fdu-core 0.3.0

//! The scan layer: walking a tree, producing observations, and applying reconciliation.
//!
//! Public scans emit upsert observations, and a revalidation sweep is the diff between
//! what the index believes and what the filesystem says. Both speak the same
//! [`Observation`] vocabulary as the watch layer. A detached one-shot index may consume
//! equivalent parent-first directory groups privately because no observer can see its
//! construction; every later mutation still crosses the shared observation boundary.
//!
//! # Status
//!
//! The portable `read_dir` plus non-following metadata reference is what every listing
//! falls back to. Parallel scans use it on most platforms; on macOS they first try a
//! measured `getattrlistbulk` backend that returns directory entries and stat-tier
//! metadata together. Unsupported filesystems, malformed results, mount points, and
//! firmlinks fail closed to the portable path for the complete containing directory.
//! On Linux with glibc every route that lists a directory (the serial and concurrent
//! walks, revalidation, reconciliation, opened discovery) first tries a reader that
//! lists with raw `getdents64` and stats with `statx` against the listing's descriptor,
//! passing `AT_NO_AUTOMOUNT`; an open or enumeration failure, a malformed record, or a
//! kernel without `statx` falls back the same way, and the fallback's own stats then
//! pass the flag too (`observe_dir_entry`), as does every stat of a path a route
//! verifies by itself (`observe_path`). Only the walk root is resolved through a mount
//! (`root_device`). So an autofs tree answers the same on every route and delivery.
//! Every backend produces the same [`Observation`] contract.

use std::collections::{BTreeMap, BTreeSet, VecDeque};
use std::ffi::{OsStr, OsString};
use std::fmt::Write as _;
use std::fs;
use std::io::Read as _;
use std::path::{Component, Path, PathBuf};

use crate::ApplyStats;
use crate::engine_contract::{
    Attrs, Commit, EntryKind, Error, Observation, ObservationOp, Op, PathExpectation, PathState,
    Result, ScanScope,
};
use crate::execution::TreeRetention;
use crate::index::{
    DetachedIndexBuilder, Index, IndexHandle, ReconcileErrors, ReconcileFinish,
    collect_child_expectations,
};
use crate::query::ScopeAxis;
use crate::stored_state::{ControlTierIdentity, EntryScope, EntryTierIdentity, SnapshotIdentity};

// Keep the FFI exception at the platform boundary. The rest of the engine, including
// every consumer of these observations, remains under the workspace's unsafe-code
// denial.
#[cfg(target_os = "macos")]
#[allow(unsafe_code)]
mod macos_bulk;

// glibc builds only: `libc` defines `struct statx` for glibc, not for default musl.
#[cfg(all(target_os = "linux", target_env = "gnu"))]
#[allow(unsafe_code)]
mod linux_dents;

#[cfg(windows)]
#[allow(unsafe_code)]
mod windows_metadata;

/// How many ops accumulate before an observation is handed to the sink.
///
/// Batching matters for more than syscall economy: consumers coalesce per path within a
/// batch and stat once per batch, and a live UI wants partial results while a large tree
/// is still being walked rather than one delta at the end.
const DEFAULT_BATCH_SIZE: usize = crate::platform_tuning::tuning().batch_size.get();

/// Largest producer batch accepted before work must be published incrementally.
pub const MAX_SCAN_BATCH_SIZE: usize = 64 * 1024;

/// Most changed paths an exclusive parallel reconciliation may defer before applying.
///
/// Workers compare against one immutable index image, so mutations wait until the wave
/// joins. Bounding that change set keeps a churned tree from turning the fast unchanged
/// path into an unbounded allocation; overflow discards the wave and retries through
/// the incremental serial reconciler.
const MAX_DEFERRED_RECONCILE_OPS: usize = MAX_SCAN_BATCH_SIZE;

/// Directories compared against one immutable index baseline before changes are applied.
///
/// The wave is large enough to amortize scoped worker creation and small enough that a
/// changed tree publishes progress throughout a long reconciliation.
const RECONCILE_WAVE_DIRECTORIES: usize =
    crate::platform_tuning::tuning().reconcile_wave_directories.get();

/// Identity of the fixed stat-tier reducer set.
const REDUCERS_FINGERPRINT: u64 = 1;

/// The order directories are visited in.
///
/// This changes *when* observations are produced, never *which* ones: both orders
/// visit every entry exactly once and leave an identical index behind. It therefore
/// stays out of [`ScanScope`] and cannot invalidate a cache, exactly like the worker
/// count.
///
/// The choice only matters to a consumer that reads the index while the walk is still
/// running, and there it matters a great deal.
///
/// # Strength of the guarantee
///
/// **These are scheduling preferences, not strict orders, whenever more than one worker
/// is running** — which is the default.
///
/// The queue is ordered, but the *claims* are not. Workers take directories from the
/// shared queue in the policy's order; a worker that finishes early can enqueue its
/// children and another worker can claim them while a slower worker still holds
/// unfinished work from a shallower level. Nothing releases a level barrier, because
/// a barrier would idle every fast worker at each level boundary and give back most of
/// the parallel producer's win.
///
/// So:
///
/// - With `threads: Some(1)`, [`ScanOrder::BreadthFirst`] is strict: no directory is
///   read before one closer to the root.
/// - With several workers it is *shallow-first*: shallow work is always preferred when
///   a worker chooses, and deeper observations can still interleave.
///
/// That weaker property is what the browser use case actually needs — every top-level
/// subtree starts filling early, so a mid-scan ranking is meaningful — and it is the
/// property the tests pin. A caller that needs strict level order must ask for one
/// worker and pay for it.
#[derive(Clone, Copy, PartialEq, Eq, Debug, Default)]
pub enum ScanOrder {
    /// Shallow directories before deep ones.
    ///
    /// The default, because it is the order whose partial results mean something.
    /// Roll-ups are maintained per directory as the walk proceeds, so a consumer that
    /// looks mid-scan sees top-level totals grow together — bars fill, rankings
    /// converge — instead of one subtree finishing while its siblings read zero.
    /// Interrupting early leaves a usefully complete picture of the top of the tree.
    ///
    /// Under several workers this is a preference rather than a guarantee; see the
    /// type-level note above.
    ///
    /// Note that totals only grow *while an additive walk is running*. Monotonicity
    /// comes from the producer being additive, not from the order — the order decides
    /// which subtrees get to grow early.
    #[default]
    BreadthFirst,
    /// One subtree toward completion before starting the next.
    ///
    /// Lower peak memory, since the frontier is bounded by depth rather than by the
    /// width of a level, and better locality within a subtree. The cost is that
    /// partial results are actively misleading: one child of the root approaches its
    /// final total while its siblings read zero, so anything ranking by size mid-scan
    /// ranks confidently and wrongly. Correct for a caller that only reads the
    /// finished index and wants the smallest footprint.
    ///
    /// Under several workers this too is a preference: several subtrees will be in
    /// flight at once, one per worker.
    DepthFirst,
}

/// Knobs for a scan.
#[derive(Clone, Debug)]
// Four booleans, each an independent admission or observation switch with its own
// semantic-scope consequence, not an enum in disguise: any combination is legal and
// means what its fields say. The lint suspects flag-soup states; this is a config
// surface whose fields are documented one by one.
#[allow(clippy::struct_excessive_bools)]
pub struct ScanConfig {
    /// Maximum relative entry depth to retain. Zero keeps only the index root and `None`
    /// means unlimited.
    pub max_depth: Option<usize>,
    /// Ops per emitted observation. Must be between one and [`MAX_SCAN_BATCH_SIZE`].
    pub batch_size: usize,
    /// Follow symlinks to directories. Off by default: following them turns a tree walk
    /// into a graph walk with cycles, and every surveyed tool defaults to off.
    pub follow_symlinks: bool,
    /// Stay on the filesystem the root lives on.
    pub one_filesystem: bool,
    /// Hidden-component admission, or `None` to retain every component.
    pub hidden: Option<std::sync::Arc<crate::admission::HiddenPolicy>>,
    /// Exclude filesystem objects other than files, directories, and symlinks.
    pub exclude_special: bool,
    /// Directory-reading worker threads.
    ///
    /// A tree walk is a pile of independent, latency-bound directory reads, so it
    /// scales with threads far better than most work does. One means the serial
    /// walker, which stays the reference implementation and the thing every result is
    /// checked against, with two exceptions that take the concurrent walker with one
    /// worker instead: the detached index build, and a transient summary that reads
    /// `.gitignore`, which must deliver each directory's control ahead of its entries
    /// as that build consumes them. [`None`] asks for a bounded default derived from the
    /// machine's available parallelism. The automatic pool starts conservatively and
    /// unlocks more latency-hiding workers only when initial chunk timing identifies a
    /// slow filesystem path.
    ///
    /// This is an operational knob, not a semantic one: it changes how fast the same
    /// observations are produced, never which observations they are. That is why it
    /// stays out of [`ScanScope`] and cannot invalidate a cache.
    pub threads: Option<usize>,
    /// The order directories are visited in. See [`ScanOrder`].
    pub order: ScanOrder,
    /// File-type rules to classify against, or `None` for the ones compiled into fdu.
    ///
    /// Unlike [`Self::threads`] this *is* semantic: a different taxonomy classifies the
    /// same tree differently, which is why its fingerprint rides in [`ScanScope`] and a
    /// change to it invalidates a snapshot. Shared rather than owned because a scan
    /// clones its config per wave and a registry is read-only once built.
    pub types: Option<std::sync::Arc<crate::classify::TypeRegistry>>,
    /// Observe `.gitignore` control files and retain ignore classification.
    ///
    /// On by default on every surface: an [`Index`] from [`crate::open`] or a scan keeps
    /// the exact control state it exposes and a watch maintains -- which entries are
    /// ignored, and the ignored and unignored partitions of every roll-up -- and a one-shot
    /// report from [`crate::prepare_report`] shows the ignored share of every row
    /// (fdu-elnn). It costs a read of every `.gitignore` in the tree. A file past the
    /// [`Self::control_limits`] is refused and named in [`Index::control_coverage`] rather
    /// than ending the scan; a file that cannot be read is an error at its path, which
    /// makes the result partial.
    ///
    /// Off, the scan performs no control-file I/O and retains no control table, and that is
    /// stamped into [`ScanScope`], so an index-returning call never serves a snapshot taken
    /// one way as the other. An [`Index`] built that way answers [`Index::is_ignored`],
    /// [`Index::controls`], and the partition accessors with
    /// [`crate::Error::ControlStateNotObserved`], never with "not ignored", and refuses
    /// control input; a report's rows carry no ignored share, and a selection by ignored
    /// state is refused. The command line spells it `--no-gitignore`.
    ///
    /// An opened root ([`crate::OpenedIndex`]) always observes control state, because its
    /// ignored and unignored partitions are part of what it serves.
    pub read_controls: bool,
    /// Which ignored population shapes retained entries and content candidates.
    pub population: crate::query::IgnoredEntries,
    /// The budget and the line limit `.gitignore` files are applied under, each a size or
    /// unbounded. See [`crate::control::ControlLimits`].
    ///
    /// A source that would take the table past the budget, or that has a line longer than
    /// the line limit, is refused: its rules do not apply, the scan continues with every
    /// size exact, and [`Index::control_coverage`] names it and the limit that fired. The
    /// command line spells these `--gitignore-budget SIZE|all` and
    /// `--gitignore-line-limit SIZE|all`; the Python API spells them `control_budget` and
    /// `control_line_limit`.
    ///
    /// Semantic, like [`Self::read_controls`]: the limits decide which rules apply, so both
    /// are part of [`ScanScope`] and a snapshot taken under other limits is not reused.
    /// Ignored when control state is not observed.
    pub control_limits: crate::control::ControlLimits,
    /// Where to report how much of the walk has been done, or `None` to report nothing.
    ///
    /// An observer rather than a knob: it changes neither which observations a walk
    /// produces nor how it produces them, so it is no part of [`ScanScope`] or of any
    /// snapshot identity, and two configs that differ only here are the same scan.
    /// Honoured by every walker in this module -- the cold scans, the summary fold,
    /// [`revalidate`], and each `reconcile` entry point -- which enter
    /// [`ProgressPhase::Scanning`](crate::ProgressPhase) or
    /// [`ProgressPhase::Revalidating`](crate::ProgressPhase) and add their counts once
    /// per chunk of directories, never per entry. See [`crate::Progress`] for what the
    /// counts mean and what holds when a walk returns.
    pub progress: Option<crate::Progress>,
}

impl Default for ScanConfig {
    fn default() -> Self {
        Self {
            max_depth: None,
            batch_size: DEFAULT_BATCH_SIZE,
            follow_symlinks: false,
            one_filesystem: false,
            hidden: None,
            exclude_special: false,
            threads: None,
            order: ScanOrder::default(),
            types: None,
            read_controls: crate::query::Request::DEFAULTS.read_controls,
            population: crate::query::IgnoredEntries::Include,
            control_limits: crate::query::Request::DEFAULTS.control_limits,
            progress: None,
        }
    }
}

/// Why watching cannot narrow its structural scan boundary, said once for every surface.
/// Ignored population is a supported retained-scope choice because control edits
/// reconcile the governing directory and rebuild that population.
///
/// The CLI used to carry this guidance and the library carried "requires event-scope
/// filtering", which names the implementation rather than the caller's next move -- so a
/// library caller hitting the same wall got jargon and the CLI user got help. Two
/// messages for one rule also drift, and the parity harness could not tell they were the
/// same rule.
///
/// The knobs are named by the calling surface: `--scan-depth` on the command line,
/// `max_depth` through the API. Everything else is identical, so the harness can verify
/// mechanically that both surfaces state the same rule.
pub const WATCH_SCOPE_GUIDANCE: &str = concat!(
    "watching requires full scope and cannot be combined with max_depth or one_filesystem: ",
    "a watcher cannot filter backend events against a narrowed boundary. Selection such as ",
    "depth, include, and modified_since does work while watching, because it filters the ",
    "retained index rather than narrowing the scan"
);

impl ScanConfig {
    /// Classify this scan with `types` and include their derived identity in its scope.
    #[must_use]
    pub fn with_types(mut self, types: std::sync::Arc<crate::classify::TypeRegistry>) -> Self {
        self.types = Some(types);
        self
    }

    /// The file-type rules in effect: the supplied registry, or the compiled default.
    pub fn types(&self) -> &crate::classify::TypeRegistry {
        match &self.types {
            Some(types) => types,
            None => crate::classify::TypeRegistry::compiled(),
        }
    }

    /// Share the file-type rules with an index that retains them.
    pub(crate) fn types_shared(&self) -> std::sync::Arc<crate::classify::TypeRegistry> {
        self.types
            .as_ref()
            .map_or_else(crate::classify::TypeRegistry::compiled_shared, std::sync::Arc::clone)
    }

    /// Hidden-component policy in effect.
    pub fn hidden(&self) -> &crate::admission::HiddenPolicy {
        self.hidden.as_deref().unwrap_or_else(|| crate::admission::HiddenPolicy::keep_all())
    }

    /// Semantic cache identity, excluding operational batching choices.
    ///
    /// Composed from [`Self::snapshot_identity`], so the scope an index records and the
    /// tier identities a snapshot of it carries are one value in two shapes.
    ///
    /// No longer `const`: the type-rule fingerprint is now a property of the registry in
    /// effect rather than a compiled-in constant, which is the whole point of letting a
    /// caller supply one. A snapshot taken under different rules must not be reused.
    pub fn scope(&self) -> ScanScope {
        self.snapshot_identity().scan_scope()
    }

    /// Which entries this scan retains, the part of its scope no `.gitignore` setting
    /// changes.
    pub fn entry_scope(&self) -> EntryScope {
        EntryScope {
            max_depth: self.max_depth,
            follow_symlinks: self.follow_symlinks,
            one_filesystem: self.one_filesystem,
            hidden_fingerprint: self.hidden().fingerprint(),
            exclude_special: self.exclude_special,
            population: self.population,
            control_fingerprint: if self.population == crate::query::IgnoredEntries::Include {
                0
            } else {
                self.control_identity().ignore_rules_fingerprint()
            },
        }
    }

    /// Whether this scan observes `.gitignore` control state, and under which limits.
    ///
    /// The limits are part of the identity only when control state is observed: a scan
    /// that reads no control file applies none, whatever [`Self::control_limits`] says.
    pub fn control_identity(&self) -> ControlTierIdentity {
        if self.read_controls {
            ControlTierIdentity::Observed { limits: self.control_limits }
        } else {
            ControlTierIdentity::NotObserved
        }
    }

    /// The identity of every tier a snapshot of this scan holds.
    pub fn snapshot_identity(&self) -> SnapshotIdentity {
        SnapshotIdentity {
            entries: EntryTierIdentity {
                engine: crate::snapshot::engine_fingerprint(),
                scope: self.entry_scope(),
                type_rules_fingerprint: self.types().fingerprint(),
                reducers_fingerprint: REDUCERS_FINGERPRINT,
            },
            controls: self.control_identity(),
        }
    }

    /// Resolve [`Self::threads`] to the workers active when a scan begins.
    #[cfg(any(target_os = "macos", test))]
    fn worker_threads(&self) -> usize {
        self.worker_pool().initial
    }

    /// Resolve the worker count for immutable-baseline reconciliation waves.
    fn reconciliation_worker_threads(&self) -> usize {
        match self.threads {
            Some(threads) => threads.clamp(1, MAX_SCAN_THREADS),
            None => std::thread::available_parallelism()
                .map_or(1, std::num::NonZero::get)
                .clamp(1, DEFAULT_RECONCILE_THREADS_CAP),
        }
    }

    /// Resolve the initial and maximum worker counts for one scan.
    #[cfg(any(target_os = "macos", test))]
    fn worker_pool(&self) -> WorkerPool {
        self.worker_pool_for(std::thread::available_parallelism().map_or(1, std::num::NonZero::get))
    }

    /// Resolve the worker pool from one captured operating-system parallelism value.
    fn worker_pool_for(&self, available_parallelism: usize) -> WorkerPool {
        match self.threads {
            Some(threads) => WorkerPool::fixed(threads.clamp(1, MAX_SCAN_THREADS)),
            None => automatic_worker_pool(available_parallelism),
        }
    }

    /// The scope axis this build cannot honour, if any.
    ///
    /// The one statement of the capability rule, so it is asked rather than restated.
    /// [`Request::validate`](crate::query::Request::validate) asks it before any stored
    /// state is read, which is what makes a scope this build cannot honour refuse the same
    /// way on every route, every cache policy, and both surfaces; [`Self::validate`] asks
    /// it for the engine-internal callers -- a bound root, a raw scan, an observation --
    /// that never carry a request.
    pub(crate) const fn unsupported_axis(&self) -> Option<ScopeAxis> {
        if self.follow_symlinks {
            return Some(ScopeAxis::FollowSymlinks);
        }
        #[cfg(not(unix))]
        if self.one_filesystem {
            return Some(ScopeAxis::OneFilesystem);
        }
        None
    }

    pub(crate) fn validate(&self) -> Result<()> {
        if !self.read_controls && self.population != crate::query::IgnoredEntries::Include {
            return Err(Error::UnsupportedScanConfig(
                "ignored population requires .gitignore observation",
            ));
        }
        if self.batch_size == 0 || self.batch_size > MAX_SCAN_BATCH_SIZE {
            return Err(Error::UnsupportedScanConfig(
                "batch_size must be nonzero and no greater than MAX_SCAN_BATCH_SIZE",
            ));
        }
        if let Some(axis) = self.unsupported_axis() {
            return Err(Error::UnsupportedScanConfig(axis.reason()));
        }
        Ok(())
    }

    pub(crate) fn validate_for_scope(&self, indexed: ScanScope) -> Result<()> {
        self.validate()?;
        let requested = self.scope();
        if indexed != requested {
            return Err(Error::ScanScopeMismatch { indexed, requested });
        }
        Ok(())
    }

    /// Scope equality, plus the boundary a watcher cannot filter its backend's events
    /// against.
    ///
    /// The rule belongs to the request model, which refuses a watch of a narrowed scope
    /// before anything is opened ([`RequestError::WatchScope`](crate::query::RequestError));
    /// this is the same rule where a watcher is bound without a request -- an opened root
    /// that observes, and each batch the adapter applies -- and it renders the one
    /// guidance string the model renders.
    #[cfg(feature = "watch")]
    pub(crate) fn validate_for_watch_scope(&self, indexed: ScanScope) -> Result<()> {
        self.validate_for_scope(indexed)?;
        if self.max_depth.is_some() || self.one_filesystem {
            return Err(Error::UnsupportedScanConfig(WATCH_SCOPE_GUIDANCE));
        }
        Ok(())
    }
}

impl Default for ScanScope {
    fn default() -> Self {
        ScanConfig::default().scope()
    }
}

/// What a scan did, including the errors it walked past.
///
/// Unreadable directories are skipped rather than aborting the scan — a permission-denied
/// subdirectory should not cost you the other 499,000 files — but they are reported
/// rather than swallowed, so a caller can tell a complete answer from a partial one.
#[derive(Debug, Default)]
pub struct ScanReport {
    /// Directories successfully listed.
    pub dirs_read: u64,
    /// Entries observed, directories included.
    pub entries: u64,
    /// Regular files whose metadata was observed.
    pub files_walked: u64,
    /// Apparent bytes represented by the regular files whose metadata was observed,
    /// saturating at `u64::MAX`.
    ///
    /// A measure of the walk's work, not an answer: a tree whose total no `u64` can hold
    /// is refused where it is counted ([`crate::Error::UnrepresentableTotal`]), and this
    /// tally stops at the bound rather than wrapping or panicking before that refusal.
    pub bytes_walked: u64,
    /// Allocated bytes of those files: what the default size metric counts, and what a
    /// sparse disk image or a clone makes far smaller than their apparent bytes. It
    /// saturates as `bytes_walked` does.
    pub allocated_walked: u64,
    /// Paths that could not be read, with the reason.
    pub errors: Vec<Error>,
    /// Where the walk's time went, summed across workers.
    pub attribution: WalkAttribution,
}

impl ScanReport {
    /// True when every directory in scope was read successfully.
    pub fn is_complete(&self) -> bool {
        self.errors.is_empty()
    }

    /// Fold one worker's share of a parallel walk into the whole-walk report.
    fn absorb(&mut self, other: Self) {
        self.dirs_read += other.dirs_read;
        self.entries += other.entries;
        self.files_walked += other.files_walked;
        self.bytes_walked = self.bytes_walked.saturating_add(other.bytes_walked);
        self.allocated_walked = self.allocated_walked.saturating_add(other.allocated_walked);
        self.errors.extend(other.errors);
        self.attribution.absorb(other.attribution);
    }

    /// Record one successfully stated directory entry.
    fn observe(&mut self, kind: EntryKind, attrs: Attrs) {
        self.entries += 1;
        if kind == EntryKind::File {
            self.files_walked += 1;
            self.bytes_walked = self.bytes_walked.saturating_add(attrs.size);
            self.allocated_walked = self.allocated_walked.saturating_add(attrs.allocated);
        }
    }
}

/// One walker's running share of the progress counters.
///
/// Each walker keeps its own [`ScanReport`]; this remembers how much of that report it
/// has already added to the shared [`crate::Progress`] cells, so each addition is the
/// difference since the last. It lives on the worker's stack beside the report rather
/// than inside it, so a report absorbed into another never carries a stale baseline.
struct ProgressTally<'a> {
    progress: Option<&'a crate::Progress>,
    directories: u64,
    files: u64,
    bytes: u64,
    allocated: u64,
}

impl<'a> ProgressTally<'a> {
    const fn new(progress: Option<&'a crate::Progress>) -> Self {
        Self { progress, directories: 0, files: 0, bytes: 0, allocated: 0 }
    }

    /// Add what `report` has counted since the last call.
    ///
    /// The one `Option` check is the whole cost when no handle is attached. Called once
    /// per chunk of directories a walker hands over, never per entry.
    fn flush(&mut self, report: &ScanReport) {
        let Some(progress) = self.progress else { return };
        let directories = report.dirs_read - self.directories;
        let files = report.files_walked - self.files;
        let bytes = report.bytes_walked - self.bytes;
        let allocated = report.allocated_walked - self.allocated;
        if directories != 0 || files != 0 || bytes != 0 || allocated != 0 {
            progress.add_walked(directories, files, bytes, allocated);
            self.directories = report.dirs_read;
            self.files = report.files_walked;
            self.bytes = report.bytes_walked;
            self.allocated = report.allocated_walked;
        }
    }

    /// Treat everything `report` holds as already added.
    ///
    /// For a walker that continues a report whose counts other workers added themselves.
    fn skip_to(&mut self, report: &ScanReport) {
        self.directories = report.dirs_read;
        self.files = report.files_walked;
        self.bytes = report.bytes_walked;
        self.allocated = report.allocated_walked;
    }
}

/// Normalize filesystem failures before one of the bounded status collectors retains them.
///
/// A walk may encounter the same inaccessible path from several worker paths. The report is
/// already the full, transient set for this pass, so sorting and deduplicating it here avoids
/// allocating or formatting a second unbounded set solely to decide which 64 details survive.
/// I/O causes are keyed by their native root-relative path and the issue category, exactly the
/// cause identity retained by an index. Other engine failures are left distinct: walker errors
/// are I/O failures, and treating arbitrary engine errors as equivalent without constructing
/// their bounded issue representation would lose information.
pub(crate) fn normalize_walk_errors(root: &Path, errors: &mut Vec<Error>) {
    errors.sort_by(|left, right| match (left, right) {
        (
            Error::Io { path: left_path, source: left_source },
            Error::Io { path: right_path, source: right_source },
        ) => left_path
            .strip_prefix(root)
            .unwrap_or(left_path)
            .cmp(right_path.strip_prefix(root).unwrap_or(right_path))
            .then_with(|| {
                walk_issue_kind_rank(left_source).cmp(&walk_issue_kind_rank(right_source))
            }),
        (Error::Io { .. }, _) => std::cmp::Ordering::Less,
        (_, Error::Io { .. }) => std::cmp::Ordering::Greater,
        _ => std::cmp::Ordering::Equal,
    });
    errors.dedup_by(|right, left| match (left, right) {
        (
            Error::Io { path: left_path, source: left_source },
            Error::Io { path: right_path, source: right_source },
        ) => {
            left_path.strip_prefix(root).unwrap_or(left_path)
                == right_path.strip_prefix(root).unwrap_or(right_path)
                && walk_issue_kind_rank(left_source) == walk_issue_kind_rank(right_source)
        }
        _ => false,
    });
}

fn walk_issue_kind_rank(error: &std::io::Error) -> u8 {
    match error.kind() {
        std::io::ErrorKind::PermissionDenied => 0,
        std::io::ErrorKind::NotFound => 1,
        std::io::ErrorKind::InvalidData | std::io::ErrorKind::InvalidInput => 2,
        _ => 5,
    }
}

/// Schema carried by [`ScanDiagnostics`].
///
/// Diagnostics are an opt-in measurement contract rather than stable human output.
/// Consumers must reject an unknown schema instead of guessing that fields retained
/// their meaning.
pub const SCAN_DIAGNOSTICS_SCHEMA: &str = "fdu-scan-diagnostics-v1";

/// Maximum policy-window records retained by one diagnostic scan.
///
/// The bound is on controller evaluations, not filesystem entries. A controller that
/// needs more history must mark the artifact truncated; claim-grade consumers reject
/// that artifact rather than silently analyzing an incomplete policy history.
const MAX_POLICY_TRACE_EVENTS: usize = 256;

/// Opt-in, run-scoped evidence about a filesystem scan.
///
/// Obtain this through [`scan_with_diagnostics`] or
/// [`scan_into_index_with_diagnostics`]. Keeping it out of [`ScanReport`] preserves the
/// existing scan API and keeps ordinary callers off the measurement path entirely.
#[derive(Clone, Debug, PartialEq, Eq)]
pub struct ScanDiagnostics {
    /// Version of this diagnostic contract.
    pub schema: &'static str,
    /// Automatic worker-controller history and queue state.
    pub worker_policy: WorkerPolicyDiagnostics,
    /// Directory-enumeration backends used by this run.
    pub backend: ScanBackendDiagnostics,
}

/// Repository-only controller variants used by the performance evidence probe.
///
/// These variants are not selected by [`scan`] or [`scan_with_diagnostics`]; both keep
/// the shipped one-shot policy. The explicit experimental APIs make candidate behavior
/// measurable without hiding a production change behind an environment variable.
#[doc(hidden)]
#[derive(Clone, Copy, Debug, Default, PartialEq, Eq)]
pub enum WorkerPolicyExperiment {
    /// The production controller: one prefix window and at most one expansion.
    #[default]
    ShippedOneShot,
    /// Re-evaluate independent windows until a slow phase requests the full reserve.
    RepeatedWindows,
    /// Re-evaluate independent windows, gate on useful frontier/handoff backlog, and grow
    /// the pool in stages.
    StagedGatedWindows,
}

/// Final state of the automatic worker controller.
#[derive(Clone, Copy, Debug, PartialEq, Eq)]
pub enum WorkerPolicyOutcome {
    /// The scan did no walking, for example because `max_depth` was zero.
    NotRun,
    /// A fixed pool had no adaptive decision to make.
    Fixed,
    /// The walk ended before an adaptive window became observable.
    Undecided,
    /// The controller measured a window and retained the initial pool.
    Held,
    /// The controller requested and activated the reserve workers.
    ScaledUp,
    /// Slow work was observed only after no useful queued or in-flight work remained.
    HeldNoUsefulWork,
}

/// One controller evaluation over a half-open range of completed entry ordinals.
#[derive(Clone, Debug, PartialEq, Eq)]
pub struct WorkerPolicyWindow {
    /// Monotonic record number within this scan.
    pub sequence: u64,
    /// First completed-entry ordinal represented by this window, inclusive.
    pub start_entry_ordinal: u64,
    /// Ordinal immediately after the last represented entry.
    pub end_entry_ordinal: u64,
    /// Entries contributing to the service-time signal.
    pub observed_entries: u64,
    /// Completed directory claims contributing to the service-time signal.
    pub observed_chunks: u64,
    /// Worker time contributing to the service-time signal.
    pub observed_work_ns: u64,
    /// Derived service time, or null when no entry made the signal observable.
    pub work_ns_per_entry: Option<u64>,
    /// Why `work_ns_per_entry` is null.
    pub work_ns_per_entry_unavailable_reason: Option<&'static str>,
    /// Directories ready to claim when the controller evaluated the window.
    pub ready_directories: usize,
    /// Claimed directories still being processed at that point.
    pub in_flight_directories: usize,
    /// Live worker threads at that point, including workers waiting for a claim.
    pub active_workers: usize,
    /// Observation batches sent but not yet received by the consumer.
    pub handoff_backlog: usize,
    /// Worker target requested by a scale decision.
    pub requested_workers: Option<usize>,
    /// What the controller concluded from this window.
    pub decision: WorkerPolicyDecision,
}

/// Decision represented by a [`WorkerPolicyWindow`].
#[derive(Clone, Copy, Debug, PartialEq, Eq)]
pub enum WorkerPolicyDecision {
    /// The walk ended before the window could support a decision.
    Undecided,
    /// The observed window retained the current pool.
    Hold,
    /// The observed window activated reserve workers.
    ScaleUp,
    /// The trigger fired after all useful work had drained.
    HoldNoUsefulWork,
    /// A complete window held because reserve workers had no useful frontier to claim.
    HoldInsufficientFrontier,
    /// A complete window held because the unbounded handoff backlog was already high.
    HoldHandoffBacklog,
    /// A post-decision observation window remained below the slow threshold.
    ObserveFast,
    /// A post-decision observation window met the slow threshold.
    ObserveSlow,
    /// A trailing partial window carried no new terminal decision.
    Incomplete,
    /// A trailing partial post-decision observation carried no policy decision.
    ObserveIncomplete,
}

/// Worker-controller configuration, trace, and terminal queue state.
#[derive(Clone, Debug, PartialEq, Eq)]
pub struct WorkerPolicyDiagnostics {
    /// Controller variant exercised by this scan.
    pub controller: &'static str,
    /// Parallelism reported by the operating system when the scan began.
    pub available_parallelism: usize,
    /// Workers in the pool before any adaptive decision.
    pub initial_workers: usize,
    /// Hard maximum workers this scan could activate.
    pub maximum_workers: usize,
    /// Entry target for an adaptive window, or null for a fixed pool.
    pub calibration_window_entries: Option<u64>,
    /// Slow-service trigger, or null for a fixed pool.
    pub slow_threshold_ns_per_entry: Option<u64>,
    /// Directory chunks folded into live controller windows.
    pub calibration_chunks: u64,
    /// Entries folded into live controller windows.
    pub calibration_entries: u64,
    /// Worker time folded into live controller windows.
    pub calibration_work_ns: u64,
    /// Expansion messages that caused the consumer to create more workers.
    pub worker_expansions: u64,
    /// Terminal policy outcome.
    pub outcome: WorkerPolicyOutcome,
    /// Explanation when no adaptive evaluation exists.
    pub outcome_reason: Option<&'static str>,
    /// Total worker threads created during the walk.
    pub workers_spawned: usize,
    /// Maximum simultaneously live worker threads, including workers waiting for work.
    pub peak_active_workers: usize,
    /// Ready directories at scan completion; a complete walk must leave zero.
    pub ready_directories_at_finish: usize,
    /// In-flight directories at scan completion; a complete walk must leave zero.
    pub in_flight_directories_at_finish: usize,
    /// Observation batches outstanding at scan completion.
    pub handoff_backlog_at_finish: usize,
    /// Maximum outstanding observation batches during the scan.
    pub handoff_backlog_high_water: usize,
    /// Bounded controller history.
    pub windows: Vec<WorkerPolicyWindow>,
    /// True when controller history exceeded the 256-event diagnostic bound.
    pub events_truncated: bool,
}

/// Directory enumeration backends used by one scan.
///
/// Each native reader's listings are counted in its own fields, and a directory it
/// declines is counted as its fallback and then as a portable attempt. On Linux, for the
/// native reader (`getdents64` and `statx`, glibc builds):
///
/// - `linux_dents_attempts` = `linux_dents_successes` + `linux_dents_fallbacks`;
/// - the report's `dirs_read` = `linux_dents_successes` + `portable_directory_reads`.
///
/// `unavailable_reason` describes only the macOS fields.
///
/// Non-exhaustive, as [`crate::counters::Counts`] is: a diagnostics record grows with the
/// backends it describes, so a later field is an additive change. Code outside the engine
/// reads its fields; only the engine builds one.
#[derive(Clone, Debug, PartialEq, Eq)]
#[non_exhaustive]
pub struct ScanBackendDiagnostics {
    /// Portable `read_dir` calls attempted.
    pub portable_attempts: u64,
    /// Portable directory listings completed successfully.
    pub portable_directory_reads: u64,
    /// macOS bulk enumeration attempts, or null off macOS.
    pub macos_bulk_attempts: Option<u64>,
    /// Successful macOS bulk listings, or null off macOS.
    pub macos_bulk_successes: Option<u64>,
    /// Bulk attempts that fell back to portable enumeration, or null off macOS.
    pub macos_bulk_fallbacks: Option<u64>,
    /// Why macOS fields are null.
    pub unavailable_reason: Option<&'static str>,
    /// Linux native `getdents64` listing attempts, or null off Linux (and on a Linux
    /// build without glibc, where the native reader is not compiled).
    pub linux_dents_attempts: Option<u64>,
    /// Successful Linux native listings, or null off Linux.
    pub linux_dents_successes: Option<u64>,
    /// Linux native attempts that fell back to portable enumeration, or null off Linux.
    pub linux_dents_fallbacks: Option<u64>,
}

impl ScanDiagnostics {
    /// Serialize this versioned diagnostic contract as compact JSON.
    ///
    /// This deliberately lives beside the contract instead of in a benchmark binary:
    /// claim-grade installed-command measurements and the repository probe must emit
    /// byte-for-byte equivalent evidence without adding a serialization dependency to
    /// the core crate.
    pub fn to_json(&self) -> String {
        let policy = &self.worker_policy;
        let backend = &self.backend;
        let mut windows = String::from("[");
        for (index, window) in policy.windows.iter().enumerate() {
            if index > 0 {
                windows.push(',');
            }
            let _ = write!(
                windows,
                concat!(
                    "{{\"active_workers\":{},\"decision\":\"{}\",",
                    "\"end_entry_ordinal\":{},\"handoff_backlog\":{},",
                    "\"in_flight_directories\":{},\"observed_chunks\":{},",
                    "\"observed_entries\":{},",
                    "\"observed_work_ns\":{},\"ready_directories\":{},",
                    "\"requested_workers\":{},\"sequence\":{},\"start_entry_ordinal\":{},",
                    "\"work_ns_per_entry\":{},",
                    "\"work_ns_per_entry_unavailable_reason\":{}}}"
                ),
                window.active_workers,
                worker_policy_decision_name(window.decision),
                window.end_entry_ordinal,
                window.handoff_backlog,
                window.in_flight_directories,
                window.observed_chunks,
                window.observed_entries,
                window.observed_work_ns,
                window.ready_directories,
                json_optional_usize(window.requested_workers),
                window.sequence,
                window.start_entry_ordinal,
                json_optional_u64(window.work_ns_per_entry),
                json_optional_string(window.work_ns_per_entry_unavailable_reason),
            );
        }
        windows.push(']');
        format!(
            concat!(
                "{{\"backend\":{{\"linux_dents_attempts\":{},",
                "\"linux_dents_fallbacks\":{},\"linux_dents_successes\":{},",
                "\"macos_bulk_attempts\":{},",
                "\"macos_bulk_fallbacks\":{},\"macos_bulk_successes\":{},",
                "\"portable_attempts\":{},\"portable_directory_reads\":{},",
                "\"unavailable_reason\":{}}},",
                "\"schema\":\"{}\",\"worker_policy\":{{",
                "\"available_parallelism\":{},\"calibration_chunks\":{},",
                "\"calibration_entries\":{},\"calibration_window_entries\":{},",
                "\"calibration_work_ns\":{},",
                "\"controller\":\"{}\",",
                "\"events_truncated\":{},\"handoff_backlog_at_finish\":{},",
                "\"handoff_backlog_high_water\":{},\"in_flight_directories_at_finish\":{},",
                "\"initial_workers\":{},\"maximum_workers\":{},\"outcome\":\"{}\",",
                "\"outcome_reason\":{},\"peak_active_workers\":{},",
                "\"ready_directories_at_finish\":{},\"slow_threshold_ns_per_entry\":{},",
                "\"windows\":{},\"worker_expansions\":{},\"workers_spawned\":{}}}}}"
            ),
            json_optional_u64(backend.linux_dents_attempts),
            json_optional_u64(backend.linux_dents_fallbacks),
            json_optional_u64(backend.linux_dents_successes),
            json_optional_u64(backend.macos_bulk_attempts),
            json_optional_u64(backend.macos_bulk_fallbacks),
            json_optional_u64(backend.macos_bulk_successes),
            backend.portable_attempts,
            backend.portable_directory_reads,
            json_optional_string(backend.unavailable_reason),
            self.schema,
            policy.available_parallelism,
            policy.calibration_chunks,
            policy.calibration_entries,
            json_optional_u64(policy.calibration_window_entries),
            policy.calibration_work_ns,
            policy.controller,
            policy.events_truncated,
            policy.handoff_backlog_at_finish,
            policy.handoff_backlog_high_water,
            policy.in_flight_directories_at_finish,
            policy.initial_workers,
            policy.maximum_workers,
            worker_policy_outcome_name(policy.outcome),
            json_optional_string(policy.outcome_reason),
            policy.peak_active_workers,
            policy.ready_directories_at_finish,
            json_optional_u64(policy.slow_threshold_ns_per_entry),
            windows,
            policy.worker_expansions,
            policy.workers_spawned,
        )
    }
}

const fn worker_policy_outcome_name(value: WorkerPolicyOutcome) -> &'static str {
    match value {
        WorkerPolicyOutcome::NotRun => "not_run",
        WorkerPolicyOutcome::Fixed => "fixed",
        WorkerPolicyOutcome::Undecided => "undecided",
        WorkerPolicyOutcome::Held => "held",
        WorkerPolicyOutcome::ScaledUp => "scaled_up",
        WorkerPolicyOutcome::HeldNoUsefulWork => "held_no_useful_work",
    }
}

const fn worker_policy_decision_name(value: WorkerPolicyDecision) -> &'static str {
    match value {
        WorkerPolicyDecision::Undecided => "undecided",
        WorkerPolicyDecision::Hold => "hold",
        WorkerPolicyDecision::ScaleUp => "scale_up",
        WorkerPolicyDecision::HoldNoUsefulWork => "hold_no_useful_work",
        WorkerPolicyDecision::HoldInsufficientFrontier => "hold_insufficient_frontier",
        WorkerPolicyDecision::HoldHandoffBacklog => "hold_handoff_backlog",
        WorkerPolicyDecision::ObserveFast => "observe_fast",
        WorkerPolicyDecision::ObserveSlow => "observe_slow",
        WorkerPolicyDecision::Incomplete => "incomplete",
        WorkerPolicyDecision::ObserveIncomplete => "observe_incomplete",
    }
}

fn json_optional_string(value: Option<&str>) -> String {
    value.map_or_else(|| "null".into(), |value| format!("\"{}\"", json_escape(value)))
}

fn json_optional_u64(value: Option<u64>) -> String {
    value.map_or_else(|| "null".into(), |value| value.to_string())
}

fn json_optional_usize(value: Option<usize>) -> String {
    value.map_or_else(|| "null".into(), |value| value.to_string())
}

fn json_escape(value: &str) -> String {
    let mut escaped = String::new();
    for character in value.chars() {
        match character {
            '"' => escaped.push_str("\\\""),
            '\\' => escaped.push_str("\\\\"),
            '\u{08}' => escaped.push_str("\\b"),
            '\u{0c}' => escaped.push_str("\\f"),
            '\n' => escaped.push_str("\\n"),
            '\r' => escaped.push_str("\\r"),
            '\t' => escaped.push_str("\\t"),
            character if character <= '\u{1f}' => {
                let _ = write!(escaped, "\\u{:04x}", u32::from(character));
            }
            character => escaped.push(character),
        }
    }
    escaped
}

/// Where a walk's time went, so "blocked" is never one undifferentiated number.
///
/// The performance loop's standing question is whether a walk is bound by disk I/O,
/// by CPU, or by coordination, and process-level counters cannot answer it: user and
/// system time say how much CPU was burned, but a fused "blocked" number cannot say
/// whether workers were waiting on the filesystem, on the queue lock, or on nothing
/// at all because the queue was empty. These counters split that out at the source.
///
/// Everything is measured in *chunks*, never per file: one timing pair per claimed
/// run of directories, per contended lock, per batch handoff. On the 60k-entry
/// reference tree that is a few thousand `Instant` reads against hundreds of
/// milliseconds of walking — the instrumentation follows the same amortization rule
/// it exists to verify.
///
/// In a parallel walk the fields sum over workers, so `wall_ns` is worker-seconds
/// (it can exceed the scan's wall clock) and every other duration is a disjoint
/// slice of it: `work_ns + starved_ns + lock_wait_ns + send_ns <= wall_ns`, with the
/// remainder being uninstrumented odds and ends (uncontended lock ops, loop
/// bookkeeping). A serial walk fills only `wall_ns`, `work_ns`, and `send_ns` —
/// there is no coordination to attribute.
#[derive(Debug, Default, Clone, Copy, PartialEq, Eq)]
pub struct WalkAttribution {
    /// Total time workers spent in the walk loop, summed across workers.
    pub wall_ns: u64,
    /// Reading directories and stating entries — the real work, syscalls plus the
    /// compute between them. Separating disk from CPU *within* this span needs the
    /// process-level user/system counters alongside; per-syscall timing would break
    /// the chunk-amortization rule.
    pub work_ns: u64,
    /// Waiting on the queue's condvar because no work was available. Starvation:
    /// either the frontier is momentarily narrower than the worker pool, or the walk
    /// is ending.
    pub starved_ns: u64,
    /// Waiting to acquire the queue lock when another worker held it. This is the
    /// contention the shared-queue design bets stays negligible; now it is measured
    /// instead of argued.
    pub lock_wait_ns: u64,
    /// Handing observation batches to the consumer: the channel send in a parallel
    /// walk, the inline sink call — which is the consumer actually running — in a
    /// serial one.
    pub send_ns: u64,
    /// Chunks of directories claimed from the queue.
    pub claims: u64,
    /// Queue lock acquisitions, contended or not.
    pub lock_ops: u64,
    /// Lock acquisitions that found the lock already held.
    pub lock_contended: u64,
}

impl WalkAttribution {
    /// Fold one worker's counters into the whole-walk totals.
    fn absorb(&mut self, other: Self) {
        self.wall_ns += other.wall_ns;
        self.work_ns += other.work_ns;
        self.starved_ns += other.starved_ns;
        self.lock_wait_ns += other.lock_wait_ns;
        self.send_ns += other.send_ns;
        self.claims += other.claims;
        self.lock_ops += other.lock_ops;
        self.lock_contended += other.lock_contended;
    }

    /// Time attributed to a named cause, as opposed to `wall_ns`'s total.
    pub fn accounted_ns(&self) -> u64 {
        self.work_ns + self.starved_ns + self.lock_wait_ns + self.send_ns
    }
}

/// Filesystem and index effects from an applying reconciliation pass.
#[derive(Debug, Default)]
pub struct ReconcileReport {
    /// Filesystem walk effects and partial errors.
    ///
    /// [`ScanReport::attribution`] remains zero for reconciliation because neither the
    /// serial nor parallel path has complete, comparable instrumentation yet. Zero
    /// means "not measured" here, not "no work".
    pub scan: ScanReport,
    /// Index arbitration and mutation effects.
    pub apply: ApplyStats,
    /// Exact producer operations considered, including no-op controls that do not
    /// increment an effect counter or create a commit.
    pub(crate) observations: u64,
    /// Directories this pass listed in full, with no error inside them, that the index did
    /// not yet hold as complete.
    ///
    /// The closing commit records each one's child set as authoritative, as discovery's
    /// own listing commit does, whether or not the rest of the pass completed: one transient
    /// child error
    /// elsewhere used to keep every directory the pass listed incomplete, and a directory
    /// first listed by such a pass stayed `Unknown { Building }` under a complete root.
    pub(crate) listed_incomplete: Vec<PathBuf>,
    /// Ownership epoch for conditional reconciliation batches.
    reconcile_epoch: Option<u64>,
    /// Retry after bounded verification evidence was superseded.
    retry_required: bool,
}

impl ReconcileReport {
    /// True when the filesystem walk was complete and no conditional observation lost
    /// a race with another producer.
    pub fn is_complete(&self) -> bool {
        self.scan.is_complete()
            && self.apply.stale == 0
            && self.apply.resource_refused == 0
            && !self.retry_required
    }

    /// True when a newer verification retired this pass's bounded evidence before it
    /// closed, so its scope is published partial and must be walked again.
    pub(crate) const fn retry_required(&self) -> bool {
        self.retry_required
    }

    /// The directories whose listings this pass can vouch for, taken out of the report.
    ///
    /// None when a conditional commit lost a race or was refused: a child of any listed
    /// directory may then be missing from the index until the retry that race earns, and
    /// the retry records completeness for what it lists.
    pub(crate) fn take_recordable_completeness(&mut self) -> Vec<PathBuf> {
        let listed = std::mem::take(&mut self.listed_incomplete);
        if self.apply.stale > 0 || self.apply.resource_refused > 0 { Vec::new() } else { listed }
    }
}

enum ReconcileTarget<'a> {
    Direct(&'a mut Index),
    Shared(&'a IndexHandle),
    Controlled { handle: &'a IndexHandle, control: &'a dyn ReconcileControl },
}

/// Lifecycle checkpoints used by an owned long-running reconciliation.
///
/// The ordinary one-shot APIs use no controller. An [`crate::OpenedIndex`] supplies one
/// so close can stop a refresh before another write, and deterministic tests can pause
/// after filesystem verification but before conditional arbitration.
pub(crate) trait ReconcileControl {
    /// Fail when the owning operation may no longer publish state.
    fn check_active(&self) -> Result<()>;

    /// Boundary after filesystem verification and before a conditional fact commit.
    fn before_conditional_commit(&self) -> Result<()>;

    /// Atomic file-retention limit shared with every producer for this opened root.
    fn max_files(&self) -> Option<u64>;
}

impl ReconcileTarget<'_> {
    fn scope(&self) -> Result<ScanScope> {
        match self {
            Self::Direct(index) => Ok(index.scope()),
            Self::Shared(handle) | Self::Controlled { handle, .. } => handle.scope(),
        }
    }

    fn root_path(&self) -> Result<PathBuf> {
        match self {
            Self::Direct(index) => Ok(index.root_path().to_path_buf()),
            Self::Shared(handle) | Self::Controlled { handle, .. } => handle.root_path(),
        }
    }

    fn expectation(&self, path: &Path) -> Result<PathExpectation> {
        match self {
            Self::Direct(index) => Ok(index.expectation(path)),
            Self::Shared(handle) | Self::Controlled { handle, .. } => handle.expectation(path),
        }
    }

    fn child_states(&self, path: &Path) -> Result<BTreeMap<OsString, PathExpectation>> {
        match self {
            Self::Direct(index) => Ok(collect_child_expectations(index, path)),
            Self::Shared(handle) | Self::Controlled { handle, .. } => handle.child_states(path),
        }
    }

    /// Child baselines for one directory listing, and whether a complete listing of it
    /// would be news to the index's directory completeness.
    ///
    /// A directory whose upsert has not been flushed yet is not held at all and counts as
    /// incomplete.
    fn listing_baseline(&self, path: &Path) -> Result<(BTreeMap<OsString, PathExpectation>, bool)> {
        match self {
            Self::Direct(index) => Ok((
                collect_child_expectations(index, path),
                index.directory_complete(path) != Some(true),
            )),
            Self::Shared(handle) | Self::Controlled { handle, .. } => handle.listing_baseline(path),
        }
    }

    fn has_control(&self, path: &Path) -> Result<bool> {
        match self {
            Self::Direct(index) => Ok(index.control_table().contains(path)),
            Self::Shared(handle) | Self::Controlled { handle, .. } => handle.has_control(path),
        }
    }

    fn control_table(&self) -> Result<crate::control::ControlTable> {
        match self {
            Self::Direct(index) => Ok(index.control_table().clone()),
            Self::Shared(handle) | Self::Controlled { handle, .. } => {
                handle.read_with(|index| index.control_table().clone())
            }
        }
    }

    fn control_classification_known(&self, path: &Path) -> Result<bool> {
        match self {
            Self::Direct(index) => Ok(index.control_classification_known(path)),
            Self::Shared(handle) | Self::Controlled { handle, .. } => {
                handle.read_with(|index| index.control_classification_known(path))
            }
        }
    }

    fn apply(&mut self, started_at: u64, observation: &Observation) -> Result<crate::ApplyOutcome> {
        match self {
            Self::Direct(index) => index.apply(observation),
            Self::Shared(handle) => handle.apply_reconcile(started_at, observation),
            Self::Controlled { handle, control } => {
                control.before_conditional_commit()?;
                handle.apply_opened_reconcile(started_at, observation, control.max_files())
            }
        }
    }

    fn direct_upsert_is_unchanged(
        &self,
        baseline: PathExpectation,
        kind: EntryKind,
        attrs: Attrs,
    ) -> bool {
        matches!(self, Self::Direct(_)) && baseline.state == (PathState::Present { kind, attrs })
    }

    fn take_pending_invalidations(&mut self) -> Result<Vec<(PathBuf, crate::InvalidateReason)>> {
        match self {
            Self::Direct(index) => Ok(index.take_pending_invalidations()),
            Self::Shared(handle) | Self::Controlled { handle, .. } => {
                handle.take_pending_invalidations()
            }
        }
    }

    fn restore_pending_invalidations(
        &mut self,
        invalidations: Vec<(PathBuf, crate::InvalidateReason)>,
    ) -> Result<()> {
        match self {
            Self::Direct(index) => index.restore_pending_invalidations(invalidations),
            Self::Shared(handle) | Self::Controlled { handle, .. } => {
                handle.restore_pending_invalidations(invalidations)?;
            }
        }
        Ok(())
    }

    /// Whether an invalidation whose reconciliation came back incomplete is queued again.
    ///
    /// A caller of the one-shot API owns its index exclusively and drains the queue when it
    /// chooses, so an unreadable subtree stays queued for it to retry. The shared API and an
    /// opened root are drained after every observed event -- by `Watcher::apply_next` and by
    /// the opened root's observer -- where that retry is a full walk of the same unreadable
    /// subtree per unrelated event, for the life of the session. There only a lost race is
    /// worth retrying: a stale conditional commit, or one the budget refused. A scan error
    /// is a settled boundary: the subtree stays partial, as it does at the observation
    /// handoff, and the report names the error once.
    fn retries_incomplete(&self, report: &ReconcileReport) -> bool {
        match self {
            Self::Direct(_) => !report.is_complete(),
            Self::Shared(_) | Self::Controlled { .. } => {
                report.apply.stale > 0 || report.apply.resource_refused > 0 || report.retry_required
            }
        }
    }

    fn begin_reconcile(&mut self, path: &Path) -> Result<(u64, Option<Commit>)> {
        match self {
            Self::Direct(index) => index.begin_reconcile(path),
            Self::Shared(handle) => handle.begin_reconcile(path),
            Self::Controlled { handle, control } => {
                control.check_active()?;
                handle.begin_reconcile(path)
            }
        }
    }

    fn finish_reconcile(
        &mut self,
        path: &Path,
        started_at: u64,
        complete: bool,
        listed_incomplete: &[PathBuf],
        failed_paths: &[PathBuf],
        errors: ReconcileErrors<'_>,
    ) -> Result<ReconcileFinish> {
        match self {
            Self::Direct(index) => index.finish_reconcile(
                path,
                started_at,
                complete,
                listed_incomplete,
                failed_paths,
                errors,
            ),
            Self::Shared(handle) => handle.finish_reconcile(
                path,
                started_at,
                complete,
                listed_incomplete,
                failed_paths,
                errors,
            ),
            Self::Controlled { handle, control } => {
                control.check_active()?;
                handle.finish_reconcile(
                    path,
                    started_at,
                    complete,
                    listed_incomplete,
                    failed_paths,
                    errors,
                )
            }
        }
    }
}

#[cfg(unix)]
pub(crate) fn metadata_for_fingerprint(entry: &fs::DirEntry) -> std::io::Result<fs::Metadata> {
    crate::counters::bump(|c| c.stats += 1);
    entry.metadata()
}

#[cfg(any(all(windows, test), not(any(unix, windows))))]
pub(crate) fn metadata_for_fingerprint(entry: &fs::DirEntry) -> std::io::Result<fs::Metadata> {
    crate::counters::bump(|c| c.stats += 1);
    // Windows serves DirEntry metadata from directory-enumeration data, which the
    // platform permits to be stale. Fingerprints need a fresh non-following query.
    fs::symlink_metadata(entry.path())
}

#[cfg(test)]
type WalkHook = std::sync::Arc<dyn Fn(WalkHookPoint<'_>) -> Option<std::io::Error> + Send + Sync>;

/// Where a test hook runs in a listing walk.
#[cfg(test)]
#[derive(Clone, Copy, Debug)]
pub(crate) enum WalkHookPoint<'a> {
    /// Before the metadata lookup of the listed child at this absolute path; an error
    /// stands in for the lookup's.
    ChildMetadata(&'a Path),
    /// After a reconciliation's listing of a directory returns its last entry; an error is
    /// read as one more listing item, which leaves the listing incomplete.
    ListingEnd,
    /// After a lookup of a directory's canonical control path, at this absolute path, has
    /// returned and before its answer is used; an error stands in for that answer.
    ControlLookup(&'a Path),
}

/// Hooks run at each [`WalkHookPoint`], each for the paths under its root.
///
/// Process-wide, because a parallel walk looks children up on its worker threads; keyed
/// by root, because tests run in parallel and each walks its own temporary directory.
#[cfg(test)]
static WALK_HOOKS: std::sync::RwLock<Vec<(Vec<PathBuf>, WalkHook)>> =
    std::sync::RwLock::new(Vec::new());

/// Removes its hook from [`WALK_HOOKS`] when dropped.
#[cfg(test)]
#[must_use = "the hook is removed as soon as the guard is dropped"]
pub(crate) struct WalkHookGuard(WalkHook);

#[cfg(test)]
impl Drop for WalkHookGuard {
    fn drop(&mut self) {
        WALK_HOOKS
            .write()
            .unwrap_or_else(std::sync::PoisonError::into_inner)
            .retain(|(_, hook)| !std::sync::Arc::ptr_eq(hook, &self.0));
    }
}

/// Run `hook` at every [`WalkHookPoint`] under `root`, on any thread, until the guard drops.
///
/// The hook may also change the tree before it returns. `root` matches as given and
/// canonical, since an opened root and a detached scan walk the canonical path.
#[cfg(test)]
pub(crate) fn install_walk_hook(
    root: &Path,
    hook: impl Fn(WalkHookPoint<'_>) -> Option<std::io::Error> + Send + Sync + 'static,
) -> WalkHookGuard {
    let mut roots = vec![root.to_path_buf()];
    if let Ok(canonical) = root.canonicalize() {
        roots.push(canonical);
    }
    let hook: WalkHook = std::sync::Arc::new(hook);
    WALK_HOOKS
        .write()
        .unwrap_or_else(std::sync::PoisonError::into_inner)
        .push((roots, std::sync::Arc::clone(&hook)));
    WalkHookGuard(hook)
}

/// Run `hook` before every listed child's metadata lookup under `root`, with the child's
/// absolute path, until the guard drops.
#[cfg(test)]
pub(crate) fn install_child_metadata_hook(
    root: &Path,
    hook: impl Fn(&Path) -> Option<std::io::Error> + Send + Sync + 'static,
) -> WalkHookGuard {
    install_walk_hook(root, move |point| match point {
        WalkHookPoint::ChildMetadata(path) => hook(path),
        WalkHookPoint::ListingEnd | WalkHookPoint::ControlLookup(_) => None,
    })
}

/// The hook installed for a root containing `path`, if any.
#[cfg(test)]
fn walk_hook(path: &Path) -> Option<WalkHook> {
    WALK_HOOKS
        .read()
        .unwrap_or_else(std::sync::PoisonError::into_inner)
        .iter()
        .find(|(roots, _)| roots.iter().any(|root| path.starts_with(root)))
        .map(|(_, hook)| std::sync::Arc::clone(hook))
}

/// A reconciliation's listing of `dir`, followed by any error a test hook injects.
///
/// Callers bind the result to `listing` and iterate it as `for … in listing`, because the
/// admission audit (`scripts/check-admission-sites.mjs`) counts routed listing loops by that
/// shape. Keep the binding when editing a call site; dropping it silently removes the loop
/// from the audit, whose expected count would then look too high rather than wrong.
#[cfg(test)]
fn reconcile_listing<'r>(listing: Listing<'r>, dir: &Path) -> impl Iterator<Item = Listed<'r>> {
    let injected = walk_hook(dir).and_then(|hook| hook(WalkHookPoint::ListingEnd));
    listing.chain(injected.map(Listed::Failed))
}

/// A reconciliation's listing of `dir`.
#[cfg(not(test))]
fn reconcile_listing<'r>(listing: Listing<'r>, _dir: &Path) -> Listing<'r> {
    listing
}

/// The error a test hook injects for the metadata lookup of the listed child at `path`.
#[cfg(all(test, not(windows)))]
fn child_metadata_hook(path: &Path) -> Option<std::io::Error> {
    walk_hook(path).and_then(|hook| hook(WalkHookPoint::ChildMetadata(path)))
}

/// Whether std's stat of a listed child, `DirEntry::file_type` of a `DT_UNKNOWN` entry
/// included, is one the kernel treats as `AT_NO_AUTOMOUNT`.
///
/// It is `fstatat` everywhere but glibc, where it is `statx` without the flag wherever
/// `statx` is served, and `fstatat` only once the reader has found it unavailable
/// (`linux_dents`).
#[cfg(all(target_os = "linux", target_env = "gnu"))]
fn std_child_stat_never_automounts() -> bool {
    linux_dents::statx_unavailable()
}

#[cfg(not(any(windows, all(target_os = "linux", target_env = "gnu"))))]
const fn std_child_stat_never_automounts() -> bool {
    true
}

/// Kind and attributes for one listed child.
///
/// The transient summary fold counts directories and ignores symlink attributes, so a
/// listing `file_type` (`d_type` on Linux) is enough for those kinds when the walk is
/// not bound to one filesystem. Files and specials still need a metadata lookup for
/// size, allocated bytes, and mtime. `one_filesystem` still stats directories because
/// descent compares `attrs.dev` to the root device, and `dev == 0` would otherwise
/// cross a mount. The listing's kind is taken only once the listing has proved its
/// directory searchable ([`Searchability`]); until then every child is stated, and the
/// skip applies to what the stat found.
///
/// Where `d_type` is `DT_UNKNOWN` (XFS without `ftype`, some FUSE/NFS mounts, older
/// ext3), std's `file_type` performs the non-following stat itself, and the skip then
/// applies to the kind it found. On glibc that stat would trigger an automount, so
/// while `statx` is served the listing's `file_type` is not consulted at all: every
/// child is stated by [`observe_dir_entry`], which never automounts, and the skip
/// applies to what that stat found, as it does to a `DT_UNKNOWN` entry. That route is
/// the fallback for a directory the native reader declined; the reader itself keeps
/// the `d_type` skip.
///
/// Windows never takes the skip: its observation contract reads every listed entry
/// through a fresh non-following handle ([`observe_dir_entry`]), so the transient fold
/// there performs exactly the observations the retained walk performs.
fn listed_child_kind_and_attrs(
    entry: &fs::DirEntry,
    policy: ListingPolicy,
    searchability: &mut Searchability,
) -> std::io::Result<Option<(EntryKind, Attrs)>> {
    #[cfg(not(windows))]
    {
        let listing_kind_suffices = policy.skip_dir_symlink_stat
            && *searchability == Searchability::Proven
            && std_child_stat_never_automounts();
        if listing_kind_suffices {
            if let Ok(file_type) = entry.file_type() {
                if file_type.is_dir() && !policy.one_filesystem {
                    return Ok(Some((EntryKind::Dir, Attrs::default())));
                }
                if file_type.is_symlink() {
                    return Ok(Some((EntryKind::Symlink, Attrs::default())));
                }
            }
        }
        let observed = observe_dir_entry(entry)?;
        if observed.is_some() {
            *searchability = Searchability::Proven;
        }
        if policy.skip_dir_symlink_stat && !listing_kind_suffices {
            return Ok(observed.map(|(kind, attrs)| match kind {
                EntryKind::Dir if !policy.one_filesystem => (kind, Attrs::default()),
                EntryKind::Symlink => (kind, Attrs::default()),
                EntryKind::Dir | EntryKind::File | EntryKind::Other => (kind, attrs),
            }));
        }
        Ok(observed)
    }
    #[cfg(windows)]
    {
        // Windows observes every listed entry through a fresh handle on both routes, so
        // the skip does not apply there; a successful observation proves the directory
        // searchable all the same, so the listing's state means the same on every host.
        let _ = (policy.skip_dir_symlink_stat, policy.one_filesystem);
        let observed = observe_dir_entry(entry)?;
        if observed.is_some() {
            *searchability = Searchability::Proven;
        }
        Ok(observed)
    }
}

/// How a listing observes its children.
#[derive(Clone, Copy, Debug)]
pub(crate) struct ListingPolicy {
    /// H72: directory and symlink kinds come from the listing without a stat, once the
    /// listing has proved its directory searchable ([`Searchability`]).
    pub(crate) skip_dir_symlink_stat: bool,
    /// Descent compares `attrs.dev`, so directories are stated even under the skip.
    pub(crate) one_filesystem: bool,
}

/// Whether one listing has proved its directory searchable, which the skip of
/// [`ListingPolicy::skip_dir_symlink_stat`] requires before it takes a child's kind from
/// the listing alone. Per listing: it starts unproven with each directory.
///
/// A stat is also an observation of failure. A directory that is readable but not
/// searchable (mode `0400`) opens and lists, and then every child's stat fails with
/// `EACCES`; a walk that stats every child reports each child as an error and holds no
/// entry for it. A `DT_DIR` child admitted on its `d_type` alone would be a directory
/// entry the full index does not have, queued and then reported again when it fails to
/// open, and a `DT_LNK` child a symlink entry with no error at all, so the folded tree
/// and the summary would count one directory more and report one error fewer than the
/// full index (R163-1). Under the skip, therefore, every child is stated until one stat
/// in the listing succeeds, which proves the directory searchable, as the skip assumes;
/// only then are directory and symlink kinds taken from the listing. The cost is one
/// stat per listing whose first children are directories or symlinks. A child that
/// vanished before its stat proves nothing here, which costs a stat and never an entry.
#[derive(Clone, Copy, Debug, Default, PartialEq, Eq)]
pub(crate) enum Searchability {
    /// No child's stat has succeeded yet, so every child is stated.
    #[default]
    Unproven,
    /// A child's stat succeeded, so the listing's kind stands for a directory or symlink.
    Proven,
}

impl ListingPolicy {
    /// Every child stated: what the retained index, a reconciliation, and discovery need.
    pub(crate) const fn every_child_stated(one_filesystem: bool) -> Self {
        Self { skip_dir_symlink_stat: false, one_filesystem }
    }
}

/// The platform readers a listing loop keeps between the directories it lists.
///
/// One per walker or reconciliation worker: the Linux reader holds a 64 KiB record
/// buffer that each of its listings borrows for the directory's duration. Elsewhere
/// there is nothing to keep, and every listing is portable.
pub(crate) struct Readers {
    #[cfg(all(target_os = "linux", target_env = "gnu"))]
    dents: linux_dents::Reader,
}

impl Readers {
    pub(crate) fn new() -> Self {
        Self {
            #[cfg(all(target_os = "linux", target_env = "gnu"))]
            dents: linux_dents::Reader::new(),
        }
    }
}

/// One item of a directory listing, whichever backend served it.
pub(crate) enum Listed<'a> {
    /// The listing failed to yield an item, so the directory is listed incompletely.
    Failed(std::io::Error),
    /// A child by name, with its kind and attributes, or `Ok(None)` when it vanished
    /// between the listing and its stat, or the error that stat failed with.
    ///
    /// `NotFound` for a name the listing just returned means the entry was deleted in
    /// between, and every walk records it as it records a name the listing never
    /// returned: a cold walk has nothing to record, and a reconciliation removes what
    /// its baseline held. Reported as an error, it would make a walk over a tree being
    /// cleaned partial, and in a reconciliation it would settle as a phantom entry with
    /// permanent partial freshness. A native listing does not yield such a child at all.
    /// Any other error means the entry is present but unreadable.
    Child {
        name: std::borrow::Cow<'a, OsStr>,
        observed: std::io::Result<Option<(EntryKind, Attrs)>>,
    },
}

/// One directory's children, natively where a reader served it and portably otherwise.
pub(crate) enum Listing<'r> {
    #[cfg(all(target_os = "linux", target_env = "gnu"))]
    Native(linux_dents::Listing<'r>),
    Portable {
        entries: fs::ReadDir,
        policy: ListingPolicy,
        /// What this listing has proved of its directory so far.
        searchability: Searchability,
        readers: std::marker::PhantomData<&'r mut Readers>,
    },
}

impl<'r> Iterator for Listing<'r> {
    type Item = Listed<'r>;

    fn next(&mut self) -> Option<Listed<'r>> {
        match self {
            #[cfg(all(target_os = "linux", target_env = "gnu"))]
            Self::Native(listing) => {
                let entry = listing.next()?;
                let observed = match entry.outcome {
                    linux_dents::Outcome::Observed { kind, attrs } => Ok(Some((kind, attrs))),
                    linux_dents::Outcome::Failed(error) => Err(error),
                };
                Some(Listed::Child { name: std::borrow::Cow::Borrowed(entry.name), observed })
            }
            Self::Portable { entries, policy, searchability, .. } => {
                let item = match entries.next()? {
                    Ok(item) => item,
                    Err(error) => return Some(Listed::Failed(error)),
                };
                crate::counters::bump(|c| c.dir_entries += 1);
                let observed = listed_child_kind_and_attrs(&item, *policy, searchability);
                Some(Listed::Child { name: std::borrow::Cow::Owned(item.file_name()), observed })
            }
        }
    }
}

/// List `abs_dir` for a walk, a reconciliation, or discovery.
///
/// On Linux with glibc the native reader serves the directory unless it declines or a
/// test hook covers the directory; the portable `read_dir` answers otherwise, and an
/// error opening it is the caller's to report. Each backend produces the same items:
/// a native listing counts itself, and the portable one is counted here, as a
/// `dir_opens` and a portable attempt in `diagnostics`.
pub(crate) fn list_directory<'r>(
    readers: &'r mut Readers,
    abs_dir: &Path,
    policy: ListingPolicy,
    diagnostics: Option<&ScanDiagnosticsRecorder>,
) -> std::io::Result<Listing<'r>> {
    #[cfg(all(target_os = "linux", target_env = "gnu"))]
    {
        // Plain `if`, not `bool::then(|| …)`: the listing borrows the reader.
        if !walk_hook_covers(abs_dir) {
            if let Some(diagnostics) = diagnostics {
                diagnostics.linux_dents_attempted();
            }
            let native = linux_dents::StatPolicy {
                skip_dir_symlink_stat: policy.skip_dir_symlink_stat,
                one_filesystem: policy.one_filesystem,
            };
            if let Some(listing) = readers.dents.read(abs_dir, native) {
                if let Some(diagnostics) = diagnostics {
                    diagnostics.linux_dents_succeeded();
                }
                return Ok(Listing::Native(listing));
            }
            // A declined directory is counted as a fallback here and as the portable
            // attempt below, as the walker counts it.
            if let Some(diagnostics) = diagnostics {
                diagnostics.linux_dents_fell_back();
            }
        }
    }
    #[cfg(not(all(target_os = "linux", target_env = "gnu")))]
    let _ = &readers;
    crate::counters::bump(|c| c.dir_opens += 1);
    if let Some(diagnostics) = diagnostics {
        diagnostics.portable_attempted();
    }
    let entries = fs::read_dir(abs_dir)?;
    if let Some(diagnostics) = diagnostics {
        diagnostics.portable_succeeded();
    }
    Ok(Listing::Portable {
        entries,
        policy,
        searchability: Searchability::Unproven,
        readers: std::marker::PhantomData,
    })
}

fn missing_as_none<T>(lookup: std::io::Result<T>) -> std::io::Result<Option<T>> {
    match lookup {
        Ok(metadata) => Ok(Some(metadata)),
        Err(error) if error.kind() == std::io::ErrorKind::NotFound => Ok(None),
        Err(error) => Err(error),
    }
}

/// Whether a test hook observes lookups or listings under `path`, which a native read
/// would not make.
#[cfg(all(test, any(target_os = "macos", all(target_os = "linux", target_env = "gnu"))))]
fn walk_hook_covers(path: &Path) -> bool {
    walk_hook(path).is_some()
}

#[cfg(all(not(test), any(target_os = "macos", all(target_os = "linux", target_env = "gnu"))))]
const fn walk_hook_covers(_path: &Path) -> bool {
    false
}

/// Owned output from the filesystem walker before it crosses a public mutation boundary.
///
/// Only the scan and opened-discovery producers construct this type. Their admission,
/// depth, filesystem, and symlink checks have already selected every operation, and the
/// index consumes the owned paths while proving their parent identities under its write
/// boundary. Public scan callers receive an [`Observation`] instead and therefore keep
/// the full public normalization and atomic-validation contract.
#[derive(Debug)]
pub(crate) struct ScannerBatch {
    ops: Vec<ObservationOp>,
    /// When set, the consumer must return `ops` through this sender instead of dropping
    /// them. Workers allocate the `PathBuf`s; returning the drained vec lets glibc free
    /// those arenas on the producing thread. The public [`scan`] path leaves this unset.
    recycle: Option<std::sync::mpsc::Sender<Vec<ObservationOp>>>,
}

impl ScannerBatch {
    pub(crate) const fn new(ops: Vec<ObservationOp>) -> Self {
        Self { ops, recycle: None }
    }

    fn with_recycle(self, recycle: std::sync::mpsc::Sender<Vec<ObservationOp>>) -> Self {
        Self { recycle: Some(recycle), ..self }
    }

    #[cfg(test)]
    pub(crate) fn from_ops(ops: Vec<Op>) -> Self {
        Self { ops: ops.into_iter().map(ObservationOp::unconditional).collect(), recycle: None }
    }

    pub(crate) fn len(&self) -> usize {
        self.ops.len()
    }

    pub(crate) fn ops(&self) -> &[ObservationOp] {
        &self.ops
    }

    pub(crate) fn into_ops(self) -> Vec<ObservationOp> {
        self.ops
    }

    fn into_observation(self) -> Observation {
        Observation::from_ops(self.ops)
    }

    fn recycle(self) {
        if let Some(recycle) = self.recycle {
            let _ = recycle.send(self.ops);
        }
    }
}

/// One direct child retained by the private detached cold-bootstrap builder.
///
/// The worker owns the component once. Unlike [`ScannerBatch`], this record does not
/// manufacture a full relative path or a public observation for every entry.
#[derive(Debug)]
pub(crate) struct DetachedChild {
    pub(crate) name: OsString,
    pub(crate) kind: EntryKind,
    pub(crate) attrs: Attrs,
    /// Enumeration order within the listing. An enumerator can repeat a name while its
    /// directory is modified, and the builder keeps the later observation, as a
    /// streaming re-upsert does.
    pub(crate) position: u32,
}

/// One directory listing retained by a worker for detached bootstrap consolidation.
///
/// `path` is paid once per directory. Its children remain grouped exactly as the
/// filesystem enumerator produced them, so consolidation resolves the parent once and
/// never reconstructs a child path for nondirectories. A fixed control is retained
/// separately so the consumer can install the directory's complete control state
/// before it classifies any sibling or makes descendants visible.
#[derive(Debug)]
pub(crate) struct DetachedDirectory {
    pub(crate) path: PathBuf,
    pub(crate) children: Vec<DetachedChild>,
    pub(crate) control: Option<Op>,
}

/// Walk `root` and emit observations describing everything found.
pub fn scan(
    root: &Path,
    config: &ScanConfig,
    sink: &mut dyn FnMut(Observation),
) -> Result<ScanReport> {
    let mut public_sink = |batch: ScannerBatch| sink(batch.into_observation());
    let (mut report, _diagnostics) = scan_internal(
        root,
        config,
        &mut public_sink,
        false,
        WorkerPolicyExperiment::ShippedOneShot,
        SinkMode::Retained,
    )?;
    normalize_walk_errors(root, &mut report.errors);
    Ok(report)
}

/// Walk `root` for the transient summary tier, folding each op without retaining it.
///
/// The public [`scan`] path hands each batch to the caller as an [`Observation`], so
/// worker-allocated `PathBuf`s are freed on the consumer thread. This path returns
/// drained batches to the producing worker so each arena is allocated and freed on one
/// thread. Tallies must match [`scan`], and so must the normalized error set: the
/// summary report's status is built from these errors exactly as a retained walk's is.
/// A scan that reads `.gitignore` delivers each directory's control ahead of every entry
/// it governs, so the fold can classify each entry as the index would
/// ([`SinkMode::groups_directories`]).
pub(crate) fn scan_summary_fold(
    root: &Path,
    config: &ScanConfig,
    fold: &mut dyn FnMut(&ObservationOp),
) -> Result<ScanReport> {
    let mut sink = |batch: ScannerBatch| {
        for op in batch.ops() {
            fold(op);
        }
        batch.recycle();
    };
    let (mut report, _diagnostics) = scan_internal(
        root,
        config,
        &mut sink,
        false,
        WorkerPolicyExperiment::ShippedOneShot,
        SinkMode::TransientFold,
    )?;
    normalize_walk_errors(root, &mut report.errors);
    Ok(report)
}

/// [`scan_summary_fold`] plus the diagnostic trace [`scan_with_diagnostics`] collects.
pub(crate) fn scan_summary_fold_with_diagnostics(
    root: &Path,
    config: &ScanConfig,
    fold: &mut dyn FnMut(&ObservationOp),
) -> Result<(ScanReport, ScanDiagnostics)> {
    let mut sink = |batch: ScannerBatch| {
        for op in batch.ops() {
            fold(op);
        }
        batch.recycle();
    };
    let (mut report, diagnostics) = scan_internal(
        root,
        config,
        &mut sink,
        true,
        WorkerPolicyExperiment::ShippedOneShot,
        SinkMode::TransientFold,
    )?;
    normalize_walk_errors(root, &mut report.errors);
    Ok((report, diagnostics.expect("diagnostic scan creates a recorder")))
}

/// Walk `root`, emitting observations and a bounded run-scoped diagnostic trace.
///
/// This is the measurement counterpart to [`scan`]. It produces the same observation
/// stream and report while recording controller and backend evidence that ordinary
/// scans intentionally do not collect.
pub fn scan_with_diagnostics(
    root: &Path,
    config: &ScanConfig,
    sink: &mut dyn FnMut(Observation),
) -> Result<(ScanReport, ScanDiagnostics)> {
    scan_with_policy_diagnostics(root, config, sink, WorkerPolicyExperiment::ShippedOneShot)
}

/// Exercise a repository-only worker-controller candidate and retain its trace.
#[doc(hidden)]
pub fn scan_with_policy_diagnostics(
    root: &Path,
    config: &ScanConfig,
    sink: &mut dyn FnMut(Observation),
    policy: WorkerPolicyExperiment,
) -> Result<(ScanReport, ScanDiagnostics)> {
    let mut public_sink = |batch: ScannerBatch| sink(batch.into_observation());
    let (mut report, diagnostics) =
        scan_internal(root, config, &mut public_sink, true, policy, SinkMode::Retained)?;
    normalize_walk_errors(root, &mut report.errors);
    Ok((report, diagnostics.expect("diagnostic scan creates a recorder")))
}

/// What the caller does with each batch of observations.
///
/// Two measured keeps hang off this one concept, and both were measured on the
/// transient fold alone: returning drained batches to the producing worker (H147,
/// exp-151) and taking directory and symlink kind from the listing without a stat
/// (H72, exp-153). They are named here as properties of the mode rather than passed as
/// one flag under one of their names, so a measurement on another platform can move
/// one without silently moving the other.
///
/// A third property is semantic rather than measured: a transient fold that observes
/// `.gitignore` classifies on its consumer, and that needs each directory's control
/// before its entries ([`Self::groups_directories`]).
#[derive(Clone, Copy, Debug, PartialEq, Eq)]
enum SinkMode {
    /// The consumer keeps the observations: the public [`scan`] and the index.
    Retained,
    /// The consumer folds each batch and drops it: the transient summary tier.
    TransientFold,
}

impl SinkMode {
    /// Drained batches go back to the worker that allocated them (H147).
    fn recycles_batches(self) -> bool {
        self == Self::TransientFold
    }

    /// Directory and symlink kind come from the listing without a stat (H72).
    fn skips_dir_symlink_stat(self) -> bool {
        self == Self::TransientFold
    }

    /// Each directory's control reaches the consumer ahead of every entry it governs
    /// (fdu-1ovb).
    ///
    /// The transient summary classifies every entry against `.gitignore` on its
    /// consumer, as the detached index builder does, and the builder applies a
    /// directory's control before it classifies any child because it receives each
    /// listing whole. A streaming batch keeps listing order, where `.gitignore` can come
    /// last, so a worker holds a listing's observations and moves the control ahead of
    /// them when the listing ends, or, if the batch fills first, reads the directory's
    /// control directly and lets that read stand for the listing
    /// (`StreamingEmission::send_if_full`). The retained stream keeps listing order: the
    /// index reclassifies the subtree a control governs when that control arrives.
    fn groups_directories(self, config: &ScanConfig) -> bool {
        self == Self::TransientFold && config.read_controls
    }
}

fn scan_internal(
    root: &Path,
    config: &ScanConfig,
    sink: &mut dyn FnMut(ScannerBatch),
    collect_diagnostics: bool,
    policy: WorkerPolicyExperiment,
    sink_mode: SinkMode,
) -> Result<(ScanReport, Option<ScanDiagnostics>)> {
    config.validate()?;
    if let Some(progress) = &config.progress {
        progress.enter(crate::ProgressPhase::Scanning);
    }
    let root_meta = {
        crate::counters::bump(|c| c.stats += 1);
        fs::symlink_metadata(root)
    }
    .map_err(|e| Error::io(root, e))?;
    if !root_meta.is_dir() {
        return Err(Error::io(
            root,
            std::io::Error::new(std::io::ErrorKind::NotADirectory, "scan root is not a directory"),
        ));
    }
    let root_dev = root_device(root, &root_meta).map_err(|error| Error::io(root, error))?;
    let available_parallelism =
        std::thread::available_parallelism().map_or(1, std::num::NonZero::get);
    // Narrow populations need control admission before child enumeration. Keep one ordered
    // producer and control table so a bounded budget makes the same decisions as the
    // index that consumes the observations.
    let pool = if config.population == crate::query::IgnoredEntries::Include {
        config.worker_pool_for(available_parallelism)
    } else {
        WorkerPool::fixed(1)
    };
    let diagnostics = collect_diagnostics
        .then(|| ScanDiagnosticsRecorder::new(pool, available_parallelism, policy));

    // The serial walk below emits in listing order and never groups a directory, so a
    // transient fold that classifies takes the concurrent walk even with one worker. That
    // is the walk the detached index takes at every worker count, and one worker visits
    // directories in the order that builder consumes them, so a control budget admits
    // the same files on both routes. A narrowed population keeps the serial walk, which
    // reads each directory's control before listing it.
    let groups = sink_mode.groups_directories(config)
        && config.population == crate::query::IgnoredEntries::Include;
    if config.max_depth != Some(0) && (pool.initial > 1 || groups) {
        let report = scan_concurrent(
            root,
            config,
            root_dev,
            sink,
            pool,
            diagnostics.as_ref(),
            policy,
            sink_mode,
        );
        return Ok((report, diagnostics.as_ref().map(|value| value.finish())));
    }

    let mut report = ScanReport::default();
    if config.max_depth == Some(0) {
        if let Some(diagnostics) = &diagnostics {
            diagnostics.mark_not_run();
            diagnostics.record_queue_finish(0, 0);
        }
        return Ok((report, diagnostics.as_ref().map(|value| value.finish())));
    }
    let worker_guard = diagnostics.as_ref().map(ScanDiagnosticsRecorder::worker_guard);
    let walk_started = std::time::Instant::now();
    let mut batch: Vec<ObservationOp> = Vec::with_capacity(config.batch_size);
    let mut queue: VecDeque<(PathBuf, usize)> = VecDeque::from(vec![(PathBuf::new(), 0)]);
    let mut controls = (config.population != crate::query::IgnoredEntries::Include)
        .then(|| crate::control::ControlTable::with_limits(config.control_limits));
    let mut unreadable_controls = std::collections::BTreeSet::new();
    let mut tally = ProgressTally::new(config.progress.as_ref());
    let mut readers = Readers::new();
    let policy = ListingPolicy {
        skip_dir_symlink_stat: sink_mode.skips_dir_symlink_stat(),
        one_filesystem: config.one_filesystem,
    };
    // Every batch leaves through here, so the batch is where the serial walk reports
    // its progress: the handoff the consumer already pays for, never the entry.
    let mut emit = |ops: Vec<ObservationOp>, report: &mut ScanReport| {
        let send_started = std::time::Instant::now();
        sink(ScannerBatch::new(ops));
        report.attribution.send_ns += elapsed_ns(send_started);
        tally.flush(report);
    };

    while let Some((rel_dir, depth)) = take_next(&mut queue, config.order) {
        let abs_dir = root.join(&rel_dir);
        if let Some(controls) = controls.as_mut() {
            let control_path = rel_dir.join(crate::control::CONTROL_FILE_NAME);
            match read_directory_control(config, root, &control_path) {
                Ok(Some(op)) => {
                    apply_discovery_control(controls, &op)?;
                    batch.push(ObservationOp::unconditional(op));
                }
                Ok(None) => {}
                Err(error) => {
                    unreadable_controls.insert(rel_dir.clone());
                    report.errors.push(error);
                }
            }
        }
        let listing = match list_directory(&mut readers, &abs_dir, policy, diagnostics.as_deref()) {
            Ok(listing) => listing,
            Err(e) => {
                report.errors.push(Error::io(abs_dir, e));
                continue;
            }
        };
        report.dirs_read += 1;

        for item in listing {
            let (name, observed) = match item {
                Listed::Child { name, observed } => (name, observed),
                Listed::Failed(e) => {
                    report.errors.push(Error::io(&abs_dir, e));
                    continue;
                }
            };
            let rel_path = rel_dir.join(&name);
            let (kind, attrs) = match observed {
                Ok(Some(observed)) => observed,
                Ok(None) => continue,
                Err(error) => {
                    report.errors.push(Error::io(abs_dir.join(&name), error));
                    continue;
                }
            };
            let disposition =
                crate::admission::decide(&name, kind, config.hidden(), config.exclude_special);
            if disposition == crate::admission::Disposition::Reject {
                continue;
            }
            if population_prunes(
                config.population,
                &rel_path,
                kind,
                disposition,
                controls.as_ref(),
                &unreadable_controls,
            ) {
                continue;
            }
            // A narrowed walk looked the directory's control up before listing it, and that
            // lookup stands for every spelling the listing shows.
            let control =
                match if controls.is_some() && crate::control::control_spelling(&name).is_some() {
                    Ok(None)
                } else {
                    read_control_op(config, root, &rel_path, kind)
                } {
                    Ok(control) => control,
                    Err(error) => {
                        report.errors.push(error);
                        None
                    }
                };
            if disposition == crate::admission::Disposition::ControlOnly {
                if let Some(control) = control {
                    batch.push(ObservationOp::unconditional(control));
                    if batch.len() >= config.batch_size {
                        emit(std::mem::take(&mut batch), &mut report);
                        batch.reserve(config.batch_size);
                    }
                }
                continue;
            }
            report.observe(kind, attrs);
            batch.push(ObservationOp::unconditional(Op::Upsert {
                path: rel_path.clone(),
                kind,
                attrs,
            }));
            if batch.len() >= config.batch_size {
                emit(std::mem::take(&mut batch), &mut report);
                batch.reserve(config.batch_size);
            }
            if let Some(control) = control {
                batch.push(ObservationOp::unconditional(control));
                if batch.len() >= config.batch_size {
                    emit(std::mem::take(&mut batch), &mut report);
                    batch.reserve(config.batch_size);
                }
            }

            if should_descend(kind, attrs, depth, root_dev, config) {
                queue.push_back((rel_path, depth + 1));
            }
        }
    }

    if !batch.is_empty() {
        emit(batch, &mut report);
    }
    // A walk whose last directories filled no batch has counted them and sent nothing.
    tally.flush(&report);
    // A serial walk has no coordination to attribute: wall is the loop, "send" is the
    // inline sink — which is the consumer actually running — and work is the rest.
    report.attribution.wall_ns = elapsed_ns(walk_started);
    report.attribution.work_ns =
        report.attribution.wall_ns.saturating_sub(report.attribution.send_ns);
    drop(worker_guard);
    if let Some(diagnostics) = &diagnostics {
        diagnostics.record_queue_finish(0, 0);
    }
    Ok((report, diagnostics.as_ref().map(|value| value.finish())))
}

/// Take the next directory in the configured order.
///
/// Both orders push to the back; only the end they are taken from differs, which is
/// what keeps this a one-line policy rather than two walkers.
fn take_next(queue: &mut VecDeque<(PathBuf, usize)>, order: ScanOrder) -> Option<(PathBuf, usize)> {
    match order {
        ScanOrder::BreadthFirst => queue.pop_front(),
        ScanOrder::DepthFirst => queue.pop_back(),
    }
}

/// Largest worker pool a caller may ask for explicitly.
///
/// Well past anything measured to help. It exists so a caller that computes a thread
/// count from something silly cannot spawn thousands of threads.
const MAX_SCAN_THREADS: usize = 32;

/// Ceiling on the workers active at the start of an automatic scan.
///
/// Measured, not guessed. On a 10-core machine walking a 60k-entry `node_modules`
/// tree, wall time fell 37% at two workers and 50% at four, then stopped improving:
/// six matched four within noise and eight was 4% worse than four. The walk becomes
/// bound by the single index consumer, so past this point extra workers buy queue
/// contention and efficiency-core scheduling rather than throughput. See
/// `docs/project/reports/report-2026-08-10-fdu-performance-experiments.md`.
///
/// That measurement was taken on macOS, and every constant in this group now reads its
/// value from [`crate::platform_tuning`], which records per platform whether the number
/// was measured there or inherited. On Linux these are inherited.
const DEFAULT_SCAN_THREADS_CAP: usize = crate::platform_tuning::tuning().scan_threads_cap.get();

/// Ceiling on automatic workers for an immutable-baseline reconciliation wave.
///
/// Reconciliation reads both filesystem and index state. Its measured knee arrives
/// before the cold producer's because additional metadata calls amplify kernel work
/// after the index comparisons already saturate the performance cores.
const DEFAULT_RECONCILE_THREADS_CAP: usize =
    crate::platform_tuning::tuning().reconcile_threads_cap.get();

/// Ceiling an automatic scan may unlock after it establishes that the tree is large.
///
/// Sixteen was the knee on the 720k-entry cache-pressure corpus in exp-015. Thirty-two
/// did not improve on it and spent substantially more worker time waiting at the end.
const ADAPTIVE_SCAN_THREADS_CAP: usize =
    crate::platform_tuning::tuning().adaptive_scan_threads_cap.get();

/// Maximum reserve depth relative to the host's reported parallelism.
const ADAPTIVE_SCAN_PARALLELISM_MULTIPLIER: usize =
    crate::platform_tuning::tuning().adaptive_scan_parallelism_multiplier.get();

/// Entries used to calibrate the initial workers' filesystem service time.
const ADAPTIVE_SCAN_CALIBRATION_ENTRIES: u64 =
    crate::platform_tuning::tuning().adaptive_scan_calibration_entries.get();

/// Average worker time per observed entry that identifies a latency-bound scan.
///
/// Whole-run attribution separated the measured regimes: roughly 18 microseconds on
/// the 60k tree, 22 on the 120k boundary, and 42 or more on the 720k cache-pressure
/// tree. Thirty leaves margin between them. The calibration uses the same chunk timing
/// already collected for attribution, so it adds no per-entry clock reads.
///
/// Those are APFS regimes. The Linux warm floor is about 1.5 µs per entry, twenty times
/// below this threshold, so the trigger may never fire there — which is exactly the kind
/// of inherited constant [`crate::platform_tuning`] exists to make visible (H84).
const ADAPTIVE_SCAN_SLOW_WORK_NS_PER_ENTRY: u64 =
    crate::platform_tuning::tuning().adaptive_scan_slow_work_ns_per_entry.get();

#[derive(Clone, Copy, Debug, PartialEq, Eq)]
struct WorkerPool {
    initial: usize,
    maximum: usize,
    calibration: Option<WorkerCalibration>,
}

#[derive(Clone, Copy, Debug, PartialEq, Eq)]
struct WorkerCalibration {
    minimum_entries: u64,
    slow_work_ns_per_entry: u64,
    entries: u64,
    work_ns: u64,
    chunks: u64,
}

#[derive(Clone, Copy, Debug)]
struct PolicyWindowSnapshot {
    sequence: u64,
    start_entry_ordinal: u64,
    end_entry_ordinal: u64,
    observed_entries: u64,
    observed_chunks: u64,
    observed_work_ns: u64,
    ready_directories: usize,
    in_flight_directories: usize,
    active_workers: usize,
    handoff_backlog: usize,
    requested_workers: Option<usize>,
    decision: WorkerPolicyDecision,
}

struct PolicyTraceState {
    outcome: WorkerPolicyOutcome,
    outcome_reason: Option<&'static str>,
    outcome_sequence: Option<u64>,
    windows: Vec<WorkerPolicyWindow>,
    events_truncated: bool,
    ready_directories_at_finish: usize,
    in_flight_directories_at_finish: usize,
}

/// Shared state used only by the opt-in diagnostic scan APIs.
///
/// Normal scans pass no recorder and therefore never touch these atomics or locks. The
/// trace mutex is deliberately separate from the directory queue: recording a policy
/// window may add diagnostic cost, but it cannot alter the queue's synchronization or
/// the controller's decision.
pub(crate) struct ScanDiagnosticsRecorder {
    available_parallelism: usize,
    pool: WorkerPool,
    policy: WorkerPolicyExperiment,
    trace: std::sync::Mutex<PolicyTraceState>,
    workers_spawned: std::sync::atomic::AtomicUsize,
    active_workers: std::sync::atomic::AtomicUsize,
    peak_active_workers: std::sync::atomic::AtomicUsize,
    handoff_backlog: std::sync::atomic::AtomicUsize,
    handoff_backlog_high_water: std::sync::atomic::AtomicUsize,
    calibration_chunks: std::sync::atomic::AtomicU64,
    calibration_entries: std::sync::atomic::AtomicU64,
    calibration_work_ns: std::sync::atomic::AtomicU64,
    worker_expansions: std::sync::atomic::AtomicU64,
    portable_attempts: std::sync::atomic::AtomicU64,
    portable_successes: std::sync::atomic::AtomicU64,
    #[cfg(target_os = "macos")]
    macos_bulk_attempts: std::sync::atomic::AtomicU64,
    #[cfg(target_os = "macos")]
    macos_bulk_successes: std::sync::atomic::AtomicU64,
    #[cfg(target_os = "macos")]
    macos_bulk_fallbacks: std::sync::atomic::AtomicU64,
    #[cfg(all(target_os = "linux", target_env = "gnu"))]
    linux_dents_attempts: std::sync::atomic::AtomicU64,
    #[cfg(all(target_os = "linux", target_env = "gnu"))]
    linux_dents_successes: std::sync::atomic::AtomicU64,
    #[cfg(all(target_os = "linux", target_env = "gnu"))]
    linux_dents_fallbacks: std::sync::atomic::AtomicU64,
}

impl ScanDiagnosticsRecorder {
    fn new(
        pool: WorkerPool,
        available_parallelism: usize,
        policy: WorkerPolicyExperiment,
    ) -> std::sync::Arc<Self> {
        let (outcome, outcome_reason) = if pool.calibration.is_some() {
            (
                WorkerPolicyOutcome::Undecided,
                Some("the adaptive calibration window has not completed"),
            )
        } else {
            (WorkerPolicyOutcome::Fixed, Some("this worker pool has no adaptive reserve"))
        };
        std::sync::Arc::new(Self {
            available_parallelism,
            pool,
            policy,
            trace: std::sync::Mutex::new(PolicyTraceState {
                outcome,
                outcome_reason,
                outcome_sequence: None,
                windows: Vec::new(),
                events_truncated: false,
                ready_directories_at_finish: 0,
                in_flight_directories_at_finish: 0,
            }),
            workers_spawned: std::sync::atomic::AtomicUsize::new(0),
            active_workers: std::sync::atomic::AtomicUsize::new(0),
            peak_active_workers: std::sync::atomic::AtomicUsize::new(0),
            handoff_backlog: std::sync::atomic::AtomicUsize::new(0),
            handoff_backlog_high_water: std::sync::atomic::AtomicUsize::new(0),
            calibration_chunks: std::sync::atomic::AtomicU64::new(0),
            calibration_entries: std::sync::atomic::AtomicU64::new(0),
            calibration_work_ns: std::sync::atomic::AtomicU64::new(0),
            worker_expansions: std::sync::atomic::AtomicU64::new(0),
            portable_attempts: std::sync::atomic::AtomicU64::new(0),
            portable_successes: std::sync::atomic::AtomicU64::new(0),
            #[cfg(target_os = "macos")]
            macos_bulk_attempts: std::sync::atomic::AtomicU64::new(0),
            #[cfg(target_os = "macos")]
            macos_bulk_successes: std::sync::atomic::AtomicU64::new(0),
            #[cfg(target_os = "macos")]
            macos_bulk_fallbacks: std::sync::atomic::AtomicU64::new(0),
            #[cfg(all(target_os = "linux", target_env = "gnu"))]
            linux_dents_attempts: std::sync::atomic::AtomicU64::new(0),
            #[cfg(all(target_os = "linux", target_env = "gnu"))]
            linux_dents_successes: std::sync::atomic::AtomicU64::new(0),
            #[cfg(all(target_os = "linux", target_env = "gnu"))]
            linux_dents_fallbacks: std::sync::atomic::AtomicU64::new(0),
        })
    }

    fn worker_guard(self: &std::sync::Arc<Self>) -> ScanWorkerGuard {
        let active = self
            .active_workers
            .fetch_add(1, std::sync::atomic::Ordering::Relaxed)
            .saturating_add(1);
        self.workers_spawned.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
        atomic_update_max(&self.peak_active_workers, active);
        ScanWorkerGuard { recorder: self.clone() }
    }

    fn record_policy_window(&self, snapshot: PolicyWindowSnapshot) {
        let mut trace = self.trace.lock().unwrap_or_else(std::sync::PoisonError::into_inner);
        let sequence = snapshot.sequence;
        let supersedes = trace.outcome_sequence.is_none_or(|current| sequence >= current);
        match snapshot.decision {
            WorkerPolicyDecision::Undecided if supersedes => {
                trace.outcome = WorkerPolicyOutcome::Undecided;
                trace.outcome_reason =
                    Some("the walk ended before the adaptive calibration window completed");
                trace.outcome_sequence = Some(sequence);
            }
            WorkerPolicyDecision::Hold
                if trace.outcome != WorkerPolicyOutcome::ScaledUp && supersedes =>
            {
                trace.outcome = WorkerPolicyOutcome::Held;
                trace.outcome_reason = None;
                trace.outcome_sequence = Some(sequence);
            }
            WorkerPolicyDecision::ScaleUp => {
                trace.outcome = WorkerPolicyOutcome::ScaledUp;
                trace.outcome_reason = None;
                trace.outcome_sequence = Some(sequence);
            }
            WorkerPolicyDecision::HoldNoUsefulWork
                if trace.outcome != WorkerPolicyOutcome::ScaledUp && supersedes =>
            {
                trace.outcome = WorkerPolicyOutcome::HeldNoUsefulWork;
                trace.outcome_reason =
                    Some("the slow trigger fired only after ready and in-flight work had drained");
                trace.outcome_sequence = Some(sequence);
            }
            WorkerPolicyDecision::HoldInsufficientFrontier
                if trace.outcome != WorkerPolicyOutcome::ScaledUp && supersedes =>
            {
                trace.outcome = WorkerPolicyOutcome::Held;
                trace.outcome_reason =
                    Some("the observed frontier could not use additional workers");
                trace.outcome_sequence = Some(sequence);
            }
            WorkerPolicyDecision::HoldHandoffBacklog
                if trace.outcome != WorkerPolicyOutcome::ScaledUp && supersedes =>
            {
                trace.outcome = WorkerPolicyOutcome::Held;
                trace.outcome_reason =
                    Some("the observation handoff backlog was already at the controller limit");
                trace.outcome_sequence = Some(sequence);
            }
            WorkerPolicyDecision::Incomplete
                if supersedes && trace.outcome == WorkerPolicyOutcome::Undecided =>
            {
                trace.outcome_reason =
                    Some("the walk ended before any adaptive calibration window completed");
                trace.outcome_sequence = Some(sequence);
            }
            _ => {}
        }
        if sequence >= MAX_POLICY_TRACE_EVENTS as u64 {
            trace.events_truncated = true;
            return;
        }
        let work_ns_per_entry = (snapshot.observed_entries > 0)
            .then(|| snapshot.observed_work_ns / snapshot.observed_entries);
        trace.windows.push(WorkerPolicyWindow {
            sequence,
            start_entry_ordinal: snapshot.start_entry_ordinal,
            end_entry_ordinal: snapshot.end_entry_ordinal,
            observed_entries: snapshot.observed_entries,
            observed_chunks: snapshot.observed_chunks,
            observed_work_ns: snapshot.observed_work_ns,
            work_ns_per_entry,
            work_ns_per_entry_unavailable_reason: work_ns_per_entry
                .is_none()
                .then_some("the window observed no entries"),
            ready_directories: snapshot.ready_directories,
            in_flight_directories: snapshot.in_flight_directories,
            active_workers: snapshot.active_workers,
            handoff_backlog: snapshot.handoff_backlog,
            requested_workers: snapshot.requested_workers,
            decision: snapshot.decision,
        });
    }

    fn mark_not_run(&self) {
        let mut trace = self.trace.lock().unwrap_or_else(std::sync::PoisonError::into_inner);
        trace.outcome = WorkerPolicyOutcome::NotRun;
        trace.outcome_reason = Some("max_depth zero requested no directory walk");
    }

    fn record_queue_finish(&self, ready_directories: usize, in_flight_directories: usize) {
        let mut trace = self.trace.lock().unwrap_or_else(std::sync::PoisonError::into_inner);
        trace.ready_directories_at_finish = ready_directories;
        trace.in_flight_directories_at_finish = in_flight_directories;
    }

    fn handoff_sent(&self) {
        let backlog = self
            .handoff_backlog
            .fetch_add(1, std::sync::atomic::Ordering::Relaxed)
            .saturating_add(1);
        atomic_update_max(&self.handoff_backlog_high_water, backlog);
    }

    fn handoff_received(&self) {
        let previous = self.handoff_backlog.fetch_sub(1, std::sync::atomic::Ordering::Relaxed);
        debug_assert!(previous > 0, "received handoff must have been sent");
    }

    fn calibration_chunk(&self, entries: u64, work_ns: u64) {
        self.calibration_chunks.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
        self.calibration_entries.fetch_add(entries, std::sync::atomic::Ordering::Relaxed);
        self.calibration_work_ns.fetch_add(work_ns, std::sync::atomic::Ordering::Relaxed);
    }

    fn worker_expanded(&self) {
        self.worker_expansions.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
    }

    fn portable_attempted(&self) {
        self.portable_attempts.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
    }

    fn portable_succeeded(&self) {
        self.portable_successes.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
    }

    #[cfg(target_os = "macos")]
    fn macos_bulk_attempted(&self) {
        self.macos_bulk_attempts.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
    }

    #[cfg(target_os = "macos")]
    fn macos_bulk_succeeded(&self) {
        self.macos_bulk_successes.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
    }

    #[cfg(target_os = "macos")]
    fn macos_bulk_fell_back(&self) {
        self.macos_bulk_fallbacks.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
    }

    #[cfg(all(target_os = "linux", target_env = "gnu"))]
    fn linux_dents_attempted(&self) {
        self.linux_dents_attempts.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
    }

    #[cfg(all(target_os = "linux", target_env = "gnu"))]
    fn linux_dents_succeeded(&self) {
        self.linux_dents_successes.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
    }

    #[cfg(all(target_os = "linux", target_env = "gnu"))]
    fn linux_dents_fell_back(&self) {
        self.linux_dents_fallbacks.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
    }

    fn finish(&self) -> ScanDiagnostics {
        let trace = self.trace.lock().unwrap_or_else(std::sync::PoisonError::into_inner);
        let calibration = self.pool.calibration;
        let backend = ScanBackendDiagnostics {
            portable_attempts: self.portable_attempts.load(std::sync::atomic::Ordering::Relaxed),
            portable_directory_reads: self
                .portable_successes
                .load(std::sync::atomic::Ordering::Relaxed),
            #[cfg(target_os = "macos")]
            macos_bulk_attempts: Some(
                self.macos_bulk_attempts.load(std::sync::atomic::Ordering::Relaxed),
            ),
            #[cfg(not(target_os = "macos"))]
            macos_bulk_attempts: None,
            #[cfg(target_os = "macos")]
            macos_bulk_successes: Some(
                self.macos_bulk_successes.load(std::sync::atomic::Ordering::Relaxed),
            ),
            #[cfg(not(target_os = "macos"))]
            macos_bulk_successes: None,
            #[cfg(target_os = "macos")]
            macos_bulk_fallbacks: Some(
                self.macos_bulk_fallbacks.load(std::sync::atomic::Ordering::Relaxed),
            ),
            #[cfg(not(target_os = "macos"))]
            macos_bulk_fallbacks: None,
            #[cfg(target_os = "macos")]
            unavailable_reason: None,
            #[cfg(not(target_os = "macos"))]
            unavailable_reason: Some(
                "macOS bulk directory enumeration is unavailable on this platform",
            ),
            #[cfg(all(target_os = "linux", target_env = "gnu"))]
            linux_dents_attempts: Some(
                self.linux_dents_attempts.load(std::sync::atomic::Ordering::Relaxed),
            ),
            #[cfg(not(all(target_os = "linux", target_env = "gnu")))]
            linux_dents_attempts: None,
            #[cfg(all(target_os = "linux", target_env = "gnu"))]
            linux_dents_successes: Some(
                self.linux_dents_successes.load(std::sync::atomic::Ordering::Relaxed),
            ),
            #[cfg(not(all(target_os = "linux", target_env = "gnu")))]
            linux_dents_successes: None,
            #[cfg(all(target_os = "linux", target_env = "gnu"))]
            linux_dents_fallbacks: Some(
                self.linux_dents_fallbacks.load(std::sync::atomic::Ordering::Relaxed),
            ),
            #[cfg(not(all(target_os = "linux", target_env = "gnu")))]
            linux_dents_fallbacks: None,
        };
        ScanDiagnostics {
            schema: SCAN_DIAGNOSTICS_SCHEMA,
            worker_policy: WorkerPolicyDiagnostics {
                controller: worker_policy_experiment_name(self.policy),
                available_parallelism: self.available_parallelism,
                initial_workers: self.pool.initial,
                maximum_workers: self.pool.maximum,
                calibration_window_entries: calibration.map(|value| value.minimum_entries),
                slow_threshold_ns_per_entry: calibration.map(|value| value.slow_work_ns_per_entry),
                calibration_chunks: self
                    .calibration_chunks
                    .load(std::sync::atomic::Ordering::Relaxed),
                calibration_entries: self
                    .calibration_entries
                    .load(std::sync::atomic::Ordering::Relaxed),
                calibration_work_ns: self
                    .calibration_work_ns
                    .load(std::sync::atomic::Ordering::Relaxed),
                worker_expansions: self
                    .worker_expansions
                    .load(std::sync::atomic::Ordering::Relaxed),
                outcome: trace.outcome,
                outcome_reason: trace.outcome_reason,
                workers_spawned: self.workers_spawned.load(std::sync::atomic::Ordering::Relaxed),
                peak_active_workers: self
                    .peak_active_workers
                    .load(std::sync::atomic::Ordering::Relaxed),
                ready_directories_at_finish: trace.ready_directories_at_finish,
                in_flight_directories_at_finish: trace.in_flight_directories_at_finish,
                handoff_backlog_at_finish: self
                    .handoff_backlog
                    .load(std::sync::atomic::Ordering::Relaxed),
                handoff_backlog_high_water: self
                    .handoff_backlog_high_water
                    .load(std::sync::atomic::Ordering::Relaxed),
                windows: {
                    let mut windows = trace.windows.clone();
                    windows.sort_by_key(|window| window.sequence);
                    windows
                },
                events_truncated: trace.events_truncated,
            },
            backend,
        }
    }
}

const fn worker_policy_experiment_name(value: WorkerPolicyExperiment) -> &'static str {
    match value {
        WorkerPolicyExperiment::ShippedOneShot => "shipped_one_shot",
        WorkerPolicyExperiment::RepeatedWindows => "repeated_windows",
        WorkerPolicyExperiment::StagedGatedWindows => "staged_gated_windows",
    }
}

struct ScanWorkerGuard {
    recorder: std::sync::Arc<ScanDiagnosticsRecorder>,
}

impl Drop for ScanWorkerGuard {
    fn drop(&mut self) {
        let previous =
            self.recorder.active_workers.fetch_sub(1, std::sync::atomic::Ordering::Relaxed);
        debug_assert!(previous > 0, "worker guard must balance worker start");
    }
}

fn atomic_update_max(target: &std::sync::atomic::AtomicUsize, value: usize) {
    let mut observed = target.load(std::sync::atomic::Ordering::Relaxed);
    while value > observed {
        match target.compare_exchange_weak(
            observed,
            value,
            std::sync::atomic::Ordering::Relaxed,
            std::sync::atomic::Ordering::Relaxed,
        ) {
            Ok(_) => break,
            Err(actual) => observed = actual,
        }
    }
}

enum WalkMessage {
    Batch(ScannerBatch),
    DetachedDirectories {
        directories: Vec<DetachedDirectory>,
        /// The worker that allocated `directories`. The consumer drains each listing
        /// into the index and sends the emptied listings back here, so their path and
        /// child buffers are freed or reused on that worker's thread (H159).
        recycle: std::sync::mpsc::Sender<Vec<DetachedDirectory>>,
    },
    ScaleUp {
        sender: std::sync::mpsc::Sender<Self>,
        target_workers: usize,
    },
}

impl WorkerPool {
    const fn fixed(workers: usize) -> Self {
        Self { initial: workers, maximum: workers, calibration: None }
    }
}

impl WorkerCalibration {
    const fn new(minimum_entries: u64, slow_work_ns_per_entry: u64) -> Self {
        Self { minimum_entries, slow_work_ns_per_entry, entries: 0, work_ns: 0, chunks: 0 }
    }

    fn observe(&mut self, entries: u64, work_ns: u64) -> Option<bool> {
        self.chunks = self.chunks.saturating_add(1);
        self.entries = self.entries.saturating_add(entries);
        self.work_ns = self.work_ns.saturating_add(work_ns);
        (self.entries >= self.minimum_entries)
            .then(|| self.work_ns / self.entries >= self.slow_work_ns_per_entry)
    }
}

#[derive(Clone, Copy, Debug)]
struct CalibrationWindow {
    start_entry_ordinal: u64,
    end_entry_ordinal: u64,
    entries: u64,
    chunks: u64,
    work_ns: u64,
    slow: bool,
}

#[derive(Debug)]
struct RepeatedCalibration {
    minimum_entries: u64,
    slow_work_ns_per_entry: u64,
    window_start: u64,
    entries: u64,
    chunks: u64,
    work_ns: u64,
    completed_windows: u64,
}

impl RepeatedCalibration {
    const fn new(calibration: WorkerCalibration) -> Self {
        Self {
            minimum_entries: calibration.minimum_entries,
            slow_work_ns_per_entry: calibration.slow_work_ns_per_entry,
            window_start: 0,
            entries: 0,
            chunks: 0,
            work_ns: 0,
            completed_windows: 0,
        }
    }

    const fn starting_at(calibration: WorkerCalibration, window_start: u64) -> Self {
        let mut repeated = Self::new(calibration);
        repeated.window_start = window_start;
        repeated
    }

    fn observe(&mut self, entries: u64, work_ns: u64) -> Option<CalibrationWindow> {
        self.chunks = self.chunks.saturating_add(1);
        self.entries = self.entries.saturating_add(entries);
        self.work_ns = self.work_ns.saturating_add(work_ns);
        if self.entries < self.minimum_entries {
            return None;
        }
        let end_entry_ordinal = self.window_start.saturating_add(self.entries);
        let window = CalibrationWindow {
            start_entry_ordinal: self.window_start,
            end_entry_ordinal,
            entries: self.entries,
            chunks: self.chunks,
            work_ns: self.work_ns,
            slow: self.work_ns / self.entries >= self.slow_work_ns_per_entry,
        };
        self.window_start = end_entry_ordinal;
        self.entries = 0;
        self.chunks = 0;
        self.work_ns = 0;
        self.completed_windows = self.completed_windows.saturating_add(1);
        Some(window)
    }
}

#[derive(Debug)]
enum WorkerController {
    OneShot(WorkerCalibration),
    Repeated { calibration: RepeatedCalibration, staged_gated: bool },
}

impl WorkerController {
    fn new(calibration: WorkerCalibration, policy: WorkerPolicyExperiment) -> Self {
        match policy {
            WorkerPolicyExperiment::ShippedOneShot => Self::OneShot(calibration),
            WorkerPolicyExperiment::RepeatedWindows => Self::Repeated {
                calibration: RepeatedCalibration::new(calibration),
                staged_gated: false,
            },
            WorkerPolicyExperiment::StagedGatedWindows => Self::Repeated {
                calibration: RepeatedCalibration::new(calibration),
                staged_gated: true,
            },
        }
    }

    fn observe(&mut self, entries: u64, work_ns: u64) -> Option<CalibrationWindow> {
        match self {
            Self::OneShot(calibration) => {
                let slow = calibration.observe(entries, work_ns)?;
                Some(CalibrationWindow {
                    start_entry_ordinal: 0,
                    end_entry_ordinal: calibration.entries,
                    entries: calibration.entries,
                    chunks: calibration.chunks,
                    work_ns: calibration.work_ns,
                    slow,
                })
            }
            Self::Repeated { calibration, .. } => calibration.observe(entries, work_ns),
        }
    }

    fn partial_window(&self) -> (CalibrationWindow, WorkerPolicyDecision) {
        match self {
            Self::OneShot(calibration) => (
                CalibrationWindow {
                    start_entry_ordinal: 0,
                    end_entry_ordinal: calibration.entries,
                    entries: calibration.entries,
                    chunks: calibration.chunks,
                    work_ns: calibration.work_ns,
                    slow: false,
                },
                WorkerPolicyDecision::Undecided,
            ),
            Self::Repeated { calibration, .. } => (
                CalibrationWindow {
                    start_entry_ordinal: calibration.window_start,
                    end_entry_ordinal: calibration.window_start.saturating_add(calibration.entries),
                    entries: calibration.entries,
                    chunks: calibration.chunks,
                    work_ns: calibration.work_ns,
                    slow: false,
                },
                if calibration.completed_windows == 0 {
                    WorkerPolicyDecision::Undecided
                } else {
                    WorkerPolicyDecision::Incomplete
                },
            ),
        }
    }

    const fn is_staged_gated(&self) -> bool {
        matches!(self, Self::Repeated { staged_gated: true, .. })
    }

    const fn is_one_shot(&self) -> bool {
        matches!(self, Self::OneShot(_))
    }

    const fn calibration_spec(&self) -> WorkerCalibration {
        match self {
            Self::OneShot(calibration) => WorkerCalibration::new(
                calibration.minimum_entries,
                calibration.slow_work_ns_per_entry,
            ),
            Self::Repeated { calibration, .. } => WorkerCalibration::new(
                calibration.minimum_entries,
                calibration.slow_work_ns_per_entry,
            ),
        }
    }
}

fn automatic_worker_pool(available: usize) -> WorkerPool {
    let initial = available.clamp(1, DEFAULT_SCAN_THREADS_CAP);
    // Preserve the serial fallback when the platform cannot report more than one
    // available processor. There is no measured basis for inventing parallelism there.
    if initial == 1 {
        return WorkerPool::fixed(1);
    }
    let maximum = available
        .saturating_mul(ADAPTIVE_SCAN_PARALLELISM_MULTIPLIER)
        .clamp(initial, ADAPTIVE_SCAN_THREADS_CAP);
    WorkerPool {
        initial,
        maximum,
        calibration: (maximum > initial).then_some(WorkerCalibration::new(
            ADAPTIVE_SCAN_CALIBRATION_ENTRIES,
            ADAPTIVE_SCAN_SLOW_WORK_NS_PER_ENTRY,
        )),
    }
}

/// Directories handed to a worker in one go.
///
/// Popping one directory at a time makes the queue lock the bottleneck on a wide,
/// shallow tree; taking a small run amortizes the lock without letting one worker
/// starve the others by hoarding the queue.
const DIR_CLAIM: usize = 4;

/// A parallel directory walk that produces exactly the observations the serial walk does.
///
/// The shape is deliberate. Workers read directories and *produce* observations; they
/// never touch an index. A single consumer — the caller's sink, on this thread —
/// applies them. That keeps the crate's one mutation contract intact: parallelism is a
/// property of the producer, and the index still sees one ordered stream of observations.
///
/// Ordering across independent subtrees is not fixed, but a directory observation is
/// published before that directory becomes claimable. The index therefore sees a
/// parent-first causal stream without imposing a global level barrier or serializing
/// filesystem work. The resulting index is byte-identical to the serial walker's,
/// which the benchmark harness re-proves on every trial by comparing engine digests
/// against an independent oracle.
#[allow(clippy::too_many_arguments)]
fn scan_concurrent(
    root: &Path,
    config: &ScanConfig,
    root_dev: u64,
    sink: &mut dyn FnMut(ScannerBatch),
    pool: WorkerPool,
    diagnostics: Option<&std::sync::Arc<ScanDiagnosticsRecorder>>,
    policy: WorkerPolicyExperiment,
    sink_mode: SinkMode,
) -> ScanReport {
    let mut consume = |message| match message {
        WalkMessage::Batch(batch) => {
            if let Some(diagnostics) = diagnostics {
                diagnostics.handoff_received();
            }
            sink(batch);
        }
        WalkMessage::DetachedDirectories { .. } => {
            unreachable!("the streaming walker never publishes detached directories")
        }
        WalkMessage::ScaleUp { .. } => {
            unreachable!("the shared runner consumes scale-up messages")
        }
    };
    run_concurrent_walk(
        root,
        config,
        root_dev,
        pool,
        diagnostics,
        policy,
        match sink_mode {
            SinkMode::Retained => walk_worker,
            SinkMode::TransientFold => walk_worker_transient_fold,
        },
        &mut consume,
    )
}

/// Parallel cold walk for a detached index that has no streaming consumer.
///
/// Workers publish directory-shaped facts before making their children claimable. The
/// caller consumes those groups into a private builder while filesystem work continues,
/// preserving parent-first causality and pipeline overlap without sending one full path
/// or public observation per entry.
fn scan_concurrent_detached(
    root: &Path,
    config: &ScanConfig,
    root_dev: u64,
    pool: WorkerPool,
    diagnostics: Option<&std::sync::Arc<ScanDiagnosticsRecorder>>,
    policy: WorkerPolicyExperiment,
    retention: Option<TreeRetention>,
) -> Result<(ScanReport, DetachedIndexBuilder)> {
    let mut builder = detached_builder(root, config, retention);
    let mut build_error = None;
    let output = {
        let mut consume = |message| match message {
            WalkMessage::Batch(_) => {
                unreachable!("the detached walker never publishes scanner batches")
            }
            WalkMessage::DetachedDirectories { mut directories, recycle } => {
                if let Some(diagnostics) = diagnostics {
                    diagnostics.handoff_received();
                }
                if build_error.is_none() {
                    for directory in &mut directories {
                        if let Err(error) = builder.push_directory(directory) {
                            build_error = Some(error);
                            break;
                        }
                    }
                }
                // A worker that has already left has dropped its receiver, and then the
                // listings are freed here, as every listing was before H159.
                let _ = recycle.send(directories);
            }
            WalkMessage::ScaleUp { .. } => {
                unreachable!("the shared runner consumes scale-up messages")
            }
        };
        // A folded index takes H72's listing policy (H185); the full index stats every
        // entry, since its directory attributes are the cache's freshness fingerprint.
        let worker: WalkWorker =
            if retention.is_some() { walk_detached_folding_worker } else { walk_detached_worker };
        run_concurrent_walk(root, config, root_dev, pool, diagnostics, policy, worker, &mut consume)
    };
    if let Some(error) = build_error {
        return Err(error);
    }
    Ok((output, builder))
}

type WalkWorker = fn(
    &Path,
    &ScanConfig,
    u64,
    &DirectoryQueue,
    &std::sync::mpsc::Sender<WalkMessage>,
    Option<&std::sync::Arc<ScanDiagnosticsRecorder>>,
) -> ScanReport;

/// Run the shared pool, scaling controller, diagnostics, and report reduction.
///
/// Streaming and detached scans differ only in their worker emission and main-thread
/// consumer. Keeping orchestration here prevents fixes to termination, diagnostics, or
/// panic handling from diverging between the two cold paths.
#[allow(clippy::too_many_arguments)]
fn run_concurrent_walk<C>(
    root: &Path,
    config: &ScanConfig,
    root_dev: u64,
    pool: WorkerPool,
    diagnostics: Option<&std::sync::Arc<ScanDiagnosticsRecorder>>,
    policy: WorkerPolicyExperiment,
    worker: WalkWorker,
    consume: &mut C,
) -> ScanReport
where
    C: FnMut(WalkMessage),
{
    let diagnostics = diagnostics.cloned();
    let queue = DirectoryQueue::new_with_policy(
        (PathBuf::new(), 0),
        config.order,
        pool.calibration,
        diagnostics.clone(),
        pool.initial,
        pool.maximum,
        policy,
    );
    let (sender, receiver) = std::sync::mpsc::channel::<WalkMessage>();

    let mut report = std::thread::scope(|scope| {
        let mut handles: Vec<_> = (0..pool.initial)
            .map(|_| {
                let sender = sender.clone();
                let queue = &queue;
                let diagnostics = diagnostics.clone();
                scope.spawn(move || {
                    worker(root, config, root_dev, queue, &sender, diagnostics.as_ref())
                })
            })
            .collect();
        // The loop below ends when every sender is gone, so this one must go first.
        drop(sender);

        let mut spawned_workers = pool.initial;
        for message in receiver {
            match message {
                WalkMessage::ScaleUp { sender, target_workers }
                    if target_workers > spawned_workers =>
                {
                    let target_workers = target_workers.min(pool.maximum);
                    record_adaptive_worker_expansion(diagnostics.as_ref());
                    for _ in spawned_workers..target_workers {
                        let sender = sender.clone();
                        let queue = &queue;
                        let diagnostics = diagnostics.clone();
                        handles.push(scope.spawn(move || {
                            worker(root, config, root_dev, queue, &sender, diagnostics.as_ref())
                        }));
                    }
                    spawned_workers = target_workers;
                }
                WalkMessage::ScaleUp { .. } => {}
                output => consume(output),
            }
        }

        // A walk that ends before its calibration window fills never observed enough to
        // decide anything. That is an *unobservable* policy, not a decision to hold the
        // initial pool, and an artifact that conflated the two would report a held pool
        // as if the walk had measured one and chosen it.
        let queue_finish = {
            let mut state = queue.lock();
            let mut trailing_window = None;
            if let Some(controller) = &state.controller {
                let (window, decision) = controller.partial_window();
                if decision == WorkerPolicyDecision::Undecided {
                    crate::counters::bump(|counts| {
                        counts.adaptive_policy_undecided =
                            counts.adaptive_policy_undecided.saturating_add(1);
                    });
                }
                if diagnostics.is_some() {
                    let sequence = state.allocate_policy_sequence();
                    trailing_window = Some(PolicyWindowSnapshot {
                        sequence,
                        start_entry_ordinal: window.start_entry_ordinal,
                        end_entry_ordinal: window.end_entry_ordinal,
                        observed_entries: window.entries,
                        observed_chunks: window.chunks,
                        observed_work_ns: window.work_ns,
                        ready_directories: state.ready_directories,
                        in_flight_directories: state.in_flight_directories,
                        active_workers: diagnostics.as_ref().map_or(0, |diagnostics| {
                            diagnostics.active_workers.load(std::sync::atomic::Ordering::Relaxed)
                        }),
                        handoff_backlog: diagnostics.as_ref().map_or(0, |diagnostics| {
                            diagnostics.handoff_backlog.load(std::sync::atomic::Ordering::Relaxed)
                        }),
                        requested_workers: None,
                        decision,
                    });
                }
            } else if diagnostics.is_some() {
                let shadow = state.shadow_calibration.as_ref().and_then(|shadow| {
                    (shadow.entries > 0).then_some((
                        shadow.window_start,
                        shadow.entries,
                        shadow.chunks,
                        shadow.work_ns,
                    ))
                });
                if let Some((window_start, entries, chunks, work_ns)) = shadow {
                    let sequence = state.allocate_policy_sequence();
                    trailing_window = Some(PolicyWindowSnapshot {
                        sequence,
                        start_entry_ordinal: window_start,
                        end_entry_ordinal: window_start.saturating_add(entries),
                        observed_entries: entries,
                        observed_chunks: chunks,
                        observed_work_ns: work_ns,
                        ready_directories: state.ready_directories,
                        in_flight_directories: state.in_flight_directories,
                        active_workers: diagnostics.as_ref().map_or(0, |diagnostics| {
                            diagnostics.active_workers.load(std::sync::atomic::Ordering::Relaxed)
                        }),
                        handoff_backlog: diagnostics.as_ref().map_or(0, |diagnostics| {
                            diagnostics.handoff_backlog.load(std::sync::atomic::Ordering::Relaxed)
                        }),
                        requested_workers: None,
                        decision: WorkerPolicyDecision::ObserveIncomplete,
                    });
                }
            }
            (state.ready_directories, state.in_flight_directories, trailing_window)
        };
        if let Some(diagnostics) = &diagnostics {
            diagnostics.record_queue_finish(queue_finish.0, queue_finish.1);
            if let Some(window) = queue_finish.2 {
                diagnostics.record_policy_window(window);
            }
        }

        let mut report = ScanReport::default();
        for handle in handles {
            match handle.join() {
                Ok(worker) => report.absorb(worker),
                Err(_) => {
                    // A worker panic leaves directories unaccounted for. Preserve that
                    // as a partial scan instead of reporting a short tree as complete.
                    report.errors.push(Error::io(
                        root,
                        std::io::Error::other("a scan worker thread panicked"),
                    ));
                }
            }
        }
        report
    });

    // Workers finish in filesystem order, so normalize errors before they escape.
    report.errors.sort_by_cached_key(ToString::to_string);
    report
}

/// Compile-time adapter for the one directory walker.
///
/// The filesystem, queue, admission, and diagnostics logic stays singular. Generic
/// emission keeps the public streaming path and private detached path branch-free in
/// their per-entry loops after monomorphization.
trait WalkEmission {
    type Directory;

    fn begin_directory(&mut self, path: &Path) -> Self::Directory;

    #[allow(clippy::too_many_arguments)]
    fn record_entry(
        &mut self,
        root: &Path,
        rel_dir: &Path,
        depth: usize,
        region: RegionId,
        name: &OsStr,
        kind: EntryKind,
        attrs: Attrs,
        root_dev: u64,
        config: &ScanConfig,
        directory: &mut Self::Directory,
        discovered: &mut Vec<(PathBuf, usize, RegionId)>,
        report: &mut ScanReport,
        sender: &std::sync::mpsc::Sender<WalkMessage>,
        chunk_send_ns: &mut u64,
        diagnostics: Option<&ScanDiagnosticsRecorder>,
    ) -> bool;

    fn finish_directory(&mut self, directory: Self::Directory);

    fn publish_before_discovery(
        &mut self,
        has_discovered: bool,
        sender: &std::sync::mpsc::Sender<WalkMessage>,
        chunk_send_ns: &mut u64,
        diagnostics: Option<&ScanDiagnosticsRecorder>,
    ) -> bool;

    fn finish(
        &mut self,
        sender: &std::sync::mpsc::Sender<WalkMessage>,
        report: &mut ScanReport,
        diagnostics: Option<&ScanDiagnosticsRecorder>,
    );

    /// The transient summary and the folded index take directory and symlink kind from
    /// the listing, with default attributes (H72, H185); every other route stats them.
    fn skip_dir_symlink_stat(&self) -> bool {
        false
    }
}

struct StreamingEmission {
    batch: Vec<ObservationOp>,
    batch_size: usize,
    recycle_tx: Option<std::sync::mpsc::Sender<Vec<ObservationOp>>>,
    recycle_rx: Option<std::sync::mpsc::Receiver<Vec<ObservationOp>>>,
    skip_dir_symlink_stat: bool,
    /// Whether each listing's control goes ahead of its entries
    /// ([`SinkMode::groups_directories`]).
    group_directories: bool,
    /// Where the listing being recorded begins in `batch`.
    directory_start: usize,
    /// Whether that listing's control already leads its observations, so the batch may be
    /// sent before the listing ends.
    listing_settled: bool,
    /// Where that listing stands on its directory's control.
    listing_control: ListingControl,
}

/// Where the listing a grouping worker is recording stands on its directory's control.
#[derive(Clone, Copy, PartialEq, Eq, Debug)]
enum ListingControl {
    /// No listed spelling of `.gitignore` has read the directory's control yet.
    Unread,
    /// A listed spelling read the directory's control, or failed to; its observation, if
    /// any, is among the held ones.
    Listed,
    /// The held observations filled the batch before any listed spelling read the
    /// control, so the directory's control was looked up directly (`read_directory_control`)
    /// and stands for the whole listing, as in the narrowed-population walk.
    ///
    /// A spelling listed afterwards is recorded as a row and not read again, so the
    /// directory's control file is read once. On a tree nothing modifies during the walk
    /// the probed file is the listed one; if the file changes between the two, the listing
    /// keeps what the probe read, and an error the second read would have met is never
    /// met.
    Probed,
}

impl StreamingEmission {
    /// An emission with the properties [`SinkMode`] names for `mode` under `config`.
    ///
    /// The recycle channel returns drained `PathBuf` arenas to this worker so glibc
    /// frees them on the thread that allocated them.
    fn for_sink(config: &ScanConfig, mode: SinkMode) -> Self {
        let batch_size = config.batch_size;
        let (recycle_tx, recycle_rx) = if mode.recycles_batches() {
            let (tx, rx) = std::sync::mpsc::channel();
            (Some(tx), Some(rx))
        } else {
            (None, None)
        };
        Self {
            batch: Vec::with_capacity(batch_size),
            batch_size,
            recycle_tx,
            recycle_rx,
            skip_dir_symlink_stat: mode.skips_dir_symlink_stat(),
            group_directories: mode.groups_directories(config),
            directory_start: 0,
            listing_settled: false,
            listing_control: ListingControl::Unread,
        }
    }

    /// Send the batch once it is full, settling a grouped listing's control first.
    ///
    /// A grouped listing's observations are held until its control leads them: at the
    /// listing's end ([`WalkEmission::finish_directory`]), or here, when the batch fills
    /// first. So a batch never holds more than `batch_size` observations and one control,
    /// however long the listing, and a fill that finds no spelling of `.gitignore` listed
    /// yet probes for it once per listing: one metadata lookup, and a read on a hit
    /// (`read_directory_control`). Returns whether the consumer is still there.
    #[allow(clippy::too_many_arguments)]
    fn send_if_full(
        &mut self,
        root: &Path,
        rel_dir: &Path,
        config: &ScanConfig,
        report: &mut ScanReport,
        sender: &std::sync::mpsc::Sender<WalkMessage>,
        chunk_send_ns: &mut u64,
        diagnostics: Option<&ScanDiagnosticsRecorder>,
    ) -> bool {
        if self.batch.len() < self.batch_size {
            return true;
        }
        if self.group_directories && !self.listing_settled {
            self.settle_listing(root, rel_dir, config, report);
        }
        let send_started = std::time::Instant::now();
        let sent = self.send_full(sender, diagnostics);
        *chunk_send_ns += elapsed_ns(send_started);
        sent
    }

    /// Put the current listing's control ahead of its held observations.
    fn settle_listing(
        &mut self,
        root: &Path,
        rel_dir: &Path,
        config: &ScanConfig,
        report: &mut ScanReport,
    ) {
        self.listing_settled = true;
        if self.listing_control != ListingControl::Unread {
            controls_first(&mut self.batch[self.directory_start..]);
            return;
        }
        self.listing_control = ListingControl::Probed;
        let control = rel_dir.join(crate::control::CONTROL_FILE_NAME);
        // The probe is the lookup the rule names, so a hit stands for the directory
        // whichever spelling it resolved to, exactly as a listed spelling's read would.
        match read_directory_control(config, root, &control) {
            Ok(Some(op)) => {
                self.batch.insert(self.directory_start, ObservationOp::unconditional(op));
            }
            Ok(None) => {}
            Err(error) => report.errors.push(error),
        }
    }

    /// Whether the entries this listing still lists must not read their control: its
    /// directory's control was already probed, and one read stands for the listing, as in
    /// the narrowed-population walk (`read_listed_control_op`).
    fn control_probed(&self) -> bool {
        self.group_directories && self.listing_control == ListingControl::Probed
    }

    /// Note that a listed entry read this listing's control, or failed to.
    fn note_listed_control(&mut self, control: Option<&Op>, control_error: Option<&Error>) {
        if control.is_some() || control_error.is_some() {
            self.listing_control = ListingControl::Listed;
        }
    }

    fn wrap(&self, ops: Vec<ObservationOp>) -> ScannerBatch {
        match &self.recycle_tx {
            Some(recycle) => ScannerBatch::new(ops).with_recycle(recycle.clone()),
            None => ScannerBatch::new(ops),
        }
    }

    fn next_vec(&self) -> Vec<ObservationOp> {
        // The retained path keeps its pre-H147 shape: an empty vec that grows by
        // doubling. Pre-sizing every batch there was never measured, and the public
        // `scan` is what a library caller pays for.
        let Some(recycle_rx) = &self.recycle_rx else {
            return Vec::new();
        };
        let mut kept = None;
        while let Ok(mut recycled) = recycle_rx.try_recv() {
            recycled.clear();
            kept = Some(recycled);
        }
        kept.unwrap_or_else(|| Vec::with_capacity(self.batch_size))
    }

    fn send_full(
        &mut self,
        sender: &std::sync::mpsc::Sender<WalkMessage>,
        diagnostics: Option<&ScanDiagnosticsRecorder>,
    ) -> bool {
        let ops = std::mem::take(&mut self.batch);
        let sent = send_scanner_batch(sender, self.wrap(ops), diagnostics);
        self.batch = self.next_vec();
        // A listing sent partway continues at the start of the new batch.
        self.directory_start = 0;
        sent
    }
}

impl WalkEmission for StreamingEmission {
    type Directory = ();

    fn begin_directory(&mut self, _path: &Path) {
        self.directory_start = self.batch.len();
        self.listing_settled = false;
        self.listing_control = ListingControl::Unread;
    }

    #[allow(clippy::too_many_arguments)]
    fn record_entry(
        &mut self,
        root: &Path,
        rel_dir: &Path,
        depth: usize,
        region: RegionId,
        name: &OsStr,
        kind: EntryKind,
        attrs: Attrs,
        root_dev: u64,
        config: &ScanConfig,
        _directory: &mut Self::Directory,
        discovered: &mut Vec<(PathBuf, usize, RegionId)>,
        report: &mut ScanReport,
        sender: &std::sync::mpsc::Sender<WalkMessage>,
        chunk_send_ns: &mut u64,
        diagnostics: Option<&ScanDiagnosticsRecorder>,
    ) -> bool {
        record_walk_entry(
            root,
            rel_dir,
            depth,
            region,
            name,
            kind,
            attrs,
            root_dev,
            config,
            self,
            discovered,
            report,
            sender,
            chunk_send_ns,
            diagnostics,
        )
    }

    fn finish_directory(&mut self, _directory: Self::Directory) {
        if self.group_directories && !self.listing_settled {
            controls_first(&mut self.batch[self.directory_start..]);
        }
    }

    fn publish_before_discovery(
        &mut self,
        has_discovered: bool,
        sender: &std::sync::mpsc::Sender<WalkMessage>,
        chunk_send_ns: &mut u64,
        diagnostics: Option<&ScanDiagnosticsRecorder>,
    ) -> bool {
        if self.batch.is_empty() || !has_discovered {
            return true;
        }
        let send_started = std::time::Instant::now();
        let sent = self.send_full(sender, diagnostics);
        *chunk_send_ns += elapsed_ns(send_started);
        sent
    }

    fn finish(
        &mut self,
        sender: &std::sync::mpsc::Sender<WalkMessage>,
        report: &mut ScanReport,
        diagnostics: Option<&ScanDiagnosticsRecorder>,
    ) {
        if self.batch.is_empty() {
            return;
        }
        let send_started = std::time::Instant::now();
        // The walk is over: do not ask `send_full` for a replacement vec that no
        // later `record_entry` would use.
        let ops = std::mem::take(&mut self.batch);
        let _ = send_scanner_batch(sender, self.wrap(ops), diagnostics);
        self.batch = Vec::new();
        report.attribution.send_ns += elapsed_ns(send_started);
    }

    fn skip_dir_symlink_stat(&self) -> bool {
        self.skip_dir_symlink_stat
    }
}

/// Emptied listings one detached worker keeps for reuse: one chunk's worth.
///
/// A bound on retention, not a tuned speed value. A chunk claims at most [`DIR_CLAIM`]
/// directories, and the worker takes returned listings back once per chunk, so this
/// covers the next chunk; a listing returned past it is freed at once, still on its own
/// thread, which is the property H159 needs. Retained listings are memory the consumer
/// can no longer reuse for the index: four chunks' worth of listings up to 256 children
/// each cost 1.0 MiB (+1.4%) of peak RSS on a 158,705-entry macOS subject and 2.3 MiB
/// (+5.0%) on a 77,159-entry one.
const DETACHED_SPARE_LISTINGS: usize = DIR_CLAIM;

/// The largest child buffer, in children, a spare listing may keep.
///
/// Also a bound on retention rather than a tuned value: most directories are small, and
/// a listing whose buffer grew past this, at about 80 bytes per child, is freed when it
/// comes back, on its own thread, instead of being pinned for the rest of the walk.
const DETACHED_SPARE_CHILD_CAPACITY: usize = 64;

/// One worker's detached emission: listings built here and published to the consumer.
///
/// Every path and child buffer in a listing is allocated on this worker's thread. The
/// consumer drains each listing into the index and sends the emptied listings back
/// (H159), and the worker reuses them or frees them itself. Under glibc a chunk freed on
/// another thread goes back to the arena that allocated it, under that arena's lock,
/// while this worker is allocating from it; the 2026-09-27 Linux comparison's allocator
/// screen and context-switch profile point at that contention for the index tier's gap
/// to its peers. A reused listing carries exactly the facts a fresh one would: the same
/// path bytes, and children and control only from this directory's listing.
struct DetachedEmission {
    /// H72's listing policy for the folded index (H185): a directory or symlink takes its
    /// kind from `d_type` and default attributes, as the transient summary does. A
    /// one-shot tree report reads no directory's or symlink's own attributes: its rows
    /// carry roll-ups, `newest_mtime_ns` is the files', symlinks and other kinds
    /// contribute nothing, and `dev` is read only under `--one-filesystem`, where the
    /// policy keeps the stat. The full index keeps every stat: its directory attributes
    /// are the cache's freshness fingerprint.
    skip_dir_symlink_stat: bool,
    directories: Vec<DetachedDirectory>,
    /// Emptied listings ready to be reused by [`WalkEmission::begin_directory`].
    spare: Vec<DetachedDirectory>,
    /// An emptied list to publish the next chunk's listings in.
    spare_list: Vec<DetachedDirectory>,
    recycle_tx: std::sync::mpsc::Sender<Vec<DetachedDirectory>>,
    recycle_rx: std::sync::mpsc::Receiver<Vec<DetachedDirectory>>,
}

impl DetachedEmission {
    fn new(skip_dir_symlink_stat: bool) -> Self {
        let (recycle_tx, recycle_rx) = std::sync::mpsc::channel();
        Self {
            skip_dir_symlink_stat,
            directories: Vec::new(),
            spare: Vec::new(),
            spare_list: Vec::new(),
            recycle_tx,
            recycle_rx,
        }
    }

    /// Take back every list the consumer has returned since the last chunk.
    ///
    /// A listing the consumer skipped, after a build error or for a repeated directory,
    /// comes back with its children and control still in it; they are dropped here, on
    /// the thread that allocated them, before the listing can be reused.
    fn collect_returned(&mut self) {
        while let Ok(mut returned) = self.recycle_rx.try_recv() {
            for mut directory in returned.drain(..) {
                directory.children.clear();
                directory.control = None;
                if self.spare.len() < DETACHED_SPARE_LISTINGS
                    && directory.children.capacity() <= DETACHED_SPARE_CHILD_CAPACITY
                {
                    self.spare.push(directory);
                }
            }
            if self.spare_list.capacity() == 0 {
                self.spare_list = returned;
            }
        }
    }
}

impl WalkEmission for DetachedEmission {
    type Directory = DetachedDirectory;

    fn skip_dir_symlink_stat(&self) -> bool {
        self.skip_dir_symlink_stat
    }

    fn begin_directory(&mut self, path: &Path) -> Self::Directory {
        let Some(mut directory) = self.spare.pop() else {
            return DetachedDirectory {
                path: path.to_path_buf(),
                children: Vec::new(),
                control: None,
            };
        };
        // The same bytes `to_path_buf` would copy, into a buffer this thread already owns.
        let buffer = directory.path.as_mut_os_string();
        buffer.clear();
        buffer.push(path);
        debug_assert!(directory.children.is_empty() && directory.control.is_none());
        directory
    }

    #[allow(clippy::too_many_arguments)]
    fn record_entry(
        &mut self,
        root: &Path,
        rel_dir: &Path,
        depth: usize,
        region: RegionId,
        name: &OsStr,
        kind: EntryKind,
        attrs: Attrs,
        root_dev: u64,
        config: &ScanConfig,
        directory: &mut Self::Directory,
        discovered: &mut Vec<(PathBuf, usize, RegionId)>,
        report: &mut ScanReport,
        _sender: &std::sync::mpsc::Sender<WalkMessage>,
        _chunk_send_ns: &mut u64,
        _diagnostics: Option<&ScanDiagnosticsRecorder>,
    ) -> bool {
        record_detached_entry(
            root,
            rel_dir,
            depth,
            region,
            name,
            kind,
            attrs,
            root_dev,
            config,
            &mut directory.children,
            &mut directory.control,
            discovered,
            report,
        );
        true
    }

    fn finish_directory(&mut self, directory: Self::Directory) {
        self.directories.push(directory);
    }

    fn publish_before_discovery(
        &mut self,
        _has_discovered: bool,
        sender: &std::sync::mpsc::Sender<WalkMessage>,
        chunk_send_ns: &mut u64,
        diagnostics: Option<&ScanDiagnosticsRecorder>,
    ) -> bool {
        if self.directories.is_empty() {
            return true;
        }
        // Taking back returned listings is handoff work, timed with the send so the
        // chunk's work time, which calibrates the worker pool, stays the walk's own.
        let send_started = std::time::Instant::now();
        self.collect_returned();
        let next = std::mem::take(&mut self.spare_list);
        let directories = std::mem::replace(&mut self.directories, next);
        let sent =
            send_detached_directories(sender, directories, self.recycle_tx.clone(), diagnostics);
        *chunk_send_ns += elapsed_ns(send_started);
        sent
    }

    fn finish(
        &mut self,
        _sender: &std::sync::mpsc::Sender<WalkMessage>,
        _report: &mut ScanReport,
        _diagnostics: Option<&ScanDiagnosticsRecorder>,
    ) {
    }
}

fn walk_detached_worker(
    root: &Path,
    config: &ScanConfig,
    root_dev: u64,
    queue: &DirectoryQueue,
    sender: &std::sync::mpsc::Sender<WalkMessage>,
    diagnostics: Option<&std::sync::Arc<ScanDiagnosticsRecorder>>,
) -> ScanReport {
    walk_detached_with(root, config, root_dev, queue, sender, diagnostics, false)
}

/// [`walk_detached_worker`] for a folded index, which describes each directory once, by
/// its own listing (H185): see [`DetachedEmission::skip_dir_symlink_stat`].
fn walk_detached_folding_worker(
    root: &Path,
    config: &ScanConfig,
    root_dev: u64,
    queue: &DirectoryQueue,
    sender: &std::sync::mpsc::Sender<WalkMessage>,
    diagnostics: Option<&std::sync::Arc<ScanDiagnosticsRecorder>>,
) -> ScanReport {
    walk_detached_with(root, config, root_dev, queue, sender, diagnostics, true)
}

fn walk_detached_with(
    root: &Path,
    config: &ScanConfig,
    root_dev: u64,
    queue: &DirectoryQueue,
    sender: &std::sync::mpsc::Sender<WalkMessage>,
    diagnostics: Option<&std::sync::Arc<ScanDiagnosticsRecorder>>,
    skip_dir_symlink_stat: bool,
) -> ScanReport {
    let report = walk_worker_with(
        root,
        config,
        root_dev,
        queue,
        sender,
        diagnostics,
        DetachedEmission::new(skip_dir_symlink_stat),
    );
    // A walker leaves only when the queue is empty with nothing in flight, or when its
    // consumer is gone, so the walk is over. The index may still be assembling the
    // listings already sent; the counters have stopped, and the phase says why.
    if let Some(progress) = &config.progress {
        progress.enter(crate::ProgressPhase::Indexing);
    }
    report
}

/// One worker's share of the public observation walk.
fn walk_worker(
    root: &Path,
    config: &ScanConfig,
    root_dev: u64,
    queue: &DirectoryQueue,
    sender: &std::sync::mpsc::Sender<WalkMessage>,
    diagnostics: Option<&std::sync::Arc<ScanDiagnosticsRecorder>>,
) -> ScanReport {
    walk_worker_with(
        root,
        config,
        root_dev,
        queue,
        sender,
        diagnostics,
        StreamingEmission::for_sink(config, SinkMode::Retained),
    )
}

/// One worker's share of the transient summary walk.
fn walk_worker_transient_fold(
    root: &Path,
    config: &ScanConfig,
    root_dev: u64,
    queue: &DirectoryQueue,
    sender: &std::sync::mpsc::Sender<WalkMessage>,
    diagnostics: Option<&std::sync::Arc<ScanDiagnosticsRecorder>>,
) -> ScanReport {
    walk_worker_with(
        root,
        config,
        root_dev,
        queue,
        sender,
        diagnostics,
        StreamingEmission::for_sink(config, SinkMode::TransientFold),
    )
}

fn walk_worker_with<E: WalkEmission>(
    root: &Path,
    config: &ScanConfig,
    root_dev: u64,
    queue: &DirectoryQueue,
    sender: &std::sync::mpsc::Sender<WalkMessage>,
    diagnostics: Option<&std::sync::Arc<ScanDiagnosticsRecorder>>,
    mut emission: E,
) -> ScanReport {
    let _counter_guard = crate::counters::thread_flush_guard();
    let _worker_guard = diagnostics.map(ScanDiagnosticsRecorder::worker_guard);
    let worker_started = std::time::Instant::now();
    let mut report = ScanReport::default();
    let mut tally = ProgressTally::new(config.progress.as_ref());
    let mut claimed: Vec<(PathBuf, usize, RegionId)> = Vec::with_capacity(DIR_CLAIM);
    let mut discovered: Vec<(PathBuf, usize, RegionId)> = Vec::new();
    let mut consumer_gone = false;
    #[cfg(target_os = "macos")]
    let mut bulk_reader = macos_bulk::Reader::new();
    #[cfg(all(target_os = "linux", target_env = "gnu"))]
    let mut dents_reader = linux_dents::Reader::new();

    'walk: while let Some(claim) = queue.claim(&mut claimed, &mut report.attribution) {
        // One timing pair per claimed chunk, never per entry: the chunk is the unit
        // the amortization argument is made in, so it is the unit the evidence is
        // collected in.
        let chunk_started = std::time::Instant::now();
        let mut chunk_send_ns: u64 = 0;
        let entries_before = report.entries;
        for (rel_dir, depth, region) in claimed.drain(..) {
            let abs_dir = root.join(&rel_dir);
            let mut directory = emission.begin_directory(&rel_dir);
            #[cfg(target_os = "macos")]
            {
                if let Some(diagnostics) = diagnostics {
                    diagnostics.macos_bulk_attempted();
                }
                if let Some(entries) =
                    (!walk_hook_covers(&abs_dir)).then(|| bulk_reader.read(&abs_dir)).flatten()
                {
                    if let Some(diagnostics) = diagnostics {
                        diagnostics.macos_bulk_succeeded();
                    }
                    report.dirs_read += 1;
                    for entry in entries {
                        if !emission.record_entry(
                            root,
                            &rel_dir,
                            depth,
                            region,
                            &entry.name,
                            entry.kind,
                            entry.attrs,
                            root_dev,
                            config,
                            &mut directory,
                            &mut discovered,
                            &mut report,
                            sender,
                            &mut chunk_send_ns,
                            diagnostics.map(AsRef::as_ref),
                        ) {
                            consumer_gone = true;
                            break 'walk;
                        }
                    }
                    emission.finish_directory(directory);
                    continue;
                }
                if let Some(diagnostics) = diagnostics {
                    diagnostics.macos_bulk_fell_back();
                }
            }
            // Counted in the Linux backend fields: an attempt, then a success or a
            // fallback, and a fallback goes on to count as a portable attempt below, so
            // `dirs_read` is the native successes plus the portable reads.
            #[cfg(all(target_os = "linux", target_env = "gnu"))]
            {
                if let Some(diagnostics) = diagnostics {
                    diagnostics.linux_dents_attempted();
                }
                let policy = linux_dents::StatPolicy {
                    skip_dir_symlink_stat: emission.skip_dir_symlink_stat(),
                    one_filesystem: config.one_filesystem,
                };
                // Plain `if`, not `bool::then(|| …)`: the listing borrows the reader for the
                // loop.
                let listing = if walk_hook_covers(&abs_dir) {
                    None
                } else {
                    dents_reader.read(&abs_dir, policy)
                };
                if let Some(listing) = listing {
                    if let Some(diagnostics) = diagnostics {
                        diagnostics.linux_dents_succeeded();
                    }
                    report.dirs_read += 1;
                    for entry in listing {
                        let (kind, attrs) = match entry.outcome {
                            linux_dents::Outcome::Observed { kind, attrs } => (kind, attrs),
                            linux_dents::Outcome::Failed(error) => {
                                // The same path std's `DirEntry::path` builds: the listing
                                // path joined with the name.
                                report.errors.push(Error::io(abs_dir.join(entry.name), error));
                                continue;
                            }
                        };
                        if !emission.record_entry(
                            root,
                            &rel_dir,
                            depth,
                            region,
                            entry.name,
                            kind,
                            attrs,
                            root_dev,
                            config,
                            &mut directory,
                            &mut discovered,
                            &mut report,
                            sender,
                            &mut chunk_send_ns,
                            diagnostics.map(AsRef::as_ref),
                        ) {
                            consumer_gone = true;
                            break 'walk;
                        }
                    }
                    emission.finish_directory(directory);
                    continue;
                }
                if let Some(diagnostics) = diagnostics {
                    diagnostics.linux_dents_fell_back();
                }
            }

            crate::counters::bump(|c| c.dir_opens += 1);
            if let Some(diagnostics) = diagnostics {
                diagnostics.portable_attempted();
            }

            let listing = match fs::read_dir(&abs_dir) {
                Ok(listing) => {
                    if let Some(diagnostics) = diagnostics {
                        diagnostics.portable_succeeded();
                    }
                    listing
                }
                Err(e) => {
                    report.errors.push(Error::io(abs_dir, e));
                    continue;
                }
            };
            report.dirs_read += 1;

            let policy = ListingPolicy {
                skip_dir_symlink_stat: emission.skip_dir_symlink_stat(),
                one_filesystem: config.one_filesystem,
            };
            let mut searchability = Searchability::Unproven;
            for item in listing {
                let item = match item {
                    Ok(item) => item,
                    Err(e) => {
                        report.errors.push(Error::io(&abs_dir, e));
                        continue;
                    }
                };
                crate::counters::bump(|c| c.dir_entries += 1);
                let name = item.file_name();
                let (kind, attrs) =
                    match listed_child_kind_and_attrs(&item, policy, &mut searchability) {
                        Ok(Some(observed)) => observed,
                        Ok(None) => continue,
                        Err(error) => {
                            report.errors.push(Error::io(item.path(), error));
                            continue;
                        }
                    };
                if !emission.record_entry(
                    root,
                    &rel_dir,
                    depth,
                    region,
                    &name,
                    kind,
                    attrs,
                    root_dev,
                    config,
                    &mut directory,
                    &mut discovered,
                    &mut report,
                    sender,
                    &mut chunk_send_ns,
                    diagnostics.map(AsRef::as_ref),
                ) {
                    // The consumer is gone; nothing further will be read.
                    consumer_gone = true;
                    break 'walk;
                }
            }
            emission.finish_directory(directory);
        }
        // Publish facts that authorize newly discovered directories before making
        // those directories claimable. Both emission modes preserve this boundary.
        if !emission.publish_before_discovery(
            !discovered.is_empty(),
            sender,
            &mut chunk_send_ns,
            diagnostics.map(AsRef::as_ref),
        ) {
            report.attribution.send_ns += chunk_send_ns;
            report.attribution.work_ns += elapsed_ns(chunk_started).saturating_sub(chunk_send_ns);
            consumer_gone = true;
            break 'walk;
        }
        report.attribution.send_ns += chunk_send_ns;
        let chunk_work_ns = elapsed_ns(chunk_started).saturating_sub(chunk_send_ns);
        report.attribution.work_ns += chunk_work_ns;
        // Progress is reported per chunk for the same reason timing is: the chunk is
        // the unit of handoff, so it is the unit the shared counters are touched in.
        tally.flush(&report);

        // Publish new work before releasing the claim so a worker that finds nothing
        // new does not hold work that others could be doing.
        if !discovered.is_empty() {
            queue.extend(discovered.drain(..), &mut report.attribution);
        }
        if let Some(target_workers) = claim.release(
            report.entries.saturating_sub(entries_before),
            chunk_work_ns,
            &mut report.attribution,
        ) {
            // Carry a sender in-band so the consumer can create the reserve workers
            // without retaining a channel endpoint that would keep a small scan alive.
            // Only the release that completes a slow calibration returns true, so one
            // message expands the pool exactly once.
            let _ = sender.send(WalkMessage::ScaleUp { sender: sender.clone(), target_workers });
        }
    }

    if !consumer_gone {
        emission.finish(sender, &mut report, diagnostics.map(AsRef::as_ref));
    }
    // A worker that left mid-chunk because its consumer was gone still read what it
    // read, and the report it returns says so.
    tally.flush(&report);
    report.attribution.wall_ns = elapsed_ns(worker_started);
    report
}

#[allow(clippy::too_many_arguments)]
fn record_detached_entry(
    root: &Path,
    rel_dir: &Path,
    depth: usize,
    region: RegionId,
    name: &OsStr,
    kind: EntryKind,
    attrs: Attrs,
    root_dev: u64,
    config: &ScanConfig,
    children: &mut Vec<DetachedChild>,
    control: &mut Option<Op>,
    discovered: &mut Vec<(PathBuf, usize, RegionId)>,
    report: &mut ScanReport,
) {
    let disposition = crate::admission::decide(name, kind, config.hidden(), config.exclude_special);
    if disposition == crate::admission::Disposition::Reject {
        return;
    }
    // Construct a full path only for a spelling of the control name. The scanner's public
    // preparation builds one for every retained entry because that path escapes in an
    // observation; this private builder keeps ordinary children component-only.
    if config.read_controls && crate::control::control_spelling(name).is_some() {
        let path = rel_dir.join(name);
        match read_control_op(config, root, &path, kind) {
            // A listing can repeat the control name while the directory changes, and a
            // case-sensitive one can list `.gitignore` beside `.GITIGNORE`, whose lookup
            // reads the same file. The later read wins, as the builder keeps the later
            // observation of the entry.
            Ok(observed) => *control = observed,
            Err(error) => report.errors.push(error),
        }
    }
    if disposition != crate::admission::Disposition::Retain {
        return;
    }
    report.observe(kind, attrs);
    // Positions only order repeated names, and no real listing reaches `u32::MAX` entries.
    let position = u32::try_from(children.len()).unwrap_or(u32::MAX);
    children.push(DetachedChild { name: name.to_os_string(), kind, attrs, position });
    if should_descend(kind, attrs, depth, root_dev, config) {
        let child_region = if depth == 0 { RegionId::UNASSIGNED } else { region };
        discovered.push((rel_dir.join(name), depth + 1, child_region));
    }
}

/// One filesystem entry after the scan's shared admission, control, and descent rules.
pub(crate) struct PreparedWalkEntry {
    pub(crate) path: PathBuf,
    pub(crate) kind: EntryKind,
    pub(crate) attrs: Attrs,
    pub(crate) retained: bool,
    pub(crate) control: Option<Op>,
    pub(crate) descend: bool,
    pub(crate) control_error: Option<Error>,
}

/// Apply the producer-independent part of a directory walk to one verified entry.
///
/// Both blocking and opened-root scans call this after obtaining non-following metadata,
/// which keeps admission, fixed controls, and traversal boundaries from drifting.
#[allow(clippy::too_many_arguments)]
pub(crate) fn prepare_walk_entry(
    root: &Path,
    rel_dir: &Path,
    depth: usize,
    name: &OsStr,
    kind: EntryKind,
    attrs: Attrs,
    root_dev: u64,
    config: &ScanConfig,
) -> Option<PreparedWalkEntry> {
    prepare_walk_entry_reading(root, rel_dir, depth, name, kind, attrs, root_dev, config, true)
}

/// [`prepare_walk_entry`], reading the entry's control only when `read_control` allows.
///
/// A grouping emission whose listing already probed its directory's control passes
/// `false`, so a `.gitignore` listed afterwards is not read a second time
/// (`StreamingEmission::control_probed`).
#[allow(clippy::too_many_arguments)]
fn prepare_walk_entry_reading(
    root: &Path,
    rel_dir: &Path,
    depth: usize,
    name: &OsStr,
    kind: EntryKind,
    attrs: Attrs,
    root_dev: u64,
    config: &ScanConfig,
    read_control: bool,
) -> Option<PreparedWalkEntry> {
    let disposition = crate::admission::decide(name, kind, config.hidden(), config.exclude_special);
    if disposition == crate::admission::Disposition::Reject {
        return None;
    }
    let path = join_listed_name(rel_dir, name);
    let (control, control_error) = if read_control {
        match read_named_control_op(config, root, &path, name, kind) {
            Ok(control) => (control, None),
            Err(error) => (None, Some(error)),
        }
    } else {
        (None, None)
    };
    Some(PreparedWalkEntry {
        path,
        kind,
        attrs,
        retained: disposition == crate::admission::Disposition::Retain,
        control,
        descend: should_descend(kind, attrs, depth, root_dev, config),
        control_error,
    })
}

/// `rel_dir.join(name)`, allocated once at the joined length.
///
/// [`Path::join`] copies `rel_dir` at its exact length and pushes onto the copy, so the
/// separator grows it and, for a name longer than the directory, so does the name: a
/// `realloc` for nearly every entry a streaming walk prepares. The summary route's
/// walkers spent 760-820 instructions per entry joining, and 410-450 in `realloc`
/// against the detached route's 85-150 (H180). The same push onto a copy that already
/// fits both makes the same path, byte for byte, on every platform.
fn join_listed_name(rel_dir: &Path, name: &OsStr) -> PathBuf {
    let mut path = PathBuf::with_capacity(rel_dir.as_os_str().len() + 1 + name.len());
    path.as_mut_os_string().push(rel_dir);
    path.push(name);
    path
}

#[allow(clippy::too_many_arguments)]
fn record_walk_entry(
    root: &Path,
    rel_dir: &Path,
    depth: usize,
    region: RegionId,
    name: &OsStr,
    kind: EntryKind,
    attrs: Attrs,
    root_dev: u64,
    config: &ScanConfig,
    emission: &mut StreamingEmission,
    discovered: &mut Vec<(PathBuf, usize, RegionId)>,
    report: &mut ScanReport,
    sender: &std::sync::mpsc::Sender<WalkMessage>,
    chunk_send_ns: &mut u64,
    diagnostics: Option<&ScanDiagnosticsRecorder>,
) -> bool {
    let read_control = !emission.control_probed();
    let Some(prepared) = prepare_walk_entry_reading(
        root,
        rel_dir,
        depth,
        name,
        kind,
        attrs,
        root_dev,
        config,
        read_control,
    ) else {
        return true;
    };
    emission.note_listed_control(prepared.control.as_ref(), prepared.control_error.as_ref());
    let (mut control, control_error) = (prepared.control, prepared.control_error);
    if let Some(error) = control_error {
        report.errors.push(error);
    }
    // A grouping emission holds an entry's control ahead of the entry itself, so no send
    // can take the entry, `.gitignore` included, before the control that governs it.
    if !prepared.retained || emission.group_directories {
        if let Some(control) = control.take() {
            emission.batch.push(ObservationOp::unconditional(control));
            if !emission.send_if_full(
                root,
                rel_dir,
                config,
                report,
                sender,
                chunk_send_ns,
                diagnostics,
            ) {
                return false;
            }
        }
        if !prepared.retained {
            return true;
        }
    }
    report.observe(kind, attrs);
    // Only a directory the walk descends into needs its path twice, in its observation
    // and in the queue. Every other entry's path moves into its observation, so it is
    // allocated once and freed with its batch rather than copied and freed at once
    // (H180).
    let (path, descend_path) = if prepared.descend {
        (prepared.path.clone(), Some(prepared.path))
    } else {
        (prepared.path, None)
    };
    emission.batch.push(ObservationOp::unconditional(Op::Upsert { path, kind, attrs }));
    if !emission.send_if_full(root, rel_dir, config, report, sender, chunk_send_ns, diagnostics) {
        return false;
    }
    if let Some(control) = control {
        emission.batch.push(ObservationOp::unconditional(control));
        if !emission.send_if_full(root, rel_dir, config, report, sender, chunk_send_ns, diagnostics)
        {
            return false;
        }
    }
    if let Some(path) = descend_path {
        // A child of the root seeds a new region; everything deeper inherits its
        // parent's. Region membership therefore costs one integer copy and never
        // inspects a path.
        let child_region = if depth == 0 { RegionId::UNASSIGNED } else { region };
        discovered.push((path, depth + 1, child_region));
    }
    true
}

/// Move one listing's control observations ahead of its entries, each kind in its order.
///
/// A listing that repeats `.gitignore` while the directory changes keeps its reads in
/// order, so the later one still wins, as it does in the detached builder. Only the rare
/// listing that holds a control moves; every other listing costs one pass over the tags
/// it has just written.
fn controls_first(listing: &mut [ObservationOp]) {
    let mut placed = 0;
    for position in 0..listing.len() {
        if matches!(listing[position].op, Op::ControlUpsert { .. } | Op::ControlRemove { .. }) {
            listing[placed..=position].rotate_right(1);
            placed += 1;
        }
    }
}

/// Observe the control an entry at `path` stands for, if the scan's policy asks for
/// control state at all.
///
/// Every control observation goes through here -- each walk and reconcile site, and the
/// watch layer's verification -- or through [`read_named_control_op`], which shares its
/// gate, so the policy cannot be forgotten at one of them. A watch must honor it like a
/// scan does: its scope has to equal the index's, the scope carries this bit, and a
/// verifier that read control files regardless would grow a partial rule set, from
/// whichever sources events touched, under a scope that says there is none.
///
/// It is also where a listed name becomes a control read, by the rule the module
/// documentation of [`crate::control`] states: the directory's control is what a lookup
/// of `<dir>/.gitignore` resolves to. An entry named exactly `.gitignore` is that
/// lookup's target on every filesystem, so it is read through its own path with the kind
/// its listing observed, and costs nothing more than it did. A case variant such as
/// `.GITIGNORE` may or may not be the target, so the canonical path is looked up
/// ([`read_directory_control`]): the entry decides only whether to look. The observation
/// names the canonical path either way, and its bytes are the ones git reads. A name
/// that spells no control is answered without a system call.
pub(crate) fn read_control_op(
    config: &ScanConfig,
    root: &Path,
    path: &Path,
    kind: EntryKind,
) -> Result<Option<Op>> {
    read_spelled_control_op(config, root, path, kind, crate::control::path_control_spelling)
}

/// [`read_control_op`] for the entry `name` a listing just produced, at `path`, the
/// listed directory joined with `name`.
///
/// The walker already holds the name, so its spelling is tested on those bytes: a length
/// comparison for nearly every entry, as in the detached builder. Parsing the last
/// component back out of the joined path cost the transient summary about 270
/// instructions for every entry, on a tree with no `.gitignore` as much as on one with
/// many (H180). A listed name is one normal component, so it is the path's last one and
/// the decision is the same.
fn read_named_control_op(
    config: &ScanConfig,
    root: &Path,
    path: &Path,
    name: &OsStr,
    kind: EntryKind,
) -> Result<Option<Op>> {
    debug_assert_eq!(path.file_name(), Some(name), "a listed name ends its path");
    read_spelled_control_op(config, root, path, kind, |_| crate::control::control_spelling(name))
}

/// [`read_control_op`], with the spelling of `path`'s last component found by `spelling`
/// once the policy allows a read at all.
fn read_spelled_control_op(
    config: &ScanConfig,
    root: &Path,
    path: &Path,
    kind: EntryKind,
    spelling: impl FnOnce(&Path) -> Option<crate::control::ControlSpelling>,
) -> Result<Option<Op>> {
    if !config.read_controls {
        return Ok(None);
    }
    match spelling(path) {
        Some(crate::control::ControlSpelling::Exact) => {
            read_control_op_unconditional(root, path, kind, config.control_limits.budget)
        }
        Some(crate::control::ControlSpelling::Variant) => {
            read_directory_control(config, root, &crate::control::sibling_control_path(path))
        }
        None => Ok(None),
    }
}

/// Read a directory's control by looking up its canonical path, `<dir>/.gitignore`.
///
/// This is the rule itself: on a case-insensitive directory the lookup resolves to
/// whichever spelling the directory stores, as git's open does, and on a case-sensitive
/// one only to the exact name. `Ok(None)` means nothing resolves. It costs one metadata
/// lookup, and a read on a hit, so it runs only where a listing cannot stand in for it: a
/// narrowed population reads each directory's control before listing it, a classifying
/// transient fold whose batch fills first probes for it (`StreamingEmission::send_if_full`),
/// and a listed case variant resolves through it ([`read_control_op`]).
fn read_directory_control(
    config: &ScanConfig,
    root: &Path,
    control_path: &Path,
) -> Result<Option<Op>> {
    if !config.read_controls {
        return Ok(None);
    }
    let absolute = control_lookup_path(root, control_path);
    let found = look_up_control(&absolute, control_path, config.control_limits.budget);
    #[cfg(test)]
    {
        if let Some(error) =
            walk_hook(&absolute).and_then(|hook| hook(WalkHookPoint::ControlLookup(&absolute)))
        {
            return Err(Error::io(&absolute, error));
        }
    }
    found
}

/// The lookup [`read_directory_control`] makes, at `absolute`, of the control it names
/// `control_path`.
fn look_up_control(
    absolute: &Path,
    control_path: &Path,
    budget: Option<usize>,
) -> Result<Option<Op>> {
    let kind = match observe_path(absolute) {
        Ok((EntryKind::File, _)) => EntryKind::File,
        Ok(_) => EntryKind::Other,
        Err(error) if error.kind() == std::io::ErrorKind::NotFound => return Ok(None),
        Err(error) => return Err(Error::io(absolute, error)),
    };
    read_control_source(absolute, control_path, kind, budget)
}

/// [`read_directory_control`], answering a miss with the removal of the directory's rules.
///
/// For a producer that saw one spelling of the control name vanish or fail: whether the
/// directory still has a control is the lookup's question, and a removal of rules the
/// table does not hold is inert.
pub(crate) fn read_directory_control_or_removal(
    config: &ScanConfig,
    root: &Path,
    control_path: &Path,
) -> Result<Option<Op>> {
    if !config.read_controls {
        return Ok(None);
    }
    Ok(Some(
        read_directory_control(config, root, control_path)?
            .unwrap_or_else(|| Op::ControlRemove { path: control_path.to_path_buf() }),
    ))
}

/// Where a lookup of the control path `control_path` under `root` goes.
#[cfg(not(test))]
fn control_lookup_path(root: &Path, control_path: &Path) -> PathBuf {
    root.join(control_path)
}

/// Where a lookup of the control path `control_path` under `root` goes, which a test may
/// resolve as a case-insensitive directory would ([`install_case_folding_control_lookup`]).
#[cfg(test)]
fn control_lookup_path(root: &Path, control_path: &Path) -> PathBuf {
    let absolute = root.join(control_path);
    let folds = CASE_FOLDING_LOOKUPS
        .read()
        .unwrap_or_else(std::sync::PoisonError::into_inner)
        .iter()
        .any(|folded| absolute.starts_with(folded));
    let Some(directory) = absolute.parent().filter(|_| folds) else {
        return absolute;
    };
    // A case-insensitive directory holds at most one spelling, and the lookup returns it.
    // A test tree may hold two only where the host is case-sensitive, and then the exact
    // name is what the host itself resolves.
    if fs::symlink_metadata(&absolute).is_ok() {
        return absolute;
    }
    let Ok(mut names) = fs::read_dir(directory) else {
        return absolute;
    };
    names
        .find_map(|item| {
            let name = item.ok()?.file_name();
            crate::control::control_spelling(&name).map(|_| directory.join(name))
        })
        .unwrap_or(absolute)
}

/// Roots whose control lookups [`control_lookup_path`] resolves case-insensitively.
#[cfg(test)]
static CASE_FOLDING_LOOKUPS: std::sync::RwLock<Vec<PathBuf>> = std::sync::RwLock::new(Vec::new());

/// Removes its root from [`CASE_FOLDING_LOOKUPS`] when dropped.
#[cfg(test)]
#[must_use = "the lookups fold only until the guard is dropped"]
pub(crate) struct CaseFoldingGuard(Vec<PathBuf>);

#[cfg(test)]
impl Drop for CaseFoldingGuard {
    fn drop(&mut self) {
        CASE_FOLDING_LOOKUPS
            .write()
            .unwrap_or_else(std::sync::PoisonError::into_inner)
            .retain(|root| !self.0.contains(root));
    }
}

/// Resolve every control lookup under `root` as a case-insensitive directory would, until
/// the guard drops: `<dir>/.gitignore` opens whichever spelling `<dir>` stores.
///
/// A case-sensitive test host cannot make a case-insensitive directory (an ext4 casefold
/// directory needs a kernel built with Unicode support and an empty directory to flag), so
/// this is how the rule's case-insensitive branch runs on every CI runner. It changes only
/// the lookup the rule depends on: listings still show stored names, and every other path
/// resolves as the host resolves it. A real case-insensitive volume is tested where the
/// temporary directory is one. `root` matches as given and canonical.
#[cfg(test)]
pub(crate) fn install_case_folding_control_lookup(root: &Path) -> CaseFoldingGuard {
    let mut roots = vec![root.to_path_buf()];
    if let Ok(canonical) = root.canonicalize() {
        if canonical != root {
            roots.push(canonical);
        }
    }
    CASE_FOLDING_LOOKUPS
        .write()
        .unwrap_or_else(std::sync::PoisonError::into_inner)
        .extend(roots.iter().cloned());
    CaseFoldingGuard(roots)
}

fn read_listed_control_op(
    config: &ScanConfig,
    root: &Path,
    path: &Path,
    kind: EntryKind,
) -> Result<Option<Op>> {
    if config.population != crate::query::IgnoredEntries::Include
        && crate::control::path_control_spelling(path).is_some()
    {
        // The exclusion walk already looked the directory's control up before any
        // sibling, whichever spelling holds it.
        Ok(None)
    } else {
        read_control_op(config, root, path, kind)
    }
}

fn population_prunes(
    population: crate::query::IgnoredEntries,
    path: &Path,
    kind: EntryKind,
    disposition: crate::admission::Disposition,
    controls: Option<&crate::control::ControlTable>,
    unreadable_controls: &std::collections::BTreeSet<PathBuf>,
) -> bool {
    // The fixed control source stays in the entry tier even for `only`: removing its
    // entry would also remove the retained rule source during a watch reconciliation.
    // Every spelling is kept, because the listing cannot tell which one a case-insensitive
    // directory resolves `.gitignore` to without the lookup the walk made before it.
    if crate::control::path_control_spelling(path).is_some() {
        return false;
    }
    let Some(table) = controls else { return false };
    if !table.classification_known(path)
        || path.ancestors().skip(1).any(|ancestor| unreadable_controls.contains(ancestor))
    {
        return false;
    }
    let ignored = table.is_ignored(path, kind.is_dir());
    match population {
        crate::query::IgnoredEntries::Include => false,
        crate::query::IgnoredEntries::Exclude => ignored,
        crate::query::IgnoredEntries::Only => {
            !ignored && !kind.is_dir() && disposition != crate::admission::Disposition::ControlOnly
        }
    }
}

fn apply_discovery_control(table: &mut crate::control::ControlTable, op: &Op) -> Result<()> {
    match op {
        Op::ControlUpsert { path, source } => {
            table.upsert(path, source.clone())?;
        }
        Op::ControlRemove { path } => {
            table.remove(path)?;
        }
        _ => unreachable!("directory control probe emits only control operations"),
    }
    Ok(())
}

/// Read one fixed control source without allowing a raced or hostile file to allocate
/// beyond the index-wide control budget.
///
/// A file longer than the budget is read only to one byte past it. No table under that
/// budget can admit a source that long, since every retained byte is charged at least
/// once, so the truncated source it sends is refused for the budget rather than parsed.
///
/// Private to this module, so no caller elsewhere can step around the policy gate in
/// `read_control_op`.
fn read_control_op_unconditional(
    root: &Path,
    path: &Path,
    kind: EntryKind,
    budget: Option<usize>,
) -> Result<Option<Op>> {
    if !crate::control::is_control_file(path) {
        return Ok(None);
    }
    read_control_source(&root.join(path), path, kind, budget)
}

/// The most a control file's read buffer is sized from its metadata length (H189): every
/// `.gitignore` in the nominated subjects is under 16 KiB, and a larger one grows its
/// buffer by reading, as any file did before.
const CONTROL_READ_RESERVE_BYTES: usize = 64 * 1024;

/// Read the control at `absolute`, whose kind is `kind`, as an observation of `path`.
///
/// `absolute` is where the lookup went and `path` the canonical control path the
/// observation names; on a real filesystem the first is `root.join(path)`.
fn read_control_source(
    absolute: &Path,
    path: &Path,
    kind: EntryKind,
    budget: Option<usize>,
) -> Result<Option<Op>> {
    if kind != EntryKind::File {
        return Ok(Some(Op::ControlRemove { path: path.to_path_buf() }));
    }
    let file = open_control_file(absolute).map_err(|error| Error::io(absolute, error))?;
    let metadata = file.metadata().map_err(|error| Error::io(absolute, error))?;
    if !metadata.file_type().is_file() {
        return Ok(Some(Op::ControlRemove { path: path.to_path_buf() }));
    }
    let read_limit = budget
        .map_or(u64::MAX, |budget| u64::try_from(budget).unwrap_or(u64::MAX).saturating_add(1));
    // Room for the length the metadata already carries, within the read limit and a
    // fixed bound, plus the byte `read_to_end` reads to see the end (H189). `take` hides
    // the length from `read_to_end`, which otherwise grows from 32 bytes by doubling: eight
    // `read` calls for the 2 KiB root `.gitignore` of a kernel tree, two now. A file that
    // grew since its stat reads on as before; a longer one is bounded by the limit as
    // before, and its buffer is grown by `read_to_end` past the reservation, not sized
    // from its length.
    let reserve = usize::try_from(metadata.len().min(read_limit))
        .unwrap_or(usize::MAX)
        .min(CONTROL_READ_RESERVE_BYTES)
        .saturating_add(1);
    let mut source = Vec::with_capacity(reserve);
    file.take(read_limit).read_to_end(&mut source).map_err(|error| Error::io(absolute, error))?;
    crate::counters::bump(|counts| counts.control_reads = counts.control_reads.saturating_add(1));
    Ok(Some(Op::ControlUpsert { path: path.to_path_buf(), source }))
}

#[cfg(unix)]
fn open_control_file(path: &Path) -> std::io::Result<fs::File> {
    use std::os::unix::fs::OpenOptionsExt as _;

    fs::OpenOptions::new().read(true).custom_flags(libc::O_NONBLOCK | libc::O_NOFOLLOW).open(path)
}

#[cfg(not(unix))]
fn open_control_file(path: &Path) -> std::io::Result<fs::File> {
    fs::File::open(path)
}

fn send_scanner_batch(
    sender: &std::sync::mpsc::Sender<WalkMessage>,
    batch: ScannerBatch,
    diagnostics: Option<&ScanDiagnosticsRecorder>,
) -> bool {
    if let Some(diagnostics) = diagnostics {
        diagnostics.handoff_sent();
    }
    let sent = sender.send(WalkMessage::Batch(batch)).is_ok();
    if !sent {
        // Balance the reservation when the receiver disappeared before accepting it.
        if let Some(diagnostics) = diagnostics {
            diagnostics.handoff_received();
        }
    }
    sent
}

fn send_detached_directories(
    sender: &std::sync::mpsc::Sender<WalkMessage>,
    directories: Vec<DetachedDirectory>,
    recycle: std::sync::mpsc::Sender<Vec<DetachedDirectory>>,
    diagnostics: Option<&ScanDiagnosticsRecorder>,
) -> bool {
    if let Some(diagnostics) = diagnostics {
        diagnostics.handoff_sent();
    }
    let sent = sender.send(WalkMessage::DetachedDirectories { directories, recycle }).is_ok();
    if !sent {
        if let Some(diagnostics) = diagnostics {
            diagnostics.handoff_received();
        }
    }
    sent
}

/// Nanoseconds since `started`, saturating rather than panicking on the absurd.
fn elapsed_ns(started: std::time::Instant) -> u64 {
    u64::try_from(started.elapsed().as_nanos()).unwrap_or(u64::MAX)
}

/// A top-level subtree, used to spread workers across the breadth of the tree.
///
/// Every directory below the root belongs to the region seeded by its depth-1
/// ancestor, inherited from its parent rather than recomputed from its path. The root
/// itself is [`RegionId::ROOT`], which exists only to bootstrap.
#[derive(Clone, Copy, PartialEq, Eq, Debug)]
struct RegionId(usize);

impl RegionId {
    /// The root's own region, which exists only to bootstrap the walk.
    const ROOT: Self = Self(0);
    /// "Allocate a fresh region for this directory." Resolved by
    /// [`DirectoryQueueState::push`], which is the only place holding the lock that
    /// owns the region table.
    const UNASSIGNED: Self = Self(usize::MAX);
}

/// Directories still to read, plus enough state to know when the walk is finished.
///
/// The termination condition is the only subtle part: the queue being empty does not
/// mean the walk is done, because a worker that is mid-directory may be about to push
/// its children. So a worker holds a claim from the moment it takes work until the
/// moment it has published everything that work produced, and the walk ends only when
/// the queue is empty *and* no claim is outstanding.
///
/// # Why breadth-first is region-scheduled rather than a global FIFO
///
/// A global FIFO orders the *queue*, but claims are unordered: workers take whatever
/// is at the front, which on a real tree means several workers grinding through the
/// same top-level subtree while others sit untouched. Measured on the branching
/// fixture, that left a global-FIFO walk starting the same 7–8 of 12 subtrees at the
/// halfway mark as depth-first did — the ordering bought nothing a consumer could see.
/// It also made the pending set hold a whole level of the tree, which is where the
/// +1.5–3.7% peak RSS in exp-012 came from.
///
/// So breadth-first keeps work in per-region buckets and hands each free worker a
/// *different* region, round-robin. Within a region the bucket is LIFO, which restores
/// depth-first's locality and spine-bounded memory. Nothing waits on a level boundary:
/// if only one region has work, every worker takes it. The result is a scheduler whose
/// shallow preference is expressed in *which subtree a worker picks up*, not in the
/// order a single queue drains — which is the property progressive consumers actually
/// need.
struct DirectoryQueue {
    state: std::sync::Mutex<DirectoryQueueState>,
    ready: std::sync::Condvar,
    order: ScanOrder,
    diagnostics: Option<std::sync::Arc<ScanDiagnosticsRecorder>>,
}

/// One outstanding claim, held for exactly as long as the worker owes the queue the
/// work it took.
///
/// Giving the claim back is the queue's liveness condition, not a courtesy: [`claim`]
/// parks every other worker on the condvar while `outstanding` is nonzero, so a single
/// claim that is never returned stops the whole walk and the scoped join that waits on
/// it. That makes `Drop` the only safe place to put the release, because the paths that
/// skip a hand-written call are exactly the ones that matter — the `break` taken when
/// the consumer disconnects, and an unwinding panic inside a directory read.
///
/// [`claim`]: DirectoryQueue::claim
struct DirectoryClaim<'a> {
    queue: &'a DirectoryQueue,
    /// Directories represented by the claim, for exact in-flight accounting.
    directories: usize,
    /// Whether the worker already returned this claim through [`Self::release`].
    released: bool,
}

impl DirectoryClaim<'_> {
    /// Return the claim at the end of a completed chunk, feeding the chunk's own
    /// measurements to the shared calibration.
    ///
    /// Returns the queue's scale-up decision, which is why the normal path cannot be
    /// `Drop`: a destructor has neither the chunk's timing nor anywhere to put an
    /// answer.
    fn release(
        mut self,
        entries: u64,
        work_ns: u64,
        timing: &mut WalkAttribution,
    ) -> Option<usize> {
        self.released = true;
        self.queue.release(self.directories, entries, work_ns, timing)
    }
}

impl Drop for DirectoryClaim<'_> {
    fn drop(&mut self) {
        if self.released {
            return;
        }
        // An abandoned chunk: the consumer went away mid-directory, or a read panicked.
        // Either way the partial timing describes an aborted chunk rather than the cost
        // of reading directories, so it must not reach the calibration that sizes the
        // worker pool. Returning the claim is the whole job.
        self.queue.abandon(self.directories);
    }
}

struct DirectoryQueueState {
    /// Depth-first's single stack. Unused under breadth-first.
    pending: VecDeque<(PathBuf, usize, RegionId)>,
    /// Breadth-first's per-region work, indexed by [`RegionId`]. Each is a LIFO stack.
    regions: Vec<Vec<(PathBuf, usize, RegionId)>>,
    /// Regions with work, in round-robin order. A region appears at most once; the
    /// flag array is what keeps that true without scanning the ring.
    ready_ring: VecDeque<RegionId>,
    /// Whether each region is currently in `ready_ring`.
    enqueued: Vec<bool>,
    /// Directories currently available for a future claim.
    ready_directories: usize,
    /// Directories held by outstanding claims.
    in_flight_directories: usize,
    outstanding: usize,
    finished: bool,
    controller: Option<WorkerController>,
    /// Observation-only windows retained after the shipped one-shot decision.
    shadow_calibration: Option<RepeatedCalibration>,
    /// Completion-order sequence assigned under the queue lock.
    next_policy_sequence: u64,
    worker_target: usize,
    maximum_workers: usize,
}

impl DirectoryQueueState {
    fn seeded(
        root: (PathBuf, usize),
        order: ScanOrder,
        calibration: Option<WorkerCalibration>,
        initial_workers: usize,
        maximum_workers: usize,
        policy: WorkerPolicyExperiment,
    ) -> Self {
        let mut state = Self {
            pending: VecDeque::new(),
            regions: Vec::new(),
            ready_ring: VecDeque::new(),
            enqueued: Vec::new(),
            ready_directories: 0,
            in_flight_directories: 0,
            outstanding: 0,
            finished: false,
            controller: calibration.map(|value| WorkerController::new(value, policy)),
            shadow_calibration: None,
            next_policy_sequence: 0,
            worker_target: initial_workers,
            maximum_workers,
        };
        state.push((root.0, root.1, RegionId::ROOT), order);
        state
    }

    /// Push one directory into the structure the order uses.
    fn push(&mut self, item: (PathBuf, usize, RegionId), order: ScanOrder) {
        self.ready_directories = self.ready_directories.saturating_add(1);
        match order {
            ScanOrder::DepthFirst => self.pending.push_back(item),
            ScanOrder::BreadthFirst => {
                let region = if item.2 == RegionId::UNASSIGNED {
                    // One region per top-level subtree, numbered as they are found.
                    self.regions.len().max(1)
                } else {
                    item.2.0
                };
                if region >= self.regions.len() {
                    self.regions.resize_with(region + 1, Vec::new);
                    self.enqueued.resize(region + 1, false);
                }
                // Resolve the id *into* the item, so every directory discovered beneath
                // this one inherits a concrete region instead of the sentinel. Without
                // this the sentinel propagates and each directory allocates a region of
                // its own, degenerating the scheduler into round-robin over the whole
                // frontier.
                let mut item = item;
                item.2 = RegionId(region);
                self.regions[region].push(item);
                if !self.enqueued[region] {
                    self.enqueued[region] = true;
                    self.ready_ring.push_back(RegionId(region));
                }
            }
        }
    }

    /// Whether any work is available.
    fn is_empty(&self, order: ScanOrder) -> bool {
        match order {
            ScanOrder::DepthFirst => self.pending.is_empty(),
            ScanOrder::BreadthFirst => self.ready_ring.is_empty(),
        }
    }

    /// Take up to `limit` directories from the next region in the round-robin ring.
    ///
    /// Every region holding work is in the ring exactly once, so popping it always
    /// finds work and always moves to a *different* subtree than the previous claim.
    /// An earlier version preferred the caller's previous region for locality, which
    /// pinned each worker to one subtree: with twelve deep chains and six workers only
    /// six subtrees ever advanced, and depth-first — whose four-directory claims
    /// happen to fan across the root's children — spread wider than breadth-first did.
    /// Locality still comes from the claim being a run of directories out of one
    /// region; it must not come from a worker refusing to leave.
    fn take(
        &mut self,
        limit: usize,
        order: ScanOrder,
        into: &mut Vec<(PathBuf, usize, RegionId)>,
    ) -> usize {
        let before = into.len();
        match order {
            ScanOrder::DepthFirst => {
                let take = self.pending.len().min(limit);
                let start = self.pending.len() - take;
                into.extend(self.pending.drain(start..));
            }
            ScanOrder::BreadthFirst => {
                let Some(region) = self.ready_ring.pop_front() else { return 0 };
                self.enqueued[region.0] = false;
                let bucket = &mut self.regions[region.0];
                let take = bucket.len().min(limit);
                let start = bucket.len() - take;
                into.extend(bucket.drain(start..));
                // Re-arm the region only if work remains and it is not already queued,
                // so a busy region cannot appear twice and starve the others.
                if !bucket.is_empty() && !self.enqueued[region.0] {
                    self.enqueued[region.0] = true;
                    self.ready_ring.push_back(region);
                }
            }
        }
        into.len().saturating_sub(before)
    }

    fn allocate_policy_sequence(&mut self) -> u64 {
        let sequence = self.next_policy_sequence;
        self.next_policy_sequence = self.next_policy_sequence.saturating_add(1);
        sequence
    }
}

impl DirectoryQueue {
    /// Seed the queue with the root, which is region zero until its children fan out.
    #[cfg(test)]
    fn new(
        root: (PathBuf, usize),
        order: ScanOrder,
        calibration: Option<WorkerCalibration>,
        diagnostics: Option<std::sync::Arc<ScanDiagnosticsRecorder>>,
    ) -> Self {
        Self::new_with_policy(
            root,
            order,
            calibration,
            diagnostics,
            1,
            2,
            WorkerPolicyExperiment::ShippedOneShot,
        )
    }

    #[allow(clippy::too_many_arguments)]
    fn new_with_policy(
        root: (PathBuf, usize),
        order: ScanOrder,
        calibration: Option<WorkerCalibration>,
        diagnostics: Option<std::sync::Arc<ScanDiagnosticsRecorder>>,
        initial_workers: usize,
        maximum_workers: usize,
        policy: WorkerPolicyExperiment,
    ) -> Self {
        let state = DirectoryQueueState::seeded(
            root,
            order,
            calibration,
            initial_workers,
            maximum_workers,
            policy,
        );
        Self {
            state: std::sync::Mutex::new(state),
            ready: std::sync::Condvar::new(),
            order,
            diagnostics,
        }
    }

    /// Take up to [`DIR_CLAIM`] directories, blocking until there is work or the walk
    /// is over. Returns `None` once no more work will ever arrive.
    ///
    /// Time spent waiting is charged to `timing`: lock acquisition to `lock_wait_ns`
    /// when contended, condvar waits to `starved_ns`. The condvar span includes the
    /// lock re-acquisition on wake, which slightly overstates starvation rather than
    /// understating contention — the fail-honest direction for the number that is
    /// supposed to stay near zero.
    fn claim<'a>(
        &'a self,
        into: &mut Vec<(PathBuf, usize, RegionId)>,
        timing: &mut WalkAttribution,
    ) -> Option<DirectoryClaim<'a>> {
        let mut state = self.lock_timed(timing);
        loop {
            if !state.is_empty(self.order) {
                let directories = state.take(DIR_CLAIM, self.order, into);
                state.ready_directories = state.ready_directories.saturating_sub(directories);
                state.in_flight_directories =
                    state.in_flight_directories.saturating_add(directories);
                state.outstanding += 1;
                timing.claims += 1;
                return Some(DirectoryClaim { queue: self, directories, released: false });
            }
            if state.finished {
                return None;
            }
            if state.outstanding == 0 {
                state.finished = true;
                self.ready.notify_all();
                return None;
            }
            let started = std::time::Instant::now();
            state = self.ready.wait(state).unwrap_or_else(std::sync::PoisonError::into_inner);
            timing.starved_ns += elapsed_ns(started);
        }
    }

    fn extend(
        &self,
        directories: impl Iterator<Item = (PathBuf, usize, RegionId)>,
        timing: &mut WalkAttribution,
    ) {
        let mut state = self.lock_timed(timing);
        for item in directories {
            state.push(item, self.order);
        }
        drop(state);
        self.ready.notify_all();
    }

    /// Give up a claim whose chunk never finished. Wakes everyone if it was the last.
    ///
    /// Reached only from [`DirectoryClaim::drop`], where there is no `WalkAttribution`
    /// to charge and nothing worth charging: an abandoned chunk read some unknown
    /// fraction of its directories, so its lock wait says nothing about contention
    /// during the walk.
    fn abandon(&self, directories: usize) {
        let mut state = self.lock();
        state.outstanding -= 1;
        state.in_flight_directories = state.in_flight_directories.saturating_sub(directories);
        if state.outstanding == 0 && state.is_empty(self.order) {
            state.finished = true;
            drop(state);
            self.ready.notify_all();
        }
    }

    /// Give up a claim taken by [`claim`]. Wakes everyone if this was the last one.
    ///
    /// The chunk's own entry count and work time feed the shared service-time
    /// calibration, so the decision uses the timing the walk already collects for
    /// attribution rather than a second clock. Returns the new worker target when a
    /// controller requests expansion. A release that ends the walk returns no target,
    /// because there is no longer useful work for a reserve worker to take.
    fn release(
        &self,
        directories: usize,
        observed_entries: u64,
        observed_work_ns: u64,
        timing: &mut WalkAttribution,
    ) -> Option<usize> {
        let mut state = self.lock_timed(timing);
        // Whether this chunk reached the calibration at all. Only chunks released while
        // it is still live are policy history; later ones are ordinary walk work.
        let calibrating = state.controller.is_some();
        let one_shot = state.controller.as_ref().is_some_and(WorkerController::is_one_shot);
        let controller_spec = state.controller.as_ref().map(WorkerController::calibration_spec);
        let staged_gated = state.controller.as_ref().is_some_and(WorkerController::is_staged_gated);
        let completed_window = state
            .controller
            .as_mut()
            .and_then(|value| value.observe(observed_entries, observed_work_ns));
        let observed_window = state
            .shadow_calibration
            .as_mut()
            .and_then(|value| value.observe(observed_entries, observed_work_ns));
        let window_completed = completed_window.is_some();
        state.outstanding -= 1;
        state.in_flight_directories = state.in_flight_directories.saturating_sub(directories);
        let finished = state.outstanding == 0 && state.is_empty(self.order);
        if finished {
            state.finished = true;
        }
        let handoff_backlog = self.diagnostics.as_ref().map_or(0, |diagnostics| {
            diagnostics.handoff_backlog.load(std::sync::atomic::Ordering::Relaxed)
        });
        let mut requested_workers = None;
        let policy_window = completed_window.map(|window| {
            let useful_frontier =
                state.ready_directories.saturating_add(state.in_flight_directories);
            let decision = if !window.slow {
                WorkerPolicyDecision::Hold
            } else if finished {
                WorkerPolicyDecision::HoldNoUsefulWork
            } else if staged_gated && useful_frontier <= state.worker_target {
                WorkerPolicyDecision::HoldInsufficientFrontier
            } else if staged_gated && handoff_backlog >= state.worker_target {
                WorkerPolicyDecision::HoldHandoffBacklog
            } else {
                let target = if staged_gated {
                    state.worker_target.saturating_mul(2).min(state.maximum_workers)
                } else {
                    state.maximum_workers
                };
                if target > state.worker_target {
                    state.worker_target = target;
                    requested_workers = Some(target);
                    WorkerPolicyDecision::ScaleUp
                } else {
                    WorkerPolicyDecision::HoldInsufficientFrontier
                }
            };
            let sequence = state.allocate_policy_sequence();
            PolicyWindowSnapshot {
                sequence,
                start_entry_ordinal: window.start_entry_ordinal,
                end_entry_ordinal: window.end_entry_ordinal,
                observed_entries: window.entries,
                observed_chunks: window.chunks,
                observed_work_ns: window.work_ns,
                ready_directories: state.ready_directories,
                in_flight_directories: state.in_flight_directories,
                active_workers: self.diagnostics.as_ref().map_or(0, |diagnostics| {
                    diagnostics.active_workers.load(std::sync::atomic::Ordering::Relaxed)
                }),
                handoff_backlog,
                requested_workers,
                decision,
            }
        });
        let shadow_window = observed_window.map(|window| {
            let sequence = state.allocate_policy_sequence();
            PolicyWindowSnapshot {
                sequence,
                start_entry_ordinal: window.start_entry_ordinal,
                end_entry_ordinal: window.end_entry_ordinal,
                observed_entries: window.entries,
                observed_chunks: window.chunks,
                observed_work_ns: window.work_ns,
                ready_directories: state.ready_directories,
                in_flight_directories: state.in_flight_directories,
                active_workers: self.diagnostics.as_ref().map_or(0, |diagnostics| {
                    diagnostics.active_workers.load(std::sync::atomic::Ordering::Relaxed)
                }),
                handoff_backlog,
                requested_workers: None,
                decision: if window.slow {
                    WorkerPolicyDecision::ObserveSlow
                } else {
                    WorkerPolicyDecision::ObserveFast
                },
            }
        });
        let controller_terminates = (one_shot || finished) && window_completed
            || requested_workers.is_some_and(|target| target == state.maximum_workers);
        if controller_terminates {
            // Continue observing after every terminal decision, including a candidate
            // that reached the maximum pool. Without this shadow history a slow-prefix
            // expansion makes a later fast phase unobservable, precisely the
            // irreversible over-expansion case the evidence matrix must detect.
            if let (Some(spec), Some(window)) = (controller_spec, completed_window) {
                if self.diagnostics.is_some() && !finished {
                    state.shadow_calibration =
                        Some(RepeatedCalibration::starting_at(spec, window.end_entry_ordinal));
                }
            }
            state.controller = None;
        }
        drop(state);

        // The sequence was assigned while the queue was locked, but the trace lock and
        // bounded-vector update stay outside that critical section. Recorder arrival
        // may differ from completion order; `finish` sorts the retained prefix.
        if let (Some(diagnostics), Some(window)) = (&self.diagnostics, policy_window) {
            diagnostics.record_policy_window(window);
        }
        if let (Some(diagnostics), Some(window)) = (&self.diagnostics, shadow_window) {
            diagnostics.record_policy_window(window);
        }

        // Recorded outside the lock: the sampling is off by default, and a disabled
        // counter must not lengthen the critical section it observes.
        if calibrating {
            record_adaptive_calibration_chunk(
                self.diagnostics.as_ref(),
                observed_entries,
                observed_work_ns,
            );
        }

        if finished {
            self.ready.notify_all();
        }
        requested_workers
    }

    /// Acquire the state lock, charging any contention to `timing`.
    ///
    /// The fast path is a `try_lock` that succeeds and costs one counter increment;
    /// only the contended path pays for reading the clock. Poisoning is tolerated for
    /// the same reason as [`Self::lock`].
    fn lock_timed(
        &self,
        timing: &mut WalkAttribution,
    ) -> std::sync::MutexGuard<'_, DirectoryQueueState> {
        timing.lock_ops += 1;
        match self.state.try_lock() {
            Ok(guard) => guard,
            Err(std::sync::TryLockError::Poisoned(poisoned)) => poisoned.into_inner(),
            Err(std::sync::TryLockError::WouldBlock) => {
                timing.lock_contended += 1;
                let started = std::time::Instant::now();
                let guard = self.lock();
                timing.lock_wait_ns += elapsed_ns(started);
                guard
            }
        }
    }

    /// A poisoned queue means a worker panicked mid-walk. The data behind the lock is
    /// a plain work list with no invariant that a panic could have broken, and the
    /// caller already reports the panic as a scan error, so recovering the list is
    /// strictly better than propagating a second panic into every other worker.
    fn lock(&self) -> std::sync::MutexGuard<'_, DirectoryQueueState> {
        self.state.lock().unwrap_or_else(std::sync::PoisonError::into_inner)
    }
}

fn record_adaptive_calibration_chunk(
    diagnostics: Option<&std::sync::Arc<ScanDiagnosticsRecorder>>,
    entries: u64,
    work_ns: u64,
) {
    if let Some(diagnostics) = diagnostics {
        diagnostics.calibration_chunk(entries, work_ns);
    }
    crate::counters::bump(|counts| {
        counts.adaptive_calibration_chunks = counts.adaptive_calibration_chunks.saturating_add(1);
        counts.adaptive_calibration_entries =
            counts.adaptive_calibration_entries.saturating_add(entries);
        counts.adaptive_calibration_work_us =
            counts.adaptive_calibration_work_us.saturating_add(work_ns / 1_000);
    });
}

fn record_adaptive_worker_expansion(diagnostics: Option<&std::sync::Arc<ScanDiagnosticsRecorder>>) {
    if let Some(diagnostics) = diagnostics {
        diagnostics.worker_expanded();
    }
    crate::counters::bump(|counts| {
        counts.adaptive_scale_ups = counts.adaptive_scale_ups.saturating_add(1);
    });
}

/// The private builder a detached walk consumes its listings into: a full index, or with
/// `retention` a folded one.
fn detached_builder(
    root: &Path,
    config: &ScanConfig,
    retention: Option<TreeRetention>,
) -> DetachedIndexBuilder {
    let builder = DetachedIndexBuilder::new(root, config.scope(), config.types_shared())
        .with_control_limits(config.control_limits);
    match retention {
        Some(retention) => builder.folding(retention),
        None => builder,
    }
}

fn scan_detached_directories(
    root: &Path,
    config: &ScanConfig,
    collect_diagnostics: bool,
    policy: WorkerPolicyExperiment,
    retention: Option<TreeRetention>,
) -> Result<(ScanReport, DetachedIndexBuilder, Option<ScanDiagnostics>)> {
    if let Some(progress) = &config.progress {
        progress.enter(crate::ProgressPhase::Scanning);
    }
    let root_metadata = {
        crate::counters::bump(|counts| counts.stats += 1);
        fs::symlink_metadata(root)
    }
    .map_err(|error| Error::io(root, error))?;
    if !root_metadata.is_dir() {
        return Err(Error::io(
            root,
            std::io::Error::new(std::io::ErrorKind::NotADirectory, "scan root is not a directory"),
        ));
    }
    let root_dev = root_device(root, &root_metadata).map_err(|error| Error::io(root, error))?;
    let available_parallelism =
        std::thread::available_parallelism().map_or(1, std::num::NonZero::get);
    let pool = config.worker_pool_for(available_parallelism);
    let diagnostics = collect_diagnostics
        .then(|| ScanDiagnosticsRecorder::new(pool, available_parallelism, policy));

    if config.max_depth == Some(0) {
        if let Some(diagnostics) = &diagnostics {
            diagnostics.mark_not_run();
            diagnostics.record_queue_finish(0, 0);
        }
        return Ok((
            ScanReport::default(),
            detached_builder(root, config, retention),
            diagnostics.as_ref().map(|value| value.finish()),
        ));
    }

    let walk_started = crate::counters::enabled().then(std::time::Instant::now);
    let (output, builder) = scan_concurrent_detached(
        root,
        config,
        root_dev,
        pool,
        diagnostics.as_ref(),
        policy,
        retention,
    )?;
    // Also reached by a walk no worker left, such as a single-threaded one, so every
    // cold index ends its walk in the same phase.
    if let Some(progress) = &config.progress {
        progress.enter(crate::ProgressPhase::Indexing);
    }
    if let Some(started) = walk_started {
        let elapsed = elapsed_ns(started) / 1_000;
        crate::counters::bump(|counts| {
            counts.detached_walk_us = counts.detached_walk_us.saturating_add(elapsed);
        });
    }
    Ok((output, builder, diagnostics.as_ref().map(|value| value.finish())))
}

fn consolidate_detached_index(
    mut output: ScanReport,
    builder: DetachedIndexBuilder,
) -> (Index, ScanReport) {
    let entries = output.entries;
    let consolidate_started = crate::counters::enabled().then(std::time::Instant::now);
    let mut index = builder.finish();
    if let Some(started) = consolidate_started {
        let elapsed = elapsed_ns(started) / 1_000;
        crate::counters::bump(|counts| {
            counts.detached_builds = counts.detached_builds.saturating_add(1);
            counts.detached_entries = counts.detached_entries.saturating_add(entries);
            counts.detached_finish_us = counts.detached_finish_us.saturating_add(elapsed);
        });
    }
    index.record_walk_errors(&mut output.errors);
    index.set_initial_detached_scan_freshness(&output.errors);
    (index, output)
}

/// Walk `root` and return a fully populated index.
pub fn scan_into_index(root: &Path, config: &ScanConfig) -> Result<(Index, ScanReport)> {
    config.validate()?;
    let root = root.canonicalize().map_err(|error| Error::io(root, error))?;
    if config.population != crate::query::IgnoredEntries::Include {
        let (index, report, _) = scan_into_index_with_scanner(
            &root,
            config,
            false,
            WorkerPolicyExperiment::ShippedOneShot,
        )?;
        return Ok((index, report));
    }
    let (output, builder, _diagnostics) = scan_detached_directories(
        &root,
        config,
        false,
        WorkerPolicyExperiment::ShippedOneShot,
        None,
    )?;
    Ok(consolidate_detached_index(output, builder))
}

/// Walk the canonical `root` into a folded index that keeps what `retention` names
/// ([`crate::execution::RetainedState::Tree`]), with the diagnostic trace when asked.
///
/// The same detached walk and builder as [`scan_into_index`], which a folded index
/// differs from only in the files it keeps. Crate-private because a folded index answers
/// one tree report and nothing else; its one caller reports from it and frees it.
pub(crate) fn scan_into_folded_index(
    root: &Path,
    config: &ScanConfig,
    retention: TreeRetention,
    collect_diagnostics: bool,
) -> Result<(Index, ScanReport, Option<ScanDiagnostics>)> {
    config.validate()?;
    debug_assert_eq!(
        config.population,
        crate::query::IgnoredEntries::Include,
        "a folded index keeps the whole population; a narrowed one takes the scanner"
    );
    let (output, builder, diagnostics) = scan_detached_directories(
        root,
        config,
        collect_diagnostics,
        WorkerPolicyExperiment::ShippedOneShot,
        Some(retention),
    )?;
    let (index, report) = consolidate_detached_index(output, builder);
    Ok((index, report, diagnostics))
}

fn scan_into_index_with_scanner(
    root: &Path,
    config: &ScanConfig,
    collect_diagnostics: bool,
    policy: WorkerPolicyExperiment,
) -> Result<(Index, ScanReport, Option<ScanDiagnostics>)> {
    let mut index = Index::new_with_scope_and_types(root, config.scope(), config.types_shared());
    index.set_control_limits(config.control_limits);
    let mut apply_error: Option<Error> = None;
    let (mut report, diagnostics) = scan_internal(
        root,
        config,
        &mut |batch| {
            if apply_error.is_none() {
                if let Err(error) = index.apply_scanner_baseline(batch) {
                    apply_error = Some(error);
                }
            }
        },
        collect_diagnostics,
        policy,
        SinkMode::Retained,
    )?;
    if let Some(error) = apply_error {
        return Err(error);
    }
    index.record_walk_errors(&mut report.errors);
    index.set_initial_scan_freshness(&report.errors);
    Ok((index, report, diagnostics))
}

#[cfg(test)]
fn scan_into_index_via_scanner(root: &Path, config: &ScanConfig) -> Result<(Index, ScanReport)> {
    let (index, report, _) =
        scan_into_index_with_scanner(root, config, false, WorkerPolicyExperiment::ShippedOneShot)?;
    Ok((index, report))
}

/// Walk `root` into an index and retain the opt-in diagnostic trace for that run.
///
/// The index and [`ScanReport`] have the same semantics as [`scan_into_index`]. The
/// additional trace is versioned independently so evidence tooling can fail closed on
/// changes without coupling cache or engine behavior to measurement details.
pub fn scan_into_index_with_diagnostics(
    root: &Path,
    config: &ScanConfig,
) -> Result<(Index, ScanReport, ScanDiagnostics)> {
    scan_into_index_with_policy_diagnostics(root, config, WorkerPolicyExperiment::ShippedOneShot)
}

/// Exercise a repository-only worker-controller candidate while building an index.
#[doc(hidden)]
pub fn scan_into_index_with_policy_diagnostics(
    root: &Path,
    config: &ScanConfig,
    policy: WorkerPolicyExperiment,
) -> Result<(Index, ScanReport, ScanDiagnostics)> {
    config.validate()?;
    let root = root.canonicalize().map_err(|error| Error::io(root, error))?;
    if config.population != crate::query::IgnoredEntries::Include {
        let (index, report, diagnostics) =
            scan_into_index_with_scanner(&root, config, true, policy)?;
        return Ok((index, report, diagnostics.expect("diagnostic scanner creates a recorder")));
    }
    let (output, builder, diagnostics) =
        scan_detached_directories(&root, config, true, policy, None)?;
    let (index, report) = consolidate_detached_index(output, builder);
    Ok((index, report, diagnostics.expect("diagnostic detached scan creates a recorder")))
}

/// Diff the filesystem against an existing index and emit conditional observations.
///
/// This is cache tier 2: after a snapshot is loaded, a sweep like this is what makes the
/// answer trustworthy rather than merely fast. Unchanged entries produce upserts whose
/// fingerprints already match, which the index discards as no-ops, so the caller can
/// apply the whole stream without filtering it first.
///
/// Entries the index holds but the filesystem no longer has become [`Op::Remove`],
/// detected per directory rather than by accumulating every visited path in memory.
///
/// This observation-only reference API assumes its emitted stream is applied to the same
/// unchanged baseline after the borrow ends. Use [`reconcile`] or [`reconcile_handle`]
/// when other producers can write concurrently; those paths capture the stronger
/// generation/revision/absence expectations returned by [`Index::expectation`].
pub fn revalidate(
    index: &Index,
    config: &ScanConfig,
    sink: &mut dyn FnMut(Observation),
) -> Result<ScanReport> {
    config.validate_for_scope(index.scope())?;
    let root = index.root_path().to_path_buf();
    let root_meta = {
        crate::counters::bump(|c| c.stats += 1);
        fs::symlink_metadata(&root)
    }
    .map_err(|error| Error::io(&root, error))?;
    if !root_meta.is_dir() {
        return Err(Error::io(
            &root,
            std::io::Error::new(
                std::io::ErrorKind::NotADirectory,
                "revalidation root is not a directory",
            ),
        ));
    }
    let root_dev = root_device(&root, &root_meta).map_err(|error| Error::io(&root, error))?;
    if let Some(progress) = &config.progress {
        progress.enter(crate::ProgressPhase::Revalidating);
    }
    let mut report = ScanReport::default();
    let mut tally = ProgressTally::new(config.progress.as_ref());
    let batch_limit = config.batch_size.max(1);
    let mut batch: Vec<ObservationOp> = Vec::with_capacity(batch_limit);
    if config.max_depth == Some(0) {
        if let Some(children) = index.children(Path::new("")) {
            for (name, _) in children {
                let path = PathBuf::from(name);
                batch.push(ObservationOp::if_state(
                    Op::Remove { path: path.clone() },
                    index.relaxed_expectation(&path),
                ));
                if batch.len() >= batch_limit {
                    sink(Observation::from_ops(std::mem::take(&mut batch)));
                    batch.reserve(batch_limit);
                }
            }
        }
        if !batch.is_empty() {
            sink(Observation::from_ops(batch));
        }
        return Ok(report);
    }
    let mut queue: VecDeque<(PathBuf, usize)> = VecDeque::from(vec![(PathBuf::new(), 0)]);
    let mut controls = (config.population != crate::query::IgnoredEntries::Include)
        .then(|| index.control_table().clone());
    let mut unreadable_controls = std::collections::BTreeSet::new();
    let mut readers = Readers::new();
    let policy = ListingPolicy::every_child_stated(config.one_filesystem);

    while let Some((rel_dir, depth)) = take_next(&mut queue, config.order) {
        let abs_dir = root.join(&rel_dir);
        let control_path = rel_dir.join(crate::control::CONTROL_FILE_NAME);
        let mut had_control = index.control_table().contains(&control_path);
        let mut listed_control = ListedControl::default();
        if let Some(table) = controls.as_mut() {
            let baseline = index.relaxed_expectation(&control_path);
            let lookup = read_directory_control(config, &root, &control_path);
            listed_control.lookup(&lookup);
            match lookup {
                Ok(Some(op)) => {
                    apply_discovery_control(table, &op)?;
                    batch.push(ObservationOp::if_state(op, baseline));
                }
                Ok(None) => {
                    table.remove(&control_path)?;
                    if had_control {
                        batch.push(ObservationOp::if_state(
                            Op::ControlRemove { path: control_path.clone() },
                            baseline,
                        ));
                        had_control = false;
                    }
                }
                Err(error) => {
                    unreadable_controls.insert(rel_dir.clone());
                    report.errors.push(error);
                }
            }
        }
        let listing = match list_directory(&mut readers, &abs_dir, policy, None) {
            Ok(listing) => listing,
            Err(e) => {
                report.errors.push(Error::io(abs_dir, e));
                continue;
            }
        };
        report.dirs_read += 1;

        let mut seen: BTreeSet<OsString> = BTreeSet::new();
        let mut listing_complete = true;
        let listing = reconcile_listing(listing, &abs_dir);
        for item in listing {
            let (name, observed) = match item {
                Listed::Child { name, observed } => (name, observed),
                Listed::Failed(e) => {
                    listing_complete = false;
                    report.errors.push(Error::io(&abs_dir, e));
                    continue;
                }
            };
            // Seeing the name proves it is not absent even when a following metadata
            // lookup fails. Record it before any fallible per-entry work so an
            // operational error cannot become a false removal in the missing sweep.
            seen.insert(name.to_os_string());
            let rel_path = rel_dir.join(&name);
            let baseline = index.relaxed_expectation(&rel_path);
            let (kind, attrs) = match observed {
                Ok(Some(observed)) => observed,
                Ok(None) => {
                    let entry_held = baseline.state != PathState::Absent;
                    for removal in
                        vanished_child_removals(&rel_dir, &name, entry_held, &mut had_control)
                            .into_iter()
                            .flatten()
                    {
                        batch.push(ObservationOp::if_state(removal, baseline));
                    }
                    if batch.len() >= batch_limit {
                        sink(Observation::from_ops(std::mem::take(&mut batch)));
                        batch.reserve(batch_limit);
                    }
                    continue;
                }
                Err(e) => {
                    listed_control.listed(&name);
                    report.errors.push(Error::io(abs_dir.join(&name), e));
                    continue;
                }
            };
            listed_control.listed(&name);
            let disposition =
                crate::admission::decide(&name, kind, config.hidden(), config.exclude_special);
            if population_prunes(
                config.population,
                &rel_path,
                kind,
                disposition,
                controls.as_ref(),
                &unreadable_controls,
            ) {
                if baseline.state != PathState::Absent {
                    batch.push(ObservationOp::if_state(Op::Remove { path: rel_path }, baseline));
                }
                continue;
            }
            let read = read_listed_control_op(config, &root, &rel_path, kind);
            listed_control.read(&name, &read);
            let control = match read {
                Ok(control) => control,
                Err(error) => {
                    report.errors.push(error);
                    None
                }
            };
            if disposition != crate::admission::Disposition::Retain {
                if baseline.state != PathState::Absent {
                    batch.push(ObservationOp::if_state(
                        Op::Remove { path: rel_path.clone() },
                        baseline,
                    ));
                }
                if disposition == crate::admission::Disposition::ControlOnly {
                    if let Some(control) = control {
                        let guard = control_guard(&rel_path, baseline, |path| {
                            index.relaxed_expectation(path)
                        });
                        batch.push(ObservationOp::if_state(control, guard));
                    }
                }
                if batch.len() >= batch_limit {
                    sink(Observation::from_ops(std::mem::take(&mut batch)));
                    batch.reserve(batch_limit);
                }
                continue;
            }
            report.observe(kind, attrs);
            batch.push(ObservationOp::if_state(
                Op::Upsert { path: rel_path.clone(), kind, attrs },
                baseline,
            ));
            if batch.len() >= batch_limit {
                sink(Observation::from_ops(std::mem::take(&mut batch)));
                batch.reserve(batch_limit);
            }
            if let Some(control) = control {
                let guard =
                    control_guard(&rel_path, baseline, |path| index.relaxed_expectation(path));
                batch.push(ObservationOp::if_state(control, guard));
                if batch.len() >= batch_limit {
                    sink(Observation::from_ops(std::mem::take(&mut batch)));
                    batch.reserve(batch_limit);
                }
            }

            if should_descend(kind, attrs, depth, root_dev, config) {
                queue.push_back((rel_path, depth + 1));
            } else if kind.is_dir() {
                if let Some(children) = index.children(&rel_path) {
                    for (child_name, _) in children {
                        let child_path = rel_path.join(child_name);
                        batch.push(ObservationOp::if_state(
                            Op::Remove { path: child_path.clone() },
                            index.relaxed_expectation(&child_path),
                        ));
                        if batch.len() >= batch_limit {
                            sink(Observation::from_ops(std::mem::take(&mut batch)));
                            batch.reserve(batch_limit);
                        }
                    }
                }
            }
        }

        // Anything the index still lists here but the filesystem did not return is gone.
        if listing_complete {
            if let Some(known) = index.children(&rel_dir) {
                for (name, _) in known {
                    if !seen.contains(name) {
                        let path = rel_dir.join(name);
                        batch.push(ObservationOp::if_state(
                            Op::Remove { path: path.clone() },
                            index.relaxed_expectation(&path),
                        ));
                        // In the same batch, so the rules the removal drops are back
                        // before any commit shows the directory without them.
                        if let Some(control) = listed_control.restatement_after_removing(name) {
                            batch.push(ObservationOp::if_state(
                                control,
                                index.relaxed_expectation(&control_path),
                            ));
                        }
                    }
                }
            }
            if had_control && !listed_control.seen {
                batch.push(ObservationOp::if_state(
                    Op::ControlRemove { path: control_path.clone() },
                    index.relaxed_expectation(&control_path),
                ));
            }
        }
        // Per directory: an unchanged tree fills no batch, so the batch cannot be the
        // unit here without the counters standing still for the whole walk.
        tally.flush(&report);
    }

    if !batch.is_empty() {
        sink(Observation::from_ops(batch));
    }
    tally.flush(&report);
    Ok(report)
}

/// Reconcile the full index and publish each exact commit as it lands.
pub fn reconcile(
    index: &mut Index,
    config: &ScanConfig,
    sink: &mut dyn FnMut(&Commit),
) -> Result<ReconcileReport> {
    reconcile_subtree(index, Path::new(""), config, sink)
}

/// Reconcile one relative subtree, applying effective changes during the walk.
///
/// If an ancestor vanished or became a non-directory, reconciliation widens to that
/// ancestor so a child invalidation can converge instead of retrying `ENOTDIR` forever.
pub fn reconcile_subtree(
    index: &mut Index,
    subtree: &Path,
    config: &ScanConfig,
    sink: &mut dyn FnMut(&Commit),
) -> Result<ReconcileReport> {
    reconcile_target(&mut ReconcileTarget::Direct(index), subtree, config, sink)
}

/// Reconcile a shared index while allowing readers between applied batches.
pub fn reconcile_handle(
    handle: &IndexHandle,
    config: &ScanConfig,
    sink: &mut dyn FnMut(&Commit),
) -> Result<ReconcileReport> {
    reconcile_subtree_handle(handle, Path::new(""), config, sink)
}

/// Reconcile one subtree of a shared index, widening to a missing/non-directory ancestor
/// when necessary.
pub fn reconcile_subtree_handle(
    handle: &IndexHandle,
    subtree: &Path,
    config: &ScanConfig,
    sink: &mut dyn FnMut(&Commit),
) -> Result<ReconcileReport> {
    reconcile_target(&mut ReconcileTarget::Shared(handle), subtree, config, sink)
}

/// Internal effects of one opened-root multi-path reconciliation.
#[derive(Debug, Default)]
pub(crate) struct ReconcilePathsReport {
    pub(crate) reconciliation: ReconcileReport,
    pub(crate) accepted: Vec<PathBuf>,
    pub(crate) rejected: Vec<crate::RejectedRefreshPath>,
}

/// Reconcile one bounded path set under an opened-root lifecycle controller.
///
/// Classification precedes I/O, overlapping descendants fold into one walk, and all
/// surviving scopes enter `Reconciling` before the first is read. `forbid_expansion`
/// is the conservative resource-stop rule: removals and same-file verification remain
/// legal, while work that could retain another file or discover children is refused.
pub(crate) fn reconcile_paths_handle_controlled(
    handle: &IndexHandle,
    paths: &[PathBuf],
    config: &ScanConfig,
    forbid_expansion: bool,
    control: &dyn ReconcileControl,
    sink: &mut dyn FnMut(&Commit),
) -> Result<ReconcilePathsReport> {
    let mut target = ReconcileTarget::Controlled { handle, control };
    reconcile_paths_target(&mut target, paths, config, forbid_expansion, sink)
}

fn reconcile_paths_target(
    target: &mut ReconcileTarget<'_>,
    paths: &[PathBuf],
    config: &ScanConfig,
    forbid_expansion: bool,
    sink: &mut dyn FnMut(&Commit),
) -> Result<ReconcilePathsReport> {
    config.validate_for_scope(target.scope()?)?;
    let mut report = ReconcilePathsReport::default();
    let mut accepted = BTreeSet::new();

    for requested in paths {
        let reject = |reason| crate::RejectedRefreshPath { path: requested.clone(), reason };
        let Ok(path) = normalize_subtree(requested) else {
            report.rejected.push(reject(crate::RefreshRejection::OutsideRoot));
            continue;
        };
        if config.max_depth.is_some_and(|maximum| path.components().count() > maximum) {
            report.rejected.push(reject(crate::RefreshRejection::BeyondDepth));
            continue;
        }
        // This is lexical admission before the final kind is observed. Treating the
        // boundary as a file preserves the fixed hidden `.gitignore` control exception;
        // the verified walk still applies the real kind and special-object policy.
        if crate::admission::decide_path(&path, EntryKind::File, config.hidden(), false)
            == crate::admission::Disposition::Reject
        {
            report.rejected.push(reject(crate::RefreshRejection::NotAdmitted));
            continue;
        }
        if forbid_expansion && refresh_may_expand(target, &path, &mut report.reconciliation.scan)? {
            report.rejected.push(reject(crate::RefreshRejection::ResourceBudget));
            continue;
        }
        accepted.insert(path);
    }

    report.accepted = accepted.into_iter().collect();
    let mut resolved = Vec::new();
    let mut unsafe_roots = Vec::new();
    for requested_root in covering_roots(report.accepted.clone()) {
        match resolve_subtree_root(target, &requested_root, config) {
            Ok(root) => resolved.push(root),
            Err(Error::SubtreeOutsideScanScope { .. }) => unsafe_roots.push(requested_root),
            Err(error) => return Err(error),
        }
    }
    if !unsafe_roots.is_empty() {
        let mut retained = Vec::with_capacity(report.accepted.len());
        for path in std::mem::take(&mut report.accepted) {
            if unsafe_roots.iter().any(|root| path.starts_with(root)) {
                report.rejected.push(crate::RejectedRefreshPath {
                    path,
                    reason: crate::RefreshRejection::UnsafeAncestry,
                });
            } else {
                retained.push(path);
            }
        }
        report.accepted = retained;
    }
    let walked = covering_roots(resolved);
    if walked.is_empty() {
        return Ok(report);
    }

    let mut opened = Vec::with_capacity(walked.len());
    for subtree in walked {
        let (started_at, commit) = target.begin_reconcile(&subtree)?;
        if let Some(commit) = commit.as_ref() {
            sink(commit);
        }
        opened.push((subtree, started_at));
    }

    // Each subtree closes on its own walk's outcome, so a subtree that could not be read
    // neither marks a verified sibling partial nor withholds the completeness its listing
    // earned.
    let mut failure = None;
    let mut outcomes = Vec::with_capacity(opened.len());
    for (subtree, started_at) in &opened {
        if failure.is_some() {
            outcomes.push((false, false));
            continue;
        }
        match reconcile_target_inner(
            target,
            subtree,
            *started_at,
            config,
            MAX_DEFERRED_RECONCILE_OPS,
            sink,
        ) {
            Ok(mut reconciliation) => {
                outcomes.push((
                    reconciliation.is_complete(),
                    reconciliation.apply.stale == 0 && reconciliation.apply.resource_refused == 0,
                ));
                reconciliation.listed_incomplete = reconciliation.take_recordable_completeness();
                merge_reconcile_report(&mut report.reconciliation, reconciliation);
            }
            Err(error) => {
                outcomes.push((false, false));
                failure = Some(error);
            }
        }
    }

    let listed_incomplete = std::mem::take(&mut report.reconciliation.listed_incomplete);
    let root = target.root_path()?;
    normalize_walk_errors(&root, &mut report.reconciliation.scan.errors);
    let failed_paths = failure_paths(target, &report.reconciliation.scan.errors)?;
    for ((subtree, started_at), (complete, disproves_old)) in opened.into_iter().zip(outcomes) {
        let commit = target.finish_reconcile(
            &subtree,
            started_at,
            complete,
            &listed_incomplete,
            &failed_paths,
            ReconcileErrors {
                errors: &report.reconciliation.scan.errors,
                terminal: failure.as_ref(),
                disproves_old,
            },
        )?;
        if let Some(commit) = commit.commit.as_ref() {
            sink(commit);
        }
        report.reconciliation.retry_required |= commit.retry;
    }

    match failure {
        Some(error) => Err(error),
        None => Ok(report),
    }
}

/// Root-relative paths whose filesystem facts a failed reconciliation could not verify.
///
/// An unscoped error returns an empty set, which makes the closer conservatively mark the
/// whole requested subtree partial. Precise I/O paths let verified siblings remain fresh.
fn failure_paths(target: &ReconcileTarget<'_>, errors: &[Error]) -> Result<Vec<PathBuf>> {
    if errors.is_empty() {
        return Ok(Vec::new());
    }
    let root = target.root_path()?;
    let mut paths = Vec::with_capacity(errors.len());
    for error in errors {
        let Some(path) = crate::Issue::from_error_under(&root, error).path else {
            return Ok(Vec::new());
        };
        if path.is_absolute() {
            return Ok(Vec::new());
        }
        paths.push(path);
    }
    paths.sort();
    paths.dedup();
    Ok(paths)
}

/// Whether verification could increase the retained-file set.
///
/// This deliberately recognizes only cases that prove non-expansion. At a resource
/// boundary, uncertainty is a refusal rather than permission to exceed the bound.
fn refresh_may_expand(
    target: &ReconcileTarget<'_>,
    path: &Path,
    work: &mut ScanReport,
) -> Result<bool> {
    let current = target.expectation(path)?.state;
    let absolute = target.root_path()?.join(path);
    let observed = match observe_path(&absolute) {
        Ok((kind, attrs)) => {
            work.observe(kind, attrs);
            Some(kind)
        }
        Err(error)
            if matches!(
                error.kind(),
                std::io::ErrorKind::NotFound | std::io::ErrorKind::NotADirectory
            ) =>
        {
            None
        }
        Err(_) => return Ok(true),
    };
    Ok(!matches!(
        (current, observed),
        (PathState::Present { kind: EntryKind::File, .. }, Some(EntryKind::File)) | (_, None)
    ))
}

/// Drop every path covered by a shallower member of the same sorted set.
fn covering_roots(mut paths: Vec<PathBuf>) -> Vec<PathBuf> {
    paths.sort();
    paths.dedup();
    if paths.first().is_some_and(|first| first.as_os_str().is_empty()) {
        return vec![PathBuf::new()];
    }
    let mut roots: Vec<PathBuf> = Vec::with_capacity(paths.len());
    for path in paths {
        if roots.last().is_some_and(|kept| path.starts_with(kept)) {
            continue;
        }
        roots.push(path);
    }
    roots
}

fn reconcile_target(
    target: &mut ReconcileTarget<'_>,
    subtree: &Path,
    config: &ScanConfig,
    sink: &mut dyn FnMut(&Commit),
) -> Result<ReconcileReport> {
    config.validate_for_scope(target.scope()?)?;
    if let Some(progress) = &config.progress {
        progress.enter(crate::ProgressPhase::Revalidating);
    }
    let subtree = normalize_subtree(subtree)?;
    if config.max_depth.is_some_and(|maximum| subtree.components().count() > maximum) {
        return Err(Error::SubtreeOutsideScanScope { path: subtree, scope: config.scope() });
    }
    let subtree = if config.population != crate::query::IgnoredEntries::Include
        && !subtree.as_os_str().is_empty()
    {
        // A narrowed tier may have pruned the requested entry or an ancestor. Start
        // from its nearest retained parent so the governing control is read before the
        // directory listing decides whether the boundary itself belongs in the tier.
        // A control-file edit also needs this parent listing to discover siblings that
        // were absent under the previous rule.
        let mut parent = subtree.parent().map_or_else(PathBuf::new, Path::to_path_buf);
        while !parent.as_os_str().is_empty()
            && (target.expectation(&parent)?.state == PathState::Absent
                || !target.control_classification_known(&parent)?)
        {
            parent = parent.parent().map_or_else(PathBuf::new, Path::to_path_buf);
        }
        parent
    } else {
        subtree
    };
    let subtree = resolve_subtree_root(target, &subtree, config)?;
    let (started_at, started) = target.begin_reconcile(&subtree)?;
    if let Some(commit) = started.as_ref() {
        sink(commit);
    }
    match reconcile_target_inner(
        target,
        &subtree,
        started_at,
        config,
        MAX_DEFERRED_RECONCILE_OPS,
        sink,
    ) {
        Ok(mut report) => {
            let root = target.root_path()?;
            normalize_walk_errors(&root, &mut report.scan.errors);
            let listed_incomplete = report.take_recordable_completeness();
            let failed_paths = failure_paths(target, &report.scan.errors)?;
            let finished = target.finish_reconcile(
                &subtree,
                started_at,
                report.is_complete(),
                &listed_incomplete,
                &failed_paths,
                ReconcileErrors {
                    errors: &report.scan.errors,
                    terminal: None,
                    disproves_old: report.apply.stale == 0 && report.apply.resource_refused == 0,
                },
            )?;
            if let Some(commit) = finished.commit.as_ref() {
                sink(commit);
            }
            report.retry_required |= finished.retry;
            Ok(report)
        }
        Err(error) => {
            let finished = target.finish_reconcile(
                &subtree,
                started_at,
                false,
                &[],
                &[],
                ReconcileErrors { errors: &[], terminal: Some(&error), disproves_old: false },
            )?;
            if let Some(commit) = finished.commit.as_ref() {
                sink(commit);
            }
            Err(error)
        }
    }
}

fn reconcile_target_inner(
    target: &mut ReconcileTarget<'_>,
    subtree: &Path,
    started_at: u64,
    config: &ScanConfig,
    max_deferred_ops: usize,
    sink: &mut dyn FnMut(&Commit),
) -> Result<ReconcileReport> {
    let root = target.root_path()?;
    let root_meta = {
        crate::counters::bump(|c| c.stats += 1);
        fs::symlink_metadata(&root)
    }
    .map_err(|error| Error::io(&root, error))?;
    if !root_meta.is_dir() {
        return Err(Error::io(
            &root,
            std::io::Error::new(
                std::io::ErrorKind::NotADirectory,
                "reconciliation root is not a directory",
            ),
        ));
    }
    let root_dev = root_device(&root, &root_meta).map_err(|error| Error::io(&root, error))?;
    let start_depth = subtree.components().count();
    let mut report =
        ReconcileReport { reconcile_epoch: Some(started_at), ..ReconcileReport::default() };
    let mut tally = ProgressTally::new(config.progress.as_ref());
    let mut retry_frontier = None;
    let mut batch: Vec<ObservationOp> = Vec::with_capacity(config.batch_size.max(1));

    if config.max_depth == Some(0) {
        remove_known_children(target, Path::new(""), config, &mut batch, sink, &mut report)?;
        return Ok(report);
    }

    if !subtree.as_os_str().is_empty() {
        let baseline = target.expectation(subtree)?;
        let absolute = root.join(subtree);
        let (kind, attrs) = match observe_path(&absolute) {
            Ok(observed) => observed,
            Err(error) if error.kind() == std::io::ErrorKind::NotFound => {
                batch.push(ObservationOp::if_state(
                    Op::Remove { path: subtree.to_path_buf() },
                    baseline,
                ));
                push_lost_spelling_control(
                    target,
                    &root,
                    config,
                    subtree,
                    baseline,
                    &mut batch,
                    &mut report,
                )?;
                flush_reconcile_batch(target, &mut batch, sink, &mut report)?;
                return Ok(report);
            }
            Err(error) => {
                report.scan.errors.push(Error::io(&absolute, error));
                if baseline.state != PathState::Absent {
                    batch.push(ObservationOp::if_state(
                        Op::Remove { path: subtree.to_path_buf() },
                        baseline,
                    ));
                }
                push_lost_spelling_control(
                    target,
                    &root,
                    config,
                    subtree,
                    baseline,
                    &mut batch,
                    &mut report,
                )?;
                flush_reconcile_batch(target, &mut batch, sink, &mut report)?;
                return Ok(report);
            }
        };
        let disposition =
            crate::admission::decide_path(subtree, kind, config.hidden(), config.exclude_special);
        if disposition != crate::admission::Disposition::Retain {
            if baseline.state != PathState::Absent {
                batch.push(ObservationOp::if_state(
                    Op::Remove { path: subtree.to_path_buf() },
                    baseline,
                ));
            }
            if disposition == crate::admission::Disposition::ControlOnly {
                let read = read_control_op(config, &root, subtree, kind);
                push_read_control(target, subtree, baseline, read, &mut batch, &mut report)?;
            }
            flush_reconcile_batch(target, &mut batch, sink, &mut report)?;
            return Ok(report);
        }
        report.scan.observe(kind, attrs);
        push_reconcile_upsert(target, subtree, kind, attrs, baseline, &mut batch, &mut report);
        // A retained control file at the root of the walk reads its rules here, as the
        // listing walk does for every retained entry it lists: a file does not descend,
        // so nothing below would read them, and the table kept the old source while the
        // pass reported complete and marked the path fresh. In the same batch as the
        // upsert, so both are arbitrated against one baseline.
        let read = read_control_op(config, &root, subtree, kind);
        push_read_control(target, subtree, baseline, read, &mut batch, &mut report)?;
        flush_reconcile_batch(target, &mut batch, sink, &mut report)?;
        if !should_descend(kind, attrs, start_depth.saturating_sub(1), root_dev, config) {
            if kind.is_dir() {
                remove_known_children(target, subtree, config, &mut batch, sink, &mut report)?;
            }
            tally.flush(&report.scan);
            return Ok(report);
        }
    }

    if subtree.as_os_str().is_empty()
        && config.population == crate::query::IgnoredEntries::Include
        && config.reconciliation_worker_threads() > 1
    {
        if let ReconcileTarget::Direct(index) = target {
            match reconcile_direct_parallel(index, &root, root_dev, config, max_deferred_ops, sink)?
            {
                DirectParallelOutcome::Complete(parallel) => return Ok(parallel),
                DirectParallelOutcome::RetrySerial { prefix, remaining } => {
                    report = prefix;
                    report.reconcile_epoch = Some(started_at);
                    retry_frontier = Some(remaining);
                    // The wave workers reported the prefix themselves, and the wave
                    // that overflowed as well: progress counts that wave's reads twice,
                    // once there and once as the serial retry rereads it, while this
                    // report counts each directory once. Work done, not the answer.
                    tally.skip_to(&report.scan);
                }
            }
        }
    }

    let mut queue: VecDeque<(PathBuf, usize)> = retry_frontier
        .unwrap_or_else(|| VecDeque::from(vec![(subtree.to_path_buf(), start_depth)]));
    let mut controls = if config.population == crate::query::IgnoredEntries::Include {
        None
    } else {
        Some(target.control_table()?)
    };
    let mut unreadable_controls = std::collections::BTreeSet::new();
    #[cfg(target_os = "macos")]
    let mut bulk_reader = (config.worker_threads() > 1).then(macos_bulk::Reader::new);
    let mut readers = Readers::new();
    let policy = ListingPolicy::every_child_stated(config.one_filesystem);
    while let Some((rel_dir, depth)) = take_next(&mut queue, config.order) {
        let errors_before = report.scan.errors.len();
        let abs_dir = root.join(&rel_dir);
        let control_path = rel_dir.join(crate::control::CONTROL_FILE_NAME);
        let mut had_control = target.has_control(&control_path)?;
        let mut listed_control = ListedControl::default();
        if let Some(table) = controls.as_mut() {
            let lookup = read_directory_control(config, &root, &control_path);
            listed_control.lookup(&lookup);
            match lookup {
                Ok(Some(op)) => {
                    let baseline = target.expectation(&control_path)?;
                    apply_discovery_control(table, &op)?;
                    batch.push(ObservationOp::if_state(op, baseline));
                }
                Ok(None) => {
                    table.remove(&control_path)?;
                    if had_control {
                        let baseline = target.expectation(&control_path)?;
                        batch.push(ObservationOp::if_state(
                            Op::ControlRemove { path: control_path.clone() },
                            baseline,
                        ));
                        had_control = false;
                    }
                }
                Err(error) => {
                    unreadable_controls.insert(rel_dir.clone());
                    report.scan.errors.push(error);
                }
            }
            flush_reconcile_batch(target, &mut batch, sink, &mut report)?;
            if report.apply.stale > 0 || report.apply.resource_refused > 0 {
                // This directory's population decisions use the source just read.
                // If its conditional control commit lost ownership, walking siblings
                // against the local copy could prune under rules the index rejected.
                report.retry_required = true;
                return Ok(report);
            }
        }
        let (mut known, records_completeness) = target.listing_baseline(&rel_dir)?;
        let mut listing_complete = true;
        let process_entry = |name: OsString,
                             kind: EntryKind,
                             attrs: Attrs,
                             baseline: PathExpectation,
                             listed_control: &mut ListedControl,
                             target: &mut ReconcileTarget<'_>,
                             queue: &mut VecDeque<(PathBuf, usize)>,
                             batch: &mut Vec<ObservationOp>,
                             sink: &mut dyn FnMut(&Commit),
                             report: &mut ReconcileReport|
         -> Result<()> {
            let rel_path = rel_dir.join(&name);
            listed_control.listed(&name);
            let disposition =
                crate::admission::decide(&name, kind, config.hidden(), config.exclude_special);
            if population_prunes(
                config.population,
                &rel_path,
                kind,
                disposition,
                controls.as_ref(),
                &unreadable_controls,
            ) {
                if baseline.state != PathState::Absent {
                    batch.push(ObservationOp::if_state(Op::Remove { path: rel_path }, baseline));
                }
                return Ok(());
            }
            if disposition != crate::admission::Disposition::Retain {
                if baseline.state != PathState::Absent {
                    batch.push(ObservationOp::if_state(
                        Op::Remove { path: rel_path.clone() },
                        baseline,
                    ));
                }
                if disposition == crate::admission::Disposition::ControlOnly {
                    let read = read_listed_control_op(config, &root, &rel_path, kind);
                    listed_control.read(&name, &read);
                    push_read_control(target, &rel_path, baseline, read, batch, report)?;
                }
                if batch.len() >= config.batch_size.max(1) {
                    flush_reconcile_batch(target, batch, sink, report)?;
                }
                return Ok(());
            }
            report.scan.observe(kind, attrs);
            push_reconcile_upsert(target, &rel_path, kind, attrs, baseline, batch, report);
            if batch.len() >= config.batch_size.max(1) {
                flush_reconcile_batch(target, batch, sink, report)?;
            }
            let read = read_listed_control_op(config, &root, &rel_path, kind);
            listed_control.read(&name, &read);
            push_read_control(target, &rel_path, baseline, read, batch, report)?;
            if batch.len() >= config.batch_size.max(1) {
                flush_reconcile_batch(target, batch, sink, report)?;
            }

            if should_descend(kind, attrs, depth, root_dev, config) {
                queue.push_back((rel_path, depth + 1));
            } else if kind.is_dir() {
                remove_known_children(target, &rel_path, config, batch, sink, report)?;
            }
            Ok(())
        };

        #[cfg(target_os = "macos")]
        let used_bulk = if walk_hook_covers(&abs_dir) {
            false
        } else if let Some(entries) = bulk_reader.as_mut().and_then(|reader| reader.read(&abs_dir))
        {
            report.scan.dirs_read += 1;
            for entry in entries {
                let baseline = match known.remove(&entry.name) {
                    Some(baseline) => baseline,
                    None => target.expectation(&rel_dir.join(&entry.name))?,
                };
                process_entry(
                    entry.name,
                    entry.kind,
                    entry.attrs,
                    baseline,
                    &mut listed_control,
                    target,
                    &mut queue,
                    &mut batch,
                    sink,
                    &mut report,
                )?;
            }
            true
        } else {
            false
        };
        #[cfg(not(target_os = "macos"))]
        let used_bulk = false;

        if !used_bulk {
            let listing = match list_directory(&mut readers, &abs_dir, policy, None) {
                Ok(listing) => listing,
                Err(error) => {
                    report.scan.errors.push(Error::io(&abs_dir, error));
                    remove_known_children(target, &rel_dir, config, &mut batch, sink, &mut report)?;
                    if had_control {
                        let baseline = target.expectation(&control_path)?;
                        batch.push(ObservationOp::if_state(
                            Op::ControlRemove { path: control_path },
                            baseline,
                        ));
                        flush_reconcile_batch(target, &mut batch, sink, &mut report)?;
                    }
                    continue;
                }
            };
            report.scan.dirs_read += 1;
            let listing = reconcile_listing(listing, &abs_dir);
            for item in listing {
                let (name, observed) = match item {
                    Listed::Child { name, observed } => (name.into_owned(), observed),
                    Listed::Failed(error) => {
                        listing_complete = false;
                        report.scan.errors.push(Error::io(&abs_dir, error));
                        continue;
                    }
                };
                // Seeing the name proves it is not absent even if the following
                // metadata lookup fails. Remove it from the missing set before that
                // fallible lookup so an operational error cannot turn an existing
                // entry into a deletion.
                let baseline = match known.remove(&name) {
                    Some(baseline) => baseline,
                    None => target.expectation(&rel_dir.join(&name))?,
                };
                let (kind, attrs) = match observed {
                    Ok(Some(observed)) => observed,
                    Ok(None) => {
                        let entry_held = baseline.state != PathState::Absent;
                        for removal in
                            vanished_child_removals(&rel_dir, &name, entry_held, &mut had_control)
                                .into_iter()
                                .flatten()
                        {
                            batch.push(ObservationOp::if_state(removal, baseline));
                        }
                        if batch.len() >= config.batch_size.max(1) {
                            flush_reconcile_batch(target, &mut batch, sink, &mut report)?;
                        }
                        continue;
                    }
                    Err(error) => {
                        listed_control.listed(&name);
                        report.scan.errors.push(Error::io(abs_dir.join(&name), error));
                        if baseline.state != PathState::Absent {
                            batch.push(ObservationOp::if_state(
                                Op::Remove { path: rel_dir.join(&name) },
                                baseline,
                            ));
                        }
                        if name == crate::control::CONTROL_FILE_NAME && had_control {
                            batch.push(ObservationOp::if_state(
                                Op::ControlRemove { path: rel_dir.join(&name) },
                                baseline,
                            ));
                            had_control = false;
                        }
                        if batch.len() >= config.batch_size.max(1) {
                            flush_reconcile_batch(target, &mut batch, sink, &mut report)?;
                        }
                        continue;
                    }
                };
                process_entry(
                    name,
                    kind,
                    attrs,
                    baseline,
                    &mut listed_control,
                    target,
                    &mut queue,
                    &mut batch,
                    sink,
                    &mut report,
                )?;
            }
        }

        for (name, baseline) in known {
            let restatement = listed_control.restatement_after_removing(&name);
            batch.push(ObservationOp::if_state(Op::Remove { path: rel_dir.join(name) }, baseline));
            // Before the batch can be flushed, so the rules the removal drops are back in the
            // same commit.
            if let Some(control) = restatement {
                let guard = target.expectation(&control_path)?;
                batch.push(ObservationOp::if_state(control, guard));
            }
            if batch.len() >= config.batch_size.max(1) {
                flush_reconcile_batch(target, &mut batch, sink, &mut report)?;
            }
        }
        if had_control && !listed_control.seen {
            let baseline = target.expectation(&control_path)?;
            batch.push(ObservationOp::if_state(Op::ControlRemove { path: control_path }, baseline));
        }
        if listing_complete {
            // Only a directory with no error inside its own processing vouches for its
            // child set; an error under a sibling or a descendant is that directory's
            // to answer for, as discovery decides completeness per directory.
            if records_completeness && report.scan.errors.len() == errors_before {
                report.listed_incomplete.push(rel_dir);
            }
        }
        // Per directory, as in `revalidate`: an unchanged tree hands the sink nothing.
        tally.flush(&report.scan);
    }

    flush_reconcile_batch(target, &mut batch, sink, &mut report)?;
    report.scan.errors.sort_by_cached_key(ToString::to_string);
    tally.flush(&report.scan);
    Ok(report)
}

#[derive(Debug, Default)]
struct DeferredReconcile {
    scan: ScanReport,
    unchanged: u64,
    operations: Vec<Op>,
    /// Each removal of a stale `.gitignore` entry whose rules a case variant now holds,
    /// followed by the restatement of those rules (`ListedControl`). Applied after the
    /// wave's sorted operations, in one batch, because the sort moves every removal after
    /// every other operation and a batch boundary could fall between the two.
    restated: Vec<Op>,
    discovered: Vec<(PathBuf, usize, RegionId)>,
    listed_incomplete: Vec<PathBuf>,
}

enum DirectParallelOutcome {
    Complete(ReconcileReport),
    RetrySerial { prefix: ReconcileReport, remaining: VecDeque<(PathBuf, usize)> },
}

/// Reconcile an exclusive full tree in bounded immutable-baseline waves.
///
/// No index write occurs while a wave's workers hold shared baseline references. That
/// lets each worker discard exact no-ops where they are observed instead of funnelling
/// every entry through one consumer. Effective changes still enter through ordinary
/// observations between waves, preserving both the index's sole mutation contract and
/// progressive delta delivery.
///
/// Unlike every other reconciliation path, the operations a wave defers are
/// unconditional: they carry no [`ObservationOp::if_state`] guard. Three properties
/// have to hold together for that to be safe, and a change to any one of them puts the
/// guards back. The target is an exclusive `&mut Index`, so no other producer can
/// commit between a worker's read and the wave's write. Nothing is applied while
/// workers run, so no baseline a worker compared against can go stale beneath it. And
/// a directory is only reconciled in a wave after the wave that discovered it has
/// committed, so a parent is never absent when its children arrive.
fn reconcile_direct_parallel(
    index: &mut Index,
    root: &Path,
    root_dev: u64,
    config: &ScanConfig,
    max_deferred_ops: usize,
    sink: &mut dyn FnMut(&Commit),
) -> Result<DirectParallelOutcome> {
    let mut frontier = DirectoryQueueState::seeded(
        (PathBuf::new(), 0),
        config.order,
        None,
        1,
        1,
        WorkerPolicyExperiment::ShippedOneShot,
    );
    let mut report = ReconcileReport::default();
    while !frontier.is_empty(config.order) {
        let mut wave = Vec::with_capacity(RECONCILE_WAVE_DIRECTORIES);
        while wave.len() < RECONCILE_WAVE_DIRECTORIES && !frontier.is_empty(config.order) {
            let remaining = RECONCILE_WAVE_DIRECTORIES - wave.len();
            frontier.take(remaining.min(DIR_CLAIM), config.order, &mut wave);
        }

        let next = std::sync::atomic::AtomicUsize::new(0);
        let deferred_count = std::sync::atomic::AtomicUsize::new(0);
        let overflowed = std::sync::atomic::AtomicBool::new(false);
        let workers = config.reconciliation_worker_threads().min(wave.len());
        let baseline: &Index = index;
        let results: Vec<DeferredReconcile> = std::thread::scope(|scope| {
            let handles: Vec<_> = (0..workers)
                .map(|_| {
                    scope.spawn(|| {
                        reconcile_wave_worker(
                            baseline,
                            root,
                            root_dev,
                            config,
                            &wave,
                            &next,
                            &deferred_count,
                            &overflowed,
                            max_deferred_ops,
                        )
                    })
                })
                .collect();

            handles
                .into_iter()
                .map(|handle| {
                    if let Ok(worker) = handle.join() {
                        return worker;
                    }
                    let mut worker = DeferredReconcile::default();
                    worker.scan.errors.push(Error::io(
                        root,
                        std::io::Error::other("a reconciliation worker thread panicked"),
                    ));
                    worker
                })
                .collect()
        });

        // Nothing from an overflowing wave was applied, so the ordinary incremental
        // reconciler can resume at that wave. Completed waves and their statistics are
        // retained exactly once; restarting from the root would count their unchanged
        // entries again and misreport the logical reconciliation pass.
        if overflowed.load(std::sync::atomic::Ordering::Relaxed) {
            let mut remaining: VecDeque<_> =
                wave.into_iter().map(|(path, depth, _region)| (path, depth)).collect();
            let mut deferred = Vec::with_capacity(DIR_CLAIM);
            while !frontier.is_empty(config.order) {
                frontier.take(DIR_CLAIM, config.order, &mut deferred);
                remaining.extend(deferred.drain(..).map(|(path, depth, _region)| (path, depth)));
            }
            return Ok(DirectParallelOutcome::RetrySerial { prefix: report, remaining });
        }

        let operation_count = deferred_count.load(std::sync::atomic::Ordering::Relaxed);
        let mut operations = Vec::with_capacity(operation_count);
        let mut restated = Vec::new();
        for worker in results {
            report.listed_incomplete.extend(worker.listed_incomplete);
            report.scan.absorb(worker.scan);
            report.apply.unchanged += worker.unchanged;
            report.observations = report.observations.saturating_add(worker.unchanged);
            operations.extend(worker.operations);
            restated.extend(worker.restated);
            for directory in worker.discovered {
                frontier.push(directory, config.order);
            }
        }
        apply_deferred_reconcile(
            index,
            &mut operations,
            restated,
            config,
            sink,
            &mut report.apply,
            &mut report.observations,
        )?;
    }
    report.scan.errors.sort_by_cached_key(ToString::to_string);
    Ok(DirectParallelOutcome::Complete(report))
}

fn apply_deferred_reconcile(
    index: &mut Index,
    operations: &mut Vec<Op>,
    restated: Vec<Op>,
    config: &ScanConfig,
    sink: &mut dyn FnMut(&Commit),
    stats: &mut ApplyStats,
    observations: &mut u64,
) -> Result<()> {
    // Parent upserts establish real directory attributes before children arrive.
    // Removals run deepest first so a parent removal never precedes an independently
    // observed descendant operation. Deterministic causal order also makes emitted
    // commits stable for callers.
    operations.sort_by(|left, right| {
        let left_remove = matches!(left, Op::Remove { .. });
        let right_remove = matches!(right, Op::Remove { .. });
        left_remove.cmp(&right_remove).then_with(|| {
            let left_depth = left.path().components().count();
            let right_depth = right.path().components().count();
            if left_remove {
                right_depth.cmp(&left_depth).then_with(|| left.path().cmp(right.path()))
            } else {
                left_depth.cmp(&right_depth).then_with(|| left.path().cmp(right.path()))
            }
        })
    });

    let batch_limit = config.batch_size.max(1);
    let mut batch = Vec::with_capacity(batch_limit.min(operations.len()));
    for operation in operations.drain(..) {
        batch.push(operation);
        if batch.len() >= batch_limit {
            flush_direct_reconcile_batch(index, &mut batch, sink, stats, observations)?;
        }
    }
    flush_direct_reconcile_batch(index, &mut batch, sink, stats, observations)?;
    // Whole, so no commit shows a directory without the rules a restatement puts back.
    let mut restated = restated;
    flush_direct_reconcile_batch(index, &mut restated, sink, stats, observations)
}

#[allow(clippy::too_many_arguments)]
fn reconcile_wave_worker(
    index: &Index,
    root: &Path,
    root_dev: u64,
    config: &ScanConfig,
    wave: &[(PathBuf, usize, RegionId)],
    next: &std::sync::atomic::AtomicUsize,
    deferred_count: &std::sync::atomic::AtomicUsize,
    overflowed: &std::sync::atomic::AtomicBool,
    max_deferred_ops: usize,
) -> DeferredReconcile {
    let _counter_guard = crate::counters::thread_flush_guard();
    let mut result = DeferredReconcile::default();
    let mut tally = ProgressTally::new(config.progress.as_ref());
    #[cfg(target_os = "macos")]
    let mut bulk_reader = macos_bulk::Reader::new();
    let mut readers = Readers::new();
    let policy = ListingPolicy::every_child_stated(config.one_filesystem);

    loop {
        let start = next.fetch_add(DIR_CLAIM, std::sync::atomic::Ordering::Relaxed);
        if start >= wave.len() {
            break;
        }
        let end = start.saturating_add(DIR_CLAIM).min(wave.len());
        for (rel_dir, depth, region) in &wave[start..end] {
            let errors_before = result.scan.errors.len();
            let mut known = collect_child_expectations(index, rel_dir);
            let abs_dir = root.join(rel_dir);
            let control_path = rel_dir.join(crate::control::CONTROL_FILE_NAME);
            let mut had_control = index.control_table().contains(&control_path);
            let mut listed_control = ListedControl::default();
            let mut control_errors = Vec::new();
            let mut vanished = Vec::new();
            let mut unverified = Vec::new();
            let mut control_read_failed = false;
            let mut listing_open_failed = false;

            {
                let mut process_entry =
                    |name: OsString,
                     kind: EntryKind,
                     attrs: Attrs,
                     baseline: PathExpectation,
                     listed_control: &mut ListedControl| {
                        let rel_path = rel_dir.join(&name);
                        listed_control.listed(&name);
                        let disposition = crate::admission::decide(
                            &name,
                            kind,
                            config.hidden(),
                            config.exclude_special,
                        );
                        if disposition == crate::admission::Disposition::Reject {
                            if baseline.state != PathState::Absent {
                                defer_reconcile_op(
                                    Op::Remove { path: rel_path },
                                    &mut result.operations,
                                    deferred_count,
                                    overflowed,
                                    max_deferred_ops,
                                );
                            }
                            return;
                        }
                        if disposition == crate::admission::Disposition::ControlOnly {
                            let read = read_control_op(config, root, &rel_path, kind);
                            listed_control.read(&name, &read);
                            match read {
                                Ok(Some(Op::ControlUpsert { path, source })) => {
                                    if !index.control_table().source_is(&path, &source) {
                                        defer_reconcile_op(
                                            Op::ControlUpsert { path, source },
                                            &mut result.operations,
                                            deferred_count,
                                            overflowed,
                                            max_deferred_ops,
                                        );
                                    }
                                }
                                Ok(Some(Op::ControlRemove { path })) => {
                                    if index.control_table().contains(&path) {
                                        defer_reconcile_op(
                                            Op::ControlRemove { path },
                                            &mut result.operations,
                                            deferred_count,
                                            overflowed,
                                            max_deferred_ops,
                                        );
                                    }
                                }
                                Ok(Some(_) | None) => {}
                                Err(error) => {
                                    control_errors.push(error);
                                    control_read_failed = true;
                                }
                            }
                            return;
                        }
                        result.scan.entries += 1;
                        if kind == EntryKind::File {
                            result.scan.files_walked += 1;
                            result.scan.bytes_walked =
                                result.scan.bytes_walked.saturating_add(attrs.size);
                            result.scan.allocated_walked =
                                result.scan.allocated_walked.saturating_add(attrs.allocated);
                        }
                        if baseline.state == (PathState::Present { kind, attrs }) {
                            result.unchanged += 1;
                        } else {
                            defer_reconcile_op(
                                Op::Upsert { path: rel_path.clone(), kind, attrs },
                                &mut result.operations,
                                deferred_count,
                                overflowed,
                                max_deferred_ops,
                            );
                        }
                        let read = read_control_op(config, root, &rel_path, kind);
                        listed_control.read(&name, &read);
                        match read {
                            Ok(Some(Op::ControlUpsert { path, source })) => {
                                if !index.control_table().source_is(&path, &source) {
                                    defer_reconcile_op(
                                        Op::ControlUpsert { path, source },
                                        &mut result.operations,
                                        deferred_count,
                                        overflowed,
                                        max_deferred_ops,
                                    );
                                }
                            }
                            // The exact name's upsert as a non-file drops its rules by
                            // itself; a case variant's does not, so its lookup's removal
                            // is sent.
                            Ok(Some(Op::ControlRemove { path }))
                                if crate::control::control_spelling(&name)
                                    == Some(crate::control::ControlSpelling::Variant)
                                    && index.control_table().contains(&path) =>
                            {
                                defer_reconcile_op(
                                    Op::ControlRemove { path },
                                    &mut result.operations,
                                    deferred_count,
                                    overflowed,
                                    max_deferred_ops,
                                );
                            }
                            Ok(Some(_) | None) => {}
                            Err(error) => {
                                control_errors.push(error);
                                // The failed read was of the directory's control, through
                                // the entry for the exact name and a lookup for a variant.
                                control_read_failed = true;
                            }
                        }

                        if should_descend(kind, attrs, *depth, root_dev, config) {
                            let child_region =
                                if *depth == 0 { RegionId::UNASSIGNED } else { *region };
                            result.discovered.push((rel_path, depth + 1, child_region));
                        } else if kind.is_dir() {
                            for name in collect_child_expectations(index, &rel_path).into_keys() {
                                defer_reconcile_op(
                                    Op::Remove { path: rel_path.join(name) },
                                    &mut result.operations,
                                    deferred_count,
                                    overflowed,
                                    max_deferred_ops,
                                );
                            }
                        }
                    };

                #[cfg(target_os = "macos")]
                let used_bulk = if let Some(entries) =
                    (!walk_hook_covers(&abs_dir)).then(|| bulk_reader.read(&abs_dir)).flatten()
                {
                    result.scan.dirs_read += 1;
                    for entry in entries {
                        let baseline = known
                            .remove(&entry.name)
                            .unwrap_or_else(|| index.expectation(&rel_dir.join(&entry.name)));
                        process_entry(
                            entry.name,
                            entry.kind,
                            entry.attrs,
                            baseline,
                            &mut listed_control,
                        );
                    }
                    true
                } else {
                    false
                };
                #[cfg(not(target_os = "macos"))]
                let used_bulk = false;

                if !used_bulk {
                    let listing = match list_directory(&mut readers, &abs_dir, policy, None) {
                        Ok(listing) => Some(listing),
                        Err(error) => {
                            result.scan.errors.push(Error::io(&abs_dir, error));
                            listing_open_failed = true;
                            None
                        }
                    };
                    if let Some(listing) = listing {
                        result.scan.dirs_read += 1;
                        let listing = reconcile_listing(listing, &abs_dir);
                        for item in listing {
                            let (name, observed) = match item {
                                Listed::Child { name, observed } => (name.into_owned(), observed),
                                Listed::Failed(error) => {
                                    result.scan.errors.push(Error::io(&abs_dir, error));
                                    continue;
                                }
                            };
                            // Match the serial path: an entry whose name was enumerated is
                            // not missing merely because its metadata could not be read.
                            let baseline = known
                                .remove(&name)
                                .unwrap_or_else(|| index.expectation(&rel_dir.join(&name)));
                            let (kind, attrs) = match observed {
                                Ok(Some(observed)) => observed,
                                Ok(None) => {
                                    // Removed once this directory's listing is done.
                                    vanished.push((name, baseline.state != PathState::Absent));
                                    continue;
                                }
                                Err(error) => {
                                    listed_control.listed(&name);
                                    result.scan.errors.push(Error::io(abs_dir.join(&name), error));
                                    unverified.push((name, baseline.state != PathState::Absent));
                                    continue;
                                }
                            };
                            process_entry(name, kind, attrs, baseline, &mut listed_control);
                        }
                    }
                }
            }
            control_read_failed |= listing_open_failed && had_control;
            result.scan.errors.append(&mut control_errors);
            if index.directory_complete(rel_dir) != Some(true)
                && result.scan.errors.len() == errors_before
            {
                result.listed_incomplete.push(rel_dir.clone());
            }
            for (name, entry_held) in vanished {
                for removal in vanished_child_removals(rel_dir, &name, entry_held, &mut had_control)
                    .into_iter()
                    .flatten()
                {
                    defer_reconcile_op(
                        removal,
                        &mut result.operations,
                        deferred_count,
                        overflowed,
                        max_deferred_ops,
                    );
                }
            }
            for (name, entry_held) in unverified {
                if entry_held {
                    defer_reconcile_op(
                        Op::Remove { path: rel_dir.join(&name) },
                        &mut result.operations,
                        deferred_count,
                        overflowed,
                        max_deferred_ops,
                    );
                }
                if name == crate::control::CONTROL_FILE_NAME {
                    control_read_failed = true;
                }
            }
            for (name, _) in known {
                let restatement = listed_control.restatement_after_removing(&name);
                let removal = Op::Remove { path: rel_dir.join(name) };
                match restatement {
                    Some(control) => {
                        for operation in [removal, control] {
                            defer_reconcile_op(
                                operation,
                                &mut result.restated,
                                deferred_count,
                                overflowed,
                                max_deferred_ops,
                            );
                        }
                    }
                    None => defer_reconcile_op(
                        removal,
                        &mut result.operations,
                        deferred_count,
                        overflowed,
                        max_deferred_ops,
                    ),
                }
            }
            if had_control && (!listed_control.seen || control_read_failed) {
                defer_reconcile_op(
                    Op::ControlRemove { path: control_path },
                    &mut result.operations,
                    deferred_count,
                    overflowed,
                    max_deferred_ops,
                );
            }
        }
        // Once per claimed chunk, as the cold walker reports, so a long wave on a slow
        // filesystem moves the counters while it runs rather than when it lands.
        tally.flush(&result.scan);
    }
    result
}

/// What one reconciliation listing has shown about its directory's control file.
///
/// A listing names the control by the spelling the directory stores. The exact name
/// `.gitignore` is the control wherever it is listed, so listing it proves a control
/// present even when its read then fails. A case variant is the control only where a
/// lookup of `.gitignore` resolves to it (`read_control_op`), so what listing one proves is
/// what that lookup returned: a control when it found one, nothing when it missed, and
/// nothing when the variant's own metadata could not be read and no lookup was made.
/// That last case leaves the rules unknown, which the walk error records
/// (`crate::control::unreadable_control`); the sweep removes rules nothing showed, so the
/// table ends as a cold walk's does, having read nothing there either.
#[derive(Default)]
struct ListedControl {
    /// The listing showed the directory's control, or failed to read it, so the closing
    /// sweep must not remove it.
    seen: bool,
    /// The listing showed a case variant of the control name, such as `.GITIGNORE`.
    variant_listed: bool,
    /// The last control a lookup of the canonical path found during this listing: a
    /// listed case variant's, or a narrowed walk's before the listing.
    ///
    /// A case-only rename on a case-insensitive volume (`.gitignore` to `.GITIGNORE`) leaves
    /// the old spelling's entry in the index, and the closing sweep removes it. Removing
    /// the entry at the canonical path drops the rules it governs
    /// (`Index::projected_controls`), yet those rules are now the variant's, so when this
    /// listing showed the variant the sweep restates this observation right after that
    /// removal ([`Self::restatement_after_removing`]).
    looked_up: Option<Op>,
}

impl ListedControl {
    /// Note that `name` was listed, before its metadata or its control is read.
    fn listed(&mut self, name: &OsStr) {
        match crate::control::control_spelling(name) {
            Some(crate::control::ControlSpelling::Exact) => self.seen = true,
            Some(crate::control::ControlSpelling::Variant) => self.variant_listed = true,
            None => {}
        }
    }

    /// Note what reading the control through the listed `name` returned. Only a case
    /// variant's read is a lookup; the exact name already counted when it was listed.
    fn read(&mut self, name: &OsStr, read: &Result<Option<Op>>) {
        if crate::control::control_spelling(name) == Some(crate::control::ControlSpelling::Variant)
        {
            self.lookup(read);
        }
    }

    /// Note what a lookup of the directory's canonical control path returned.
    fn lookup(&mut self, lookup: &Result<Option<Op>>) {
        match lookup {
            Ok(Some(observed)) => {
                self.seen = true;
                self.looked_up = Some(observed.clone());
            }
            Ok(None) => {}
            Err(_) => self.seen = true,
        }
    }

    /// The observation to push right after the sweep removes the entry `name`: what the
    /// lookup found, when `name` is the stale exact spelling and this listing showed a case
    /// variant in its place.
    ///
    /// Only a listed variant can hold rules the removal would drop. Without one, the exact
    /// file is simply gone: a narrowed walk's lookup found it before the listing and it was
    /// deleted in between, and the removal leaves the table as bare as the directory, as a
    /// cold walk would. Restating that lookup would keep rules for a file that no longer
    /// exists until the next pass. That window remains only where a case-sensitive
    /// directory stores a variant beside the exact file deleted in it, a variant that never
    /// held the rules there.
    fn restatement_after_removing(&self, name: &OsStr) -> Option<Op> {
        if name == crate::control::CONTROL_FILE_NAME && self.variant_listed {
            self.looked_up.clone()
        } else {
            None
        }
    }
}

/// The expectation that guards the control observation a listed entry at `path` produced,
/// given the entry's own, `entry`.
///
/// A guard is the expectation of the path its operation names. That is the entry's path
/// for the exact name, but a case variant's observation names the canonical control path,
/// so only a variant pays for `expect` to read that path's expectation. Asked only once an
/// observation exists, so an ordinary entry never pays for it.
fn control_guard<T>(path: &Path, entry: T, expect: impl FnOnce(&Path) -> T) -> T {
    match crate::control::path_control_spelling(path) {
        Some(crate::control::ControlSpelling::Variant) => {
            expect(&crate::control::sibling_control_path(path))
        }
        _ => entry,
    }
}

/// Push what reading the control through the entry at `path` returned: the observation,
/// guarded by the expectation of the path it names ([`control_guard`]), or, when the read
/// failed, the removal of rules the table holds there, since nothing verifies them now.
fn push_read_control(
    target: &ReconcileTarget<'_>,
    path: &Path,
    baseline: PathExpectation,
    read: Result<Option<Op>>,
    batch: &mut Vec<ObservationOp>,
    report: &mut ReconcileReport,
) -> Result<()> {
    match read {
        Ok(Some(control)) => {
            let guard = control_guard(path, Ok(baseline), |named| target.expectation(named))?;
            batch.push(ObservationOp::if_state(control, guard));
        }
        Ok(None) => {}
        Err(error) => {
            // The failed read was of the entry itself for the exact name, and of the
            // canonical path a case variant looked up: the canonical path either way.
            let control_path = crate::control::sibling_control_path(path);
            if target.has_control(&control_path)? {
                let guard = control_guard(path, Ok(baseline), |named| target.expectation(named))?;
                batch
                    .push(ObservationOp::if_state(Op::ControlRemove { path: control_path }, guard));
            }
            report.scan.errors.push(error);
        }
    }
    Ok(())
}

/// Push what a reconcile root at `path` that vanished, or could not be observed, leaves of
/// its directory's control, when `path` spells the control name.
///
/// The exact name was the lookup's target, so its loss removes the rules the table holds,
/// as it always has. A case variant's loss says nothing by itself: another spelling may
/// still resolve, or the variant was never the control, so the canonical path is looked
/// up and its answer stands, a miss removing the rules.
fn push_lost_spelling_control(
    target: &ReconcileTarget<'_>,
    root: &Path,
    config: &ScanConfig,
    path: &Path,
    baseline: PathExpectation,
    batch: &mut Vec<ObservationOp>,
    report: &mut ReconcileReport,
) -> Result<()> {
    match crate::control::path_control_spelling(path) {
        Some(crate::control::ControlSpelling::Exact) => {
            if target.has_control(path)? {
                batch.push(ObservationOp::if_state(
                    Op::ControlRemove { path: path.to_path_buf() },
                    baseline,
                ));
            }
        }
        Some(crate::control::ControlSpelling::Variant) => {
            let control_path = crate::control::sibling_control_path(path);
            let read = read_directory_control_or_removal(config, root, &control_path);
            push_read_control(target, path, baseline, read, batch, report)?;
        }
        None => {}
    }
    Ok(())
}

/// What a listed child gone at its stat removes: its entry, if the baseline holds one, and
/// its rules, if it is the directory's control file and the table holds them.
///
/// The stat's `NotFound` is positive evidence that both are gone, so neither waits for a
/// complete listing; only a name the listing never returned has to, because it may merely
/// be unread. Removing a retained control file's entry drops its rules too, but a
/// hidden-pruned one has no entry, and without its own removal one unreadable sibling would
/// leave its rules applied with no file behind them. Clears `had_control` once the rules
/// are removed, so the listing's closing removals do not repeat it.
fn vanished_child_removals(
    dir: &Path,
    name: &OsStr,
    entry_held: bool,
    had_control: &mut bool,
) -> [Option<Op>; 2] {
    let path = dir.join(name);
    let rules = (*had_control && name == crate::control::CONTROL_FILE_NAME).then(|| {
        *had_control = false;
        Op::ControlRemove { path: path.clone() }
    });
    [entry_held.then_some(Op::Remove { path }), rules]
}

fn defer_reconcile_op(
    operation: Op,
    operations: &mut Vec<Op>,
    deferred_count: &std::sync::atomic::AtomicUsize,
    overflowed: &std::sync::atomic::AtomicBool,
    max_deferred_ops: usize,
) {
    let position = deferred_count.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
    if position < max_deferred_ops {
        operations.push(operation);
    } else {
        overflowed.store(true, std::sync::atomic::Ordering::Relaxed);
    }
}

fn flush_direct_reconcile_batch(
    index: &mut Index,
    batch: &mut Vec<Op>,
    sink: &mut dyn FnMut(&Commit),
    stats: &mut ApplyStats,
    observations: &mut u64,
) -> Result<()> {
    if batch.is_empty() {
        return Ok(());
    }
    *observations = observations.saturating_add(u64::try_from(batch.len()).unwrap_or(u64::MAX));
    let outcome = index.apply(&Observation::new(std::mem::take(batch)))?;
    merge_apply_stats(stats, outcome.stats);
    if let Some(commit) = outcome.commit.as_ref() {
        sink(commit);
    }
    Ok(())
}

/// Drain and reconcile every pending invalidation, collapsing nested requests.
///
/// An invalidation whose reconciliation comes back incomplete -- a subtree that could not
/// be read, or a conditional commit that lost a race -- is queued again, so the next call
/// retries it. That suits a caller that drains when it chooses. A caller that drains after
/// every event would re-walk an unreadable subtree each time, and belongs on
/// [`reconcile_pending_handle`], which settles it instead.
pub fn reconcile_pending(
    index: &mut Index,
    config: &ScanConfig,
    sink: &mut dyn FnMut(&Commit),
) -> Result<ReconcileReport> {
    let mut target = ReconcileTarget::Direct(index);
    reconcile_pending_target(&mut target, config, sink)
}

/// Drain and reconcile invalidations on a shared index.
///
/// Unlike [`reconcile_pending`], a subtree that could not be read is not queued again.
/// `Watcher::apply_next` drains after every event, and a retry there re-walks the same
/// unreadable subtree on each unrelated one, for the life of the watch. The error is a
/// settled boundary instead: the subtree stays [`crate::Freshness::Partial`], the returned
/// report names the error once, and the watcher retains its cause as an issue. Only a lost
/// race -- a stale conditional commit -- is queued for the next call. Invalidate the subtree
/// again to retry it deliberately.
pub fn reconcile_pending_handle(
    handle: &IndexHandle,
    config: &ScanConfig,
    sink: &mut dyn FnMut(&Commit),
) -> Result<ReconcileReport> {
    let mut target = ReconcileTarget::Shared(handle);
    reconcile_pending_target(&mut target, config, sink)
}

/// Drain and reconcile invalidations under an opened-root lifecycle and resource bound.
#[cfg(feature = "watch")]
pub(crate) fn reconcile_pending_handle_controlled(
    handle: &IndexHandle,
    config: &ScanConfig,
    control: &dyn ReconcileControl,
    sink: &mut dyn FnMut(&Commit),
) -> Result<ReconcileReport> {
    let mut target = ReconcileTarget::Controlled { handle, control };
    reconcile_pending_target(&mut target, config, sink)
}

fn reconcile_pending_target(
    target: &mut ReconcileTarget<'_>,
    config: &ScanConfig,
    sink: &mut dyn FnMut(&Commit),
) -> Result<ReconcileReport> {
    config.validate_for_scope(target.scope()?)?;
    let roots = take_invalidation_roots(target)?;
    let mut combined = ReconcileReport::default();
    for (position, (root, reason)) in roots.iter().enumerate() {
        match reconcile_target(target, root, config, sink) {
            Ok(report) => {
                if target.retries_incomplete(&report) {
                    target.restore_pending_invalidations(vec![(root.clone(), *reason)])?;
                }
                merge_reconcile_report(&mut combined, report);
            }
            Err(error) => {
                target.restore_pending_invalidations(roots[position..].to_vec())?;
                return Err(error);
            }
        }
    }
    Ok(combined)
}

fn take_invalidation_roots(
    target: &mut ReconcileTarget<'_>,
) -> Result<Vec<(PathBuf, crate::InvalidateReason)>> {
    let mut pending = target.take_pending_invalidations()?;
    pending.sort_by(|(left, _), (right, _)| {
        left.components().count().cmp(&right.components().count()).then_with(|| left.cmp(right))
    });
    let mut roots: Vec<(PathBuf, crate::InvalidateReason)> = Vec::new();
    for (path, reason) in pending {
        if roots.iter().any(|(root, _)| path.starts_with(root)) {
            continue;
        }
        roots.push((path, reason));
    }

    Ok(roots)
}

fn remove_known_children(
    target: &mut ReconcileTarget<'_>,
    path: &Path,
    config: &ScanConfig,
    batch: &mut Vec<ObservationOp>,
    sink: &mut dyn FnMut(&Commit),
    report: &mut ReconcileReport,
) -> Result<()> {
    for (name, baseline) in target.child_states(path)? {
        batch.push(ObservationOp::if_state(Op::Remove { path: path.join(name) }, baseline));
        if batch.len() >= config.batch_size.max(1) {
            flush_reconcile_batch(target, batch, sink, report)?;
        }
    }
    flush_reconcile_batch(target, batch, sink, report)
}

fn push_reconcile_upsert(
    target: &ReconcileTarget<'_>,
    path: &Path,
    kind: EntryKind,
    attrs: Attrs,
    baseline: PathExpectation,
    batch: &mut Vec<ObservationOp>,
    report: &mut ReconcileReport,
) {
    // An exclusive Index borrow cannot race another index producer. If filesystem
    // metadata exactly matches the captured state, applying this upsert can only be a
    // no-op, so avoid allocating an owned op and walking the index again. Shared
    // reconciliation keeps the conditional observation so ABA arbitration remains
    // authoritative between its read and write lock boundaries.
    if target.direct_upsert_is_unchanged(baseline, kind, attrs) {
        report.observations = report.observations.saturating_add(1);
        report.apply.unchanged = report.apply.unchanged.saturating_add(1);
        return;
    }
    batch.push(ObservationOp::if_state(
        Op::Upsert { path: path.to_path_buf(), kind, attrs },
        baseline,
    ));
}

fn flush_reconcile_batch(
    target: &mut ReconcileTarget<'_>,
    batch: &mut Vec<ObservationOp>,
    sink: &mut dyn FnMut(&Commit),
    report: &mut ReconcileReport,
) -> Result<()> {
    if batch.is_empty() {
        return Ok(());
    }
    report.observations =
        report.observations.saturating_add(u64::try_from(batch.len()).unwrap_or(u64::MAX));
    let started_at = report.reconcile_epoch.expect("reconciliation report has an owner");
    let outcome = target.apply(started_at, &Observation::from_ops(std::mem::take(batch)))?;
    merge_apply_stats(&mut report.apply, outcome.stats);
    if let Some(commit) = outcome.commit.as_ref() {
        sink(commit);
    }
    Ok(())
}

fn merge_apply_stats(total: &mut ApplyStats, addition: ApplyStats) {
    total.inserted += addition.inserted;
    total.updated += addition.updated;
    total.removed += addition.removed;
    total.unchanged += addition.unchanged;
    total.invalidated += addition.invalidated;
    total.controls += addition.controls;
    total.reclassified += addition.reclassified;
    total.stale += addition.stale;
    total.resource_refused += addition.resource_refused;
}

fn merge_reconcile_report(total: &mut ReconcileReport, addition: ReconcileReport) {
    total.retry_required |= addition.retry_required;
    total.scan.dirs_read += addition.scan.dirs_read;
    total.scan.entries += addition.scan.entries;
    total.scan.files_walked += addition.scan.files_walked;
    total.scan.bytes_walked = total.scan.bytes_walked.saturating_add(addition.scan.bytes_walked);
    total.scan.allocated_walked =
        total.scan.allocated_walked.saturating_add(addition.scan.allocated_walked);
    total.scan.errors.extend(addition.scan.errors);
    total.observations = total.observations.saturating_add(addition.observations);
    merge_apply_stats(&mut total.apply, addition.apply);
    total.listed_incomplete.extend(addition.listed_incomplete);
}

fn should_descend(
    kind: EntryKind,
    attrs: Attrs,
    parent_depth: usize,
    root_dev: u64,
    config: &ScanConfig,
) -> bool {
    crate::admission::should_descend(
        kind,
        attrs,
        parent_depth,
        root_dev,
        config.max_depth,
        config.one_filesystem,
    )
}

pub(crate) fn normalize_subtree(path: &Path) -> Result<PathBuf> {
    let mut normalized = PathBuf::new();
    for component in path.components() {
        match component {
            Component::Normal(part) => normalized.push(part),
            Component::CurDir => {}
            Component::ParentDir | Component::RootDir | Component::Prefix(_) => {
                return Err(Error::PathEscapesRoot(path.to_path_buf()));
            }
        }
    }
    Ok(normalized)
}

fn resolve_subtree_root(
    target: &ReconcileTarget<'_>,
    subtree: &Path,
    config: &ScanConfig,
) -> Result<PathBuf> {
    if subtree.as_os_str().is_empty() {
        return Ok(PathBuf::new());
    }
    let root = target.root_path()?;
    crate::counters::bump(|c| c.stats += 1);
    let Ok(root_metadata) = fs::symlink_metadata(&root) else {
        // The applying pass reports operational root failures as partial.
        return Ok(subtree.to_path_buf());
    };
    if !root_metadata.is_dir() {
        return Ok(subtree.to_path_buf());
    }
    let Ok(root_dev) = root_device(&root, &root_metadata) else {
        return Ok(subtree.to_path_buf());
    };
    let mut prefix = PathBuf::new();
    let mut components = subtree.components().peekable();
    while let Some(component) = components.next() {
        if components.peek().is_none() {
            break; // The boundary entry itself remains visible even when descent stops.
        }
        prefix.push(component.as_os_str());
        let (kind, attrs) = match observe_path(&root.join(&prefix)) {
            Ok(observed) => observed,
            Err(error)
                if matches!(
                    error.kind(),
                    std::io::ErrorKind::NotFound | std::io::ErrorKind::NotADirectory
                ) =>
            {
                return Ok(prefix);
            }
            Err(_) => break, // The applying pass records operational failures as partial.
        };
        if kind == EntryKind::Symlink {
            return Err(Error::SubtreeOutsideScanScope {
                path: subtree.to_path_buf(),
                scope: config.scope(),
            });
        }
        if kind != EntryKind::Dir {
            return Ok(prefix);
        }
        if config.one_filesystem && root_dev != 0 && attrs.dev != 0 && attrs.dev != root_dev {
            return Err(Error::SubtreeOutsideScanScope {
                path: subtree.to_path_buf(),
                scope: config.scope(),
            });
        }
    }
    Ok(subtree.to_path_buf())
}

/// Read an entry's kind and roll-up attributes out of its metadata.
///
/// Exposed so the watch layer verifies entries exactly the way the walker records them —
/// two stat interpretations that could drift would show up as an index that disagrees
/// with itself depending on which producer last touched a path.
///
/// On Windows the observation comes from a fresh non-following handle, and `meta` is
/// what answers for an entry whose handle cannot be opened because it is locked or
/// access is denied — the same fallback std's `metadata` makes, with identity and change
/// time unavailable for that entry.
pub fn observe(path: &Path, meta: &fs::Metadata) -> std::io::Result<(EntryKind, Attrs)> {
    #[cfg(windows)]
    {
        windows_metadata::observe(path, || Ok(meta.clone()))
    }
    #[cfg(not(windows))]
    {
        Ok((kind_from(meta), attrs_from(path, meta)?))
    }
}

pub(crate) fn observe_dir_entry(
    entry: &fs::DirEntry,
) -> std::io::Result<Option<(EntryKind, Attrs)>> {
    #[cfg(windows)]
    {
        crate::counters::bump(|c| c.stats += 1);
        #[cfg(test)]
        {
            let path = entry.path();
            if let Some(error) =
                walk_hook(&path).and_then(|hook| hook(WalkHookPoint::ChildMetadata(&path)))
            {
                return missing_as_none(Err(error));
            }
        }
        // The listing already holds the entry's enumeration data; it is read only when
        // the handle cannot be opened, so the ordinary path allocates nothing more.
        missing_as_none(windows_metadata::observe(&entry.path(), || entry.metadata()))
    }
    #[cfg(not(windows))]
    {
        #[cfg(test)]
        {
            let path = entry.path();
            if let Some(error) = child_metadata_hook(&path) {
                return missing_as_none(Err(error));
            }
        }
        #[cfg(all(target_os = "linux", target_env = "gnu"))]
        {
            // std's `DirEntry::metadata` is `statx` without `AT_NO_AUTOMOUNT` wherever
            // `statx` is served, and std keeps the listing's descriptor to itself, so the
            // reader's stat answers by path (fdu-d2fn). This is the route of a directory
            // the reader declined, so the whole-path resolution is paid rarely.
            if !linux_dents::statx_unavailable() {
                crate::counters::bump(|c| c.stats += 1);
                if let Some(observed) = linux_dents::stat_path(&entry.path()) {
                    return missing_as_none(observed);
                }
                // `statx` has just been found unavailable: std answers, with `fstatat`.
            }
        }
        let Some(meta) = missing_as_none(metadata_for_fingerprint(entry))? else {
            return Ok(None);
        };
        Ok(Some((kind_from(&meta), attrs_from(Path::new(""), &meta)?)))
    }
}

/// The non-following observation of `path` itself, as [`observe`] reads it from
/// `symlink_metadata`, and on Linux one that never triggers an automount.
///
/// For a path a route holds as an entry and verifies by itself: a reconciliation's
/// subtree and the prefixes above it, a change the watch verifies, a control looked up
/// by name. The walk root is not one; [`root_device`] says why.
pub(crate) fn observe_path(path: &Path) -> std::io::Result<(EntryKind, Attrs)> {
    crate::counters::bump(|c| c.stats += 1);
    #[cfg(all(target_os = "linux", target_env = "gnu"))]
    if let Some(observed) = linux_dents::stat_path(path) {
        return observed;
    }
    let metadata = fs::symlink_metadata(path)?;
    observe(path, &metadata)
}

#[cfg(not(windows))]
fn kind_from(meta: &fs::Metadata) -> EntryKind {
    let file_type = meta.file_type();
    if file_type.is_symlink() {
        EntryKind::Symlink
    } else if file_type.is_dir() {
        EntryKind::Dir
    } else if file_type.is_file() {
        EntryKind::File
    } else {
        EntryKind::Other
    }
}

#[cfg(unix)]
#[allow(clippy::unnecessary_wraps)] // Windows observation is fallible; keep one call contract.
pub(crate) fn attrs_from(_path: &Path, meta: &fs::Metadata) -> std::io::Result<Attrs> {
    use std::os::unix::fs::MetadataExt;
    Ok(Attrs {
        size: meta.size(),
        // st_blocks is in 512-byte units by POSIX convention regardless of the
        // filesystem's own block size.
        allocated: meta.blocks().saturating_mul(512),
        mtime_ns: compose_ns(meta.mtime(), meta.mtime_nsec()),
        ctime_ns: compose_ns(meta.ctime(), meta.ctime_nsec()),
        inode: meta.ino(),
        dev: meta.dev(),
    })
}

#[cfg(unix)]
fn compose_ns(secs: i64, nanos: i64) -> i64 {
    secs.saturating_mul(1_000_000_000).saturating_add(nanos)
}

// Windows observation goes through `windows_metadata::observe` on every route; only a
// test still derives attributes from metadata alone.
#[cfg(all(windows, test))]
pub(crate) fn attrs_from(path: &Path, meta: &fs::Metadata) -> std::io::Result<Attrs> {
    windows_metadata::observe(path, || Ok(meta.clone())).map(|(_, attrs)| attrs)
}

#[cfg(not(any(unix, windows)))]
#[allow(clippy::unnecessary_wraps)] // Windows observation is fallible; keep one call contract.
pub(crate) fn attrs_from(_path: &Path, meta: &fs::Metadata) -> std::io::Result<Attrs> {
    let mtime_ns = meta.modified().map_or(0, system_time_ns);
    Ok(Attrs {
        size: meta.len(),
        // No allocated size without platform-specific calls; apparent size is the
        // honest fallback rather than a guess at block rounding.
        allocated: meta.len(),
        mtime_ns,
        // Windows has no ctime in the Unix sense. Leaving it zero means the fingerprint
        // degrades to size + mtime there, which is what every portable tool does.
        ctime_ns: 0,
        inode: 0,
        dev: 0,
    })
}

/// The device a walk's root is on, which bounds a one-filesystem walk.
///
/// Only the device is needed, and on Windows it is read without demanding a consistent
/// observation of the root's times, which change whenever a child is created or removed.
/// The device that bounds a `one_filesystem` walk from `root`, whose non-following
/// metadata is `meta`.
///
/// The walk root is resolved. The user named it and the listing that follows opens it,
/// and on Linux an open mounts an unmounted autofs trigger where a non-following stat
/// need not: `symlink_metadata` does on glibc (std's `statx` without `AT_NO_AUTOMOUNT`)
/// and does not on musl (`fstatat`). A trigger's own device would then exclude every
/// entry the listing finds under it, so on Linux the device is read from an opened
/// descriptor on every route and under either libc, and `meta`'s device answers only
/// when the open fails, as the listing then fails the same way. Every other stat of the
/// tree never mounts ([`observe_dir_entry`], [`observe_path`], `linux_dents`).
pub(crate) fn root_device(root: &Path, meta: &fs::Metadata) -> std::io::Result<u64> {
    #[cfg(windows)]
    {
        let _ = meta;
        windows_metadata::volume_serial(root)
    }
    #[cfg(target_os = "linux")]
    {
        use std::os::unix::fs::OpenOptionsExt as _;
        let opened = fs::OpenOptions::new()
            .read(true)
            .custom_flags(libc::O_DIRECTORY | libc::O_NOFOLLOW)
            .open(root)
            .and_then(|directory| directory.metadata());
        attrs_from(root, opened.as_ref().unwrap_or(meta)).map(|attrs| attrs.dev)
    }
    #[cfg(not(any(windows, target_os = "linux")))]
    {
        attrs_from(root, meta).map(|attrs| attrs.dev)
    }
}

pub(crate) fn attrs_from_file(file: &fs::File, meta: &fs::Metadata) -> std::io::Result<Attrs> {
    #[cfg(windows)]
    {
        let _ = meta;
        windows_metadata::attrs_from_file(file)
    }
    #[cfg(not(windows))]
    {
        let _ = file;
        attrs_from(Path::new(""), meta)
    }
}

#[cfg(any(not(any(unix, windows)), test))]
fn system_time_ns(time: std::time::SystemTime) -> i64 {
    match time.duration_since(std::time::UNIX_EPOCH) {
        Ok(duration) => i64::try_from(duration.as_nanos()).unwrap_or(i64::MAX),
        Err(error) => {
            i64::try_from(error.duration().as_nanos()).map_or(i64::MIN, i64::saturating_neg)
        }
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use std::fs::File;
    use std::io::Write;

    fn write_file(path: &Path, contents: &[u8]) {
        if let Some(parent) = path.parent() {
            fs::create_dir_all(parent).expect("create parent");
        }
        let mut f = File::create(path).expect("create file");
        f.write_all(contents).expect("write");
    }

    fn sample_tree() -> tempfile::TempDir {
        let dir = tempfile::tempdir().expect("tempdir");
        write_file(&dir.path().join("a.txt"), b"hello");
        write_file(&dir.path().join("src/main.rs"), b"fn main() {}");
        write_file(&dir.path().join("src/deep/nested.rs"), b"// nested");
        crate::test_support::settle_allocations(dir.path());
        dir
    }

    /// A counter that silently reads zero is worse than a missing one, because a report
    /// full of zeroes invites the conclusion that the work did not happen.
    ///
    /// This has already gone wrong twice: once when the per-entry counter was added to
    /// the serial walk while the parallel walk went uninstrumented, and once when a
    /// clippy fix hoisted a `read_dir` out of a match scrutinee and took the counter
    /// with it. Both builds compiled, passed every other test, and reported zero. This
    /// asserts the relationships a real walk must satisfy, so the next such edit fails
    /// here instead of in a report someone believes.
    #[test]
    fn a_walk_moves_every_counter_it_should() {
        // Both walkers, because they are separate loops with separate call sites. The
        // first version of this test only exercised the parallel one, and deleting the
        // serial walker's counter still passed — a guard covering one path gives false
        // confidence about the other.
        let _serial = crate::counters::test_serial();
        crate::counters::enable(true);
        for threads in [Some(1), Some(4)] {
            let dir = sample_tree();
            let config = ScanConfig { threads, ..ScanConfig::default() };

            // Deltas around the scan, not absolute totals. The counters are
            // process-global, so a test running beside this one can add to them — and
            // `test_serial` cannot prevent that, since it only serializes tests that
            // take it, not every test that happens to walk a tree.
            let before = crate::counters::snapshot();
            let report = scan(dir.path(), &config, &mut |_| {}).expect("scan");
            crate::counters::flush_thread();
            let after = crate::counters::snapshot();
            let observed_entries = after.dir_entries - before.dir_entries;
            let observed_opens = after.dir_opens - before.dir_opens;
            let observed_stats = after.stats - before.stats;

            // `>=` rather than `==`, and the direction is the whole point: concurrent
            // work can only inflate these, never deflate them. So a counter that is too
            // low means a path ran uninstrumented, which is the failure worth catching
            // and the one that has actually happened — the macOS bulk reader reported
            // zero opens against three real ones. Equality would catch double-counting
            // too, and would be flaky for it.
            assert!(
                observed_entries >= report.entries,
                "every enumerated entry is counted at {threads:?}: {observed_entries} < {}",
                report.entries
            );
            assert!(
                observed_opens >= report.dirs_read,
                "every directory open is counted at {threads:?}: {observed_opens} < {}",
                report.dirs_read
            );
            assert!(
                observed_stats >= report.entries,
                "every entry is stated at {threads:?}: {observed_stats} < {}",
                report.entries
            );
            #[cfg(any(target_os = "macos", all(target_os = "linux", target_env = "gnu")))]
            {
                let observed_enum = after.dir_enumeration_calls - before.dir_enumeration_calls;
                // The serial walker is the portable `read_dir` path, which cannot see
                // getdents multiplicity. Enumeration calls are a native-backend fact.
                if threads != Some(1) {
                    assert!(
                        observed_enum >= report.dirs_read,
                        "every successful native directory issues at least one enumeration \
                         call at {threads:?}: {observed_enum} < {}",
                        report.dirs_read
                    );
                }
            }

            // Deliberately not asserted: `allocs` stays zero in a library test, because
            // allocation counting needs a binary to install `CountingAlloc` as its
            // global allocator and a test harness installs its own. The probe covers
            // that half; this covers the counters the library itself drives.
        }
        crate::counters::enable(false);
    }

    #[test]
    fn summary_fold_skips_stat_on_directories_and_symlinks() {
        let _serial = crate::counters::test_serial();
        crate::counters::enable(true);
        let dir = tempfile::tempdir().expect("tempdir");
        fs::create_dir(dir.path().join("src")).expect("directory");
        write_file(&dir.path().join("a.txt"), b"hi");
        #[cfg(unix)]
        std::os::unix::fs::symlink("a.txt", dir.path().join("link")).expect("symlink");
        let config = ScanConfig { threads: Some(1), read_controls: false, ..ScanConfig::default() };

        crate::counters::test_thread_reset();
        let scan_report = scan(dir.path(), &config, &mut |_| {}).expect("scan");
        let scan_stats = crate::counters::test_thread_snapshot().stats;

        crate::counters::test_thread_reset();
        let fold_report = scan_summary_fold(dir.path(), &config, &mut |_| {}).expect("fold");
        let fold_stats = crate::counters::test_thread_snapshot().stats;
        crate::counters::enable(false);

        assert_eq!(fold_report.entries, scan_report.entries);
        assert_eq!(fold_report.files_walked, scan_report.files_walked);
        assert_eq!(fold_report.bytes_walked, scan_report.bytes_walked);
        // Windows observes every listed entry through a fresh handle on both paths, so the
        // fold performs exactly the retained walk's observations there; the skip is a
        // non-Windows saving.
        #[cfg(unix)]
        assert!(
            fold_stats < scan_stats,
            "fold {fold_stats} should skip directory/symlink stats versus scan {scan_stats}"
        );
        // The fold stats `a.txt` and, when the listing yields `src` or `link` before it,
        // the first of those, whose stat proves the directory searchable
        // ([`Searchability`]); the others take their kind from the listing.
        #[cfg(not(windows))]
        let proof = u64::from(first_listed(dir.path()) != "a.txt");
        #[cfg(unix)]
        assert_eq!(scan_stats.saturating_sub(fold_stats), 2 - proof);
        #[cfg(not(any(unix, windows)))]
        assert_eq!(scan_stats.saturating_sub(fold_stats), 1 - proof);
        #[cfg(windows)]
        assert_eq!(fold_stats, scan_stats);
    }

    /// The name the filesystem lists first in `dir`, read rather than assumed: under the
    /// skip a listing stats its children until one stat succeeds ([`Searchability`]), so
    /// how many stats the fold saves depends on the enumeration order.
    #[cfg(not(windows))]
    fn first_listed(dir: &Path) -> std::ffi::OsString {
        fs::read_dir(dir).expect("listing").next().expect("an entry").expect("entry").file_name()
    }

    #[cfg(unix)]
    #[test]
    fn summary_fold_still_stats_directories_when_bound_to_one_filesystem() {
        let _serial = crate::counters::test_serial();
        crate::counters::enable(true);
        let dir = tempfile::tempdir().expect("tempdir");
        fs::create_dir(dir.path().join("src")).expect("directory");
        write_file(&dir.path().join("a.txt"), b"hi");
        std::os::unix::fs::symlink("a.txt", dir.path().join("link")).expect("symlink");
        let config = ScanConfig {
            threads: Some(1),
            read_controls: false,
            one_filesystem: true,
            ..ScanConfig::default()
        };

        crate::counters::test_thread_reset();
        let scan_report = scan(dir.path(), &config, &mut |_| {}).expect("scan");
        let scan_stats = crate::counters::test_thread_snapshot().stats;

        crate::counters::test_thread_reset();
        let fold_report = scan_summary_fold(dir.path(), &config, &mut |_| {}).expect("fold");
        let fold_stats = crate::counters::test_thread_snapshot().stats;
        crate::counters::enable(false);

        assert_eq!(fold_report.entries, scan_report.entries);
        // `src` and `a.txt` are stated on both routes; `link` takes its kind from the
        // listing only once one of them has been stated before it, which proves the
        // directory searchable ([`Searchability`]).
        let proof = u64::from(first_listed(dir.path()) == "link");
        assert_eq!(scan_stats.saturating_sub(fold_stats), 1 - proof);
    }

    #[test]
    fn summary_fold_reuses_cleared_recycled_batches() {
        // Four workers and a batch of three force StreamingEmission to send more than
        // once per worker on this tree. Without `recycled.clear()`, the next send
        // re-folds the previous ops and files/bytes/dirs double-count.
        const DIRS: usize = 16;
        const FILES_PER_DIR: usize = 40;
        let dir = tempfile::tempdir().expect("tempdir");
        let mut expected_bytes = 0u64;
        for directory in 0..DIRS {
            let child = dir.path().join(format!("d{directory:02}"));
            fs::create_dir(&child).expect("directory");
            for file in 0..FILES_PER_DIR {
                let size = directory * FILES_PER_DIR + file + 1;
                expected_bytes += size as u64;
                write_file(&child.join(format!("f{file:02}.dat")), &vec![b'x'; size]);
            }
        }
        let expected_files = (DIRS * FILES_PER_DIR) as u64;
        let expected_dirs = DIRS as u64;
        let expected_entries = expected_files + expected_dirs;
        let config = ScanConfig {
            threads: Some(4),
            batch_size: 3,
            read_controls: false,
            ..ScanConfig::default()
        };
        let mut files = 0u64;
        let mut bytes = 0u64;
        let mut dirs = 0u64;
        let mut ops = 0u64;
        let report = scan_summary_fold(dir.path(), &config, &mut |observed| {
            ops += 1;
            let Op::Upsert { kind, attrs, .. } = &observed.op else {
                return;
            };
            match kind {
                EntryKind::File => {
                    files += 1;
                    bytes += attrs.size;
                }
                EntryKind::Dir => dirs += 1,
                EntryKind::Symlink | EntryKind::Other => {}
            }
        })
        .expect("fold");
        assert_eq!(files, expected_files);
        assert_eq!(bytes, expected_bytes);
        assert_eq!(dirs, expected_dirs);
        assert_eq!(ops, report.entries);
        assert_eq!(report.entries, expected_entries);
        assert_eq!(report.files_walked, expected_files);
        assert_eq!(report.bytes_walked, expected_bytes);
    }

    /// Publish one listing through `emission` and receive what the consumer would.
    fn publish_detached(
        emission: &mut DetachedEmission,
        directory: DetachedDirectory,
        sender: &std::sync::mpsc::Sender<WalkMessage>,
        receiver: &std::sync::mpsc::Receiver<WalkMessage>,
    ) -> (Vec<DetachedDirectory>, std::sync::mpsc::Sender<Vec<DetachedDirectory>>) {
        emission.finish_directory(directory);
        let mut send_ns = 0;
        assert!(emission.publish_before_discovery(true, sender, &mut send_ns, None));
        match receiver.try_recv().expect("a published chunk") {
            WalkMessage::DetachedDirectories { directories, recycle } => (directories, recycle),
            _ => panic!("the detached walker publishes only listings"),
        }
    }

    /// H185: a folded index describes each directory once, by its own listing, so its
    /// directory and symlink entries carry the listing's default attributes, as the
    /// transient summary's do (H72); the full index keeps the parent's stat, which is the
    /// cache's freshness fingerprint. Under `--one-filesystem` descent reads each
    /// directory's device, so the folded walk keeps that stat too. macOS lists every
    /// child's attributes in bulk and Windows never takes the skip, so neither shows it.
    #[cfg(all(unix, not(target_os = "macos")))]
    #[test]
    fn a_folded_index_takes_directory_and_symlink_kinds_from_the_listing() {
        let root = tempfile::tempdir().expect("temp root");
        write_file(&root.path().join("dir/file.txt"), b"contents");
        write_file(&root.path().join("dir/nested/deep.txt"), b"more");
        write_file(&root.path().join("top.txt"), b"top");
        std::os::unix::fs::symlink("top.txt", root.path().join("link")).expect("symlink");
        let canonical = root.path().canonicalize().expect("canonical root");
        let retention = crate::execution::TreeRetention {
            largest_files: 100,
            size: crate::query::SizeMetric::Allocated,
        };
        let attrs_by_kind = |index: &Index| -> Vec<(EntryKind, bool)> {
            let mut seen = Vec::new();
            let mut stack = vec![(PathBuf::new(), crate::EntryId::ROOT)];
            while let Some((path, id)) = stack.pop() {
                let Some(children) = index.children_of(id) else { continue };
                for (name, child) in children {
                    let kind = index.kind_of(child).expect("live child");
                    let attrs = index.attrs_of(child).expect("live child");
                    seen.push((kind, *attrs == Attrs::default()));
                    if kind.is_dir() {
                        stack.push((path.join(name), child));
                    }
                }
            }
            seen.sort_by_key(|(kind, defaulted)| (*kind as u8, *defaulted));
            seen
        };

        for threads in [1, 4] {
            let config = ScanConfig { threads: Some(threads), ..ScanConfig::default() };
            let (folded, _, _) =
                scan_into_folded_index(&canonical, &config, retention, false).expect("folded");
            let (full, _) = scan_into_index(&canonical, &config).expect("full");
            for (kind, defaulted) in attrs_by_kind(&folded) {
                assert_eq!(
                    defaulted,
                    matches!(kind, EntryKind::Dir | EntryKind::Symlink),
                    "{threads} workers: a folded {kind:?} takes its kind from the listing"
                );
            }
            assert!(
                attrs_by_kind(&full).iter().all(|(_, defaulted)| !defaulted),
                "{threads} workers: the full index stats every entry"
            );

            let bound = ScanConfig { one_filesystem: true, ..config };
            let (folded, _, _) =
                scan_into_folded_index(&canonical, &bound, retention, false).expect("folded");
            for (kind, defaulted) in attrs_by_kind(&folded) {
                assert_eq!(
                    defaulted,
                    kind == EntryKind::Symlink,
                    "{threads} workers, one filesystem: directories keep their device"
                );
            }
        }
    }

    #[test]
    fn detached_emission_reuses_returned_listings_as_fresh_ones() {
        // H159: the consumer hands drained listings back to the worker that allocated
        // them. A reused listing must be indistinguishable from a fresh one, including
        // after the consumer returned it unapplied, as it does after a build error.
        let (sender, receiver) = std::sync::mpsc::channel();
        let mut emission = DetachedEmission::new(false);

        let mut skipped = emission.begin_directory(Path::new("a/much/longer/relative/path"));
        skipped.children.push(DetachedChild {
            name: OsString::from("stale.txt"),
            kind: EntryKind::File,
            attrs: Attrs::default(),
            position: 0,
        });
        skipped.control = Some(Op::ControlRemove { path: PathBuf::from("a/.gitignore") });
        let (directories, recycle) = publish_detached(&mut emission, skipped, &sender, &receiver);
        let returned_list = directories.as_ptr();
        let returned_children = directories[0].children.as_ptr();
        recycle.send(directories).expect("the worker still listens");

        // Returned listings are taken back at the next publish, so this one is fresh.
        let mut large = emission.begin_directory(Path::new("large"));
        assert_eq!(large.children.capacity(), 0);
        large.children.reserve_exact(DETACHED_SPARE_CHILD_CAPACITY + 1);
        let (directories, recycle) = publish_detached(&mut emission, large, &sender, &receiver);
        recycle.send(directories).expect("the worker still listens");

        let reused = emission.begin_directory(Path::new("b"));
        assert_eq!(reused.path.as_os_str(), OsStr::new("b"));
        assert!(reused.children.is_empty(), "a returned listing's children are dropped");
        assert!(reused.control.is_none(), "a returned listing's control is dropped");
        assert_eq!(reused.children.as_ptr(), returned_children, "the child buffer is reused");
        let (directories, _recycle) = publish_detached(&mut emission, reused, &sender, &receiver);
        assert_eq!(directories.as_ptr(), returned_list, "the published list is reused");

        // The large listing came back past the retention bound and was freed instead.
        let fresh = emission.begin_directory(Path::new("c"));
        assert_eq!(fresh.path.as_os_str(), OsStr::new("c"));
        assert_eq!(fresh.children.capacity(), 0);
    }

    /// A transient fold that reads `.gitignore` receives each directory's control, once,
    /// ahead of every entry in that directory, whatever the batch size, the worker count,
    /// or where `.gitignore` falls in the listing (fdu-1ovb). Its consumer classifies each
    /// entry as it arrives, so an entry folded before its directory's control would miss
    /// that control's rules. Batches smaller than a listing force the probe path; the
    /// default batch holds whole listings and moves the listed control.
    #[test]
    fn a_classifying_summary_fold_receives_each_control_ahead_of_its_entries() {
        const DIRS: usize = 12;
        const FILES_PER_DIR: usize = 30;
        let dir = tempfile::tempdir().expect("tempdir");
        write_file(&dir.path().join(".gitignore"), b"*.log\n");
        for directory in 0..DIRS {
            let child = dir.path().join(format!("d{directory:02}"));
            write_file(&child.join(".gitignore"), b"*.tmp\n");
            for file in 0..FILES_PER_DIR {
                write_file(&child.join(format!("f{file:02}.dat")), b"x");
            }
            write_file(&child.join("nested/n.dat"), b"n");
        }
        let default_batch = ScanConfig::default().batch_size;
        for (threads, batch_size) in [
            (Some(1), 1),
            (Some(4), 1),
            (Some(4), 3),
            (None, 7),
            (Some(4), 16),
            (None, default_batch),
        ] {
            let config = ScanConfig { batch_size, threads, ..ScanConfig::default() };
            assert!(config.read_controls, "observation is the default");
            // Per directory: the positions of its control observations and of its entries.
            let mut seen: std::collections::BTreeMap<PathBuf, (Vec<usize>, Vec<usize>)> =
                std::collections::BTreeMap::new();
            let mut position = 0;
            scan_summary_fold(dir.path(), &config, &mut |observed| {
                let (path, control) = match &observed.op {
                    Op::Upsert { path, .. } => (path, false),
                    Op::ControlUpsert { path, .. } | Op::ControlRemove { path } => (path, true),
                    other => panic!("a cold walk emitted {other:?}"),
                };
                let directory = path.parent().unwrap_or_else(|| Path::new("")).to_path_buf();
                let (controls, entries) = seen.entry(directory).or_default();
                if control { controls } else { entries }.push(position);
                position += 1;
            })
            .expect("fold");

            assert_eq!(seen.len(), 1 + 2 * DIRS, "the root, each child, and each nested");
            for (directory, (controls, entries)) in &seen {
                let label = format!("{directory:?} with {threads:?} workers, batch {batch_size}");
                let governed = directory.as_os_str().is_empty()
                    || directory.file_name().is_some_and(|name| name != "nested");
                assert_eq!(controls.len(), usize::from(governed), "{label}: one control read");
                if let (Some(last_control), Some(first_entry)) =
                    (controls.iter().max(), entries.iter().min())
                {
                    assert!(last_control < first_entry, "{label}: its control comes first");
                }
            }
        }
    }

    /// A directory's control is what a lookup of `<dir>/.gitignore` resolves to, as git
    /// opens it (fdu-0w1b). A listed `.GITIGNORE` reads exactly what that lookup finds,
    /// named by the canonical path: its rules where the directory is case-insensitive, and
    /// nothing where it is not. The probe a transient fold makes is that same lookup.
    #[test]
    fn a_directory_control_is_what_a_lookup_of_its_canonical_path_resolves_to() {
        use crate::test_support::CaseLookups;

        let dir = tempfile::tempdir().expect("tempdir");
        write_file(&dir.path().join("exact/.gitignore"), b"*.log\n");
        write_file(&dir.path().join("variant/.GITIGNORE"), b"*.tmp\n");
        fs::create_dir_all(dir.path().join("shape/.GitIgnore")).expect("a directory so named");
        write_file(&dir.path().join("absent/README"), b"no rules");
        let config = ScanConfig::default();
        let upsert = |path: &str, source: &[u8]| {
            Some(Op::ControlUpsert { path: PathBuf::from(path), source: source.to_vec() })
        };

        for (lookups, insensitive) in CaseLookups::on_this_host(dir.path()) {
            let _lookups = lookups.install(dir.path());
            let label = format!("{lookups:?} lookups");
            let lookup = |directory: &str| {
                read_directory_control(
                    &config,
                    dir.path(),
                    &Path::new(directory).join(".gitignore"),
                )
                .expect("lookup")
            };
            let listed = |path: &str, kind| {
                let path = Path::new(path);
                let read = read_control_op(&config, dir.path(), path, kind).expect("listed read");
                let name = path.file_name().expect("a listed name");
                assert_eq!(
                    read_named_control_op(&config, dir.path(), path, name, kind).expect("named"),
                    read,
                    "{label}: a walker's read by the listed name decides as the path's (H180)"
                );
                read
            };

            assert_eq!(lookup("exact"), upsert("exact/.gitignore", b"*.log\n"), "{label}");
            assert_eq!(listed("exact/.gitignore", EntryKind::File), lookup("exact"), "{label}");
            assert_eq!(
                lookup("variant"),
                if insensitive { upsert("variant/.gitignore", b"*.tmp\n") } else { None },
                "{label}"
            );
            assert_eq!(
                listed("variant/.GITIGNORE", EntryKind::File),
                lookup("variant"),
                "{label}: a listed variant reads what the lookup finds"
            );
            assert_eq!(
                listed("shape/.GitIgnore", EntryKind::Dir),
                insensitive.then(|| Op::ControlRemove { path: PathBuf::from("shape/.gitignore") }),
                "{label}: a directory is resolved to, and holds no rules"
            );
            assert_eq!(lookup("absent"), None, "{label}");
            assert_eq!(listed("absent/README", EntryKind::File), None, "{label}");

            let blind = ScanConfig { read_controls: false, ..ScanConfig::default() };
            assert_eq!(
                read_control_op(
                    &blind,
                    dir.path(),
                    Path::new("variant/.GITIGNORE"),
                    EntryKind::File
                )
                .expect("read"),
                None,
                "{label}: a scan that reads no rules looks nothing up"
            );
            assert_eq!(
                read_named_control_op(
                    &blind,
                    dir.path(),
                    Path::new("exact/.gitignore"),
                    OsStr::new(".gitignore"),
                    EntryKind::File
                )
                .expect("read"),
                None,
                "{label}: nor does a read by the listed name"
            );
        }
    }

    /// The walker's join makes [`Path::join`]'s path byte for byte, including under the
    /// root, whose relative directory is empty, and for a name longer than its directory
    /// (H180).
    #[test]
    fn a_listed_name_joins_as_path_join_does() {
        for (rel_dir, name) in [
            ("", "file"),
            ("", ".gitignore"),
            ("a", "b"),
            ("a/b", ".GITIGNORE"),
            ("node_modules/x", "a-name-longer-than-twice-its-directory.js"),
        ] {
            let joined = join_listed_name(Path::new(rel_dir), OsStr::new(name));
            assert_eq!(
                joined.as_os_str(),
                Path::new(rel_dir).join(name).as_os_str(),
                "{rel_dir:?} and {name:?}"
            );
        }
    }

    /// Every entry's kind and ignored classification, and every retained control source:
    /// what a route's index says about a tree's `.gitignore` rules.
    type Classification = (BTreeMap<PathBuf, (EntryKind, Option<bool>)>, Vec<(PathBuf, Vec<u8>)>);

    fn classification(index: &Index) -> Classification {
        let mut entries = BTreeMap::new();
        let mut pending = vec![PathBuf::new()];
        while let Some(directory) = pending.pop() {
            let children: Vec<(PathBuf, crate::EntryId)> = index
                .children(&directory)
                .into_iter()
                .flatten()
                .map(|(name, id)| (directory.join(name), id))
                .collect();
            for (path, id) in children {
                let kind = index.kind_of(id).expect("a live child");
                entries.insert(path.clone(), (kind, index.is_ignored(&path).expect("observed")));
                if kind.is_dir() {
                    pending.push(path);
                }
            }
        }
        let sources = index
            .controls()
            .expect("observed")
            .sources()
            .map(|(path, source)| (path, source.to_vec()))
            .collect();
        (entries, sources)
    }

    /// A tree whose `up` directory holds its rules only in a case variant of the control
    /// name, beside a root `.gitignore` those rules partly override.
    fn case_variant_tree() -> tempfile::TempDir {
        let dir = tempfile::tempdir().expect("tempdir");
        write_file(&dir.path().join(".gitignore"), b"*.log\n");
        write_file(&dir.path().join("a.log"), b"root rule");
        write_file(&dir.path().join("keep.txt"), b"kept");
        write_file(&dir.path().join("up/.GITIGNORE"), b"*.tmp\n!keep.log\nbuild/\n");
        write_file(&dir.path().join("up/x.tmp"), b"variant rule");
        write_file(&dir.path().join("up/keep.log"), b"negated by the variant");
        write_file(&dir.path().join("up/y.log"), b"root rule");
        write_file(&dir.path().join("up/build/out.bin"), b"variant rule");
        write_file(&dir.path().join("up/deep/z.tmp"), b"variant rule");
        for file in 0..12 {
            write_file(&dir.path().join(format!("up/f{file:02}.dat")), b"unignored");
        }
        dir
    }

    /// What [`case_variant_tree`] classifies where its variant does and does not govern.
    fn assert_case_variant_classification(classified: &Classification, governs: bool, label: &str) {
        let (entries, sources) = classified;
        let ignored = |path: &str| entries.get(Path::new(path)).map(|(_, ignored)| *ignored);
        assert_eq!(ignored("a.log"), Some(Some(true)), "{label}");
        assert_eq!(ignored("up/y.log"), Some(Some(true)), "{label}");
        assert_eq!(ignored("up/x.tmp"), Some(Some(governs)), "{label}");
        assert_eq!(ignored("up/deep/z.tmp"), Some(Some(governs)), "{label}");
        assert_eq!(ignored("up/build"), Some(Some(governs)), "{label}");
        assert_eq!(ignored("up/build/out.bin"), Some(Some(governs)), "{label}");
        assert_eq!(ignored("up/keep.log"), Some(Some(!governs)), "{label}: negation");
        assert_eq!(ignored("up/f00.dat"), Some(Some(false)), "{label}");
        let governing: Vec<&Path> = sources.iter().map(|(path, _)| path.as_path()).collect();
        let expected: Vec<&Path> = if governs {
            vec![Path::new(".gitignore"), Path::new("up/.gitignore")]
        } else {
            vec![Path::new(".gitignore")]
        };
        assert_eq!(governing, expected, "{label}: rules are recorded by the canonical path");
    }

    /// The detached builder, the retained scanner stream (serial and concurrent), and the
    /// narrowed-population walks classify a tree whose rules sit in `.GITIGNORE` alike, and
    /// apply those rules exactly where a lookup of `.gitignore` resolves to it (fdu-0w1b).
    /// Hidden pruning keeps the variant a control signal, as it keeps `.gitignore` one.
    #[test]
    fn every_cold_route_classifies_a_case_variant_control_alike() {
        use crate::test_support::CaseLookups;

        let dir = case_variant_tree();
        let root = dir.path();
        let pruned = Some(std::sync::Arc::new(crate::HiddenPolicy::prune_hidden(Vec::<
            std::ffi::OsString,
        >::new())));
        for (lookups, governs) in CaseLookups::on_this_host(root) {
            let _lookups = lookups.install(root);
            let reference = ScanConfig { threads: Some(1), ..ScanConfig::default() };
            let (index, report) = scan_into_index(root, &reference).expect("detached scan");
            assert!(report.is_complete(), "{:?}", report.errors);
            let expected = classification(&index);
            assert_case_variant_classification(&expected, governs, &format!("{lookups:?}"));

            for threads in [Some(1), Some(4)] {
                for batch_size in [ScanConfig::default().batch_size, 1, 3] {
                    let config = ScanConfig { batch_size, threads, ..ScanConfig::default() };
                    let label = format!("{lookups:?}, {threads:?} workers, batch {batch_size}");
                    let (detached, _) = scan_into_index(root, &config).expect("detached");
                    assert_eq!(classification(&detached), expected, "{label}: detached");
                    let (streamed, _) = scan_into_index_via_scanner(root, &config).expect("stream");
                    assert_eq!(classification(&streamed), expected, "{label}: scanner stream");
                }
            }

            // Pruning hidden entries drops the variant's row, never its rules.
            let hidden = ScanConfig { hidden: pruned.clone(), ..ScanConfig::default() };
            let (without_hidden, _) = scan_into_index(root, &hidden).expect("hidden pruned");
            let (entries, sources) = classification(&without_hidden);
            assert_eq!(sources, expected.1, "{lookups:?}: hidden pruning keeps the rules");
            for (path, fact) in &entries {
                assert_eq!(expected.0.get(path), Some(fact), "{lookups:?}: {path:?}");
            }
            assert!(!entries.contains_key(Path::new("up/.GITIGNORE")), "{lookups:?}");

            // A narrowed population reads each control before listing its directory, which
            // is the lookup itself; it keeps every spelling's row, as it keeps `.gitignore`.
            let spelled = |path: &Path| crate::control::path_control_spelling(path).is_some();
            for population in
                [crate::query::IgnoredEntries::Exclude, crate::query::IgnoredEntries::Only]
            {
                let config = ScanConfig { population, ..ScanConfig::default() };
                let (narrowed, report) = scan_into_index(root, &config).expect("narrowed");
                assert!(report.is_complete(), "{:?}", report.errors);
                let (entries, sources) = classification(&narrowed);
                let label = format!("{lookups:?}, {population:?}");
                assert_eq!(sources, expected.1, "{label}");
                let files =
                    |entries: &BTreeMap<PathBuf, (EntryKind, Option<bool>)>| -> Vec<PathBuf> {
                        entries
                            .iter()
                            .filter(|(_, (kind, _))| *kind == EntryKind::File)
                            .map(|(path, _)| path.clone())
                            .collect()
                    };
                let wanted: BTreeMap<PathBuf, (EntryKind, Option<bool>)> = expected
                    .0
                    .iter()
                    .filter(|(path, (_, ignored))| {
                        spelled(path)
                            || match population {
                                crate::query::IgnoredEntries::Exclude => *ignored == Some(false),
                                _ => *ignored == Some(true),
                            }
                    })
                    .map(|(path, fact)| (path.clone(), *fact))
                    .collect();
                assert_eq!(files(&entries), files(&wanted), "{label}: retained files");
            }
        }
    }

    /// Reconciliation keeps a case-variant control exactly as a cold walk finds it through
    /// every change: a case-only rename of `.gitignore` to `.GITIGNORE`, an edit, a
    /// removal, and a refresh of the variant's own path after it is created and after it
    /// is gone. Serial and parallel reconciliation, and revalidation, agree with a cold
    /// walk after each (fdu-0w1b).
    ///
    /// The rename is the case a canonical path cannot follow on its own: the index still
    /// holds the old `.gitignore` entry, and removing an entry at the canonical path drops
    /// the rules it governs, which on a case-insensitive directory the variant still holds.
    #[test]
    fn reconciliation_follows_a_case_variant_control_through_every_change() {
        use crate::test_support::CaseLookups;

        let probe = tempfile::tempdir().expect("tempdir");
        for (lookups, governs) in CaseLookups::on_this_host(probe.path()) {
            for threads in [Some(1), Some(4)] {
                let dir = tempfile::tempdir().expect("tempdir");
                let root = dir.path().canonicalize().expect("canonical root");
                let _lookups = lookups.install(&root);
                let config = ScanConfig { threads, batch_size: 3, ..ScanConfig::default() };
                write_file(&root.join(".gitignore"), b"*.log\n");
                write_file(&root.join("up/.gitignore"), b"*.tmp\n");
                write_file(&root.join("up/x.tmp"), b"governed");
                write_file(&root.join("up/y.bin"), b"governed after the edit");
                for file in 0..8 {
                    write_file(&root.join(format!("up/f{file}.dat")), b"unignored");
                }
                let (mut index, _) = scan_into_index(&root, &config).expect("cold scan");
                let cold = |label: &str| {
                    let (cold, report) = scan_into_index(&root, &config).expect("cold");
                    assert!(report.is_complete(), "{label}: {:?}", report.errors);
                    classification(&cold)
                };
                let reconciled = |index: &mut Index, label: &str| {
                    let report = reconcile(index, &config, &mut |_| {}).expect("reconcile");
                    assert!(report.is_complete(), "{label}: {:?}", report.scan.errors);
                    assert_eq!(report.apply.stale, 0, "{label}: no observation lost a race");
                };
                let revalidated = |snapshot: &Index, label: &str| {
                    let mut revalidated = snapshot.clone();
                    let mut observations = Vec::new();
                    revalidate(&revalidated, &config, &mut |observation| {
                        observations.push(observation);
                    })
                    .expect("revalidate");
                    for observation in &observations {
                        let outcome = revalidated.apply(observation).expect("apply");
                        assert_eq!(outcome.stats.stale, 0, "{label}: revalidation lost a race");
                    }
                    classification(&revalidated)
                };
                let governed = |label: &str| {
                    let (entries, _) = cold(label);
                    entries.get(Path::new("up/x.tmp")).map(|(_, ignored)| *ignored)
                };

                let before_rename = index.clone();
                fs::rename(root.join("up/.gitignore"), root.join("up/.GITIGNORE")).expect("recase");
                let label = format!("{lookups:?}, {threads:?} workers: case-only rename");
                assert_eq!(governed(&label), Some(Some(governs)), "{label}");
                assert_eq!(
                    revalidated(&before_rename, &label),
                    cold(&label),
                    "{label}: revalidate"
                );
                reconciled(&mut index, &label);
                assert_eq!(classification(&index), cold(&label), "{label}");

                write_file(&root.join("up/.GITIGNORE"), b"*.bin\n");
                let label = format!("{lookups:?}, {threads:?} workers: edit");
                reconciled(&mut index, &label);
                assert_eq!(classification(&index), cold(&label), "{label}");

                fs::remove_file(root.join("up/.GITIGNORE")).expect("remove the variant");
                let label = format!("{lookups:?}, {threads:?} workers: removal");
                reconciled(&mut index, &label);
                assert_eq!(classification(&index), cold(&label), "{label}");

                // A refresh of the variant's own path looks the directory's control up
                // whether the variant is there or gone.
                write_file(&root.join("up/.GITIGNORE"), b"*.tmp\n");
                for (step, label) in [
                    (None, format!("{lookups:?}, {threads:?} workers: refresh of a new variant")),
                    (
                        Some(()),
                        format!("{lookups:?}, {threads:?} workers: refresh of a removed variant"),
                    ),
                ] {
                    if step.is_some() {
                        fs::remove_file(root.join("up/.GITIGNORE")).expect("remove");
                    }
                    let report = reconcile_subtree(
                        &mut index,
                        Path::new("up/.GITIGNORE"),
                        &config,
                        &mut |_| {},
                    )
                    .expect("refresh");
                    assert!(report.is_complete(), "{label}: {:?}", report.scan.errors);
                    assert_eq!(classification(&index), cold(&label), "{label}");
                }
            }
        }
    }

    /// Revalidate a copy of `index` and apply every observation it sends, none of which may
    /// lose a race.
    fn revalidated_copy(index: &Index, config: &ScanConfig, label: &str) -> Index {
        let mut revalidated = index.clone();
        let mut observations = Vec::new();
        let report = revalidate(&revalidated, config, &mut |observation| {
            observations.push(observation);
        })
        .expect("revalidate");
        assert!(report.is_complete(), "{label}: {:?}", report.errors);
        for observation in &observations {
            let outcome = revalidated.apply(observation).expect("apply");
            assert_eq!(outcome.stats.stale, 0, "{label}: revalidation lost a race");
        }
        revalidated
    }

    /// A narrowed population's reconciliation keeps a case-variant control through a
    /// case-only rename, as a cold walk finds it. Its lookup before the listing is what
    /// finds the variant's rules, and the listing shows the variant, so the sweep restates
    /// the rules its removal of the stale `.gitignore` entry drops (fdu-0w1b).
    #[test]
    fn a_narrowed_reconciliation_follows_a_case_only_rename() {
        use crate::test_support::CaseLookups;

        let probe = tempfile::tempdir().expect("tempdir");
        for (lookups, governs) in CaseLookups::on_this_host(probe.path()) {
            for population in
                [crate::query::IgnoredEntries::Exclude, crate::query::IgnoredEntries::Only]
            {
                let label = format!("{lookups:?}, {population:?}");
                let dir = tempfile::tempdir().expect("tempdir");
                let root = dir.path().canonicalize().expect("canonical root");
                let _lookups = lookups.install(&root);
                let config = ScanConfig { population, batch_size: 3, ..ScanConfig::default() };
                write_file(&root.join("up/.gitignore"), b"*.tmp\n");
                write_file(&root.join("up/x.tmp"), b"governed");
                for file in 0..4 {
                    write_file(&root.join(format!("up/f{file}.dat")), b"unignored");
                }
                let (mut index, _) = scan_into_index(&root, &config).expect("cold scan");
                fs::rename(root.join("up/.gitignore"), root.join("up/.GITIGNORE")).expect("recase");
                let (cold, report) = scan_into_index(&root, &config).expect("cold");
                assert!(report.is_complete(), "{label}: {:?}", report.errors);
                let cold = classification(&cold);
                assert_eq!(
                    cold.1.iter().any(|(path, _)| path == Path::new("up/.gitignore")),
                    governs,
                    "{label}: the variant governs exactly where the lookup resolves to it"
                );

                let revalidated = revalidated_copy(&index, &config, &label);
                assert_eq!(classification(&revalidated), cold, "{label}: revalidate");
                let report = reconcile(&mut index, &config, &mut |_| {}).expect("reconcile");
                assert!(report.is_complete(), "{label}: {:?}", report.scan.errors);
                assert_eq!(report.apply.stale, 0, "{label}: no observation lost a race");
                assert_eq!(classification(&index), cold, "{label}: reconcile");
            }
        }
    }

    /// A narrowed population's reconciliation looks a directory's control up before
    /// listing it. When the `.gitignore` that lookup found is deleted before the listing,
    /// the sweep removes the stale entry, and the rules go with it: no case variant was
    /// listed to hold them, so nothing is restated, and the table is as bare as the
    /// directory. The next pass agrees with a cold walk.
    #[test]
    fn a_control_deleted_between_its_lookup_and_the_listing_leaves_no_rules() {
        for population in
            [crate::query::IgnoredEntries::Exclude, crate::query::IgnoredEntries::Only]
        {
            for revalidating in [true, false] {
                let label = format!("{population:?}, revalidating: {revalidating}");
                let dir = tempfile::tempdir().expect("tempdir");
                let root = dir.path().canonicalize().expect("canonical root");
                let config = ScanConfig { population, ..ScanConfig::default() };
                write_file(&root.join("up/.gitignore"), b"*.tmp\n");
                write_file(&root.join("up/x.tmp"), b"governed");
                write_file(&root.join("up/y.dat"), b"unignored");
                let control_path = Path::new("up/.gitignore");
                let (mut index, _) = scan_into_index(&root, &config).expect("cold scan");
                assert!(index.control_table().contains(control_path), "{label}");

                let control = root.join(control_path);
                let deleted = std::sync::Arc::new(std::sync::atomic::AtomicBool::new(false));
                let hook_deleted = std::sync::Arc::clone(&deleted);
                let race = install_walk_hook(&root, move |point| {
                    if let WalkHookPoint::ControlLookup(looked_up) = point {
                        if looked_up == control && fs::remove_file(looked_up).is_ok() {
                            hook_deleted.store(true, std::sync::atomic::Ordering::SeqCst);
                        }
                    }
                    None
                });
                if revalidating {
                    index = revalidated_copy(&index, &config, &label);
                } else {
                    let report = reconcile(&mut index, &config, &mut |_| {}).expect("reconcile");
                    assert!(report.is_complete(), "{label}: {:?}", report.scan.errors);
                }
                drop(race);
                assert!(
                    deleted.load(std::sync::atomic::Ordering::SeqCst),
                    "{label}: the lookup found the control before it was deleted"
                );
                assert_eq!(index.path_state(control_path), PathState::Absent, "{label}");
                assert!(
                    !index.control_table().contains(control_path),
                    "{label}: no rules remain for a file that is gone"
                );

                let report = reconcile(&mut index, &config, &mut |_| {}).expect("reconcile");
                assert!(report.is_complete(), "{label}: {:?}", report.scan.errors);
                let (cold, _) = scan_into_index(&root, &config).expect("cold");
                assert_eq!(classification(&index), classification(&cold), "{label}");
            }
        }
    }

    /// The sweep restates what a lookup found only after removing the stale `.gitignore`
    /// entry of a listing that showed a case variant, the one spelling that can still hold
    /// the rules that removal drops. Without a listed variant the exact file is gone, and
    /// so are its rules.
    #[test]
    fn a_listing_restates_looked_up_rules_only_after_showing_a_case_variant() {
        let found =
            Op::ControlUpsert { path: PathBuf::from("up/.gitignore"), source: b"*.tmp\n".to_vec() };
        let exact = OsStr::new(".gitignore");

        let mut deleted = ListedControl::default();
        deleted.lookup(&Ok(Some(found.clone())));
        deleted.listed(OsStr::new("x.tmp"));
        assert_eq!(deleted.restatement_after_removing(exact), None, "the exact file is gone");

        let mut recased = ListedControl::default();
        recased.lookup(&Ok(Some(found.clone())));
        recased.listed(OsStr::new(".GITIGNORE"));
        assert_eq!(recased.restatement_after_removing(exact), Some(found.clone()));
        assert_eq!(
            recased.restatement_after_removing(OsStr::new(".GitIgnore")),
            None,
            "only removing the canonical entry drops rules"
        );
        assert_eq!(recased.restatement_after_removing(OsStr::new("x.tmp")), None);

        // An included population looks the control up through the listed variant itself.
        let mut read = ListedControl::default();
        read.listed(OsStr::new(".GitIgnore"));
        read.read(OsStr::new(".GitIgnore"), &Ok(Some(found.clone())));
        assert_eq!(read.restatement_after_removing(exact), Some(found));

        let mut missed = ListedControl::default();
        missed.lookup(&Ok(None));
        missed.listed(OsStr::new(".GITIGNORE"));
        assert_eq!(missed.restatement_after_removing(exact), None, "nothing was found");
    }

    /// Where a directory is case-sensitive it can list `.gitignore` beside `.GITIGNORE`,
    /// and only the exact name governs, on every route and through every change: the
    /// variant's lookup resolves to the exact file, and removing the variant leaves the
    /// rules while removing the exact name takes them (fdu-0w1b).
    #[test]
    fn a_case_sensitive_directory_listing_both_spellings_takes_only_the_exact_name() {
        let dir = tempfile::tempdir().expect("tempdir");
        let root = dir.path().canonicalize().expect("canonical root");
        write_file(&root.join(".gitignore"), b"*.log\n");
        write_file(&root.join(".GITIGNORE"), b"*.tmp\n");
        let listed = fs::read_dir(&root).expect("list").count();
        if listed != 2 {
            eprintln!(
                "skipped: the temporary directory is case-insensitive, so it cannot hold both \
                 spellings"
            );
            return;
        }
        write_file(&root.join("x.log"), b"exact rule");
        write_file(&root.join("y.tmp"), b"variant, not a rule here");
        let ignored =
            |index: &Index, path: &str| index.is_ignored(Path::new(path)).expect("observed");

        for threads in [Some(1), Some(4)] {
            let config = ScanConfig { threads, batch_size: 1, ..ScanConfig::default() };
            let (detached, _) = scan_into_index(&root, &config).expect("detached");
            let (streamed, _) = scan_into_index_via_scanner(&root, &config).expect("stream");
            for index in [&detached, &streamed] {
                assert_eq!(ignored(index, "x.log"), Some(true), "{threads:?}");
                assert_eq!(ignored(index, "y.tmp"), Some(false), "{threads:?}");
                assert_eq!(classification(index), classification(&detached), "{threads:?}");
            }
        }

        let config = ScanConfig::default();
        let (mut index, _) = scan_into_index(&root, &config).expect("cold");
        fs::remove_file(root.join(".GITIGNORE")).expect("remove the variant");
        reconcile(&mut index, &config, &mut |_| {}).expect("reconcile");
        assert_eq!(ignored(&index, "x.log"), Some(true), "the exact name still governs");
        write_file(&root.join(".GITIGNORE"), b"*.tmp\n");
        fs::remove_file(root.join(".gitignore")).expect("remove the exact name");
        reconcile(&mut index, &config, &mut |_| {}).expect("reconcile");
        assert_eq!(ignored(&index, "x.log"), Some(false), "no rules remain");
        assert_eq!(ignored(&index, "y.tmp"), Some(false), "the variant never governs here");
        let (cold, _) = scan_into_index(&root, &config).expect("cold");
        assert_eq!(classification(&index), classification(&cold));
    }

    /// A listing whose control was already probed does not read its `.gitignore` again:
    /// the entry is prepared with no control and, here, no error from a file that cannot
    /// be read, where reading it would have produced one.
    #[test]
    #[cfg(unix)]
    fn a_probed_listing_prepares_its_control_entry_without_reading_it() {
        use std::os::unix::fs::PermissionsExt;
        if !crate::test_support::require_permission_bits() {
            return;
        }
        let dir = tempfile::tempdir().expect("tempdir");
        let control = dir.path().join(".gitignore");
        write_file(&control, b"*.log\n");
        fs::set_permissions(&control, fs::Permissions::from_mode(0o000)).expect("deny");
        let config = ScanConfig::default();
        let prepare = |read_control| {
            prepare_walk_entry_reading(
                dir.path(),
                Path::new(""),
                0,
                OsStr::new(".gitignore"),
                EntryKind::File,
                Attrs::default(),
                0,
                &config,
                read_control,
            )
            .expect("admitted")
        };
        let read = prepare(true);
        let skipped = prepare(false);
        fs::set_permissions(&control, fs::Permissions::from_mode(0o600)).expect("restore");

        assert!(read.control.is_none() && read.control_error.is_some(), "the read was made");
        assert!(skipped.control.is_none() && skipped.control_error.is_none(), "no read");
        assert!(skipped.retained, "the entry is still a row");
    }

    /// An automatic walk too short to fill its calibration window must say so.
    ///
    /// The failure this guards is quiet: such a walk runs on its initial pool, which is
    /// indistinguishable in the artifacts from a walk that measured the filesystem and
    /// chose to hold — unless the undecided case is recorded separately. Reading the
    /// first as the second is how a policy with no evidence behind it comes to look
    /// like a policy with evidence behind it.
    #[test]
    fn a_short_automatic_walk_records_an_undecided_policy() {
        let _serial = crate::counters::test_serial();
        let available = std::thread::available_parallelism().map_or(1, std::num::NonZeroUsize::get);
        if automatic_worker_pool(available).calibration.is_none() {
            // A host reporting one processor has no reserve to unlock, so there is no
            // policy here to leave undecided.
            return;
        }

        crate::counters::enable(true);
        let dir = sample_tree();
        let config = ScanConfig { threads: None, ..ScanConfig::default() };
        let before = crate::counters::snapshot();
        scan(dir.path(), &config, &mut |_| {}).expect("scan");
        crate::counters::flush_thread();
        let after = crate::counters::snapshot();
        crate::counters::enable(false);

        // A strict increase, so a counter inflated by a test running beside this one
        // cannot turn the assertion into a false pass.
        assert!(
            after.adaptive_policy_undecided > before.adaptive_policy_undecided,
            "a three-file tree cannot fill a {ADAPTIVE_SCAN_CALIBRATION_ENTRIES}-entry window"
        );
    }

    #[test]
    fn diagnostics_make_a_fixed_pool_and_backend_choice_explicit() {
        let dir = sample_tree();
        let config = ScanConfig { threads: Some(1), ..ScanConfig::default() };

        let (report, diagnostics) =
            scan_with_diagnostics(dir.path(), &config, &mut |_| {}).expect("diagnostic scan");

        assert_eq!(diagnostics.schema, SCAN_DIAGNOSTICS_SCHEMA);
        assert_eq!(diagnostics.worker_policy.outcome, WorkerPolicyOutcome::Fixed);
        assert_eq!(diagnostics.worker_policy.initial_workers, 1);
        assert_eq!(diagnostics.worker_policy.maximum_workers, 1);
        assert_eq!(diagnostics.worker_policy.peak_active_workers, 1);
        assert!(diagnostics.worker_policy.windows.is_empty());
        assert!(!diagnostics.worker_policy.events_truncated);
        assert_eq!(diagnostics.worker_policy.ready_directories_at_finish, 0);
        assert_eq!(diagnostics.worker_policy.in_flight_directories_at_finish, 0);
        // On glibc the serial walk lists through the native reader too, and the Linux
        // backend fields count it: every attempt is a success or a fallback, and the
        // directories read are the native successes plus the portable reads.
        #[cfg(all(target_os = "linux", target_env = "gnu"))]
        {
            let backend = &diagnostics.backend;
            assert_eq!(backend.portable_attempts, backend.portable_directory_reads);
            assert!(
                backend.portable_directory_reads < report.dirs_read,
                "native listings are not portable reads: {backend:?}, {} read",
                report.dirs_read
            );
            let (Some(attempts), Some(successes), Some(fallbacks)) = (
                backend.linux_dents_attempts,
                backend.linux_dents_successes,
                backend.linux_dents_fallbacks,
            ) else {
                panic!("Linux native counts are present: {backend:?}");
            };
            assert_eq!(attempts, successes + fallbacks);
            assert!(successes > 0, "the serial walk lists natively: {backend:?}");
            assert_eq!(successes + backend.portable_directory_reads, report.dirs_read);
        }
        #[cfg(not(all(target_os = "linux", target_env = "gnu")))]
        assert_eq!(diagnostics.backend.portable_directory_reads, report.dirs_read);

        #[cfg(target_os = "macos")]
        {
            assert_eq!(diagnostics.backend.macos_bulk_attempts, Some(0));
            assert_eq!(diagnostics.backend.macos_bulk_successes, Some(0));
            assert_eq!(diagnostics.backend.macos_bulk_fallbacks, Some(0));
            assert!(diagnostics.backend.unavailable_reason.is_none());
        }
        #[cfg(not(target_os = "macos"))]
        {
            assert_eq!(diagnostics.backend.macos_bulk_attempts, None);
            assert_eq!(diagnostics.backend.macos_bulk_successes, None);
            assert_eq!(diagnostics.backend.macos_bulk_fallbacks, None);
            assert_eq!(
                diagnostics.backend.unavailable_reason,
                Some("macOS bulk directory enumeration is unavailable on this platform")
            );
        }

        // A parallel walk counts the same way, worker by worker.
        #[cfg(all(target_os = "linux", target_env = "gnu"))]
        {
            let config = ScanConfig { threads: Some(4), ..ScanConfig::default() };
            let (report, diagnostics) =
                scan_with_diagnostics(dir.path(), &config, &mut |_| {}).expect("diagnostic scan");
            let backend = &diagnostics.backend;
            assert!(report.is_complete(), "{:?}", report.errors);
            assert_eq!(backend.portable_attempts, backend.portable_directory_reads);
            assert!(
                backend.portable_directory_reads < report.dirs_read,
                "native listings are not portable reads: {backend:?}, {} read",
                report.dirs_read
            );
            assert_eq!(
                backend.unavailable_reason,
                Some("macOS bulk directory enumeration is unavailable on this platform")
            );
            let (Some(attempts), Some(successes), Some(fallbacks)) = (
                backend.linux_dents_attempts,
                backend.linux_dents_successes,
                backend.linux_dents_fallbacks,
            ) else {
                panic!("Linux native counts are present: {backend:?}");
            };
            assert_eq!(attempts, successes + fallbacks);
            assert!(successes > 0, "a parallel walk lists natively: {backend:?}");
            assert_eq!(successes + backend.portable_directory_reads, report.dirs_read);
        }
        #[cfg(not(all(target_os = "linux", target_env = "gnu")))]
        {
            assert_eq!(diagnostics.backend.linux_dents_attempts, None);
            assert_eq!(diagnostics.backend.linux_dents_successes, None);
            assert_eq!(diagnostics.backend.linux_dents_fallbacks, None);
        }
    }

    #[test]
    fn diagnostics_fail_closed_when_an_automatic_window_is_incomplete() {
        let available = std::thread::available_parallelism().map_or(1, std::num::NonZeroUsize::get);
        let pool = automatic_worker_pool(available);
        if pool.calibration.is_none() {
            return;
        }
        let dir = sample_tree();
        let config = ScanConfig { threads: None, ..ScanConfig::default() };

        let (report, diagnostics) =
            scan_with_diagnostics(dir.path(), &config, &mut |_| {}).expect("diagnostic scan");

        assert_eq!(diagnostics.worker_policy.outcome, WorkerPolicyOutcome::Undecided);
        assert_eq!(diagnostics.worker_policy.available_parallelism, available);
        assert_eq!(diagnostics.worker_policy.initial_workers, pool.initial);
        assert_eq!(diagnostics.worker_policy.maximum_workers, pool.maximum);
        assert_eq!(
            diagnostics.worker_policy.calibration_window_entries,
            Some(ADAPTIVE_SCAN_CALIBRATION_ENTRIES)
        );
        assert_eq!(
            diagnostics.worker_policy.slow_threshold_ns_per_entry,
            Some(ADAPTIVE_SCAN_SLOW_WORK_NS_PER_ENTRY)
        );
        assert_eq!(diagnostics.worker_policy.windows.len(), 1);
        let window = &diagnostics.worker_policy.windows[0];
        assert_eq!(window.sequence, 0);
        assert_eq!(window.start_entry_ordinal, 0);
        assert_eq!(window.end_entry_ordinal, report.entries);
        assert_eq!(window.observed_entries, report.entries);
        assert_eq!(window.decision, WorkerPolicyDecision::Undecided);
        assert!(window.end_entry_ordinal < ADAPTIVE_SCAN_CALIBRATION_ENTRIES);
        assert!(window.active_workers <= diagnostics.worker_policy.peak_active_workers);
        assert_eq!(diagnostics.worker_policy.ready_directories_at_finish, 0);
        assert_eq!(diagnostics.worker_policy.in_flight_directories_at_finish, 0);
        assert!(diagnostics.worker_policy.handoff_backlog_high_water >= 1);
    }

    #[test]
    fn diagnostic_trace_is_bounded_and_marks_truncation() {
        let recorder = ScanDiagnosticsRecorder::new(
            WorkerPool::fixed(2),
            2,
            WorkerPolicyExperiment::ShippedOneShot,
        );
        for sequence in 0..=MAX_POLICY_TRACE_EVENTS {
            recorder.record_policy_window(PolicyWindowSnapshot {
                sequence: sequence as u64,
                start_entry_ordinal: sequence as u64,
                end_entry_ordinal: sequence as u64 + 1,
                observed_entries: 1,
                observed_chunks: 1,
                observed_work_ns: 10,
                ready_directories: 1,
                in_flight_directories: 1,
                active_workers: 1,
                handoff_backlog: 0,
                requested_workers: None,
                decision: WorkerPolicyDecision::Hold,
            });
        }

        let diagnostics = recorder.finish();
        assert_eq!(diagnostics.worker_policy.windows.len(), MAX_POLICY_TRACE_EVENTS);
        assert!(diagnostics.worker_policy.events_truncated);
    }

    #[test]
    fn diagnostic_trace_preserves_queue_order_when_recorders_arrive_out_of_order() {
        let pool =
            WorkerPool { initial: 2, maximum: 4, calibration: Some(WorkerCalibration::new(1, 1)) };
        let recorder =
            ScanDiagnosticsRecorder::new(pool, 2, WorkerPolicyExperiment::RepeatedWindows);
        let snapshot = |sequence, decision| PolicyWindowSnapshot {
            sequence,
            start_entry_ordinal: sequence,
            end_entry_ordinal: sequence + 1,
            observed_entries: 1,
            observed_chunks: 1,
            observed_work_ns: 1,
            ready_directories: 0,
            in_flight_directories: 0,
            active_workers: 1,
            handoff_backlog: 0,
            requested_workers: None,
            decision,
        };

        recorder.record_policy_window(snapshot(1, WorkerPolicyDecision::HoldNoUsefulWork));
        recorder.record_policy_window(snapshot(0, WorkerPolicyDecision::Hold));

        let diagnostics = recorder.finish();
        assert_eq!(
            diagnostics
                .worker_policy
                .windows
                .iter()
                .map(|window| window.sequence)
                .collect::<Vec<_>>(),
            vec![0, 1]
        );
        assert_eq!(diagnostics.worker_policy.outcome, WorkerPolicyOutcome::HeldNoUsefulWork);
    }

    #[test]
    fn diagnostic_policy_aggregates_cross_check_runtime_counters() {
        let _serial = crate::counters::test_serial();
        let pool = WorkerPool {
            initial: 2,
            maximum: 4,
            calibration: Some(WorkerCalibration::new(17, 100)),
        };
        let recorder =
            ScanDiagnosticsRecorder::new(pool, 2, WorkerPolicyExperiment::ShippedOneShot);

        crate::counters::enable(true);
        let before = crate::counters::snapshot();
        record_adaptive_calibration_chunk(Some(&recorder), 17, 2_100);
        record_adaptive_worker_expansion(Some(&recorder));
        crate::counters::flush_thread();
        let after = crate::counters::snapshot();
        crate::counters::enable(false);

        recorder.record_policy_window(PolicyWindowSnapshot {
            sequence: 0,
            start_entry_ordinal: 0,
            end_entry_ordinal: 17,
            observed_entries: 17,
            observed_chunks: 1,
            observed_work_ns: 2_100,
            ready_directories: 2,
            in_flight_directories: 2,
            active_workers: 2,
            handoff_backlog: 0,
            requested_workers: Some(4),
            decision: WorkerPolicyDecision::ScaleUp,
        });
        let diagnostics = recorder.finish();
        let policy = diagnostics.worker_policy;
        assert_eq!(policy.calibration_chunks, 1);
        assert_eq!(policy.calibration_entries, 17);
        assert_eq!(policy.calibration_work_ns, 2_100);
        assert_eq!(policy.worker_expansions, 1);
        assert_eq!(policy.windows[0].observed_chunks, policy.calibration_chunks);
        assert_eq!(policy.windows[0].observed_entries, policy.calibration_entries);
        assert_eq!(policy.windows[0].observed_work_ns, policy.calibration_work_ns);

        // Other tests can record while this process-global interval is enabled, so the
        // counter delta may be larger but must never be smaller than this run-scoped
        // trace. The shared helpers above make the two observations one event.
        assert!(
            after.adaptive_calibration_chunks - before.adaptive_calibration_chunks
                >= policy.calibration_chunks
        );
        assert!(
            after.adaptive_calibration_entries - before.adaptive_calibration_entries
                >= policy.calibration_entries
        );
        assert!(
            after.adaptive_calibration_work_us - before.adaptive_calibration_work_us
                >= policy.calibration_work_ns / 1_000
        );
        assert!(after.adaptive_scale_ups - before.adaptive_scale_ups >= policy.worker_expansions);
    }

    #[test]
    fn diagnostic_index_scan_preserves_the_regular_result() {
        let dir = branching_tree();
        let config = ScanConfig { threads: Some(4), ..ScanConfig::default() };
        let (plain, plain_report) = scan_into_index(dir.path(), &config).expect("plain scan");
        let (diagnostic, diagnostic_report, diagnostics) =
            scan_into_index_with_diagnostics(dir.path(), &config).expect("diagnostic scan");

        assert_eq!(index_fingerprint(&plain), index_fingerprint(&diagnostic));
        assert_eq!(plain_report.entries, diagnostic_report.entries);
        assert_eq!(diagnostics.worker_policy.outcome, WorkerPolicyOutcome::Fixed);
    }

    #[test]
    fn detached_bootstrap_matches_the_streaming_reducer_for_each_worker_count() {
        let dir = branching_tree();
        for threads in 1..=4 {
            let config = ScanConfig {
                read_controls: false,
                threads: Some(threads),
                ..ScanConfig::default()
            };
            let _ = detached_and_streaming_indexes(dir.path(), &config);
        }
    }

    #[test]
    fn detached_control_bootstrap_matches_the_streaming_reducer_for_each_worker_count() {
        let dir = controlled_branching_tree();
        for threads in 1..=4 {
            let config =
                ScanConfig { read_controls: true, threads: Some(threads), ..ScanConfig::default() };
            let _ = detached_and_streaming_indexes(dir.path(), &config);
        }
    }

    #[test]
    fn detached_bootstrap_preserves_the_exact_first_mutation() {
        let dir = branching_tree();
        let config = ScanConfig { read_controls: false, threads: Some(4), ..ScanConfig::default() };
        let (mut detached, mut streaming) = detached_and_streaming_indexes(dir.path(), &config);
        let created = dir.path().join("t3/m2/after-bootstrap.rs");
        write_file(&created, b"new fact");
        let attrs =
            attrs_from(&created, &fs::symlink_metadata(&created).expect("new file metadata"))
                .expect("observe new file");
        let observation = Observation::new(vec![Op::Upsert {
            path: PathBuf::from("t3/m2/after-bootstrap.rs"),
            kind: EntryKind::File,
            attrs,
        }]);

        let detached_outcome = detached.apply(&observation).expect("detached mutation");
        let streaming_outcome = streaming.apply(&observation).expect("streaming mutation");
        assert_eq!(detached_outcome, streaming_outcome);
        assert_indexes_equal(&detached, &streaming);
    }

    #[test]
    fn detached_control_bootstrap_preserves_the_exact_first_mutation() {
        let dir = controlled_branching_tree();
        let config = ScanConfig { read_controls: true, threads: Some(4), ..ScanConfig::default() };
        let (mut detached, mut streaming) = detached_and_streaming_indexes(dir.path(), &config);
        let observation = Observation::new(vec![Op::ControlUpsert {
            path: PathBuf::from(".gitignore"),
            source: b"leaf-2.dat\n".to_vec(),
        }]);

        let detached_outcome = detached.apply(&observation).expect("detached control mutation");
        let streaming_outcome = streaming.apply(&observation).expect("streaming control mutation");
        assert_eq!(detached_outcome, streaming_outcome);
        assert_indexes_equal(&detached, &streaming);
    }

    fn observed_coverage(index: &Index) -> crate::control::ControlObservation {
        match index.control_coverage() {
            crate::control::ControlCoverage::Observed(observation) => observation,
            crate::control::ControlCoverage::NotObserved => panic!("controls were observed"),
        }
    }

    /// Both bootstrap lanes refuse a line over the limit and a file over the budget, and
    /// neither ends the scan or makes it partial. Both refusals are order-independent, so
    /// the lanes agree on exactly which files they refused.
    #[test]
    fn both_bootstrap_lanes_refuse_over_bound_controls_without_ending_the_scan() {
        let dir = tempfile::tempdir().expect("tempdir");
        let mut long_line = b"*.log\n".to_vec();
        long_line.extend(std::iter::repeat_n(b'x', crate::control::DEFAULT_CONTROL_LINE_LIMIT + 1));
        write_file(&dir.path().join("guarded/.gitignore"), &long_line);
        write_file(&dir.path().join("guarded/kept.log"), b"guarded");
        write_file(
            &dir.path().join("huge/.gitignore"),
            &b"x\n".repeat(crate::control::DEFAULT_CONTROL_BUDGET / 2),
        );
        write_file(&dir.path().join("applied/.gitignore"), b"*.log\n");
        write_file(&dir.path().join("applied/dropped.log"), b"applied");
        let config = ScanConfig { read_controls: true, threads: Some(4), ..ScanConfig::default() };

        let (detached, _) = detached_and_streaming_indexes(dir.path(), &config);
        let (_, report) = scan_into_index(dir.path(), &config).expect("scan");

        assert!(report.is_complete(), "{:?}", report.errors);
        let coverage = observed_coverage(&detached);
        assert_eq!((coverage.applied, coverage.refused), (1, 2));
        assert_eq!(
            coverage.refusals,
            vec![
                crate::control::RefusedControl {
                    path: PathBuf::from("guarded/.gitignore"),
                    reason: crate::control::ControlRefusalReason::LineLimit,
                },
                crate::control::RefusedControl {
                    path: PathBuf::from("huge/.gitignore"),
                    reason: crate::control::ControlRefusalReason::Budget,
                },
            ]
        );
        assert_eq!(detached.is_ignored(Path::new("guarded/kept.log")).expect("observed"), None);
        assert_eq!(
            detached.is_ignored(Path::new("applied/dropped.log")).expect("observed"),
            Some(true)
        );
    }

    /// The control counters attribute what a scan's control state cost: files read, sources
    /// refused, and sources that shared a retained content instead of parsing their own.
    ///
    /// Off by default and compiled in, like every counter, so the numbers a speed check
    /// reads come from the shipped path rather than an instrumented build.
    #[test]
    fn control_counters_attribute_reads_refusals_and_sharing() {
        let _serial = crate::counters::test_serial();
        let dir = tempfile::tempdir().expect("tempdir");
        let shared = b"*.log\n".to_vec();
        write_file(&dir.path().join(".gitignore"), &shared);
        write_file(&dir.path().join("twin/.gitignore"), &shared);
        let mut long_line = b"*.tmp\n".to_vec();
        long_line.extend(std::iter::repeat_n(b'x', crate::control::DEFAULT_CONTROL_LINE_LIMIT + 1));
        write_file(&dir.path().join("guarded/.gitignore"), &long_line);
        let config = ScanConfig { read_controls: true, threads: Some(1), ..ScanConfig::default() };

        crate::counters::enable(true);
        // Deltas around the scan rather than absolute totals, for the reason
        // `a_walk_moves_every_counter_it_should` gives: the counters are process-global,
        // `test_serial` only serializes the tests that take it, and every report in this
        // binary now reads `.gitignore` by default, so a test running beside this one can
        // add control reads of its own.
        let before = crate::counters::snapshot();
        let (index, report) = scan_into_index(dir.path(), &config).expect("scan");
        crate::counters::flush_thread();
        let after = crate::counters::snapshot();
        crate::counters::enable(false);

        assert!(report.is_complete(), "{:?}", report.errors);
        assert_eq!(observed_coverage(&index).refused, 1);
        // `>=` in the one direction concurrency can move them. A count that is too low
        // means a path ran uninstrumented, which is the defect worth catching; too high
        // is another test's tree, which is not.
        for (label, observed, expected) in [
            ("one read per .gitignore", after.control_reads - before.control_reads, 3),
            ("the line over the limit", after.control_refused - before.control_refused, 1),
            (
                "the twin shares one parsed content",
                after.control_sources_shared - before.control_sources_shared,
                1,
            ),
        ] {
            assert!(observed >= expected, "{label}: counted {observed}, expected {expected}");
        }
    }

    /// Both limits are part of the scope, and each lifts only its own refusals: no budget
    /// still refuses a long line, and no line limit still refuses a file past the budget.
    #[test]
    fn each_control_limit_is_scope_and_lifts_only_its_own_refusals() {
        use crate::control::{ControlLimits, ControlRefusalReason, RefusedControl};

        let with = |limits| ScanConfig { control_limits: limits, ..ScanConfig::default() };
        let defaults = ControlLimits::default();
        let default = ScanConfig::default();
        let no_budget = with(ControlLimits { budget: None, ..defaults });
        let no_line_limit = with(ControlLimits { line_limit: None, ..defaults });
        let configs = [
            default.clone(),
            with(ControlLimits { budget: Some(16 * 1024 * 1024), ..defaults }),
            no_budget.clone(),
            with(ControlLimits { line_limit: Some(64 * 1024), ..defaults }),
            no_line_limit.clone(),
            with(ControlLimits { budget: None, line_limit: None }),
            // The same values in each other's places are a different scope.
            with(ControlLimits { budget: defaults.line_limit, line_limit: defaults.budget }),
        ];
        let scopes: Vec<ScanScope> = configs.iter().map(ScanConfig::scope).collect();
        for (index, scope) in scopes.iter().enumerate() {
            assert!(scope.observes_controls());
            assert!(scopes[index + 1..].iter().all(|other| other != scope), "{scopes:?}");
        }
        for config in &configs {
            let blind = ScanConfig { read_controls: false, ..config.clone() };
            assert_eq!(blind.scope().ignore_rules_fingerprint, 0, "unobserved has one scope");
        }

        let dir = tempfile::tempdir().expect("tempdir");
        let mut long_line = b"*.log\n".to_vec();
        long_line.extend(std::iter::repeat_n(b'x', crate::control::DEFAULT_CONTROL_LINE_LIMIT + 1));
        write_file(&dir.path().join("guarded/.gitignore"), &long_line);
        write_file(&dir.path().join("guarded/dropped.log"), b"log");
        write_file(
            &dir.path().join("huge/.gitignore"),
            &b"x\n".repeat(crate::control::DEFAULT_CONTROL_BUDGET / 2),
        );
        let refused = |path: &str, reason| RefusedControl { path: PathBuf::from(path), reason };
        let (bounded, _) = scan_into_index(dir.path(), &default).expect("default scan");
        assert_eq!(observed_coverage(&bounded).refused, 2);

        let (budget_lifted, _) = detached_and_streaming_indexes(dir.path(), &no_budget);
        let coverage = observed_coverage(&budget_lifted);
        assert_eq!(coverage.limits, no_budget.control_limits);
        assert_eq!(
            coverage.refusals,
            [refused("guarded/.gitignore", ControlRefusalReason::LineLimit)]
        );
        assert_eq!(
            budget_lifted.is_ignored(Path::new("guarded/dropped.log")).expect("observed"),
            None
        );

        let (line_limit_lifted, _) = detached_and_streaming_indexes(dir.path(), &no_line_limit);
        let coverage = observed_coverage(&line_limit_lifted);
        assert_eq!(coverage.refusals, [refused("huge/.gitignore", ControlRefusalReason::Budget)]);
        assert_eq!(
            line_limit_lifted.is_ignored(Path::new("guarded/dropped.log")).expect("observed"),
            Some(true)
        );
        assert_eq!(line_limit_lifted.scope(), no_line_limit.scope());
    }

    /// The synthetic tree that ended a cold scan (fdu-1onj): 1,105 directories, each with
    /// a distinct 510-byte `.gitignore` of short rules, plus one line over the limit. The
    /// scan completes with every size exact and names what it refused, on both lanes.
    #[test]
    fn a_tree_past_both_control_bounds_completes_with_exact_sizes() {
        const DIRECTORIES: usize = 1_105;
        let dir = tempfile::tempdir().expect("tempdir");
        for directory in 0..DIRECTORIES {
            let mut source = Vec::new();
            for line in 0..63 {
                source.extend(format!("p{directory:04}{line:02}\n").bytes());
            }
            source.extend(format!("q{directory:04}\n").bytes());
            assert_eq!(source.len(), 510);
            let root = dir.path().join(format!("d{directory:04}"));
            write_file(&root.join(".gitignore"), &source);
            write_file(&root.join("file.txt"), b"contents");
        }
        write_file(
            &dir.path().join("a-guard/.gitignore"),
            &vec![b'x'; crate::control::DEFAULT_CONTROL_LINE_LIMIT + 1],
        );
        crate::test_support::settle_allocations(dir.path());
        let observing =
            ScanConfig { read_controls: true, threads: Some(4), ..ScanConfig::default() };
        let blind = ScanConfig { read_controls: false, ..observing.clone() };

        let (unobserved, _) = scan_into_index(dir.path(), &blind).expect("controls-off scan");
        let canonical = dir.path().canonicalize().expect("canonical root");
        let lanes = [
            scan_into_index(dir.path(), &observing).expect("detached scan"),
            scan_into_index_via_scanner(&canonical, &observing).expect("streaming scan"),
        ];
        for (index, report) in &lanes {
            assert!(report.is_complete(), "{:?}", report.errors);
            assert_eq!(index.total(), unobserved.total(), "sizes do not depend on controls");
            let coverage = observed_coverage(index);
            assert!(coverage.refused > 1, "the budget refused sources: {coverage:?}");
            assert_eq!(
                coverage.applied + coverage.refused,
                u64::try_from(DIRECTORIES + 1).expect("small")
            );
            assert_eq!(coverage.refusals.len(), crate::MAX_RETAINED_ISSUES);
            assert!(!coverage.lists_every_refusal());
            assert_eq!(
                coverage.refusals[0],
                crate::control::RefusedControl {
                    path: PathBuf::from("a-guard/.gitignore"),
                    reason: crate::control::ControlRefusalReason::LineLimit,
                }
            );
            assert!(
                index.control_table().retained_cost() <= crate::control::DEFAULT_CONTROL_BUDGET
            );
        }
    }

    #[test]
    fn fingerprint_metadata_observes_mutation_after_directory_enumeration() {
        let dir = tempfile::tempdir().expect("tempdir");
        let path = dir.path().join("changing.bin");
        write_file(&path, b"before");
        let entry = fs::read_dir(dir.path())
            .expect("read directory")
            .next()
            .expect("one entry")
            .expect("read entry");

        write_file(&path, b"after mutation");

        let metadata = metadata_for_fingerprint(&entry).expect("fresh metadata");
        assert_eq!(metadata.len(), b"after mutation".len() as u64);
    }

    /// A tree wide and deep enough that workers genuinely interleave.
    ///
    /// A three-file fixture would pass every one of these tests with a broken queue,
    /// because one worker would finish before another started.
    fn branching_tree() -> tempfile::TempDir {
        let dir = tempfile::tempdir().expect("tempdir");
        for top in 0..12 {
            for middle in 0..6 {
                for leaf in 0..7 {
                    write_file(
                        &dir.path().join(format!("t{top}/m{middle}/leaf-{leaf}.dat")),
                        &vec![b'x'; leaf * 13],
                    );
                }
            }
            // A deep chain alongside the wide fan-out, so depth and width are both
            // exercised by the same walk.
            write_file(&dir.path().join(format!("t{top}/a/b/c/d/e/deep.txt")), b"deep");
        }
        crate::test_support::settle_allocations(dir.path());
        dir
    }

    fn controlled_branching_tree() -> tempfile::TempDir {
        let dir = branching_tree();
        write_file(&dir.path().join(".gitignore"), b"leaf-1.dat\nt7/\n");
        write_file(&dir.path().join("t3/.gitignore"), b"!m2/leaf-1.dat\n*.tmp\n");
        write_file(&dir.path().join("t3/m2/generated.tmp"), b"ignored by nested control");
        write_file(&dir.path().join("t7/.gitignore"), b"!m0/leaf-1.dat\n");
        fs::create_dir_all(dir.path().join("t5/.gitignore")).expect("non-file control directory");
        write_file(&dir.path().join("t5/.gitignore/ordinary.txt"), b"ordinary child");
        crate::test_support::settle_allocations(dir.path());
        dir
    }

    fn index_fingerprint(index: &Index) -> Vec<(PathBuf, EntryKind, Attrs)> {
        let mut entries: Vec<(PathBuf, EntryKind, Attrs)> = Vec::new();
        let mut queue = vec![PathBuf::new()];
        while let Some(path) = queue.pop() {
            let Some(children) = index.children(&path) else {
                continue;
            };
            let names: Vec<PathBuf> = children.map(|(name, _id)| path.join(name)).collect();
            for child_path in names {
                let kind = index.kind(&child_path).expect("child has a kind");
                let attrs = *index.attrs(&child_path).expect("child has attrs");
                entries.push((child_path.clone(), kind, attrs));
                if kind.is_dir() {
                    queue.push(child_path);
                }
            }
        }
        entries.sort_by(|left, right| left.0.cmp(&right.0));
        entries
    }

    fn detached_and_streaming_indexes(root: &Path, config: &ScanConfig) -> (Index, Index) {
        let canonical = root.canonicalize().expect("canonical test root");
        let (streaming, streaming_report) =
            scan_into_index_via_scanner(&canonical, config).expect("streaming oracle");
        let (detached, detached_report) = scan_into_index(root, config).expect("detached scan");

        assert_eq!(detached_report.dirs_read, streaming_report.dirs_read);
        assert_eq!(detached_report.entries, streaming_report.entries);
        assert_eq!(detached_report.files_walked, streaming_report.files_walked);
        assert_eq!(detached_report.bytes_walked, streaming_report.bytes_walked);
        assert_eq!(
            detached_report.errors.iter().map(ToString::to_string).collect::<Vec<_>>(),
            streaming_report.errors.iter().map(ToString::to_string).collect::<Vec<_>>()
        );
        assert_indexes_equal(&detached, &streaming);
        (detached, streaming)
    }

    fn assert_indexes_equal(left: &Index, right: &Index) {
        assert_eq!(index_fingerprint(left), index_fingerprint(right));
        assert_eq!(left.total(), right.total());
        assert_eq!(left.partition_total().ok(), right.partition_total().ok());
        assert_eq!(left.scope(), right.scope());
        assert_eq!(left.freshness(), right.freshness());
        assert_eq!(left.state(), right.state());
        assert_eq!(left.clock(), right.clock());
        assert_eq!(left.len(), right.len());
        assert_eq!(left.issues(), right.issues());
        assert_eq!(left.observes_controls(), right.observes_controls());
        assert_eq!(left.control_coverage(), right.control_coverage());
        assert_eq!(
            left.control_table()
                .sources()
                .map(|(path, source)| (path, source.to_vec()))
                .collect::<Vec<_>>(),
            right
                .control_table()
                .sources()
                .map(|(path, source)| (path, source.to_vec()))
                .collect::<Vec<_>>()
        );
        for (path, _, _) in index_fingerprint(left) {
            assert_eq!(left.is_ignored(&path).ok(), right.is_ignored(&path).ok(), "{path:?}");
        }
    }

    /// A small tree whose mutation crosses every structural reconciliation boundary.
    fn reconciliation_transition_tree() -> tempfile::TempDir {
        let dir = tempfile::tempdir().expect("tempdir");
        write_file(&dir.path().join("changed.txt"), b"before");
        write_file(&dir.path().join("removed.txt"), b"remove me");
        write_file(&dir.path().join("directory-to-file/old.rs"), b"old child");
        write_file(&dir.path().join("file-to-directory"), b"old file");
        write_file(&dir.path().join("removed-tree/nested/gone.md"), b"gone");
        write_file(&dir.path().join("stable/deep/kept.rs"), b"kept");
        crate::test_support::settle_allocations(dir.path());
        dir
    }

    fn mutate_reconciliation_transition_tree(root: &Path) {
        write_file(&root.join("changed.txt"), b"after, with a distinct size");
        fs::remove_file(root.join("removed.txt")).expect("remove root file");

        fs::remove_dir_all(root.join("directory-to-file")).expect("remove old directory");
        write_file(&root.join("directory-to-file"), b"replacement file");

        fs::remove_file(root.join("file-to-directory")).expect("remove old file");
        write_file(&root.join("file-to-directory/new.txt"), b"replacement child");

        fs::remove_dir_all(root.join("removed-tree")).expect("remove nested tree");
        write_file(&root.join("added-tree/nested/new.md"), b"new nested file");
        crate::test_support::settle_allocations(root);
    }

    fn effective_ops(commits: &[Commit]) -> Vec<Op> {
        let mut operations: Vec<_> = commits
            .iter()
            .flat_map(|commit| commit.changes.iter())
            .filter_map(|change| match change {
                crate::EffectiveChange::Inserted { path, kind, attrs } => {
                    Some(Op::Upsert { path: path.clone(), kind: *kind, attrs: *attrs })
                }
                crate::EffectiveChange::Updated { path, kind, current, .. } => {
                    Some(Op::Upsert { path: path.clone(), kind: *kind, attrs: *current })
                }
                crate::EffectiveChange::Removed { path, .. } => {
                    Some(Op::Remove { path: path.clone() })
                }
                crate::EffectiveChange::Invalidated { path, reason } => {
                    Some(Op::InvalidateSubtree { path: path.clone(), reason: *reason })
                }
                crate::EffectiveChange::ControlUpdated { .. }
                | crate::EffectiveChange::ControlRefusalUpdated { .. }
                | crate::EffectiveChange::Reclassified { .. } => None,
            })
            .collect();
        operations.sort_by(|left, right| left.path().cmp(right.path()));
        operations
    }

    fn commit_touches(commit: &Commit, path: &Path) -> bool {
        commit.changes.iter().any(|change| change.path() == path)
    }

    #[test]
    fn parallel_and_serial_walks_produce_the_same_index() {
        let dir = branching_tree();
        let serial_config = ScanConfig { threads: Some(1), ..ScanConfig::default() };
        let (serial, serial_report) =
            scan_into_index(dir.path(), &serial_config).expect("serial scan");
        assert!(serial_report.is_complete());

        for threads in [2_usize, 3, 8] {
            let config = ScanConfig { threads: Some(threads), ..ScanConfig::default() };
            let (parallel, report) = scan_into_index(dir.path(), &config).expect("parallel scan");
            assert!(report.is_complete(), "{threads} threads reported errors");
            assert_eq!(report.entries, serial_report.entries, "{threads} threads");
            assert_eq!(report.dirs_read, serial_report.dirs_read, "{threads} threads");
            assert_eq!(report.files_walked, serial_report.files_walked, "{threads} threads");
            assert_eq!(report.bytes_walked, serial_report.bytes_walked, "{threads} threads");
            // Public roll-ups carry extension names even though the internal merge path
            // uses ids whose assignment order differs between serial and parallel walks.
            let (serial_total, parallel_total) = (serial.total(), parallel.total());
            assert_eq!(
                (
                    parallel_total.files,
                    parallel_total.dirs,
                    parallel_total.bytes,
                    parallel_total.allocated,
                    parallel_total.newest_mtime_ns,
                ),
                (
                    serial_total.files,
                    serial_total.dirs,
                    serial_total.bytes,
                    serial_total.allocated,
                    serial_total.newest_mtime_ns,
                ),
                "{threads} threads roll-up"
            );
            assert_eq!(
                parallel_total.by_ext, serial_total.by_ext,
                "{threads} threads per-extension roll-up"
            );
            assert_eq!(
                index_fingerprint(&parallel),
                index_fingerprint(&serial),
                "{threads} threads produced a different index"
            );
        }
    }

    #[test]
    fn parallel_walk_emits_every_entry_exactly_once() {
        let dir = branching_tree();
        let config = ScanConfig { threads: Some(4), batch_size: 16, ..ScanConfig::default() };
        let mut seen: BTreeMap<PathBuf, usize> = BTreeMap::new();
        let report = scan(dir.path(), &config, &mut |observation| {
            for op in &observation.ops {
                if let Op::Upsert { path, .. } = &op.op {
                    *seen.entry(path.clone()).or_default() += 1;
                }
            }
        })
        .expect("parallel scan");

        assert!(report.is_complete());
        assert_eq!(seen.len() as u64, report.entries, "entry count disagrees with the report");
        let duplicated: Vec<_> =
            seen.iter().filter(|(_path, count)| **count != 1).map(|(path, _)| path).collect();
        assert!(duplicated.is_empty(), "paths emitted more than once: {duplicated:?}");
    }

    #[test]
    fn parallel_walk_honours_max_depth() {
        let dir = branching_tree();
        for threads in [1_usize, 4] {
            let config =
                ScanConfig { threads: Some(threads), max_depth: Some(2), ..ScanConfig::default() };
            let (index, report) = scan_into_index(dir.path(), &config).expect("scan");
            assert!(report.is_complete());
            for (path, _kind, _attrs) in index_fingerprint(&index) {
                assert!(
                    path.components().count() <= 2,
                    "{threads} threads kept {path:?} past the depth limit"
                );
            }
        }
    }

    #[test]
    fn scan_order_never_changes_the_resulting_index() {
        let dir = branching_tree();
        let depth_first =
            ScanConfig { order: ScanOrder::DepthFirst, threads: Some(1), ..ScanConfig::default() };
        let (expected, expected_report) =
            scan_into_index(dir.path(), &depth_first).expect("depth-first scan");

        for (order, threads) in
            [(ScanOrder::BreadthFirst, 1), (ScanOrder::BreadthFirst, 4), (ScanOrder::DepthFirst, 4)]
        {
            let config = ScanConfig { order, threads: Some(threads), ..ScanConfig::default() };
            let (index, report) = scan_into_index(dir.path(), &config).expect("scan");
            assert_eq!(report.entries, expected_report.entries, "{order:?}/{threads}");
            assert_eq!(report.dirs_read, expected_report.dirs_read, "{order:?}/{threads}");
            // Public roll-ups resolve internal ids, so their named maps are stable even
            // when traversal order changes id assignment.
            let (totals, expected_totals) = (index.total(), expected.total());
            assert_eq!(
                (totals.files, totals.dirs, totals.bytes, totals.allocated),
                (
                    expected_totals.files,
                    expected_totals.dirs,
                    expected_totals.bytes,
                    expected_totals.allocated
                ),
                "{order:?}/{threads} roll-up"
            );
            assert_eq!(
                totals.newest_mtime_ns, expected_totals.newest_mtime_ns,
                "{order:?}/{threads} newest mtime"
            );
            assert_eq!(
                totals.by_ext, expected_totals.by_ext,
                "{order:?}/{threads} extension tallies"
            );
            assert_eq!(
                index_fingerprint(&index),
                index_fingerprint(&expected),
                "{order:?}/{threads} produced a different index"
            );
        }
    }

    #[test]
    fn a_single_worker_breadth_first_walk_is_strictly_level_ordered() {
        // The strict guarantee, which holds only with one worker. With several, the
        // queue is ordered but the claims are not: a fast worker can enqueue and claim
        // depth d+2 while a slow worker still holds depth d+1. See
        // `breadth_first_starts_every_top_level_subtree_early` for the property the
        // default configuration actually provides, which is the one consumers rely on.
        let dir = branching_tree();
        let config = ScanConfig {
            order: ScanOrder::BreadthFirst,
            threads: Some(1),
            batch_size: 1,
            ..ScanConfig::default()
        };
        let mut depths_in_order: Vec<usize> = Vec::new();
        scan(dir.path(), &config, &mut |observation| {
            for op in &observation.ops {
                if let Op::Upsert { path, kind, .. } = &op.op {
                    if kind.is_dir() {
                        depths_in_order.push(path.components().count());
                    }
                }
            }
        })
        .expect("scan");

        assert!(depths_in_order.len() > 10, "fixture should have many directories");
        assert!(
            depths_in_order.windows(2).all(|pair| pair[0] <= pair[1]),
            "directory depths were not non-decreasing: {depths_in_order:?}"
        );
    }

    /// How many of the fixture's twelve top-level subtrees have received any file by
    /// the time half the files have been emitted.
    ///
    /// This is the product metric — "is a mid-scan ranking meaningful?" — rather than
    /// first-touch, which cannot distinguish the orders at all: reading the root
    /// enumerates all twelve children at once either way. What a ranking needs is that
    /// the subtrees grow *together*.
    fn subtrees_started_at_halfway(order: ScanOrder, threads: usize, dir: &Path) -> usize {
        let config =
            ScanConfig { order, batch_size: 1, threads: Some(threads), ..ScanConfig::default() };

        let mut files: Vec<PathBuf> = Vec::new();
        scan(dir, &config, &mut |observation| {
            for op in &observation.ops {
                if let Op::Upsert { path, kind, .. } = &op.op {
                    if !kind.is_dir() {
                        files.push(path.clone());
                    }
                }
            }
        })
        .expect("scan");

        let halfway = files.len() / 2;
        let mut started: BTreeSet<PathBuf> = BTreeSet::new();
        for path in files.iter().take(halfway) {
            if let Some(top) = path.components().next() {
                started.insert(PathBuf::from(top.as_os_str()));
            }
        }
        started.len()
    }

    #[test]
    fn a_parallel_walk_accounts_for_where_its_time_went() {
        // The attribution identity: every named cause is a disjoint slice of worker
        // wall time, so the parts can never exceed the whole, and the counters that
        // amortization depends on are actually incremented. This is the instrument
        // the scheduler experiments will read; if it drifts, they measure noise.
        let dir = branching_tree();
        let config = ScanConfig { threads: Some(4), batch_size: 64, ..ScanConfig::default() };
        let report = scan(dir.path(), &config, &mut |_| {}).expect("scan");
        let a = report.attribution;

        assert!(a.claims > 0, "a parallel walk claims chunks: {a:?}");
        assert!(a.work_ns > 0, "reading directories takes time: {a:?}");
        assert!(a.wall_ns > 0);
        // claim() locks at least once per successful claim, and release() locks once
        // per claim cycle too.
        assert!(a.lock_ops >= a.claims * 2, "lock ops out of step with claims: {a:?}");
        assert!(
            a.accounted_ns() <= a.wall_ns,
            "attributed slices are disjoint intervals inside worker wall: {a:?}"
        );
    }

    #[test]
    fn a_serial_walk_has_no_coordination_to_attribute() {
        // Serial semantics: wall is the loop, "send" is the inline sink (the consumer
        // actually running), work is the rest — and the coordination counters stay
        // zero because there is no queue lock and no channel.
        let dir = branching_tree();
        let config = ScanConfig { threads: Some(1), batch_size: 64, ..ScanConfig::default() };
        let mut observations = 0usize;
        let report = scan(dir.path(), &config, &mut |_| observations += 1).expect("scan");
        let a = report.attribution;

        assert!(observations > 0, "the sink ran, so send_ns measured something real");
        assert!(a.work_ns > 0 && a.wall_ns >= a.work_ns);
        assert_eq!(
            (a.claims, a.lock_ops, a.lock_contended, a.starved_ns, a.lock_wait_ns),
            (0, 0, 0, 0, 0),
            "no queue, no lock, nothing to wait on: {a:?}"
        );
    }

    /// Twelve top-level subtrees, each a branching tree several levels deep.
    ///
    /// Branching matters: an earlier fixture gave every level exactly one child, which
    /// pinned the frontier at twelve directories and made both orders behave
    /// identically — a LIFO cannot dive when there is nothing to dive into. With two
    /// children per level, depth-first pushes siblings and immediately descends into
    /// the last one, which is the behaviour that leaves other subtrees behind.
    ///
    /// It is also deliberately uniform. A version using one deep spur beside shallow
    /// siblings made the result depend on whether `readdir` returned the spur early:
    /// it passed on APFS and failed on ext4.
    fn deep_forest() -> tempfile::TempDir {
        let dir = tempfile::tempdir().expect("tempdir");
        for top in 0..12 {
            let mut level: Vec<PathBuf> = vec![dir.path().join(format!("t{top}"))];
            for _ in 0..5 {
                let mut next = Vec::new();
                for parent in &level {
                    for child in 0..2 {
                        let path = parent.join(format!("c{child}"));
                        for file in 0..3 {
                            write_file(&path.join(format!("f{file}.dat")), b"xxxxxxxxxx");
                        }
                        next.push(path);
                    }
                }
                level = next;
            }
        }
        dir
    }

    /// Files accumulated by the *least advanced* top-level subtree in the first
    /// quarter of the walk.
    ///
    /// Counting subtrees merely *started* cannot discriminate on a tree whose root
    /// fans out twelve ways: every scheduler touches all twelve immediately, because
    /// reading the root enumerates them. What differs is whether they then advance
    /// together, so the question is how far behind the laggard is.
    fn leanest_subtree_early(order: ScanOrder, threads: usize, dir: &Path) -> usize {
        let config =
            ScanConfig { order, batch_size: 1, threads: Some(threads), ..ScanConfig::default() };
        let mut files: Vec<PathBuf> = Vec::new();
        scan(dir, &config, &mut |observation| {
            for op in &observation.ops {
                if let Op::Upsert { path, kind, .. } = &op.op {
                    if !kind.is_dir() {
                        files.push(path.clone());
                    }
                }
            }
        })
        .expect("scan");

        let quarter = files.len() / 4;
        let mut per_top: BTreeMap<PathBuf, usize> = BTreeMap::new();
        for path in files.iter().take(quarter) {
            if let Some(top) = path.components().next() {
                *per_top.entry(PathBuf::from(top.as_os_str())).or_default() += 1;
            }
        }
        (0..12)
            .map(|top| per_top.get(&PathBuf::from(format!("t{top}"))).copied().unwrap_or(0))
            .min()
            .unwrap_or(0)
    }

    #[test]
    fn deep_subtrees_do_not_delay_their_siblings() {
        // The orientation property, and the reason breadth-first is the default: when
        // every top-level subtree is deep, depth-first pours its early effort down
        // whichever ones it picked up and leaves the rest at zero, while the region
        // scheduler advances all twelve together. A user watching the top level fill
        // in sees a meaningful ranking in the first case and a misleading one in the
        // second.
        //
        // Asserted at one worker only, and that bound is deliberate. This metric reads
        // *emission* order, and under several workers emission reflects which worker
        // finished first as much as which region was claimed — so it varies with core
        // count. Measured on a six-core machine the margin is wide (33-37 files against
        // 6); on a CI runner with fewer cores both orders can report zero. That makes it
        // a benchmark-grade observation, recorded in exp-013, not a unit-test assertion.
        //
        // The scheduling property itself *is* asserted deterministically, against the
        // queue rather than through a walk, by
        // `the_region_scheduler_spreads_workers_over_distinct_subtrees`.
        let dir = deep_forest();
        let breadth = leanest_subtree_early(ScanOrder::BreadthFirst, 1, dir.path());
        let depth = leanest_subtree_early(ScanOrder::DepthFirst, 1, dir.path());
        assert!(
            breadth > depth,
            "breadth-first should leave its least advanced top-level subtree further \
             along: {breadth} files against {depth}"
        );
    }

    #[test]
    fn the_region_scheduler_spreads_workers_over_distinct_subtrees() {
        // The scheduler invariant, checked directly on the queue rather than through a
        // walk: consecutive claims by *different* workers must land in different
        // regions while several regions have work. This is what the round-robin ready
        // ring buys, and it is the thing a global FIFO could not promise.
        let queue = DirectoryQueue::new((PathBuf::new(), 0), ScanOrder::BreadthFirst, None, None);
        let mut timing = WalkAttribution::default();

        // Bootstrap: drain the root, then seed four top-level regions.
        let mut claimed = Vec::new();
        let root = queue.claim(&mut claimed, &mut timing).expect("root is claimable");
        claimed.clear();
        queue.extend(
            (0..4).map(|top| (PathBuf::from(format!("t{top}")), 1, RegionId::UNASSIGNED)),
            &mut timing,
        );
        assert!(root.release(0, 0, &mut timing).is_none());

        // Four workers with no affinity must each be handed a different region. The
        // claims are held for the whole loop, as four concurrent workers would hold
        // them, because releasing between them would let one worker take every region.
        let mut regions = BTreeSet::new();
        let mut held = Vec::new();
        for _ in 0..4 {
            let mut claimed = Vec::new();
            held.push(queue.claim(&mut claimed, &mut timing).expect("a region has work"));
            regions.insert(claimed[0].2.0);
            assert_eq!(claimed.len(), 1, "one directory per region so far");
        }
        assert_eq!(regions.len(), 4, "each claim took a distinct region: {regions:?}");
    }

    #[test]
    fn breadth_first_spreads_early_work_across_top_level_subtrees() {
        // The justification for making breadth-first the default: at the halfway point
        // more of the tree's top-level subtrees have started filling, so a consumer
        // ranking by size mid-scan is comparing partial values rather than a mix of
        // final values and zeros.
        //
        // Pinned with one worker, where the ordering guarantee is strict and the result
        // is deterministic. The multi-worker case is deliberately NOT asserted here:
        // measured on this fixture the advantage disappears under the default worker
        // count (both orders start 7-8 subtrees, run to run), because emission order is
        // then dominated by worker scheduling rather than by queue order. That is a
        // real limitation of the current design, recorded in the plan and tracked
        // rather than papered over with a test tuned until it passed.
        let dir = branching_tree();
        let breadth = subtrees_started_at_halfway(ScanOrder::BreadthFirst, 1, dir.path());
        let depth = subtrees_started_at_halfway(ScanOrder::DepthFirst, 1, dir.path());

        assert!(
            breadth > depth,
            "breadth-first should have more top-level subtrees underway at the halfway \
             point, but started {breadth} against depth-first's {depth}"
        );
    }

    #[test]
    fn scan_order_does_not_change_the_cache_scope() {
        // Order is operational, like the worker count: it changes when observations
        // appear, never which ones, so it must not be able to invalidate a snapshot.
        let breadth = ScanConfig { order: ScanOrder::BreadthFirst, ..ScanConfig::default() };
        let depth = ScanConfig { order: ScanOrder::DepthFirst, ..ScanConfig::default() };
        assert_eq!(breadth.scope(), depth.scope());
    }

    #[test]
    fn worker_threads_are_bounded_and_never_zero() {
        let zero = ScanConfig { threads: Some(0), ..ScanConfig::default() };
        assert_eq!(zero.worker_threads(), 1, "zero threads must fall back to the serial walk");
        let absurd = ScanConfig { threads: Some(usize::MAX), ..ScanConfig::default() };
        assert_eq!(absurd.worker_threads(), MAX_SCAN_THREADS);
        // The automatic choice is capped well below what a caller may request, because
        // the measured knee is far below the core count on a large machine.
        let automatic = ScanConfig { threads: None, ..ScanConfig::default() };
        assert!((1..=DEFAULT_SCAN_THREADS_CAP).contains(&automatic.worker_threads()));
    }

    #[test]
    fn automatic_worker_pool_keeps_a_conservative_start_and_bounded_reserve() {
        assert_eq!(automatic_worker_pool(1), WorkerPool::fixed(1));
        assert_eq!(
            automatic_worker_pool(4),
            WorkerPool {
                initial: 4,
                maximum: 8,
                calibration: Some(WorkerCalibration::new(
                    ADAPTIVE_SCAN_CALIBRATION_ENTRIES,
                    ADAPTIVE_SCAN_SLOW_WORK_NS_PER_ENTRY,
                )),
            }
        );
        assert_eq!(
            automatic_worker_pool(10),
            WorkerPool {
                initial: DEFAULT_SCAN_THREADS_CAP,
                maximum: ADAPTIVE_SCAN_THREADS_CAP,
                calibration: Some(WorkerCalibration::new(
                    ADAPTIVE_SCAN_CALIBRATION_ENTRIES,
                    ADAPTIVE_SCAN_SLOW_WORK_NS_PER_ENTRY,
                )),
            }
        );
    }

    #[test]
    fn an_abandoned_claim_does_not_strand_the_other_workers() {
        // The liveness property behind `DirectoryClaim`. A worker that stops mid-chunk
        // — consumer gone, or a panic unwinding through the directory read — still owes
        // the queue its claim, and `claim` parks everyone else until `outstanding`
        // reaches zero. Before the claim was an RAII guard both of those exits skipped
        // the release, and every remaining worker waited on the condvar forever while
        // the scoped join waited on them.
        let queue = std::sync::Arc::new(DirectoryQueue::new(
            (PathBuf::new(), 0),
            ScanOrder::BreadthFirst,
            None,
            None,
        ));
        let mut timing = WalkAttribution::default();
        let mut claimed = Vec::new();

        // One worker takes the root and abandons it without publishing anything.
        drop(queue.claim(&mut claimed, &mut timing).expect("root is claimable"));

        // A second worker must now be told the walk is over rather than parking.
        let waiter = queue.clone();
        let (done, finished) = std::sync::mpsc::sync_channel(1);
        std::thread::spawn(move || {
            let mut timing = WalkAttribution::default();
            let mut claimed = Vec::new();
            let outcome = waiter.claim(&mut claimed, &mut timing).is_some();
            done.send(outcome).expect("publish the claim outcome");
        });

        assert_eq!(
            finished.recv_timeout(std::time::Duration::from_secs(5)),
            Ok(false),
            "the queue must report the walk finished instead of parking the worker"
        );
    }

    #[test]
    fn automatic_queue_activates_its_reserve_only_for_slow_initial_work() {
        let slow = WorkerCalibration::new(3, 10);
        let queue =
            DirectoryQueue::new((PathBuf::new(), 0), ScanOrder::BreadthFirst, Some(slow), None);
        let mut timing = WalkAttribution::default();
        let mut claimed = Vec::new();

        let claim = queue.claim(&mut claimed, &mut timing).expect("root is claimable");
        queue.extend([(PathBuf::from("child"), 1, RegionId::UNASSIGNED)].into_iter(), &mut timing);
        assert!(claim.release(2, 20, &mut timing).is_none());

        claimed.clear();
        let claim = queue.claim(&mut claimed, &mut timing).expect("child is claimable");
        queue.extend([(PathBuf::from("grandchild"), 2, claimed[0].2)].into_iter(), &mut timing);
        assert_eq!(claim.release(1, 10, &mut timing), Some(2));

        let fast = WorkerCalibration::new(3, 11);
        let queue =
            DirectoryQueue::new((PathBuf::new(), 0), ScanOrder::BreadthFirst, Some(fast), None);
        let mut timing = WalkAttribution::default();
        let mut claimed = Vec::new();
        let claim = queue.claim(&mut claimed, &mut timing).expect("root is claimable");
        assert!(claim.release(3, 30, &mut timing).is_none());
        assert!(queue.lock().controller.is_none(), "calibration decides only once");
    }

    #[test]
    fn repeated_windows_reconsider_a_late_slow_phase_in_entry_order() {
        let calibration = WorkerCalibration::new(4, 10);
        let mut controller =
            WorkerController::new(calibration, WorkerPolicyExperiment::RepeatedWindows);

        let fast = controller.observe(4, 20).expect("first complete window");
        let slow = controller.observe(4, 80).expect("second complete window");

        assert!(!fast.slow);
        assert!(slow.slow);
        assert_eq!((fast.start_entry_ordinal, fast.end_entry_ordinal), (0, 4));
        assert_eq!((slow.start_entry_ordinal, slow.end_entry_ordinal), (4, 8));
    }

    #[test]
    fn shipped_trace_retains_post_decision_windows_without_changing_policy() {
        let calibration = WorkerCalibration::new(2, 10);
        let pool = WorkerPool { initial: 2, maximum: 4, calibration: Some(calibration) };
        let recorder =
            ScanDiagnosticsRecorder::new(pool, 2, WorkerPolicyExperiment::ShippedOneShot);
        let queue = DirectoryQueue::new_with_policy(
            (PathBuf::new(), 0),
            ScanOrder::BreadthFirst,
            Some(calibration),
            Some(recorder.clone()),
            pool.initial,
            pool.maximum,
            WorkerPolicyExperiment::ShippedOneShot,
        );
        let mut timing = WalkAttribution::default();
        let mut claimed = Vec::new();

        let first = queue.claim(&mut claimed, &mut timing).expect("first window");
        queue.extend([(PathBuf::from("late"), 1, RegionId::UNASSIGNED)].into_iter(), &mut timing);
        assert_eq!(first.release(2, 10, &mut timing), None, "fast prefix holds");

        claimed.clear();
        let late = queue.claim(&mut claimed, &mut timing).expect("late phase");
        queue.extend([(PathBuf::from("tail"), 2, RegionId::UNASSIGNED)].into_iter(), &mut timing);
        assert_eq!(late.release(2, 40, &mut timing), None, "shadow cannot scale");

        let diagnostics = recorder.finish();
        assert_eq!(diagnostics.worker_policy.outcome, WorkerPolicyOutcome::Held);
        assert_eq!(
            diagnostics
                .worker_policy
                .windows
                .iter()
                .map(|window| window.decision)
                .collect::<Vec<_>>(),
            vec![WorkerPolicyDecision::Hold, WorkerPolicyDecision::ObserveSlow]
        );
    }

    #[test]
    fn staged_controller_requires_a_useful_frontier_then_stays_bounded() {
        let calibration = WorkerCalibration::new(1, 10);
        let pool = WorkerPool { initial: 2, maximum: 8, calibration: Some(calibration) };
        let recorder =
            ScanDiagnosticsRecorder::new(pool, 4, WorkerPolicyExperiment::StagedGatedWindows);
        let queue = DirectoryQueue::new_with_policy(
            (PathBuf::new(), 0),
            ScanOrder::BreadthFirst,
            Some(calibration),
            Some(recorder.clone()),
            pool.initial,
            pool.maximum,
            WorkerPolicyExperiment::StagedGatedWindows,
        );
        let mut timing = WalkAttribution::default();
        let mut claimed = Vec::new();

        let root = queue.claim(&mut claimed, &mut timing).expect("root");
        queue.extend([(PathBuf::from("narrow"), 1, RegionId::UNASSIGNED)].into_iter(), &mut timing);
        assert_eq!(root.release(1, 20, &mut timing), None);

        claimed.clear();
        let narrow = queue.claim(&mut claimed, &mut timing).expect("narrow child");
        queue.extend(
            (0..9).map(|index| (PathBuf::from(format!("wide-{index}")), 2, RegionId::UNASSIGNED)),
            &mut timing,
        );
        assert_eq!(narrow.release(1, 20, &mut timing), Some(4));

        claimed.clear();
        let wide = queue.claim(&mut claimed, &mut timing).expect("wide claim");
        queue.extend(
            (0..9).map(|index| (PathBuf::from(format!("wider-{index}")), 3, RegionId::UNASSIGNED)),
            &mut timing,
        );
        assert_eq!(wide.release(1, 20, &mut timing), Some(8));

        let diagnostics = recorder.finish();
        let decisions: Vec<_> = diagnostics
            .worker_policy
            .windows
            .iter()
            .map(|window| (window.decision, window.requested_workers))
            .collect();
        assert_eq!(
            decisions,
            vec![
                (WorkerPolicyDecision::HoldInsufficientFrontier, None),
                (WorkerPolicyDecision::ScaleUp, Some(4)),
                (WorkerPolicyDecision::ScaleUp, Some(8)),
            ]
        );
        assert!(
            diagnostics
                .worker_policy
                .windows
                .windows(2)
                .all(|pair| pair[0].end_entry_ordinal <= pair[1].start_entry_ordinal)
        );
        assert!(diagnostics.worker_policy.windows.iter().all(|window| {
            window.requested_workers.is_none_or(|workers| workers <= pool.maximum)
        }));
    }

    #[test]
    fn staged_controller_does_not_add_producers_to_a_delayed_handoff() {
        let calibration = WorkerCalibration::new(1, 10);
        let pool = WorkerPool { initial: 2, maximum: 8, calibration: Some(calibration) };
        let recorder =
            ScanDiagnosticsRecorder::new(pool, 4, WorkerPolicyExperiment::StagedGatedWindows);
        recorder.handoff_sent();
        recorder.handoff_sent();
        let queue = DirectoryQueue::new_with_policy(
            (PathBuf::new(), 0),
            ScanOrder::BreadthFirst,
            Some(calibration),
            Some(recorder.clone()),
            pool.initial,
            pool.maximum,
            WorkerPolicyExperiment::StagedGatedWindows,
        );
        let mut timing = WalkAttribution::default();
        let mut claimed = Vec::new();
        let claim = queue.claim(&mut claimed, &mut timing).expect("root");
        queue.extend(
            (0..9).map(|index| (PathBuf::from(format!("ready-{index}")), 1, RegionId::UNASSIGNED)),
            &mut timing,
        );

        assert_eq!(claim.release(1, 20, &mut timing), None);
        recorder.handoff_received();
        recorder.handoff_received();
        let diagnostics = recorder.finish();
        assert_eq!(
            diagnostics.worker_policy.windows[0].decision,
            WorkerPolicyDecision::HoldHandoffBacklog
        );
    }

    #[test]
    fn candidate_retains_post_expansion_shadow_history() {
        let calibration = WorkerCalibration::new(1, 10);
        let pool = WorkerPool { initial: 2, maximum: 4, calibration: Some(calibration) };
        let recorder =
            ScanDiagnosticsRecorder::new(pool, 4, WorkerPolicyExperiment::RepeatedWindows);
        let queue = DirectoryQueue::new_with_policy(
            (PathBuf::new(), 0),
            ScanOrder::BreadthFirst,
            Some(calibration),
            Some(recorder.clone()),
            pool.initial,
            pool.maximum,
            WorkerPolicyExperiment::RepeatedWindows,
        );
        let mut timing = WalkAttribution::default();
        let mut claimed = Vec::new();

        let slow = queue.claim(&mut claimed, &mut timing).expect("slow prefix");
        queue.extend(
            (0..4).map(|index| (PathBuf::from(format!("fast-{index}")), 1, RegionId::UNASSIGNED)),
            &mut timing,
        );
        assert_eq!(slow.release(1, 20, &mut timing), Some(4));

        claimed.clear();
        let fast = queue.claim(&mut claimed, &mut timing).expect("fast suffix");
        queue.extend([(PathBuf::from("tail"), 2, RegionId::UNASSIGNED)].into_iter(), &mut timing);
        assert_eq!(fast.release(1, 1, &mut timing), None);

        let diagnostics = recorder.finish();
        assert_eq!(
            diagnostics
                .worker_policy
                .windows
                .iter()
                .map(|window| window.decision)
                .collect::<Vec<_>>(),
            vec![WorkerPolicyDecision::ScaleUp, WorkerPolicyDecision::ObserveFast]
        );
    }

    #[test]
    fn every_experimental_controller_preserves_exactness_and_shutdown() {
        let dir = branching_tree();
        let serial = ScanConfig { threads: Some(1), ..ScanConfig::default() };
        let (reference, _) = scan_into_index(dir.path(), &serial).expect("serial reference");
        let automatic = ScanConfig { threads: None, ..ScanConfig::default() };

        for policy in [
            WorkerPolicyExperiment::ShippedOneShot,
            WorkerPolicyExperiment::RepeatedWindows,
            WorkerPolicyExperiment::StagedGatedWindows,
        ] {
            let (index, report, diagnostics) =
                scan_into_index_with_policy_diagnostics(dir.path(), &automatic, policy)
                    .expect("candidate scan finishes");
            assert!(report.is_complete(), "{policy:?}: {:?}", report.errors);
            assert_eq!(index_fingerprint(&reference), index_fingerprint(&index), "{policy:?}");
            assert_eq!(diagnostics.worker_policy.ready_directories_at_finish, 0);
            assert_eq!(diagnostics.worker_policy.in_flight_directories_at_finish, 0);
            assert_eq!(diagnostics.worker_policy.handoff_backlog_at_finish, 0);
            assert!(
                diagnostics.worker_policy.workers_spawned
                    <= diagnostics.worker_policy.maximum_workers
            );
        }
    }

    /// A deterministic model of the automatic worker policy under *completion* order.
    ///
    /// The scaling decision is driven by chunk releases, and chunks complete in whatever
    /// order the filesystem and the workers produce them — not in traversal order. On a
    /// homogeneous tree that distinction is invisible, because every prefix looks like
    /// every other. On a heterogeneous one it decides the answer.
    ///
    /// These tests exist because the alternative is a stopwatch on a real tree, which
    /// measures one host on one day and cannot separate a policy defect from ambient
    /// noise. Replaying an explicit completion order through the shipped calibration
    /// isolates the policy exactly, and does so identically on every platform.
    ///
    /// They characterize behavior; they do not endorse a replacement. Which controller
    /// is *faster* is a question only the held-out Apple Silicon/APFS matrix can answer.
    mod completion_order {
        use super::{ADAPTIVE_SCAN_SLOW_WORK_NS_PER_ENTRY, WorkerCalibration};

        /// One chunk release: entries observed and worker time spent observing them.
        #[derive(Clone, Copy, Debug)]
        struct Chunk {
            entries: u64,
            work_ns: u64,
        }

        impl Chunk {
            /// A run of `entries` entries costing `per_entry_ns` each.
            const fn at(entries: u64, per_entry_ns: u64) -> Self {
                Self { entries, work_ns: entries.saturating_mul(per_entry_ns) }
            }
        }

        /// What a policy concluded over one completion order.
        #[derive(Debug, PartialEq, Eq)]
        enum Outcome {
            /// The policy found the filesystem slow and expanded the reserve.
            ScaledUp { after_chunks: usize },
            /// The policy found the filesystem fast and held the initial pool.
            Held { after_chunks: usize },
            /// The walk ended before the policy observed enough to conclude anything.
            ///
            /// Distinct from [`Outcome::Held`] on purpose: nothing was measured, so a
            /// held pool here is an absence of evidence rather than a decision.
            Undecided,
        }

        /// Mean cost per entry over a whole trace, which is what the threshold *means*.
        fn whole_trace_ns_per_entry(trace: &[Chunk]) -> u64 {
            let entries: u64 = trace.iter().map(|chunk| chunk.entries).sum();
            let work_ns: u64 = trace.iter().map(|chunk| chunk.work_ns).sum();
            assert!(entries > 0, "a trace must observe entries");
            work_ns / entries
        }

        /// Replay a completion order through the *shipped* calibration.
        ///
        /// This drives [`WorkerCalibration::observe`] itself rather than restating its
        /// arithmetic, so the model cannot quietly drift from the policy it is evidence
        /// about. The loop mirrors `DirectoryQueue::release`: fold each chunk in, and
        /// stop at the first one that produces a verdict.
        fn shipped(window: u64, threshold_ns: u64, trace: &[Chunk]) -> Outcome {
            let mut calibration = WorkerCalibration::new(window, threshold_ns);
            for (index, chunk) in trace.iter().enumerate() {
                if let Some(slow) = calibration.observe(chunk.entries, chunk.work_ns) {
                    let after_chunks = index + 1;
                    return if slow {
                        Outcome::ScaledUp { after_chunks }
                    } else {
                        Outcome::Held { after_chunks }
                    };
                }
            }
            Outcome::Undecided
        }

        /// Entries per chunk in the traces below. Four fill the 16,384-entry window.
        const CHUNK: u64 = 4_096;
        /// A shallow, cache-warm phase: metadata already resident.
        const FAST: Chunk = Chunk::at(CHUNK, 2_000);
        /// A deep, cold phase: the latency-bound regime the reserve exists to hide.
        const SLOW: Chunk = Chunk::at(CHUNK, 90_000);

        #[test]
        fn completion_order_alone_flips_the_shipped_decision() {
            // The defect, stated as an experiment: hold the *tree* constant and vary
            // only the order its chunks complete in. Both traces contain the same four
            // fast and four slow chunks, so they describe the same filesystem work.
            let fast_phase_first = [FAST, FAST, FAST, FAST, SLOW, SLOW, SLOW, SLOW];
            let interleaved = [SLOW, FAST, SLOW, FAST, SLOW, FAST, SLOW, FAST];

            let window = 4 * CHUNK;
            let threshold = ADAPTIVE_SCAN_SLOW_WORK_NS_PER_ENTRY;

            // Whole-walk truth is identical, and by the policy's own threshold both
            // walks are latency-bound: 46 µs per entry against a 30 µs trigger.
            let truth = whole_trace_ns_per_entry(&fast_phase_first);
            assert_eq!(truth, whole_trace_ns_per_entry(&interleaved));
            assert!(
                truth >= threshold,
                "both traces are slow walks by the shipped threshold: {truth} < {threshold}"
            );

            // Yet the decision depends entirely on which chunks happened to finish
            // first. One walk hides latency; the other runs the whole slow phase on the
            // starting pool, having concluded from an unrepresentative prefix.
            assert_eq!(
                shipped(window, threshold, &fast_phase_first),
                Outcome::Held { after_chunks: 4 },
                "a fast prefix holds the pool for a walk that is slow overall"
            );
            assert_eq!(
                shipped(window, threshold, &interleaved),
                Outcome::ScaledUp { after_chunks: 4 },
                "the same tree scales up when its slow chunks land in the window"
            );
        }

        #[test]
        fn a_slow_phase_after_the_window_is_never_reconsidered() {
            // The heterogeneous-tree case from the field report. A small fast region
            // fills the window, and everything after it is slow — but the calibration
            // is already gone, so no amount of later evidence can reopen the decision.
            let mut trace = vec![FAST; 4];
            trace.extend(std::iter::repeat_n(SLOW, 400));

            let observed = whole_trace_ns_per_entry(&trace);
            assert!(
                observed >= 2 * ADAPTIVE_SCAN_SLOW_WORK_NS_PER_ENTRY,
                "the walk is overwhelmingly latency-bound: {observed} ns per entry"
            );

            assert_eq!(
                shipped(4 * CHUNK, ADAPTIVE_SCAN_SLOW_WORK_NS_PER_ENTRY, &trace),
                Outcome::Held { after_chunks: 4 },
                "1% of the walk decided the worker policy for the other 99%"
            );
        }

        #[test]
        fn slow_in_flight_work_is_censored_by_fast_completions() {
            // Four slow chunks have already been claimed, but their filesystem calls
            // remain in flight while four cache-warm chunks complete. Completion-order
            // calibration cannot see owed work: the fast completions close the window
            // and permanently hold before any slow claim returns.
            let completed_before_slow_returns = [FAST, FAST, FAST, FAST];
            assert_eq!(
                shipped(
                    4 * CHUNK,
                    ADAPTIVE_SCAN_SLOW_WORK_NS_PER_ENTRY,
                    &completed_before_slow_returns,
                ),
                Outcome::Held { after_chunks: 4 }
            );

            let mut eventual_completions = completed_before_slow_returns.to_vec();
            eventual_completions.extend([SLOW, SLOW, SLOW, SLOW]);
            assert!(
                whole_trace_ns_per_entry(&eventual_completions)
                    >= ADAPTIVE_SCAN_SLOW_WORK_NS_PER_ENTRY
            );
            assert_eq!(
                shipped(4 * CHUNK, ADAPTIVE_SCAN_SLOW_WORK_NS_PER_ENTRY, &eventual_completions,),
                Outcome::Held { after_chunks: 4 }
            );
        }

        #[test]
        fn a_slow_prefix_can_scale_a_walk_that_is_fast_overall() {
            // The mirror-image error. A one-way expansion reacts correctly to the
            // prefix by its local threshold, but the prefix is under 1% of this walk
            // and the whole trace is firmly in the fast regime. A repeated trigger
            // alone cannot undo an expansion; staged growth limits exposure but does
            // not make reversible parking unnecessary.
            let mut trace = vec![SLOW; 4];
            trace.extend(std::iter::repeat_n(FAST, 400));
            let observed = whole_trace_ns_per_entry(&trace);
            assert!(observed < ADAPTIVE_SCAN_SLOW_WORK_NS_PER_ENTRY);
            assert_eq!(
                shipped(4 * CHUNK, ADAPTIVE_SCAN_SLOW_WORK_NS_PER_ENTRY, &trace),
                Outcome::ScaledUp { after_chunks: 4 }
            );
            assert_eq!(
                sliding(4 * CHUNK, ADAPTIVE_SCAN_SLOW_WORK_NS_PER_ENTRY, &trace),
                Outcome::ScaledUp { after_chunks: 4 }
            );
        }

        #[test]
        fn a_walk_shorter_than_the_window_decides_nothing() {
            // Fails closed rather than reporting a held pool: a walk this short never
            // observed enough to have an opinion, and an artifact that recorded `Held`
            // would claim a measurement that was never taken.
            let trace = [FAST, SLOW];
            assert_eq!(
                shipped(4 * CHUNK, ADAPTIVE_SCAN_SLOW_WORK_NS_PER_ENTRY, &trace),
                Outcome::Undecided
            );
        }

        /// A screening-only candidate: a window that slides instead of closing once.
        ///
        /// Present as *evidence about a design*, not as a proposed change. It keeps the
        /// shipped trigger and pool bounds and alters only when the question is asked,
        /// which is the narrowest edit that could address the order sensitivity above.
        /// Whether it is faster on a real tree is unmeasured here and unmeasurable in a
        /// virtualized non-APFS environment; selecting it would need the held-out Apple
        /// Silicon matrix that this workstream has not yet been able to run.
        struct SlidingWindow {
            window_entries: u64,
            threshold_ns: u64,
            recent: std::collections::VecDeque<Chunk>,
            entries: u64,
            work_ns: u64,
        }

        impl SlidingWindow {
            fn new(window_entries: u64, threshold_ns: u64) -> Self {
                Self {
                    window_entries,
                    threshold_ns,
                    recent: std::collections::VecDeque::new(),
                    entries: 0,
                    work_ns: 0,
                }
            }

            /// Fold in a chunk and re-ask the question over the trailing window.
            fn observe(&mut self, chunk: Chunk) -> Option<bool> {
                self.recent.push_back(chunk);
                self.entries = self.entries.saturating_add(chunk.entries);
                self.work_ns = self.work_ns.saturating_add(chunk.work_ns);

                // Drop from the front while the window stays full without the oldest
                // chunk, so the answer describes recent work rather than the whole walk.
                while let Some(oldest) = self.recent.front().copied() {
                    if self.entries - oldest.entries < self.window_entries {
                        break;
                    }
                    self.recent.pop_front();
                    self.entries -= oldest.entries;
                    self.work_ns -= oldest.work_ns;
                }

                (self.entries >= self.window_entries)
                    .then(|| self.work_ns / self.entries >= self.threshold_ns)
            }
        }

        /// Replay a completion order through the candidate, stopping at its first
        /// scale-up. The shipped pool only grows, so a later verdict cannot undo one.
        fn sliding(window: u64, threshold_ns: u64, trace: &[Chunk]) -> Outcome {
            let mut policy = SlidingWindow::new(window, threshold_ns);
            let mut decided = None;
            for (index, chunk) in trace.iter().enumerate() {
                if let Some(slow) = policy.observe(*chunk) {
                    let after_chunks = index + 1;
                    if slow {
                        return Outcome::ScaledUp { after_chunks };
                    }
                    decided.get_or_insert(Outcome::Held { after_chunks });
                }
            }
            decided.unwrap_or(Outcome::Undecided)
        }

        #[test]
        fn screening_a_sliding_window_against_the_order_sensitivity() {
            let window = 4 * CHUNK;
            let threshold = ADAPTIVE_SCAN_SLOW_WORK_NS_PER_ENTRY;

            // The pair that splits the shipped policy reaches one answer here, and it
            // is the answer the whole-trace mean supports in both orders.
            let fast_phase_first = [FAST, FAST, FAST, FAST, SLOW, SLOW, SLOW, SLOW];
            let interleaved = [SLOW, FAST, SLOW, FAST, SLOW, FAST, SLOW, FAST];
            // Both reach the same verdict; they differ only in how long the fast prefix
            // delays it, which is the behavior a trailing window is supposed to have.
            assert_eq!(
                sliding(window, threshold, &fast_phase_first),
                Outcome::ScaledUp { after_chunks: 6 }
            );
            assert_eq!(
                sliding(window, threshold, &interleaved),
                Outcome::ScaledUp { after_chunks: 4 }
            );

            // And the late slow phase is reached rather than missed: two slow chunks
            // after the window closes are enough to pull the trailing mean over.
            let mut late = vec![FAST; 4];
            late.extend(std::iter::repeat_n(SLOW, 400));
            assert_eq!(sliding(window, threshold, &late), Outcome::ScaledUp { after_chunks: 6 });

            // A genuinely fast tree must still hold the pool: the candidate has to keep
            // the property the shipped policy gets right, or it is not a candidate.
            let uniformly_fast = vec![FAST; 40];
            assert_eq!(
                sliding(window, threshold, &uniformly_fast),
                Outcome::Held { after_chunks: 4 }
            );

            // A short walk still decides nothing, for the same reason as above.
            assert_eq!(sliding(window, threshold, &[FAST, SLOW]), Outcome::Undecided);
        }
    }

    #[test]
    fn thread_count_does_not_change_the_cache_scope() {
        // Threads are an operational choice. If they leaked into the scope, changing
        // the pool size would invalidate every snapshot on disk.
        let serial = ScanConfig { threads: Some(1), ..ScanConfig::default() };
        let parallel = ScanConfig { threads: Some(8), ..ScanConfig::default() };
        assert_eq!(serial.scope(), parallel.scope());
    }

    #[test]
    fn scan_populates_an_index_end_to_end() {
        let dir = sample_tree();
        let (index, report) = scan_into_index(dir.path(), &ScanConfig::default()).expect("scan");

        assert!(report.is_complete(), "unexpected errors: {:?}", report.errors);
        let total = index.total();
        assert_eq!(total.files, 3);
        assert_eq!(total.dirs, 2);
        assert_eq!(total.bytes, 5 + 12 + 9);
        assert_eq!(total.by_ext[".rs"].files, 2);
        assert_eq!(total.by_ext[".txt"].files, 1);

        let src = index.rollup(Path::new("src")).expect("src");
        assert_eq!(src.files, 2);
        assert_eq!(src.dirs, 1);
    }

    #[test]
    fn cold_scan_routes_control_sources_through_both_walkers() {
        let dir = tempfile::tempdir().expect("tempdir");
        write_file(&dir.path().join(".gitignore"), b"*.log\n");
        write_file(&dir.path().join("debug.log"), b"ignored");
        write_file(&dir.path().join("keep.rs"), b"visible");

        for threads in [1, 4] {
            let config =
                ScanConfig { read_controls: true, threads: Some(threads), ..ScanConfig::default() };
            let (index, report) = scan_into_index(dir.path(), &config).expect("scan");

            assert!(report.is_complete(), "unexpected errors: {:?}", report.errors);
            assert!(
                index
                    .controls()
                    .expect("control state observed")
                    .source_is(Path::new(".gitignore"), b"*.log\n")
            );
            assert_eq!(
                index.is_ignored(Path::new("debug.log")).expect("control state observed"),
                Some(true)
            );
            assert_eq!(
                index.is_ignored(Path::new("keep.rs")).expect("control state observed"),
                Some(false)
            );
            let partitions = index.partition_total().expect("control state observed");
            assert_eq!(partitions.all.files, 3);
            assert_eq!(partitions.unignored.files, 2);
        }
    }

    #[cfg(unix)]
    #[test]
    fn raced_fifo_control_source_is_rejected_without_blocking() {
        let dir = tempfile::tempdir().expect("tempdir");
        let control = dir.path().join(".gitignore");
        let status = match std::process::Command::new("mkfifo").arg(&control).status() {
            Ok(status) => status,
            Err(error) if error.kind() == std::io::ErrorKind::NotFound => return,
            Err(error) => panic!("create fifo: {error}"),
        };
        assert!(status.success(), "mkfifo exited with {status}");

        let root = dir.path().to_path_buf();
        let (sender, receiver) = std::sync::mpsc::channel();
        std::thread::spawn(move || {
            let result = read_control_op_unconditional(
                &root,
                Path::new(".gitignore"),
                EntryKind::File,
                Some(crate::control::DEFAULT_CONTROL_BUDGET),
            );
            sender.send(result).ok();
        });
        let result = receiver
            .recv_timeout(std::time::Duration::from_secs(1))
            .expect("a raced FIFO must not block the scan worker")
            .expect("the non-regular replacement is a normal control removal");

        assert!(matches!(result, Some(Op::ControlRemove { .. })));
    }

    #[test]
    fn hidden_admission_keeps_exact_allowlist_and_control_signals_only() {
        let dir = tempfile::tempdir().expect("tempdir");
        write_file(&dir.path().join(".gitignore"), b"*.log\n");
        write_file(&dir.path().join("debug.log"), b"ignored");
        write_file(&dir.path().join(".secret/token"), b"hidden");
        write_file(&dir.path().join(".github/workflows/check.yml"), b"visible");
        let hidden = std::sync::Arc::new(crate::HiddenPolicy::prune_hidden([".github"]));

        for threads in [1, 4] {
            let config = ScanConfig {
                hidden: Some(std::sync::Arc::clone(&hidden)),
                threads: Some(threads),
                read_controls: true,
                ..ScanConfig::default()
            };
            let (mut index, report) = scan_into_index(dir.path(), &config).expect("scan");

            assert!(report.is_complete(), "unexpected errors: {:?}", report.errors);
            assert!(index.lookup(Path::new(".gitignore")).is_none());
            assert!(index.lookup(Path::new(".secret")).is_none());
            assert!(index.lookup(Path::new(".secret/token")).is_none());
            assert!(index.lookup(Path::new(".github/workflows/check.yml")).is_some());
            assert!(
                index
                    .controls()
                    .expect("control state observed")
                    .source_is(Path::new(".gitignore"), b"*.log\n")
            );
            assert_eq!(
                index.is_ignored(Path::new("debug.log")).expect("control state observed"),
                Some(true)
            );

            fs::remove_file(dir.path().join(".gitignore")).expect("remove control");
            if threads > 1 {
                fs::create_dir(dir.path().join(".gitignore")).expect("replace with directory");
            }
            let reconciled = reconcile(&mut index, &config, &mut |_| {}).expect("reconcile");
            assert!(reconciled.is_complete());
            assert!(index.controls().expect("control state observed").is_empty());
            if threads > 1 {
                fs::remove_dir(dir.path().join(".gitignore")).expect("remove directory");
            }
            write_file(&dir.path().join(".gitignore"), b"*.log\n");
        }
    }

    #[test]
    fn excluded_ignored_directory_is_not_enumerated_and_rule_edits_reconcile_it() {
        let root = tempfile::tempdir().expect("root");
        write_file(&root.path().join(".gitignore"), b"target/\n");
        write_file(&root.path().join("target/deep/file.rs"), b"code");
        write_file(&root.path().join("keep.rs"), b"kept");
        let config = ScanConfig {
            population: crate::query::IgnoredEntries::Exclude,
            threads: Some(4),
            ..ScanConfig::default()
        };

        let (mut index, cold) = scan_into_index(root.path(), &config).expect("cold scan");
        assert!(cold.is_complete(), "{:?}", cold.errors);
        assert_eq!(cold.dirs_read, 1, "ignored target was not opened");
        assert!(index.lookup(Path::new("target")).is_none());
        assert!(index.lookup(Path::new("keep.rs")).is_some());

        write_file(&root.path().join(".gitignore"), b"");
        let exposed = reconcile(&mut index, &config, &mut |_| {}).expect("reconcile exposure");
        assert!(exposed.is_complete(), "{:?}", exposed.scan.errors);
        assert!(index.lookup(Path::new("target/deep/file.rs")).is_some());
        assert!(exposed.scan.dirs_read >= 3);

        write_file(&root.path().join(".gitignore"), b"target/\n");
        let excluded = reconcile(&mut index, &config, &mut |_| {}).expect("reconcile exclusion");
        assert!(excluded.is_complete(), "{:?}", excluded.scan.errors);
        assert!(index.lookup(Path::new("target")).is_none());
        assert_eq!(excluded.scan.dirs_read, 1, "ignored target was not reopened");
    }

    #[test]
    fn excluded_subtree_refresh_keeps_ignored_file_and_directory_out_of_scope() {
        let root = tempfile::tempdir().expect("root");
        write_file(&root.path().join(".gitignore"), b"target/\n*.log\n");
        write_file(&root.path().join("target/deep/file.rs"), b"code");
        write_file(&root.path().join("debug.log"), b"ignored");
        let config = ScanConfig {
            population: crate::query::IgnoredEntries::Exclude,
            ..ScanConfig::default()
        };
        let (mut index, cold) = scan_into_index(root.path(), &config).expect("cold scan");
        assert!(cold.is_complete());
        assert!(index.lookup(Path::new("target")).is_none());
        assert!(index.lookup(Path::new("debug.log")).is_none());

        for path in ["target", "debug.log"] {
            let refresh = reconcile_subtree(&mut index, Path::new(path), &config, &mut |_| {})
                .expect("subtree refresh");
            assert!(refresh.is_complete(), "{path}: {:?}", refresh.scan.errors);
            assert!(index.lookup(Path::new(path)).is_none(), "{path} is outside scope");
        }
        assert_eq!(index.total().dirs, 0);
    }

    #[test]
    fn handled_subtree_refresh_recovers_pruned_ancestry_after_control_edit() {
        let root = tempfile::tempdir().expect("root");
        write_file(&root.path().join(".gitignore"), b"target/\n");
        write_file(&root.path().join("target/deep/file.rs"), b"code");
        let config = ScanConfig {
            population: crate::query::IgnoredEntries::Exclude,
            ..ScanConfig::default()
        };
        let (index, cold) = scan_into_index(root.path(), &config).expect("cold scan");
        assert!(cold.is_complete());
        let handle = IndexHandle::new(index);

        let hidden = reconcile_subtree_handle(
            &handle,
            Path::new("target/deep/file.rs"),
            &config,
            &mut |_| {},
        )
        .expect("refresh pruned descendant");
        assert!(hidden.is_complete());
        assert!(
            !handle
                .read_with(|index| index.lookup(Path::new("target")).is_some())
                .expect("read after hidden refresh")
        );

        write_file(&root.path().join(".gitignore"), b"");
        let exposed = reconcile_subtree_handle(&handle, Path::new("target"), &config, &mut |_| {})
            .expect("refresh changed control");
        assert!(exposed.is_complete());
        assert!(
            handle
                .read_with(|index| index.lookup(Path::new("target/deep/file.rs")).is_some())
                .expect("read after control recovery")
        );

        write_file(&root.path().join(".gitignore"), b"target/\n");
        let hidden_again =
            reconcile_subtree_handle(&handle, Path::new("target"), &config, &mut |_| {})
                .expect("refresh restored control");
        assert!(hidden_again.is_complete());
        assert!(
            !handle
                .read_with(|index| index.lookup(Path::new("target")).is_some())
                .expect("read after restored control")
        );
    }

    #[cfg(unix)]
    #[test]
    fn targeted_refresh_keeps_unknown_population_below_unreadable_ancestor_control() {
        use std::os::unix::fs::PermissionsExt;

        if !crate::test_support::require_permission_bits() {
            return;
        }
        let root = tempfile::tempdir().expect("root");
        let control = root.path().join(".gitignore");
        write_file(&control, b"*.log\n");
        write_file(&root.path().join("a/keep.rs"), b"unknown membership");
        let config =
            ScanConfig { population: crate::query::IgnoredEntries::Only, ..ScanConfig::default() };
        fs::set_permissions(&control, fs::Permissions::from_mode(0o000)).expect("unreadable");
        let (mut index, cold) = scan_into_index(root.path(), &config).expect("cold scan");
        assert!(!cold.is_complete());
        assert!(index.lookup(Path::new("a/keep.rs")).is_some());
        assert_eq!(index.ignored_classification(Path::new("a/keep.rs")), None);

        let refreshed = reconcile_subtree(&mut index, Path::new("a/keep.rs"), &config, &mut |_| {});
        fs::set_permissions(&control, fs::Permissions::from_mode(0o644)).expect("restore control");
        let refreshed = refreshed.expect("targeted refresh");
        assert!(!refreshed.is_complete(), "unreadable governing control was not visited");
        assert!(index.lookup(Path::new("a/keep.rs")).is_some(), "unknown must stay retained");
        assert_eq!(index.ignored_classification(Path::new("a/keep.rs")), None);
    }

    #[test]
    fn exclusion_honors_nested_negation_and_refused_rule_changes() {
        let root = tempfile::tempdir().expect("root");
        write_file(&root.path().join("nested/.gitignore"), b"*.log\n!keep.log\n");
        write_file(&root.path().join("nested/keep.log"), b"negated");
        write_file(&root.path().join("nested/drop.log"), b"ignored");
        let config = ScanConfig {
            population: crate::query::IgnoredEntries::Exclude,
            control_limits: crate::control::ControlLimits {
                line_limit: Some(20),
                ..crate::control::ControlLimits::default()
            },
            ..ScanConfig::default()
        };
        let (mut index, cold) = scan_into_index(root.path(), &config).expect("cold");
        assert!(cold.is_complete(), "{:?}", cold.errors);
        assert!(index.lookup(Path::new("nested/keep.log")).is_some());
        assert!(index.lookup(Path::new("nested/drop.log")).is_none());

        write_file(&root.path().join("nested/.gitignore"), b"this-line-is-over-the-limit\n");
        let changed =
            reconcile(&mut index, &config, &mut |_| {}).expect("reconcile refused source");
        assert!(changed.is_complete(), "{:?}", changed.scan.errors);
        assert!(index.lookup(Path::new("nested/drop.log")).is_some(), "unknown cannot be pruned");
        assert_eq!(index.ignored_classification(Path::new("nested/drop.log")), None);
    }

    #[test]
    fn only_population_skips_nonignored_content_candidates_and_refusals_are_unknown() {
        let root = tempfile::tempdir().expect("root");
        write_file(&root.path().join(".gitignore"), b"*.log\n");
        write_file(&root.path().join("keep.rs"), b"code");
        write_file(&root.path().join("debug.log"), b"ignored");
        write_file(&root.path().join("nested/keep.rs"), b"kept");
        write_file(&root.path().join("nested/debug.log"), b"ignored below nonignored dir");
        let only =
            ScanConfig { population: crate::query::IgnoredEntries::Only, ..ScanConfig::default() };
        let (index, report) = scan_into_index(root.path(), &only).expect("scan");
        assert!(report.is_complete());
        assert!(index.lookup(Path::new("keep.rs")).is_none());
        assert!(index.lookup(Path::new("nested")).is_some());
        assert!(index.lookup(Path::new("nested/keep.rs")).is_none());
        assert!(index.lookup(Path::new("nested/debug.log")).is_some());
        let candidates = index.analysis_candidates(crate::content::AnalysisSet::NONE.with_lines());
        assert_eq!(candidates.len(), 2);
        assert!(
            candidates.iter().any(|candidate| candidate.relative_path == Path::new("debug.log"))
        );
        assert!(
            candidates
                .iter()
                .any(|candidate| candidate.relative_path == Path::new("nested/debug.log"))
        );

        let refused = ScanConfig {
            population: crate::query::IgnoredEntries::Exclude,
            control_limits: crate::control::ControlLimits {
                line_limit: Some(1),
                ..crate::control::ControlLimits::default()
            },
            ..ScanConfig::default()
        };
        let (index, report) = scan_into_index(root.path(), &refused).expect("refused scan");
        assert!(report.is_complete());
        assert!(index.lookup(Path::new("debug.log")).is_some(), "unknown is not pruned");
        assert_eq!(index.ignored_classification(Path::new("debug.log")), None);
        assert!(
            index.analysis_candidates(crate::content::AnalysisSet::NONE.with_lines()).is_empty()
        );
    }

    #[test]
    fn only_population_content_is_complete_with_retained_control_file() {
        let root = tempfile::tempdir().expect("root");
        write_file(&root.path().join(".gitignore"), b"vendor/\n");
        write_file(&root.path().join("main.rs"), b"fn main() {}\n");
        write_file(&root.path().join("vendor/lib.rs"), b"fn lib() {}\n");
        let config =
            ScanConfig { population: crate::query::IgnoredEntries::Only, ..ScanConfig::default() };
        let (mut index, scan) = scan_into_index(root.path(), &config).expect("scan");
        assert!(scan.is_complete());
        let profile = crate::content::AnalysisSet::NONE.with_lines().with_code();
        let analyzed = crate::content::analyze_index(
            &mut index,
            crate::content::AnalysisRequest {
                profile,
                ..crate::content::AnalysisRequest::default()
            },
        );
        assert!(analyzed.is_complete(), "{analyzed:?}");
        assert_eq!(analyzed.candidates, 1);
        assert!(!index.content_has_pending(profile));
    }

    #[test]
    fn only_population_reconciles_rule_changes_without_losing_traversal() {
        let root = tempfile::tempdir().expect("root");
        write_file(&root.path().join("nested/.gitignore"), b"*.log\n");
        write_file(&root.path().join("nested/first.log"), b"first");
        write_file(&root.path().join("nested/second.txt"), b"second");
        let config =
            ScanConfig { population: crate::query::IgnoredEntries::Only, ..ScanConfig::default() };
        let (mut index, cold) = scan_into_index(root.path(), &config).expect("cold");
        assert!(cold.is_complete());
        assert!(index.lookup(Path::new("nested")).is_some());
        assert!(index.lookup(Path::new("nested/first.log")).is_some());
        assert!(index.lookup(Path::new("nested/second.txt")).is_none());

        write_file(&root.path().join("nested/.gitignore"), b"*.txt\n");
        let changed = reconcile(&mut index, &config, &mut |_| {}).expect("rule change");
        assert!(changed.is_complete(), "{:?}", changed.scan.errors);
        assert!(index.lookup(Path::new("nested/first.log")).is_none());
        assert!(index.lookup(Path::new("nested/second.txt")).is_some());
    }

    #[cfg(unix)]
    #[test]
    fn excluded_special_objects_never_enter_cold_or_reconciled_facts() {
        use std::os::unix::net::UnixListener;

        let dir = tempfile::tempdir().expect("tempdir");
        let socket_path = dir.path().join("service.sock");
        let _listener = UnixListener::bind(&socket_path).expect("bind socket");
        write_file(&dir.path().join("replacement"), b"ordinary");
        crate::test_support::settle_allocations(dir.path());
        let (kept, kept_report) =
            scan_into_index(dir.path(), &ScanConfig::default()).expect("default scan");
        assert!(kept_report.is_complete());
        assert_eq!(kept.kind(Path::new("service.sock")), Some(EntryKind::Other));

        let serial_config =
            ScanConfig { exclude_special: true, threads: Some(1), ..ScanConfig::default() };
        let parallel_config =
            ScanConfig { exclude_special: true, threads: Some(4), ..ScanConfig::default() };
        let (mut serial, serial_report) =
            scan_into_index(dir.path(), &serial_config).expect("serial scan");
        let (mut parallel, parallel_report) =
            scan_into_index(dir.path(), &parallel_config).expect("parallel scan");

        for (index, report) in [(&serial, &serial_report), (&parallel, &parallel_report)] {
            assert!(report.is_complete(), "unexpected errors: {:?}", report.errors);
            assert!(index.lookup(Path::new("service.sock")).is_none());
            assert!(index.lookup(Path::new("replacement")).is_some());
        }

        fs::remove_file(dir.path().join("replacement")).expect("remove file");
        let _replacement =
            UnixListener::bind(dir.path().join("replacement")).expect("bind replacement socket");
        let serial_reconciled =
            reconcile(&mut serial, &serial_config, &mut |_| {}).expect("serial reconcile");
        let parallel_reconciled =
            reconcile(&mut parallel, &parallel_config, &mut |_| {}).expect("parallel reconcile");

        assert!(serial_reconciled.is_complete());
        assert!(parallel_reconciled.is_complete());
        assert!(serial.lookup(Path::new("replacement")).is_none());
        assert!(parallel.lookup(Path::new("replacement")).is_none());
        assert_eq!(index_fingerprint(&serial), index_fingerprint(&parallel));
    }

    #[test]
    fn control_sources_respect_a_single_operation_batch_bound() {
        let dir = tempfile::tempdir().expect("tempdir");
        write_file(&dir.path().join(".gitignore"), b"*.log\n");
        write_file(&dir.path().join("debug.log"), b"ignored");

        for threads in [1, 4] {
            let config = ScanConfig {
                read_controls: true,
                threads: Some(threads),
                batch_size: 1,
                ..ScanConfig::default()
            };
            let mut largest = 0;
            let report = scan(dir.path(), &config, &mut |observation| {
                largest = largest.max(observation.len());
            })
            .expect("scan");

            assert!(report.is_complete(), "unexpected errors: {:?}", report.errors);
            assert_eq!(largest, 1);
        }
    }

    #[test]
    fn cold_scan_matches_the_metabrowser_nested_control_fixture() {
        let dir = tempfile::tempdir().expect("tempdir");
        write_file(&dir.path().join(".gitignore"), b"node_modules/\n*.pyc\n");
        write_file(&dir.path().join("src/app.py"), b"x");
        write_file(&dir.path().join("src/thing.pyc"), b"x");
        write_file(&dir.path().join("src/generated/.gitignore"), b"*.gen\n");
        write_file(&dir.path().join("src/generated/out.gen"), b"x");
        write_file(&dir.path().join("node_modules/.gitignore"), b"!keep-me.py\n");
        write_file(&dir.path().join("node_modules/keep-me.py"), b"x");

        let (index, report) = scan_into_index(
            dir.path(),
            &ScanConfig { read_controls: true, threads: Some(4), ..ScanConfig::default() },
        )
        .expect("scan fixture");

        assert!(report.is_complete(), "unexpected errors: {:?}", report.errors);
        assert_eq!(
            index.is_ignored(Path::new("src/app.py")).expect("control state observed"),
            Some(false)
        );
        assert_eq!(
            index.is_ignored(Path::new("src/thing.pyc")).expect("control state observed"),
            Some(true)
        );
        assert_eq!(
            index.is_ignored(Path::new("src/generated")).expect("control state observed"),
            Some(false)
        );
        assert_eq!(
            index.is_ignored(Path::new("src/generated/out.gen")).expect("control state observed"),
            Some(true)
        );
        assert_eq!(
            index.is_ignored(Path::new("node_modules")).expect("control state observed"),
            Some(true)
        );
        assert_eq!(
            index.is_ignored(Path::new("node_modules/keep-me.py")).expect("control state observed"),
            Some(true)
        );
    }

    #[test]
    fn reconciliation_observes_same_metadata_control_edits_and_last_deletion() {
        let dir = tempfile::tempdir().expect("tempdir");
        write_file(&dir.path().join(".gitignore"), b"*.log\n");
        write_file(&dir.path().join("debug.log"), b"ignored");
        let config = ScanConfig { read_controls: true, threads: Some(1), ..ScanConfig::default() };
        let (mut index, report) = scan_into_index(dir.path(), &config).expect("scan");
        assert!(report.is_complete());
        assert_eq!(
            index.is_ignored(Path::new("debug.log")).expect("control state observed"),
            Some(true)
        );

        // Same-length content proves control identity is not inferred from stat-tier
        // metadata, which can remain unchanged on coarse filesystems.
        write_file(&dir.path().join(".gitignore"), b"*.tmp\n");
        let edited = reconcile(&mut index, &config, &mut |_| {}).expect("edit reconcile");
        assert!(edited.is_complete());
        assert_eq!(edited.apply.controls, 1);
        assert_eq!(edited.apply.reclassified, 1);
        assert!(
            index
                .controls()
                .expect("control state observed")
                .source_is(Path::new(".gitignore"), b"*.tmp\n")
        );
        assert_eq!(
            index.is_ignored(Path::new("debug.log")).expect("control state observed"),
            Some(false)
        );

        fs::remove_file(dir.path().join(".gitignore")).expect("remove control");
        let removed = reconcile(&mut index, &config, &mut |_| {}).expect("remove reconcile");
        assert!(removed.is_complete());
        assert_eq!(removed.apply.controls, 1);
        assert!(index.controls().expect("control state observed").is_empty());
        let partitions = index.partition_total().expect("control state observed");
        assert_eq!(partitions.all, partitions.unignored);
    }

    #[cfg(unix)]
    #[test]
    fn directory_entry_metadata_does_not_follow_symlinks() {
        use std::os::unix::fs::symlink;

        let root = tempfile::tempdir().expect("root");
        let outside = tempfile::tempdir().expect("outside");
        write_file(&outside.path().join("must-not-be-scanned.txt"), b"outside");
        symlink(outside.path(), root.path().join("link")).expect("symlink");

        let (index, report) = scan_into_index(root.path(), &ScanConfig::default()).expect("scan");

        assert!(report.is_complete(), "unexpected errors: {:?}", report.errors);
        assert_eq!(index.kind(Path::new("link")), Some(EntryKind::Symlink));
        assert!(index.lookup(Path::new("link/must-not-be-scanned.txt")).is_none());
        assert_eq!(index.total().files, 0);
    }

    #[test]
    fn cold_scan_establishes_a_baseline_without_change_history() {
        let dir = sample_tree();
        let (index, _) = scan_into_index(dir.path(), &ScanConfig::default()).expect("scan");

        assert_eq!(index.clock(), crate::Clock::ZERO);
        assert!(index.since(crate::Clock::ZERO).commits.is_empty());
    }

    #[test]
    fn max_depth_stops_descent() {
        let dir = sample_tree();
        let config = ScanConfig { max_depth: Some(1), ..ScanConfig::default() };
        let (index, _) = scan_into_index(dir.path(), &config).expect("scan");

        assert!(index.lookup(Path::new("src")).is_some());
        assert!(index.lookup(Path::new("src/main.rs")).is_none());
    }

    #[test]
    fn zero_max_depth_keeps_only_the_index_root() {
        let dir = sample_tree();
        let config = ScanConfig { max_depth: Some(0), ..ScanConfig::default() };
        let (index, report) = scan_into_index(dir.path(), &config).expect("scan");

        assert!(index.is_empty());
        assert_eq!(report.entries, 0);
        assert_eq!(report.dirs_read, 0);
    }

    #[test]
    fn direct_scan_records_the_canonical_root() {
        let dir = sample_tree();
        let aliased = dir.path().join(".");
        let (index, _) = scan_into_index(&aliased, &ScanConfig::default()).expect("scan");

        assert_eq!(index.root_path(), dir.path().canonicalize().expect("canonical root"));
    }

    #[test]
    fn unsupported_symlink_following_is_rejected_on_cold_and_warm_paths() {
        let dir = sample_tree();
        let unsupported = ScanConfig { follow_symlinks: true, ..ScanConfig::default() };
        assert!(matches!(
            scan_into_index(dir.path(), &unsupported),
            Err(Error::UnsupportedScanConfig(_))
        ));

        let (index, _) = scan_into_index(dir.path(), &ScanConfig::default()).expect("scan");
        assert!(matches!(
            revalidate(&index, &unsupported, &mut |_| {}),
            Err(Error::UnsupportedScanConfig(_))
        ));
    }

    #[test]
    fn revalidation_uses_the_same_depth_boundary_as_cold_scan() {
        let dir = sample_tree();
        let config = ScanConfig { max_depth: Some(1), ..ScanConfig::default() };
        let (mut index, _) = scan_into_index(dir.path(), &config).expect("scan");
        write_file(&dir.path().join("src/added-after-scan.txt"), b"new");

        let mut observations = Vec::new();
        revalidate(&index, &config, &mut |observation| observations.push(observation))
            .expect("revalidate");
        for observation in &observations {
            index.apply_ok(observation);
        }

        assert!(index.lookup(Path::new("src/added-after-scan.txt")).is_none());
    }

    #[test]
    fn zero_depth_revalidation_prunes_cached_root_children() {
        let dir = tempfile::tempdir().expect("tempdir");
        let config = ScanConfig { max_depth: Some(0), ..ScanConfig::default() };
        let mut index = Index::new_with_scope(dir.path(), config.scope());
        index.apply_baseline_ok(&Observation::new(vec![Op::Upsert {
            path: PathBuf::from("stale.txt"),
            kind: EntryKind::File,
            attrs: Attrs::default(),
        }]));

        let mut observations = Vec::new();
        let report = revalidate(&index, &config, &mut |observation| {
            observations.push(observation);
        })
        .expect("revalidate");
        for observation in &observations {
            index.apply_ok(observation);
        }

        assert!(index.is_empty());
        assert_eq!(report.dirs_read, 0);
    }

    #[test]
    fn zero_depth_applying_reconciliation_prunes_cached_root_children() {
        let dir = tempfile::tempdir().expect("tempdir");
        let config = ScanConfig { max_depth: Some(0), ..ScanConfig::default() };
        let mut index = Index::new_with_scope(dir.path(), config.scope());
        index.apply_baseline_ok(&Observation::new(vec![Op::Upsert {
            path: PathBuf::from("stale.txt"),
            kind: EntryKind::File,
            attrs: Attrs::default(),
        }]));

        let report = reconcile(&mut index, &config, &mut |_| {}).expect("reconcile");

        assert!(index.is_empty());
        assert_eq!(report.scan.dirs_read, 0);
    }

    #[test]
    fn filesystem_boundary_is_part_of_the_shared_descent_policy() {
        let config = ScanConfig { one_filesystem: true, ..ScanConfig::default() };
        let attrs = Attrs { dev: 22, ..Attrs::default() };
        assert!(!should_descend(EntryKind::Dir, attrs, 0, 11, &config));
        assert!(should_descend(EntryKind::Dir, Attrs { dev: 11, ..attrs }, 0, 11, &config,));
    }

    /// A cold scan's index records its own pass start, the stamp a snapshot of it writes:
    /// never earlier than an instant taken before the scan, so it is not a stale or zero
    /// stamp, and never later than one taken after it. The builder constructs the index,
    /// and so takes the stamp, before the walk begins.
    #[test]
    fn a_cold_scan_stamps_its_own_pass_start() {
        let nanos = || {
            i64::try_from(
                std::time::SystemTime::now()
                    .duration_since(std::time::UNIX_EPOCH)
                    .expect("after the epoch")
                    .as_nanos(),
            )
            .expect("nanoseconds")
        };
        let dir = sample_tree();
        let before = nanos();
        let (index, report) = scan_into_index(dir.path(), &ScanConfig::default()).expect("scan");
        let after = nanos();
        assert!(report.is_complete() && report.entries > 0, "{report:?}");
        let stamp = index.writing_pass_started_at_ns();
        assert!(before <= stamp && stamp <= after, "{before} <= {stamp} <= {after}");
    }

    #[test]
    fn scanning_a_file_is_an_error_not_a_panic() {
        let dir = sample_tree();
        let err = scan_into_index(&dir.path().join("a.txt"), &ScanConfig::default());
        assert!(err.is_err());
    }

    #[test]
    fn deltas_arrive_in_batches_of_the_configured_size() {
        let dir = tempfile::tempdir().expect("tempdir");
        for i in 0..25 {
            write_file(&dir.path().join(format!("f{i}.txt")), b"x");
        }
        let config = ScanConfig { batch_size: 10, ..ScanConfig::default() };
        let mut sizes = Vec::new();
        scan(dir.path(), &config, &mut |d| sizes.push(d.len())).expect("scan");

        assert!(sizes.len() >= 3, "expected several batches, got {sizes:?}");
        assert!(sizes.iter().all(|&n| n <= 10));
        assert_eq!(sizes.iter().sum::<usize>(), 25);
    }

    #[test]
    fn invalid_batch_sizes_are_rejected_before_allocation() {
        let zero = ScanConfig { batch_size: 0, ..ScanConfig::default() };
        let unbounded = ScanConfig { batch_size: usize::MAX, ..ScanConfig::default() };

        assert!(matches!(zero.validate(), Err(Error::UnsupportedScanConfig(_))));
        assert!(matches!(unbounded.validate(), Err(Error::UnsupportedScanConfig(_))));
    }

    #[test]
    fn reconciliation_scope_budget_publishes_before_returning_retry() {
        let directory = tempfile::tempdir().expect("root");
        let config = ScanConfig::default();
        let (index, _) = scan_into_index(directory.path(), &config).expect("scan root-only tree");
        let handle = IndexHandle::new(index);
        let mut started = false;
        let mut published_partial = false;
        let report = reconcile_handle(&handle, &config, &mut |commit| {
            if !started {
                started = true;
                // No entries are added: distinct absent children must not grow history
                // for the paused root pass without bound.
                for child in ["missing-a", "missing-b", "missing-c"] {
                    let nested = reconcile_subtree_handle(&handle, Path::new(child), &config, &mut |_| {}).expect("newer absent scope");
                    assert!(nested.is_complete());
                }
            }
            if commit.state.iter().any(|state| matches!(state,
                crate::StateTransition::IndexState { current, .. }
                    if current.coverage == crate::Coverage::Partial(crate::CoverageReason::Inaccessible))) {
                assert_eq!(handle.read_with(Index::state).expect("coherent state").coverage,
                    crate::Coverage::Partial(crate::CoverageReason::Inaccessible));
                published_partial = true;
            }
        }).expect("interrupted pass returns retryable report");
        assert!(!report.is_complete());
        assert!(report.retry_required);
        assert!(published_partial, "the transition precedes the caller's retry result");
        let recovered = reconcile_handle(&handle, &config, &mut |_| {}).expect("retry");
        assert!(recovered.is_complete());
        assert_eq!(
            handle.read_with(Index::state).expect("recovered state").coverage,
            crate::Coverage::Complete
        );
    }

    #[test]
    fn stale_arbitration_keeps_a_reconciliation_incomplete() {
        let report = ReconcileReport {
            scan: ScanReport::default(),
            apply: ApplyStats { stale: 1, ..ApplyStats::default() },
            observations: 1,
            ..ReconcileReport::default()
        };

        assert!(!report.is_complete());
    }

    #[test]
    fn portable_system_time_conversion_preserves_pre_epoch_values() {
        let before_epoch = std::time::UNIX_EPOCH
            // Windows timestamps have 100 ns granularity, so use a duration that every
            // supported platform can represent without rounding back to the epoch.
            .checked_sub(std::time::Duration::from_secs(1))
            .expect("represent pre-epoch fixture");

        assert_eq!(system_time_ns(before_epoch), -1_000_000_000);
        assert_eq!(system_time_ns(std::time::UNIX_EPOCH), 0);
    }

    #[cfg(not(unix))]
    #[test]
    fn one_filesystem_fails_when_device_identity_is_unavailable() {
        let config = ScanConfig { one_filesystem: true, ..ScanConfig::default() };

        assert!(matches!(config.validate(), Err(Error::UnsupportedScanConfig(_))));
    }

    #[test]
    fn revalidate_is_a_no_op_against_an_unchanged_tree() {
        let dir = sample_tree();
        let (mut index, _) = scan_into_index(dir.path(), &ScanConfig::default()).expect("scan");
        let before = index.total();

        let mut deltas = Vec::new();
        revalidate(&index, &ScanConfig::default(), &mut |d| deltas.push(d)).expect("revalidate");
        let mut unchanged = 0;
        for delta in &deltas {
            unchanged += index.apply_ok(delta).unchanged;
        }

        assert_eq!(unchanged, 5, "3 files + 2 dirs all already known");
        assert_eq!(index.total(), before);
    }

    #[cfg(windows)]
    #[test]
    fn windows_reconcile_detects_same_size_rewrite_with_preserved_mtime() {
        let root = tempfile::tempdir().expect("tempdir");
        let path = root.path().join("same.txt");
        write_file(&path, b"first");
        let modified = fs::metadata(&path).expect("metadata").modified().expect("mtime");
        let (mut index, _) =
            scan_into_index(root.path(), &ScanConfig::default()).expect("initial scan");
        let before = *index.attrs(Path::new("same.txt")).expect("initial attrs");

        // NTFS stamps change time from the system clock, which advances in ticks of up to
        // 15.625 ms, so a rewrite stamped in the same tick as the first write is
        // indistinguishable from it. Wait for the clock to leave that tick rather than for
        // a fixed interval; the precise clock `SystemTime` reads runs at most one tick ahead.
        let stamped = std::time::UNIX_EPOCH
            + std::time::Duration::from_nanos(
                u64::try_from(before.ctime_ns).expect("change time after the epoch"),
            );
        while std::time::SystemTime::now() <= stamped + std::time::Duration::from_millis(20) {
            std::thread::sleep(std::time::Duration::from_millis(5));
        }
        write_file(&path, b"other");
        File::options()
            .write(true)
            .open(&path)
            .expect("open rewritten file")
            .set_times(std::fs::FileTimes::new().set_modified(modified))
            .expect("restore mtime");
        reconcile(&mut index, &ScanConfig::default(), &mut |_| {}).expect("reconcile");

        let after = *index.attrs(Path::new("same.txt")).expect("rewritten attrs");
        assert_eq!((after.size, after.mtime_ns), (before.size, before.mtime_ns));
        assert_ne!(after.ctime_ns, before.ctime_ns, "change time detects the rewrite");
        assert_ne!(after.fingerprint(), before.fingerprint());
    }

    #[cfg(windows)]
    #[test]
    fn windows_reconcile_detects_path_identity_replacement() {
        let root = tempfile::tempdir().expect("tempdir");
        let path = root.path().join("replace.txt");
        let displaced = root.path().join("displaced.txt");
        write_file(&path, b"first");
        let modified = fs::metadata(&path).expect("metadata").modified().expect("mtime");
        let (mut index, _) =
            scan_into_index(root.path(), &ScanConfig::default()).expect("initial scan");
        let before = *index.attrs(Path::new("replace.txt")).expect("initial attrs");

        fs::rename(&path, &displaced).expect("retain old file identity");
        write_file(&path, b"other");
        File::options()
            .write(true)
            .open(&path)
            .expect("open replacement")
            .set_times(std::fs::FileTimes::new().set_modified(modified))
            .expect("restore mtime");
        reconcile(&mut index, &ScanConfig::default(), &mut |_| {}).expect("reconcile");

        let after = *index.attrs(Path::new("replace.txt")).expect("replacement attrs");
        assert_eq!((after.size, after.mtime_ns), (before.size, before.mtime_ns));
        assert_ne!(
            (after.dev, after.inode),
            (before.dev, before.inode),
            "volume serial and file index identify the replacement"
        );
        assert_ne!(after.fingerprint(), before.fingerprint());
    }

    #[test]
    fn direct_reconciliation_counts_unchanged_entries_and_publishes_state_commits() {
        let dir = sample_tree();
        let (mut index, _) = scan_into_index(dir.path(), &ScanConfig::default()).expect("scan");
        let before_total = index.total();
        let before_clock = index.clock();
        let mut commits = Vec::new();

        let report = reconcile(&mut index, &ScanConfig::default(), &mut |commit| {
            commits.push(commit.clone());
        })
        .expect("reconcile");

        assert!(report.is_complete());
        assert_eq!(report.apply.unchanged, 5, "3 files + 2 dirs all already known");
        assert_eq!(commits.len(), 2);
        assert!(commits.iter().all(|commit| commit.changes.is_empty()));
        assert_eq!(
            index.clock(),
            crate::Clock(before_clock.0 + 2),
            "start and finish are state commits"
        );
        let commits = index.since(before_clock).commits;
        assert_eq!(commits.len(), 2);
        assert!(commits.iter().all(|commit| commit.changes.is_empty()));
        assert_eq!(index.total(), before_total);
    }

    #[test]
    fn parallel_and_serial_reconciliation_produce_the_same_index() {
        let dir = sample_tree();
        let portable_config =
            ScanConfig { threads: Some(1), batch_size: 2, ..ScanConfig::default() };
        let (mut portable, _) =
            scan_into_index(dir.path(), &portable_config).expect("portable baseline");
        let mut bulk = portable.clone();

        fs::remove_file(dir.path().join("a.txt")).expect("remove file");
        write_file(&dir.path().join("src/main.rs"), b"fn main() { much longer }");
        write_file(&dir.path().join("src/added.md"), b"new file");
        crate::test_support::settle_allocations(dir.path());

        let portable_report = reconcile(&mut portable, &portable_config, &mut |_| {})
            .expect("portable reconciliation");
        let bulk_config = ScanConfig { threads: Some(2), ..portable_config };
        let bulk_report =
            reconcile(&mut bulk, &bulk_config, &mut |_| {}).expect("bulk reconciliation");

        assert!(portable_report.is_complete());
        assert!(bulk_report.is_complete());
        assert_eq!(bulk_report.scan.attribution, WalkAttribution::default());
        assert_eq!(bulk_report.scan.entries, portable_report.scan.entries);
        assert_eq!(bulk_report.scan.dirs_read, portable_report.scan.dirs_read);
        assert_eq!(bulk_report.apply, portable_report.apply);
        assert_eq!(index_fingerprint(&bulk), index_fingerprint(&portable));
        assert_eq!(bulk.total(), portable.total());
    }

    #[test]
    fn parallel_reconciliation_workers_publish_directory_counters() {
        let _serial = crate::counters::test_serial();
        let dir = sample_tree();
        let baseline = ScanConfig { threads: Some(1), ..ScanConfig::default() };
        let (mut index, scan) = scan_into_index(dir.path(), &baseline).expect("baseline scan");
        assert!(scan.is_complete());

        crate::counters::enable(true);
        crate::counters::reset();
        let config = ScanConfig { threads: Some(4), ..baseline };
        let report = reconcile(&mut index, &config, &mut |_| {}).expect("reconciliation");
        crate::counters::flush_thread();
        let counts = crate::counters::snapshot();
        crate::counters::reset();
        crate::counters::enable(false);

        assert!(report.is_complete());
        assert!(
            counts.dir_opens >= report.scan.dirs_read,
            "parallel worker directory opens were folded: {counts:?}, report={report:?}"
        );
    }

    #[test]
    fn parallel_reconciliation_matches_serial_across_structural_transitions() {
        for max_depth in [None, Some(1), Some(2)] {
            for order in [ScanOrder::BreadthFirst, ScanOrder::DepthFirst] {
                let dir = reconciliation_transition_tree();
                let reference_config = ScanConfig {
                    order,
                    max_depth,
                    threads: Some(1),
                    batch_size: 2,
                    ..ScanConfig::default()
                };
                let (baseline, baseline_report) =
                    scan_into_index(dir.path(), &reference_config).expect("baseline scan");
                assert!(baseline_report.is_complete());
                mutate_reconciliation_transition_tree(dir.path());

                let mut serial = baseline.clone();
                let mut serial_commits = Vec::new();
                let serial_report = reconcile(&mut serial, &reference_config, &mut |commit| {
                    serial_commits.push(commit.clone());
                })
                .expect("serial reconciliation");
                let (fresh, fresh_report) =
                    scan_into_index(dir.path(), &reference_config).expect("fresh oracle");
                assert!(serial_report.is_complete(), "serial {order:?}/{max_depth:?}");
                assert!(fresh_report.is_complete(), "fresh {order:?}/{max_depth:?}");
                assert_eq!(
                    index_fingerprint(&serial),
                    index_fingerprint(&fresh),
                    "serial did not converge to a fresh scan for {order:?}/{max_depth:?}"
                );

                for workers in [2, 4] {
                    let mut parallel = baseline.clone();
                    let config = ScanConfig { threads: Some(workers), ..reference_config.clone() };
                    let mut parallel_commits = Vec::new();
                    let report = reconcile(&mut parallel, &config, &mut |commit| {
                        parallel_commits.push(commit.clone());
                    })
                    .expect("parallel reconciliation");
                    let context = format!("{order:?}/{max_depth:?}/{workers} workers");

                    assert!(report.is_complete(), "{context}: unexpected partial report");
                    assert_eq!(report.scan.entries, serial_report.scan.entries, "{context}");
                    assert_eq!(report.scan.dirs_read, serial_report.scan.dirs_read, "{context}");
                    assert_eq!(report.apply, serial_report.apply, "{context}");
                    assert_eq!(
                        effective_ops(&parallel_commits),
                        effective_ops(&serial_commits),
                        "{context}: effective delta differs"
                    );
                    assert_eq!(
                        index_fingerprint(&parallel),
                        index_fingerprint(&serial),
                        "{context}: final index differs"
                    );
                    let (parallel_total, serial_total) = (parallel.total(), serial.total());
                    assert_eq!(
                        (
                            parallel_total.files,
                            parallel_total.dirs,
                            parallel_total.bytes,
                            parallel_total.allocated,
                            parallel_total.newest_mtime_ns,
                        ),
                        (
                            serial_total.files,
                            serial_total.dirs,
                            serial_total.bytes,
                            serial_total.allocated,
                            serial_total.newest_mtime_ns,
                        ),
                        "{context}: roll-up differs"
                    );
                    assert_eq!(
                        parallel_total.by_ext, serial_total.by_ext,
                        "{context}: extension roll-up differs"
                    );
                }
            }
        }
    }

    #[cfg(unix)]
    #[test]
    fn revalidation_metadata_errors_do_not_delete_enumerated_entries() {
        use std::os::unix::fs::PermissionsExt;

        if !crate::test_support::require_permission_bits() {
            return;
        }

        let dir = sample_tree();
        let config = ScanConfig::default();
        let (mut index, baseline_report) =
            scan_into_index(dir.path(), &config).expect("baseline scan");
        assert!(baseline_report.is_complete());
        let before = index_fingerprint(&index);
        let original_permissions = fs::metadata(dir.path()).expect("root metadata").permissions();

        fs::set_permissions(dir.path(), fs::Permissions::from_mode(0o400))
            .expect("remove search permission");
        let mut observations = Vec::new();
        let outcome = revalidate(&index, &config, &mut |observation| {
            observations.push(observation);
        });
        fs::set_permissions(dir.path(), original_permissions).expect("restore permissions");

        let report = outcome.expect("operational metadata errors are a partial report");
        assert!(!report.errors.is_empty(), "the fixture did not induce metadata errors");
        for observation in &observations {
            index.apply_ok(observation);
        }
        assert_eq!(index_fingerprint(&index), before);
        assert!(index.attrs(Path::new("a.txt")).is_some(), "existing entry was removed");
    }

    #[cfg(unix)]
    #[test]
    fn reconciliation_metadata_errors_drop_unverified_entries_like_a_cold_scan() {
        use std::os::unix::fs::PermissionsExt;

        if !crate::test_support::require_permission_bits() {
            return;
        }

        // macOS's parallel path may satisfy the whole directory through
        // getattrlistbulk even without search permission. One worker pins the portable
        // fallback there; Linux also exercises the parallel portable worker.
        let worker_counts = if cfg!(target_os = "macos") { vec![1] } else { vec![1, 2] };
        for workers in worker_counts {
            let dir = sample_tree();
            let config =
                ScanConfig { threads: Some(workers), batch_size: 2, ..ScanConfig::default() };
            let (mut index, baseline_report) =
                scan_into_index(dir.path(), &config).expect("baseline scan");
            assert!(baseline_report.is_complete());
            let before = index_fingerprint(&index);
            let original_permissions =
                fs::metadata(dir.path()).expect("root metadata").permissions();

            // Reading names requires read permission; looking up their metadata also
            // requires search permission. This makes enumeration succeed and each
            // metadata lookup fail, the boundary where an encountered name used to be
            // misclassified as a deletion.
            fs::set_permissions(dir.path(), fs::Permissions::from_mode(0o400))
                .expect("remove search permission");
            let outcome = reconcile(&mut index, &config, &mut |_| {});
            let cold = scan_into_index(dir.path(), &config);
            fs::set_permissions(dir.path(), original_permissions).expect("restore permissions");

            let report = outcome.expect("operational metadata errors are a partial report");
            let (cold, cold_report) = cold.expect("cold partial scan");
            assert!(!report.scan.errors.is_empty(), "the fixture did not induce metadata errors");
            assert!(!report.is_complete());
            assert!(!cold_report.is_complete());
            assert_eq!(index_fingerprint(&index), index_fingerprint(&cold));
            assert!(
                index_fingerprint(&index).is_empty(),
                "neither warm nor cold may retain attributes it could not verify"
            );
            assert_eq!(index.directory_complete(Path::new("")), Some(false));
            assert_eq!(cold.directory_complete(Path::new("")), Some(false));
            assert!(!before.is_empty(), "the fixture began with retained facts");
        }
    }

    #[cfg(unix)]
    #[test]
    fn failed_listing_withdraws_retained_completeness_and_recovers() {
        use std::os::unix::fs::PermissionsExt;

        if !crate::test_support::require_permission_bits() {
            return;
        }
        for workers in [1, 2] {
            let dir = tempfile::tempdir().expect("root");
            write_file(&dir.path().join("ancestor/blocked/unknown.txt"), b"unknown");
            write_file(&dir.path().join("healthy/known.txt"), b"known");
            let config = ScanConfig { threads: Some(workers), ..ScanConfig::default() };
            let (mut warm, baseline) = scan_into_index(dir.path(), &config).expect("baseline");
            assert!(baseline.is_complete());
            let blocked = dir.path().join("ancestor/blocked");
            fs::set_permissions(&blocked, fs::Permissions::from_mode(0o000)).expect("deny listing");
            let probe = fs::read_dir(&blocked);
            let mut commits = Vec::new();
            let refreshed =
                reconcile(&mut warm, &config, &mut |commit| commits.push(commit.clone()));
            let cold = scan_into_index(dir.path(), &config);
            fs::set_permissions(&blocked, fs::Permissions::from_mode(0o700)).expect("restore");
            assert_eq!(
                probe.expect_err("real denied listing").kind(),
                std::io::ErrorKind::PermissionDenied
            );
            assert!(!refreshed.expect("partial refresh").is_complete());
            let (cold, report) = cold.expect("partial cold scan");
            assert!(!report.is_complete());
            for index in [&warm, &cold] {
                for path in ["", "ancestor", "healthy"] {
                    assert_eq!(
                        index.directory_complete(Path::new(path)),
                        Some(true),
                        "workers={workers}: {path}"
                    );
                }
                assert_eq!(index.directory_complete(Path::new("ancestor/blocked")), Some(false));
                assert_eq!(index.freshness_at(Path::new("ancestor")), crate::Freshness::Partial);
            }
            assert!(commits.iter().flat_map(|commit| &commit.state).any(|state| matches!(state,
                crate::StateTransition::IndexState { current, .. } if current.coverage != crate::Coverage::Complete
            )), "failure is published");
            assert!(reconcile(&mut warm, &config, &mut |_| {}).expect("recovery").is_complete());
            assert_eq!(warm.directory_complete(Path::new("ancestor/blocked")), Some(true));
            assert_eq!(warm.state().coverage, crate::Coverage::Complete);
        }
    }

    #[test]
    fn unreadable_directory_warm_answer_matches_cold_verified_tree() {
        for workers in [1, 2] {
            let dir = tempfile::tempdir().expect("tempdir");
            write_file(&dir.path().join("blocked/old.txt"), b"old");
            write_file(&dir.path().join("verified.txt"), b"verified");
            crate::test_support::settle_allocations(dir.path());
            let config = ScanConfig { threads: Some(workers), ..ScanConfig::default() };
            let (mut warm, baseline) = scan_into_index(dir.path(), &config).expect("baseline");
            assert!(baseline.is_complete());

            let blocked = dir.path().join("blocked").canonicalize().expect("blocked path");
            let hook = install_child_metadata_hook(dir.path(), move |path| {
                (path == blocked)
                    .then(|| std::io::Error::from(std::io::ErrorKind::PermissionDenied))
            });
            let warm_report = reconcile(&mut warm, &config, &mut |_| {}).expect("warm partial");
            let (cold, cold_report) = scan_into_index(dir.path(), &config).expect("cold partial");
            drop(hook);

            assert!(!warm_report.is_complete(), "workers={workers}");
            assert!(!cold_report.is_complete(), "workers={workers}");
            assert!(warm.lookup(Path::new("blocked")).is_none(), "workers={workers}");
            assert!(warm.lookup(Path::new("blocked/old.txt")).is_none(), "workers={workers}");
            assert_eq!(index_fingerprint(&warm), index_fingerprint(&cold), "workers={workers}");
            assert!(warm.lookup(Path::new("verified.txt")).is_some(), "workers={workers}");
        }
    }

    #[test]
    fn deferred_change_overflow_retries_without_applying_a_partial_wave() {
        let dir = sample_tree();
        let config = ScanConfig { threads: Some(2), batch_size: 2, ..ScanConfig::default() };
        let (mut index, _) = scan_into_index(dir.path(), &config).expect("baseline");
        let before = index_fingerprint(&index);

        fs::remove_file(dir.path().join("a.txt")).expect("remove file");
        write_file(&dir.path().join("added.md"), b"new file");
        write_file(&dir.path().join("src/main.rs"), b"fn main() { much longer }");
        crate::test_support::settle_allocations(dir.path());

        let root = index.root_path().to_path_buf();
        let root_meta = {
            crate::counters::bump(|c| c.stats += 1);
            fs::symlink_metadata(&root)
        }
        .expect("root metadata");
        let mut commits = Vec::new();
        let outcome = reconcile_direct_parallel(
            &mut index,
            &root,
            root_device(&root, &root_meta).expect("root device"),
            &config,
            1,
            &mut |commit| commits.push(commit.clone()),
        )
        .expect("parallel attempt");
        let DirectParallelOutcome::RetrySerial { prefix, remaining } = outcome else {
            panic!("the deliberately tiny deferred budget must trigger the retry");
        };

        assert_eq!(prefix.apply, ApplyStats::default());
        assert_eq!(remaining, VecDeque::from([(PathBuf::new(), 0)]));
        assert!(commits.is_empty());
        assert_eq!(index_fingerprint(&index), before);

        let serial = ScanConfig { threads: Some(1), ..config };
        let report = reconcile(&mut index, &serial, &mut |_| {}).expect("serial retry");
        let (expected, expected_report) = scan_into_index(dir.path(), &serial).expect("oracle");
        assert!(report.is_complete());
        assert!(expected_report.is_complete());
        assert_eq!(index_fingerprint(&index), index_fingerprint(&expected));
        assert_eq!(index.total().files, expected.total().files);
        assert_eq!(index.total().dirs, expected.total().dirs);
        assert_eq!(index.total().bytes, expected.total().bytes);
        assert_eq!(index.total().allocated, expected.total().allocated);
        assert_eq!(index.total().newest_mtime_ns, expected.total().newest_mtime_ns);
        assert_eq!(index.total().by_ext, expected.total().by_ext);
    }

    #[test]
    fn late_overflow_resumes_without_double_counting_completed_waves() {
        let dir = tempfile::tempdir().expect("tempdir");
        // The root wave discovers more than one full wave of directories. A change in
        // the second wave then forces the serial fallback only after the first wave's
        // unchanged entries have already been counted.
        for directory in 0..=RECONCILE_WAVE_DIRECTORIES {
            write_file(&dir.path().join(format!("d{directory:04}/file.txt")), b"unchanged");
        }
        let parallel = ScanConfig { threads: Some(2), ..ScanConfig::default() };
        let (baseline, _) = scan_into_index(dir.path(), &parallel).expect("baseline");
        let mut candidate = baseline.clone();
        let mut serial_oracle = baseline;

        for directory in 0..=RECONCILE_WAVE_DIRECTORIES {
            write_file(
                &dir.path().join(format!("d{directory:04}/file.txt")),
                b"changed after the first wave",
            );
        }
        crate::test_support::settle_allocations(dir.path());

        let candidate_report = reconcile_target_inner(
            &mut ReconcileTarget::Direct(&mut candidate),
            Path::new(""),
            0,
            &parallel,
            0,
            &mut |_| {},
        )
        .expect("late-overflow reconciliation");
        let serial = ScanConfig { threads: Some(1), ..parallel };
        let oracle_report =
            reconcile(&mut serial_oracle, &serial, &mut |_| {}).expect("serial oracle");

        assert_eq!(candidate_report.apply, oracle_report.apply);
        assert_eq!(candidate_report.scan.entries, oracle_report.scan.entries);
        assert_eq!(candidate_report.scan.dirs_read, oracle_report.scan.dirs_read);
        assert_eq!(index_fingerprint(&candidate), index_fingerprint(&serial_oracle));
    }

    /// The four counts a walk reports, in the order [`crate::ProgressSnapshot`] shows them.
    fn walked(report: &ScanReport) -> (u64, u64, u64, u64) {
        (report.dirs_read, report.files_walked, report.bytes_walked, report.allocated_walked)
    }

    fn reported(progress: &crate::Progress) -> (u64, u64, u64, u64) {
        let snapshot = progress.snapshot();
        (snapshot.directories, snapshot.files, snapshot.bytes, snapshot.allocated)
    }

    /// Each walker is a separate loop with its own reporting sites, so each is checked:
    /// the detached cold walk, the streaming walk, the transient summary fold, the
    /// reference revalidation, exclusive reconciliation serial and in parallel waves,
    /// and shared-handle reconciliation. The tree is wide enough that every parallel
    /// walker claims several chunks and the small batch size fills several batches, so a
    /// walker that reported only its final state would still fail on the counts a
    /// mid-walk chunk added twice or not at all.
    #[test]
    fn every_walker_reports_exactly_what_its_report_counts() {
        let dir = tempfile::tempdir().expect("tempdir");
        for directory in 0..12 {
            for file in 0..5 {
                write_file(
                    &dir.path().join(format!("d{directory}/f{file}.txt")),
                    &vec![b'x'; directory * 5 + file + 1],
                );
            }
        }
        for threads in [1, 4] {
            let context = format!("threads={threads}");
            let progress = crate::Progress::new();
            let cold_config = ScanConfig {
                threads: Some(threads),
                batch_size: 4,
                progress: Some(progress.clone()),
                ..ScanConfig::default()
            };
            let (mut index, cold) = scan_into_index(dir.path(), &cold_config).expect("cold scan");
            assert_eq!(cold.dirs_read, 13, "{context}: the root and twelve children");
            assert_eq!(
                progress.snapshot().phase,
                crate::ProgressPhase::Indexing,
                "{context}: the detached walk ends by assembling the index"
            );
            assert_eq!(reported(&progress), walked(&cold), "{context}: detached cold walk");

            let progress = crate::Progress::new();
            let config = ScanConfig { progress: Some(progress.clone()), ..cold_config.clone() };
            let streamed = scan(dir.path(), &config, &mut |_| {}).expect("streaming scan");
            assert_eq!(walked(&streamed), walked(&cold), "{context}");
            assert_eq!(reported(&progress), walked(&streamed), "{context}: streaming walk");

            let progress = crate::Progress::new();
            let config = ScanConfig {
                read_controls: false,
                progress: Some(progress.clone()),
                ..cold_config.clone()
            };
            let folded = scan_summary_fold(dir.path(), &config, &mut |_| {}).expect("fold");
            assert_eq!(walked(&folded), walked(&cold), "{context}");
            assert_eq!(reported(&progress), walked(&folded), "{context}: summary fold");

            let progress = crate::Progress::new();
            let config = ScanConfig { progress: Some(progress.clone()), ..cold_config.clone() };
            let revalidated = revalidate(&index, &config, &mut |_| {}).expect("revalidate");
            assert_eq!(progress.snapshot().phase, crate::ProgressPhase::Revalidating, "{context}");
            assert_eq!(reported(&progress), walked(&revalidated), "{context}: revalidate");

            // Changes, so reconciliation defers and applies operations rather than
            // discarding every entry as unchanged.
            write_file(&dir.path().join(format!("d0/new{threads}.txt")), b"added");
            fs::remove_file(dir.path().join(format!("d1/f{}.txt", threads - 1))).expect("remove");
            write_file(&dir.path().join("d2/f0.txt"), &vec![b'y'; 40 + threads]);
            // An independent walk of the changed tree. The handle and the reconcile's own
            // report both come from the walker's counts, so agreeing with each other
            // would not show that the walker counted anything; agreeing with this does.
            let (_, fresh) = scan_into_index(dir.path(), &ScanConfig::default()).expect("fresh");
            let progress = crate::Progress::new();
            let config = ScanConfig { progress: Some(progress.clone()), ..cold_config.clone() };
            let reconciled = reconcile(&mut index, &config, &mut |_| {}).expect("reconcile");
            assert!(reconciled.apply.mutated(), "{context}: the changes were applied");
            assert_eq!(progress.snapshot().phase, crate::ProgressPhase::Revalidating, "{context}");
            assert_eq!(reported(&progress), walked(&reconciled.scan), "{context}: reconcile");
            assert_eq!(walked(&reconciled.scan), walked(&fresh), "{context}: the whole tree");

            let handle = crate::IndexHandle::new(index);
            let progress = crate::Progress::new();
            let config = ScanConfig { progress: Some(progress.clone()), ..cold_config };
            let shared = reconcile_handle(&handle, &config, &mut |_| {}).expect("shared");
            assert_eq!(reported(&progress), walked(&shared.scan), "{context}: shared handle");
        }
    }

    /// A wave that overflows its deferred-operation budget is thrown away and rewalked
    /// serially. The report counts each directory once, as the logical pass does, and
    /// progress counts the wave's reads both times, as the filesystem did them: work
    /// done, not the answer. This is the one walker relation that is not equality, and
    /// the difference is exactly the rewalked wave.
    #[test]
    fn progress_counts_a_rewalked_wave_twice_where_the_report_counts_it_once() {
        let dir = tempfile::tempdir().expect("tempdir");
        for directory in 0..=RECONCILE_WAVE_DIRECTORIES {
            write_file(&dir.path().join(format!("d{directory:04}/file.txt")), b"unchanged");
        }
        // A file in the wave that completes, so the serial rewalk has counts of that wave
        // to carry forward as already added rather than add again.
        write_file(&dir.path().join("root.txt"), b"counted by the wave that completes");
        let parallel = ScanConfig { threads: Some(2), ..ScanConfig::default() };
        let (mut index, _) = scan_into_index(dir.path(), &parallel).expect("baseline");
        // Larger than any filesystem stores inline in the inode, so each copy occupies
        // blocks of its own wherever the test runs.
        let changed = &vec![b'c'; 8_193];
        for directory in 0..=RECONCILE_WAVE_DIRECTORIES {
            write_file(&dir.path().join(format!("d{directory:04}/file.txt")), changed);
        }
        crate::test_support::settle_allocations(dir.path());

        let progress = crate::Progress::new();
        let observed = ScanConfig { progress: Some(progress.clone()), ..parallel };
        let report = reconcile_target_inner(
            &mut ReconcileTarget::Direct(&mut index),
            Path::new(""),
            0,
            &observed,
            0,
            &mut |_| {},
        )
        .expect("late-overflow reconciliation");

        // The root wave changes nothing and completes; the second wave holds exactly
        // one full wave of changed directories, overflows, and is rewalked with the one
        // directory the wave left behind.
        let rewalked = u64::try_from(RECONCILE_WAVE_DIRECTORIES).expect("fits");
        let snapshot = progress.snapshot();
        assert_eq!(snapshot.directories, report.scan.dirs_read + rewalked);
        assert_eq!(snapshot.files, report.scan.files_walked + rewalked);
        assert_eq!(
            snapshot.bytes,
            report.scan.bytes_walked + rewalked * u64::try_from(changed.len()).expect("fits")
        );
        let allocated = index.attrs(Path::new("d0000/file.txt")).expect("indexed").allocated;
        assert!(allocated > 0, "a file with content occupies blocks");
        assert_eq!(snapshot.allocated, report.scan.allocated_walked + rewalked * allocated);
    }

    /// Several invalidated roots are reconciled one at a time and their reports summed,
    /// so every walked count has to survive the sum, allocated bytes included.
    #[test]
    fn a_multi_root_reconcile_sums_every_walked_count() {
        let dir = tempfile::tempdir().expect("tempdir");
        for directory in ["a", "b", "c"] {
            for file in 0..3 {
                write_file(&dir.path().join(format!("{directory}/f{file}.txt")), b"before");
            }
        }
        let (mut index, _) = scan_into_index(dir.path(), &ScanConfig::default()).expect("scan");
        for directory in ["a", "b"] {
            write_file(&dir.path().join(format!("{directory}/f0.txt")), &vec![b'x'; 5_000]);
        }
        index.apply_ok(&Observation::new(
            ["a", "b"]
                .into_iter()
                .map(|directory| Op::InvalidateSubtree {
                    path: PathBuf::from(directory),
                    reason: crate::InvalidateReason::Requested,
                })
                .collect(),
        ));
        let report =
            reconcile_pending(&mut index, &ScanConfig::default(), &mut |_| {}).expect("reconcile");

        let attrs: Vec<Attrs> = ["a", "b"]
            .into_iter()
            .flat_map(|directory| (0..3).map(move |file| format!("{directory}/f{file}.txt")))
            .map(|path| *index.attrs(Path::new(&path)).expect("indexed"))
            .collect();
        assert_eq!(report.scan.files_walked, 6, "the two roots' files, and not c's");
        assert_eq!(report.scan.bytes_walked, attrs.iter().map(|attrs| attrs.size).sum::<u64>());
        assert_eq!(
            report.scan.allocated_walked,
            attrs.iter().map(|attrs| attrs.allocated).sum::<u64>()
        );
    }

    #[test]
    fn shared_reconciliation_retains_conditional_no_op_arbitration() {
        let dir = sample_tree();
        let (index, _) = scan_into_index(dir.path(), &ScanConfig::default()).expect("scan");
        let handle = crate::IndexHandle::new(index);
        let before_clock = handle.clock().expect("clock");
        let mut commits = Vec::new();

        let report = reconcile_handle(&handle, &ScanConfig::default(), &mut |commit| {
            commits.push(commit.clone());
        })
        .expect("reconcile");

        assert!(report.is_complete());
        assert_eq!(report.apply.unchanged, 5, "3 files + 2 dirs all already known");
        assert_eq!(commits.len(), 2);
        assert!(commits.iter().all(|commit| commit.changes.is_empty()));
        assert_eq!(
            handle.clock().expect("clock"),
            crate::Clock(before_clock.0 + 2),
            "start and finish are state commits"
        );
        let commits = handle.since(before_clock).expect("state commits").commits;
        assert_eq!(commits.len(), 2);
        assert!(commits.iter().all(|commit| commit.changes.is_empty()));
    }

    #[test]
    fn revalidate_detects_additions_edits_and_deletions() {
        let dir = sample_tree();
        let (mut index, _) = scan_into_index(dir.path(), &ScanConfig::default()).expect("scan");

        fs::remove_file(dir.path().join("a.txt")).expect("remove");
        write_file(&dir.path().join("src/main.rs"), b"fn main() { longer }");
        write_file(&dir.path().join("added.md"), b"new");

        let mut deltas = Vec::new();
        revalidate(&index, &ScanConfig::default(), &mut |d| deltas.push(d)).expect("revalidate");
        let mut stats = crate::index::ApplyStats::default();
        for delta in &deltas {
            let s = index.apply_ok(delta);
            stats.inserted += s.inserted;
            stats.updated += s.updated;
            stats.removed += s.removed;
        }

        assert_eq!(stats.inserted, 1, "added.md");
        assert_eq!(stats.updated, 1, "main.rs grew");
        assert_eq!(stats.removed, 1, "a.txt is gone");

        let total = index.total();
        assert_eq!(total.files, 3);
        assert_eq!(total.bytes, 20 + 9 + 3);
        assert!(!total.by_ext.contains_key(".txt"));
        assert_eq!(total.by_ext[".md"].files, 1);
    }

    #[test]
    fn revalidate_removes_a_whole_vanished_directory() {
        let dir = sample_tree();
        let (mut index, _) = scan_into_index(dir.path(), &ScanConfig::default()).expect("scan");
        fs::remove_dir_all(dir.path().join("src")).expect("remove dir");

        let mut deltas = Vec::new();
        revalidate(&index, &ScanConfig::default(), &mut |d| deltas.push(d)).expect("revalidate");
        for delta in &deltas {
            index.apply_ok(delta);
        }

        let total = index.total();
        assert_eq!(total.files, 1);
        assert_eq!(total.dirs, 0);
        assert!(index.lookup(Path::new("src")).is_none());
    }

    #[test]
    fn pending_invalidation_reconciles_the_requested_subtree() {
        let dir = sample_tree();
        let (mut index, _) = scan_into_index(dir.path(), &ScanConfig::default()).expect("scan");
        write_file(&dir.path().join("src/added.rs"), b"new");
        index.apply_ok(&Observation::new(vec![Op::InvalidateSubtree {
            path: PathBuf::from("src"),
            reason: crate::InvalidateReason::Requested,
        }]));
        assert_eq!(index.freshness_at(Path::new("src")), crate::Freshness::Stale);

        let mut applied = Vec::new();
        let report = reconcile_pending(&mut index, &ScanConfig::default(), &mut |delta| {
            applied.push(delta.clone());
        })
        .expect("reconcile pending");

        assert!(report.is_complete());
        assert!(index.lookup(Path::new("src/added.rs")).is_some());
        assert_eq!(index.freshness_at(Path::new("src")), crate::Freshness::Fresh);
        assert!(index.take_pending_invalidations().is_empty());
        assert!(applied.iter().any(|commit| commit_touches(commit, Path::new("src/added.rs"))));
    }

    /// A retained `.gitignore` reconciled as the root of its own walk re-reads its rules.
    /// A file does not descend, so the subtree-root branch was the only place that could
    /// read them, and it did not: the table kept `*.log` while the pass reported complete.
    #[test]
    fn reconciling_a_retained_control_file_as_the_subtree_root_rereads_its_rules() {
        let dir = tempfile::tempdir().expect("tempdir");
        write_file(&dir.path().join(".gitignore"), b"*.log\n");
        write_file(&dir.path().join("a.log"), b"log");
        let config = ScanConfig { read_controls: true, ..ScanConfig::default() };
        let (mut index, _) = scan_into_index(dir.path(), &config).expect("scan");
        assert_eq!(
            index.is_ignored(Path::new("a.log")).expect("control state observed"),
            Some(true)
        );

        write_file(&dir.path().join(".gitignore"), b"# nothing is ignored now\n");
        index.apply_ok(&Observation::new(vec![Op::InvalidateSubtree {
            path: PathBuf::from(".gitignore"),
            reason: crate::InvalidateReason::Requested,
        }]));
        let report = reconcile_pending(&mut index, &config, &mut |_| {}).expect("reconcile");

        assert!(report.is_complete(), "{:?}", report.scan.errors);
        assert_eq!(index.freshness_at(Path::new(".gitignore")), crate::Freshness::Fresh);
        assert_eq!(
            index.is_ignored(Path::new("a.log")).expect("control state observed"),
            Some(false)
        );
    }

    /// A control file the pass cannot verify contributes neither stale rules nor an entry.
    #[cfg(unix)]
    #[test]
    fn reconciling_an_unreadable_control_file_root_drops_its_rules_and_stays_partial() {
        use std::os::unix::fs::PermissionsExt;

        if !crate::test_support::require_permission_bits() {
            return;
        }

        let dir = tempfile::tempdir().expect("tempdir");
        let control = dir.path().join(".gitignore");
        write_file(&control, b"*.log\n");
        write_file(&dir.path().join("a.log"), b"log");
        let config = ScanConfig { read_controls: true, ..ScanConfig::default() };
        let (mut index, _) = scan_into_index(dir.path(), &config).expect("scan");

        write_file(&control, b"# rewritten, then made unreadable\n");
        fs::set_permissions(&control, fs::Permissions::from_mode(0o000)).expect("chmod");
        index.apply_ok(&Observation::new(vec![Op::InvalidateSubtree {
            path: PathBuf::from(".gitignore"),
            reason: crate::InvalidateReason::Requested,
        }]));
        let report = reconcile_pending(&mut index, &config, &mut |_| {});
        fs::set_permissions(&control, fs::Permissions::from_mode(0o644)).expect("restore");
        let report = report.expect("reconcile");

        assert!(!report.is_complete());
        assert_eq!(report.scan.errors.len(), 1, "{:?}", report.scan.errors);
        assert_eq!(index.freshness_at(Path::new(".gitignore")), crate::Freshness::Partial);
        assert!(
            !index.controls().expect("control state observed").contains(Path::new(".gitignore"))
        );
        assert_eq!(index.is_ignored(Path::new("a.log")).expect("control state observed"), None);
    }

    #[cfg(unix)]
    #[test]
    fn unreadable_control_keeps_new_excluded_file_unknown_and_out_of_analysis() {
        use std::os::unix::fs::PermissionsExt;

        if !crate::test_support::require_permission_bits() {
            return;
        }
        let dir = tempfile::tempdir().expect("tempdir");
        let control = dir.path().join(".gitignore");
        write_file(&control, b"*.log\n");
        write_file(&dir.path().join("keep.rs"), b"code");
        let config = ScanConfig {
            population: crate::query::IgnoredEntries::Exclude,
            ..ScanConfig::default()
        };
        let (mut index, cold) = scan_into_index(dir.path(), &config).expect("cold");
        assert!(cold.is_complete());
        write_file(&dir.path().join("debug.log"), b"must not analyze");
        fs::set_permissions(&control, fs::Permissions::from_mode(0o000)).expect("chmod");
        let report = reconcile(&mut index, &config, &mut |_| {});
        fs::set_permissions(&control, fs::Permissions::from_mode(0o644)).expect("restore");
        let report = report.expect("reconcile");
        assert!(!report.is_complete());
        assert!(index.lookup(Path::new("debug.log")).is_some(), "unknown is retained");
        assert_eq!(index.ignored_classification(Path::new("debug.log")), None);
        assert!(!index.ignored_classification_complete_below(Path::new("")));
        let candidates = index.analysis_candidates(crate::content::AnalysisSet::NONE.with_lines());
        assert!(
            candidates.iter().all(|candidate| candidate.relative_path != Path::new("debug.log"))
        );
        let repaired =
            reconcile(&mut index, &config, &mut |_| {}).expect("reconcile repaired control");
        assert!(repaired.is_complete(), "{:?}", repaired.scan.errors);
        assert!(index.ignored_classification_complete_below(Path::new("")));
        assert!(index.lookup(Path::new("debug.log")).is_none(), "known ignored file is pruned");
    }

    #[test]
    fn handle_reconciliation_publishes_after_each_delta_is_applied() {
        let dir = sample_tree();
        let (index, _) = scan_into_index(dir.path(), &ScanConfig::default()).expect("scan");
        let handle = crate::IndexHandle::new(index);
        let reader = handle.clone();
        write_file(&dir.path().join("added.md"), b"new");

        let mut observed_after_apply = false;
        reconcile_handle(&handle, &ScanConfig::default(), &mut |commit| {
            if commit_touches(commit, Path::new("added.md")) {
                observed_after_apply =
                    reader.kind(Path::new("added.md")).expect("query index").is_some();
            }
        })
        .expect("reconcile handle");

        assert!(observed_after_apply);
    }

    /// Delete `name` under `root` from inside its own metadata lookup, after the listing
    /// returned it, on whichever thread performs the lookup.
    fn delete_between_listing_and_stat(root: &Path, name: &'static str) -> WalkHookGuard {
        install_child_metadata_hook(root, move |path| {
            delete_if_named(path, name);
            None
        })
    }

    /// As [`delete_between_listing_and_stat`], and every reconciliation listing under `root`
    /// also ends in an error, so none of them is complete.
    fn delete_between_listing_and_stat_in_a_failing_listing(
        root: &Path,
        name: &'static str,
    ) -> WalkHookGuard {
        install_walk_hook(root, move |point| match point {
            WalkHookPoint::ChildMetadata(path) => {
                delete_if_named(path, name);
                None
            }
            WalkHookPoint::ListingEnd => Some(std::io::Error::other("injected listing error")),
            WalkHookPoint::ControlLookup(_) => None,
        })
    }

    fn delete_if_named(path: &Path, name: &str) {
        if path.file_name() == Some(OsStr::new(name)) {
            fs::remove_file(path).expect("delete between listing and stat");
        }
    }

    /// A name the listing returned that is gone by the time it is stat'd was deleted, on
    /// the serial path and on the parallel waves a full-root pass takes by default. The
    /// walk removes it; recorded as an error, it would settle as a phantom entry with
    /// permanent partial freshness.
    #[test]
    fn a_child_deleted_between_listing_and_stat_is_removed_rather_than_reported() {
        for threads in [1, 4] {
            let dir = tempfile::tempdir().expect("tempdir");
            write_file(&dir.path().join("keep.txt"), b"keep");
            write_file(&dir.path().join("gone.txt"), b"gone");
            let config = ScanConfig { threads: Some(threads), ..ScanConfig::default() };
            let (mut index, _) = scan_into_index(dir.path(), &config).expect("scan");
            assert!(index.lookup(Path::new("gone.txt")).is_some());

            index.apply_ok(&Observation::new(vec![Op::InvalidateSubtree {
                path: PathBuf::new(),
                reason: crate::InvalidateReason::Requested,
            }]));
            let hook = delete_between_listing_and_stat(dir.path(), "gone.txt");
            let report = reconcile_pending(&mut index, &config, &mut |_| {});
            drop(hook);
            let report = report.expect("reconcile");

            assert!(report.is_complete(), "threads {threads}: {:?}", report.scan.errors);
            assert_eq!(report.apply.removed, 1, "threads {threads}");
            assert!(index.lookup(Path::new("gone.txt")).is_none(), "threads {threads}");
            assert!(index.lookup(Path::new("keep.txt")).is_some(), "threads {threads}");
            assert_eq!(index.freshness(), crate::Freshness::Fresh, "threads {threads}");
        }
    }

    /// A vanished control file takes its rules with it on both reconcile paths, even when the
    /// rest of its listing fails: the stat's `NotFound` is the evidence. A retained file's
    /// rules go with its entry's removal; a hidden-pruned one has no entry to remove, so
    /// without its own removal its rules would go on ignoring its siblings.
    #[test]
    fn a_control_file_deleted_between_listing_and_stat_takes_its_rules_with_it() {
        for prune_hidden in [false, true] {
            for (threads, listing_fails) in [(1, false), (4, false), (1, true), (4, true)] {
                let case = format!(
                    "prune hidden {prune_hidden}, threads {threads}, listing fails {listing_fails}"
                );
                let dir = tempfile::tempdir().expect("tempdir");
                write_file(&dir.path().join(".gitignore"), b"*.log\n");
                write_file(&dir.path().join("a.log"), b"log");
                let config = ScanConfig {
                    read_controls: true,
                    threads: Some(threads),
                    hidden: prune_hidden
                        .then(|| std::sync::Arc::new(crate::HiddenPolicy::prune_hidden([""; 0]))),
                    ..ScanConfig::default()
                };
                let (mut index, _) = scan_into_index(dir.path(), &config).expect("scan");
                assert_eq!(
                    index.is_ignored(Path::new("a.log")).expect("control state observed"),
                    Some(true),
                    "{case}"
                );

                index.apply_ok(&Observation::new(vec![Op::InvalidateSubtree {
                    path: PathBuf::new(),
                    reason: crate::InvalidateReason::Requested,
                }]));
                let hook = if listing_fails {
                    delete_between_listing_and_stat_in_a_failing_listing(dir.path(), ".gitignore")
                } else {
                    delete_between_listing_and_stat(dir.path(), ".gitignore")
                };
                let report = reconcile_pending(&mut index, &config, &mut |_| {});
                drop(hook);
                let report = report.expect("reconcile");

                assert_eq!(
                    report.is_complete(),
                    !listing_fails,
                    "{case}: {:?}",
                    report.scan.errors
                );
                assert!(index.lookup(Path::new(".gitignore")).is_none(), "{case}");
                assert!(
                    !index
                        .controls()
                        .expect("control state observed")
                        .contains(Path::new(".gitignore")),
                    "{case}"
                );
                assert_eq!(
                    index.is_ignored(Path::new("a.log")).expect("control state observed"),
                    Some(false),
                    "{case}"
                );
            }
        }
    }

    /// A cold walk records a name gone by its stat as it records a name the listing never
    /// returned: not at all, and without an error that would make the walk partial.
    #[test]
    fn a_cold_walk_omits_a_child_deleted_between_listing_and_stat() {
        for threads in [1, 4] {
            let dir = tempfile::tempdir().expect("tempdir");
            write_file(&dir.path().join("keep.txt"), b"keep");
            write_file(&dir.path().join("gone.txt"), b"gone");
            let config = ScanConfig { threads: Some(threads), ..ScanConfig::default() };

            let hook = delete_between_listing_and_stat(dir.path(), "gone.txt");
            let scanned = scan_into_index_via_scanner(dir.path(), &config);
            drop(hook);
            let (index, report) = scanned.expect("scan");

            assert!(report.is_complete(), "threads {threads}: {:?}", report.errors);
            assert!(index.lookup(Path::new("gone.txt")).is_none(), "threads {threads}");
            assert!(index.lookup(Path::new("keep.txt")).is_some(), "threads {threads}");
        }
    }

    /// Revalidation emits the removal a reconciliation would apply.
    #[test]
    fn revalidation_removes_a_child_deleted_between_listing_and_stat() {
        let dir = tempfile::tempdir().expect("tempdir");
        write_file(&dir.path().join("keep.txt"), b"keep");
        write_file(&dir.path().join("gone.txt"), b"gone");
        let (mut index, _) = scan_into_index(dir.path(), &ScanConfig::default()).expect("scan");

        let hook = delete_between_listing_and_stat(dir.path(), "gone.txt");
        let mut observations = Vec::new();
        let report = revalidate(&index, &ScanConfig::default(), &mut |observation| {
            observations.push(observation);
        });
        drop(hook);
        let report = report.expect("revalidate");
        for observation in &observations {
            index.apply_ok(observation);
        }

        assert!(report.is_complete(), "{:?}", report.errors);
        assert!(index.lookup(Path::new("gone.txt")).is_none());
        assert!(index.lookup(Path::new("keep.txt")).is_some());
    }

    /// Revalidation emits the same rules removal when the rest of the listing fails.
    #[test]
    fn revalidation_removes_the_rules_of_a_control_file_deleted_in_a_failing_listing() {
        for prune_hidden in [false, true] {
            let dir = tempfile::tempdir().expect("tempdir");
            write_file(&dir.path().join(".gitignore"), b"*.log\n");
            write_file(&dir.path().join("a.log"), b"log");
            let config = ScanConfig {
                read_controls: true,
                hidden: prune_hidden
                    .then(|| std::sync::Arc::new(crate::HiddenPolicy::prune_hidden([""; 0]))),
                ..ScanConfig::default()
            };
            let (mut index, _) = scan_into_index(dir.path(), &config).expect("scan");
            assert_eq!(
                index.is_ignored(Path::new("a.log")).expect("control state observed"),
                Some(true)
            );

            let hook =
                delete_between_listing_and_stat_in_a_failing_listing(dir.path(), ".gitignore");
            let mut observations = Vec::new();
            let report = revalidate(&index, &config, &mut |observation| {
                observations.push(observation);
            });
            drop(hook);
            let report = report.expect("revalidate");
            for observation in &observations {
                index.apply_ok(observation);
            }

            assert!(!report.is_complete(), "prune hidden {prune_hidden}");
            assert!(index.lookup(Path::new(".gitignore")).is_none(), "prune hidden {prune_hidden}");
            assert!(
                !index
                    .controls()
                    .expect("control state observed")
                    .contains(Path::new(".gitignore")),
                "prune hidden {prune_hidden}"
            );
            assert_eq!(
                index.is_ignored(Path::new("a.log")).expect("control state observed"),
                Some(false),
                "prune hidden {prune_hidden}"
            );
        }
    }

    #[test]
    fn reconciliation_does_not_clear_a_newer_invalidation() {
        let dir = sample_tree();
        let (index, _) = scan_into_index(dir.path(), &ScanConfig::default()).expect("scan");
        let handle = crate::IndexHandle::new(index);
        write_file(&dir.path().join("added.md"), b"new");

        let invalidator = handle.clone();
        let mut saw_reconciling = false;
        reconcile_handle(&handle, &ScanConfig::default(), &mut |commit| {
            if commit_touches(commit, Path::new("added.md")) {
                saw_reconciling =
                    invalidator.freshness().expect("query") == crate::Freshness::Reconciling;
                invalidator
                    .apply(&Observation::new(vec![Op::InvalidateSubtree {
                        path: PathBuf::new(),
                        reason: crate::InvalidateReason::WatchOverflow,
                    }]))
                    .expect("new invalidation");
            }
        })
        .expect("reconcile handle");

        assert!(saw_reconciling);
        assert_eq!(handle.freshness().expect("query"), crate::Freshness::Stale);
    }

    #[test]
    fn failed_reconciliation_marks_the_scope_partial() {
        let dir = sample_tree();
        let (mut index, _) = scan_into_index(dir.path(), &ScanConfig::default()).expect("scan");
        fs::remove_dir_all(dir.path()).expect("remove root");

        assert!(reconcile(&mut index, &ScanConfig::default(), &mut |_| {}).is_err());
        assert_eq!(index.freshness(), crate::Freshness::Partial);
    }

    #[test]
    fn successful_subtree_retry_restores_complete_root_coverage() {
        let dir = tempfile::tempdir().expect("tempdir");
        write_file(&dir.path().join("blocked/known.txt"), b"known");
        let config = ScanConfig::default();
        let (mut index, report) = scan_into_index(dir.path(), &config).expect("scan");
        assert!(report.is_complete());
        let blocked = dir.path().join("blocked");
        let fault = install_walk_hook(&blocked, |_| {
            Some(std::io::Error::new(
                std::io::ErrorKind::PermissionDenied,
                "deterministic subtree refusal",
            ))
        });

        let failed = reconcile_subtree(&mut index, Path::new("blocked"), &config, &mut |_| {})
            .expect("partial");
        assert!(!failed.scan.is_complete());
        assert_eq!(
            index.state().coverage,
            crate::Coverage::Partial(crate::CoverageReason::Inaccessible)
        );
        drop(fault);

        let recovered = reconcile_subtree(&mut index, Path::new("blocked"), &config, &mut |_| {})
            .expect("retry");

        assert!(recovered.scan.is_complete());
        assert_eq!(index.state().coverage, crate::Coverage::Complete);
    }

    #[test]
    fn failed_pending_reconciliation_remains_queued_for_retry() {
        let dir = tempfile::tempdir().expect("tempdir");
        let (mut index, _) = scan_into_index(dir.path(), &ScanConfig::default()).expect("scan");
        index.apply_ok(&Observation::new(vec![Op::InvalidateSubtree {
            path: PathBuf::new(),
            reason: crate::InvalidateReason::Requested,
        }]));
        fs::remove_dir_all(dir.path()).expect("remove root");

        assert!(reconcile_pending(&mut index, &ScanConfig::default(), &mut |_| {}).is_err());
        assert_eq!(
            index.take_pending_invalidations(),
            vec![(PathBuf::new(), crate::InvalidateReason::Requested)]
        );
        assert_eq!(index.freshness(), crate::Freshness::Partial);
    }

    #[cfg(unix)]
    #[test]
    fn partial_cold_scan_keeps_verified_siblings_complete() {
        use std::os::unix::fs::PermissionsExt;
        if !crate::test_support::require_permission_bits() {
            return;
        }
        let root = tempfile::tempdir().expect("root");
        write_file(&root.path().join("blocked/unknown.txt"), b"unread");
        write_file(&root.path().join("healthy/nested/known.txt"), b"known");
        let blocked = root.path().join("blocked");
        fs::set_permissions(&blocked, fs::Permissions::from_mode(0o000)).expect("deny reads");
        let scan = ScanConfig::default();
        let detached = scan_into_index(root.path(), &scan);
        let streamed = scan_into_index_via_scanner(root.path(), &scan);
        fs::set_permissions(&blocked, fs::Permissions::from_mode(0o700)).expect("restore reads");
        for result in [detached, streamed] {
            let (index, report) = result.expect("partial scan still returns its facts");
            assert!(!report.is_complete(), "permission fixture must fail the blocked listing");
            assert!(
                index
                    .issues()
                    .iter()
                    .any(|issue| issue.path.as_deref() == Some(Path::new("blocked")))
            );
            assert_eq!(index.freshness_at(Path::new("")), crate::Freshness::Partial);
            assert_eq!(index.directory_complete(Path::new("")), Some(true));
            assert_eq!(index.directory_complete(Path::new("blocked")), Some(false));
            assert_eq!(index.freshness_at(Path::new("blocked")), crate::Freshness::Partial);
            assert!(index.lookup(Path::new("blocked/unknown.txt")).is_none());
            for sibling in ["healthy", "healthy/nested"] {
                assert_eq!(index.directory_complete(Path::new(sibling)), Some(true), "{sibling}");
                assert_eq!(
                    index.freshness_at(Path::new(sibling)),
                    crate::Freshness::Fresh,
                    "{sibling}"
                );
            }
            assert!(index.lookup(Path::new("healthy/nested/known.txt")).is_some());
            assert!(
                !crate::stored_state::entries_writable(&index),
                "partial root cannot persist metadata"
            );
        }
    }

    #[cfg(unix)]
    #[test]
    fn partial_pending_reconciliation_remains_queued_for_retry() {
        use std::os::unix::fs::PermissionsExt;

        if !crate::test_support::require_permission_bits() {
            return;
        }
        let dir = tempfile::tempdir().expect("tempdir");
        write_file(&dir.path().join("blocked/known.txt"), b"known");
        let (mut index, _) = scan_into_index(dir.path(), &ScanConfig::default()).expect("scan");
        let blocked = dir.path().join("blocked");
        fs::set_permissions(&blocked, fs::Permissions::from_mode(0o000)).expect("deny reads");
        index.apply_ok(&Observation::new(vec![Op::InvalidateSubtree {
            path: PathBuf::from("blocked"),
            reason: crate::InvalidateReason::VerificationFailed,
        }]));

        let report = reconcile_pending(&mut index, &ScanConfig::default(), &mut |_| {})
            .expect("permission failure is a partial report");
        let pending = index.take_pending_invalidations();
        fs::set_permissions(&blocked, fs::Permissions::from_mode(0o700)).expect("restore reads");
        assert!(!report.is_complete(), "permission fixture must make reconciliation partial");
        assert_eq!(
            pending,
            vec![(PathBuf::from("blocked"), crate::InvalidateReason::VerificationFailed)]
        );
        assert_eq!(index.freshness_at(Path::new("blocked")), crate::Freshness::Partial);
    }

    /// The shared API settles an unreadable subtree instead of queueing it again.
    ///
    /// Its per-event driver, `Watcher::apply_next`, drains after every event, so a retry
    /// re-walked the same unreadable subtree on each unrelated event, forever. The subtree
    /// stays partial and the report still names the error, once.
    #[cfg(unix)]
    #[test]
    fn partial_shared_pending_reconciliation_settles_instead_of_retrying() {
        use std::os::unix::fs::PermissionsExt;

        if !crate::test_support::require_permission_bits() {
            return;
        }
        let dir = tempfile::tempdir().expect("tempdir");
        write_file(&dir.path().join("blocked/known.txt"), b"known");
        let (index, _) = scan_into_index(dir.path(), &ScanConfig::default()).expect("scan");
        let handle = crate::IndexHandle::new(index);
        let blocked = dir.path().join("blocked");
        fs::set_permissions(&blocked, fs::Permissions::from_mode(0o000)).expect("deny reads");
        handle
            .apply(&Observation::new(vec![Op::InvalidateSubtree {
                path: PathBuf::from("blocked"),
                reason: crate::InvalidateReason::VerificationFailed,
            }]))
            .expect("invalidate");

        let report = reconcile_pending_handle(&handle, &ScanConfig::default(), &mut |_| {})
            .expect("permission failure is a partial report");
        let pending = handle.take_pending_invalidations().expect("pending");
        let freshness = handle.freshness_at(Path::new("blocked")).expect("freshness");
        fs::set_permissions(&blocked, fs::Permissions::from_mode(0o700)).expect("restore reads");
        assert!(!report.is_complete(), "permission fixture must make reconciliation partial");
        assert!(pending.is_empty(), "{pending:?}");
        assert_eq!(freshness, crate::Freshness::Partial);
        assert!(!report.scan.errors.is_empty());
    }

    #[test]
    fn pending_scope_mismatch_does_not_drain_the_retry_queue() {
        let dir = tempfile::tempdir().expect("tempdir");
        let shallow = ScanConfig { max_depth: Some(1), ..ScanConfig::default() };
        let (mut index, _) = scan_into_index(dir.path(), &shallow).expect("scan");
        index.apply_ok(&Observation::new(vec![Op::InvalidateSubtree {
            path: PathBuf::new(),
            reason: crate::InvalidateReason::Requested,
        }]));

        let error = reconcile_pending(&mut index, &ScanConfig::default(), &mut |_| {})
            .expect_err("mismatched scope must fail");

        assert!(matches!(error, Error::ScanScopeMismatch { .. }));
        assert_eq!(
            index.take_pending_invalidations(),
            vec![(PathBuf::new(), crate::InvalidateReason::Requested)]
        );
    }

    #[test]
    fn reconciliation_rejects_a_scope_mismatch_before_mutating() {
        let dir = tempfile::tempdir().expect("tempdir");
        write_file(&dir.path().join("deep/nested.txt"), b"nested");
        let shallow = ScanConfig { max_depth: Some(1), ..ScanConfig::default() };
        let (mut index, _) = scan_into_index(dir.path(), &shallow).expect("scan");
        assert!(index.lookup(Path::new("deep/nested.txt")).is_none());

        let error = reconcile(&mut index, &ScanConfig::default(), &mut |_| {})
            .expect_err("mismatched scope must fail");

        assert!(matches!(error, Error::ScanScopeMismatch { .. }));
        assert!(index.lookup(Path::new("deep/nested.txt")).is_none());
        assert_eq!(index.freshness(), crate::Freshness::Fresh);
    }

    #[test]
    fn subtree_reconciliation_rejects_a_path_beyond_the_depth_scope() {
        let dir = tempfile::tempdir().expect("tempdir");
        write_file(&dir.path().join("deep/nested.txt"), b"nested");
        let shallow = ScanConfig { max_depth: Some(1), ..ScanConfig::default() };
        let (mut index, _) = scan_into_index(dir.path(), &shallow).expect("scan");

        let result =
            reconcile_subtree(&mut index, Path::new("deep/nested.txt"), &shallow, &mut |_| {});

        assert!(matches!(result, Err(Error::SubtreeOutsideScanScope { .. })));
        assert!(index.lookup(Path::new("deep/nested.txt")).is_none());
        assert_eq!(index.freshness(), crate::Freshness::Fresh);
    }

    #[cfg(unix)]
    #[test]
    fn subtree_reconciliation_does_not_follow_an_ancestor_symlink() {
        use std::os::unix::fs::symlink;

        let root = tempfile::tempdir().expect("root");
        let outside = tempfile::tempdir().expect("outside");
        write_file(&outside.path().join("secret.txt"), b"secret");
        symlink(outside.path(), root.path().join("link")).expect("symlink");
        let config = ScanConfig::default();
        let (mut index, _) = scan_into_index(root.path(), &config).expect("scan");

        let result =
            reconcile_subtree(&mut index, Path::new("link/secret.txt"), &config, &mut |_| {});

        assert!(matches!(result, Err(Error::SubtreeOutsideScanScope { .. })));
        assert!(index.lookup(Path::new("link/secret.txt")).is_none());
        assert_eq!(index.freshness(), crate::Freshness::Fresh);
    }

    #[test]
    fn subtree_reconciliation_widens_to_a_non_directory_ancestor() {
        let root = tempfile::tempdir().expect("root");
        write_file(&root.path().join("parent/child.txt"), b"old");
        let config = ScanConfig::default();
        let (mut index, _) = scan_into_index(root.path(), &config).expect("scan");
        fs::remove_dir_all(root.path().join("parent")).expect("remove directory");
        write_file(&root.path().join("parent"), b"replacement");

        let report =
            reconcile_subtree(&mut index, Path::new("parent/child.txt"), &config, &mut |_| {})
                .expect("reconcile widened ancestor");

        assert!(report.is_complete());
        assert_eq!(index.kind(Path::new("parent")), Some(EntryKind::File));
        assert!(index.lookup(Path::new("parent/child.txt")).is_none());
        assert_eq!(index.freshness(), crate::Freshness::Fresh);
    }

    #[test]
    fn subtree_reconciliation_widens_to_a_missing_ancestor() {
        let root = tempfile::tempdir().expect("root");
        write_file(&root.path().join("parent/child.txt"), b"old");
        let config = ScanConfig::default();
        let (mut index, _) = scan_into_index(root.path(), &config).expect("scan");
        fs::remove_dir_all(root.path().join("parent")).expect("remove directory");

        let report =
            reconcile_subtree(&mut index, Path::new("parent/child.txt"), &config, &mut |_| {})
                .expect("reconcile widened ancestor");

        assert!(report.is_complete());
        assert!(index.lookup(Path::new("parent")).is_none());
        assert_eq!(index.freshness(), crate::Freshness::Fresh);
    }

    #[test]
    fn observation_only_revalidation_rejects_a_scope_mismatch() {
        let dir = tempfile::tempdir().expect("tempdir");
        let shallow = ScanConfig { max_depth: Some(1), ..ScanConfig::default() };
        let (index, _) = scan_into_index(dir.path(), &shallow).expect("scan");
        let mut observations = Vec::new();

        let error = revalidate(&index, &ScanConfig::default(), &mut |observation| {
            observations.push(observation);
        })
        .expect_err("mismatched scope must fail");

        assert!(matches!(error, Error::ScanScopeMismatch { .. }));
        assert!(observations.is_empty());
    }

    #[cfg(unix)]
    #[test]
    fn a_new_filesystem_boundary_prunes_cached_descendants() {
        use std::os::unix::fs::MetadataExt;

        let root = Path::new("/");
        let root_dev = {
            crate::counters::bump(|c| c.stats += 1);
            fs::symlink_metadata(root)
        }
        .expect("stat root")
        .dev();
        let Some(mount) = [Path::new("/dev"), Path::new("/proc"), Path::new("/sys")]
            .into_iter()
            .find(|candidate| {
                fs::symlink_metadata(candidate)
                    .is_ok_and(|metadata| metadata.is_dir() && metadata.dev() != root_dev)
            })
        else {
            return; // This host exposes no convenient cross-device directory.
        };
        let relative = mount.strip_prefix(root).expect("mount is below root");
        let stale_child = relative.join(".fdu-stale-snapshot-entry");
        let config = ScanConfig { one_filesystem: true, ..ScanConfig::default() };
        let mount_meta = fs::symlink_metadata(mount).expect("stat mount");
        let mut index = Index::new_with_scope(root, config.scope());
        index.apply_baseline_ok(&Observation::new(vec![
            Op::Upsert {
                path: relative.to_path_buf(),
                kind: EntryKind::Dir,
                attrs: attrs_from(mount, &mount_meta).expect("mount attrs"),
            },
            Op::Upsert {
                path: stale_child.clone(),
                kind: EntryKind::File,
                attrs: Attrs { size: 10, allocated: 10, ..Attrs::default() },
            },
        ]));

        let error = reconcile_subtree(&mut index, &stale_child, &config, &mut |_| {})
            .expect_err("a descendant below the mount boundary is outside scope");
        assert!(matches!(error, Error::SubtreeOutsideScanScope { .. }));
        assert!(index.lookup(&stale_child).is_some());

        reconcile_subtree(&mut index, relative, &config, &mut |_| {}).expect("reconcile mount");

        assert!(index.lookup(relative).is_some(), "the mount point itself stays visible");
        assert!(index.lookup(&stale_child).is_none(), "out-of-scope descendants are pruned");
    }

    #[test]
    fn subtree_reconciliation_rejects_paths_outside_the_root() {
        let dir = sample_tree();
        let (mut index, _) = scan_into_index(dir.path(), &ScanConfig::default()).expect("scan");

        assert!(matches!(
            reconcile_subtree(
                &mut index,
                Path::new("../outside"),
                &ScanConfig::default(),
                &mut |_| {},
            ),
            Err(Error::PathEscapesRoot(_))
        ));
        assert_eq!(index.freshness(), crate::Freshness::Fresh);
    }

    #[test]
    fn normalized_walk_errors_keep_index_and_one_shot_status_in_lockstep() {
        let root = Path::new("/root");
        let mut order: Vec<_> = (0..66).rev().collect();
        order.push(65);
        let mut errors = order
            .into_iter()
            .map(|number| {
                Error::io(
                    root.join(format!("file-{number:02}")),
                    std::io::Error::new(std::io::ErrorKind::PermissionDenied, "denied"),
                )
            })
            .collect::<Vec<_>>();

        normalize_walk_errors(root, &mut errors);
        assert_eq!(errors.len(), 66, "the repeated cause is removed once");

        let mut index = crate::Index::new(root);
        index.record_walk_errors(&mut errors);
        let status = crate::query::TreeStatus::of_walk(
            root,
            &mut ScanReport { errors, ..ScanReport::default() },
        );

        assert_eq!(status.errors, index.issues());
        assert_eq!(status.errors_omitted, index.state().issues.omitted);
        assert_eq!(status.errors.len(), crate::MAX_RETAINED_ISSUES);
        assert_eq!(status.errors_omitted, 2);
    }

    #[cfg(target_os = "linux")]
    #[test]
    fn scan_and_revalidate_keep_non_utf8_names_distinct() {
        use std::ffi::OsString;
        use std::os::unix::ffi::OsStringExt;

        let dir = tempfile::tempdir().expect("tempdir");
        let first = PathBuf::from(OsString::from_vec(vec![b'n', 0x80]));
        let second = PathBuf::from(OsString::from_vec(vec![b'n', 0x81]));
        write_file(&dir.path().join(&first), b"a");
        write_file(&dir.path().join(&second), b"bb");

        let (mut index, _) = scan_into_index(dir.path(), &ScanConfig::default()).expect("scan");
        assert_eq!(index.total().files, 2);
        assert_eq!(index.total().bytes, 3);

        fs::remove_file(dir.path().join(&first)).expect("remove first");
        let mut observations = Vec::new();
        revalidate(&index, &ScanConfig::default(), &mut |observation| {
            observations.push(observation);
        })
        .expect("revalidate");
        for observation in &observations {
            index.apply_ok(observation);
        }
        assert!(index.lookup(&first).is_none());
        assert!(index.lookup(&second).is_some());
        assert_eq!(index.total().bytes, 2);
    }
}