Io {
                    path: manifest_path,
                    source,
                });
            }
        };
        // A persisted store gets one bounded header probe through the
        // Store-owned VFS before mmap becomes the query data plane. Internal
        // manifest-only remaps deliberately skip this probe so their exact
        // zero-segment-read accounting contracts remain intact.
        let mut snapshot = PublishedSnapshot::load_for_open_on_vfs(
            path,
            &accounting,
            vfs.as_ref(),
            private_preparation.as_ref(),
        )?;
        #[cfg(feature = "graph-cypher")]
        if options.access_mode == AccessMode::ReadOnly
            && snapshot.graph_enabled
            && reader_store_lock.is_none()
        {
            // Graph enable raced the unlocked v2 probe. Retry open under the
            // shared graph lock rather than admitting this unlocked snapshot.
            return Err(StoreError::StoreBusy {
                path: path.join(lock::STORE_LOCK_FILE),
            });
        }
        let (schema, schema_evolved) = resolve_open_schema(
            manifest_exists,
            &snapshot,
            options.schema.as_ref(),
            options.access_mode,
        )?;
        let persisted_epoch = snapshot.epoch_alias();
        let declared_epoch = if options.inspection {
            persisted_epoch
        } else {
            options
                .epoch
                .as_ref()
                .map(crate::epoch::StoreEpoch::identity)
        };
        validate_epoch_identity(
            persisted_epoch,
            declared_epoch,
            manifest_exists,
            options.access_mode,
        )?;
        let tokenizer = crate::fts::tokenizer::Analyzer::new(
            options
                .tokenizer
                .clone()
                .unwrap_or_else(crate::fts::tokenizer::TokenizerConfig::text_default),
        )
        .map_err(StoreError::Tokenizer)?;
        validate_tokenizer_epoch(persisted_epoch.or(declared_epoch), &tokenizer)?;
        let absorbed_through = snapshot.absorbed_through();
        let wal_path = path.join("wal.ze");
        let recovered = crate::ingest::ActiveState::recover(
            vfs.as_ref(),
            &wal_path,
            snapshot.generation(),
            absorbed_through,
            snapshot.graph_absorbed_through,
            #[cfg(feature = "graph-cypher")]
            snapshot.graph_manifest.as_ref().map(|graph| &***graph),
            #[cfg(feature = "graph-cypher")]
            snapshot.graph_enabled,
            &accounting,
            &schema,
            &tokenizer,
            private_preparation.as_ref(),
        )?;
        let mut active = recovered.active;
        let recovered_wal = recovered.wal;
        let sealed_tombstones = recovered.tombstones;
        #[cfg(feature = "graph-cypher")]
        let graph_replay = recovered.graph;
        let mut open_publication = None;
        #[cfg(feature = "graph-cypher")]
        let mut pending_graph_schema = None;
        if schema_evolved {
            let boundary = recovered_wal
                .as_ref()
                .map_or(absorbed_through, crate::wal::CleanWalReader::durable_end);
            let generation = active
                .generation
                .checked_add(1)
                .ok_or(StoreError::GenerationOverflow)?;
            let mut manifest =
                crate::manifest::io::load_manifest(vfs.as_ref(), &manifest_path, boundary)
                    .map_err(StoreError::Manifest)?;
            manifest.generation = generation;
            manifest.schema = schema.clone();
            #[cfg(feature = "graph-cypher")]
            manifest
                .record_generation_bump(boundary)
                .map_err(StoreError::Manifest)?;
            let mut publication = crate::ingest::ManifestPublication::before_writer_creation();
            #[cfg(feature = "graph-cypher")]
            let defer_commit = snapshot.graph_enabled;
            #[cfg(not(feature = "graph-cypher"))]
            let defer_commit = false;
            if !defer_commit {
                publication
                    .commit_manifest(vfs.as_ref(), path, &manifest, durability_policy)
                    .map_err(StoreError::Manifest)?;
            }
            snapshot =
                PublishedSnapshot::from_manifest(vfs.as_ref(), path, &manifest, &accounting)?;
            active.generation = generation;
            #[cfg(feature = "graph-cypher")]
            if defer_commit {
                pending_graph_schema = Some(manifest);
            }
            open_publication = Some(publication);
        }
        if options.access_mode == AccessMode::ReadWrite
            && !manifest_exists
            && (options.epoch.is_some() || options.schema.is_some())
        {
            let mut publication = crate::ingest::ManifestPublication::before_writer_creation();
            publication
                .commit_manifest(
                    vfs.as_ref(),
                    path,
                    &crate::manifest::Manifest {
                        #[cfg(feature = "graph-cypher")]
                        graph: None,
                        generation: active.generation,
                        log_seq: 0,
                        segments: Vec::new(),
                        epochs: options
                            .epoch
                            .as_ref()
                            .cloned()
                            .map(crate::manifest::EpochMeta::from)
                            .into_iter()
                            .collect(),
                        epoch_alias: options
                            .epoch
                            .as_ref()
                            .map(crate::epoch::StoreEpoch::identity),
                        schema: schema.clone(),
                    },
                    durability_policy,
                )
                .map_err(StoreError::Manifest)?;
            open_publication = Some(publication);
        }
        let wal_writer = match options.access_mode {
            AccessMode::ReadWrite => Some(match recovered_wal {
                Some(recovered) => crate::ingest::StoreWal::resume(
                    vfs.as_ref(),
                    &wal_path,
                    recovered,
                    durability_policy,
                    absorbed_through,
                    &accounting,
                )?,
                None => crate::ingest::StoreWal::create(
                    Arc::clone(&vfs),
                    path,
                    &wal_path,
                    durability_policy,
                    &accounting,
                )?,
            }),
            AccessMode::ReadOnly => {
                drop(recovered_wal);
                None
            }
        };
        if let Some(publication) = open_publication.as_mut() {
            publication.bind_writer(wal_writer.as_ref().ok_or(StoreError::ReadOnly)?);
        }
        let background = match options.access_mode {
            AccessMode::ReadWrite => Some(BackgroundThread::start()?),
            AccessMode::ReadOnly => None,
        };
        #[cfg(feature = "graph-cypher")]
        let native_graph =
            native_graph::NativeGraphPublication::new(&accounting, snapshot.graph_enabled)?;
        #[cfg(test)]
        let (snapshot, background, teardown_probe) = {
            let mut snapshot = snapshot;
            let mut background = background;
            let teardown_probe = Arc::new(close::TeardownProbe::new());
            snapshot.set_teardown_probe(Arc::clone(&teardown_probe));
            if let Some(background) = background.as_mut() {
                background.set_teardown_probe(Arc::clone(&teardown_probe));
            }
            (snapshot, background, teardown_probe)
        };
        let mut store = Self {
            private_preparation,
            open_migrations: OpenMigrations {
                schema_added: schema_evolved,
                wal_tail_cut,
                generation: active.generation,
            },
            directory: path.to_path_buf(),
            vfs,
            clock,
            state: Mutex::new(StoreState::Open),
            state_changed: Condvar::new(),
            background: Mutex::new(background),
            query_pool: Mutex::new(None),
            lexical_worker: Mutex::new(None),
            #[cfg(feature = "graph-cypher")]
            native_graph,
            #[cfg(feature = "graph-cypher")]
            graph_enable_pending: std::sync::atomic::AtomicBool::new(false),
            snapshot: RwLock::new(None),
            active: Mutex::new(Some(active)),
            wal_writer: Mutex::new(wal_writer),
            writer_lock: Mutex::new(writer_lock.map(Arc::new)),
            #[cfg(feature = "graph-cypher")]
            reader_store_lock: Mutex::new(reader_store_lock),
            snapshot_pins: Arc::new(AtomicU64::new(0)),
            snapshot_pin: Mutex::new(None),
            logical_writer_lock: Mutex::new(None),
            reclamation_pin: Mutex::new(reclamation_pin),
            maintenance: Mutex::new(()),
            health_state: Mutex::new(crate::diag::HealthState::default()),
            durability_policy,
            reader_drain_timeout: options.reader_drain_timeout,
            accounting,
            active_queries: AtomicU64::new(0),
            #[cfg(any(test, feature = "test-seams"))]
            text_materialization_work: materialize::TestMaterializationWork::default(),
            epoch_alias: crate::epoch::EpochAliasCell::new(persisted_epoch.or(declared_epoch)),
            epoch: options.epoch,
            tokenizer,
            schema,
            lexical_index_cache: LexicalIndexCache::new(),
            #[cfg(any(test, feature = "test-seams"))]
            ingest_retention_fault_controller,
            #[cfg(any(test, feature = "test-seams"))]
            hybrid_leg_fault: Mutex::new(hybrid_leg_fault),
            #[cfg(any(test, feature = "test-seams"))]
            hybrid_execution_receipt: Mutex::new(None),
            #[cfg(any(test, feature = "test-seams"))]
            metadata_test_controller,
            #[cfg(any(test, feature = "test-seams"))]
            vector_fault_controller,
            #[cfg(any(test, feature = "test-seams"))]
            kernel_fault_controller,
            #[cfg(any(test, feature = "test-seams"))]
            vector_seal_scheme,
            #[cfg(test)]
            teardown_probe,
        };
        #[cfg(feature = "graph-cypher")]
        if let Some(graph) = snapshot.graph_manifest.as_ref() {
            native_graph::recovery::install_unified(
                &store,
                graph,
                snapshot.generation(),
                &graph_replay,
                options.access_mode,
                options.graph_document.clone(),
            )?;
        }
        #[cfg(feature = "graph-cypher")]
        if let Some(manifest) = pending_graph_schema {
            // Graph admission must accept the prospective snapshot before its
            // schema and generation become durable. The bound guard arms at
            // the first manifest write and completes after snapshot publication.
            open_publication
                .as_mut()
                .ok_or(StoreError::Synchronization {
                    component: "schema publication",
                })?
                .commit_manifest(store.vfs.as_ref(), path, &manifest, durability_policy)
                .map_err(StoreError::Manifest)?;
        }
        if options.access_mode == AccessMode::ReadWrite && manifest_exists {
            // An adopted manifest may be the survivor of a commit interrupted
            // between rename and directory sync. Make its dirent durable before
            // any new generation is acknowledged on top of it.
            match durability_policy.directory_sync() {
                SyncRequirement::Skip => {}
                SyncRequirement::Sync(kind) => {
                    store
                        .vfs
                        .sync(path, kind)
                        .map_err(|source| StoreError::Io {
                            path: path.to_path_buf(),
                            source,
                        })?;
                }
            }
        }
        #[cfg(feature = "graph-cypher")]
        let repair_before_purge = snapshot.graph_enabled;
        #[cfg(not(feature = "graph-cypher"))]
        let repair_before_purge = false;
        store.publish_snapshot(snapshot)?;
        if let Some(publication) = open_publication {
            publication.complete();
        }
        if repair_before_purge {
            // Graph folds must not absorb unrepaired sealed deletes.
            store.recover_sealed_tombstones(&sealed_tombstones)?;
        }
        #[cfg(feature = "graph-cypher")]
        if options.access_mode == AccessMode::ReadWrite
            && store.native_graph.has_pending_reclaim()?
        {
            for _ in 0..2 {
                let admission = store.admit_native_graph_maintenance().map_err(|error| {
                    StoreError::Manifest(crate::manifest::ManifestError::Decode(error.to_string()))
                })?;
                match store.commit_native_graph_maintenance(
                    &admission,
                    &QueryControl::Cancel(CancelToken::new()),
                ) {
                    Ok(_) => break,
                    Err(native_graph::NativeGraphError::StalePreparation) => continue,
                    Err(error) => {
                        return Err(StoreError::Manifest(
                            crate::manifest::ManifestError::Decode(error.to_string()),
                        ));
                    }
                }
            }
            if store.native_graph.has_pending_reclaim()? {
                return Err(StoreError::Manifest(
                    crate::manifest::ManifestError::Decode(
                        "pending reclaim recovery did not finish after its manifest fold"
                            .to_owned(),
                    ),
                ));
            }
        }
        store
            .recover_pending_physical_purge()
            .map_err(|error| StoreError::PurgeRecovery {
                detail: error.to_string(),
            })?;
        if !repair_before_purge {
            // Preserve main's exact graph-free generation, bytes and refusals.
            store.recover_sealed_tombstones(&sealed_tombstones)?;
        }
        if options.access_mode == AccessMode::ReadWrite {
            let cleanup_report = cleanup_open_orphans(&store)?;
            #[cfg(any(test, feature = "test-seams"))]
            if let Some(controller) = storage_fault_controller.as_ref() {
                controller.record_cleanup_report(cleanup_report);
            }
            #[cfg(not(any(test, feature = "test-seams")))]
            let _ = cleanup_report;
        }
        #[cfg(any(test, feature = "test-seams"))]
        if let Some(controller) = storage_fault_controller.as_ref() {
            let published = store
                .snapshot
                .read()
                .map_err(|_| StoreError::Synchronization {
                    component: "published snapshot",
                })?;
            let snapshot = published.as_ref().ok_or(StoreError::Closed)?;
            for segment in snapshot.all_segments() {
                register_storage_segment_controller(
                    &store.directory,
                    segment.meta().id,
                    controller,
                );
            }
        }
        store.open_migrations.generation = store
            .active
            .lock()
            .map_err(|_| StoreError::Synchronization {
                component: "active segment",
            })?
            .as_ref()
            .ok_or(StoreError::Closed)?
            .generation;
        Ok(store)
    }

    /// Opens a store with deterministic infrastructure for adversarial tests.
    #[cfg(any(test, feature = "test-seams"))]
    #[doc(hidden)]
    pub fn open_with_test_dependencies(
        path: impl AsRef<Path>,
        options: OpenOptions,
        dependencies: StoreTestDependencies,
    ) -> Result<Self, StoreError> {
        let controller = dependencies.storage_fault_controller.clone();
        with_storage_open_controller(controller.as_ref(), || {
            Self::open_with_infrastructure(
                path,
                options,
                dependencies.vfs,
                dependencies.clock,
                None,
                dependencies.hybrid_leg_fault,
                dependencies.storage_fault_controller,
                dependencies.ingest_retention_fault_controller,
                dependencies.metadata_test_controller,
                dependencies.vector_fault_controller,
                dependencies.kernel_fault_controller,
                dependencies.vector_seal_scheme,
            )
        })
    }

    /// Takes the most recent Store-owned hybrid execution receipt.
    #[cfg(any(test, feature = "test-seams"))]
    #[doc(hidden)]
    pub fn take_hybrid_execution_receipt(&self) -> Option<HybridExecutionReceipt> {
        self.hybrid_execution_receipt
            .lock()
            .ok()
            .and_then(|mut receipt| receipt.take())
    }

    /// Carries the committed epoch registry forward unchanged.
    ///
    /// A declared epoch is stamped once, during open, before any write is
    /// admitted, so every later commit only propagates what is already
    /// committed. Replacement-snapshot paths may not have a prior manifest
    /// value in hand, so a declared store reconstructs its already-committed
    /// singleton entry; an unstamped store still returns an empty registry.
    pub(crate) fn epoch_registry(
        &self,
        prior: &[crate::manifest::EpochMeta],
    ) -> Vec<crate::manifest::EpochMeta> {
        if prior.is_empty() {
            self.epoch
                .as_ref()
                .map(crate::manifest::EpochMeta::from)
                .into_iter()
                .collect()
        } else {
            prior.to_vec()
        }
    }

    /// Returns the immutable report of changes completed by this open.
    pub fn open_migrations(&self) -> OpenMigrations {
        self.open_migrations
    }

    /// Returns the published embedding and tokenizer identity, or `None`
    /// for a store that carries no stamped epoch.
    #[must_use]
    pub fn epoch_identity(&self) -> Option<crate::epoch::EpochIdentity> {
        self.epoch_alias.load()
    }

    /// Returns the immutable schema declared for this store.
    #[must_use]
    pub const fn schema(&self) -> &crate::meta::Schema {
        &self.schema
    }

    /// Returns the current explicit lifecycle state.
    pub fn state(&self) -> Result<StoreState, StoreError> {
        self.state
            .lock()
            .map(|state| *state)
            .map_err(|_| StoreError::Synchronization { component: "state" })
    }

    /// Acquires the complete immutable snapshot admitted for one read.
    pub fn snapshot(&self) -> Result<SnapshotLease, StoreError> {
        let state = self
            .state
            .lock()
            .map_err(|_| StoreError::Synchronization { component: "state" })?;
        match *state {
            StoreState::Open => {}
            StoreState::Closing => return Err(StoreError::Closing),
            StoreState::Closed => return Err(StoreError::Closed),
        }
        let active = self
            .active
            .lock()
            .map_err(|_| StoreError::Synchronization {
                component: "active segment",
            })?;
        let generation = active.as_ref().ok_or(StoreError::Closed)?.generation;
        let snapshot = self
            .snapshot
            .read()
            .map_err(|_| StoreError::Synchronization {
                component: "published snapshot",
            })?
            .as_ref()
            .cloned()
            .ok_or(StoreError::Closed)?;
        drop(active);
        drop(state);
        Ok(SnapshotLease::new_at(snapshot, generation))
    }

    /// Reads one public result row's typed metadata through the owning Store boundary.
    ///
    /// This hidden test-support adapter lets independent persistence tests pair
    /// row identities returned by a public query with the values the reopened
    /// Store actually exposes. It is not compiled into the shipping API.
    #[cfg(any(test, feature = "test-seams"))]
    #[doc(hidden)]
    pub fn test_metadata_row_values(
        &self,
        row_id: crate::ingest::GlobalRowId,
    ) -> Result<Vec<(crate::meta::ColumnId, Option<crate::meta::PredicateValue>)>, StoreError> {
        match row_id.source() {
            crate::ingest::RowSource::Active => {
                let row = usize::try_from(row_id.local_row())
                    .map_err(|_| StoreError::ActiveRowOverflow)?;
                let active = self
                    .active
                    .lock()
                    .map_err(|_| StoreError::Synchronization {
                        component: "active segment",
                    })?;
                let segment = &active.as_ref().ok_or(StoreError::Closed)?.segment;
                let timestamp = segment
                    .timestamps()
                    .get(row)
                    .copied()
                    .ok_or(StoreError::ActiveRowOverflow)?;
                let mut values = segment.column_values(row)?;
                values.push((
                    crate::meta::TIMESTAMP_COLUMN,
                    crate::meta::PredicateValue::I64(timestamp),
                ));
                self.schema
                    .columns()
                    .iter()
                    .map(|definition| {
                        let value = values
                            .iter()
                            .find(|(column, _)| *column == definition.id())
                            .map(|(_, value)| value.clone());
                        Ok((definition.id(), value))
                    })
                    .collect()
            }
            crate::ingest::RowSource::Sealed(segment_id) => {
                let snapshot = self.snapshot()?;
                let segment = snapshot
                    .segments()
                    .iter()
                    .find(|segment| segment.meta().id == segment_id)
                    .ok_or_else(|| {
                        StoreError::Segment(crate::segment::SegmentError::Geometry(format!(
                            "public metadata row names absent segment {segment_id}"
                        )))
                    })?;
                let columns = segment.columns().map_err(StoreError::Segment)?;
                if row_id.local_row() >= columns.row_count() {
                    return Err(StoreError::Segment(crate::segment::SegmentError::Geometry(
                        format!(
                            "public metadata row {} is outside {} rows",
                            row_id.local_row(),
                            columns.row_count()
                        ),
                    )));
                }
                columns
                    .schema()
                    .columns()
                    .iter()
                    .map(|definition| {
                        let column = columns.column(definition.id()).ok_or_else(|| {
                            StoreError::Segment(crate::segment::SegmentError::Geometry(format!(
                                "public metadata row is missing column {}",
                                definition.id().get()
                            )))
                        })?;
                        let value = match column {
                            crate::meta::Column::Id128(column) => column
                                .get(row_id.local_row())
                                .map(crate::meta::PredicateValue::Id128),
                            crate::meta::Column::U64(column) => column
                                .get(row_id.local_row())
                                .map(crate::meta::PredicateValue::U64),
                            crate::meta::Column::I64(column) => column
                                .get(row_id.local_row())
                                .map(crate::meta::PredicateValue::I64),
                            crate::meta::Column::F64(column) => column
                                .get(row_id.local_row())
                                .map(crate::meta::PredicateValue::F64),
                            crate::meta::Column::Bool(column) => column
                                .get(row_id.local_row())
                                .map(crate::meta::PredicateValue::Bool),
                            crate::meta::Column::DictionaryString(column) => column
                                .get(row_id.local_row())
                                .map(|value| crate::meta::PredicateValue::String(value.to_owned())),
                            crate::meta::Column::RawString(column) => column
                                .get(row_id.local_row())
                                .map(|value| crate::meta::PredicateValue::String(value.to_owned())),
                        };
                        Ok((definition.id(), value))
                    })
                    .collect()
            }
        }
    }

    /// Evaluates one source's actual typed metadata state for independent tests.
    #[cfg(any(test, feature = "test-seams"))]
    #[doc(hidden)]
    pub fn test_metadata_evaluate_source(
        &self,
        source: crate::ingest::RowSource,
        predicate: &crate::meta::Predicate,
    ) -> Result<(u32, Vec<u32>, Vec<u32>), StoreError> {
        let evaluate = |columns: &crate::meta::ColumnStore,
                        alive: &crate::meta::AliveSet|
         -> Result<(u32, Vec<u32>, Vec<u32>), StoreError> {
            let matches = crate::meta::evaluate(predicate, columns, alive).map_err(|error| {
                StoreError::Segment(crate::segment::SegmentError::Columns(format!(
                    "test metadata source evaluation: {error}"
                )))
            })?;
            Ok((
                columns.row_count(),
                alive.iter_alive().collect(),
                matches.iter().collect(),
            ))
        };
        match source {
            crate::ingest::RowSource::Active => {
                let active = self
                    .active
                    .lock()
                    .map_err(|_| StoreError::Synchronization {
                        component: "active segment",
                    })?;
                let segment = &active.as_ref().ok_or(StoreError::Closed)?.segment;
                let mut builder = crate::meta::ColumnStoreBuilder::new(self.schema.clone());
                for (row, timestamp) in segment.timestamps().iter().copied().enumerate() {
                    let values = segment.column_values(row)?;
                    builder
                        .push_row(timestamp, &crate::ingest::column_inputs(&values))
                        .map_err(|error| {
                            StoreError::Segment(crate::segment::SegmentError::Columns(
                                error.to_string(),
                            ))
                        })?;
                }
                let columns = builder.finish().map_err(|error| {
                    StoreError::Segment(crate::segment::SegmentError::Columns(error.to_string()))
                })?;
                let alive = segment.alive()?;
                evaluate(&columns, &alive)
            }
            crate::ingest::RowSource::Sealed(segment_id) => {
                let snapshot = self.snapshot()?;
                let segment = snapshot
                    .segments()
                    .iter()
                    .find(|segment| segment.meta().id == segment_id)
                    .ok_or_else(|| {
                        StoreError::Segment(crate::segment::SegmentError::Geometry(format!(
                            "metadata source evaluation names absent segment {segment_id}"
                        )))
                    })?;
                let columns = segment.columns().map_err(StoreError::Segment)?;
                let alive = segment.alive().map_err(StoreError::Segment)?;
                evaluate(&columns, &alive)
            }
        }
    }

    pub(crate) fn publish_snapshot(&self, snapshot: PublishedSnapshot) -> Result<(), StoreError> {
        let state = self
            .state
            .lock()
            .map_err(|_| StoreError::Synchronization { component: "state" })?;
        match *state {
            StoreState::Open => {}
            StoreState::Closing => return Err(StoreError::Closing),
            StoreState::Closed => return Err(StoreError::Closed),
        }
        let generation = snapshot.generation();
        let mut active = self
            .active
            .lock()
            .map_err(|_| StoreError::Synchronization {
                component: "active segment",
            })?;
        let active_generation = &mut active.as_mut().ok_or(StoreError::Closed)?.generation;
        *active_generation = (*active_generation).max(generation);
        let mut published = self
            .snapshot
            .write()
            .map_err(|_| StoreError::Synchronization {
                component: "published snapshot",
            })?;
        *published = Some(Arc::new(snapshot));
        drop(active);
        drop(state);
        Ok(())
    }

    /// Runs one exact parallel scan admitted against this store's snapshot.
    ///
    /// The query holds one snapshot lease for its complete execution and uses
    /// the store's lazily started persistent worker pool.
    pub fn top_k_with_options(
        &self,
        request: crate::scan::ScanRequest<'_>,
        k: usize,
        options: crate::scan::ScanOptions,
        control: QueryControl,
    ) -> Result<crate::scan::ScanOutcome, QueryError> {
        let control = control.with_clock(Arc::clone(&self.clock));
        let state = self
            .state
            .lock()
            .map_err(|_| QueryError::Store(StoreError::Synchronization { component: "state" }))?;
        match *state {
            StoreState::Open => {}
            StoreState::Closing => return Err(QueryError::Store(StoreError::Closing)),
            StoreState::Closed => return Err(QueryError::Store(StoreError::Closed)),
        }
        let snapshot = self
            .snapshot
            .read()
            .map_err(|_| {
                QueryError::Store(StoreError::Synchronization {
                    component: "published snapshot",
                })
            })?
            .as_ref()
            .cloned()
            .ok_or(QueryError::Store(StoreError::Closed))?;
        let mut pool_slot = self.query_pool.lock().map_err(|_| {
            QueryError::Store(StoreError::Synchronization {
                component: "query pool",
            })
        })?;
        if pool_slot.is_none() {
            let capacity = crate::scan::physical_thread_capacity().map_err(|error| {
                QueryError::Store(StoreError::QueryPoolCapacity {
                    source: error.to_string(),
                })
            })?;
            let pool =
                pool::QueryPool::start(capacity, &self.accounting).map_err(QueryError::Store)?;
            *pool_slot = Some(Arc::new(pool));
        }
        let pool = pool_slot.as_ref().cloned().ok_or({
            QueryError::Store(StoreError::Synchronization {
                component: "query pool initialization",
            })
        })?;
        self.active_queries.fetch_add(1, Ordering::Relaxed);
        let active = ActiveQuery {
            count: &self.active_queries,
        };
        drop(pool_slot);
        drop(state);

        let result = pool
            .execute(request, k, options, control, SnapshotLease::new(snapshot))
            .map(|mut outcome| {
                outcome.candidates.truncate(k);
                outcome
            });
        drop(active);
        result
    }

    /// Injects one persistent query-worker panic and joins the pool.
    ///
    /// This is test-support-only and exists so the lifecycle campaign reaches
    /// the real worker teardown/error path.
    #[cfg(any(test, feature = "test-seams"))]
    pub fn panic_query_worker_for_test(&self) -> Result<(), StoreError> {
        let pool = self
            .query_pool
            .lock()
            .map_err(|_| StoreError::Synchronization {
                component: "query pool",
            })?
            .as_ref()
            .cloned()
            .ok_or(StoreError::Synchronization {
                component: "query pool initialization",
            })?;
        pool.panic_one_and_join()
    }

    /// Searches the active segment plus every immutable segment in one pinned
    /// store generation, then merges one global top-k.
    ///
    /// Scan-tier segments use the persistent query pool. Explicit graph-tier
    /// segments traverse serially on the calling thread, while the active
    /// segment is still scanned. Both tiers share one cancellation state and
    /// `(RowSource, local_row)` addressing, so immutable row zero never collides
    /// across segments and manifest reordering cannot rename a sealed row.
    pub fn search(
        &self,
        request: crate::ingest::SearchRequest<'_>,
        k: usize,
        options: impl Into<SearchOptions>,
        control: QueryControl,
    ) -> Result<crate::ingest::SearchOutcome, QueryError> {
        self.search_with_graph_bound_mode(
            request,
            k,
            options.into(),
            control,
            GraphBoundMode::Shared,
        )
    }

    fn admit_lexical_query(&self) -> Result<AdmittedLexicalQuery<'_>, QueryError> {
        #[cfg(any(test, feature = "test-seams"))]
        self.text_materialization_work
            .admissions
            .fetch_add(1, Ordering::Relaxed);
        #[cfg(all(feature = "query-timing", any(test, feature = "test-seams")))]
        let lock_started = std::time::Instant::now();
        let state = self
            .state
            .lock()
            .map_err(|_| QueryError::Store(StoreError::Synchronization { component: "state" }))?;
        #[cfg(all(feature = "query-timing", any(test, feature = "test-seams")))]
        self.text_materialization_work
            .observe_lexical_lock(lock_started);
        match *state {
            StoreState::Open => {}
            StoreState::Closing => return Err(QueryError::Store(StoreError::Closing)),
            StoreState::Closed => return Err(QueryError::Store(StoreError::Closed)),
        }
        #[cfg(all(feature = "query-timing", any(test, feature = "test-seams")))]
        let lock_started = std::time::Instant::now();
        let active_guard = self.active.lock().map_err(|_| {
            QueryError::Store(StoreError::Synchronization {
                component: "active segment",
            })
        })?;
        #[cfg(all(feature = "query-timing", any(test, feature = "test-seams")))]
        self.text_materialization_work
            .observe_lexical_lock(lock_started);
        let active_state = active_guard
            .as_ref()
            .ok_or(QueryError::Store(StoreError::Closed))?;
        let generation = active_state.generation;
        let active = Arc::clone(&active_state.segment);
        #[cfg(all(feature = "query-timing", any(test, feature = "test-seams")))]
        let lock_started = std::time::Instant::now();
        let snapshot = self
            .snapshot
            .read()
            .map_err(|_| {
                QueryError::Store(StoreError::Synchronization {
                    component: "published snapshot",
                })
            })
            .inspect(|_| {
                #[cfg(all(feature = "query-timing", any(test, feature = "test-seams")))]
                self.text_materialization_work
                    .observe_lexical_lock(lock_started);
            })?
            .as_ref()
            .cloned()
            .ok_or(QueryError::Store(StoreError::Closed))?;
        self.active_queries.fetch_add(1, Ordering::Relaxed);
        let active_query = ActiveQuery {
            count: &self.active_queries,
        };
        drop(active_guard);
        drop(state);
        Ok(AdmittedLexicalQuery {
            generation,
            active,
            snapshot: snapshot::ReadSnapshot::new(snapshot),
            active_query,
        })
    }

    fn admit_document_read(&self) -> Result<AdmittedDocumentRead<'_>, StoreError> {
        let state = self
            .state
            .lock()
            .map_err(|_| StoreError::Synchronization { component: "state" })?;
        match *state {
            StoreState::Open => {}
            StoreState::Closing => return Err(StoreError::Closing),
            StoreState::Closed => return Err(StoreError::Closed),
        }
        let active_guard = self
            .active
            .lock()
            .map_err(|_| StoreError::Synchronization {
                component: "active segment",
            })?;
        let active_state = active_guard.as_ref().ok_or(StoreError::Closed)?;
        let generation = active_state.generation;
        let active = Arc::clone(&active_state.segment);
        let snapshot = self
            .snapshot
            .read()
            .map_err(|_| StoreError::Synchronization {
                component: "published snapshot",
            })?
            .as_ref()
            .cloned()
            .ok_or(StoreError::Closed)?;
        self.active_queries.fetch_add(1, Ordering::Relaxed);
        let active_query = ActiveQuery {
            count: &self.active_queries,
        };
        drop(active_guard);
        drop(state);
        Ok(AdmittedDocumentRead {
            generation,
            active,
            snapshot: snapshot::ReadSnapshot::new(snapshot),
            active_query,
        })
    }

    /// Prepares the current lexical assembly and prefix vocabulary without ranking.
    ///
    /// Call after open to move lazy assembly preparation out of the first query.
    /// Mutations can invalidate both caches. Cancellation may retain a complete
    /// assembly, but never publishes a partial vocabulary.
    pub fn warm_lexical(
        &self,
        control: QueryControl,
    ) -> Result<(), crate::ingest::StoreLexicalError> {
        let control = control.with_clock(Arc::clone(&self.clock));
        let admission = self.admit_lexical_query()?;
        let lease = SnapshotLease::new_at(Arc::clone(&admission.snapshot), admission.generation);
        let cancellation = QueryCancellation::new(&control, &lease);
        cancellation.check_graph().map_err(QueryError::Scan)?;
        let receipt = assemble_unfiltered_lexical_index(
            LexicalInputs {
                filter: None,
                generation: admission.generation,
                cache: &self.lexical_index_cache,
                snapshot: &admission.snapshot,
                active: &admission.active,
                accounting: &self.accounting,
            },
            true,
            Some(&cancellation),
        )
        .map_err(map_store_lexical_assembly_error)?;
        receipt.vocabulary(&self.accounting, &cancellation)?;
        Ok(())
    }

    /// Runs an exact structured lexical query over active and sealed text.
    pub fn search_lexical(
        &self,
        query: &crate::fts::search::TermQuery,
        k: usize,
        control: QueryControl,
    ) -> Result<crate::ingest::StoreLexicalSearchOutcome, crate::ingest::StoreLexicalError> {
        self.search_lexical_then(query, k, control, None, |outcome, _, _, _, _, _| {
            Ok(outcome)
        })
    }

    /// Runs lexical ranking with scan-compatible eligibility constraints.
    pub fn search_lexical_filtered(
        &self,
        query: &crate::fts::search::TermQuery,
        k: usize,
        control: QueryControl,
        filter: Option<&QueryFilter>,
    ) -> Result<crate::ingest::StoreLexicalSearchOutcome, crate::ingest::StoreLexicalError> {
        self.search_lexical_then(query, k, control, filter, |outcome, _, _, _, _, _| {
            Ok(outcome)
        })
    }

    fn search_lexical_then<R>(
        &self,
        query: &crate::fts::search::TermQuery,
        k: usize,
        control: QueryControl,
        filter: Option<&QueryFilter>,
        finish: impl FnOnce(
            crate::ingest::StoreLexicalSearchOutcome,
            &PublishedSnapshot,
            &crate::ingest::ActiveSegment,
            &LexicalAssembly,
            &[crate::fts::search::ScoredDoc],
            &QueryCancellation<'_>,
        ) -> Result<R, crate::ingest::StoreLexicalError>,
    ) -> Result<R, crate::ingest::StoreLexicalError> {
        let control = control.with_clock(Arc::clone(&self.clock));
        let started = self.clock.now();
        let admission_started = timing_start(self.clock.as_ref());
        let AdmittedLexicalQuery {
            generation,
            active,
            snapshot,
            active_query,
        } = self.admit_lexical_query()?;
        let admission_time = timing_elapsed(self.clock.as_ref(), admission_started);
        let lexical_started = timing_start(self.clock.as_ref());

        let lease = SnapshotLease::new_at(Arc::clone(&snapshot), generation);
        let cancellation = QueryCancellation::new(&control, &lease);
        cancellation.check_graph().map_err(QueryError::Scan)?;
        let assembly = assemble_lexical_index(
            LexicalInputs {
                filter,
                generation,
                cache: &self.lexical_index_cache,
                snapshot: &snapshot,
                active: &active,
                accounting: &self.accounting,
            },
            true,
            Some(&cancellation),
        )
        .map_err(map_store_lexical_assembly_error)?;
        let mut work = crate::fts::control::WorkCheck::new(|| {
            cancellation
                .check_graph()
                .map_err(QueryError::Scan)
                .map_err(crate::ingest::StoreLexicalError::from)
        });
        work.check_now()?;
        let mut allow_lists = Vec::with_capacity(assembly.alive_sets.len());
        for alive in &assembly.alive_sets {
            work.step()?;
            allow_lists.push(alive.alive_bitmap());
        }
        let lexical = crate::planner::search_lexical_filtered_refs_controlled(
            &assembly.index,
            query,
            k,
            crate::fts::bm25::Bm25Params::beir(),
            &allow_lists,
            None,
            || {
                cancellation
                    .check_graph()
                    .map_err(QueryError::Scan)
                    .map_err(crate::ingest::StoreLexicalError::from)
            },
        )?;
        cancellation.check_graph().map_err(QueryError::Scan)?;
        let mut candidates = Vec::with_capacity(lexical.result.hits.len());
        for hit in &lexical.result.hits {
            work.step()?;
            let document =
                structured_lexical_document(&snapshot, &active, &assembly.sources, hit.doc, true)
                    .map_err(map_store_lexical_document_error)?
                    .ok_or(QueryError::Store(StoreError::ActiveRowOverflow))?;
            candidates.push(crate::ingest::LexicalCandidate {
                document,
                score: hit.score,
            });
        }
        let mut diagnostics =
            crate::diag::QueryDiagnostics::lexical(crate::diag::LexicalDiagnostics {
                snapshot_generation: generation,
                indexed_through_seq: active
                    .indexed_through_seq()
                    .max(crate::wal::LogSeq::new(snapshot.absorbed_through())),
                requested_k: k,
                returned: candidates.len(),
                counters: lexical.result.counters,
                elapsed: self.clock.now().saturating_duration_since(started),
            });
        diagnostics.counters.lexical_cache_hits = usize::from(assembly.cache_hit);
        diagnostics.counters.lexical_cache_builds = usize::from(!assembly.cache_hit);
        diagnostics.timings =
            crate::diag::QUERY_TIMING_ENABLED.then(|| crate::diag::QueryTimings {
                admission: admission_time,
                lexical: timing_elapsed(self.clock.as_ref(), lexical_started),
                ..Default::default()
            });
        let outcome = crate::ingest::StoreLexicalSearchOutcome {
            candidates,
            generation,
            diagnostics,
        };
        work.check_now()?;
        let result = finish(
            outcome,
            &snapshot,
            &active,
            &assembly,
            &lexical.result.hits,
            &cancellation,
        );
        drop(active_query);
        result
    }

    /// Returns the stored UTF-8 body for one exact document revision.
    ///
    /// Active memory and immutable region 14 are searched without adding a
    /// second text format. `None` means either that the exact revision is not
    /// present or that its text field was absent.
    pub fn stored_text(
        &self,
        version: crate::ingest::DocumentVersion,
    ) -> Result<Option<String>, crate::ingest::StoreLexicalError> {
        let AdmittedLexicalQuery {
            active,
            snapshot,
            active_query,
            ..
        } = self.admit_lexical_query()?;
        #[cfg(any(test, feature = "test-seams"))]
        self.text_materialization_work
            .reverse_version_lookups
            .fetch_add(1, Ordering::Relaxed);
        if let Some((row, active_version, _)) = active.existing(version.doc_id())
            && active_version == version
        {
            let text = active
                .text(row)
                .map_err(crate::ingest::StoreLexicalError::from)?;
            drop(active_query);
            return Ok(text.map(str::to_owned));
        }
        for segment in snapshot.segments() {
            #[cfg(any(test, feature = "test-seams"))]
            self.text_materialization_work
                .reverse_version_lookups
                .fetch_add(1, Ordering::Relaxed);
            let Some(row) = segment
                .query_row_for_document_version(version)
                .map_err(crate::ingest::StoreLexicalError::from)?
            else {
                continue;
            };
            let text = segment
                .query_stored_text()
                .map_err(StoreError::Segment)
                .map_err(crate::ingest::StoreLexicalError::from)?
                .and_then(|rows| rows.row(row).flatten())
                .map(str::to_owned);
            drop(active_query);
            return Ok(text);
        }
        drop(active_query);
        Ok(None)
    }

    /// Reads live documents by stable id from one pinned generation.
    ///
    /// The result preserves request order and length. Unknown and tombstoned
    /// ids produce `None`; duplicate requested ids produce duplicate results.
    pub fn get_documents(
        &self,
        ids: &[crate::ingest::DocId],
        fields: DocumentFields,
    ) -> Result<Vec<Option<StoredDocument>>, StoreError> {
        self.get_documents_with_generation(ids, fields)
            .map(|(_, documents)| documents)
    }

    /// Reads documents and returns the exact generation pinned by the call.
    ///
    /// This is the binding-facing form of [`Self::get_documents`].
    #[doc(hidden)]
    pub fn get_documents_with_generation(
        &self,
        ids: &[crate::ingest::DocId],
        fields: DocumentFields,
    ) -> Result<(u64, Vec<Option<StoredDocument>>), StoreError> {
        let AdmittedDocumentRead {
            generation,
            active,
            snapshot,
            active_query,
        } = self.admit_document_read()?;
        let mut documents = Vec::new();
        documents
            .try_reserve_exact(ids.len())
            .map_err(|_| StoreError::AllocationFailed {
                needed: allocation_bytes::<Option<StoredDocument>>(ids.len()),
                component: "get documents",
            })?;
        for id in ids {
            let mut document = active
                .existing(*id)
                .filter(|(row, _, _)| !active.is_tombstoned(*row))
                .map(|(row, _, _)| materialize_active_document(&active, row, fields))
                .transpose()?;
            for segment in snapshot.segments() {
                for row in segment.query_rows_for_doc_id(*id)? {
                    if document.is_some() {
                        return Err(StoreError::DuplicateLiveDocument { doc_id: *id });
                    }
                    document = Some(materialize_sealed_document(segment, row, fields)?);
                }
            }
            documents.push(document);
        }
        drop(active_query);
        Ok((generation, documents))
    }

    /// Enumerates one deterministic, bounded page of live documents.
    ///
    /// The complete call is pinned to one generation. A continuation from a
    /// different generation is rejected with [`StoreError::ScanStale`].
    pub fn scan_documents(
        &self,
        request: DocumentScanRequest<'_>,
    ) -> Result<DocumentScanPage, QueryError> {
        let DocumentScanRequest {
            limit,
            fields,
            order,
            cursor,
            timestamp_range,
            predicate,
            control,
        } = request;
        if limit == 0 || limit > MAX_DOCUMENT_SCAN_LIMIT {
            return Err(QueryError::Store(StoreError::InvalidScan {
                detail: format!("limit must be in 1..={MAX_DOCUMENT_SCAN_LIMIT}, received {limit}"),
            }));
        }
        if let ScanOrder::Attribute { column, .. } = order {
            validate_scan_order_attribute(&self.schema, column)?;
        }
        if let Some(cursor) = cursor.filter(|cursor| cursor.order != order) {
            return Err(QueryError::Store(StoreError::InvalidScan {
                detail: format!(
                    "scan cursor was issued for order {:?}, but the request orders by {order:?}",
                    cursor.order
                ),
            }));
        }
        let predicate = document_scan_predicate(predicate, timestamp_range)?;
        if let Some(predicate) = &predicate {
            crate::planner::validate_predicate(predicate, &self.schema).map_err(|error| {
                QueryError::Store(StoreError::InvalidScan {
                    detail: error.to_string(),
                })
            })?;
        }
        let control = control.with_clock(Arc::clone(&self.clock));
        control.checkpoint()?;
        let AdmittedDocumentRead {
            generation,
            active,
            snapshot,
            active_query,
        } = self.admit_document_read().map_err(QueryError::Store)?;
        if let Some(cursor) = cursor.filter(|cursor| cursor.generation != generation) {
            return Err(QueryError::Store(StoreError::ScanStale {
                cursor_generation: cursor.generation,
                current_generation: generation,
            }));
        }
        let lease = SnapshotLease::new_at(Arc::clone(&snapshot), generation);
        let mut work = 0_usize;
        let rows = match order {
            ScanOrder::Storage => collect_storage_scan_rows(
                &snapshot,
                &active,
                &self.schema,
                predicate.as_ref(),
                cursor,
                limit,
                &control,
                &lease,
                &mut work,
            )?,
            ScanOrder::TimestampAscending
            | ScanOrder::TimestampDescending
            | ScanOrder::Attribute { .. } => collect_sorted_scan_rows(
                &snapshot,
                &active,
                &self.schema,
                predicate.as_ref(),
                cursor,
                limit,
                order,
                &control,
                &lease,
                &mut work,
            )?,
        };
        let continuation = rows.get(limit).map(|row| DocumentScanCursor {
            generation,
            order,
            source: row.source,
            next_row: row.row,
        });
        let result_len = rows.len().min(limit);
        let mut documents = Vec::new();
        documents.try_reserve_exact(result_len).map_err(|_| {
            QueryError::Store(StoreError::AllocationFailed {
                needed: allocation_bytes::<StoredDocument>(result_len),
                component: "scan documents",
            })
        })?;
        for row in rows.into_iter().take(result_len) {
            scan_control_checkpoint(&control, &lease, &mut work)?;
            let document = match row.source {
                DocumentScanSource::Active => {
                    materialize_active_document(&active, row.row as usize, fields)
                        .map_err(QueryError::Store)?
                }
                DocumentScanSource::Sealed(segment_id) => {
                    let segment = snapshot
                        .segments()
                        .iter()
                        .find(|segment| segment.meta().id == segment_id)
                        .ok_or_else(|| {
                            QueryError::Store(StoreError::InvalidScan {
                                detail: format!("scan row names absent segment {segment_id}"),
                            })
                        })?;
                    materialize_sealed_document(segment, row.row as usize, fields)
                        .map_err(QueryError::Store)?
                }
            };
            documents.push(document);
        }
        drop(active_query);
        Ok(DocumentScanPage {
            documents,
            generation,
            continuation,
        })
    }

    /// Counts live documents matching an optional predicate and timestamp range.
    pub fn count_documents(
        &self,
        predicate: Option<&crate::meta::Predicate>,
        timestamp_range: Option<(i64, i64)>,
    ) -> Result<DocumentCount, QueryError> {
        let predicate = document_scan_predicate(predicate, timestamp_range)?;
        if let Some(predicate) = &predicate {
            crate::planner::validate_predicate(predicate, &self.schema).map_err(|error| {
                QueryError::Store(StoreError::InvalidScan {
                    detail: error.to_string(),
                })
            })?;
        }
        let AdmittedDocumentRead {
            generation,
            active,
            snapshot,
            active_query,
        } = self.admit_document_read().map_err(QueryError::Store)?;
        let mut count = 0_u64;
        for segment in snapshot.segments() {
            count = count
                .checked_add(sealed_scan_rows(segment, predicate.as_ref())?.cardinality())
                .ok_or(QueryError::Scan(crate::scan::ScanError::ArithmeticOverflow))?;
        }
        count = count
            .checked_add(active_scan_rows(&active, &self.schema, predicate.as_ref())?.cardinality())
            .ok_or(QueryError::Scan(crate::scan::ScanError::ArithmeticOverflow))?;
        drop(active_query);
        Ok(DocumentCount { count, generation })
    }

    /// Ascending ids of the live documents in `active` and `segments` that
    /// match an already validated `predicate`, from the same per-segment
    /// bitmaps [`Self::count_documents`] sums. The caller pins that state.
    pub(crate) fn live_document_ids_matching(
        &self,
        active: &crate::ingest::ActiveSegment,
        segments: &[crate::segment::reader::SegmentReader],
        predicate: &crate::meta::Predicate,
    ) -> Result<Vec<crate::ingest::DocId>, QueryError> {
        let mut ids = Vec::new();
        for segment in segments {
            for row in sealed_scan_rows(segment, Some(predicate))?.iter() {
                ids.push(sealed_document_scan_row(segment, row)?.doc_id);
            }
        }
        for row in active_scan_rows(active, &self.schema, Some(predicate))?.iter() {
            ids.push(active_document_scan_row(active, row)?.doc_id);
        }
        ids.sort_unstable();
        ids.dedup();
        Ok(ids)
    }

    /// Runs a structured lexical query and returns provenance plus snippets
    /// copied from the exact row text pinned for this generation.
    pub fn search_lexical_structured(
        &self,
        query: &crate::fts::query::LexicalQuery,
        k: usize,
        snippet_bytes: usize,
        control: QueryControl,
    ) -> Result<crate::ingest::StoreStructuredLexicalSearchOutcome, crate::ingest::StoreLexicalError>
    {
        self.search_lexical_structured_filtered(query, k, snippet_bytes, control, None)
    }

    /// Runs structured lexical ranking with scan-compatible eligibility constraints.
    #[allow(clippy::too_many_arguments)]
    pub fn search_lexical_structured_filtered(
        &self,
        query: &crate::fts::query::LexicalQuery,
        k: usize,
        snippet_bytes: usize,
        control: QueryControl,
        filter: Option<&QueryFilter>,
    ) -> Result<crate::ingest::StoreStructuredLexicalSearchOutcome, crate::ingest::StoreLexicalError>
    {
        let control = control.with_clock(Arc::clone(&self.clock));
        let started = self.clock.now();
        let admission_started = timing_start(self.clock.as_ref());
        let AdmittedLexicalQuery {
            generation,
            active,
            snapshot,
            active_query,
        } = self.admit_lexical_query()?;
        let admission_time = timing_elapsed(self.clock.as_ref(), admission_started);
        let lexical_started = timing_start(self.clock.as_ref());

        let lease = SnapshotLease::new_at(Arc::clone(&snapshot), generation);
        let cancellation = QueryCancellation::new(&control, &lease);
        cancellation.check_graph().map_err(QueryError::Scan)?;
        let assembly = assemble_lexical_index(
            LexicalInputs {
                filter,
                generation,
                cache: &self.lexical_index_cache,
                snapshot: &snapshot,
                active: &active,
                accounting: &self.accounting,
            },
            true,
            Some(&cancellation),
        )
        .map_err(map_store_lexical_assembly_error)?;
        let index = &assembly.index;
        let sources = &assembly.sources;
        let mut expansions = assembly
            .expand(query, &self.accounting, &cancellation)
            .map_err(|error| match error {
                expansion::ExpansionError::Query(error) => {
                    crate::ingest::StoreLexicalError::Query(error)
                }
                expansion::ExpansionError::Index(error) => crate::ingest::StoreLexicalError::from(
                    crate::planner::LexicalFilterError::from(error),
                ),
                expansion::ExpansionError::Shape(error) => {
                    crate::ingest::StoreLexicalError::Structured(error)
                }
            })?;
        let allow_lists = assembly
            .alive_sets
            .iter()
            .map(|alive| alive.alive_bitmap())
            .collect::<Vec<_>>();
        let mut prepared_memory =
            stats::AccountedCounter::new(&self.accounting, stats::AllocationComponent::Temporary)
                .map_err(QueryError::Store)?;
        let mut work = crate::fts::control::WorkCheck::new(|| cancellation.check_graph());
        let fields = query
            .fields_controlled(&mut work)
            .map_err(QueryError::Scan)?;
        prepared_memory
            .set(
                crate::fts::query::PreparedWeightedQuery::allocation_bytes_controlled(
                    &expansions,
                    &fields,
                    &mut work,
                )
                .map_err(QueryError::Scan)?
                .ok_or(QueryError::Store(StoreError::ActiveRowOverflow))?,
            )
            .map_err(QueryError::Store)?;
        let mut prepared = if index.segments().is_empty() || expansions.is_empty() {
            None
        } else {
            Some(
                crate::fts::query::PreparedWeightedQuery::new_controlled(
                    index,
                    std::mem::take(&mut expansions),
                    fields,
                    || cancellation.check_graph(),
                )
                .map_err(|error| match error {
                    crate::fts::search::ControlledSearchError::Index(error) => {
                        crate::ingest::StoreLexicalError::from(
                            crate::planner::LexicalFilterError::from(error),
                        )
                    }
                    crate::fts::search::ControlledSearchError::Control(error) => {
                        crate::ingest::StoreLexicalError::from(QueryError::Scan(error))
                    }
                })?,
            )
        };
        let mut scratch_memory =
            stats::AccountedCounter::new(&self.accounting, stats::AllocationComponent::Temporary)
                .map_err(QueryError::Store)?;
        let mut phrase = prepared_lexical::PhraseEligibility::new(
            index,
            Some(query),
            &self.accounting,
            &cancellation,
        )?;
        let result = match &prepared {
            None => crate::fts::prune::weighted::WeightedResult::default(),
            Some(prepared) => crate::fts::prune::weighted::search(
                index,
                prepared,
                k,
                &allow_lists,
                || {
                    cancellation
                        .check_graph()
                        .map_err(QueryError::Scan)
                        .map_err(crate::ingest::StoreLexicalError::from)
                },
                |doc| {
                    phrase
                        .matches(doc, &cancellation)
                        .map_err(crate::ingest::StoreLexicalError::from)
                },
                |bytes| {
                    prepared_lexical::reserve_weighted_scratch(&mut scratch_memory, bytes)
                        .map_err(crate::ingest::StoreLexicalError::from)
                },
            )
            .map_err(|error| match error {
                crate::fts::search::ControlledSearchError::Index(error) => {
                    crate::ingest::StoreLexicalError::from(
                        crate::planner::LexicalFilterError::from(error),
                    )
                }
                crate::fts::search::ControlledSearchError::Control(error) => error,
            })?,
        };
        let mut counters = result.counters;
        let scored = result.hits;
        let mut provenance_memory =
            stats::AccountedCounter::new(&self.accounting, stats::AllocationComponent::Temporary)
                .map_err(QueryError::Store)?;
        let mut candidates = Vec::with_capacity(scored.len());
        let mut provenance_work = crate::fts::control::WorkCheck::new(|| {
            cancellation.check_graph().map_err(QueryError::Scan)
        });
        for (doc, score) in scored {
            let mut provenance = Vec::new();
            let segment = index
                .segments()
                .get(doc.segment as usize)
                .ok_or(QueryError::Store(StoreError::ActiveRowOverflow))?;
            let scoring = prepared
                .as_ref()
                .ok_or(QueryError::Store(StoreError::ActiveRowOverflow))?;
            for expansion in scoring.expansions() {
                cancellation.check_graph().map_err(QueryError::Scan)?;
                let fields = &scoring.scoring().query().fields;
                prepared_lexical::reserve_weighted_scratch(
                    &mut provenance_memory,
                    crate::fts::sealed::TermStream::allocation_bytes(
                        segment,
                        &expansion.term,
                        fields,
                    ),
                )?;
                if let Some(mut stream) = crate::fts::sealed::TermStream::open_sized_controlled(
                    segment,
                    &expansion.term,
                    fields,
                    &mut provenance_work,
                )? {
                    stream.seek_controlled(doc.row, &mut provenance_work)?;
                    if stream.current_row() == Some(doc.row) {
                        provenance.push(expansion.clone());
                        counters.postings_decoded = counters.postings_decoded.saturating_add(1);
                    }
                    counters.blocks_decoded = counters
                        .blocks_decoded
                        .saturating_add(stream.blocks_decoded());
                    counters.blocks_skipped = counters
                        .blocks_skipped
                        .saturating_add(stream.blocks_skipped());
                }
                provenance_memory.set(0).map_err(QueryError::Store)?;
            }
            cancellation.check_graph().map_err(QueryError::Scan)?;
            let (text, document) = structured_lexical_row(&snapshot, &active, sources, doc)?;
            let terms = provenance
                .iter()
                .map(|entry| entry.term.clone())
                .collect::<Vec<_>>();
            let snippet = crate::fts::snippet::best_window(
                &self.tokenizer,
                text,
                &terms,
                snippet_bytes,
                true,
            )?
            .ok_or(crate::ingest::StoreLexicalError::MissingSnippetMatch {
                segment: doc.segment,
                row: doc.row,
            })?;
            let snippet_text = snippet
                .text(text)
                .ok_or(QueryError::Store(StoreError::ActiveRowOverflow))?;
            candidates.push(crate::ingest::ExplainedLexicalCandidate {
                document,
                score,
                provenance,
                snippet: crate::fts::query::OwnedLexicalSnippet {
                    text: snippet_text.to_owned(),
                    source: snippet.window,
                    highlights: snippet.highlights,
                    source_len: text.len(),
                },
            });
        }
        cancellation.check_graph().map_err(QueryError::Scan)?;
        let mut diagnostics =
            crate::diag::QueryDiagnostics::lexical(crate::diag::LexicalDiagnostics {
                snapshot_generation: generation,
                indexed_through_seq: active
                    .indexed_through_seq()
                    .max(crate::wal::LogSeq::new(snapshot.absorbed_through())),
                requested_k: k,
                returned: candidates.len(),
                counters,
                elapsed: self.clock.now().saturating_duration_since(started),
            });
        diagnostics.counters.lexical_cache_hits = usize::from(assembly.cache_hit);
        diagnostics.counters.lexical_cache_builds = usize::from(!assembly.cache_hit);
        diagnostics.timings =
            crate::diag::QUERY_TIMING_ENABLED.then(|| crate::diag::QueryTimings {
                admission: admission_time,
                lexical: timing_elapsed(self.clock.as_ref(), lexical_started),
                ..Default::default()
            });
        drop(active_query);
        Ok(crate::ingest::StoreStructuredLexicalSearchOutcome {
            candidates,
            expansions: prepared
                .as_mut()
                .map_or(expansions, |query| query.take_expansions()),
            generation,
            diagnostics,
        })
    }

    /// Runs the store vector engine and exact lexical leg against one pinned
    /// generation, then delegates every blend decision to `crate::fusion`.
    pub fn search_hybrid(
        &self,
        vector_query: crate::ingest::SearchRequest<'_>,
        lexical_query: &crate::fts::search::TermQuery,
        hybrid_query: &crate::fusion::HybridQuery,
        options: impl Into<SearchOptions>,
        control: QueryControl,
    ) -> Result<crate::ingest::StoreHybridSearchOutcome, crate::fusion::FusionError> {
        self.search_hybrid_inner(
            vector_query,
            PinnedLexicalQuery::Term(lexical_query),
            hybrid_query,
            options.into(),
            control,
        )
    }

    /// Runs vector work with any structured lexical operator against the
    /// same pinned snapshot and generation.
    pub fn search_hybrid_structured(
        &self,
        vector_query: crate::ingest::SearchRequest<'_>,
        lexical_query: &crate::fts::query::LexicalQuery,
        hybrid_query: &crate::fusion::HybridQuery,
        options: impl Into<SearchOptions>,
        control: QueryControl,
    ) -> Result<crate::ingest::StoreHybridSearchOutcome, crate::fusion::FusionError> {
        self.search_hybrid_inner(
            vector_query,
            PinnedLexicalQuery::Structured(lexical_query),
            hybrid_query,
            options.into(),
            control,
        )
    }

    fn search_hybrid_inner(
        &self,
        vector_query: crate::ingest::SearchRequest<'_>,
        lexical_query: PinnedLexicalQuery<'_>,
        hybrid_query: &crate::fusion::HybridQuery,
        options: SearchOptions,
        control: QueryControl,
    ) -> Result<crate::ingest::StoreHybridSearchOutcome, crate::fusion::FusionError> {
        self.search_hybrid_then(
            vector_query,
            lexical_query,
            hybrid_query,
            options,
            control,
            false,
            |outcome, _, _, _, _| Ok(outcome),
        )
    }

    #[allow(clippy::too_many_arguments)] // The existing query arguments plus its scoped continuation.
    fn search_hybrid_then<R>(
        &self,
        vector_query: crate::ingest::SearchRequest<'_>,
        lexical_query: PinnedLexicalQuery<'_>,
        hybrid_query: &crate::fusion::HybridQuery,
        options: SearchOptions,
        control: QueryControl,
        capture_rows: bool,
        finish: impl FnOnce(
            crate::ingest::StoreHybridSearchOutcome,
            &PublishedSnapshot,
            &crate::ingest::ActiveSegment,
            &QueryCancellation<'_>,
            Option<&materialize::HybridAddresses>,
        ) -> Result<R, crate::fusion::FusionError>,
    ) -> Result<R, crate::fusion::FusionError> {
        self.search_hybrid_prepared_then(
            || Ok::<_, std::convert::Infallible>(vector_query),
            vector_query.filter,
            vector_query.eligible,
            lexical_query,
            hybrid_query,
            options,
            control,
            capture_rows,
            None,
            finish,
        )
        .map_err(|error| match error {
            HybridPreparationError::Preparation(never) => match never {},
            HybridPreparationError::Search(error) => error,
        })
    }

    #[allow(clippy::too_many_arguments)] // The existing query arguments plus its scoped continuation.
    fn search_hybrid_prepared_then<'vector, E, R>(
        &self,
        prepare_vector: impl FnOnce() -> Result<crate::ingest::SearchRequest<'vector>, E>,
        filter: Option<&QueryFilter>,
        known_eligible: Option<&[crate::ingest::DocId]>,
        lexical_query: PinnedLexicalQuery<'_>,
        hybrid_query: &crate::fusion::HybridQuery,
        options: SearchOptions,
        control: QueryControl,
        capture_rows: bool,
        pinned: Option<AdmittedVectorSearch<'_>>,
        finish: impl FnOnce(
            crate::ingest::StoreHybridSearchOutcome,
            &PublishedSnapshot,
            &crate::ingest::ActiveSegment,
            &QueryCancellation<'_>,
            Option<&materialize::HybridAddresses>,
        ) -> Result<R, crate::fusion::FusionError>,
    ) -> Result<R, HybridPreparationError<E>> {
        let control = if pinned.is_some() {
            control
        } else {
            control.with_clock(Arc::clone(&self.clock))
        };
        let started = self.clock.now();
        // Resolve known restrictions here, before either producer starts. Deferred
        // producers may supply ids later; their lexical ranking is refreshed below.
        let eligible_filter = known_eligible.map(|ids| {
            filter.cloned().map_or_else(
                || QueryFilter::eligible(&self.schema, ids),
                |filter| filter.with_eligible(ids),
            )
        });
        let filter = eligible_filter.as_ref().or(filter);
        let admission_started = timing_start(self.clock.as_ref());
        let mut options = options;
        let requested_tier = options.explicit_tier();
        // Hybrid once forced SearchTier::Exact unconditionally, because
        // fusion needs an anchor for unseen scores and only an exhaustive
        // scan could name the farthest alive row. The deterministic
        // ceiling now supplies that anchor on every tier, so hybrid
        // follows the caller's tier and reaches the graph like the dense
        // leg does.
        //
        // A segment with no published graph is different. There `Auto`
        // resolves to the Bit4 scan, whose scores are estimates, and
        // fusion may only fuse exactly rescored scores. Graph traversal
        // rescores its retained pool exactly, so it is safe; the Bit4
        // scan is not. When the caller stated no preference and no
        // segment has earned a graph yet, hybrid therefore still selects
        // Exact for itself. This is the adaptive ladder choosing the
        // tier that can honour the contract, not a fallback hiding one.
        let pinned_epoch = pinned
            .as_ref()
            .map(|admitted| admitted.snapshot.epoch_alias());
        let admitted = match pinned {
            Some(admitted) => admitted,
            None => self
                .admit_vector_search_for(options, requested_tier.is_none())
                .map_err(crate::fusion::FusionError::from)?,
        };
        options = admitted.execution_options;
        let tier_resolution = if requested_tier.is_none() {
            if snapshot_has_graph(&admitted.snapshot) {
                Some(crate::diag::HybridTierResolution::Auto)
            } else {
                options = options.with_tier(SearchTier::Exact);
                Some(crate::diag::HybridTierResolution::Exact)
            }
        } else {
            None
        };
        let corpus_rows = hybrid::corpus_rows(&admitted.snapshot, &admitted.active_segment)?;
        // Round zero asks each producer for the window plus one row, so the
        // element just past the window is the stability bound's unseen value.
        // A caller that disabled widening asks for the corpus at once, which
        // is the complete-list fusion this path used to run unconditionally.
        let mut width = if hybrid_query.max_rounds == 0 {
            corpus_rows
        } else {
            hybrid::hybrid_window(hybrid_query.k, corpus_rows)?.width
        };
        let panic_vector = self.consume_hybrid_test_fault(crate::fusion::FusionLeg::Vector);
        let panic_lexical = self.consume_hybrid_test_fault(crate::fusion::FusionLeg::Lexical);
        let lexical_worker = self.ensure_lexical_worker()?;
        let mut timings = crate::diag::QueryTimings {
            admission: timing_elapsed(self.clock.as_ref(), admission_started),
            ..Default::default()
        };
        let mut lexical_inputs = LexicalInputs {
            filter,
            generation: admitted.generation,
            cache: &self.lexical_index_cache,
            snapshot: &admitted.snapshot,
            active: &admitted.active_segment,
            accounting: &self.accounting,
        };
        #[cfg(any(test, feature = "test-seams"))]
        let preparation_observer = crate::fts::preparation_observer::current();
        let mut lexical_preparation = prepared_lexical::PreparedLexicalQuery::new(lexical_query);
        #[cfg(any(test, feature = "test-seams"))]
        let fresh_preparation = hybrid::fresh_round_test_control();
        let run_lexical_leg = |lexical_inputs: LexicalInputs<'_>,
                               preparation: &mut prepared_lexical::PreparedLexicalQuery<'_>,
                               bound,
                               queued| {
            #[cfg(any(test, feature = "test-seams"))]
            if fresh_preparation {
                preparation.reset_for_test();
            }
            #[cfg(any(test, feature = "test-seams"))]
            let _preparation_scope =
                crate::fts::preparation_observer::enter(preparation_observer.clone());
            let queue_time = timing_elapsed(self.clock.as_ref(), queued);
            let lexical_started = timing_start(self.clock.as_ref());
            let name = std::thread::current().name().map(str::to_owned);
            let result = (|| {
                maybe_trigger_hybrid_leg_panic(panic_lexical, "injected lexical hybrid leg panic");
                let lease =
                    SnapshotLease::new_at(Arc::clone(&admitted.snapshot), admitted.generation);
                let cancellation = QueryCancellation::new(&control, &lease);
                cancellation
                    .check_graph()
                    .map_err(QueryError::Scan)
                    .map_err(crate::fusion::FusionError::from)?;
                match lexical_query {
                    PinnedLexicalQuery::Term(_) => {
                        exact_lexical_leg(lexical_inputs, bound, &cancellation, preparation)
                    }
                    PinnedLexicalQuery::Structured(query) => exact_structured_lexical_leg(
                        lexical_inputs,
                        &self.tokenizer,
                        query,
                        bound,
                        &cancellation,
                        preparation,
                    ),
                }
            })();
            (
                name,
                result,
                queue_time,
                timing_elapsed(self.clock.as_ref(), lexical_started),
            )
        };
        let caller_chose_tier = requested_tier.is_some();
        let run_vector_leg = |vector_query: crate::ingest::SearchRequest<'_>,
                              vector_preparation: &mut prepared::PreparedVectorQuery<'_>,
                              k: usize| {
            let graph_available = snapshot_has_graph(&admitted.snapshot);
            let scan_reason_override = if caller_chose_tier || !graph_available {
                None
            } else if k > corpus_rows {
                Some(crate::planner::ScanReason::FullMaterialization)
            } else if graph_round_is_worth_it(k, corpus_rows) {
                None
            } else {
                Some(crate::planner::ScanReason::WideningCap {
                    ef: graph_round_ef(k),
                    rows: corpus_rows,
                })
            };
            search_pinned(
                admitted.pool.as_deref(),
                &admitted.snapshot,
                &admitted.active_segment,
                &self.accounting,
                admitted.generation,
                pinned_epoch.unwrap_or_else(|| self.epoch_identity()),
                vector_query,
                k,
                // A graph traversal only pays for itself while its
                // candidate pool is a small fraction of the segment. Once
                // a widening round would visit more than
                // `GRAPH_WIDENING_CAP_NUMERATOR/DENOMINATOR` of the rows,
                // the traversal is a scan wearing a graph costume: it
                // pays the graph's indirection and then reads most of the
                // segment anyway. Such a round runs as Exact instead,
                // which is both faster and exactly rescored.
                // Only when the caller stated no preference. An explicit
                // tier is a contract: honour it, and let fusion reject it
                // loudly if it cannot supply exact scores.
                if caller_chose_tier || graph_round_is_worth_it(k, corpus_rows) {
                    options
                } else {
                    options.with_tier(SearchTier::Exact)
                },
                control.clone(),
                GraphBoundMode::Shared,
                requested_tier,
                scan_reason_override,
                started,
                self.clock.as_ref(),
                vector_preparation,
                #[cfg(any(test, feature = "test-seams"))]
                self.vector_fault_controller.as_ref(),
            )
            .map_err(crate::fusion::FusionError::from)
        };
        let lexical_queued = timing_start(self.clock.as_ref());
        let (prepared_vector_result, lexical_result) = lexical_worker.run_scoped(
            || {
                run_lexical_leg(
                    lexical_inputs,
                    &mut lexical_preparation,
                    width.saturating_add(1),
                    lexical_queued,
                )
            },
            || {
                std::panic::catch_unwind(std::panic::AssertUnwindSafe(|| {
                    let vector_query =
                        prepare_vector().map_err(HybridPreparationError::Preparation)?;
                    if vector_query.filter.is_some() && filter.is_none() {
                        return Err(HybridPreparationError::Search(
                            QueryError::Store(StoreError::InvalidScan {
                                detail:
                                    "deferred vector preparation cannot introduce a query filter"
                                        .to_owned(),
                            })
                            .into(),
                        ));
                    }
                    let deferred_filter = if known_eligible.is_none() {
                        vector_query.eligible.map(|ids| {
                            filter.cloned().map_or_else(
                                || QueryFilter::eligible(&self.schema, ids),
                                |filter| filter.with_eligible(ids),
                            )
                        })
                    } else {
                        None
                    };
                    let effective_vector_query = vector_query
                        .with_filter(deferred_filter.as_ref().or(filter).or(vector_query.filter));
                    let mut vector_preparation = prepared::PreparedVectorQuery::new(
                        vector_query.vector(),
                        pinned_epoch.unwrap_or_else(|| self.epoch_identity()),
                        &self.accounting,
                    );
                    maybe_trigger_hybrid_leg_panic(
                        panic_vector,
                        "injected vector hybrid leg panic",
                    );
                    let result = run_vector_leg(
                        effective_vector_query,
                        &mut vector_preparation,
                        width.saturating_add(1),
                    );
                    Ok((vector_query, deferred_filter, vector_preparation, result))
                }))
                .unwrap_or(Err(HybridPreparationError::Search(
                    crate::fusion::FusionError::LegPanic {
                        leg: crate::fusion::FusionLeg::Vector,
                        detail: "vector hybrid leg panicked",
                    },
                )))
            },
        )?;
        let (lexical_thread_name, lexical_result, queue_time, lexical_time) = lexical_result
            .unwrap_or_else(|error| {
                (
                    Some("zeppelin-fts".to_owned()),
                    Err(error),
                    std::time::Duration::ZERO,
                    std::time::Duration::ZERO,
                )
            });
        #[cfg(any(test, feature = "test-seams"))]
        if let Ok(mut receipt) = self.hybrid_execution_receipt.lock() {
            *receipt = Some(HybridExecutionReceipt {
                vector_thread: std::thread::current().id(),
                lexical_thread_name,
                vector_completed: true,
                lexical_completed: true,
            });
        }
        #[cfg(not(any(test, feature = "test-seams")))]
        let _ = lexical_thread_name;
        // run_scoped has joined lexical work even when vector preparation
        // failed or panicked. A stopped admission takes precedence over a late
        // producer error; otherwise retain the caller's original error type.
        let (vector_query, deferred_filter, mut vector_preparation, vector_result) =
            match prepared_vector_result {
                Ok(prepared) => prepared,
                Err(error) => {
                    let lease =
                        SnapshotLease::new_at(Arc::clone(&admitted.snapshot), admitted.generation);
                    QueryCancellation::new(&control, &lease)
                        .check_graph()
                        .map_err(map_scan_error)?;
                    if let Err(lexical_error) = lexical_result
                        && matches!(
                            lexical_error,
                            crate::fusion::FusionError::ReadCancelled { .. }
                                | crate::fusion::FusionError::Timeout { .. }
                                | crate::fusion::FusionError::Cancelled { .. }
                        )
                    {
                        return Err(HybridPreparationError::Search(lexical_error));
                    }
                    return Err(error);
                }
            };
        let vector_query =
            vector_query.with_filter(deferred_filter.as_ref().or(filter).or(vector_query.filter));
        let (
            mut vector_outcome,
            (mut lexical_hits, mut lexical_assembly, mut lexical_counters, mut lexical_cache_hit),
        ) = resolve_hybrid_leg_results(vector_result, lexical_result)?;

        timings.lexical_queue = queue_time;
        timings.lexical = lexical_time;
        let mut deferred_initial_cache_hit = None;
        if let Some(filter) = deferred_filter.as_ref() {
            // Preserve lexical/vector overlap for deferred embedding. Eligibility
            // was unavailable to the first lexical pass: discard that ranking,
            // reset its prepared assembly, and rank the eligible rows before fusion.
            lexical_inputs.filter = Some(filter);
            lexical_preparation = prepared_lexical::PreparedLexicalQuery::new(lexical_query);
            let queued = timing_start(self.clock.as_ref());
            let (_, reranked) = lexical_worker.run_scoped(
                || {
                    run_lexical_leg(
                        lexical_inputs,
                        &mut lexical_preparation,
                        width.saturating_add(1),
                        queued,
                    )
                },
                || (),
            )?;
            let (_, reranked, queue_time, lexical_time) = reranked?;
            let (hits, assembly, mut counters, cache_hit) = reranked?;
            accumulate_search_counters(&mut counters, &lexical_counters);
            deferred_initial_cache_hit = Some(lexical_cache_hit);
            lexical_hits = hits;
            lexical_assembly = assembly;
            lexical_counters = counters;
            lexical_cache_hit = cache_hit;
            timings.lexical_queue += queue_time;
            timings.lexical += lexical_time;
        }
        let anchors = hybrid::HybridAnchors {
            vector_ceiling: vector_outcome.vector_ceiling,
            lexical_maximum: lexical_hits.first().map_or(0.0, |hit| hit.bm25),
        };
        let fusion_lease =
            SnapshotLease::new_at(Arc::clone(&admitted.snapshot), admitted.generation);
        let fusion_cancellation = QueryCancellation::new(&control, &fusion_lease);
        let mut work = hybrid::HybridWork::new();
        if let Some(cache_hit) = deferred_initial_cache_hit {
            work.lexical_cache_hits += usize::from(cache_hit);
            work.lexical_cache_builds += usize::from(!cache_hit);
        }
        let mut cross_scores = hybrid::CrossScoreCache::new(vector_outcome.epoch);
        let mut round_buffers = hybrid::HybridRoundBuffers::new();
        let mut fusion_scratch =
            crate::fusion::StoreFusionScratch::new(Arc::clone(&self.accounting));
        let mut rescore_counters = crate::diag::ScanRescoreCounters::default();
        let mut rounds = 1_usize;
        let mut budget_exhausted = false;
        let (fused, report_window, addresses) = loop {
            if let Some(vector_timings) = vector_outcome.diagnostics.timings {
                timings.vector += vector_timings.vector;
            }
            let fusion_started = timing_start(self.clock.as_ref());
            let round = hybrid::build_round(
                &admitted.snapshot,
                &admitted.active_segment,
                &lexical_assembly,
                &lexical_preparation,
                &self.tokenizer,
                lexical_query,
                vector_query.vector(),
                &vector_outcome.candidates,
                anchors,
                &lexical_hits,
                width,
                &fusion_cancellation,
                capture_rows,
                &self.accounting,
                &mut cross_scores,
                &mut round_buffers,
            )?;
            // Cross-fill reads full-precision rows on the caller thread,
            // outside the vector producer. Include those actual reads in
            // the same scan work receipt as the producer's scored rows.
            let cross_fill_dims = u64::try_from(round.vector_scores_computed)
                .ok()
                .and_then(|rows| rows.checked_mul(vector_query.vector().len() as u64))
                .ok_or(QueryError::Scan(crate::scan::ScanError::ArithmeticOverflow))?;
            let cross_fill_bytes = cross_fill_dims
                .checked_mul(4)
                .ok_or(QueryError::Scan(crate::scan::ScanError::ArithmeticOverflow))?;
            vector_outcome.stats.dims_touched = vector_outcome
                .stats
                .dims_touched
                .checked_add(cross_fill_dims)
                .ok_or(QueryError::Scan(crate::scan::ScanError::ArithmeticOverflow))?;
            vector_outcome.stats.bytes_read = vector_outcome
                .stats
                .bytes_read
                .checked_add(cross_fill_bytes)
                .ok_or(QueryError::Scan(crate::scan::ScanError::ArithmeticOverflow))?;
            if cross_fill_dims != 0 {
                let caller = std::thread::current().id();
                if !vector_outcome.stats.worker_thread_ids.contains(&caller) {
                    vector_outcome.stats.worker_thread_ids.push(caller);
                }
                vector_outcome.stats.threads_used = vector_outcome.stats.worker_thread_ids.len();
            }
            accumulate_search_counters(&mut lexical_counters, &round.lexical_counters);
            if let Some(counters) = vector_outcome.diagnostics.scan_rescore {
                rescored_scan::accumulate(&mut rescore_counters, counters)?;
            }
            work.complete_round(
                &vector_outcome.stats,
                vector_outcome.graph_stats,
                lexical_counters,
            )?;
            work.lexical_cache_hits += usize::from(lexical_cache_hit);
            work.lexical_cache_builds += usize::from(!lexical_cache_hit);
            work.plans.append(&mut vector_outcome.diagnostics.plan);
            work.cross_filled_vector = work
                .cross_filled_vector
                .checked_add(round.vector_scores_computed)
                .ok_or(QueryError::Scan(crate::scan::ScanError::ArithmeticOverflow))?;
            work.cross_filled_lexical = work
                .cross_filled_lexical
                .checked_add(round.lexical_scores_computed)
                .ok_or(QueryError::Scan(crate::scan::ScanError::ArithmeticOverflow))?;
            work.vector_candidates_produced = work
                .vector_candidates_produced
                .checked_add(vector_outcome.candidates.len())
                .ok_or(QueryError::Scan(crate::scan::ScanError::ArithmeticOverflow))?;
            work.lexical_candidates_produced = work
                .lexical_candidates_produced
                .checked_add(lexical_hits.len())
                .ok_or(QueryError::Scan(crate::scan::ScanError::ArithmeticOverflow))?;
            let provenance = hybrid::round_provenance(&vector_outcome, lexical_hits.len(), width);
            let fused = hybrid::fuse_round(
                hybrid_query,
                &round,
                provenance,
                requested_tier == Some(SearchTier::Exact),
                &mut fusion_scratch,
            )?;
            timings.fusion_cross_fill += timing_elapsed(self.clock.as_ref(), fusion_started);
            if fused.report.termination != crate::fusion::FusionTermination::WindowUnproven
                || width >= corpus_rows
            {
                break (
                    fused,
                    crate::diag::HybridReport {
                        normalization_policy_version:
                            crate::fusion::HYBRID_NORMALIZATION_POLICY_VERSION,
                        lexical_full_materializations: 0,
                        provenance,
                        window: width,
                        vector_returned: round.vector.len(),
                        lexical_returned: round.lexical.len(),
                        cross_filled_vector: round.cross_filled_vector,
                        cross_filled_lexical: round.cross_filled_lexical,
                        total_cross_filled_vector: work.cross_filled_vector,
                        total_cross_filled_lexical: work.cross_filled_lexical,
                        vector_candidates_produced: work.vector_candidates_produced,
                        lexical_candidates_produced: work.lexical_candidates_produced,
                    },
                    round.addresses,
                );
            }
            width = if rounds >= hybrid_query.max_rounds {
                budget_exhausted = true;
                corpus_rows
            } else {
                width.saturating_mul(2).min(corpus_rows)
            };
            rounds = rounds.saturating_add(1);
            let lexical_queued = timing_start(self.clock.as_ref());
            let (vector_result, lexical_result) = lexical_worker.run_scoped(
                || {
                    run_lexical_leg(
                        lexical_inputs,
                        &mut lexical_preparation,
                        width.saturating_add(1),
                        lexical_queued,
                    )
                },
                || {
                    run_vector_leg(
                        vector_query,
                        &mut vector_preparation,
                        width.saturating_add(1),
                    )
                },
            )?;
            let (_, lexical_result, queue_time, lexical_time) =
                lexical_result.unwrap_or_else(|error| {
                    (
                        Some("zeppelin-fts".to_owned()),
                        Err(error),
                        std::time::Duration::ZERO,
                        std::time::Duration::ZERO,
                    )
                });
            let (next_vector, (next_hits, next_sources, next_counters, next_cache_hit)) =
                resolve_hybrid_leg_results(vector_result, lexical_result)?;
            timings.lexical_queue += queue_time;
            timings.lexical += lexical_time;
            vector_outcome = next_vector;
            lexical_hits = next_hits;
            lexical_assembly = next_sources;
            lexical_counters = next_counters;
            lexical_cache_hit = next_cache_hit;
        };

        let crate::ingest::SearchOutcome {
            candidates: _,
            vector_ceiling: _,
            stats: _,
            graph_stats: _,
            generation,
            epoch,
            diagnostics: vector_diagnostics,
        } = vector_outcome;
        let mut report = fused.report;
        report.rounds = rounds;
        if budget_exhausted {
            report.budget_exhausted = true;
            report.termination = crate::fusion::FusionTermination::BudgetFullMaterialization;
        }
        let mut diagnostics =
            crate::diag::QueryDiagnostics::hybrid(crate::diag::HybridDiagnostics {
                snapshot_generation: vector_diagnostics.snapshot_generation,
                indexed_through_seq: vector_diagnostics.indexed_through_seq,
                plan: work.plans,
                approximate: vector_diagnostics.approximate
                    || report.termination
                        == crate::fusion::FusionTermination::ApproximateCandidates,
                exact_rescore: vector_diagnostics.exact_rescore,
                requested_k: hybrid_query.k,
                returned: fused.hits.len(),
                scan: work.scan,
                graph: work.graph,
                lexical: work.lexical,
                report,
                hybrid: report_window,
                tier_resolution,
                epoch,
                elapsed: self.clock.now().saturating_duration_since(started),
            });
        diagnostics.counters.lexical_cache_hits = work.lexical_cache_hits;
        diagnostics.scan_rescore = options.scan_rescore().map(|_| rescore_counters);
        diagnostics.counters.lexical_cache_builds = work.lexical_cache_builds;
        diagnostics.timings = crate::diag::QUERY_TIMING_ENABLED.then_some(timings);
        let outcome = crate::ingest::StoreHybridSearchOutcome {
            hits: fused.hits,
            lexical_expansions: lexical_preparation.take_expansions()?,
            generation,
            diagnostics,
        };
        finish(
            outcome,
            &admitted.snapshot,
            &admitted.active_segment,
            &fusion_cancellation,
            addresses.as_ref(),
        )
        .map_err(HybridPreparationError::Search)
    }

    /// Returns `(hits, builds)` for the assembled-lexical-index cache.
    ///
    /// Exists so a test can prove the cache is actually reused, and actually
    /// rebuilt when its inputs move, rather than inferring either from a
    /// timing.
    #[cfg(any(test, feature = "test-seams"))]
    #[must_use]
    pub fn lexical_index_cache_counters(&self) -> (u64, u64) {
        (
            self.lexical_index_cache.hits.load(Ordering::Relaxed),
            self.lexical_index_cache.builds.load(Ordering::Relaxed),
        )
    }

    /// `(hits, builds)` of immutable lexical statistics contributions.
    #[cfg(any(test, feature = "test-seams"))]
    #[must_use]
    pub fn lexical_contribution_cache_counters(&self) -> (u64, u64) {
        (
            self.lexical_index_cache
                .contribution_hits
                .load(Ordering::Relaxed),
            self.lexical_index_cache
                .contribution_builds
                .load(Ordering::Relaxed),
        )
    }

    fn consume_hybrid_test_fault(&self, leg: crate::fusion::FusionLeg) -> bool {
        #[cfg(any(test, feature = "test-seams"))]
        {
            let Ok(mut fault) = self.hybrid_leg_fault.lock() else {
                return false;
            };
            if matches!(*fault, Some(HybridLegTestFault::Panic(candidate)) if candidate == leg) {
                *fault = None;
                return true;
            }
        }
        #[cfg(not(any(test, feature = "test-seams")))]
        let _ = leg;
        false
    }

    fn search_with_graph_bound_mode(
        &self,
        request: crate::ingest::SearchRequest<'_>,
        k: usize,
        options: SearchOptions,
        control: QueryControl,
        graph_bound_mode: GraphBoundMode,
    ) -> Result<crate::ingest::SearchOutcome, QueryError> {
        self.search_with_graph_bound_mode_then(
            request,
            k,
            options,
            control,
            graph_bound_mode,
            |outcome, _, _| Ok(outcome),
        )
    }

    fn search_with_graph_bound_mode_then<R>(
        &self,
        request: crate::ingest::SearchRequest<'_>,
        k: usize,
        options: SearchOptions,
        control: QueryControl,
        graph_bound_mode: GraphBoundMode,
        finish: impl FnOnce(
            crate::ingest::SearchOutcome,
            &AdmittedVectorSearch<'_>,
            &QueryControl,
        ) -> Result<R, QueryError>,
    ) -> Result<R, QueryError> {
        let eligible_filter = request.eligible.map(|ids| {
            request.filter.cloned().map_or_else(
                || QueryFilter::eligible(&self.schema, ids),
                |filter| filter.with_eligible(ids),
            )
        });
        let request = request.with_filter(eligible_filter.as_ref().or(request.filter));
        let control = control.with_clock(Arc::clone(&self.clock));
        let started = self.clock.now();
        let admission_started = timing_start(self.clock.as_ref());
        let admitted = self.admit_vector_search(options)?;
        let admission_time = timing_elapsed(self.clock.as_ref(), admission_started);
        let mut vector_preparation = prepared::PreparedVectorQuery::new(
            request.vector(),
            self.epoch_identity(),
            &self.accounting,
        );
        let score = || {
            let mut outcome = search_pinned(
                admitted.pool.as_deref(),
                &admitted.snapshot,
                &admitted.active_segment,
                &self.accounting,
                admitted.generation,
                self.epoch_identity(),
                request,
                k,
                options,
                control.clone(),
                graph_bound_mode,
                options.explicit_tier(),
                None,
                started,
                self.clock.as_ref(),
                &mut vector_preparation,
                #[cfg(any(test, feature = "test-seams"))]
                self.vector_fault_controller.as_ref(),
            )?;
            if let Some(timings) = &mut outcome.diagnostics.timings {
                timings.admission = admission_time;
            }
            finish(outcome, &admitted, &control)
        };
        #[cfg(any(test, feature = "test-seams"))]
        let result = crate::kernels::vector_fault::run_store_scoring(
            self.kernel_fault_controller.as_ref(),
            score,
        );
        #[cfg(not(any(test, feature = "test-seams")))]
        let result = score();
        #[cfg(any(test, feature = "test-seams"))]
        {
            if let Some(controller) = self.vector_fault_controller.as_ref() {
                controller.finalize_search(result.is_ok());
            }
        }
        result
    }

    fn ensure_query_pool(&self) -> Result<Arc<pool::QueryPool>, QueryError> {
        let mut pool_slot = self.query_pool.lock().map_err(|_| {
            QueryError::Store(StoreError::Synchronization {
                component: "query pool",
            })
        })?;
        if pool_slot.is_none() {
            let capacity = crate::scan::physical_thread_capacity().map_err(|error| {
                QueryError::Store(StoreError::QueryPoolCapacity {
                    source: error.to_string(),
                })
            })?;
            let pool =
                pool::QueryPool::start(capacity, &self.accounting).map_err(QueryError::Store)?;
            *pool_slot = Some(Arc::new(pool));
        }
        let pool = pool_slot.as_ref().cloned().ok_or({
            QueryError::Store(StoreError::Synchronization {
                component: "query pool initialization",
            })
        })?;
        drop(pool_slot);
        Ok(pool)
    }

    fn ensure_lexical_worker(
        &self,
    ) -> Result<Arc<pool::LexicalWorker>, crate::fusion::FusionError> {
        let mut worker_slot =
            self.lexical_worker
                .lock()
                .map_err(|_| crate::fusion::FusionError::LegThreadStart {
                    leg: crate::fusion::FusionLeg::Lexical,
                    detail: "pooled lexical worker synchronization was poisoned".to_owned(),
                })?;
        if worker_slot.is_none() {
            *worker_slot = Some(Arc::new(pool::LexicalWorker::start()?));
        }
        let worker = worker_slot.as_ref().cloned().ok_or_else(|| {
            crate::fusion::FusionError::LegThreadStart {
                leg: crate::fusion::FusionLeg::Lexical,
                detail: "pooled lexical worker initialization failed".to_owned(),
            }
        })?;
        drop(worker_slot);
        Ok(worker)
    }

    fn admit_vector_search(
        &self,
        options: SearchOptions,
    ) -> Result<AdmittedVectorSearch<'_>, QueryError> {
        self.admit_vector_search_for(options, false)
    }

    fn admit_vector_search_for(
        &self,
        options: SearchOptions,
        hybrid_default: bool,
    ) -> Result<AdmittedVectorSearch<'_>, QueryError> {
        #[cfg(any(test, feature = "test-seams"))]
        self.text_materialization_work
            .admissions
            .fetch_add(1, Ordering::Relaxed);
        let state = self
            .state
            .lock()
            .map_err(|_| QueryError::Store(StoreError::Synchronization { component: "state" }))?;
        match *state {
            StoreState::Open => {}
            StoreState::Closing => return Err(QueryError::Store(StoreError::Closing)),
            StoreState::Closed => return Err(QueryError::Store(StoreError::Closed)),
        }
        let active_guard = self.active.lock().map_err(|_| {
            QueryError::Store(StoreError::Synchronization {
                component: "active segment",
            })
        })?;
        let active_state = active_guard
            .as_ref()
            .ok_or(QueryError::Store(StoreError::Closed))?;
        let generation = active_state.generation;
        let active_segment = Arc::clone(&active_state.segment);
        let snapshot = self
            .snapshot
            .read()
            .map_err(|_| {
                QueryError::Store(StoreError::Synchronization {
                    component: "published snapshot",
                })
            })?
            .as_ref()
            .cloned()
            .ok_or(QueryError::Store(StoreError::Closed))?;
        let execution_options = if hybrid_default && !snapshot_has_graph(&snapshot) {
            options.with_tier(SearchTier::Exact)
        } else {
            options
        };
        // Default hybrid may move from graph traversal to Exact while widening.
        // Admit that resource while the Store is still Open under this lock.
        let needs_query_pool = !active_segment.is_empty()
            || (hybrid_default && !snapshot.segments().is_empty())
            || match execution_options.tier() {
                SearchTier::Auto => snapshot.segments().iter().any(|segment| {
                    !segment.directory().iter().any(|entry| {
                        entry.kind == crate::segment::layout::RegionKind::GraphNodeBlocks.id()
                    })
                }),
                SearchTier::Exact => !snapshot.segments().is_empty(),
                SearchTier::Scan => true,
                SearchTier::Graph(_) => false,
            };
        let pool = if needs_query_pool {
            Some(self.ensure_query_pool()?)
        } else {
            None
        };
        self.active_queries.fetch_add(1, Ordering::Relaxed);
        let active_query = ActiveQuery {
            count: &self.active_queries,
        };
        drop(active_guard);
        drop(state);
        Ok(AdmittedVectorSearch {
            pool,
            execution_options,
            snapshot: snapshot::ReadSnapshot::new(snapshot),
            active_segment,
            generation,
            _active_query: Some(active_query),
        })
    }

    #[cfg(test)]
    fn search_independent_for_test(
        &self,
        request: crate::ingest::SearchRequest<'_>,
        k: usize,
        options: SearchOptions,
        control: QueryControl,
    ) -> Result<crate::ingest::SearchOutcome, QueryError> {
        self.search_with_graph_bound_mode(request, k, options, control, GraphBoundMode::Independent)
    }
}

#[derive(Clone, Copy)]
pub(crate) enum StructuredLexicalSource {
    Sealed(usize),
    Active,
}

#[derive(Clone, Debug)]
struct EligibleIds {
    numeric: Vec<crate::ingest::DocId>,
    // The sealed reader permutation orders little-endian identity bytes, not u128s.
    // Sort once per query, shared by every sealed source and both hybrid legs.
    byte_order: Vec<[u8; 16]>,
}

const ELIGIBLE_BYTE_ORDER_MIN_IDS: usize = 1_000;

/// Validated scan-compatible query constraints. Ranking statistics remain corpus-wide.
#[derive(Clone, Debug)]
pub struct QueryFilter {
    predicate: Option<crate::meta::Predicate>,
    schema: crate::meta::Schema,
    eligible_ids: Option<EligibleIds>,
}

impl QueryFilter {
    /// Upper bound for the owned backing of a fresh eligibility-only filter.
    /// Reserve before construction; the schema clone cannot exceed its source.
    #[cfg(feature = "graph-cypher")]
    pub(crate) fn eligible_capacity_bytes(
        schema: &crate::meta::Schema,
        count: usize,
    ) -> Option<usize> {
        let numeric = count.checked_mul(std::mem::size_of::<crate::ingest::DocId>())?;
        let byte_order = if count >= ELIGIBLE_BYTE_ORDER_MIN_IDS {
            count.checked_mul(std::mem::size_of::<[u8; 16]>())?
        } else {
            0
        };
        std::mem::size_of::<Self>()
            .checked_add(schema.resident_bytes()?)?
            .checked_add(numeric)?
            .checked_add(byte_order)
    }

    /// Validates attributes and a half-open timestamp range against this schema.
    pub fn new(
        schema: &crate::meta::Schema,
        predicate: Option<&crate::meta::Predicate>,
        timestamp_range: Option<(i64, i64)>,
    ) -> Result<Option<Self>, QueryError> {
        let Some(predicate) = document_scan_predicate(predicate, timestamp_range)? else {
            return Ok(None);
        };
        crate::planner::validate_predicate(&predicate, schema).map_err(|error| {
            QueryError::Store(StoreError::InvalidScan {
                detail: error.to_string(),
            })
        })?;
        Ok(Some(Self {
            predicate: Some(predicate),
            schema: schema.clone(),
            eligible_ids: None,
        }))
    }

    fn active_rows(
        &self,
        active: &crate::ingest::ActiveSegment,
    ) -> Result<crate::meta::DocBitmap, QueryError> {
        let mut rows = active_scan_rows(active, &self.schema, self.predicate.as_ref())?;
        if let Some(ids) = &self.eligible_ids {
            // Active rows have no identity index. Scan their bounded, u32-addressed
            // in-memory ids; no additional index or persisted state is needed.
            let mut eligible = crate::meta::DocBitmap::new();
            for (row, id) in active.doc_ids().iter().enumerate() {
                if ids.numeric.binary_search(id).is_ok() {
                    eligible.insert(
                        u32::try_from(row)
                            .map_err(|_| QueryError::Store(StoreError::ActiveRowOverflow))?,
                    );
                }
            }
            rows.intersect_with(&eligible);
        }
        Ok(rows)
    }

    /// Restricts all query legs to these document ids, before top-k and fusion.
    /// An empty set selects no documents. Input order and duplicates do not matter.
    #[must_use]
    pub fn eligible(schema: &crate::meta::Schema, ids: &[crate::ingest::DocId]) -> Self {
        Self {
            predicate: None,
            schema: schema.clone(),
            eligible_ids: None,
        }
        .with_eligible(ids)
    }

    /// Intersects these ids with existing eligibility and attribute/time constraints.
    #[must_use]
    pub fn with_eligible(mut self, ids: &[crate::ingest::DocId]) -> Self {
        let mut ids = ids.to_vec();
        ids.sort_unstable();
        ids.dedup();
        if let Some(existing) = &self.eligible_ids {
            ids.retain(|id| existing.numeric.binary_search(id).is_ok());
        }
        let mut byte_order = if ids.len() >= ELIGIBLE_BYTE_ORDER_MIN_IDS {
            ids.iter()
                .map(|id| id.get().to_le_bytes())
                .collect::<Vec<_>>()
        } else {
            Vec::new()
        };
        byte_order.sort_unstable();
        self.eligible_ids = Some(EligibleIds {
            numeric: ids,
            byte_order,
        });
        self
    }

    fn sealed_rows(
        &self,
        segment: &crate::segment::reader::SegmentReader,
    ) -> Result<crate::meta::DocBitmap, QueryError> {
        let mut rows = sealed_scan_rows(segment, self.predicate.as_ref())?;
        if let Some(ids) = &self.eligible_ids {
            rows.intersect_with(
                &segment
                    .query_rows_for_eligible_ids(&ids.numeric, &ids.byte_order)
                    .map_err(QueryError::Store)?,
            );
        }
        Ok(rows)
    }
}

struct LexicalAssembly {
    index: crate::fts::index::LexicalIndex,
    alive_sets: Vec<Arc<crate::meta::AliveSet>>,
    sources: Vec<StructuredLexicalSource>,
    contributions: Vec<Arc<CachedLexicalContribution>>,
    vocabulary: Arc<Mutex<Option<Arc<CachedVocabulary>>>>,
    // Filtered readers retain the owner of the shared vocabulary-slot charge.
    _base: Option<Arc<LexicalAssembly>>,
    // The last cache/query owner releases this charge, even after eviction.
    memory: stats::AccountedCounter,
}

struct CachedVocabulary {
    view: crate::fts::vocabulary::Vocabulary,
    phonetic: Mutex<Option<Arc<expansion::CachedPhonetic>>>,
    // Charged for as long as any query or assembly retains the dictionary.
    _memory: stats::AccountedCounter,
}

impl LexicalAssembly {
    fn vocabulary(
        &self,
        accounting: &Arc<stats::Accounting>,
        cancellation: &QueryCancellation<'_>,
    ) -> Result<Arc<CachedVocabulary>, QueryError> {
        let mut cached = cancellation.cache_lock(&self.vocabulary, "lexical vocabulary cache")?;
        if let Some(value) = cached.as_ref() {
            return Ok(Arc::clone(value));
        }
        let mut memory =
            stats::AccountedCounter::new(accounting, stats::AllocationComponent::Cache)
                .map_err(QueryError::Store)?;
        let mut temporary =
            stats::AccountedCounter::new(accounting, stats::AllocationComponent::Temporary)
                .map_err(QueryError::Store)?;
        let mut work = crate::fts::control::WorkCheck::new(|| {
            cancellation.check_graph().map_err(QueryError::Scan)
        });
        let view = crate::fts::vocabulary::Vocabulary::build_controlled(
            self.index.vocabulary_terms(),
            |scratch, owned| {
                cancellation.check_graph().map_err(QueryError::Scan)?;
                prepared_lexical::reserve_weighted_scratch(&mut temporary, scratch)?;
                let bytes = owned
                    .and_then(|bytes| bytes.checked_add(std::mem::size_of::<CachedVocabulary>()))
                    .ok_or(QueryError::Store(StoreError::BudgetExceeded {
                        needed: u64::MAX,
                        budget: u64::MAX,
                        component: "lexical vocabulary cache",
                    }))?;
                memory.set(bytes).map_err(QueryError::Store)
            },
            &mut work,
        )?;
        let value = Arc::new(CachedVocabulary {
            view,
            phonetic: Mutex::new(None),
            _memory: memory,
        });
        *cached = Some(Arc::clone(&value));
        Ok(value)
    }

    /// Assembly spines only: contribution reservations own their shared
    /// counters/bitmaps, and current readers own decoded postings/alive views.
    fn owned_bytes(&self) -> Option<usize> {
        std::mem::size_of::<Self>()
            .checked_add(self.index.resident_spine_bytes()?)?
            .checked_add(
                self.alive_sets
                    .capacity()
                    .checked_mul(std::mem::size_of::<Arc<crate::meta::AliveSet>>())?,
            )?
            .checked_add(
                self.sources
                    .capacity()
                    .checked_mul(std::mem::size_of::<StructuredLexicalSource>())?,
            )?
            .checked_add(
                self.contributions
                    .capacity()
                    .checked_mul(std::mem::size_of::<Arc<CachedLexicalContribution>>())?,
            )
    }
}

enum LexicalContributionKey {
    // Reader objects are recreated on publication. Same immutable file plus
    // freshly verified live membership identifies the same contribution.
    Sealed(crate::segment::SegmentMeta, (u64, u64)),
    Active(std::sync::Weak<crate::ingest::ActiveSegment>),
}

struct CachedLexicalContribution {
    key: LexicalContributionKey,
    statistics: Arc<crate::fts::index::LiveSegmentStatistics>,
    // Only active alive sets are newly allocated here. Sealed sets remain the
    // current reader's owned/counted query view, never a retired reader's view.
    active_alive: Option<Arc<crate::meta::AliveSet>>,
    _memory: stats::AccountedCounter,
}

impl CachedLexicalContribution {
    fn build(
        key: LexicalContributionKey,
        postings: &crate::fts::sealed::SealedSegment,
        alive: &Arc<crate::meta::AliveSet>,
        ordinal: usize,
        accounting: &Arc<stats::Accounting>,
        work: &mut crate::fts::control::WorkCheck<impl FnMut() -> Result<(), LexicalAssemblyError>>,
    ) -> Result<Arc<Self>, LexicalAssemblyError> {
        work.check_now()?;
        let active = matches!(key, LexicalContributionKey::Active(_));
        let alive_bytes = if active {
            alive
                .resident_bytes()
                .and_then(|n| n.checked_add(std::mem::size_of::<crate::meta::AliveSet>()))
                .and_then(|n| n.checked_add(2 * std::mem::size_of::<usize>()))
        } else {
            Some(0)
        };
        let bytes =
            crate::fts::index::LiveSegmentStatistics::allocation_bytes(alive.alive_bitmap())
                .and_then(|n| n.checked_add(std::mem::size_of::<Self>()))
                .and_then(|n| n.checked_add(2 * std::mem::size_of::<usize>()))
                .and_then(|n| n.checked_add(alive_bytes?))
                .ok_or(LexicalAssemblyError::Store(StoreError::BudgetExceeded {
                    needed: u64::MAX,
                    budget: u64::MAX,
                    component: "cache",
                }))?;
        let mut memory =
            stats::AccountedCounter::new(accounting, stats::AllocationComponent::Cache)
                .map_err(LexicalAssemblyError::Store)?;
        memory.set(bytes).map_err(LexicalAssemblyError::Store)?;
        let mut statistics = crate::fts::index::LiveSegmentStatistics::build_controlled(
            ordinal,
            postings,
            alive.alive_bitmap(),
            work,
        )?;
        if statistics.rows.cardinality() != 0
            && statistics.rows.cardinality() != u64::from(postings.row_count())
        {
            statistics.frequency_cache = Some(
                crate::fts::live_df::LiveFrequencyCache::new(accounting)
                    .map_err(LexicalAssemblyError::Store)?,
            );
        }
        let statistics = Arc::new(statistics);
        work.check_now()?;
        Ok(Arc::new(Self {
            key,
            statistics,
            active_alive: active.then(|| Arc::clone(alive)),
            _memory: memory,
        }))
    }
}

enum LexicalAssemblyError {
    Store(StoreError),
    Lexical(crate::fts::index::IndexError),
    MissingDocumentIdentity(crate::segment::SegmentId),
    Cancelled(crate::scan::ScanError),
}

impl From<crate::fts::index::IndexError> for LexicalAssemblyError {
    fn from(error: crate::fts::index::IndexError) -> Self {
        Self::Lexical(error)
    }
}

impl From<StoreError> for LexicalAssemblyError {
    fn from(error: StoreError) -> Self {
        Self::Store(error)
    }
}

/// One assembled lexical index, held for as long as the inputs it was built
/// from are still the store's inputs.
///
/// A complete assembly is reused while its snapshot/active identities match.
/// After mutation, independently reserved contributions reuse unchanged sealed
/// counters and membership while rebuilding current source ordinals and totals.
///
/// Reusing it is safe because both inputs are copy-on-write behind an `Arc`
/// and the borrow checker, not a convention, enforces that. A
/// `PublishedSnapshot` is replaced wholesale in `Store::snapshot` when a
/// generation is published, and `ActiveState::segment` is replaced with a
/// freshly built `Arc<ActiveSegment>` on every mutation; neither is ever
/// reached through `Arc::get_mut` or `Arc::make_mut`, so no shared value can
/// be edited in place. Pointer identity is therefore exactly "the same
/// inputs", and it is the same immutability that already lets
/// `SegmentReader::query_postings`, `SegmentReader::query_alive` and
/// `ActiveSegment::sealed_lexical` cache with no invalidation hook at all.
/// A key derived from observable state instead -- the generation, a row
/// count, a log sequence -- would have to enumerate every way the inputs can
/// move; pointer identity cannot miss one.
///
/// The handles are `Weak`, not `Arc`, for two reasons. A strong reference
/// would pin a retired snapshot's mappings and a retired active segment's
/// buffers past the point the store released them. And a `Weak` still keeps
/// the allocation reserved, so a dropped input's address can never be reused
/// by a different value and compare equal here.
struct CachedLexicalAssembly {
    generation: u64,
    snapshot: std::sync::Weak<PublishedSnapshot>,
    active: std::sync::Weak<crate::ingest::ActiveSegment>,
    /// Set when this assembly was built, or has since been proven, under
    /// `require_document_identity`. The hybrid legs do not require it and the
    /// standalone paths do, so an entry one path built is never handed to the
    /// other until the proof has actually run.
    document_identity_verified: bool,
    assembly: Arc<LexicalAssembly>,
}

impl CachedLexicalAssembly {
    fn matches(
        &self,
        snapshot: &Arc<PublishedSnapshot>,
        active: &Arc<crate::ingest::ActiveSegment>,
    ) -> bool {
        self.snapshot
            .upgrade()
            .is_some_and(|cached| Arc::ptr_eq(&cached, snapshot))
            && self
                .active
                .upgrade()
                .is_some_and(|cached| Arc::ptr_eq(&cached, active))
    }
}

/// The store's single-entry cache of assembled lexical indexes.
///
/// One entry bounds retained history. A builder holds the entry lock through
/// construction; old admissions may build their own answer, but cannot replace
/// a cached newer generation. Only immutable contribution Arcs cross rebuilds.
pub(crate) struct LexicalIndexCache {
    entry: Mutex<Option<CachedLexicalAssembly>>,
    #[cfg(any(test, feature = "test-seams"))]
    hits: AtomicU64,
    #[cfg(any(test, feature = "test-seams"))]
    builds: AtomicU64,
    #[cfg(any(test, feature = "test-seams"))]
    contribution_hits: AtomicU64,
    #[cfg(any(test, feature = "test-seams"))]
    contribution_builds: AtomicU64,
}

impl LexicalIndexCache {
    pub(crate) const fn new() -> Self {
        Self {
            entry: Mutex::new(None),
            #[cfg(any(test, feature = "test-seams"))]
            hits: AtomicU64::new(0),
            #[cfg(any(test, feature = "test-seams"))]
            builds: AtomicU64::new(0),
            #[cfg(any(test, feature = "test-seams"))]
            contribution_hits: AtomicU64::new(0),
            #[cfg(any(test, feature = "test-seams"))]
            contribution_builds: AtomicU64::new(0),
        }
    }

    #[cfg(any(test, feature = "test-seams"))]
    fn record_hit(&self) {
        self.hits.fetch_add(1, Ordering::Relaxed);
    }

    #[cfg(not(any(test, feature = "test-seams")))]
    const fn record_hit(&self) {}

    #[cfg(any(test, feature = "test-seams"))]
    fn record_build(&self) {
        self.builds.fetch_add(1, Ordering::Relaxed);
    }

    #[cfg(not(any(test, feature = "test-seams")))]
    const fn record_build(&self) {}

    fn record_contribution(&self, hit: bool) {
        #[cfg(any(test, feature = "test-seams"))]
        if hit {
            self.contribution_hits.fetch_add(1, Ordering::Relaxed);
        } else {
            self.contribution_builds.fetch_add(1, Ordering::Relaxed);
        }
        #[cfg(not(any(test, feature = "test-seams")))]
        let _ = hit;
    }

    fn lock(
        &self,
    ) -> Result<std::sync::MutexGuard<'_, Option<CachedLexicalAssembly>>, LexicalAssemblyError>
    {
        self.entry.lock().map_err(|_| {
            LexicalAssemblyError::Store(StoreError::Synchronization {
                component: "lexical index cache",
            })
        })
    }

    fn lock_controlled(
        &self,
        cancellation: Option<&QueryCancellation<'_>>,
    ) -> Result<std::sync::MutexGuard<'_, Option<CachedLexicalAssembly>>, LexicalAssemblyError>
    {
        let Some(cancellation) = cancellation else {
            return self.lock();
        };
        cancellation
            .check_graph()
            .map_err(LexicalAssemblyError::Cancelled)?;
        let guard = loop {
            match self.entry.try_lock() {
                Ok(guard) => break guard,
                Err(std::sync::TryLockError::Poisoned(_)) => {
                    return Err(LexicalAssemblyError::Store(StoreError::Synchronization {
                        component: "lexical index cache",
                    }));
                }
                Err(std::sync::TryLockError::WouldBlock) => {}
            }
            cancellation
                .check_graph()
                .map_err(LexicalAssemblyError::Cancelled)?;
            std::thread::park_timeout(Duration::from_millis(1));
        };
        cancellation
            .check_graph()
            .map_err(LexicalAssemblyError::Cancelled)?;
        Ok(guard)
    }
}

/// Everything an assembled lexical index is built from, and the cache that
/// holds the one built last. Grouped because the two hybrid legs and the two
/// standalone paths all pass exactly this set through unchanged.
#[derive(Clone, Copy)]
struct LexicalInputs<'a> {
    filter: Option<&'a QueryFilter>,
    generation: u64,
    cache: &'a LexicalIndexCache,
    snapshot: &'a Arc<PublishedSnapshot>,
    active: &'a Arc<crate::ingest::ActiveSegment>,
    accounting: &'a Arc<stats::Accounting>,
}

struct LexicalAssemblyReceipt {
    assembly: Arc<LexicalAssembly>,
    cache_hit: bool,
}

impl std::ops::Deref for LexicalAssemblyReceipt {
    type Target = LexicalAssembly;
    fn deref(&self) -> &Self::Target {
        &self.assembly
    }
}

fn map_eligibility_error(error: QueryError) -> LexicalAssemblyError {
    match error {
        QueryError::Store(error) => LexicalAssemblyError::Store(error),
        QueryError::Scan(error) => LexicalAssemblyError::Cancelled(error),
        error => LexicalAssemblyError::Store(StoreError::InvalidScan {
            detail: error.to_string(),
        }),
    }
}

fn assemble_lexical_index(
    inputs: LexicalInputs<'_>,
    require_document_identity: bool,
    cancellation: Option<&QueryCancellation<'_>>,
) -> Result<LexicalAssemblyReceipt, LexicalAssemblyError> {
    let receipt =
        assemble_unfiltered_lexical_index(inputs, require_document_identity, cancellation)?;
    let Some(filter) = inputs.filter else {
        return Ok(receipt);
    };
    let mut alive_sets = Vec::with_capacity(receipt.alive_sets.len());
    for (source, alive) in receipt.sources.iter().zip(&receipt.alive_sets) {
        if let Some(control) = cancellation {
            control
                .check_graph()
                .map_err(LexicalAssemblyError::Cancelled)?;
        }
        let rows = match source {
            StructuredLexicalSource::Active => filter.active_rows(inputs.active),
            StructuredLexicalSource::Sealed(ordinal) => inputs
                .snapshot
                .segments()
                .get(*ordinal)
                .ok_or(QueryError::Store(StoreError::ActiveRowOverflow))
                .and_then(|segment| filter.sealed_rows(segment)),
        }
        .map_err(map_eligibility_error)?;
        let mut eligible = (**alive).clone();
        #[cfg(any(test, feature = "test-seams"))]
        crate::fts::preparation_observer::record_filter_work(0, 0, 1);
        eligible.retain(&rows);
        alive_sets.push(Arc::new(eligible));
    }
    if let Some(control) = cancellation {
        control
            .check_graph()
            .map_err(LexicalAssemblyError::Cancelled)?;
    }
    let memory =
        stats::AccountedCounter::new(inputs.accounting, stats::AllocationComponent::Temporary)
            .map_err(LexicalAssemblyError::Store)?;
    let mut assembly = LexicalAssembly {
        index: receipt.index.clone(),
        alive_sets,
        sources: receipt.sources.clone(),
        contributions: receipt.contributions.clone(),
        vocabulary: Arc::clone(&receipt.vocabulary),
        _base: Some(Arc::clone(&receipt.assembly)),
        memory,
    };
    let bytes = assembly
        .alive_sets
        .iter()
        .fold(assembly.owned_bytes(), |bytes, alive| {
            bytes?
                .checked_add(alive.resident_bytes()?)?
                .checked_add(std::mem::size_of::<crate::meta::AliveSet>())?
                .checked_add(2 * std::mem::size_of::<usize>())
        })
        .ok_or(LexicalAssemblyError::Store(StoreError::ActiveRowOverflow))?;
    assembly
        .memory
        .set(bytes)
        .map_err(LexicalAssemblyError::Store)?;
    Ok(LexicalAssemblyReceipt {
        cache_hit: receipt.cache_hit,
        assembly: Arc::new(assembly),
    })
}

fn assemble_unfiltered_lexical_index(
    inputs: LexicalInputs<'_>,
    require_document_identity: bool,
    cancellation: Option<&QueryCancellation<'_>>,
) -> Result<LexicalAssemblyReceipt, LexicalAssemblyError> {
    let LexicalInputs {
        filter: _,
        generation,
        cache,
        snapshot,
        active,
        accounting,
    } = inputs;
    let mut entry = cache.lock_controlled(cancellation)?;
    if let Some(cached) = entry.as_mut()
        && cached.matches(snapshot, active)
    {
        if require_document_identity && !cached.document_identity_verified {
            verify_document_identity(snapshot, cancellation)?;
            cached.document_identity_verified = true;
        }
        let assembly = Arc::clone(&cached.assembly);
        cache.record_hit();
        return Ok(LexicalAssemblyReceipt {
            assembly,
            cache_hit: true,
        });
    }
    // One builder at a time bounds concurrent allocation and prevents a late
    // old admission from repeatedly replacing a newer assembly. The old entry
    // remains accounted during construction and is replaced only on success.
    let previous = entry.as_ref().map(|cached| Arc::clone(&cached.assembly));
    let contributions = previous
        .as_ref()
        .map_or(&[][..], |a| a.contributions.as_slice());
    let (mut assembly, owned_bytes) = build_lexical_assembly(
        snapshot,
        active,
        accounting,
        require_document_identity,
        cancellation,
        cache,
        contributions,
    )?;
    cache.record_build();
    assembly
        .memory
        .set(owned_bytes)
        .map_err(LexicalAssemblyError::Store)?;
    if let Some(cancellation) = cancellation {
        cancellation
            .check_graph()
            .map_err(LexicalAssemblyError::Cancelled)?;
    }
    let assembly = Arc::new(assembly);
    if entry
        .as_ref()
        .is_none_or(|cached| generation >= cached.generation)
    {
        *entry = Some(CachedLexicalAssembly {
            generation,
            snapshot: Arc::downgrade(snapshot),
            active: Arc::downgrade(active),
            document_identity_verified: require_document_identity,
            assembly: Arc::clone(&assembly),
        });
    }
    drop(entry);
    Ok(LexicalAssemblyReceipt {
        assembly,
        cache_hit: false,
    })
}

fn check_segment_document_identity(
    segment: &crate::segment::reader::SegmentReader,
    postings: &crate::fts::sealed::SealedSegment,
) -> Result<(), LexicalAssemblyError> {
    if postings.row_count() != 0
        && segment
            .document_version(0)
            .map_err(StoreError::Segment)
            .map_err(LexicalAssemblyError::Store)?
            .is_none()
    {
        return Err(LexicalAssemblyError::MissingDocumentIdentity(
            segment.meta().id,
        ));
    }
    Ok(())
}

/// Proves document identity over a snapshot whose assembly was built without
/// that requirement.
///
/// Every segment's postings were already decoded to build that assembly, so
/// this reads the same cached values and adds no region decode.
fn verify_document_identity(
    snapshot: &PublishedSnapshot,
    cancellation: Option<&QueryCancellation<'_>>,
) -> Result<(), LexicalAssemblyError> {
    for segment in snapshot.segments() {
        if let Some(cancellation) = cancellation {
            cancellation
                .check_graph()
                .map_err(LexicalAssemblyError::Cancelled)?;
        }
        if let Some(postings) = segment
            .query_postings()
            .map_err(LexicalAssemblyError::Store)?
        {
            check_segment_document_identity(segment, &postings)?;
        }
    }
    Ok(())
}

fn build_lexical_assembly(
    snapshot: &PublishedSnapshot,
    active: &Arc<crate::ingest::ActiveSegment>,
    accounting: &Arc<stats::Accounting>,
    require_document_identity: bool,
    cancellation: Option<&QueryCancellation<'_>>,
    cache: &LexicalIndexCache,
    previous: &[Arc<CachedLexicalContribution>],
) -> Result<(LexicalAssembly, usize), LexicalAssemblyError> {
    let mut work = crate::fts::control::WorkCheck::new(|| {
        cancellation
            .map_or(Ok(()), |control| control.check_graph())
            .map_err(LexicalAssemblyError::Cancelled)
    });
    work.check_now()?;
    let mut index = crate::fts::index::LexicalIndex::new();
    let mut alive_sets = Vec::new();
    let mut sources = Vec::new();
    let mut contributions = Vec::new();
    for (ordinal, segment) in snapshot.segments().iter().enumerate() {
        if let Some(cancellation) = cancellation {
            cancellation
                .check_graph()
                .map_err(LexicalAssemblyError::Cancelled)?;
        }
        if let Some(postings) = segment
            .query_postings()
            .map_err(LexicalAssemblyError::Store)?
        {
            if require_document_identity {
                check_segment_document_identity(segment, &postings)?;
            }
            let alive = segment.query_alive().map_err(LexicalAssemblyError::Store)?;
            let identity = segment
                .lexical_postings_identity()
                .map_err(StoreError::Segment)
                .map_err(LexicalAssemblyError::Store)?;
            let reused = previous
                .binary_search_by(|value| match &value.key {
                    LexicalContributionKey::Sealed(meta, _) => meta.id.cmp(&segment.meta().id),
                    LexicalContributionKey::Active(_) => std::cmp::Ordering::Greater,
                })
                .ok()
                .and_then(|slot| previous.get(slot))
                .filter(|value| {
                    matches!(&value.key, LexicalContributionKey::Sealed(meta, stored)
                    if meta.same_segment_file(segment.meta()) && *stored == identity)
                        && value.statistics.rows == *alive.alive_bitmap()
                });
            cache.record_contribution(reused.is_some());
            let contribution = match reused {
                Some(value) => Arc::clone(value),
                None => CachedLexicalContribution::build(
                    LexicalContributionKey::Sealed(segment.meta().clone(), identity),
                    &postings,
                    &alive,
                    index.segments().len(),
                    accounting,
                    &mut work,
                )?,
            };
            index.push_shared_with_statistics(postings, Arc::clone(&contribution.statistics));
            contributions.push(contribution);
            alive_sets.push(alive);
            sources.push(StructuredLexicalSource::Sealed(ordinal));
        }
    }
    if active.has_text_controlled(&mut work)? {
        let sealed = active.sealed_lexical_controlled(accounting, &mut work)?;
        let reused = previous.last().filter(|value| {
            matches!(&value.key, LexicalContributionKey::Active(cached)
                if cached.upgrade().is_some_and(|cached| Arc::ptr_eq(&cached, active)))
        });
        cache.record_contribution(reused.is_some());
        let contribution = match reused {
            Some(value) => Arc::clone(value),
            None => {
                let alive = Arc::new(active.alive_controlled(&mut work)?);
                CachedLexicalContribution::build(
                    LexicalContributionKey::Active(Arc::downgrade(active)),
                    &sealed,
                    &alive,
                    index.segments().len(),
                    accounting,
                    &mut work,
                )?
            }
        };
        let alive = contribution
            .active_alive
            .as_ref()
            .ok_or(LexicalAssemblyError::Store(StoreError::Synchronization {
                component: "active lexical contribution",
            }))?;
        index.push_shared_with_statistics(sealed, Arc::clone(&contribution.statistics));
        alive_sets.push(Arc::clone(alive));
        contributions.push(contribution);
        sources.push(StructuredLexicalSource::Active);
    }
    crate::fts::control::sort_by(
        &mut contributions,
        |left, right| match (&left.key, &right.key) {
            (LexicalContributionKey::Sealed(a, _), LexicalContributionKey::Sealed(b, _)) => {
                a.id.cmp(&b.id)
            }
            (LexicalContributionKey::Active(_), LexicalContributionKey::Active(_)) => {
                std::cmp::Ordering::Equal
            }
            (LexicalContributionKey::Active(_), _) => std::cmp::Ordering::Greater,
            (_, LexicalContributionKey::Active(_)) => std::cmp::Ordering::Less,
        },
        &mut work,
    )?;
    let assembly = LexicalAssembly {
        index,
        alive_sets,
        sources,
        contributions,
        vocabulary: Arc::new(Mutex::new(None)),
        _base: None,
        memory: stats::AccountedCounter::new(accounting, stats::AllocationComponent::Cache)
            .map_err(LexicalAssemblyError::Store)?,
    };
    let owned_bytes = assembly
        .owned_bytes()
        .and_then(|n| {
            n.checked_add(
                std::mem::size_of::<Mutex<Option<Arc<CachedVocabulary>>>>()
                    + 2 * std::mem::size_of::<usize>(),
            )
        })
        .ok_or(LexicalAssemblyError::Store(StoreError::BudgetExceeded {
            needed: u64::MAX,
            budget: u64::MAX,
            component: "cache",
        }))?;
    work.check_now()?;
    Ok((assembly, owned_bytes))
}

fn map_store_lexical_assembly_error(
    error: LexicalAssemblyError,
) -> crate::ingest::StoreLexicalError {
    match error {
        LexicalAssemblyError::Store(error) => QueryError::Store(error).into(),
        LexicalAssemblyError::Lexical(error) => {
            crate::planner::LexicalFilterError::from(error).into()
        }
        LexicalAssemblyError::MissingDocumentIdentity(segment_id) => {
            crate::ingest::StoreLexicalError::MissingDocumentIdentity { segment_id }
        }
        LexicalAssemblyError::Cancelled(error) => QueryError::Scan(error).into(),
    }
}

fn map_fusion_lexical_assembly_error(error: LexicalAssemblyError) -> crate::fusion::FusionError {
    match error {
        LexicalAssemblyError::Store(error) => crate::fusion::FusionError::Leg {
            leg: crate::fusion::FusionLeg::Lexical,
            kind: crate::fusion::LegFailureKind::Store(error.kind()),
            detail: error.to_string(),
        },
        LexicalAssemblyError::Lexical(error) => crate::fusion::FusionError::Leg {
            leg: crate::fusion::FusionLeg::Lexical,
            kind: crate::fusion::LegFailureKind::Lexical,
            detail: error.to_string(),
        },
        LexicalAssemblyError::MissingDocumentIdentity(segment_id) => {
            crate::fusion::FusionError::Leg {
                leg: crate::fusion::FusionLeg::Lexical,
                kind: crate::fusion::LegFailureKind::Invariant,
                detail: format!("sealed lexical segment {segment_id} has no document identity"),
            }
        }
        LexicalAssemblyError::Cancelled(error) => {
            crate::fusion::FusionError::from(QueryError::Scan(error))
        }
    }
}

fn accumulate_search_counters(
    total: &mut crate::fts::search::SearchCounters,
    delta: &crate::fts::search::SearchCounters,
) {
    total.docs_evaluated = total.docs_evaluated.saturating_add(delta.docs_evaluated);
    total.postings_decoded = total
        .postings_decoded
        .saturating_add(delta.postings_decoded);
    total.blocks_decoded = total.blocks_decoded.saturating_add(delta.blocks_decoded);
    total.blocks_skipped = total.blocks_skipped.saturating_add(delta.blocks_skipped);
}

fn map_fusion_controlled_lexical_error(
    error: crate::fts::search::ControlledSearchError<crate::scan::ScanError>,
) -> crate::fusion::FusionError {
    match error {
        crate::fts::search::ControlledSearchError::Index(error) => {
            crate::fusion::FusionError::Leg {
                leg: crate::fusion::FusionLeg::Lexical,
                kind: crate::fusion::LegFailureKind::Lexical,
                detail: error.to_string(),
            }
        }
        crate::fts::search::ControlledSearchError::Control(error) => {
            crate::fusion::FusionError::from(QueryError::Scan(error))
        }
    }
}

#[derive(Clone, Copy)]
enum PinnedLexicalQuery<'a> {
    Term(&'a crate::fts::search::TermQuery),
    Structured(&'a crate::fts::query::LexicalQuery),
}

enum LexicalDocumentError {
    SourceOrdinalOutOfRange,
    RowOverflow,
    SealedSourceAbsent,
    Segment(crate::segment::SegmentError),
    MissingDocumentIdentity(crate::segment::SegmentId),
    ActiveRowAbsent,
}

fn structured_lexical_document(
    snapshot: &PublishedSnapshot,
    active: &crate::ingest::ActiveSegment,
    sources: &[StructuredLexicalSource],
    doc: crate::fts::search::GlobalDocId,
    require_document_identity: bool,
) -> Result<Option<crate::ingest::DocumentVersion>, LexicalDocumentError> {
    let source = usize::try_from(doc.segment)
        .ok()
        .and_then(|slot| sources.get(slot))
        .ok_or(LexicalDocumentError::SourceOrdinalOutOfRange)?;
    let row = usize::try_from(doc.row).map_err(|_| LexicalDocumentError::RowOverflow)?;
    match source {
        StructuredLexicalSource::Sealed(ordinal) => {
            let segment = snapshot
                .segments()
                .get(*ordinal)
                .ok_or(LexicalDocumentError::SealedSourceAbsent)?;
            let document = segment
                .document_version(row)
                .map_err(LexicalDocumentError::Segment)?;
            if require_document_identity && document.is_none() {
                return Err(LexicalDocumentError::MissingDocumentIdentity(
                    segment.meta().id,
                ));
            }
            Ok(document)
        }
        StructuredLexicalSource::Active => {
            let document = active.document(row);
            if require_document_identity && document.is_none() {
                return Err(LexicalDocumentError::ActiveRowAbsent);
            }
            Ok(document)
        }
    }
}

fn map_store_lexical_document_error(
    error: LexicalDocumentError,
) -> crate::ingest::StoreLexicalError {
    match error {
        LexicalDocumentError::SourceOrdinalOutOfRange
        | LexicalDocumentError::RowOverflow
        | LexicalDocumentError::SealedSourceAbsent
        | LexicalDocumentError::ActiveRowAbsent => {
            QueryError::Store(StoreError::ActiveRowOverflow).into()
        }
        LexicalDocumentError::Segment(error) => {
            QueryError::Store(StoreError::Segment(error)).into()
        }
        LexicalDocumentError::MissingDocumentIdentity(segment_id) => {
            crate::ingest::StoreLexicalError::MissingDocumentIdentity { segment_id }
        }
    }
}

fn map_fusion_lexical_document_error(
    error: LexicalDocumentError,
    invariant_kind: crate::fusion::LegFailureKind,
    segment_kind: crate::fusion::LegFailureKind,
) -> crate::fusion::FusionError {
    let (kind, detail) = match error {
        LexicalDocumentError::SourceOrdinalOutOfRange => (
            invariant_kind,
            "lexical source ordinal is out of range".to_owned(),
        ),
        LexicalDocumentError::RowOverflow => {
            (invariant_kind, "lexical row exceeds usize".to_owned())
        }
        LexicalDocumentError::SealedSourceAbsent => {
            (invariant_kind, "lexical sealed source is absent".to_owned())
        }
        LexicalDocumentError::Segment(error) => (segment_kind, error.to_string()),
        LexicalDocumentError::MissingDocumentIdentity(segment_id) => (
            invariant_kind,
            format!("sealed lexical segment {segment_id} has no document identity"),
        ),
        LexicalDocumentError::ActiveRowAbsent => (
            invariant_kind,
            "lexical active source row is absent".to_owned(),
        ),
    };
    crate::fusion::FusionError::Leg {
        leg: crate::fusion::FusionLeg::Lexical,
        kind,
        detail,
    }
}

fn map_structured_leg_document_error(error: LexicalDocumentError) -> crate::fusion::FusionError {
    map_fusion_lexical_document_error(
        error,
        crate::fusion::LegFailureKind::Lexical,
        crate::fusion::LegFailureKind::Lexical,
    )
}

fn map_term_leg_document_error(error: LexicalDocumentError) -> crate::fusion::FusionError {
    map_fusion_lexical_document_error(
        error,
        crate::fusion::LegFailureKind::Invariant,
        crate::fusion::LegFailureKind::Segment,
    )
}

fn structured_lexical_row<'a>(
    snapshot: &'a PublishedSnapshot,
    active: &'a crate::ingest::ActiveSegment,
    sources: &[StructuredLexicalSource],
    doc: crate::fts::search::GlobalDocId,
) -> Result<(&'a str, crate::ingest::DocumentVersion), crate::ingest::StoreLexicalError> {
    let source = usize::try_from(doc.segment)
        .ok()
        .and_then(|slot| sources.get(slot))
        .ok_or(QueryError::Store(StoreError::ActiveRowOverflow))?;
    let row =
        usize::try_from(doc.row).map_err(|_| QueryError::Store(StoreError::ActiveRowOverflow))?;
    match source {
        StructuredLexicalSource::Sealed(ordinal) => {
            let segment = snapshot
                .segments()
                .get(*ordinal)
                .ok_or(QueryError::Store(StoreError::ActiveRowOverflow))?;
            let text = segment
                .query_stored_text()
                .map_err(StoreError::Segment)
                .map_err(QueryError::Store)?
                .and_then(|rows| rows.row(row).flatten())
                .ok_or(crate::ingest::StoreLexicalError::MissingStoredText {
                    segment_id: segment.meta().id,
                    row: doc.row,
                })?;
            let document = segment
                .document_version(row)
                .map_err(StoreError::Segment)
                .map_err(QueryError::Store)?
                .ok_or(crate::ingest::StoreLexicalError::MissingDocumentIdentity {
                    segment_id: segment.meta().id,
                })?;
            Ok((text, document))
        }
        StructuredLexicalSource::Active => {
            let text = active
                .text(row)
                .map_err(QueryError::Store)?
                .ok_or(QueryError::Store(StoreError::ActiveRowOverflow))?;
            let document = active
                .document(row)
                .ok_or(QueryError::Store(StoreError::ActiveRowOverflow))?;
            Ok((text, document))
        }
    }
}

type ExactLexicalLeg = (
    Vec<hybrid::LexicalHit>,
    Arc<LexicalAssembly>,
    crate::fts::search::SearchCounters,
    bool,
);

fn resolve_hybrid_leg_results<Vector, Lexical>(
    vector: Result<Vector, crate::fusion::FusionError>,
    lexical: Result<Lexical, crate::fusion::FusionError>,
) -> Result<(Vector, Lexical), crate::fusion::FusionError> {
    fn control_priority(error: &crate::fusion::FusionError) -> Option<u8> {
        match error {
            crate::fusion::FusionError::ReadCancelled { .. } => Some(0),
            crate::fusion::FusionError::Timeout { .. } => Some(1),
            crate::fusion::FusionError::Cancelled { .. } => Some(2),
            _ => None,
        }
    }

    match (vector, lexical) {
        (Ok(vector), Ok(lexical)) => Ok((vector, lexical)),
        (Err(vector), Ok(_)) => Err(vector),
        (Ok(_), Err(lexical)) => Err(lexical),
        (Err(vector), Err(lexical)) => {
            match (control_priority(&vector), control_priority(&lexical)) {
                (Some(vector_priority), Some(lexical_priority)) => {
                    if vector_priority <= lexical_priority {
                        Err(vector)
                    } else {
                        Err(lexical)
                    }
                }
                (Some(_), None) => Err(vector),
                (None, Some(_)) => Err(lexical),
                (None, None) => Err(vector),
            }
        }
    }
}

fn exact_structured_lexical_leg(
    inputs: LexicalInputs<'_>,
    _analyzer: &crate::fts::tokenizer::Analyzer,
    query: &crate::fts::query::LexicalQuery,
    bound: usize,
    cancellation: &QueryCancellation<'_>,
    preparation: &mut prepared_lexical::PreparedLexicalQuery<'_>,
) -> Result<ExactLexicalLeg, crate::fusion::FusionError> {
    let (snapshot, active) = (inputs.snapshot, inputs.active);
    let lexical_error = |detail: String| crate::fusion::FusionError::Leg {
        leg: crate::fusion::FusionLeg::Lexical,
        kind: crate::fusion::LegFailureKind::Lexical,
        detail,
    };
    let (prepared, cache_hit) = preparation.prepare_structured(inputs, cancellation)?;
    let assembly = &prepared.assembly;
    let index = &assembly.index;
    let sources = &assembly.sources;
    let Some(query_scoring) = prepared.query.as_ref() else {
        return Ok((
            Vec::new(),
            Arc::clone(assembly),
            crate::fts::search::SearchCounters::default(),
            cache_hit,
        ));
    };
    let allow_lists = assembly
        .alive_sets
        .iter()
        .map(|alive| alive.alive_bitmap())
        .collect::<Vec<_>>();
    let mut scratch_memory =
        stats::AccountedCounter::new(inputs.accounting, stats::AllocationComponent::Temporary)
            .map_err(QueryError::Store)?;
    let mut phrase = prepared_lexical::PhraseEligibility::new(
        index,
        Some(query),
        inputs.accounting,
        cancellation,
    )?;
    let result = crate::fts::prune::weighted::search(
        index,
        query_scoring,
        bound,
        &allow_lists,
        || {
            cancellation
                .check_graph()
                .map_err(QueryError::Scan)
                .map_err(crate::fusion::FusionError::from)
        },
        |doc| {
            phrase
                .matches(doc, cancellation)
                .map_err(crate::fusion::FusionError::from)
        },
        |bytes| {
            prepared_lexical::reserve_weighted_scratch(&mut scratch_memory, bytes)
                .map_err(crate::fusion::FusionError::from)
        },
    )
    .map_err(|error| match error {
        crate::fts::search::ControlledSearchError::Index(error) => lexical_error(error.to_string()),
        crate::fts::search::ControlledSearchError::Control(error) => error,
    })?;
    let counters = result.counters;
    let scored = result.hits;
    let mut joined = Vec::with_capacity(scored.len());
    for (doc, score) in scored {
        let document = structured_lexical_document(snapshot, active, sources, doc, false)
            .map_err(map_structured_leg_document_error)?;
        joined.push(hybrid::LexicalHit {
            doc,
            document,
            bm25: score,
        });
    }
    Ok((joined, Arc::clone(assembly), counters, cache_hit))
}

fn exact_lexical_leg(
    inputs: LexicalInputs<'_>,
    bound: usize,
    cancellation: &QueryCancellation<'_>,
    preparation: &mut prepared_lexical::PreparedLexicalQuery<'_>,
) -> Result<ExactLexicalLeg, crate::fusion::FusionError> {
    let (snapshot, active) = (inputs.snapshot, inputs.active);
    let (prepared, cache_hit) = preparation.prepare_term(inputs, cancellation)?;
    let assembly = &prepared.assembly;
    let index = &assembly.index;
    let sources = &assembly.sources;
    if index.segments().is_empty() {
        return Ok((
            Vec::new(),
            Arc::clone(assembly),
            crate::fts::search::SearchCounters::default(),
            cache_hit,
        ));
    }
    let k = bound.min(usize::try_from(index.document_count()).unwrap_or(usize::MAX));
    let allow_lists = assembly
        .alive_sets
        .iter()
        .map(|alive| alive.alive_bitmap())
        .collect::<Vec<_>>();
    let result =
        exact_hybrid_lexical_search(index, prepared.scoring()?, k, &allow_lists, cancellation)?;
    let mut joined = Vec::with_capacity(result.hits.len());
    for hit in result.hits {
        let document = structured_lexical_document(snapshot, active, sources, hit.doc, false)
            .map_err(map_term_leg_document_error)?;
        joined.push(hybrid::LexicalHit {
            doc: hit.doc,
            document,
            bm25: hit.score,
        });
    }
    Ok((joined, Arc::clone(assembly), result.counters, cache_hit))
}

fn exact_hybrid_lexical_search(
    index: &crate::fts::index::LexicalIndex,
    prepared: &crate::fts::search::PreparedTermQuery,
    k: usize,
    allow_lists: &[&crate::meta::DocBitmap],
    cancellation: &QueryCancellation<'_>,
) -> Result<crate::fts::search::SearchResult, crate::fusion::FusionError> {
    let allowed = allow_lists
        .iter()
        .map(|allow_list| allow_list.cardinality())
        .fold(0_u64, u64::saturating_add);
    if allowed.saturating_mul(crate::planner::LEXICAL_ALLOW_LIST_DIVISOR) <= index.document_count()
    {
        return crate::fts::search::search_allow_list_prepared_controlled(
            index,
            prepared,
            k,
            allow_lists,
            || cancellation.check_graph(),
        )
        .map_err(map_fusion_controlled_lexical_error);
    }
    crate::fts::prune::search_pruned_prepared_controlled(index, prepared, k, allow_lists, || {
        cancellation.check_graph()
    })
    .map_err(map_fusion_controlled_lexical_error)
}

#[allow(clippy::too_many_arguments)]
fn search_pinned(
    pool: Option<&pool::QueryPool>,
    snapshot: &Arc<PublishedSnapshot>,
    active: &crate::ingest::ActiveSegment,
    accounting: &Arc<stats::Accounting>,
    generation: u64,
    epoch: Option<crate::epoch::EpochIdentity>,
    request: crate::ingest::SearchRequest<'_>,
    k: usize,
    options: SearchOptions,
    control: QueryControl,
    graph_bound_mode: GraphBoundMode,
    requested_tier: Option<SearchTier>,
    hybrid_scan_reason: Option<crate::planner::ScanReason>,
    started: std::time::Instant,
    clock: &dyn MonotonicClock,
    vector_preparation: &mut prepared::PreparedVectorQuery<'_>,
    #[cfg(any(test, feature = "test-seams"))] vector_fault_controller: Option<
        &crate::scan::vector_fault::VectorFaultController,
    >,
) -> Result<crate::ingest::SearchOutcome, QueryError> {
    use crate::ingest::{GraphSearchStats, RowSource, SearchOutcome};
    use crate::quant::prepare_bit4_query;
    use crate::scan::{ScanQuery, ScanRequest, ScanRows, ScanStats};

    let vector_started = timing_start(clock);
    let scan_options = options.scan();
    let rescore_options = options.scan_rescore();
    let mut rescore_counters = crate::diag::ScanRescoreCounters::default();
    if let Some(options) = rescore_options {
        options.frontier(k, 0)?;
    }
    let query = request.vector();
    prepared::validate_vector_coordinates(query, crate::kernels::MAX_DOT_I8_DIMENSION, |_| {
        Ok::<(), std::convert::Infallible>(())
    })
    .map_err(|error| match error {
        prepared::VectorValidationError::Data(error) => {
            QueryError::Scan(crate::scan::ScanError::Quant(error))
        }
        prepared::VectorValidationError::Control(never) => match never {},
    })?;
    vector_preparation.validate_binding(query, epoch)?;
    let bit4_query = std::cell::OnceCell::new();
    let int8_query = std::cell::OnceCell::new();
    let mut candidates = Vec::new();
    let mut dims_touched = 0_u64;
    let mut bytes_read = 0_u64;
    let mut worst_squared_l2: Option<f64> = None;
    let mut worst_exhaustive = true;
    let mut worker_thread_ids = Vec::new();
    let mut graph_stats = GraphSearchStats::default();
    let mut plans = Vec::new();
    let auto_graph_options = if matches!(options.tier(), SearchTier::Auto)
        && snapshot.segments().iter().any(|segment| {
            segment
                .directory()
                .iter()
                .any(|entry| entry.kind == crate::segment::layout::RegionKind::GraphNodeBlocks.id())
        }) {
        Some(auto_graph_search_options(snapshot)?)
    } else {
        None
    };

    if matches!(options.tier(), SearchTier::Graph(_)) {
        for segment in snapshot.segments() {
            if !segment
                .directory()
                .iter()
                .any(|entry| entry.kind == crate::segment::layout::RegionKind::GraphNodeBlocks.id())
            {
                return Err(QueryError::Store(StoreError::GraphUnavailable {
                    segment_id: segment.meta().id,
                }));
            }
        }
    }
    let full_precision = matches!(options.tier(), SearchTier::Exact | SearchTier::Graph(_))
        || (matches!(options.tier(), SearchTier::Auto)
            && auto_uses_full_precision(snapshot, active));

    // Reuse one bounded worker lane across this pinned query's segments.
    let exact_lane = if full_precision {
        pool.map_or(0, |pool| {
            pool.reserve_exact_lane(scan_options.thread_budget)
        })
    } else {
        0
    };

    if !active.is_empty() {
        let mut alive = active.alive().map_err(QueryError::Store)?;
        if let Some(filter) = request.filter {
            alive.retain(&filter.active_rows(active)?);
        }
        let mut exact_score = full_precision;
        let mut approximate = false;
        let frontier = rescore_options
            .map(|options| options.frontier(k, alive.live_count()))
            .transpose()?
            .unwrap_or(k);
        let mut exact_memory = ExactScanMemory::new(accounting)?;
        let mut outcome = match options.tier() {
            SearchTier::Auto if full_precision => {
                let lease = SnapshotLease::new_at(Arc::clone(snapshot), generation);
                scan_active_squared_l2(
                    active,
                    &alive,
                    request.vector(),
                    frontier,
                    ExactScanDispatch {
                        pool,
                        accounting,
                        control: &control,
                        lease: &lease,
                        thread_budget: scan_options.thread_budget,
                        lane: exact_lane,
                    },
                    &mut exact_memory,
                    #[cfg(any(test, feature = "test-seams"))]
                    vector_fault_controller,
                    #[cfg(any(test, feature = "test-seams"))]
                    vector_fault_tier(options.tier()),
                )?
            }
            SearchTier::Auto | SearchTier::Scan => {
                let query_pool = pool.ok_or(QueryError::Store(StoreError::Synchronization {
                    component: "active scan-tier query pool",
                }))?;
                let bit4_query = bit4_query
                    .get_or_init(|| prepare_bit4_query(request.vector(), 0))
                    .as_ref()
                    .map_err(|error| crate::scan::ScanError::Quant(error.clone()))
                    .map_err(QueryError::Scan)?;
                execute_store_scan(
                    query_pool,
                    ScanRequest {
                        query: ScanQuery::Bit4(bit4_query),
                        rows: ScanRows::Bit4RowMajor {
                            codes: active.codes(),
                            factors: active.factors(),
                        },
                        row_mask: Some(alive.scan_mask()),
                    },
                    frontier,
                    scan_options,
                    control.clone(),
                    SnapshotLease::new_at(Arc::clone(snapshot), generation),
                    #[cfg(any(test, feature = "test-seams"))]
                    vector_fault_controller,
                    #[cfg(any(test, feature = "test-seams"))]
                    crate::scan::vector_fault::VectorRowSource::Active,
                    #[cfg(any(test, feature = "test-seams"))]
                    vector_fault_tier(options.tier()),
                )?
            }
            SearchTier::Exact | SearchTier::Graph(_) => {
                let lease = SnapshotLease::new_at(Arc::clone(snapshot), generation);
                scan_active_squared_l2(
                    active,
                    &alive,
                    request.vector(),
                    frontier,
                    ExactScanDispatch {
                        pool,
                        accounting,
                        control: &control,
                        lease: &lease,
                        thread_budget: scan_options.thread_budget,
                        lane: exact_lane,
                    },
                    &mut exact_memory,
                    #[cfg(any(test, feature = "test-seams"))]
                    vector_fault_controller,
                    #[cfg(any(test, feature = "test-seams"))]
                    vector_fault_tier(options.tier()),
                )?
            }
        };
        if let Some(rescore) = rescore_options {
            let lease = SnapshotLease::new_at(Arc::clone(snapshot), generation);
            let cancellation = QueryCancellation::new(&control, &lease);
            let (counters, exhaustive, worst) = rescored_scan::rescore(
                &mut outcome.candidates,
                &mut outcome.stats,
                active.vectors(),
                query,
                k,
                alive.live_count(),
                rescore,
                &cancellation,
                accounting,
                &mut exact_memory,
            )?;
            outcome.worst_score = worst;
            exact_score = true;
            approximate = !exhaustive;
            rescored_scan::accumulate(&mut rescore_counters, counters)?;
        }
        fold_worst_squared_l2(&mut worst_squared_l2, &mut worst_exhaustive, &outcome);
        merge_store_outcome(
            outcome,
            RowSource::Active,
            |row| Ok(active.document(row)),
            &mut candidates,
            &mut dims_touched,
            &mut bytes_read,
            &mut worker_thread_ids,
            exact_score,
            #[cfg(any(test, feature = "test-seams"))]
            vector_fault_controller,
            #[cfg(any(test, feature = "test-seams"))]
            vector_fault_tier(options.tier()),
        )?;
        let mut plan = crate::planner::SegmentPlan::unfiltered_scan(
            RowSource::Active,
            crate::planner::SegmentTier::ActiveScan,
            alive.live_count(),
            crate::planner::ScanReason::ActiveSegment,
        );
        plan.approximate = approximate;
        plans.push(plan);
        if matches!(graph_bound_mode, GraphBoundMode::Shared) {
            retain_global_top_k(&mut candidates, k);
        }
    }

    let mut ordered_segments = snapshot.segments().iter().collect::<Vec<_>>();
    if matches!(graph_bound_mode, GraphBoundMode::Shared) {
        // Larger immutable segments have more opportunities to supply the
        // first competitive top-k. Row count is already in the manifest, so
        // this ordering tightens the bound without query-time artifact I/O.
        ordered_segments.sort_unstable_by(|left, right| {
            right
                .meta()
                .row_count
                .cmp(&left.meta().row_count)
                .then_with(|| left.meta().id.cmp(&right.meta().id))
        });
    }

    for segment in ordered_segments {
        let original_alive = segment.query_alive().map_err(QueryError::Store)?;
        let mut filtered_alive;
        let alive = if let Some(filter) = request.filter {
            filtered_alive = (*original_alive).clone();
            filtered_alive.retain(&filter.sealed_rows(segment)?);
            &filtered_alive
        } else {
            original_alive.as_ref()
        };
        let source = RowSource::Sealed(segment.meta().id);
        // Automatic tiering follows the artifact that is atomically published
        // now, not the tier policy's desired future state. A due-but-unbuilt
        // graph is therefore scanned silently and correctly: that is the
        // adaptive ladder working, not a fallback hiding a broken contract.
        let graph_options = match options.tier() {
            SearchTier::Graph(graph_options) => Some(graph_options),
            SearchTier::Auto
                if segment.directory().iter().any(|entry| {
                    entry.kind == crate::segment::layout::RegionKind::GraphNodeBlocks.id()
                }) =>
            {
                auto_graph_options
            }
            SearchTier::Auto | SearchTier::Exact | SearchTier::Scan => None,
        };
        if let Some(graph_options) = graph_options {
            worst_exhaustive = false;
            let competitive_distance = global_competitive_distance(&candidates, k);
            let MergedSegmentGraph {
                plan,
                traversed,
                _result,
                _scratch,
                _lease,
            } = traverse_segment_graph(
                segment,
                alive,
                source,
                graph_options,
                graph_bound_mode,
                request,
                k,
                vector_preparation,
                accounting,
                snapshot,
                generation,
                &control,
                competitive_distance,
                #[cfg(any(test, feature = "test-seams"))]
                vector_fault_controller,
                options.tier(),
            )?
            .merge_into(
                segment,
                alive,
                source,
                &mut candidates,
                &mut dims_touched,
                &mut bytes_read,
                &mut worker_thread_ids,
                &mut graph_stats,
                #[cfg(any(test, feature = "test-seams"))]
                vector_fault_controller,
                options.tier(),
            )?;
            plans.push(plan);
            if traversed && matches!(graph_bound_mode, GraphBoundMode::Shared) {
                retain_global_top_k(&mut candidates, k);
            }
            continue;
        }

        let mut exact_memory = ExactScanMemory::new(accounting)?;
        let frontier = rescore_options
            .map(|options| options.frontier(k, alive.live_count()))
            .transpose()?
            .unwrap_or(k);
        let mut outcome = scan_sealed_segment(
            pool,
            snapshot,
            segment,
            alive,
            generation,
            request,
            frontier,
            scan_options,
            &control,
            &bit4_query,
            &int8_query,
            full_precision,
            exact_lane,
            accounting,
            &mut exact_memory,
            source,
            options.tier(),
            #[cfg(any(test, feature = "test-seams"))]
            vector_fault_controller,
        )?;
        let mut exact_score = full_precision || segment.meta().scheme == 0;
        let mut approximate = false;
        if let Some(rescore) = rescore_options {
            let lease = SnapshotLease::new_at(Arc::clone(snapshot), generation);
            let cancellation = QueryCancellation::new(&control, &lease);
            let vectors = exact_rescore_rows_for_search(
                segment,
                #[cfg(any(test, feature = "test-seams"))]
                vector_fault_controller,
                #[cfg(any(test, feature = "test-seams"))]
                vector_fault_tier(options.tier()),
            )?;
            let (counters, exhaustive, worst) = rescored_scan::rescore(
                &mut outcome.candidates,
                &mut outcome.stats,
                vectors,
                query,
                k,
                alive.live_count(),
                rescore,
                &cancellation,
                accounting,
                &mut exact_memory,
            )?;
            outcome.worst_score = worst;
            exact_score = true;
            approximate = !exhaustive;
            rescored_scan::accumulate(&mut rescore_counters, counters)?;
        }
        fold_worst_squared_l2(&mut worst_squared_l2, &mut worst_exhaustive, &outcome);
        merge_store_outcome(
            outcome,
            source,
            |row| {
                segment
                    .document_version(row)
                    .map_err(StoreError::Segment)
                    .map_err(QueryError::Store)
            },
            &mut candidates,
            &mut dims_touched,
            &mut bytes_read,
            &mut worker_thread_ids,
            exact_score,
            #[cfg(any(test, feature = "test-seams"))]
            vector_fault_controller,
            #[cfg(any(test, feature = "test-seams"))]
            vector_fault_tier(options.tier()),
        )?;
        let tier = if segment
            .directory()
            .iter()
            .any(|entry| entry.kind == crate::segment::layout::RegionKind::GraphNodeBlocks.id())
        {
            crate::planner::SegmentTier::SealedGraph
        } else {
            crate::planner::SegmentTier::SealedScan
        };
        let mut plan = crate::planner::SegmentPlan::unfiltered_scan(
            source,
            tier,
            alive.live_count(),
            sealed_scan_reason(segment, requested_tier, hybrid_scan_reason),
        );
        plan.approximate = approximate;
        plans.push(plan);
        if matches!(graph_bound_mode, GraphBoundMode::Shared) {
            retain_global_top_k(&mut candidates, k);
        }
    }

    retain_global_top_k(&mut candidates, k);
    let stats = ScanStats {
        dims_touched,
        bytes_read,
        threads_used: worker_thread_ids.len(),
        worker_thread_ids,
    };
    accounting.record_plans(&plans).map_err(QueryError::Store)?;
    // Ceiling construction is part of the vector query, including the first
    // immutable norm-enclosure pass. Finalize elapsed only after it completes.
    let vector_ceiling = Some(exact_vector_ceiling(snapshot, active, query)?);
    let mut diagnostics = crate::diag::QueryDiagnostics::vector(crate::diag::VectorDiagnostics {
        snapshot_generation: generation,
        indexed_through_seq: active
            .indexed_through_seq()
            .max(crate::wal::LogSeq::new(snapshot.absorbed_through())),
        approximate: plans.iter().any(|plan| plan.approximate),
        exact_rescore: candidates.iter().all(|candidate| candidate.exact_score()),
        requested_k: k,
        returned: candidates.len(),
        budget_exhausted: false,
        plan: plans,
        scan: stats.clone(),
        graph: graph_stats,
        epoch,
        elapsed: clock.now().saturating_duration_since(started),
    });
    diagnostics.scan_rescore = rescore_options.map(|_| rescore_counters);
    diagnostics.timings = crate::diag::QUERY_TIMING_ENABLED.then(|| crate::diag::QueryTimings {
        vector: timing_elapsed(clock, vector_started),
        ..Default::default()
    });
    Ok(SearchOutcome {
        candidates,
        // Every tier reports the same deterministic ceiling. A graph
        // traversal cannot name the farthest alive row, and letting the
        // anchor depend on the tier would let the tier change ranking,
        // so the ceiling is the anchor everywhere.
        vector_ceiling,
        stats,
        graph_stats,
        generation,
        epoch,
        diagnostics,
    })
}

/// Reports whether any sealed segment has a published graph region.
///
/// Only a published artifact counts. A graph that policy says is due but
/// that maintenance has not built yet is not usable by a query.
/// Widening rounds visit `ef` rows per segment, and `ef` grows with the
/// requested `k`. Past this share of the corpus a traversal reads most of
/// the segment anyway, so the round is cheaper and exact as a scan.
const GRAPH_WIDENING_CAP_NUMERATOR: usize = 1;
/// Denominator of [`GRAPH_WIDENING_CAP_NUMERATOR`]; the cap is 25%.
const GRAPH_WIDENING_CAP_DENOMINATOR: usize = 4;

/// Reports whether a widening round of width `k` should still use the graph.
///
/// `ef` follows `graph::search`: at least the SIFT floor, and otherwise a
/// small multiple of `k`. The angular multiple is used so the cap is
/// evaluated against the widest pool any profile would request, which
/// keeps the decision independent of the distance metric in play.
fn graph_round_is_worth_it(k: usize, corpus_rows: usize) -> bool {
    if corpus_rows == 0 {
        return true;
    }
    let ef = graph_round_ef(k);
    let cap =
        corpus_rows.saturating_mul(GRAPH_WIDENING_CAP_NUMERATOR) / GRAPH_WIDENING_CAP_DENOMINATOR;
    ef <= cap
}

fn graph_round_ef(k: usize) -> usize {
    k.saturating_mul(4).max(140)
}

fn sealed_scan_reason(
    segment: &crate::segment::reader::SegmentReader,
    requested_tier: Option<SearchTier>,
    hybrid_scan_reason: Option<crate::planner::ScanReason>,
) -> crate::planner::ScanReason {
    use crate::planner::{ExplicitScanTier, ScanReason};

    let has_graph = segment
        .directory()
        .iter()
        .any(|entry| entry.kind == crate::segment::layout::RegionKind::GraphNodeBlocks.id());
    if has_graph && let Some(reason) = hybrid_scan_reason {
        return reason;
    }
    match requested_tier {
        Some(SearchTier::Exact) => return ScanReason::ExplicitTier(ExplicitScanTier::Exact),
        Some(SearchTier::Scan) => return ScanReason::ExplicitTier(ExplicitScanTier::Scan),
        Some(SearchTier::Auto | SearchTier::Graph(_)) | None => {}
    }
    let rows = segment.meta().row_count;
    let min_rows = crate::tier::thresholds::for_bucket(segment.meta().dims, segment.meta().scheme)
        .graph_min_rows;
    if rows < min_rows {
        ScanReason::BelowGraphThreshold { rows, min_rows }
    } else {
        ScanReason::GraphPending
    }
}

fn snapshot_has_graph(snapshot: &PublishedSnapshot) -> bool {
    snapshot.segments().iter().any(|segment| {
        segment
            .directory()
            .iter()
            .any(|entry| entry.kind == crate::segment::layout::RegionKind::GraphNodeBlocks.id())
    })
}

fn exact_vector_ceiling(
    snapshot: &PublishedSnapshot,
    active: &crate::ingest::ActiveSegment,
    query: &[f32],
) -> Result<f64, QueryError> {
    fold_exact_vector_ceiling(
        snapshot,
        active,
        query,
        remembered_segment_vector_ceiling_norm_range,
    )
}

/// Returns a sealed segment's norm enclosure, walking its region only once.
///
/// The enclosure is the query-independent half of the ceiling: it depends on
/// the segment's immutable factor or rescore bytes alone. Every tier and every
/// hybrid widening round would otherwise rewalk every factor record of every
/// segment to reach the same two numbers.
fn remembered_segment_vector_ceiling_norm_range(
    segment: &crate::segment::reader::SegmentReader,
) -> Result<crate::graph::search::GraphSegmentNormRange, QueryError> {
    if let Some(range) = segment.cached_vector_ceiling_norm_range() {
        return Ok(range);
    }
    let range = compute_segment_vector_ceiling_norm_range(segment)?;
    segment.remember_vector_ceiling_norm_range(range);
    Ok(range)
}

/// Folds every segment's norm enclosure into the query-wide ceiling.
///
/// `segment_range` supplies one sealed segment's query-independent enclosure.
/// Production passes the remembering source; the differential test passes the
/// recomputing one, so both walk identical fold arithmetic.
fn fold_exact_vector_ceiling(
    snapshot: &PublishedSnapshot,
    active: &crate::ingest::ActiveSegment,
    query: &[f32],
    mut segment_range: impl FnMut(
        &crate::segment::reader::SegmentReader,
    )
        -> Result<crate::graph::search::GraphSegmentNormRange, QueryError>,
) -> Result<f64, QueryError> {
    let mut ceiling = 0.0_f64;
    if !active.is_empty() {
        // The active segment is the one mutable row source, so its enclosure
        // is recomputed on every query and is never remembered.
        ceiling = fold_vector_ceiling(
            ceiling,
            crate::graph::search::GraphSegmentNormRange::from_factors(active.factors()),
            query,
        )?;
    }
    for segment in snapshot.segments() {
        let range = segment_range(segment)?;
        ceiling = fold_vector_ceiling(ceiling, range, query)?;
    }
    Ok(ceiling)
}

/// Walks one sealed segment's factor or rescore region for its norm enclosure.
fn compute_segment_vector_ceiling_norm_range(
    segment: &crate::segment::reader::SegmentReader,
) -> Result<crate::graph::search::GraphSegmentNormRange, QueryError> {
    match segment.meta().scheme {
        4 => Ok(crate::graph::search::GraphSegmentNormRange::from_factors(
            segment
                .query_bit4_factors()
                .map_err(StoreError::Segment)
                .map_err(QueryError::Store)?,
        )),
        0 | 2 => Ok(
            crate::graph::search::GraphSegmentNormRange::from_exact_rows(
                segment
                    .query_rescore_f32()
                    .map_err(StoreError::Segment)
                    .map_err(QueryError::Store)?,
                segment.meta().dims as usize,
            ),
        ),
        scheme => Err(QueryError::Store(StoreError::Segment(
            crate::segment::SegmentError::Geometry(format!(
                "vector norm enclosure does not support scheme {scheme}"
            )),
        ))),
    }
}

fn fold_vector_ceiling(
    current: f64,
    range: crate::graph::search::GraphSegmentNormRange,
    query: &[f32],
) -> Result<f64, QueryError> {
    let ceiling = range.squared_l2_upper_bound(query);
    if !ceiling.is_finite() {
        return Err(QueryError::Store(StoreError::Segment(
            crate::segment::SegmentError::Geometry(
                "vector norm enclosure did not produce a finite squared-L2 ceiling".to_owned(),
            ),
        )));
    }
    Ok(current.max(ceiling))
}

/// Folds one segment's exhaustive worst score into the query-wide farthest
/// distance. A scan that returned candidates without a worst score did not
/// visit every allowed row, so the query-wide anchor is not available.
fn fold_worst_squared_l2(
    worst: &mut Option<f64>,
    exhaustive: &mut bool,
    outcome: &crate::scan::ScanOutcome,
) {
    match outcome.worst_score {
        Some(score) => {
            let squared_l2 = -f64::from(score);
            *worst = Some(worst.map_or(squared_l2, |current: f64| current.max(squared_l2)));
        }
        None => {
            if !outcome.candidates.is_empty() {
                *exhaustive = false;
            }
        }
    }
}

// Boxing the traversed variant would add an allocation to every graph segment.
#[allow(clippy::large_enum_variant)]
enum SegmentGraphResult<'a> {
    Pruned {
        plan: crate::planner::SegmentPlan,
        graph_stats: crate::ingest::GraphSearchStats,
        _lease: SnapshotLease,
    },
    Traversed {
        result: crate::graph::search::GraphSearchResult,
        plan: crate::planner::SegmentPlan,
        dims_touched: u64,
        bytes_read: u64,
        caller_thread: std::thread::ThreadId,
        graph_stats: crate::ingest::GraphSearchStats,
        _scratch: graph_cache::GraphScratchLease<'a>,
        _lease: SnapshotLease,
    },
}

struct MergedSegmentGraph<'a> {
    plan: crate::planner::SegmentPlan,
    traversed: bool,
    _result: Option<crate::graph::search::GraphSearchResult>,
    _scratch: Option<graph_cache::GraphScratchLease<'a>>,
    _lease: SnapshotLease,
}

impl<'a> SegmentGraphResult<'a> {
    #[allow(clippy::too_many_arguments)]
    fn merge_into(
        self,
        segment: &crate::segment::reader::SegmentReader,
        alive: &crate::meta::AliveSet,
        source: crate::ingest::RowSource,
        candidates: &mut Vec<crate::ingest::SearchCandidate>,
        dims_touched: &mut u64,
        bytes_read: &mut u64,
        worker_thread_ids: &mut Vec<std::thread::ThreadId>,
        graph_stats: &mut crate::ingest::GraphSearchStats,
        #[cfg(any(test, feature = "test-seams"))] vector_fault_controller: Option<
            &crate::scan::vector_fault::VectorFaultController,
        >,
        _tier: SearchTier,
    ) -> Result<MergedSegmentGraph<'a>, QueryError> {
        match self {
            Self::Pruned {
                plan,
                graph_stats: delta,
                _lease,
            } => {
                add_traversal_stats(graph_stats, delta)?;
                Ok(MergedSegmentGraph {
                    plan,
                    traversed: false,
                    _result: None,
                    _scratch: None,
                    _lease,
                })
            }
            Self::Traversed {
                result,
                plan,
                dims_touched: traversal_dims_touched,
                bytes_read: traversal_bytes_read,
                caller_thread,
                graph_stats: delta,
                _scratch,
                _lease,
            } => {
                *dims_touched = dims_touched
                    .checked_add(traversal_dims_touched)
                    .ok_or(QueryError::Scan(crate::scan::ScanError::ArithmeticOverflow))?;
                *bytes_read = bytes_read
                    .checked_add(traversal_bytes_read)
                    .ok_or(QueryError::Scan(crate::scan::ScanError::ArithmeticOverflow))?;
                if !worker_thread_ids.contains(&caller_thread) {
                    worker_thread_ids.push(caller_thread);
                }
                add_traversal_stats(graph_stats, delta)?;
                reserve_global_candidates(
                    candidates,
                    result
                        .candidates()
                        .iter()
                        .filter(|candidate| alive.is_alive(candidate.row_id()))
                        .count(),
                    #[cfg(any(test, feature = "test-seams"))]
                    vector_fault_controller,
                    #[cfg(any(test, feature = "test-seams"))]
                    vector_fault_tier(_tier),
                )?;
                for candidate in result.candidates() {
                    if !alive.is_alive(candidate.row_id()) {
                        continue;
                    }
                    let score = -(candidate.distance() as f32);
                    if !score.is_finite() {
                        return Err(QueryError::Graph(
                            crate::graph::search::GraphSearchError::Geometry(format!(
                                "graph result distance for row {} is not representable as f32",
                                candidate.row_id()
                            )),
                        ));
                    }
                    candidates.push(crate::ingest::SearchCandidate::new(
                        crate::ingest::GlobalRowId::new(source, candidate.row_id()),
                        segment
                            .document_version(candidate.row_id() as usize)
                            .map_err(StoreError::Segment)
                            .map_err(QueryError::Store)?,
                        score,
                        true,
                    ));
                }
                Ok(MergedSegmentGraph {
                    plan,
                    traversed: true,
                    _result: Some(result),
                    _scratch: Some(_scratch),
                    _lease,
                })
            }
        }
    }
}

#[allow(clippy::too_many_arguments)]
fn traverse_segment_graph<'a>(
    segment: &'a crate::segment::reader::SegmentReader,
    alive: &crate::meta::AliveSet,
    source: crate::ingest::RowSource,
    graph_options: GraphSearchOptions,
    graph_bound_mode: GraphBoundMode,
    request: crate::ingest::SearchRequest<'_>,
    k: usize,
    vector_preparation: &mut prepared::PreparedVectorQuery<'_>,
    accounting: &Arc<stats::Accounting>,
    snapshot: &Arc<PublishedSnapshot>,
    generation: u64,
    control: &QueryControl,
    competitive_distance: Option<f32>,
    #[cfg(any(test, feature = "test-seams"))] vector_fault_controller: Option<
        &crate::scan::vector_fault::VectorFaultController,
    >,
    _tier: SearchTier,
) -> Result<SegmentGraphResult<'a>, QueryError> {
    let lease = SnapshotLease::new_at(Arc::clone(snapshot), generation);
    let cancellation = QueryCancellation::new(control, &lease);
    cancellation.check_graph().map_err(map_scan_error)?;
    let (graph, graph_validated, prepared_entry_seed_discovered, norm_range) =
        match graph_bound_mode {
            GraphBoundMode::Shared => {
                let prepared = segment
                    .graph_search_cache
                    .prepare_shared(segment, &cancellation)
                    .map_err(map_graph_cache_error)?;
                (
                    prepared.graph,
                    prepared.graph_validated,
                    prepared.entry_seed_discovered,
                    Some(prepared.norm_range),
                )
            }
            #[cfg(test)]
            GraphBoundMode::Independent => {
                let (graph, graph_validated) = segment
                    .graph_search_cache
                    .bind_graph(segment)
                    .map_err(map_graph_cache_error)?;
                (graph, graph_validated, false, None)
            }
        };
    let rescore = query_rescore_rows_for_search(
        segment,
        #[cfg(any(test, feature = "test-seams"))]
        vector_fault_controller,
        #[cfg(any(test, feature = "test-seams"))]
        vector_fault_tier(_tier),
    )?;
    let node_count = graph.node_count() as usize;
    let segment_k = k.min(node_count);
    let has_tombstones = alive.tombstone_count() != 0;
    let target_live_k = if has_tombstones {
        usize::try_from(alive.live_count())
            .map_err(|_| QueryError::Scan(crate::scan::ScanError::ArithmeticOverflow))?
            .min(k)
    } else {
        segment_k
    };
    let maximum_candidate_k = if has_tombstones {
        let tombstone_count = usize::try_from(alive.tombstone_count())
            .map_err(|_| QueryError::Scan(crate::scan::ScanError::ArithmeticOverflow))?;
        target_live_k
            .checked_add(tombstone_count)
            .ok_or(QueryError::Scan(crate::scan::ScanError::ArithmeticOverflow))?
            .min(node_count)
            .max(segment_k)
    } else {
        segment_k
    };
    if let Some(ef) = graph_options.ef()
        && ef < segment_k
    {
        return Err(QueryError::Graph(
            crate::graph::search::GraphSearchError::AdaptiveEf(
                crate::graph::search::AdaptiveEfError::ExplicitBelowK { k: segment_k, ef },
            ),
        ));
    }
    let request_for = |candidate_k| {
        let graph_request = crate::graph::search::GraphSearchRequest::new(
            request.vector(),
            candidate_k,
            graph_options.seed(),
        )
        .with_profile(graph_options.profile());
        // An explicit ef controls the caller's live-candidate window.
        // Tombstones are engine state, so reserve only the additional
        // width needed to replace deleted rows instead of turning a
        // previously valid ef=k request into an internal below-k error.
        graph_options.ef().map_or(graph_request, |ef| {
            graph_request.with_ef(ef.max(candidate_k))
        })
    };
    let mut candidate_k = segment_k;
    let mut graph_request = request_for(candidate_k);
    let effective_ef = graph_request
        .effective_ef(graph.node_count() as usize)
        .map_err(crate::graph::search::GraphSearchError::AdaptiveEf)
        .map_err(QueryError::Graph)?;
    let requestable_max_k = maximum_candidate_k;
    let scratch_ef = if has_tombstones && requestable_max_k > candidate_k {
        request_for(requestable_max_k)
            .effective_ef(node_count)
            .map_err(crate::graph::search::GraphSearchError::AdaptiveEf)
            .map_err(QueryError::Graph)?
    } else {
        effective_ef
    };
    if let (Some(norm_range), Some(competitive_distance)) = (norm_range, competitive_distance)
        && norm_range.squared_l2_upper_bound(request.vector()) <= f64::from(f32::MAX)
        && (norm_range.squared_l2_lower_bound(request.vector()) as f32) > competitive_distance
    {
        cancellation.check_graph().map_err(map_scan_error)?;
        let graph_stats = crate::ingest::GraphSearchStats {
            graph_validations: usize::from(graph_validated),
            entry_seed_discoveries: usize::from(prepared_entry_seed_discovered),
            segments_pruned_by_bound: 1,
            ..crate::ingest::GraphSearchStats::default()
        };
        let plan = crate::planner::SegmentPlan::unfiltered_pruned(source, alive.live_count());
        return Ok(SegmentGraphResult::Pruned {
            plan,
            graph_stats,
            _lease: lease,
        });
    }
    let mut scratch = segment
        .graph_search_cache
        .checkout(graph, scratch_ef, accounting, &cancellation)
        .map_err(map_graph_cache_error)?;
    let entry_seed_discovered = prepared_entry_seed_discovered || scratch.entry_seed_discovered();
    let entries = scratch.entries();
    let mut searcher = crate::graph::search::GraphSearcher::with_entry_row_ids(
        graph,
        rescore,
        entries,
        scratch.scratch_mut().map_err(map_graph_error)?,
    )
    .map_err(map_graph_error)?
    .with_rescore_validator(segment);
    #[cfg(any(test, feature = "test-seams"))]
    if let Some(after_hops) = graph_options.cancel_after_hops() {
        let token = cancellation.cancel_token_for_test().ok_or_else(|| {
            QueryError::Graph(crate::graph::search::GraphSearchError::Geometry(
                "test hop cancellation requires a Cancel query control".to_owned(),
            ))
        })?;
        let _observed_hops = searcher.cancel_after_hops(after_hops, token);
    }
    #[cfg(any(test, feature = "test-seams"))]
    if let Some(controller) = vector_fault_controller {
        searcher = searcher.with_vector_fault_controller(
            controller,
            vector_fault_source(source),
            crate::scan::vector_fault::VectorSearchTier::Graph,
        );
    }
    let mut traversal_dims_touched = 0_u64;
    let mut traversal_bytes_read = 0_u64;
    let mut traversal_epoch_clears = 0_usize;
    let mut traversal_candidates_scored = 0_usize;
    let mut traversal_candidates_rescored = 0_usize;
    let result = loop {
        let result = searcher.search_reusing(
            graph_request,
            Some(&cancellation),
            vector_preparation,
            segment,
            source,
        )?;
        let counters = result.counters();
        traversal_dims_touched = traversal_dims_touched
            .checked_add(counters.dims_touched())
            .ok_or(QueryError::Scan(crate::scan::ScanError::ArithmeticOverflow))?;
        traversal_bytes_read = traversal_bytes_read
            .checked_add(counters.bytes_read())
            .ok_or(QueryError::Scan(crate::scan::ScanError::ArithmeticOverflow))?;
        traversal_epoch_clears = traversal_epoch_clears
            .checked_add(usize::from(counters.visited_epoch_cleared()))
            .ok_or(QueryError::Scan(crate::scan::ScanError::ArithmeticOverflow))?;
        traversal_candidates_scored = traversal_candidates_scored
            .checked_add(counters.candidates_scored())
            .ok_or(QueryError::Scan(crate::scan::ScanError::ArithmeticOverflow))?;
        traversal_candidates_rescored = traversal_candidates_rescored
            .checked_add(counters.candidates_rescored())
            .ok_or(QueryError::Scan(crate::scan::ScanError::ArithmeticOverflow))?;

        if !has_tombstones {
            break result;
        }
        let retained_live = result
            .candidates()
            .iter()
            .filter(|candidate| alive.is_alive(candidate.row_id()))
            .count();
        if retained_live >= target_live_k {
            break result;
        }
        if candidate_k >= requestable_max_k {
            return Err(QueryError::Graph(
                crate::graph::search::GraphSearchError::Geometry(format!(
                    "graph traversal exhausted {candidate_k} candidates but retained {retained_live} live rows, expected {target_live_k}"
                )),
            ));
        }
        let widened_k = candidate_k
            .saturating_mul(2)
            .max(candidate_k.saturating_add(1))
            .min(requestable_max_k);
        if widened_k <= candidate_k {
            return Err(QueryError::Graph(
                crate::graph::search::GraphSearchError::Geometry(format!(
                    "graph traversal could not widen beyond {candidate_k} candidates"
                )),
            ));
        }
        candidate_k = widened_k;
        graph_request = request_for(candidate_k);
    };
    let graph_stats = crate::ingest::GraphSearchStats {
        segments_traversed: 1,
        graph_validations: usize::from(graph_validated),
        entry_seed_discoveries: usize::from(entry_seed_discovered),
        visited_epoch_clears: traversal_epoch_clears,
        candidates_scored: traversal_candidates_scored,
        candidates_rescored: traversal_candidates_rescored,
        segments_pruned_by_bound: 0,
    };
    let plan = crate::planner::SegmentPlan::unfiltered_graph(
        source,
        alive.live_count(),
        graph_options.ef(),
        effective_ef,
        graph_options.profile(),
    );
    let caller_thread = std::thread::current().id();
    #[cfg(any(test, feature = "test-seams"))]
    drop(searcher);
    Ok(SegmentGraphResult::Traversed {
        result,
        plan,
        dims_touched: traversal_dims_touched,
        bytes_read: traversal_bytes_read,
        caller_thread,
        graph_stats,
        _scratch: scratch,
        _lease: lease,
    })
}

fn add_traversal_stats(
    stats: &mut crate::ingest::GraphSearchStats,
    delta: crate::ingest::GraphSearchStats,
) -> Result<(), QueryError> {
    stats.segments_traversed = stats
        .segments_traversed
        .checked_add(delta.segments_traversed)
        .ok_or(QueryError::Scan(crate::scan::ScanError::ArithmeticOverflow))?;
    stats.graph_validations = stats
        .graph_validations
        .checked_add(delta.graph_validations)
        .ok_or(QueryError::Scan(crate::scan::ScanError::ArithmeticOverflow))?;
    stats.entry_seed_discoveries = stats
        .entry_seed_discoveries
        .checked_add(delta.entry_seed_discoveries)
        .ok_or(QueryError::Scan(crate::scan::ScanError::ArithmeticOverflow))?;
    stats.visited_epoch_clears = stats
        .visited_epoch_clears
        .checked_add(delta.visited_epoch_clears)
        .ok_or(QueryError::Scan(crate::scan::ScanError::ArithmeticOverflow))?;
    stats.candidates_scored = stats
        .candidates_scored
        .checked_add(delta.candidates_scored)
        .ok_or(QueryError::Scan(crate::scan::ScanError::ArithmeticOverflow))?;
    stats.candidates_rescored = stats
        .candidates_rescored
        .checked_add(delta.candidates_rescored)
        .ok_or(QueryError::Scan(crate::scan::ScanError::ArithmeticOverflow))?;
    stats.segments_pruned_by_bound = stats
        .segments_pruned_by_bound
        .checked_add(delta.segments_pruned_by_bound)
        .ok_or(QueryError::Scan(crate::scan::ScanError::ArithmeticOverflow))?;
    Ok(())
}

#[allow(clippy::too_many_arguments)]
fn scan_sealed_segment(
    pool: Option<&pool::QueryPool>,
    snapshot: &Arc<PublishedSnapshot>,
    segment: &crate::segment::reader::SegmentReader,
    alive: &crate::meta::AliveSet,
    generation: u64,
    request: crate::ingest::SearchRequest<'_>,
    k: usize,
    scan_options: crate::scan::ScanOptions,
    control: &QueryControl,
    bit4_query: &std::cell::OnceCell<Result<crate::quant::Bit4Query, crate::quant::QuantError>>,
    int8_query: &std::cell::OnceCell<Result<crate::quant::Int8Query, crate::quant::QuantError>>,
    full_precision: bool,
    exact_lane: usize,
    accounting: &Arc<stats::Accounting>,
    exact_memory: &mut ExactScanMemory,
    _source: crate::ingest::RowSource,
    _tier: SearchTier,
    #[cfg(any(test, feature = "test-seams"))] vector_fault_controller: Option<
        &crate::scan::vector_fault::VectorFaultController,
    >,
) -> Result<crate::scan::ScanOutcome, QueryError> {
    use crate::scan::{ScanQuery, ScanRequest, ScanRows};

    if full_precision {
        let lease = SnapshotLease::new_at(Arc::clone(snapshot), generation);
        let vectors = exact_rescore_rows_for_search(
            segment,
            #[cfg(any(test, feature = "test-seams"))]
            vector_fault_controller,
            #[cfg(any(test, feature = "test-seams"))]
            vector_fault_tier(_tier),
        )?;
        return scan_squared_l2(
            vectors,
            segment.meta().row_count as usize,
            alive,
            request.vector(),
            k,
            ExactScanDispatch {
                pool,
                accounting,
                control,
                lease: &lease,
                thread_budget: scan_options.thread_budget,
                lane: exact_lane,
            },
            exact_memory,
            #[cfg(any(test, feature = "test-seams"))]
            vector_fault_controller,
            #[cfg(any(test, feature = "test-seams"))]
            vector_fault_source(_source),
            #[cfg(any(test, feature = "test-seams"))]
            vector_fault_tier(_tier),
        );
    }

    let query_pool = pool.ok_or(QueryError::Store(StoreError::Synchronization {
        component: "sealed scan-tier query pool",
    }))?;
    match segment.meta().scheme {
        0 => execute_store_scan(
            query_pool,
            ScanRequest {
                query: ScanQuery::F32(request.vector()),
                rows: ScanRows::F32BorrowedRowMajor(
                    segment
                        .query_f32_codes()
                        .map_err(StoreError::Segment)
                        .map_err(QueryError::Store)?,
                ),
                row_mask: Some(alive.scan_mask()),
            },
            k,
            scan_options,
            control.clone(),
            SnapshotLease::new_at(Arc::clone(snapshot), generation),
            #[cfg(any(test, feature = "test-seams"))]
            vector_fault_controller,
            #[cfg(any(test, feature = "test-seams"))]
            vector_fault_source(_source),
            #[cfg(any(test, feature = "test-seams"))]
            vector_fault_tier(_tier),
        ),
        4 => execute_store_scan(
            query_pool,
            ScanRequest {
                query: ScanQuery::Bit4(
                    bit4_query
                        .get_or_init(|| crate::quant::prepare_bit4_query(request.vector(), 0))
                        .as_ref()
                        .map_err(|error| crate::scan::ScanError::Quant(error.clone()))
                        .map_err(QueryError::Scan)?,
                ),
                rows: ScanRows::Bit4RowMajor {
                    codes: segment
                        .query_bit4_codes()
                        .map_err(StoreError::Segment)
                        .map_err(QueryError::Store)?,
                    factors: segment
                        .query_bit4_factors()
                        .map_err(StoreError::Segment)
                        .map_err(QueryError::Store)?,
                },
                row_mask: Some(alive.scan_mask()),
            },
            k,
            scan_options,
            control.clone(),
            SnapshotLease::new_at(Arc::clone(snapshot), generation),
            #[cfg(any(test, feature = "test-seams"))]
            vector_fault_controller,
            #[cfg(any(test, feature = "test-seams"))]
            vector_fault_source(_source),
            #[cfg(any(test, feature = "test-seams"))]
            vector_fault_tier(_tier),
        ),
        2 => {
            let factors = segment.query_int8_factors().map_err(QueryError::Store)?;
            execute_store_scan(
                query_pool,
                ScanRequest {
                    query: ScanQuery::Int8(
                        int8_query
                            .get_or_init(|| crate::quant::prepare_int8_query(request.vector()))
                            .as_ref()
                            .map_err(|error| crate::scan::ScanError::Quant(error.clone()))
                            .map_err(QueryError::Scan)?,
                    ),
                    rows: ScanRows::Int8RowMajor {
                        codes: segment
                            .query_int8_codes()
                            .map_err(StoreError::Segment)
                            .map_err(QueryError::Store)?,
                        factors: factors.as_slice(),
                    },
                    row_mask: Some(alive.scan_mask()),
                },
                k,
                scan_options,
                control.clone(),
                SnapshotLease::new_at(Arc::clone(snapshot), generation),
                #[cfg(any(test, feature = "test-seams"))]
                vector_fault_controller,
                #[cfg(any(test, feature = "test-seams"))]
                vector_fault_source(_source),
                #[cfg(any(test, feature = "test-seams"))]
                vector_fault_tier(_tier),
            )
        }
        scheme => Err(QueryError::Store(StoreError::Segment(
            crate::segment::SegmentError::Geometry(format!(
                "store search does not support sealed scheme {scheme}"
            )),
        ))),
    }
}

#[allow(clippy::too_many_arguments)]
fn execute_store_scan(
    pool: &pool::QueryPool,
    request: crate::scan::ScanRequest<'_>,
    k: usize,
    options: crate::scan::ScanOptions,
    control: QueryControl,
    lease: SnapshotLease,
    #[cfg(any(test, feature = "test-seams"))] controller: Option<
        &crate::scan::vector_fault::VectorFaultController,
    >,
    #[cfg(any(test, feature = "test-seams"))] source: crate::scan::vector_fault::VectorRowSource,
    #[cfg(any(test, feature = "test-seams"))] tier: crate::scan::vector_fault::VectorSearchTier,
) -> Result<crate::scan::ScanOutcome, QueryError> {
    #[cfg(any(test, feature = "test-seams"))]
    if let Some(controller) = controller {
        let geometry = crate::scan::scan_geometry(request).map_err(QueryError::Scan)?;
        let cancellation = QueryCancellation::new(&control, &lease);
        let partition = crate::scan::scan_partition_with_vector_faults(
            request,
            k,
            0..geometry.row_count,
            Some(&cancellation),
            controller,
            source,
            tier,
        )
        .map_err(map_scan_error)?;
        return Ok(crate::scan::ScanOutcome {
            candidates: partition.candidates,
            worst_score: None,
            stats: crate::scan::ScanStats {
                dims_touched: partition.dims_touched,
                bytes_read: partition.bytes_read,
                threads_used: 1,
                worker_thread_ids: vec![partition.worker_thread_id],
            },
        });
    }
    pool.execute(request, k, options, control, lease)
}

#[cfg(any(test, feature = "test-seams"))]
fn vector_fault_tier(tier: SearchTier) -> crate::scan::vector_fault::VectorSearchTier {
    match tier {
        SearchTier::Exact => crate::scan::vector_fault::VectorSearchTier::Exact,
        SearchTier::Scan => crate::scan::vector_fault::VectorSearchTier::Scan,
        SearchTier::Graph(_) => crate::scan::vector_fault::VectorSearchTier::Graph,
        SearchTier::Auto => crate::scan::vector_fault::VectorSearchTier::Auto,
    }
}

#[cfg(any(test, feature = "test-seams"))]
fn vector_fault_source(
    source: crate::ingest::RowSource,
) -> crate::scan::vector_fault::VectorRowSource {
    match source {
        crate::ingest::RowSource::Active => crate::scan::vector_fault::VectorRowSource::Active,
        crate::ingest::RowSource::Sealed(segment) => {
            crate::scan::vector_fault::VectorRowSource::Sealed(*segment.as_bytes())
        }
    }
}

pub(crate) fn auto_uses_full_precision(
    snapshot: &PublishedSnapshot,
    active: &crate::ingest::ActiveSegment,
) -> bool {
    let mut has_exact = false;
    let mut has_estimated = !active.is_empty();
    for segment in snapshot.segments() {
        if segment
            .directory()
            .iter()
            .any(|entry| entry.kind == crate::segment::layout::RegionKind::GraphNodeBlocks.id())
        {
            return true;
        }
        if segment.meta().row_count == 0 {
            continue;
        }
        match segment.meta().scheme {
            0 => has_exact = true,
            2 | 4 => has_estimated = true,
            _ => {}
        }
    }
    has_exact && has_estimated
}

/// Returns the f32 rescore rows for a query-time exact score.
///
/// This is the query path, so it uses the validate-once accessor. The
/// verifying `rescore_f32` re-hashes the whole rescore region on every
/// call, and hybrid cross-fill calls this once per cross-filled
/// document: 50 calls over a 181 MB region cost about 219 ms of a
/// 221 ms hybrid query on 58,980 FiQA rows, while both legs together
/// cost under 3 ms. Uncached `rescore_f32` remains for the validation
/// and diagnostic callers.
pub(crate) fn exact_rescore_rows(
    segment: &crate::segment::reader::SegmentReader,
) -> Result<&[f32], QueryError> {
    segment
        .query_rescore_f32()
        .map_err(|source| QueryError::Store(StoreError::Segment(source)))
}

#[cfg(not(any(test, feature = "test-seams")))]
use exact_rescore_rows as exact_rescore_rows_for_search;

#[cfg(any(test, feature = "test-seams"))]
fn exact_rescore_rows_for_search<'a>(
    segment: &'a crate::segment::reader::SegmentReader,
    #[cfg(any(test, feature = "test-seams"))] controller: Option<
        &crate::scan::vector_fault::VectorFaultController,
    >,
    #[cfg(any(test, feature = "test-seams"))] tier: crate::scan::vector_fault::VectorSearchTier,
) -> Result<&'a [f32], QueryError> {
    #[cfg(any(test, feature = "test-seams"))]
    if let Some(available_rows) = controller.and_then(|controller| {
        controller.missing_rescore_rows(
            crate::scan::vector_fault::VectorRowSource::Sealed(*segment.meta().id.as_bytes()),
            tier,
            crate::scan::vector_fault::MissingRescoreSite::ExactRescoreRows,
            segment.meta().row_count,
        )
    }) {
        return Err(missing_rescore_rows_error(segment, available_rows));
    }
    exact_rescore_rows(segment)
}

pub(crate) fn query_rescore_rows(
    segment: &crate::segment::reader::SegmentReader,
) -> Result<&[f32], QueryError> {
    segment
        .query_rescore_f32()
        .map_err(|source| QueryError::Store(StoreError::Segment(source)))
}

#[cfg(not(any(test, feature = "test-seams")))]
use query_rescore_rows as query_rescore_rows_for_search;

#[cfg(any(test, feature = "test-seams"))]
fn query_rescore_rows_for_search<'a>(
    segment: &'a crate::segment::reader::SegmentReader,
    #[cfg(any(test, feature = "test-seams"))] controller: Option<
        &crate::scan::vector_fault::VectorFaultController,
    >,
    #[cfg(any(test, feature = "test-seams"))] tier: crate::scan::vector_fault::VectorSearchTier,
) -> Result<&'a [f32], QueryError> {
    #[cfg(any(test, feature = "test-seams"))]
    if let Some(available_rows) = controller.and_then(|controller| {
        controller.missing_rescore_rows(
            crate::scan::vector_fault::VectorRowSource::Sealed(*segment.meta().id.as_bytes()),
            tier,
            crate::scan::vector_fault::MissingRescoreSite::QueryRescoreRows,
            segment.meta().row_count,
        )
    }) {
        return Err(missing_rescore_rows_error(segment, available_rows));
    }
    query_rescore_rows(segment)
}

#[cfg(any(test, feature = "test-seams"))]
fn missing_rescore_rows_error(
    segment: &crate::segment::reader::SegmentReader,
    available_rows: u32,
) -> QueryError {
    QueryError::Store(StoreError::Segment(crate::segment::SegmentError::Geometry(
        format!(
            "exact scores unavailable for segment {}: expected {} rows, got {available_rows}",
            segment.meta().id,
            segment.meta().row_count,
        ),
    )))
}

fn retain_global_top_k(candidates: &mut Vec<crate::ingest::SearchCandidate>, k: usize) {
    candidates.sort_unstable_by(crate::ingest::compare_search_candidates);
    candidates.truncate(k);
}

fn global_competitive_distance(
    candidates: &[crate::ingest::SearchCandidate],
    k: usize,
) -> Option<f32> {
    if k == 0 || candidates.len() < k {
        return None;
    }
    candidates
        .get(k.saturating_sub(1))
        .map(|candidate| -candidate.score())
        .filter(|distance| distance.is_finite() && *distance >= 0.0)
}

fn map_graph_cache_error(error: graph_cache::GraphCacheError) -> QueryError {
    match error {
        graph_cache::GraphCacheError::Store(error) => QueryError::Store(error),
        graph_cache::GraphCacheError::Search(error) => map_graph_error(error),
    }
}

pub(crate) fn map_graph_error(error: crate::graph::search::GraphSearchError) -> QueryError {
    match error {
        crate::graph::search::GraphSearchError::ExactRescoreUnavailable(detail) => {
            QueryError::Store(StoreError::Segment(crate::segment::SegmentError::Geometry(
                detail,
            )))
        }
        crate::graph::search::GraphSearchError::Cancelled { partial } => {
            QueryError::Cancelled { partial }
        }
        crate::graph::search::GraphSearchError::Timeout { partial } => {
            QueryError::Timeout { partial }
        }
        crate::graph::search::GraphSearchError::ReadCancelled { partial } => {
            QueryError::ReadCancelled { partial }
        }
        crate::graph::search::GraphSearchError::Scan(error) => QueryError::Scan(error),
        error => QueryError::Graph(error),
    }
}

#[allow(
    clippy::too_many_arguments,
    reason = "scoped dispatch and test source identity stay explicit"
)]
fn scan_active_squared_l2(
    active: &crate::ingest::ActiveSegment,
    alive: &crate::meta::AliveSet,
    query: &[f32],
    k: usize,
    dispatch: ExactScanDispatch<'_>,
    memory: &mut ExactScanMemory,
    #[cfg(any(test, feature = "test-seams"))] controller: Option<
        &crate::scan::vector_fault::VectorFaultController,
    >,
    #[cfg(any(test, feature = "test-seams"))] tier: crate::scan::vector_fault::VectorSearchTier,
) -> Result<crate::scan::ScanOutcome, QueryError> {
    scan_squared_l2(
        active.vectors(),
        active.row_count(),
        alive,
        query,
        k,
        dispatch,
        memory,
        #[cfg(any(test, feature = "test-seams"))]
        controller,
        #[cfg(any(test, feature = "test-seams"))]
        crate::scan::vector_fault::VectorRowSource::Active,
        #[cfg(any(test, feature = "test-seams"))]
        tier,
    )
}

#[derive(Debug)]
enum ExactScanError {
    Query(QueryError),
    Narrowing { row_id: usize, score: f64 },
}

impl From<QueryError> for ExactScanError {
    fn from(error: QueryError) -> Self {
        Self::Query(error)
    }
}

impl ExactScanError {
    fn into_query(self) -> QueryError {
        match self {
            Self::Query(error) => error,
            Self::Narrowing { row_id, .. } => {
                QueryError::Scan(crate::scan::ScanError::NonFiniteScore { row_id })
            }
        }
    }
}

pub(crate) struct ExactScanMemory {
    candidates: stats::AccountedCounter,
    worker_ids: stats::AccountedCounter,
}

impl ExactScanMemory {
    pub(crate) fn new(accounting: &Arc<stats::Accounting>) -> Result<Self, QueryError> {
        Ok(Self {
            candidates: stats::AccountedCounter::new(
                accounting,
                stats::AllocationComponent::Temporary,
            )
            .map_err(QueryError::Store)?,
            worker_ids: stats::AccountedCounter::new(
                accounting,
                stats::AllocationComponent::Temporary,
            )
            .map_err(QueryError::Store)?,
        })
    }
}

struct ExactScanDispatch<'a> {
    pool: Option<&'a pool::QueryPool>,
    accounting: &'a Arc<stats::Accounting>,
    control: &'a QueryControl,
    lease: &'a SnapshotLease,
    thread_budget: usize,
    lane: usize,
}

#[allow(
    clippy::too_many_arguments,
    reason = "scoped dispatch and test source identity stay explicit"
)]
fn scan_squared_l2(
    vectors: &[f32],
    row_count: usize,
    alive: &crate::meta::AliveSet,
    query: &[f32],
    k: usize,
    dispatch: ExactScanDispatch<'_>,
    memory: &mut ExactScanMemory,
    #[cfg(any(test, feature = "test-seams"))] controller: Option<
        &crate::scan::vector_fault::VectorFaultController,
    >,
    #[cfg(any(test, feature = "test-seams"))] source: crate::scan::vector_fault::VectorRowSource,
    #[cfg(any(test, feature = "test-seams"))] tier: crate::scan::vector_fault::VectorSearchTier,
) -> Result<crate::scan::ScanOutcome, QueryError> {
    if let Some(pool) = dispatch.pool
        && pool.exact_workers(row_count, query.len(), dispatch.thread_budget) > 1
    {
        return pool.execute_exact(
            vectors,
            row_count,
            alive,
            query,
            k,
            dispatch.thread_budget,
            dispatch.lane,
            dispatch.control,
            dispatch.lease,
            dispatch.accounting,
            memory,
            #[cfg(any(test, feature = "test-seams"))]
            controller,
            #[cfg(any(test, feature = "test-seams"))]
            source,
            #[cfg(any(test, feature = "test-seams"))]
            tier,
        );
    }
    let cancellation = QueryCancellation::new(dispatch.control, dispatch.lease);
    #[cfg(any(test, feature = "test-seams"))]
    if controller.is_some_and(|controller| controller.has_exact_row_hooks()) {
        return scan_exact_partition::<false>(
            vectors,
            row_count,
            0..row_count,
            alive,
            query,
            k,
            &cancellation,
            memory,
            #[cfg(any(test, feature = "test-seams"))]
            controller,
            #[cfg(any(test, feature = "test-seams"))]
            source,
            #[cfg(any(test, feature = "test-seams"))]
            tier,
        )
        .map_err(ExactScanError::into_query);
    }
    scan_serial_squared_l2(
        vectors,
        row_count,
        alive,
        query,
        k,
        &cancellation,
        &mut memory.candidates,
        &mut memory.worker_ids,
        #[cfg(any(test, feature = "test-seams"))]
        controller,
        #[cfg(any(test, feature = "test-seams"))]
        source,
        #[cfg(any(test, feature = "test-seams"))]
        tier,
    )
}

#[allow(clippy::too_many_arguments)]
fn scan_exact_partition<const OFFSET_ROWS: bool>(
    vectors: &[f32],
    row_count: usize,
    range: std::ops::Range<usize>,
    alive: &crate::meta::AliveSet,
    query: &[f32],
    k: usize,
    cancellation: &QueryCancellation<'_>,
    memory: &mut ExactScanMemory,
    #[cfg(any(test, feature = "test-seams"))] controller: Option<
        &crate::scan::vector_fault::VectorFaultController,
    >,
    #[cfg(any(test, feature = "test-seams"))] source: crate::scan::vector_fault::VectorRowSource,
    #[cfg(any(test, feature = "test-seams"))] tier: crate::scan::vector_fault::VectorSearchTier,
) -> Result<crate::scan::ScanOutcome, ExactScanError> {
    if query.is_empty() {
        return Err(QueryError::Scan(crate::scan::ScanError::ZeroDimension).into());
    }
    let expected = row_count
        .checked_mul(query.len())
        .ok_or(QueryError::Scan(crate::scan::ScanError::ArithmeticOverflow))?;
    if vectors.len() != expected {
        return Err(QueryError::Scan(crate::scan::ScanError::RowDataLength {
            dimension: query.len(),
            actual: vectors.len(),
        })
        .into());
    }
    // Preserve the original physical-row cancellation preflight without storing
    // the rows. Scoring checks subsequently count eligible rows, as before.
    if range.start > range.end || range.end > row_count || (!OFFSET_ROWS && range.start != 0) {
        return Err(QueryError::Scan(crate::scan::ScanError::ArithmeticOverflow).into());
    }
    let mut live_rows = 0;
    for row in range.clone() {
        if row.is_multiple_of(64) {
            cancellation.check_graph().map_err(map_scan_error)?;
        }
        let local_row = u32::try_from(row)
            .map_err(|_| QueryError::Scan(crate::scan::ScanError::ArithmeticOverflow))?;
        live_rows += usize::from(alive.is_alive(local_row));
    }
    cancellation.check_graph().map_err(map_scan_error)?;
    let mut collector = crate::scan::topk::ExactTopK::try_new(k, live_rows, &mut memory.candidates)
        .map_err(QueryError::Store)?;
    let mut worst_score: Option<f32> = None;
    let mut narrowing_error: Option<(usize, f64)> = None;
    let start = range
        .start
        .checked_mul(query.len())
        .ok_or(QueryError::Scan(crate::scan::ScanError::ArithmeticOverflow))?;
    let end = range
        .end
        .checked_mul(query.len())
        .ok_or(QueryError::Scan(crate::scan::ScanError::ArithmeticOverflow))?;
    let partition_vectors = vectors
        .get(start..end)
        .ok_or(QueryError::Scan(crate::scan::ScanError::ArithmeticOverflow))?;
    let rows = partition_vectors
        .chunks_exact(query.len())
        .enumerate()
        .map(|(local, vector)| {
            (
                if OFFSET_ROWS {
                    range.start + local
                } else {
                    local
                },
                vector,
            )
        })
        .filter(|(row, _)| u32::try_from(*row).is_ok_and(|row| alive.is_alive(row)));
    #[cfg(any(test, feature = "test-seams"))]
    let mut trace =
        controller.and_then(|controller| controller.trace_rows(source, tier, range.clone()));
    let scored_rows = crate::quant::exact_squared_l2_with_sink(
        query,
        rows,
        |_row, is_checkpoint| {
            #[cfg(any(test, feature = "test-seams"))]
            if let Some(controller) = controller {
                if let Some(trace) = &mut trace {
                    trace.checked(_row);
                }
                if is_checkpoint {
                    cancellation.check_graph().map_err(map_scan_error)?;
                }
                if controller.after_eligible_row(source, tier, _row) {
                    return Err(QueryError::Cancelled { partial: false });
                }
                return Ok(());
            }
            if is_checkpoint {
                cancellation.check_graph().map_err(map_scan_error)?;
            }
            Ok(())
        },
        |row_id, exact| {
            let score = exact as f32;
            if !score.is_finite() {
                // The old full f64 sort chose the best overflowing score, then
                // physical row ID. Defer this failure until every row check and
                // f64 rejection has run, including the final cancellation check.
                if narrowing_error.is_none_or(|(row, previous)| {
                    exact.total_cmp(&previous).is_gt()
                        || (exact.total_cmp(&previous).is_eq() && row_id < row)
                }) {
                    narrowing_error = Some((row_id, exact));
                }
                return Ok(());
            }
            if worst_score.is_none_or(|worst| score.total_cmp(&worst).is_lt()) {
                worst_score = Some(score);
            }
            collector
                .try_push(
                    crate::scan::ScanCandidate { row_id, score },
                    &mut memory.candidates,
                )
                .map_err(QueryError::Store)
        },
    )
    .map_err(|error| match error {
        crate::quant::RescoreCheckError::Rescore(error) => map_l2_rescore_error(error),
        crate::quant::RescoreCheckError::Check(error) => error,
    })?;
    cancellation.check_graph().map_err(map_scan_error)?;
    if let Some((row_id, score)) = narrowing_error {
        return Err(ExactScanError::Narrowing { row_id, score });
    }
    let (candidates, _collector_capacity) = collector
        .try_into_sorted_with_ties(&mut memory.candidates)
        .map_err(QueryError::Store)?;
    #[cfg(any(test, feature = "test-seams"))]
    if let Some(controller) = controller {
        controller.record_exact_scan(crate::scan::vector_fault::ExactScanWork {
            sorted_items: candidates.len(),
            collector_capacity: _collector_capacity,
            scored_rows,
            worst_score,
            ..Default::default()
        });
    }
    let dims = u64::try_from(query.len())
        .map_err(|_| QueryError::Scan(crate::scan::ScanError::ArithmeticOverflow))?;
    let scored_rows = u64::try_from(scored_rows)
        .map_err(|_| QueryError::Scan(crate::scan::ScanError::ArithmeticOverflow))?;
    let dims_touched = scored_rows
        .checked_mul(dims)
        .ok_or(QueryError::Scan(crate::scan::ScanError::ArithmeticOverflow))?;
    let bytes_read = dims_touched
        .checked_mul(std::mem::size_of::<f32>() as u64)
        .ok_or(QueryError::Scan(crate::scan::ScanError::ArithmeticOverflow))?;
    memory
        .worker_ids
        .set(std::mem::size_of::<std::thread::ThreadId>())
        .map_err(QueryError::Store)?;
    let mut worker_thread_ids = Vec::new();
    worker_thread_ids.try_reserve_exact(1).map_err(|_| {
        QueryError::Store(StoreError::AllocationFailed {
            needed: memory.worker_ids.bytes(),
            component: "exact worker IDs",
        })
    })?;
    worker_thread_ids.push(std::thread::current().id());
    Ok(crate::scan::ScanOutcome {
        candidates,
        worst_score,
        stats: crate::scan::ScanStats {
            dims_touched,
            bytes_read,
            threads_used: 1,
            worker_thread_ids,
        },
    })
}

// Retain the plan-04 serial iterator and QueryError path. Partition offsets
// and narrowing-error aggregation belong only to worker scans.
#[allow(clippy::too_many_arguments)]
fn scan_serial_squared_l2(
    vectors: &[f32],
    row_count: usize,
    alive: &crate::meta::AliveSet,
    query: &[f32],
    k: usize,
    cancellation: &QueryCancellation<'_>,
    memory: &mut stats::AccountedCounter,
    worker_memory: &mut stats::AccountedCounter,
    #[cfg(any(test, feature = "test-seams"))] controller: Option<
        &crate::scan::vector_fault::VectorFaultController,
    >,
    #[cfg(any(test, feature = "test-seams"))] source: crate::scan::vector_fault::VectorRowSource,
    #[cfg(any(test, feature = "test-seams"))] tier: crate::scan::vector_fault::VectorSearchTier,
) -> Result<crate::scan::ScanOutcome, QueryError> {
    if query.is_empty() {
        return Err(QueryError::Scan(crate::scan::ScanError::ZeroDimension));
    }
    let expected = row_count
        .checked_mul(query.len())
        .ok_or(QueryError::Scan(crate::scan::ScanError::ArithmeticOverflow))?;
    if vectors.len() != expected {
        return Err(QueryError::Scan(crate::scan::ScanError::RowDataLength {
            dimension: query.len(),
            actual: vectors.len(),
        }));
    }
    // Preserve the original physical-row cancellation preflight without storing
    // the rows. Scoring checks subsequently count eligible rows, as before.
    let mut live_rows = 0;
    for row in 0..row_count {
        if row.is_multiple_of(64) {
            cancellation.check_graph().map_err(map_scan_error)?;
        }
        let local_row = u32::try_from(row)
            .map_err(|_| QueryError::Scan(crate::scan::ScanError::ArithmeticOverflow))?;
        live_rows += usize::from(alive.is_alive(local_row));
    }
    cancellation.check_graph().map_err(map_scan_error)?;
    let mut collector =
        crate::scan::topk::ExactTopK::try_new(k, live_rows, memory).map_err(QueryError::Store)?;
    let mut worst_score: Option<f32> = None;
    let mut narrowing_error: Option<(usize, f64)> = None;
    let rows = vectors
        .chunks_exact(query.len())
        .enumerate()
        .filter(|(row, _)| u32::try_from(*row).is_ok_and(|row| alive.is_alive(row)));
    let scored_rows = crate::quant::exact_squared_l2_with_sink(
        query,
        rows,
        |_row, is_checkpoint| {
            if is_checkpoint {
                cancellation.check_graph().map_err(map_scan_error)?;
            }
            #[cfg(any(test, feature = "test-seams"))]
            if controller
                .is_some_and(|controller| controller.after_eligible_row(source, tier, _row))
            {
                return Err(QueryError::Cancelled { partial: false });
            }
            Ok(())
        },
        |row_id, exact| {
            let score = exact as f32;
            if !score.is_finite() {
                // The old full f64 sort chose the best overflowing score, then
                // physical row ID. Defer this failure until every row check and
                // f64 rejection has run, including the final cancellation check.
                if narrowing_error.is_none_or(|(row, previous)| {
                    exact.total_cmp(&previous).is_gt()
                        || (exact.total_cmp(&previous).is_eq() && row_id < row)
                }) {
                    narrowing_error = Some((row_id, exact));
                }
                return Ok(());
            }
            if worst_score.is_none_or(|worst| score.total_cmp(&worst).is_lt()) {
                worst_score = Some(score);
            }
            collector
                .try_push(crate::scan::ScanCandidate { row_id, score }, memory)
                .map_err(QueryError::Store)
        },
    )
    .map_err(|error| match error {
        crate::quant::RescoreCheckError::Rescore(error) => map_l2_rescore_error(error),
        crate::quant::RescoreCheckError::Check(error) => error,
    })?;
    cancellation.check_graph().map_err(map_scan_error)?;
    if let Some((row_id, _)) = narrowing_error {
        return Err(QueryError::Scan(crate::scan::ScanError::NonFiniteScore {
            row_id,
        }));
    }
    let (candidates, _collector_capacity) = collector
        .try_into_sorted_with_ties(memory)
        .map_err(QueryError::Store)?;
    #[cfg(any(test, feature = "test-seams"))]
    if let Some(controller) = controller {
        controller.record_exact_scan(crate::scan::vector_fault::ExactScanWork {
            sorted_items: candidates.len(),
            collector_capacity: _collector_capacity,
            scored_rows,
            worst_score,
            ..Default::default()
        });
    }
    let dims = u64::try_from(query.len())
        .map_err(|_| QueryError::Scan(crate::scan::ScanError::ArithmeticOverflow))?;
    let scored_rows = u64::try_from(scored_rows)
        .map_err(|_| QueryError::Scan(crate::scan::ScanError::ArithmeticOverflow))?;
    let dims_touched = scored_rows
        .checked_mul(dims)
        .ok_or(QueryError::Scan(crate::scan::ScanError::ArithmeticOverflow))?;
    let bytes_read = dims_touched
        .checked_mul(std::mem::size_of::<f32>() as u64)
        .ok_or(QueryError::Scan(crate::scan::ScanError::ArithmeticOverflow))?;
    worker_memory
        .set(std::mem::size_of::<std::thread::ThreadId>())
        .map_err(QueryError::Store)?;
    let mut worker_thread_ids = Vec::new();
    worker_thread_ids.try_reserve_exact(1).map_err(|_| {
        QueryError::Store(StoreError::AllocationFailed {
            needed: worker_memory.bytes(),
            component: "exact worker IDs",
        })
    })?;
    worker_thread_ids.push(std::thread::current().id());
    Ok(crate::scan::ScanOutcome {
        candidates,
        worst_score,
        stats: crate::scan::ScanStats {
            dims_touched,
            bytes_read,
            threads_used: 1,
            worker_thread_ids,
        },
    })
}

fn map_l2_rescore_error(error: crate::quant::RescoreError) -> QueryError {
    let error = match error {
        crate::quant::RescoreError::ZeroDimension => crate::scan::ScanError::ZeroDimension,
        crate::quant::RescoreError::QueryDimension { expected, actual }
        | crate::quant::RescoreError::CoarseScoreCount { expected, actual }
        | crate::quant::RescoreError::CandidateRowCount { expected, actual } => {
            crate::scan::ScanError::RowDataLength {
                dimension: expected,
                actual,
            }
        }
        crate::quant::RescoreError::RowDataLength { dimension, actual } => {
            crate::scan::ScanError::RowDataLength { dimension, actual }
        }
        crate::quant::RescoreError::CandidateRowOutOfRange {
            row_index,
            row_count,
            ..
        } => crate::scan::ScanError::RowDataLength {
            dimension: row_count,
            actual: row_index,
        },
        crate::quant::RescoreError::NonFiniteCoarseScore { index }
        | crate::quant::RescoreError::NonFiniteExactScore { row_index: index } => {
            crate::scan::ScanError::NonFiniteScore { row_id: index }
        }
        crate::quant::RescoreError::ZeroK
        | crate::quant::RescoreError::ZeroOversample
        | crate::quant::RescoreError::InsufficientCandidates { .. }
        | crate::quant::RescoreError::ArithmeticOverflow => {
            crate::scan::ScanError::ArithmeticOverflow
        }
        error @ (crate::quant::RescoreError::ZeroCandidateLimit
        | crate::quant::RescoreError::CandidateLimitExceeded { .. }) => {
            crate::scan::ScanError::Rescore(error)
        }
    };
    QueryError::Scan(error)
}

fn map_scan_error(error: crate::scan::ScanError) -> QueryError {
    match error {
        crate::scan::ScanError::Cancelled { partial } => QueryError::Cancelled { partial },
        crate::scan::ScanError::Timeout { partial } => QueryError::Timeout { partial },
        crate::scan::ScanError::ReadCancelled { partial } => QueryError::ReadCancelled { partial },
        error => QueryError::Scan(error),
    }
}

#[allow(clippy::too_many_arguments)]
fn merge_store_outcome(
    outcome: crate::scan::ScanOutcome,
    source: crate::ingest::RowSource,
    document: impl Fn(usize) -> Result<Option<crate::ingest::DocumentVersion>, QueryError>,
    candidates: &mut Vec<crate::ingest::SearchCandidate>,
    dims_touched: &mut u64,
    bytes_read: &mut u64,
    worker_thread_ids: &mut Vec<std::thread::ThreadId>,
    exact_score: bool,
    #[cfg(any(test, feature = "test-seams"))] vector_fault_controller: Option<
        &crate::scan::vector_fault::VectorFaultController,
    >,
    #[cfg(any(test, feature = "test-seams"))]
    vector_fault_tier: crate::scan::vector_fault::VectorSearchTier,
) -> Result<(), QueryError> {
    *dims_touched = dims_touched
        .checked_add(outcome.stats.dims_touched)
        .ok_or(QueryError::Scan(crate::scan::ScanError::ArithmeticOverflow))?;
    *bytes_read = bytes_read
        .checked_add(outcome.stats.bytes_read)
        .ok_or(QueryError::Scan(crate::scan::ScanError::ArithmeticOverflow))?;
    for worker in outcome.stats.worker_thread_ids {
        if !worker_thread_ids.contains(&worker) {
            worker_thread_ids.push(worker);
        }
    }
    reserve_global_candidates(
        candidates,
        outcome.candidates.len(),
        #[cfg(any(test, feature = "test-seams"))]
        vector_fault_controller,
        #[cfg(any(test, feature = "test-seams"))]
        vector_fault_tier,
    )?;
    for candidate in outcome.candidates {
        let local_row = u32::try_from(candidate.row_id)
            .map_err(|_| QueryError::Scan(crate::scan::ScanError::ArithmeticOverflow))?;
        candidates.push(crate::ingest::SearchCandidate::new(
            crate::ingest::GlobalRowId::new(source, local_row),
            document(candidate.row_id)?,
            candidate.score,
            exact_score,
        ));
    }
    Ok(())
}

pub(crate) fn reserve_global_candidates(
    candidates: &mut Vec<crate::ingest::SearchCandidate>,
    additional: usize,
    #[cfg(any(test, feature = "test-seams"))] vector_fault_controller: Option<
        &crate::scan::vector_fault::VectorFaultController,
    >,
    #[cfg(any(test, feature = "test-seams"))]
    vector_fault_tier: crate::scan::vector_fault::VectorSearchTier,
) -> Result<(), QueryError> {
    if additional == 0 {
        return Ok(());
    }
    let items = u64::try_from(additional)
        .map_err(|_| QueryError::Scan(crate::scan::ScanError::ArithmeticOverflow))?;
    let candidate_bytes = u64::try_from(std::mem::size_of::<crate::ingest::SearchCandidate>())
        .map_err(|_| QueryError::Scan(crate::scan::ScanError::ArithmeticOverflow))?;
    let needed = items
        .checked_mul(candidate_bytes)
        .ok_or(QueryError::Scan(crate::scan::ScanError::ArithmeticOverflow))?;
    #[cfg(any(test, feature = "test-seams"))]
    if vector_fault_controller.is_some_and(|controller| {
        controller.deny_global_candidate_allocation(vector_fault_tier, items, needed)
    }) {
        return Err(QueryError::Store(StoreError::AllocationFailed {
            needed,
            component: "vector search global candidates",
        }));
    }
    candidates.try_reserve_exact(additional).map_err(|_| {
        QueryError::Store(StoreError::AllocationFailed {
            needed,
            component: "vector search global candidates",
        })
    })
}

struct AdmittedLexicalQuery<'a> {
    generation: u64,
    active: Arc<crate::ingest::ActiveSegment>,
    snapshot: snapshot::ReadSnapshot,
    active_query: ActiveQuery<'a>,
}

struct AdmittedDocumentRead<'a> {
    generation: u64,
    active: Arc<crate::ingest::ActiveSegment>,
    snapshot: snapshot::ReadSnapshot,
    active_query: ActiveQuery<'a>,
}

struct AdmittedVectorSearch<'a> {
    execution_options: SearchOptions,
    pool: Option<Arc<pool::QueryPool>>,
    snapshot: snapshot::ReadSnapshot,
    active_segment: Arc<crate::ingest::ActiveSegment>,
    generation: u64,
    _active_query: Option<ActiveQuery<'a>>,
}

struct ActiveQuery<'a> {
    count: &'a AtomicU64,
}

impl Drop for ActiveQuery<'_> {
    fn drop(&mut self) {
        self.count.fetch_sub(1, Ordering::Relaxed);
    }
}

impl Drop for Store {
    fn drop(&mut self) {
        self.close_best_effort();
    }
}

#[derive(Clone, Copy, Debug, Eq, PartialEq)]
struct DocumentScanRow {
    source: DocumentScanSource,
    row: u32,
    timestamp: i64,
    doc_id: crate::ingest::DocId,
    /// Order-preserving key of the [`ScanOrder::Attribute`] value, or `None`
    /// when the value is missing or NaN or the order is not by attribute.
    attribute_key: Option<u64>,
}

fn validate_scan_order_attribute(
    schema: &crate::meta::Schema,
    column: crate::meta::ColumnId,
) -> Result<(), QueryError> {
    use crate::meta::ColumnType;

    let id = column.get();
    let detail = if column == crate::meta::TIMESTAMP_COLUMN {
        Some(format!(
            "scan order attribute {id} is the document timestamp; use a timestamp order"
        ))
    } else {
        match schema
            .column(column)
            .map(crate::meta::ColumnDefinition::column_type)
        {
            None => Some(format!(
                "scan order attribute {id} is not a declared attribute"
            )),
            Some(ColumnType::U64 | ColumnType::I64 | ColumnType::F64) => None,
            Some(other) => Some(format!(
                "scan order attribute {id} has type {other:?}; only u64, i64 and f64 attributes are orderable"
            )),
        }
    };
    detail.map_or(Ok(()), |detail| {
        Err(QueryError::Store(StoreError::InvalidScan { detail }))
    })
}

/// Maps one attribute value to a `u64` whose unsigned order is the value
/// order: `U64` unchanged, `I64` with the sign bit flipped, and `F64` by the
/// IEEE-754 total-order bit transform after folding `-0.0` into `+0.0`.
/// Missing and NaN values have no key.
fn attribute_order_key(column: &crate::meta::Column, row: u32) -> Result<Option<u64>, QueryError> {
    use crate::meta::Column;
    const SIGN: u64 = 1 << 63;

    match column {
        Column::U64(values) => Ok(values.get(row)),
        Column::I64(values) => Ok(values.get(row).map(|value| value.cast_unsigned() ^ SIGN)),
        Column::F64(values) => Ok(values
            .get(row)
            .filter(|value| !value.is_nan())
            .map(|value| {
                // Adding +0.0 folds -0.0 into +0.0 and leaves every other
                // non-NaN value unchanged.
                let bits = (value + 0.0).to_bits();
                if bits & SIGN == 0 { bits | SIGN } else { !bits }
            })),
        Column::Id128(_) | Column::Bool(_) | Column::DictionaryString(_) | Column::RawString(_) => {
            Err(QueryError::Store(StoreError::InvalidScan {
                detail: "scan order attribute column is not numeric".to_owned(),
            }))
        }
    }
}

fn keyed_scan_row(
    mut row: DocumentScanRow,
    columns: Option<(&crate::meta::ColumnStore, crate::meta::ColumnId)>,
) -> Result<DocumentScanRow, QueryError> {
    if let Some((columns, column)) = columns {
        let values = columns.column(column).ok_or_else(|| {
            QueryError::Store(StoreError::InvalidScan {
                detail: format!("scan order attribute {} has no stored column", column.get()),
            })
        })?;
        row.attribute_key = attribute_order_key(values, row.row)?;
    }
    Ok(row)
}

fn document_scan_predicate(
    predicate: Option<&crate::meta::Predicate>,
    timestamp_range: Option<(i64, i64)>,
) -> Result<Option<crate::meta::Predicate>, QueryError> {
    use crate::meta::{Predicate, PredicateValue, RangeBound, RangePredicate, TIMESTAMP_COLUMN};

    let timestamp = timestamp_range
        .map(|(start, end)| {
            if start > end {
                return Err(QueryError::Store(StoreError::InvalidScan {
                    detail: format!("timestamp range start {start} exceeds end {end}"),
                }));
            }
            Ok(Predicate::Range(RangePredicate {
                column: TIMESTAMP_COLUMN,
                lower: Some(RangeBound::inclusive(PredicateValue::I64(start))),
                upper: Some(RangeBound::exclusive(PredicateValue::I64(end))),
            }))
        })
        .transpose()?;
    Ok(match (predicate, timestamp) {
        (Some(predicate), Some(timestamp)) => {
            Some(Predicate::And(vec![predicate.clone(), timestamp]))
        }
        (Some(predicate), None) => Some(predicate.clone()),
        (None, Some(timestamp)) => Some(timestamp),
        (None, None) => None,
    })
}

fn scan_control_checkpoint(
    control: &QueryControl,
    lease: &SnapshotLease,
    work: &mut usize,
) -> Result<(), QueryError> {
    *work = work.saturating_add(1);
    if (*work).is_multiple_of(64) {
        control.checkpoint()?;
        lease.check_active().map_err(QueryError::Store)?;
    }
    Ok(())
}

fn active_scan_columns(
    active: &crate::ingest::ActiveSegment,
    schema: &crate::meta::Schema,
) -> Result<crate::meta::ColumnStore, QueryError> {
    let mut builder = crate::meta::ColumnStoreBuilder::new(schema.clone());
    for (row, timestamp) in active.timestamps().iter().copied().enumerate() {
        let values = active.column_values(row).map_err(QueryError::Store)?;
        builder
            .push_row(timestamp, &crate::ingest::column_inputs(&values))
            .map_err(|error| {
                QueryError::Store(StoreError::Segment(crate::segment::SegmentError::Columns(
                    error.to_string(),
                )))
            })?;
    }
    builder.finish().map_err(|error| {
        QueryError::Store(StoreError::Segment(crate::segment::SegmentError::Columns(
            error.to_string(),
        )))
    })
}

fn sealed_scan_rows(
    segment: &crate::segment::reader::SegmentReader,
    predicate: Option<&crate::meta::Predicate>,
) -> Result<crate::meta::DocBitmap, QueryError> {
    if predicate.is_some_and(|predicate| {
        !crate::planner::segment_may_match(segment.meta().clustering_key_range, predicate)
    }) {
        return Ok(crate::meta::DocBitmap::new());
    }
    let alive = segment.query_alive().map_err(QueryError::Store)?;
    match predicate {
        Some(predicate) => {
            let columns = segment.query_columns().map_err(QueryError::Store)?;
            crate::meta::evaluate(predicate, &columns, &alive).map_err(|error| {
                QueryError::Store(StoreError::Segment(crate::segment::SegmentError::Columns(
                    error.to_string(),
                )))
            })
        }
        None => Ok(alive.alive_bitmap().clone()),
    }
}

fn active_scan_rows(
    active: &crate::ingest::ActiveSegment,
    schema: &crate::meta::Schema,
    predicate: Option<&crate::meta::Predicate>,
) -> Result<crate::meta::DocBitmap, QueryError> {
    let alive = active.alive().map_err(QueryError::Store)?;
    match predicate {
        Some(predicate) => {
            let columns = active_scan_columns(active, schema)?;
            crate::meta::evaluate(predicate, &columns, &alive).map_err(|error| {
                QueryError::Store(StoreError::Segment(crate::segment::SegmentError::Columns(
                    error.to_string(),
                )))
            })
        }
        None => Ok(alive.alive_bitmap().clone()),
    }
}

fn sealed_document_scan_row(
    segment: &crate::segment::reader::SegmentReader,
    row: u32,
) -> Result<DocumentScanRow, QueryError> {
    let version = segment
        .document_version(row as usize)
        .map_err(StoreError::Segment)
        .map_err(QueryError::Store)?
        .ok_or_else(|| {
            QueryError::Store(StoreError::Segment(crate::segment::SegmentError::Geometry(
                format!("sealed document row {row} has no identity"),
            )))
        })?;
    let columns = segment.query_columns().map_err(QueryError::Store)?;
    let timestamp = columns.timestamp(row).ok_or_else(|| {
        QueryError::Store(StoreError::Segment(crate::segment::SegmentError::Geometry(
            format!("sealed document row {row} has no timestamp"),
        )))
    })?;
    Ok(DocumentScanRow {
        source: DocumentScanSource::Sealed(segment.meta().id),
        row,
        timestamp,
        doc_id: version.doc_id(),
        attribute_key: None,
    })
}

fn active_document_scan_row(
    active: &crate::ingest::ActiveSegment,
    row: u32,
) -> Result<DocumentScanRow, QueryError> {
    let index = row as usize;
    let version = active
        .document(index)
        .ok_or(QueryError::Store(StoreError::ActiveRowOverflow))?;
    let timestamp = active
        .timestamps()
        .get(index)
        .copied()
        .ok_or(QueryError::Store(StoreError::ActiveRowOverflow))?;
    Ok(DocumentScanRow {
        source: DocumentScanSource::Active,
        row,
        timestamp,
        doc_id: version.doc_id(),
        attribute_key: None,
    })
}

#[allow(clippy::too_many_arguments)]
fn collect_storage_scan_rows(
    snapshot: &PublishedSnapshot,
    active: &crate::ingest::ActiveSegment,
    schema: &crate::meta::Schema,
    predicate: Option<&crate::meta::Predicate>,
    cursor: Option<DocumentScanCursor>,
    limit: usize,
    control: &QueryControl,
    lease: &SnapshotLease,
    work: &mut usize,
) -> Result<Vec<DocumentScanRow>, QueryError> {
    let capacity = limit.saturating_add(1);
    let mut result = Vec::new();
    result.try_reserve_exact(capacity).map_err(|_| {
        QueryError::Store(StoreError::AllocationFailed {
            needed: allocation_bytes::<DocumentScanRow>(capacity),
            component: "scan positions",
        })
    })?;
    let mut started = cursor.is_none();
    'segments: for segment in snapshot.segments() {
        let source = DocumentScanSource::Sealed(segment.meta().id);
        let rows = sealed_scan_rows(segment, predicate)?;
        for row in rows.iter() {
            scan_control_checkpoint(control, lease, work)?;
            if !started {
                started =
                    cursor.is_some_and(|cursor| cursor.source == source && cursor.next_row == row);
                if !started {
                    continue;
                }
            }
            result.push(sealed_document_scan_row(segment, row)?);
            if result.len() == capacity {
                break 'segments;
            }
        }
    }
    if result.len() < capacity {
        let source = DocumentScanSource::Active;
        let rows = active_scan_rows(active, schema, predicate)?;
        for row in rows.iter() {
            scan_control_checkpoint(control, lease, work)?;
            if !started {
                started =
                    cursor.is_some_and(|cursor| cursor.source == source && cursor.next_row == row);
                if !started {
                    continue;
                }
            }
            result.push(active_document_scan_row(active, row)?);
            if result.len() == capacity {
                break;
            }
        }
    }
    if !started {
        return Err(QueryError::Store(StoreError::InvalidScan {
            detail: "scan cursor does not name a matching row".to_owned(),
        }));
    }
    Ok(result)
}

fn document_scan_row_cmp(
    left: &DocumentScanRow,
    right: &DocumentScanRow,
    order: ScanOrder,
) -> std::cmp::Ordering {
    let primary = match order {
        ScanOrder::Storage | ScanOrder::TimestampAscending => left.timestamp.cmp(&right.timestamp),
        ScanOrder::TimestampDescending => right.timestamp.cmp(&left.timestamp),
        ScanOrder::Attribute { direction, .. } => match (left.attribute_key, right.attribute_key) {
            (Some(left), Some(right)) => match direction {
                ScanDirection::Ascending => left.cmp(&right),
                ScanDirection::Descending => right.cmp(&left),
            },
            (Some(_), None) => std::cmp::Ordering::Less,
            (None, Some(_)) => std::cmp::Ordering::Greater,
            (None, None) => std::cmp::Ordering::Equal,
        },
    };
    primary
        .then_with(|| left.doc_id.cmp(&right.doc_id))
        .then_with(|| document_scan_source_cmp(left.source, right.source))
        .then_with(|| left.row.cmp(&right.row))
}

fn document_scan_source_cmp(
    left: DocumentScanSource,
    right: DocumentScanSource,
) -> std::cmp::Ordering {
    match (left, right) {
        (DocumentScanSource::Sealed(left), DocumentScanSource::Sealed(right)) => left.cmp(&right),
        (DocumentScanSource::Sealed(_), DocumentScanSource::Active) => std::cmp::Ordering::Less,
        (DocumentScanSource::Active, DocumentScanSource::Sealed(_)) => std::cmp::Ordering::Greater,
        (DocumentScanSource::Active, DocumentScanSource::Active) => std::cmp::Ordering::Equal,
    }
}

fn insert_timestamp_scan_row(
    rows: &mut Vec<DocumentScanRow>,
    candidate: DocumentScanRow,
    capacity: usize,
    order: ScanOrder,
) {
    let position = rows.partition_point(|row| {
        document_scan_row_cmp(row, &candidate, order) != std::cmp::Ordering::Greater
    });
    if position < capacity {
        rows.insert(position, candidate);
        if rows.len() > capacity {
            let _ = rows.pop();
        }
    }
}

fn clustering_range_cmp(
    left: crate::segment::ClusteringKeyRange,
    right: crate::segment::ClusteringKeyRange,
    order: ScanOrder,
) -> std::cmp::Ordering {
    use crate::segment::ClusteringKeyRange::{Bounded, Empty, Unstamped};

    match (left, right) {
        (Bounded { min_ts: left, .. }, Bounded { min_ts: right, .. })
            if order == ScanOrder::TimestampAscending =>
        {
            left.cmp(&right)
        }
        (Bounded { max_ts: left, .. }, Bounded { max_ts: right, .. }) => right.cmp(&left),
        (Unstamped, Unstamped) | (Empty, Empty) => std::cmp::Ordering::Equal,
        (Unstamped, _) => std::cmp::Ordering::Less,
        (_, Unstamped) => std::cmp::Ordering::Greater,
        (Empty, _) => std::cmp::Ordering::Greater,
        (_, Empty) => std::cmp::Ordering::Less,
    }
}

/// Collects the first `limit + 1` rows at or after `cursor` under a timestamp
/// or attribute order; every row is compared, so segment order only affects
/// how early the bounded buffer fills.
#[allow(clippy::too_many_arguments)]
fn collect_sorted_scan_rows(
    snapshot: &PublishedSnapshot,
    active: &crate::ingest::ActiveSegment,
    schema: &crate::meta::Schema,
    predicate: Option<&crate::meta::Predicate>,
    cursor: Option<DocumentScanCursor>,
    limit: usize,
    order: ScanOrder,
    control: &QueryControl,
    lease: &SnapshotLease,
    work: &mut usize,
) -> Result<Vec<DocumentScanRow>, QueryError> {
    let capacity = limit.saturating_add(1);
    let mut result = Vec::new();
    result.try_reserve_exact(capacity).map_err(|_| {
        QueryError::Store(StoreError::AllocationFailed {
            needed: allocation_bytes::<DocumentScanRow>(capacity),
            component: "scan positions",
        })
    })?;
    let attribute = match order {
        ScanOrder::Attribute { column, .. } => Some(column),
        ScanOrder::Storage | ScanOrder::TimestampAscending | ScanOrder::TimestampDescending => None,
    };
    let active_columns = attribute
        .map(|_| active_scan_columns(active, schema))
        .transpose()?;
    let active_keys = active_columns.as_ref().zip(attribute);
    let sealed_columns = |segment: &crate::segment::reader::SegmentReader| {
        attribute
            .map(|_| segment.query_columns().map_err(QueryError::Store))
            .transpose()
    };
    let cursor_row = cursor
        .map(|cursor| match cursor.source {
            DocumentScanSource::Active => keyed_scan_row(
                active_document_scan_row(active, cursor.next_row)?,
                active_keys,
            ),
            DocumentScanSource::Sealed(segment_id) => {
                let segment = snapshot
                    .segments()
                    .iter()
                    .find(|segment| segment.meta().id == segment_id)
                    .ok_or_else(|| {
                        QueryError::Store(StoreError::InvalidScan {
                            detail: format!("scan cursor names absent segment {segment_id}"),
                        })
                    })?;
                let columns = sealed_columns(segment)?;
                keyed_scan_row(
                    sealed_document_scan_row(segment, cursor.next_row)?,
                    columns.as_deref().zip(attribute),
                )
            }
        })
        .transpose()?;
    let mut segments = snapshot.segments().iter().collect::<Vec<_>>();
    segments.sort_by(|left, right| {
        clustering_range_cmp(
            left.meta().clustering_key_range,
            right.meta().clustering_key_range,
            order,
        )
        .then_with(|| left.meta().id.cmp(&right.meta().id))
    });
    for segment in segments {
        let rows = sealed_scan_rows(segment, predicate)?;
        let columns = sealed_columns(segment)?;
        for row in rows.iter() {
            scan_control_checkpoint(control, lease, work)?;
            let candidate = keyed_scan_row(
                sealed_document_scan_row(segment, row)?,
                columns.as_deref().zip(attribute),
            )?;
            if cursor_row.as_ref().is_none_or(|cursor_row| {
                document_scan_row_cmp(&candidate, cursor_row, order) != std::cmp::Ordering::Less
            }) {
                insert_timestamp_scan_row(&mut result, candidate, capacity, order);
            }
        }
    }
    for row in active_scan_rows(active, schema, predicate)?.iter() {
        scan_control_checkpoint(control, lease, work)?;
        let candidate = keyed_scan_row(active_document_scan_row(active, row)?, active_keys)?;
        if cursor_row.as_ref().is_none_or(|cursor_row| {
            document_scan_row_cmp(&candidate, cursor_row, order) != std::cmp::Ordering::Less
        }) {
            insert_timestamp_scan_row(&mut result, candidate, capacity, order);
        }
    }
    if let Some(cursor) = cursor {
        let starts_at_cursor = result
            .first()
            .is_some_and(|row| row.source == cursor.source && row.row == cursor.next_row);
        if !starts_at_cursor {
            return Err(QueryError::Store(StoreError::InvalidScan {
                detail: "scan cursor does not name a matching row".to_owned(),
            }));
        }
    }
    Ok(result)
}

fn materialize_active_document(
    active: &crate::ingest::ActiveSegment,
    row: usize,
    fields: DocumentFields,
) -> Result<StoredDocument, StoreError> {
    let version = active.document(row).ok_or(StoreError::ActiveRowOverflow)?;
    let timestamp = active
        .timestamps()
        .get(row)
        .copied()
        .ok_or(StoreError::ActiveRowOverflow)?;
    let vector = if fields.contains(DocumentFields::VECTOR) {
        let dimensions = active.dims().ok_or(StoreError::ActiveRowOverflow)?;
        let start = row
            .checked_mul(dimensions)
            .ok_or(StoreError::ActiveRowOverflow)?;
        let end = start
            .checked_add(dimensions)
            .ok_or(StoreError::ActiveRowOverflow)?;
        let values = active
            .vectors()
            .get(start..end)
            .ok_or(StoreError::ActiveRowOverflow)?;
        Some(copy_slice(values, "get vector")?)
    } else {
        None
    };
    let text = if fields.contains(DocumentFields::TEXT) {
        active
            .text(row)?
            .map(|value| copy_string(value, "get text"))
            .transpose()?
    } else {
        None
    };
    let metadata = if fields.contains(DocumentFields::METADATA) {
        active
            .metadata(row)
            .map(|value| copy_slice(value, "get metadata"))
            .transpose()?
    } else {
        None
    };
    let attributes = if fields.contains(DocumentFields::ATTRIBUTES) {
        Some(active.column_values(row)?)
    } else {
        None
    };
    Ok(StoredDocument {
        doc_id: version.doc_id(),
        revision: version.revision(),
        timestamp,
        vector,
        text,
        metadata,
        attributes,
    })
}

fn materialize_sealed_document(
    segment: &crate::segment::reader::SegmentReader,
    row: usize,
    fields: DocumentFields,
) -> Result<StoredDocument, StoreError> {
    let row_u32 = u32::try_from(row).map_err(|_| {
        StoreError::Segment(crate::segment::SegmentError::Geometry(
            "sealed document row exceeds u32".to_owned(),
        ))
    })?;
    let version = segment
        .document_version(row)
        .map_err(StoreError::Segment)?
        .ok_or_else(|| {
            StoreError::Segment(crate::segment::SegmentError::Geometry(format!(
                "sealed document row {row} has no identity"
            )))
        })?;
    let columns = segment.query_columns()?;
    let timestamp = columns.timestamp(row_u32).ok_or_else(|| {
        StoreError::Segment(crate::segment::SegmentError::Geometry(format!(
            "sealed document row {row} has no timestamp"
        )))
    })?;
    let vector = if fields.contains(DocumentFields::VECTOR) {
        segment
            .validate_rescore_rows(&[row_u32])
            .map_err(StoreError::Segment)?;
        let dimensions = segment.meta().dims as usize;
        let start = row.checked_mul(dimensions).ok_or_else(|| {
            StoreError::Segment(crate::segment::SegmentError::Geometry(
                "sealed vector row offset overflow".to_owned(),
            ))
        })?;
        let end = start.checked_add(dimensions).ok_or_else(|| {
            StoreError::Segment(crate::segment::SegmentError::Geometry(
                "sealed vector row end overflow".to_owned(),
            ))
        })?;
        let rescore = segment.query_rescore_f32().map_err(StoreError::Segment)?;
        let values = rescore.get(start..end).ok_or_else(|| {
            StoreError::Segment(crate::segment::SegmentError::Geometry(format!(
                "sealed vector row {row} is missing"
            )))
        })?;
        Some(copy_slice(values, "get vector")?)
    } else {
        None
    };
    let text = if fields.contains(DocumentFields::TEXT) {
        match segment.query_stored_text().map_err(StoreError::Segment)? {
            Some(rows) => rows
                .row(row)
                .ok_or_else(|| {
                    StoreError::Segment(crate::segment::SegmentError::Geometry(format!(
                        "stored-text row {row} is invalid"
                    )))
                })?
                .map(|value| copy_string(value, "get text"))
                .transpose()?,
            None => None,
        }
    } else {
        None
    };
    let metadata = if fields.contains(DocumentFields::METADATA) {
        match segment
            .query_stored_metadata()
            .map_err(StoreError::Segment)?
        {
            Some(rows) => Some(copy_slice(
                rows.row(row).ok_or_else(|| {
                    StoreError::Segment(crate::segment::SegmentError::Geometry(format!(
                        "stored-metadata row {row} is invalid"
                    )))
                })?,
                "get metadata",
            )?),
            None => None,
        }
    } else {
        None
    };
    let attributes = fields
        .contains(DocumentFields::ATTRIBUTES)
        .then(|| materialize_sealed_attributes(&columns, row_u32))
        .transpose()?;
    Ok(StoredDocument {
        doc_id: version.doc_id(),
        revision: version.revision(),
        timestamp,
        vector,
        text,
        metadata,
        attributes,
    })
}

fn materialize_sealed_attributes(
    columns: &crate::meta::ColumnStore,
    row: u32,
) -> Result<Vec<(crate::meta::ColumnId, crate::meta::PredicateValue)>, StoreError> {
    let mut attributes = Vec::new();
    attributes
        .try_reserve_exact(columns.schema().user_column_count())
        .map_err(|_| StoreError::AllocationFailed {
            needed: allocation_bytes::<(crate::meta::ColumnId, crate::meta::PredicateValue)>(
                columns.schema().user_column_count(),
            ),
            component: "get attributes",
        })?;
    for definition in columns
        .schema()
        .columns()
        .iter()
        .filter(|definition| definition.id() != crate::meta::TIMESTAMP_COLUMN)
    {
        let column = columns.column(definition.id()).ok_or_else(|| {
            StoreError::Segment(crate::segment::SegmentError::Geometry(format!(
                "sealed attribute column {} is missing",
                definition.id().get()
            )))
        })?;
        let value = predicate_value_for_row(column, row)?;
        if value.is_none() && !definition.is_nullable() {
            return Err(StoreError::Segment(crate::segment::SegmentError::Geometry(
                format!(
                    "sealed required attribute {} is null at row {row}",
                    definition.id().get()
                ),
            )));
        }
        if let Some(value) = value {
            attributes.push((definition.id(), value));
        }
    }
    Ok(attributes)
}

fn predicate_value_for_row(
    column: &crate::meta::Column,
    row: u32,
) -> Result<Option<crate::meta::PredicateValue>, StoreError> {
    use crate::meta::{Column, PredicateValue};

    match column {
        Column::Id128(values) => Ok(values.get(row).map(PredicateValue::Id128)),
        Column::U64(values) => Ok(values.get(row).map(PredicateValue::U64)),
        Column::I64(values) => Ok(values.get(row).map(PredicateValue::I64)),
        Column::F64(values) => Ok(values.get(row).map(PredicateValue::F64)),
        Column::Bool(values) => Ok(values.get(row).map(PredicateValue::Bool)),
        Column::DictionaryString(values) => values
            .get(row)
            .map(|value| copy_string(value, "get attribute string").map(PredicateValue::String))
            .transpose(),
        Column::RawString(values) => values
            .get(row)
            .map(|value| copy_string(value, "get attribute string").map(PredicateValue::String))
            .transpose(),
    }
}

fn copy_slice<T: Copy>(values: &[T], component: &'static str) -> Result<Vec<T>, StoreError> {
    let mut copied = Vec::new();
    copied
        .try_reserve_exact(values.len())
        .map_err(|_| StoreError::AllocationFailed {
            needed: allocation_bytes::<T>(values.len()),
            component,
        })?;
    copied.extend_from_slice(values);
    Ok(copied)
}

fn copy_string(value: &str, component: &'static str) -> Result<String, StoreError> {
    let mut copied = String::new();
    copied
        .try_reserve_exact(value.len())
        .map_err(|_| StoreError::AllocationFailed {
            needed: u64::try_from(value.len()).unwrap_or(u64::MAX),
            component,
        })?;
    copied.push_str(value);
    Ok(copied)
}

fn allocation_bytes<T>(count: usize) -> u64 {
    count
        .checked_mul(std::mem::size_of::<T>())
        .and_then(|bytes| u64::try_from(bytes).ok())
        .unwrap_or(u64::MAX)
}

#[cfg(test)]
mod lexical_assembly_tests;

#[cfg(test)]
mod get_tests;
#[cfg(test)]
mod group_count_tests;
#[cfg(test)]
mod scan_attribute_order_tests;
#[cfg(test)]
mod scan_document_tests;

#[cfg(test)]
#[allow(clippy::expect_used, clippy::panic)]
mod tests {
    use std::error::Error;
    use std::sync::Arc;

    use tempfile::tempdir;

    use crate::lifecycle::durability::DurabilityPolicy;
    use crate::manifest::io::commit_manifest;
    use crate::manifest::{Manifest, ManifestError};
    use crate::meta::Schema;
    use crate::vfs::StdVfs;
    #[cfg(feature = "graph-cypher")]
    use crate::vfs::Vfs;
    use crate::vfs::crash::MemoryVfs;

    use super::durability::{CommitTier, DurabilityMode, DurabilityPolicyError};
    use super::{
        CancelToken, GraphSearchOptions, ManualMonotonicClock, OpenOptions, QueryControl,
        SearchOptions, SearchTier, Store, StoreError, StoreTestDependencies,
        resolve_hybrid_leg_results,
    };

    #[cfg(feature = "graph-cypher")]
    #[test]
    fn graph_free_initial_manifest_bytes_survive_a_failed_writer_open_as_on_main() {
        use super::native_graph::tests::publication::{FaultPoint, RecordingVfs};
        let directory = tempfile::tempdir().expect("fixture");
        let schema = crate::meta::Schema::new(vec![crate::meta::ColumnDefinition::new(
            crate::meta::ColumnId::new(1),
            "extra",
            crate::meta::ColumnType::U64,
            true,
        )])
        .expect("fixture");
        let vfs = Arc::new(RecordingVfs::default());
        vfs.arm_fault(FaultPoint::OpenAppend);
        assert!(
            Store::open_with_test_dependencies(
                directory.path(),
                OpenOptions::default().with_schema(schema.clone()),
                StoreTestDependencies::new(vfs.clone(), Arc::new(super::SystemMonotonicClock)),
            )
            .is_err()
        );
        vfs.assert_fired_once();
        let expected = crate::manifest::encode_manifest(&crate::manifest::Manifest {
            graph: None,
            generation: 0,
            log_seq: 0,
            segments: Vec::new(),
            epochs: Vec::new(),
            epoch_alias: None,
            schema,
        })
        .expect("fixture");
        assert_eq!(
            StdVfs
                .read(&directory.path().join("manifest.ze"))
                .expect("main publishes the initial manifest before opening the writer"),
            expected
        );
    }

    #[cfg(feature = "graph-cypher")]
    #[test]
    fn graph_free_schema_evolution_bytes_survive_a_failed_writer_open_as_on_main() {
        use super::native_graph::tests::publication::{FaultPoint, RecordingVfs};
        let directory = tempdir().expect("directory");
        let store = Store::open(directory.path(), OpenOptions::default()).expect("store");
        store
            .ingest(crate::ingest::IngestBatch::new(vec![
                crate::ingest::IngestDocument::new(
                    crate::ingest::DocumentVersion::new(
                        crate::ingest::DocId::new(91),
                        crate::ingest::Revision::new(1),
                    ),
                    vec![1.0],
                ),
            ]))
            .expect("document");
        store.seal().expect("seal");
        store.close().expect("close");
        let schema = Schema::new(vec![crate::meta::ColumnDefinition::new(
            crate::meta::ColumnId::new(1),
            "extra",
            crate::meta::ColumnType::U64,
            true,
        )])
        .expect("schema");
        let mut expected = crate::manifest::io::load_manifest(
            &StdVfs,
            &directory.path().join("manifest.ze"),
            u64::MAX,
        )
        .expect("prior manifest");
        expected.generation += 1;
        expected.schema = schema.clone();
        let snapshot = || {
            std::fs::read_dir(directory.path())
                .expect("files")
                .map(|entry| {
                    let entry = entry.expect("entry");
                    (
                        entry.file_name(),
                        StdVfs.read(&entry.path()).expect("bytes"),
                    )
                })
                .collect::<std::collections::BTreeMap<_, _>>()
        };
        let mut expected_files = snapshot();
        expected_files.insert(
            std::ffi::OsString::from("manifest.ze"),
            crate::manifest::encode_manifest(&expected).expect("main bytes"),
        );
        let vfs = Arc::new(RecordingVfs::default());
        vfs.arm_fault(FaultPoint::OpenAppend);
        assert!(
            Store::open_with_test_dependencies(
                directory.path(),
                OpenOptions::default().with_schema(schema),
                StoreTestDependencies::new(vfs.clone(), Arc::new(super::SystemMonotonicClock)),
            )
            .is_err()
        );
        vfs.assert_fired_once();
        assert_eq!(
            crate::manifest::io::load_manifest(
                &StdVfs,
                &directory.path().join("manifest.ze"),
                u64::MAX
            )
            .expect("published schema")
            .generation,
            expected.generation,
            "main commits schema before opening the writer"
        );
        assert_eq!(snapshot(), expected_files);
    }

    #[cfg(feature = "graph-cypher")]
    #[test]
    pub(super) fn replacement_snapshot_backup_reopens_with_nonzero_watermarks() {
        use crate::ingest::{DocId, DocumentVersion, IngestBatch, IngestDocument, Revision};
        use crate::meta::{AliveSet, ColumnStoreBuilder, Schema};
        let directory = tempdir().expect("source");
        let parent = tempdir().expect("backup parent");
        let options =
            OpenOptions::new().with_durability(DurabilityMode::Durable, CommitTier::Durable);
        let store = Store::open(directory.path(), options.clone()).expect("open source");
        assert_eq!(store.enable_graph().expect("enable graph"), 1);
        let document = |id| {
            IngestDocument::new(
                DocumentVersion::new(DocId::new(id), Revision::new(1)),
                vec![1.0, 0.0],
            )
        };
        let ack = store
            .ingest(IngestBatch::new(vec![document(1)]))
            .expect("ingest");
        assert_eq!(ack.seq().get(), 1);
        assert_eq!(store.seal().expect("seal"), 3);
        let columns = ColumnStoreBuilder::new(Schema::timestamp_only())
            .finish()
            .expect("columns");
        let alive = AliveSet::new(0);
        let prepared = store
            .prepare_segment(super::InMemorySegment {
                id: crate::segment::SegmentId::new(1, [2; 10]),
                scheme: 4,
                dims: 2,
                codes: Vec::new(),
                factors: super::InMemorySegmentFactors::Bit4(Vec::new()),
                rescore: Vec::new(),
                columns: &columns,
                alive: &alive,
            })
            .expect("prepare empty replacement");
        assert_eq!(store.seal_snapshot(prepared).expect("replace snapshot"), 4);
        let target = parent.path().join("backup");
        assert_eq!(
            store.write_snapshot(&target).expect("acknowledge backup"),
            4
        );
        // No crash or source close: a successfully acknowledged backup must open.
        for access in [super::AccessMode::ReadOnly, super::AccessMode::ReadWrite] {
            let backup = Store::open(&target, options.clone().with_access_mode(access))
                .expect("acknowledged replacement backup must open");
            assert_eq!(backup.snapshot().expect("snapshot").generation(), 4);
            assert_eq!(backup.count_documents(None, None).expect("count").count, 0);
            assert!(backup.admit_native_read().is_ok());
            backup.close().expect("close backup");
        }
        let manifest = crate::manifest::io::load_manifest(&StdVfs, &target.join("manifest.ze"), 1)
            .expect("backup manifest");
        assert_eq!(manifest.log_seq, 1);
        assert_eq!(
            manifest
                .graph
                .expect("graph section")
                .graph_absorbed_through,
            1
        );
        let backup = Store::open(&target, options.clone()).expect("writable backup");
        let ack = backup
            .ingest(IngestBatch::new(vec![document(2)]))
            .expect("continue backup WAL");
        assert_eq!(ack.seq().get(), 2);
        assert_eq!(ack.generation(), 5);
        backup.close().expect("close continued backup");
        store.close().expect("close source");
        let source = Store::open(directory.path(), options).expect("reopen replacement source");
        assert_eq!(
            source
                .count_documents(None, None)
                .expect("source count")
                .count,
            0
        );
        assert_eq!(source.snapshot().expect("source snapshot").generation(), 4);
    }

    #[cfg(feature = "graph-cypher")]
    #[test]
    fn enable_graph_retry_completes_a_failed_install() {
        enable_graph_retry_completes(
            super::native_graph::tests::publication::FaultPoint::Enumeration,
        );
    }

    #[cfg(feature = "graph-cypher")]
    #[test]
    fn enable_graph_retry_completes_a_failed_directory_sync() {
        enable_graph_retry_completes(
            super::native_graph::tests::publication::FaultPoint::SelectorSync,
        );
    }

    #[cfg(feature = "graph-cypher")]
    #[test]
    fn enable_graph_retry_survives_a_failed_writer_reopen() {
        enable_graph_retry_with_writer_reopen_failure(
            super::native_graph::tests::publication::FaultPoint::Enumeration,
            true,
        );
    }

    #[cfg(feature = "graph-cypher")]
    pub(super) fn enable_graph_retry_completes(
        point: super::native_graph::tests::publication::FaultPoint,
    ) {
        enable_graph_retry_with_writer_reopen_failure(point, false)
    }

    #[cfg(feature = "graph-cypher")]
    fn enable_graph_retry_with_writer_reopen_failure(
        point: super::native_graph::tests::publication::FaultPoint,
        fail_writer_reopen: bool,
    ) {
        use super::native_graph::tests::publication::{DurabilityEvent, RecordingVfs};
        let directory = tempdir().expect("directory");
        let vfs = Arc::new(RecordingVfs::default());
        let store = Store::open_with_test_dependencies(
            directory.path(),
            OpenOptions::new().with_max_resident_bytes(256 * 1024 * 1024),
            StoreTestDependencies::new(vfs.clone(), Arc::new(super::SystemMonotonicClock)),
        )
        .expect("open");
        let wal = StdVfs.read(&directory.path().join("wal.ze")).expect("WAL");
        vfs.arm_fault(point);
        assert!(store.enable_graph().is_err());
        vfs.assert_fired_once();
        assert!(
            store
                .ingest(crate::ingest::IngestBatch::new(vec![
                    crate::ingest::IngestDocument::new(
                        crate::ingest::DocumentVersion::new(
                            crate::ingest::DocId::new(99),
                            crate::ingest::Revision::new(1)
                        ),
                        vec![1.0, 0.0],
                    )
                ]))
                .is_err(),
            "failed enable must fence document writes until the retry completes"
        );
        if fail_writer_reopen {
            vfs.arm_fault(super::native_graph::tests::publication::FaultPoint::OpenAppend);
            assert!(store.enable_graph().is_err());
            vfs.assert_fired_once();
        }
        vfs.clear_events();
        assert_eq!(store.enable_graph().expect("retry completes upgrade"), 1);
        assert_eq!(store.snapshot().expect("snapshot").generation(), 1);
        assert!(
            store.admit_native_read().is_ok(),
            "retry must install the bundle"
        );
        assert!(
            vfs.take().iter().any(|event| matches!(event,
                DurabilityEvent::Sync(path, crate::vfs::SyncKind::Full) if path == directory.path()
            )),
            "retry must complete the directory durability barrier"
        );
        assert_eq!(
            StdVfs.read(&directory.path().join("wal.ze")).expect("WAL"),
            wal
        );
        assert_eq!(store.enable_graph().expect("idempotent"), 1);
        let ack = store
            .ingest(crate::ingest::IngestBatch::new(vec![
                crate::ingest::IngestDocument::new(
                    crate::ingest::DocumentVersion::new(
                        crate::ingest::DocId::new(91),
                        crate::ingest::Revision::new(1),
                    ),
                    vec![1.0, 0.0],
                ),
            ]))
            .expect("document writer repaired");
        assert_eq!(ack.generation(), 2);
        drop(store);
        let reopened = Store::open(
            directory.path(),
            OpenOptions::new().with_max_resident_bytes(256 * 1024 * 1024),
        )
        .expect("reopen");
        assert_eq!(reopened.snapshot().expect("snapshot").generation(), 2);
        assert!(reopened.admit_native_read().is_ok());
    }

    #[test]
    fn astra_12_vocabulary_charge_survives_eviction_and_close() {
        use crate::fts::{index::DEFAULT_FIELD, query::LexicalQuery};
        use crate::ingest::{DocId, DocumentVersion, IngestBatch, IngestDocument, Revision};
        let directory = tempdir().expect("directory");
        let store = Store::open(directory.path(), OpenOptions::default()).expect("open");
        store
            .ingest(IngestBatch::new(vec![
                IngestDocument::new(
                    DocumentVersion::new(DocId::new(1), Revision::new(1)),
                    vec![1.0, 0.0],
                )
                .with_text("alpha alphabet beta"),
            ]))
            .expect("ingest");
        store
            .search_lexical_structured(
                &LexicalQuery::prefix(b"alph".to_vec(), DEFAULT_FIELD),
                1,
                64,
                QueryControl::Cancel(CancelToken::new()),
            )
            .expect("build cache");
        let dictionary = {
            let entry = store.lexical_index_cache.entry.lock().expect("assembly");
            let vocabulary = entry
                .as_ref()
                .expect("assembly")
                .assembly
                .vocabulary
                .lock()
                .expect("vocabulary");
            Arc::clone(vocabulary.as_ref().expect("cached"))
        };
        let bytes = dictionary._memory.bytes();
        assert!(bytes > 17);
        let second_owner = Arc::clone(&dictionary);
        drop(
            store
                .lexical_index_cache
                .entry
                .lock()
                .expect("evict")
                .take(),
        );
        assert_eq!(store.stats().expect("held bytes").cache_bytes, bytes);
        store
            .search_lexical_structured(
                &LexicalQuery::prefix(b"alph".to_vec(), DEFAULT_FIELD),
                1,
                64,
                QueryControl::Cancel(CancelToken::new()),
            )
            .expect("replacement cache");
        assert!(store.stats().expect("both dictionaries").cache_bytes > bytes);
        store.close().expect("close");
        assert_eq!(
            store
                .accounting
                .audit()
                .expect("closed accounting")
                .cache_bytes,
            bytes
        );
        drop(dictionary);
        assert_eq!(
            store.accounting.audit().expect("second owner").cache_bytes,
            bytes
        );
        assert_eq!(
            second_owner.view.iter().collect::<Vec<_>>(),
            vec![b"alpha".as_slice(), b"alphabet", b"beta"]
        );
        drop(second_owner);
        assert_eq!(store.accounting.audit().expect("last owner").cache_bytes, 0);
        assert_eq!(
            store.accounting.audit().expect("scratch").temporary_bytes,
            0
        );
    }

    #[test]
    fn astra_12_refused_vocabulary_is_not_published() {
        use crate::fts::{index::DEFAULT_FIELD, query::LexicalQuery};
        use crate::ingest::{
            DocId, DocumentVersion, IngestBatch, IngestDocument, Revision, StoreLexicalError,
        };
        let directory = tempdir().expect("directory");
        let writer = Store::open(directory.path(), OpenOptions::default()).expect("open");
        writer
            .ingest(IngestBatch::new(vec![
                IngestDocument::new(
                    DocumentVersion::new(DocId::new(1), Revision::new(1)),
                    vec![1.0, 0.0],
                )
                .with_text("alpha alphabet beta"),
            ]))
            .expect("ingest");
        writer.seal().expect("seal");
        writer.close().expect("close writer");
        let store = Store::open(
            directory.path(),
            OpenOptions::default().with_max_temp_bytes(1),
        )
        .expect("small budget");
        for _ in 0..2 {
            let result = store.search_lexical_structured(
                &LexicalQuery::prefix(b"alph".to_vec(), DEFAULT_FIELD),
                1,
                64,
                QueryControl::Cancel(CancelToken::new()),
            );
            assert!(
                matches!(
                    result,
                    Err(StoreLexicalError::Query(super::QueryError::Store(
                        StoreError::BudgetExceeded {
                            component: "temporary",
                            ..
                        }
                    )))
                ),
                "{result:?}"
            );
            assert_eq!(store.stats().expect("released").temporary_bytes, 0);
            let entry = store.lexical_index_cache.entry.lock().expect("assembly");
            assert!(
                entry
                    .as_ref()
                    .expect("assembled")
                    .assembly
                    .vocabulary
                    .lock()
                    .expect("dictionary")
                    .is_none()
            );
        }
        store.close().expect("close");
        assert_eq!(
            store
                .accounting
                .audit()
                .expect("released cache")
                .cache_bytes,
            0
        );
    }

    #[test]
    fn astra_03_lexical_assembly_charge_survives_cache_eviction() {
        use crate::fts::index::DEFAULT_FIELD;
        use crate::fts::search::TermQuery;
        use crate::ingest::{DocId, DocumentVersion, IngestBatch, IngestDocument, Revision};
        let directory = tempdir().expect("directory");
        let store = Store::open(directory.path(), OpenOptions::default()).expect("open");
        store
            .ingest(IngestBatch::new(vec![
                IngestDocument::new(
                    DocumentVersion::new(DocId::new(1), Revision::new(1)),
                    vec![1.0, 0.0],
                )
                .with_text("alpha beta"),
            ]))
            .expect("ingest");
        store
            .search_lexical(
                &TermQuery::flat(vec![b"alpha".to_vec()], &[DEFAULT_FIELD]),
                1,
                QueryControl::Cancel(CancelToken::new()),
            )
            .expect("build lexical cache");
        let (first, second) = {
            let entry = store.lexical_index_cache.entry.lock().expect("cache lock");
            let assembly = &entry.as_ref().expect("cached assembly").assembly;
            (Arc::clone(assembly), Arc::clone(assembly))
        };
        let charged = store.stats().expect("charged stats").cache_bytes;
        let evicted = store
            .lexical_index_cache
            .entry
            .lock()
            .expect("cache lock")
            .take();
        drop(evicted);
        assert_eq!(
            first.index.document_count(),
            1,
            "eviction left the query data alive"
        );
        assert_eq!(
            store.stats().expect("held stats").cache_bytes,
            charged,
            "cache eviction must not release a query-owned assembly's charge"
        );
        drop(first);
        assert_eq!(
            store.stats().expect("second owner stats").cache_bytes,
            charged
        );
        drop(second);
        let released = store.stats().expect("released stats").cache_bytes;
        assert_eq!(
            released, 0,
            "the final owner must release the full assembly charge"
        );
        println!("assembly cache bytes: retained={charged}, after final owner={released}");
        store.close().expect("close");
    }

    /// Vectors whose norms differ enough that the ceiling is not degenerate.
    const CEILING_ROWS: [[f32; 4]; 3] = [
        [0.5, -0.25, 0.75, 0.125],
        [3.0, 1.5, -2.5, 0.25],
        [0.0, 0.0, 0.0, 0.0],
    ];

    /// Publishes a scheme-0 segment that no seal path can produce.
    fn publish_f32_segment(directory: &std::path::Path) {
        let created = Store::open(directory, OpenOptions::default()).expect("create f32 store");
        created.close().expect("close f32 store");

        let schema = Schema::timestamp_only();
        let mut columns = crate::meta::ColumnStoreBuilder::new(schema.clone());
        for row in 0..CEILING_ROWS.len() {
            columns
                .push_row(i64::try_from(row).expect("row timestamp"), &[])
                .expect("timestamp row");
        }
        let columns = columns.finish().expect("columns");
        let alive =
            crate::meta::AliveSet::new(u32::try_from(CEILING_ROWS.len()).expect("row count"));
        let rescore = CEILING_ROWS.iter().flatten().copied().collect::<Vec<f32>>();
        let codes = rescore
            .iter()
            .flat_map(|value| value.to_bits().to_le_bytes())
            .collect::<Vec<u8>>();
        let policy = DurabilityPolicy::new(DurabilityMode::Derived, CommitTier::None)
            .expect("f32 fixture durability");
        let meta = crate::segment::writer::write_segment(
            &StdVfs,
            directory,
            crate::segment::writer::SegmentBuild {
                id: crate::segment::SegmentId::new(11, [7; 10]),
                scheme: 0,
                dims: 4,
                codes: &codes,
                factors: crate::segment::writer::SegmentFactors::F32,
                rescore: &rescore,
                columns: &columns,
                alive: &alive,
            },
            policy,
        )
        .expect("write scheme-0 segment");
        commit_manifest(
            &StdVfs,
            directory,
            &Manifest {
                #[cfg(feature = "graph-cypher")]
                graph: None,
                generation: 1,
                log_seq: 0,
                segments: vec![meta],
                epochs: Vec::new(),
                epoch_alias: None,
                schema,
            },
            policy,
        )
        .expect("publish scheme-0 segment");
    }

    /// Ingests `CEILING_ROWS`, sealing every row except an optional tail that
    /// stays in the mutable active segment.
    fn ingest_ceiling_rows(store: &Store, keep_active: usize, one_segment_per_row: bool) {
        use crate::ingest::{DocId, DocumentVersion, IngestBatch, IngestDocument, Revision};

        let sealed = CEILING_ROWS.len() - keep_active;
        for (index, row) in CEILING_ROWS.iter().enumerate() {
            let version = DocumentVersion::new(
                DocId::new(u128::try_from(index).expect("document id")),
                Revision::new(1),
            );
            store
                .ingest(IngestBatch::new(vec![IngestDocument::new(
                    version,
                    row.to_vec(),
                )]))
                .expect("ingest ceiling row");
            let seal_here = if one_segment_per_row {
                index < sealed
            } else {
                index + 1 == sealed
            };
            if seal_here {
                store.seal().expect("seal ceiling rows");
            }
        }
    }

    /// Proves the remembered per-segment enclosure reproduces the ceiling
    /// exactly, for one store, over several queries.
    fn assert_remembered_ceiling_is_exact(store: &Store, expected_segments: usize) {
        let queries: [Vec<f32>; 3] = [
            vec![0.0, 0.0, 0.0, 0.0],
            vec![1.0, -1.0, 0.5, 2.0],
            vec![7.5, 0.25, -3.5, 0.125],
        ];
        let admitted = store
            .admit_vector_search(SearchOptions::default())
            .expect("admit ceiling query");
        let snapshot = &admitted.snapshot;
        let active = &admitted.active_segment;
        assert_eq!(snapshot.segments().len(), expected_segments);
        for segment in snapshot.segments() {
            assert!(
                segment.cached_vector_ceiling_norm_range().is_none(),
                "a segment remembered an enclosure before any query walked it"
            );
        }

        for query in &queries {
            let cold = super::exact_vector_ceiling(snapshot, active, query).expect("cold ceiling");
            for segment in snapshot.segments() {
                let remembered = segment
                    .cached_vector_ceiling_norm_range()
                    .expect("a walked segment must remember its enclosure");
                let recomputed = super::compute_segment_vector_ceiling_norm_range(segment)
                    .expect("recomputed enclosure");
                assert_eq!(
                    remembered.endpoint_bits(),
                    recomputed.endpoint_bits(),
                    "remembered enclosure differs from the recomputed one"
                );
            }
            let warm = super::exact_vector_ceiling(snapshot, active, query).expect("warm ceiling");
            let recomputed = super::fold_exact_vector_ceiling(
                snapshot,
                active,
                query,
                super::compute_segment_vector_ceiling_norm_range,
            )
            .expect("recomputed ceiling");
            assert_eq!(
                cold.to_bits(),
                warm.to_bits(),
                "the remembered ceiling changed between queries"
            );
            assert_eq!(
                cold.to_bits(),
                recomputed.to_bits(),
                "the remembered ceiling is not the recomputed ceiling"
            );
        }
    }

    #[test]
    fn astra_00_elapsed_includes_ceiling_work() {
        use std::time::{Duration, Instant};
        struct CeilingClock {
            base: Instant,
            snapshot: Arc<super::PublishedSnapshot>,
        }
        impl super::MonotonicClock for CeilingClock {
            fn now(&self) -> Instant {
                // Advance exactly when the production ceiling has completed
                // its immutable enclosure. No sleeping or CPU timing oracle.
                if self
                    .snapshot
                    .segments()
                    .iter()
                    .all(|segment| segment.cached_vector_ceiling_norm_range().is_some())
                {
                    self.base + Duration::from_secs(7)
                } else {
                    self.base
                }
            }
        }
        let directory = tempdir().expect("ceiling timer store");
        let mut store = Store::open(directory.path(), OpenOptions::default()).expect("open");
        ingest_ceiling_rows(&store, 0, false);
        let admitted = store
            .admit_vector_search(SearchOptions::default())
            .expect("snapshot");
        assert_eq!(admitted.snapshot.segments().len(), 1);
        assert!(
            admitted
                .snapshot
                .segments()
                .iter()
                .all(|segment| { segment.cached_vector_ceiling_norm_range().is_none() })
        );
        let snapshot = Arc::clone(&admitted.snapshot);
        drop(admitted);
        store.clock = Arc::new(CeilingClock {
            base: Instant::now(),
            snapshot,
        });
        let outcome = store
            .search(
                crate::ingest::SearchRequest::new(&[1.0, 0.0, 0.0, 0.0]),
                1,
                SearchOptions::default().with_tier(SearchTier::Exact),
                QueryControl::Cancel(CancelToken::new()),
            )
            .expect("exact query");
        assert_eq!(outcome.diagnostics.elapsed, Duration::from_secs(7));
        // Release the test clock's snapshot before close waits for all pins.
        store.clock = Arc::new(super::SystemMonotonicClock);
        store.close().expect("close");
    }

    #[test]
    fn the_remembered_vector_ceiling_is_the_recomputed_ceiling_for_every_scheme() {
        let bit4_directory = tempdir().expect("Bit4 ceiling store directory");
        let bit4 = Store::open(bit4_directory.path(), OpenOptions::default())
            .expect("open Bit4 ceiling store");
        ingest_ceiling_rows(&bit4, 0, false);
        assert_remembered_ceiling_is_exact(&bit4, 1);
        bit4.close().expect("close Bit4 ceiling store");

        let split_directory = tempdir().expect("multi-segment ceiling store directory");
        let split = Store::open(split_directory.path(), OpenOptions::default())
            .expect("open multi-segment ceiling store");
        ingest_ceiling_rows(&split, 0, true);
        assert_remembered_ceiling_is_exact(&split, CEILING_ROWS.len());
        split.close().expect("close multi-segment ceiling store");

        let int8_directory = tempdir().expect("Int8 ceiling store directory");
        let int8 = Store::open_with_test_dependencies(
            int8_directory.path(),
            OpenOptions::default(),
            StoreTestDependencies::new(Arc::new(StdVfs), Arc::new(super::SystemMonotonicClock))
                .with_vector_seal_scheme(crate::quant::QuantScheme::Int8),
        )
        .expect("open Int8 ceiling store");
        ingest_ceiling_rows(&int8, 0, false);
        assert_remembered_ceiling_is_exact(&int8, 1);
        int8.close().expect("close Int8 ceiling store");

        let f32_directory = tempdir().expect("F32 ceiling store directory");
        publish_f32_segment(f32_directory.path());
        let f32_store = Store::open(f32_directory.path(), OpenOptions::default())
            .expect("open F32 ceiling store");
        assert_remembered_ceiling_is_exact(&f32_store, 1);
        f32_store.close().expect("close F32 ceiling store");

        let active_directory = tempdir().expect("active ceiling store directory");
        let active = Store::open(active_directory.path(), OpenOptions::default())
            .expect("open active ceiling store");
        ingest_ceiling_rows(&active, 2, false);
        assert!(
            !active
                .admit_vector_search(SearchOptions::default())
                .expect("admit active ceiling query")
                .active_segment
                .is_empty(),
            "the active-segment case sealed every row"
        );
        assert_remembered_ceiling_is_exact(&active, 1);
        active.close().expect("close active ceiling store");
    }

    #[test]
    fn exact_and_graph_tiers_preserve_quantizer_validation_errors() {
        let directory = tempdir().expect("store directory");
        let store = Store::open(directory.path(), OpenOptions::default()).expect("open");
        let query = vec![0.0_f32; crate::kernels::MAX_DOT_I8_DIMENSION + 1];

        for tier in [
            SearchTier::Exact,
            SearchTier::Graph(GraphSearchOptions::default()),
        ] {
            let error = store
                .search(
                    crate::ingest::SearchRequest::new(&query),
                    1,
                    SearchOptions::default().with_tier(tier),
                    QueryControl::Cancel(CancelToken::new()),
                )
                .expect_err("oversized quantized query must retain its error");
            assert!(matches!(
                error,
                super::QueryError::Scan(crate::scan::ScanError::Quant(
                    crate::quant::QuantError::DimensionTooLarge { actual, maximum }
                )) if actual == query.len()
                    && maximum == crate::kernels::MAX_DOT_I8_DIMENSION
            ));
        }
    }

    #[cfg(feature = "allocation-audit")]
    #[test]
    fn exact_tier_query_does_not_prepare_quantized_forms() {
        let directory = tempdir().expect("store directory");
        let store = Store::open(directory.path(), OpenOptions::default()).expect("open");
        let query = [0.25_f32; 128];
        let options = SearchOptions::default().with_tier(SearchTier::Exact);
        let warm = store
            .search(
                crate::ingest::SearchRequest::new(&query),
                1,
                options,
                QueryControl::Cancel(CancelToken::new()),
            )
            .expect("warm exact-tier search");
        assert!(warm.candidates.is_empty());
        let control = QueryControl::Cancel(CancelToken::new());

        let (outcome, report) = crate::allocation_audit::audit_engine_path(|| {
            store.search(
                crate::ingest::SearchRequest::new(&query),
                1,
                options,
                control,
            )
        });
        let outcome = outcome.expect("audited exact-tier search");
        assert!(outcome.candidates.is_empty());
        assert_eq!(report.allocations, 0, "Exact tier prepared quantized forms");
        assert_eq!(report.attributed_bytes, 0);
        assert_eq!(report.unattributed_bytes, 0);
    }

    #[test]
    fn astra_18_scoped_work_is_joined_before_cancel_returns() {
        use std::sync::atomic::{AtomicBool, Ordering};
        use std::sync::mpsc;
        let worker = super::pool::LexicalWorker::start().expect("worker");
        let token = CancelToken::new();
        let work_token = token.clone();
        let finished = AtomicBool::new(false);
        let mut borrowed_rows = Vec::new();
        let (started_tx, started_rx) = mpsc::sync_channel(0);
        let (release_tx, release_rx) = mpsc::sync_channel(0);
        let worker_rows = &mut borrowed_rows;
        let worker_finished = &finished;
        let (caller, lexical) = worker
            .run_scoped(
                move || {
                    started_tx.send(()).expect("entered borrowed job");
                    release_rx.recv().expect("caller has cancelled");
                    assert!(work_token.is_cancelled());
                    worker_rows.push(42_u32);
                    worker_finished.store(true, Ordering::Release);
                    Err::<(), _>(crate::fusion::FusionError::Cancelled { partial: false })
                },
                || {
                    started_rx.recv().expect("worker entered");
                    token.cancel();
                    assert!(!finished.load(Ordering::Acquire));
                    release_tx.send(()).expect("release borrowed job");
                    Err::<(), _>(crate::fusion::FusionError::Cancelled { partial: false })
                },
            )
            .expect("scoped dispatch");
        assert!(finished.load(Ordering::Acquire));
        assert_eq!(borrowed_rows, [42]);
        assert_eq!(
            resolve_hybrid_leg_results(caller, lexical.expect("worker returned")),
            Err(crate::fusion::FusionError::Cancelled { partial: false }),
        );
        worker.stop_and_join().expect("join worker");
    }

    #[test]
    fn astra_18_control_error_precedence_is_completion_order_independent() {
        use crate::fusion::{FusionError, FusionLeg, LegFailureKind};
        use std::sync::mpsc;
        let failures = [
            FusionError::Leg {
                leg: FusionLeg::Vector,
                kind: LegFailureKind::Scan,
                detail: "scan".to_owned(),
            },
            FusionError::Cancelled { partial: false },
            FusionError::Timeout { partial: false },
            FusionError::ReadCancelled { partial: false },
        ];
        let worker = super::pool::LexicalWorker::start().expect("worker");
        for (a, vector) in failures.iter().enumerate() {
            for (b, lexical) in failures.iter().enumerate() {
                for lexical_first in [false, true] {
                    let (lexical_tx, lexical_rx) = mpsc::sync_channel(1);
                    let (vector_tx, vector_rx) = mpsc::sync_channel(1);
                    let (vector_result, lexical_result) = worker
                        .run_scoped(
                            move || {
                                if !lexical_first {
                                    vector_rx.recv().expect("vector completed");
                                }
                                if lexical_first {
                                    lexical_tx.send(()).expect("lexical completed");
                                }
                                Err::<(), _>(lexical.clone())
                            },
                            || {
                                if lexical_first {
                                    lexical_rx.recv().expect("lexical completed");
                                }
                                if !lexical_first {
                                    vector_tx.send(()).expect("vector completed");
                                }
                                Err::<(), _>(vector.clone())
                            },
                        )
                        .expect("dispatch");
                    assert_eq!(
                        resolve_hybrid_leg_results(vector_result, lexical_result.expect("joined")),
                        Err(failures.get(a.max(b)).expect("known failure index").clone()),
                        "vector={a} lexical={b} lexical_first={lexical_first}",
                    );
                }
            }
        }
        worker.stop_and_join().expect("join worker");
    }

    #[test]
    fn hybrid_dual_failure_precedence_is_independent_of_leg_completion_order() {
        use crate::fusion::{FusionError, FusionLeg, LegFailureKind};

        let vector_failure = FusionError::Leg {
            leg: FusionLeg::Vector,
            kind: LegFailureKind::Scan,
            detail: "vector failure".to_owned(),
        };
        let lexical_failure = FusionError::Leg {
            leg: FusionLeg::Lexical,
            kind: LegFailureKind::Lexical,
            detail: "lexical failure".to_owned(),
        };
        assert_eq!(
            resolve_hybrid_leg_results::<(), ()>(
                Err(vector_failure.clone()),
                Err(lexical_failure.clone()),
            ),
            Err(vector_failure.clone()),
        );
        assert_eq!(
            resolve_hybrid_leg_results::<(), ()>(
                Err(vector_failure.clone()),
                Err(FusionError::Cancelled { partial: false }),
            ),
            Err(FusionError::Cancelled { partial: false }),
        );
        assert_eq!(
            resolve_hybrid_leg_results::<(), ()>(
                Err(FusionError::Timeout { partial: false }),
                Err(FusionError::ReadCancelled { partial: false }),
            ),
            Err(FusionError::ReadCancelled { partial: false }),
        );
        assert_eq!(
            resolve_hybrid_leg_results::<(), ()>(
                Err(FusionError::Cancelled { partial: false }),
                Err(FusionError::Timeout { partial: false }),
            ),
            Err(FusionError::Timeout { partial: false }),
        );
        assert_eq!(
            resolve_hybrid_leg_results::<(), ()>(
                Err(FusionError::ReadCancelled { partial: false }),
                Err(lexical_failure),
            ),
            Err(FusionError::ReadCancelled { partial: false }),
        );
    }

    #[test]
    fn accounting_errors_report_the_typed_context() {
        let statistics = StoreError::Statistics {
            component: "mapped resident bytes",
            source: std::io::Error::other("mincore failed"),
        };
        assert_eq!(
            statistics.to_string(),
            "store statistics mapped resident bytes: mincore failed"
        );
        assert_eq!(
            statistics.source().map(ToString::to_string),
            Some("mincore failed".to_owned())
        );

        let budget = StoreError::BudgetExceeded {
            needed: 65,
            budget: 64,
            component: "wal",
        };
        assert_eq!(
            budget.to_string(),
            "store wal allocation needs 65 bytes, budget is 64 bytes"
        );
        assert!(budget.source().is_none());

        let allocation = StoreError::AllocationFailed {
            needed: 23,
            component: "snapshot",
        };
        assert_eq!(
            allocation.to_string(),
            "store snapshot allocator rejected 23 bytes"
        );
        assert!(allocation.source().is_none());
    }

    #[test]
    fn open_options_resolve_durability_before_filesystem_mutation() {
        let parent = tempdir().expect("parent directory");
        let store_path = parent.path().join("not-created");
        let options =
            OpenOptions::new().with_durability(DurabilityMode::Attached, CommitTier::Durable);

        let error = match Store::open(&store_path, options) {
            Ok(store) => {
                drop(store);
                panic!("attached durability unexpectedly opened")
            }
            Err(error) => error,
        };

        assert!(matches!(
            error,
            StoreError::Durability(DurabilityPolicyError::AttachedNotYetSupported)
        ));
        assert!(!store_path.exists(), "rejected open created store files");
    }

    #[test]
    fn memory_vfs_writable_open_requires_host_lease_directory() {
        let parent = tempdir().expect("parent directory");
        let store_path = parent.path().join("memory-only-store");
        let dependencies = StoreTestDependencies::new(
            Arc::new(MemoryVfs::new()),
            Arc::new(ManualMonotonicClock::new()),
        );
        let result =
            Store::open_with_test_dependencies(&store_path, OpenOptions::default(), dependencies);
        assert!(
            matches!(result, Err(StoreError::Lock(super::lock::StoreLockError::Io { ref source, .. })) if source.kind() == std::io::ErrorKind::NotFound),
            "writable open must refuse the missing host lease directory"
        );
        assert!(!store_path.exists());
    }

    #[test]
    fn memory_vfs_store_retains_real_reader_lease() {
        let directory = tempdir().expect("host lease directory");
        let dependencies = StoreTestDependencies::new(
            Arc::new(MemoryVfs::new()),
            Arc::new(ManualMonotonicClock::new()),
        );
        let store = Store::open_with_test_dependencies(
            directory.path(),
            OpenOptions::default(),
            dependencies,
        )
        .expect("MemoryVfs store with host lease");
        let snapshot = store.open_snapshot().expect("snapshot view");
        assert!(matches!(
            super::lock::StoreLock::reclaim_exclusive(directory.path()),
            Err(super::lock::StoreLockError::Io { source, .. })
                if source.kind() == std::io::ErrorKind::WouldBlock
        ));
        drop(store);
        assert!(matches!(
            super::lock::StoreLock::reclaim_exclusive(directory.path()),
            Err(super::lock::StoreLockError::Io { source, .. })
                if source.kind() == std::io::ErrorKind::WouldBlock
        ));
        drop(snapshot);
        drop(super::lock::StoreLock::reclaim_exclusive(directory.path()).expect("released lease"));
    }

    #[test]
    fn open_refuses_a_manifest_ahead_of_the_checked_wal() {
        let directory = tempdir().expect("store directory");
        let policy = DurabilityPolicy::new(DurabilityMode::Derived, CommitTier::Ordered)
            .expect("derived policy");
        commit_manifest(
            &StdVfs,
            directory.path(),
            &Manifest {
                #[cfg(feature = "graph-cypher")]
                graph: None,
                generation: 1,
                log_seq: 1,
                segments: Vec::new(),
                epochs: Vec::new(),
                epoch_alias: None,
                schema: Schema::new(Vec::new()).expect("schema"),
            },
            policy,
        )
        .expect("manifest");

        let error = match Store::open(directory.path(), OpenOptions::default()) {
            Ok(store) => {
                drop(store);
                panic!("manifest ahead of missing WAL unexpectedly opened")
            }
            Err(error) => error,
        };

        assert!(matches!(
            error,
            StoreError::Manifest(ManifestError::AheadOfLog {
                snapshot: 1,
                durable: 0
            })
        ));
    }
}

#[cfg(test)]
mod shared_bound_tests;

#[cfg(test)]
mod hybrid_overlap_tests;